essays in public finance and industrial organization a dissertation ...
essays in public finance and industrial organization a dissertation ...
essays in public finance and industrial organization a dissertation ...
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
ESSAYS IN PUBLIC FINANCE<br />
AND<br />
INDUSTRIAL ORGANIZATION<br />
A DISSERTATION<br />
SUBMITTED TO THE DEPARTMENT OF ECONOMICS<br />
AND THE COMMITTEE ON GRADUATE STUDIES<br />
OF STANFORD UNIVERSITY<br />
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS<br />
FOR THE DEGREE OF<br />
DOCTOR OF PHILOSOPHY<br />
Neale Mahoney<br />
May 2011
© 2011 by Neale Ashok Mahoney. All Rights Reserved.<br />
Re-distributed by Stanford University under license with the author.<br />
This work is licensed under a Creative Commons Attribution-<br />
Noncommercial 3.0 United States License.<br />
http://creativecommons.org/licenses/by-nc/3.0/us/<br />
This <strong>dissertation</strong> is onl<strong>in</strong>e at: http://purl.stanford.edu/fx992pz6459<br />
ii
I certify that I have read this <strong>dissertation</strong> <strong>and</strong> that, <strong>in</strong> my op<strong>in</strong>ion, it is fully adequate<br />
<strong>in</strong> scope <strong>and</strong> quality as a <strong>dissertation</strong> for the degree of Doctor of Philosophy.<br />
Jonathan Lev<strong>in</strong>, Primary Adviser<br />
I certify that I have read this <strong>dissertation</strong> <strong>and</strong> that, <strong>in</strong> my op<strong>in</strong>ion, it is fully adequate<br />
<strong>in</strong> scope <strong>and</strong> quality as a <strong>dissertation</strong> for the degree of Doctor of Philosophy.<br />
Liran E<strong>in</strong>av<br />
I certify that I have read this <strong>dissertation</strong> <strong>and</strong> that, <strong>in</strong> my op<strong>in</strong>ion, it is fully adequate<br />
<strong>in</strong> scope <strong>and</strong> quality as a <strong>dissertation</strong> for the degree of Doctor of Philosophy.<br />
Approved for the Stanford University Committee on Graduate Studies.<br />
Carol<strong>in</strong>e Hoxby<br />
Patricia J. Gumport, Vice Provost Graduate Education<br />
This signature page was generated electronically upon submission of this <strong>dissertation</strong> <strong>in</strong><br />
electronic format. An orig<strong>in</strong>al signed hard copy of the signature page is on file <strong>in</strong><br />
University Archives.<br />
iii
Preface<br />
This <strong>dissertation</strong> has four chapters. The first three chapters exam<strong>in</strong>e health <strong>in</strong>surance<br />
markets <strong>in</strong> the U.S., focus<strong>in</strong>g <strong>in</strong> particular on contexts where there are important<br />
<strong>in</strong>teractions between health <strong>in</strong>surance plans. The fourth chapter is on the U.S. budget,<br />
exam<strong>in</strong><strong>in</strong>g the implications of annual budget cycles on the quantity <strong>and</strong> quality of<br />
end-of-year spend<strong>in</strong>g.<br />
Chapter 1, entitled “Bankruptcy as Implicit Health Insurance” exam<strong>in</strong>es the <strong>in</strong>ter-<br />
action between health <strong>in</strong>surance <strong>and</strong> the implicit <strong>in</strong>surance that people have because<br />
they can file (or threaten to file) for bankruptcy. With a simple model that captures<br />
key <strong>in</strong>stitutional features, I demonstrate that the f<strong>in</strong>ancial risk from medical shocks<br />
is capped by the assets that could be seized <strong>in</strong> bankruptcy. For households with mod-<br />
est seizable assets, this implicit “bankruptcy <strong>in</strong>surance” can crowd out conventional<br />
health <strong>in</strong>surance. I test these predictions us<strong>in</strong>g variation <strong>in</strong> the state laws that specify<br />
the type <strong>and</strong> level of assets that can be seized <strong>in</strong> bankruptcy. Because of the differ<strong>in</strong>g<br />
laws, people who have the same assets <strong>and</strong> receive the same medical care face different<br />
losses <strong>in</strong> bankruptcy. Exploit<strong>in</strong>g the variation <strong>in</strong> seizable assets that is orthogonal to<br />
wealth <strong>and</strong> other household characteristics, I show that households with fewer seiz-<br />
able assets are more likely to be un<strong>in</strong>sured. This f<strong>in</strong>d<strong>in</strong>g is consistent with another:<br />
un<strong>in</strong>sured households with fewer seizable assets end up mak<strong>in</strong>g lower out-of-pocket<br />
medical payments. The estimates suggest that if the laws of the least debtor-friendly<br />
state of Delaware were applied nationally, 16.3 percent of the un<strong>in</strong>sured would buy<br />
health <strong>in</strong>surance. Achiev<strong>in</strong>g the same <strong>in</strong>crease <strong>in</strong> coverage would require a premium<br />
subsidy of approximately 44.0 percent. To shed light on puzzles <strong>in</strong> the literature <strong>and</strong><br />
exam<strong>in</strong>e policy counterfactuals, I calibrate a utility-based, micro-simulation model of<br />
iv
<strong>in</strong>surance choice. Among other th<strong>in</strong>gs, simulations show that “bankruptcy <strong>in</strong>surance”<br />
expla<strong>in</strong>s the low take-up of high-deductible health <strong>in</strong>surance.<br />
Chapter 2, entitled “Pric<strong>in</strong>g <strong>and</strong> Welfare <strong>in</strong> Health Plan Choice”, is coauthored<br />
with M. Kate Bundorf <strong>and</strong> Jonathan Lev<strong>in</strong>. The start<strong>in</strong>g po<strong>in</strong>t for the paper is the<br />
simple observation that when <strong>in</strong>surance premiums do not reflect <strong>in</strong>dividual differences<br />
<strong>in</strong> expected costs, consumers may choose plans <strong>in</strong>efficiently. We study this problem <strong>in</strong><br />
health <strong>in</strong>surance markets, a sett<strong>in</strong>g <strong>in</strong> which prices often do not <strong>in</strong>corporate observable<br />
differences <strong>in</strong> expected costs. We develop a simple model <strong>and</strong> estimate it us<strong>in</strong>g data<br />
on small employers. In this sett<strong>in</strong>g, the welfare loss compared to the feasible risk-<br />
rated benchmark is around 2-11% of coverage costs. Three-quarters of this is due<br />
to restrictions on risk-rat<strong>in</strong>g employee contributions; the rest is due to <strong>in</strong>efficient<br />
contribution choices. Despite the <strong>in</strong>efficiency, the benefits from plan choice relative<br />
to each of the s<strong>in</strong>gle-plan options are substantial.<br />
Chapter 3, entitled “The Private Coverage <strong>and</strong> Public Costs: Identify<strong>in</strong>g the Effect<br />
of Private Supplemental Insurance on Medicare Spend<strong>in</strong>g,” is coauthored with Marika<br />
Cabral. While most elderly Americans have health <strong>in</strong>surance coverage through Medi-<br />
care, traditional Medicare policies leave <strong>in</strong>dividuals exposed to significant f<strong>in</strong>ancial<br />
risk. Private supplemental <strong>in</strong>surance to “fill the gaps” of Medicare, known as Medi-<br />
gap, is very popular. In this Chapter, we estimate the impact of this supplemental<br />
<strong>in</strong>surance on total medical spend<strong>in</strong>g us<strong>in</strong>g an <strong>in</strong>strumental variables strategy that<br />
leverages discont<strong>in</strong>uities <strong>in</strong> Medigap premiums at state boundaries. Our estimates<br />
suggest that Medigap <strong>in</strong>creases medical spend<strong>in</strong>g by 57 percent—or about 40 percent<br />
more than previous estimates. Back-of-the-envelope calculations <strong>in</strong>dicate that a 20<br />
percent tax on premiums would generate comb<strong>in</strong>ed revenue <strong>and</strong> sav<strong>in</strong>gs of 6.2 percent<br />
of basel<strong>in</strong>e costs; a Pigovian tax that fully accounts for the fiscal externality would<br />
yield sav<strong>in</strong>gs of 18.1 percent.<br />
Chapter 4, entitled “Do Expir<strong>in</strong>g Budgets Lead to Wasteful Year-End Spend<strong>in</strong>g?<br />
Evidence from Federal Procurement,” is coauthored with Jeffrey Liebman. Many<br />
<strong>organization</strong>s fund their spend<strong>in</strong>g out of a fixed budget that expires at year’s end.<br />
Faced with uncerta<strong>in</strong>ty over future spend<strong>in</strong>g dem<strong>and</strong>s, these <strong>organization</strong>s have an<br />
<strong>in</strong>centive to build a buffer stock of funds over the front end of the budget cycle.<br />
v
When dem<strong>and</strong> does not materialize, they then rush to spend these funds on lower<br />
quality projects at the end of the year. We test these predictions us<strong>in</strong>g data on<br />
procurement spend<strong>in</strong>g by the U.S. federal government. Us<strong>in</strong>g data on all federal<br />
contracts from 2004 through 2009, we document that spend<strong>in</strong>g spikes <strong>in</strong> all major<br />
federal agencies dur<strong>in</strong>g the 52nd week of the year as the agencies rush to exhaust<br />
expir<strong>in</strong>g budget authority. Spend<strong>in</strong>g <strong>in</strong> the last week of the year is 4.9 times higher<br />
than the rest-of-the-year weekly average. We exam<strong>in</strong>e the relative quality of year-end<br />
spend<strong>in</strong>g us<strong>in</strong>g a newly available dataset that tracks the quality of $130 billion <strong>in</strong><br />
<strong>in</strong>formation technology (I.T.) projects made by federal agencies. Consistent with the<br />
model, average project quality falls at the end of the year. Quality scores <strong>in</strong> the last<br />
week of the year are 2.2 to 5.6 times more likely to be below the central value. To<br />
explore the impact of allow<strong>in</strong>g agencies to roll unused spend<strong>in</strong>g over <strong>in</strong>to subsequent<br />
fiscal years, we study the I.T. contracts of an agency with special authority to roll<br />
over unused fund<strong>in</strong>g. We show that there is only a small end-of-year I.T. spend<strong>in</strong>g<br />
spike <strong>in</strong> this agency <strong>and</strong> that the one major I.T. contract this agency issued <strong>in</strong> the<br />
52nd week of the year has a quality rat<strong>in</strong>g that is well above average.<br />
vi
Acknowledgements<br />
I am deeply thankful to my <strong>dissertation</strong> committee for their advice <strong>and</strong> support. I<br />
thank Jonathan Lev<strong>in</strong> for his mentorship. I could not ask for a better adviser or role<br />
model. I thank Carol<strong>in</strong>e Hoxby <strong>and</strong> Liran E<strong>in</strong>av for their exceptional guidance. They<br />
say that your graduate advisers are your advisers for life. Hav<strong>in</strong>g these advisers for<br />
my years at Stanford is already more than I deserve.<br />
I am also grateful to the many faculty members who provided valuable comments<br />
<strong>and</strong> suggestions for my research dur<strong>in</strong>g my graduate studies. Among them are Ran<br />
Abramitzky, Doug Bernheim, Jay Bhattacharya, Nick Bloom, Tim Bresnahan, Kate<br />
Bundorf, Jacub Kastl, <strong>and</strong> Jeffrey Leibman. I thank Kyna Fong for jo<strong>in</strong><strong>in</strong>g my oral<br />
exam committee.<br />
I acknowledge the Ric Weil<strong>and</strong> Graduate Fellowship, Kapnick Fellowship, <strong>and</strong><br />
Shultz Scholarship for f<strong>in</strong>ancial support.<br />
My time at Stanford has been a case study <strong>in</strong> the virtuous cycle of positive peer<br />
effects. I thank Albert Bollard <strong>and</strong> Max Floetotto for their friendship. I am lucky to<br />
have had Aaron Bodoh-Creed, Marika Cabral, Alex Hirsch, Carlos Lever, Marcello<br />
Miccoli, Izi S<strong>in</strong>, <strong>and</strong> Gui Woolston as colleagues <strong>and</strong> friends. I thank Aless<strong>and</strong>ra<br />
Voena most of all.<br />
I am also grateful to my advisers at Brown University for <strong>in</strong>troduc<strong>in</strong>g me to the<br />
practice of economic research. I thank Herschel Grossman for countless hours of<br />
debate <strong>and</strong> discussion <strong>and</strong> Roberto Serrano for his wisdom <strong>in</strong> urg<strong>in</strong>g me to pursue<br />
graduate studies.<br />
F<strong>in</strong>ally, I am immensely thankful to my family, my parents Raymond Mahoney<br />
<strong>and</strong> Nal<strong>in</strong>i Shah-Mahoney <strong>and</strong> my brother Col<strong>in</strong>, for their unconditional love <strong>and</strong><br />
vii
support <strong>and</strong> the abundant happ<strong>in</strong>ess they br<strong>in</strong>g to my life.<br />
viii
Contents<br />
Preface iv<br />
Acknowledgements vii<br />
1 Bankruptcy as Implicit Health Insurance 1<br />
1.1 Introduction ................................ 1<br />
1.2 Bankruptcy as a Form of High-Deductible Health Insurance ..... 6<br />
1.2.1 Institutional Background ..................... 6<br />
1.2.2 A Model of Bankruptcy as High-Deductible Health Insurance . 8<br />
1.3 Data Overview .............................. 11<br />
1.3.1 Asset Exemptions ......................... 12<br />
1.3.2 Seizable Assets .......................... 12<br />
1.3.3 Medical Costs ........................... 14<br />
1.3.4 Insurance Premiums ....................... 15<br />
1.4 Empirical Strategy ............................ 16<br />
1.4.1 Coverage Equation ........................ 16<br />
1.4.2 Costs Equation .......................... 16<br />
1.4.3 Cross-State Variation ....................... 17<br />
1.4.4 Historical Homestead Exemptions ................ 19<br />
1.5 The Effect on Insurance Coverage .................... 21<br />
1.5.1 First Stage Estimates ....................... 21<br />
1.5.2 Basel<strong>in</strong>e Coverage Estimates ................... 22<br />
1.5.3 Sensitivity Analysis ........................ 23<br />
ix
1.5.4 Heterogeneity <strong>in</strong> the Effect on Coverage ............ 24<br />
1.5.5 The Effect on Wealth <strong>and</strong> Premiums .............. 25<br />
1.5.6 Summary <strong>and</strong> Interpretation ................... 26<br />
1.6 The Effect on Costs ............................ 27<br />
1.6.1 Basel<strong>in</strong>e Estimates ........................ 27<br />
1.6.2 Sensitivity Analysis ........................ 28<br />
1.6.3 Summary <strong>and</strong> Interpretation ................... 29<br />
1.7 Micro-Simulation Model ......................... 29<br />
1.8 Puzzles <strong>and</strong> Policy ............................ 31<br />
1.8.1 Puzzles ............................... 31<br />
1.8.2 Policy Implications ........................ 33<br />
1.9 Conclusion ................................. 35<br />
1.10 Appendix to Chapter 1 .......................... 36<br />
1.10.1 Model Predictions ......................... 36<br />
1.10.2 Seizable Assets Calculation Details ............... 37<br />
1.10.3 Micro-Simulation Details ..................... 38<br />
2 Pric<strong>in</strong>g <strong>and</strong> Welfare <strong>in</strong> Health Plan Choice 65<br />
2.1 Introduction ................................ 65<br />
2.2 Health Plan Pric<strong>in</strong>g <strong>and</strong> Market Efficiency ............... 68<br />
2.3 Data <strong>and</strong> Environment .......................... 72<br />
2.3.1 Institutional Sett<strong>in</strong>g ....................... 72<br />
2.3.2 Data <strong>and</strong> Descriptive Statistics ................. 74<br />
2.4 Econometric Model ............................ 77<br />
2.4.1 Consumer Preferences, Plan Costs <strong>and</strong> Market Behavior ... 77<br />
2.4.2 Discussion of Model <strong>and</strong> Identification ............. 82<br />
2.4.3 Estimation Strategy ....................... 85<br />
2.4.4 Welfare Measurement ....................... 86<br />
2.5 Empirical Results ............................. 87<br />
2.5.1 Model Estimates ......................... 88<br />
2.5.2 Quantify<strong>in</strong>g Social Welfare Inefficiencies ............ 93<br />
x
2.5.3 The Value of Plan Choice .................... 96<br />
2.5.4 Discussion <strong>and</strong> Sensitivity Analysis ............... 97<br />
2.6 Conclusion ................................. 98<br />
2.7 Appendix to Chapter 2 .......................... 99<br />
2.7.1 Comparison to Dem<strong>and</strong> Estimates <strong>in</strong> Other Sett<strong>in</strong>gs ..... 99<br />
2.7.2 Alternative Specifications of the Dem<strong>and</strong> Model ........ 100<br />
3 Private Coverage <strong>and</strong> Public Costs: Identify<strong>in</strong>g the Effect of Private<br />
Supplemental Insurance on Medicare Spend<strong>in</strong>g 117<br />
3.1 Introduction ................................ 117<br />
3.2 Background ................................ 122<br />
3.3 Data .................................... 124<br />
3.4 Empirical Model ............................. 125<br />
3.4.1 Estimat<strong>in</strong>g equations ....................... 125<br />
3.4.2 Estimation approach ....................... 127<br />
3.5 Identify<strong>in</strong>g Variation ........................... 128<br />
3.5.1 The Instrument .......................... 128<br />
3.5.2 Potential Concerns ........................ 130<br />
3.6 Results <strong>and</strong> Discussion .......................... 132<br />
3.6.1 Premiums <strong>and</strong> Medigap Choice ................. 132<br />
3.6.2 Causal Impact on Costs ..................... 134<br />
3.6.3 Robustness <strong>and</strong> External Validity ................ 136<br />
3.7 Policy Counterfactuals .......................... 137<br />
3.8 Conclusion ................................. 139<br />
4 Do Expir<strong>in</strong>g Budgets Lead to Wasteful Year-End Spend<strong>in</strong>g? Evi-<br />
dence from Federal Procurement 153<br />
4.1 Introduction ................................ 153<br />
4.2 A Model of Wasteful Year-End Spend<strong>in</strong>g ................ 157<br />
4.2.1 The Basel<strong>in</strong>e Model ........................ 157<br />
4.2.2 Rollover Budget Authority .................... 159<br />
4.2.3 Extend<strong>in</strong>g the Model To Allow For Rollover .......... 160<br />
xi
4.3 Does Spend<strong>in</strong>g Spike at the End of the Year? ............. 161<br />
4.3.1 The Federal Procurement Data System ............. 163<br />
4.3.2 The With<strong>in</strong>-Year Pattern of Government Procurement Spend<strong>in</strong>g 164<br />
4.3.3 The Impact of Appropriations Tim<strong>in</strong>g on the With<strong>in</strong>-Year Pat-<br />
tern of Government Procurement Spend<strong>in</strong>g ........... 165<br />
4.4 Is End of Year Spend<strong>in</strong>g of Lower Quality? ............... 167<br />
4.4.1 I.T. Dashboard .......................... 168<br />
4.4.2 Data <strong>and</strong> Summary Statistics .................. 169<br />
4.4.3 The Relative Quality of Year-End I.T. Contracts ....... 172<br />
4.4.4 Sensitivity Analysis ........................ 174<br />
4.4.5 Why Are Year End Contracts of Lower Quality? ........ 176<br />
4.5 Do Rollover Provisions Raise Spend<strong>in</strong>g Quality? ............ 179<br />
4.5.1 The DOJ’s Rollover Authority .................. 179<br />
4.6 Conclusion ................................. 182<br />
xii
List of Tables<br />
1.1 Asset Exemption Laws by State ..................... 52<br />
1.2 Implied First Stage Estimates ...................... 53<br />
1.3 Basel<strong>in</strong>e Coverage Estimates ....................... 54<br />
1.4 Sensitivity Analysis of the Effect on Coverage ............. 55<br />
1.5 Heterogeneity <strong>in</strong> the Effect on Coverage ................. 56<br />
1.6 The Effect on Assets ........................... 57<br />
1.7 The Effect on Premiums ......................... 58<br />
1.8 Basel<strong>in</strong>e Costs Estimates ......................... 59<br />
1.9 Sensitivity Analysis of the Effect on Costs ............... 60<br />
1.10 Micro-Simulation Estimates of Percent Covered Without <strong>and</strong> With<br />
Bankruptcy Insurance by Insurance Status ............... 61<br />
1.11 Policy Counterfactuals .......................... 62<br />
1.12 Summary Statistics: Seizable Assets by Insurance Status ....... 63<br />
1.13 Summary Statistics: Medical Costs by Insurance Status ........ 63<br />
1.14 Summary Statistics: Premiums ..................... 64<br />
1.15 Costs <strong>and</strong> Premiums by Deductible Level ................ 64<br />
2.1 Risk <strong>and</strong> Demographics .......................... 102<br />
2.2 Plan Characteristics ........................... 103<br />
2.3 Risk <strong>and</strong> Demographics by Plan ..................... 104<br />
2.4 Dem<strong>and</strong> Model .............................. 105<br />
2.5 Costs <strong>and</strong> Bids .............................. 106<br />
2.6 Match<strong>in</strong>g <strong>and</strong> Welfare under Alternative Contribution Policies .... 107<br />
2.7 Match<strong>in</strong>g <strong>and</strong> Welfare by Risk Score Qu<strong>in</strong>tile ............. 108<br />
xiii
2.8 The Value of Plan Choice ........................ 109<br />
2.9 Alternative Dem<strong>and</strong> Model Specifications ................ 115<br />
3.1 Medicare Cost-Shar<strong>in</strong>g .......................... 141<br />
3.2 Summary Statistics by Supplemental Insurance Type ......... 142<br />
3.3 Medigap Enrollment <strong>and</strong> Plan Characteristics by Letter ........ 143<br />
3.4 HRR Level-Medicare Costs Tabulated By State ............ 144<br />
3.5 First Stage: OLS Estimates of Premiums on State <strong>and</strong> HRR Costs .. 145<br />
3.6 Second Stage: Marg<strong>in</strong>al Effects for Medigap Choice from Mult<strong>in</strong>omial<br />
Logit Model ................................ 146<br />
3.7 Third Stage: OLS <strong>and</strong> Full Model Estimates of Medical Costs on Medi-<br />
gap Indicator ............................... 147<br />
3.8 Robustness Checks: Key Parameter Estimates from Alternative Spec-<br />
ifications <strong>and</strong> Samples .......................... 148<br />
3.9 Policy Counterfactuals for Subsidies <strong>and</strong> Taxes of Medigap Premiums 149<br />
3.10 OLS Estimates of Premiums on HRR- <strong>and</strong> State-Level Medicare Costs 150<br />
3.11 Second Stage: 2SLS Estimates of Medigap Choice from L<strong>in</strong>ear Proba-<br />
bility Model ................................ 151<br />
3.12 Third Stage: 3SLS Estimates of Medical Costs on Medigap Indicator 152<br />
4.1 Summary Statistics: Federal Contract<strong>in</strong>g, Pooled 2004 to 2009 FPDS 187<br />
4.2 Year-End Contract Spend<strong>in</strong>g by Agency, Pooled 2004 to 2009 FPDS . 188<br />
4.3 Year-End Contract Spend<strong>in</strong>g By Selected Product or Service Code,<br />
Pooled 2004 to 2009 FPDS ........................ 189<br />
4.4 Summary Statistics: Major I.T. Projects as of March, 2010 ...... 190<br />
4.5 Summary Statistics: Quality Indexes <strong>and</strong> Project Characteristics for<br />
Major I.T. Projects ............................ 191<br />
4.6 Ordered Logit Regressions of Overall Rat<strong>in</strong>g on Last Week <strong>and</strong> Controls192<br />
4.7 Alternative Overall Rat<strong>in</strong>gs Specifications ............... 193<br />
4.8 Difference-<strong>in</strong>-Differences Estimates of Overall Rat<strong>in</strong>g on Justice <strong>and</strong><br />
Last Week ................................. 194<br />
xiv
4.9 First Week Contract Spend<strong>in</strong>g for Selected Product or Service Codes,<br />
Pooled 2004 to 2009 FPDS ........................ 195<br />
4.10 Ordered Logit Regressions of Sub<strong>in</strong>dexes on Last Week <strong>and</strong> Controls 196<br />
4.11 Year-End Contract Characteristics Regressions ............. 197<br />
4.12 Percent of Projects <strong>in</strong> I.T. Dashboard Data ............... 198<br />
xv
List of Figures<br />
1.1 Histogram of Seizable Assets by Insurance Status ........... 41<br />
1.2 Payments vs. Charges by Insurance Status ............... 42<br />
1.3 Provider Bill<strong>in</strong>g Decision: Submitted Bill vs. List Price of Medical Care 43<br />
1.4 Payments vs. Charges by Seizable Assets ................ 44<br />
1.5 Simulated Instrument by State ...................... 45<br />
1.6 Log Seizable Assets Percentiles by State for a Constant Sample of<br />
Households ................................ 46<br />
1.7 Seizable Homestead Equity <strong>in</strong> 2005 vs. 1920 for a Constant Sample of<br />
Households ................................ 47<br />
1.8 Insurance Coverage vs. Seizable Assets ................. 48<br />
1.9 Effect on Coverage on Samples Exclud<strong>in</strong>g Each State ......... 49<br />
1.10 Micro-Simulation Estimates of Percent Covered Without <strong>and</strong> With<br />
Bankruptcy Insurance by Deductible .................. 50<br />
1.11 The Insurance Generosity Gap ...................... 51<br />
2.3 Contributions <strong>and</strong> Bids Relative to Integrated HMO .......... 111<br />
2.4 Employer Contributions <strong>and</strong> Employee Characteristics ........ 112<br />
2.5 Costs by Risk Score ............................ 113<br />
2.6 Bids by Risk Score ............................ 114<br />
3.1 Example of Identify<strong>in</strong>g Variation .................... 140<br />
4.1 Federal Contract<strong>in</strong>g by Week, Pooled 2004 to 2009 FPDS ....... 183<br />
4.2 Year-End Spend<strong>in</strong>g by Appropriations Date .............. 184<br />
xvi
4.3 I.T. Contract<strong>in</strong>g by Week ........................ 185<br />
4.4 Year-End <strong>and</strong> Rest-of-Year Overall Rat<strong>in</strong>gs ............... 186<br />
xvii
Chapter 1<br />
Bankruptcy as Implicit Health<br />
Insurance<br />
1.1 Introduction<br />
There is a large literature <strong>in</strong> economics evaluat<strong>in</strong>g the effects of government policy<br />
on health <strong>in</strong>surance coverage <strong>in</strong> the United States. 1 The question of why households<br />
choose to be un<strong>in</strong>sured is less well understood. 2 To better underst<strong>and</strong> the <strong>in</strong>surance<br />
coverage decision, this paper exam<strong>in</strong>es a mechanism that has received little attention:<br />
implicit <strong>in</strong>surance from the threat-po<strong>in</strong>t of personal bankruptcy.<br />
The implicit <strong>in</strong>surance from bankruptcy arises from the confluence of three factors.<br />
First, due to federal law, hospitals are required to provide emergency treatment on<br />
credit—<strong>and</strong> typically provide non-emergency care without any upfront payment as<br />
well. Second, under Chapter 7 of the U.S. bankruptcy code, households can discharge<br />
medical debt, giv<strong>in</strong>g up assets above asset exemption limits <strong>in</strong> return. 3 Third, because<br />
1 See ? for review of the take-up <strong>and</strong> crowd-out effects of <strong>public</strong> <strong>in</strong>surance expansions. See ? for<br />
a review of the impact of tax subsidies on the employer provision of <strong>in</strong>surance. See ? for a review<br />
of the effects of tax policy on <strong>in</strong>surance take-up <strong>in</strong> the non-group market.<br />
2 In a review of the literature ? concludes, “there are a variety of hypotheses for why so many<br />
<strong>in</strong>dividuals are un<strong>in</strong>sured, but no clear sense that this set of explanations can account for the 47<br />
million <strong>in</strong>dividuals.”<br />
3 The Bankruptcy Abuse Prevention <strong>and</strong> Consumer Protection Act (BAPCPA) of 2005 was implemented<br />
after the period I analyze. It prevents households with more than the state median <strong>in</strong>come<br />
1
CHAPTER 1. BANKRUPTCY 2<br />
of the deadweight cost of the bankruptcy process, households <strong>and</strong> creditors have an<br />
<strong>in</strong>centive to negotiate payments without a formal bankruptcy fil<strong>in</strong>g.<br />
Bankruptcy, as a result, provides households with a form of high-deductible health<br />
<strong>in</strong>surance. Households are exposed to the f<strong>in</strong>ancial risk from medical shocks up to<br />
the level of assets that can be seized <strong>in</strong> bankruptcy <strong>and</strong> <strong>in</strong>sured aga<strong>in</strong>st f<strong>in</strong>ancial risk<br />
above this level.<br />
Summary data on the un<strong>in</strong>sured suggest that this mechanism could be important.<br />
Figure 1.1 shows that un<strong>in</strong>sured households have vastly fewer seizable assets than<br />
households with private <strong>in</strong>surance. Sixty-three percent of the un<strong>in</strong>sured would give up<br />
less than $5,000 <strong>in</strong> a bankruptcy fil<strong>in</strong>g, compared to only 28 percent of the privately<br />
<strong>in</strong>sured. Figure 1.2 shows that payments by the un<strong>in</strong>sured are substantially lower<br />
when receiv<strong>in</strong>g a high volume of medical care. While payments by the privately<br />
<strong>in</strong>sured scale up proportionally with medical charges, payments by the un<strong>in</strong>sured are<br />
capped on average at just over $5,000.<br />
To more rigorously exam<strong>in</strong>e the mechanism, I construct a simple model of bankruptcy,<br />
medical bill<strong>in</strong>g, <strong>and</strong> <strong>in</strong>surance choice. The model predicts that, conditional on wealth,<br />
out-of-pocket medical payments should be decreas<strong>in</strong>g <strong>in</strong> the level of seizable assets for<br />
a given volume of medical care received. Hold<strong>in</strong>g wealth constant, households with<br />
fewer seizable assets should be less likely to purchase conventional coverage.<br />
I test these predictions us<strong>in</strong>g variation <strong>in</strong> the state-level asset exemption laws that<br />
specify the type <strong>and</strong> level of assets that can be seized <strong>in</strong> bankruptcy <strong>and</strong> detailed<br />
asset data from the restricted access Medical Expenditure Panel Survey (MEPS) <strong>and</strong><br />
the Panel Survey of Income Dynamics (PSID). The degree of cross-state variation<br />
<strong>in</strong> the asset exemption laws <strong>in</strong> substantial. Kansas, for example, allows households<br />
to exempt an unlimited amount of home equity <strong>and</strong> up to $40,000 <strong>in</strong> vehicle equity.<br />
Neighbor<strong>in</strong>g Nebraska allows households to keep no more than $12,500 <strong>in</strong> home equity<br />
or a $5,000 wildcard of any type of asset. Both states allow households to keep<br />
retirement assets. I construct a simulated <strong>in</strong>strument that isolates the variation <strong>in</strong><br />
seizable assets solely due to these laws, mechanically purg<strong>in</strong>g variation due to wealth<br />
from fil<strong>in</strong>g under Chapter 7 <strong>in</strong> most circumstances. The households most affected by the reform are<br />
unlikely to be marg<strong>in</strong>al to the mechanism I analyze.
CHAPTER 1. BANKRUPTCY 3<br />
<strong>and</strong> other household characteristics.<br />
Us<strong>in</strong>g this source of variation <strong>and</strong> cost data from the MEPS, I f<strong>in</strong>d that un<strong>in</strong>sured<br />
households with fewer seizable assets make lower out-of-pocket payments for a given<br />
level of medical care received. My preferred estimate <strong>in</strong>dicates that a log po<strong>in</strong>t drop<br />
<strong>in</strong> seizable assets reduces out-of-pocket payments by 37 percent. Consistent with the<br />
high-deductible nature of this <strong>in</strong>surance, the drop is larger for households that utilize<br />
more medical services as the “deductible” of this implicit <strong>in</strong>surance is more likely to<br />
b<strong>in</strong>d.<br />
Us<strong>in</strong>g the same source of variation <strong>and</strong> MEPS <strong>and</strong> PSID data, I f<strong>in</strong>d that house-<br />
holds with fewer seizable assets are less likely to have <strong>in</strong>surance. The estimates suggest<br />
that if the bankruptcy laws of the least debtor-friendly state of Delaware were applied<br />
nationally, 16.3 percent of the un<strong>in</strong>sured would buy health <strong>in</strong>surance. With a take-up<br />
semi-elasticity of -0.084 from the literature, achiev<strong>in</strong>g the same <strong>in</strong>crease <strong>in</strong> take-up<br />
would require a premium subsidy of 44.0 percent. 4<br />
I use three strategies to address the concern that asset exemption laws may be<br />
correlated with unobserved state-level factors. The first strategy is to use variation<br />
due to 1920 homestead exemptions as an <strong>in</strong>strument. By us<strong>in</strong>g homestead exemptions<br />
from before the era of widespread health <strong>in</strong>surance, the <strong>in</strong>strument alleviates potential<br />
bias from factors that might have simultaneously caused changes <strong>in</strong> asset exemption<br />
law <strong>and</strong> the dependent variables over the course of the twentieth century. Moreover,<br />
historical evidence (e.g., ?) show<strong>in</strong>g that 1920 homestead exemption levels resulted<br />
from an idiosyncratic array of n<strong>in</strong>eteenth century historical circumstances dim<strong>in</strong>ishes<br />
the concern that this variable merely proxies for a persistent state characteristic (such<br />
as the strength of a pro-debtor political movement).<br />
The second strategy is to sequentially add fixed effects for Census Regions (e.g.,<br />
Northeast) <strong>and</strong> Census Divisions (e.g., New Engl<strong>and</strong>) to the ma<strong>in</strong> specification. If a<br />
spatially correlated unobserved factor is driv<strong>in</strong>g the f<strong>in</strong>d<strong>in</strong>gs, the results should change<br />
with the <strong>in</strong>clusion of these covariates. Stable estimates across these specifications<br />
mitigate this concern. The third strategy is to add controls for a rich set of potentially<br />
4 The -0.084 estimates is taken from ? <strong>and</strong> is based on premium variation due to state-level<br />
community rat<strong>in</strong>g <strong>and</strong> premium compression regulations. As I discuss below, this estimate is <strong>in</strong> the<br />
center of the range <strong>in</strong> the literature.
CHAPTER 1. BANKRUPTCY 4<br />
relevant legislative factors. The fact that the estimates are uncharged by the <strong>in</strong>clusion<br />
of these covariates provides further support for the case that the identify<strong>in</strong>g variation<br />
is uncorrelated with unobserved state-level determ<strong>in</strong>ants of costs <strong>and</strong> coverage.<br />
I take the analysis a step further by calibrat<strong>in</strong>g a utility-based, micro-simulation<br />
model of <strong>in</strong>surance choice. The model is based on a nationally representative sample<br />
of households. Households face household-specific medical shock distributions that<br />
depend on the age <strong>and</strong> sex of each household member. To maximize their expected<br />
utility over wealth, households choose to either purchase conventional <strong>in</strong>surance at<br />
market premiums or rely on the high-deductible <strong>in</strong>surance from bankruptcy.<br />
I use this model to shed light on puzzles <strong>in</strong> the health policy literature. One of the<br />
puzzles I exam<strong>in</strong>e is the low take-up of high-deductible health plans (HDHP) by the<br />
un<strong>in</strong>sured (?). Proponents of these plans have argued that by offer<strong>in</strong>g lower premi-<br />
ums, HDHPs would exp<strong>and</strong> coverage among the un<strong>in</strong>sured. But because the implicit<br />
<strong>in</strong>surance from bankruptcy often resembles a high-deductible policy, HDHPs are rel-<br />
atively more likely to be crowded out by this mechanism. In the micro-simulation<br />
model, account<strong>in</strong>g for bankruptcy reduces the percentage of un<strong>in</strong>sured households<br />
projected to purchase a $1,000 deductible plan by 13 percentage po<strong>in</strong>ts. For a $5,000<br />
deductible plan, bankruptcy reduces dem<strong>and</strong> by 37 percentage po<strong>in</strong>ts, or from 43 to<br />
6percent.<br />
A second puzzle I exam<strong>in</strong>e is heterogeneity <strong>in</strong> the dem<strong>and</strong> for coverage. Without<br />
bankruptcy, households with <strong>and</strong> without <strong>in</strong>surance coverage are difficult to separate.<br />
Us<strong>in</strong>g variation <strong>in</strong> medical risk, tax exemptions, <strong>and</strong> adm<strong>in</strong>istrative costs, the model<br />
can separate the predicted coverage levels of un<strong>in</strong>sured <strong>and</strong> <strong>in</strong>sured households by only<br />
14 percentage po<strong>in</strong>ts. Because there are large differences <strong>in</strong> seizable assets across these<br />
groups, account<strong>in</strong>g for bankruptcy has a substantial <strong>in</strong>cremental effect, widen<strong>in</strong>g the<br />
gap <strong>in</strong> predicted coverage from 14 to 51 percentage po<strong>in</strong>ts.<br />
The mechanism I study may be relevant to policy design. On the one h<strong>and</strong>, the<br />
implicit <strong>in</strong>surance from bankruptcy has obvious <strong>in</strong>efficiencies. Un<strong>in</strong>sured households<br />
receive a substantial amount of non-emergency medical care <strong>in</strong> emergency rooms,<br />
which are obviously not optimized for this purpose (?). At the same time, these<br />
households receive less preventative care than they would with conventional health
CHAPTER 1. BANKRUPTCY 5<br />
<strong>in</strong>surance (?). There are deadweight costs to negotiation <strong>and</strong> collections under the<br />
threat-po<strong>in</strong>t of bankruptcy. And the fact that un<strong>in</strong>sured households are not exposed<br />
to the social cost of this implicit <strong>in</strong>surance means that too many households choose<br />
to be un<strong>in</strong>sured.<br />
Yet conventional health <strong>in</strong>surance has <strong>in</strong>efficiencies of its own. A particularly<br />
<strong>in</strong>terest<strong>in</strong>g <strong>in</strong>efficiency relative to bankruptcy <strong>in</strong>surance is moral hazard. With con-<br />
ventional <strong>in</strong>surance, medical providers <strong>and</strong> patients often have <strong>in</strong>centives to supply<br />
<strong>and</strong> dem<strong>and</strong> excess medical care. Under the implicit <strong>in</strong>surance from bankruptcy,<br />
however, physicians are more likely to be exposed to the social cost of their deci-<br />
sions <strong>and</strong> patients have little leverage to dem<strong>and</strong> excess treatment. The result is that<br />
bankruptcy may be a lower moral hazard form of social <strong>in</strong>surance. 5<br />
While a comprehensive analysis of the costs <strong>and</strong> benefits of bankruptcy is overly<br />
ambitious, I can exam<strong>in</strong>e one key tension between these forms of <strong>in</strong>surance with the<br />
micro-simulation model. In particular, the model allows me to trade off the benefit of<br />
bankruptcy as a lower moral hazard form of <strong>in</strong>surance aga<strong>in</strong>st the <strong>in</strong>efficiency due to<br />
households not fac<strong>in</strong>g the full social cost from be<strong>in</strong>g un<strong>in</strong>sured. This tradeoff suggests<br />
a corrective system of “Pigovian penalties” that expose households to the full social<br />
cost of the implicit <strong>in</strong>surance they receive. 6<br />
With the optimal penalty of $218 per person on average, about 7 percent of<br />
the un<strong>in</strong>sured take up coverage <strong>and</strong> aggregate surplus <strong>in</strong>creases by a small $4 to $5<br />
per person. Analyzed <strong>in</strong> this framework, the penalties under the Patient Projection<br />
<strong>and</strong> Affordable Care Act (PPACA), which average $418 per person, are too large,<br />
decreas<strong>in</strong>g aggregate surplus by $9 to $13 per person on average. By effectively<br />
elim<strong>in</strong>at<strong>in</strong>g the low moral hazard option, the counterfactual of mak<strong>in</strong>g medical debt<br />
non-dischargeable <strong>in</strong> bankruptcy reduces surplus by an average of $36 to $43 per<br />
person. While the exact welfare numbers should be viewed with some caution, the<br />
exercise suggests that dramatically reduc<strong>in</strong>g bankruptcy <strong>in</strong>surance—while hav<strong>in</strong>g the<br />
5 Estimates of the moral hazard effects from conventional health <strong>in</strong>surance, when identified off<br />
large medical costs, are <strong>in</strong>cremental to any moral hazard effects from the implicit <strong>in</strong>surance from<br />
bankruptcy.<br />
6 For this exercise, I assume that un<strong>in</strong>sured households are not already subsidized through the<br />
tax code or some other channel.
CHAPTER 1. BANKRUPTCY 6<br />
superficial benefit of <strong>in</strong>creas<strong>in</strong>g conventional coverage—may not be socially desirable.<br />
By study<strong>in</strong>g the <strong>in</strong>teraction between implicit <strong>and</strong> conventional <strong>in</strong>surance, this pa-<br />
per is closely related to the literature on long-term care <strong>in</strong>surance <strong>and</strong> the implicit<br />
<strong>in</strong>surance from spend<strong>in</strong>g down assets <strong>and</strong> qualify<strong>in</strong>g for Medicare, such as ?. Like<br />
them, I f<strong>in</strong>d that implicit <strong>in</strong>surance can cause substantial crowd-out. It is more gen-<br />
erally related to a literature <strong>in</strong> macroeconomics that assesses the equilibrium effects<br />
of bankruptcy as a form of consumption <strong>in</strong>surance aga<strong>in</strong>st a range of different shocks,<br />
<strong>in</strong>clud<strong>in</strong>g those related to earn<strong>in</strong>gs, divorce, childbear<strong>in</strong>g, lawsuits, <strong>and</strong> medical bills<br />
(??). By exam<strong>in</strong><strong>in</strong>g unpaid care, this paper is also related to ? <strong>and</strong> ?, who f<strong>in</strong>d a<br />
negative association between measures of charity care <strong>and</strong> <strong>in</strong>surance coverage. And<br />
this paper shares similarities with a literature that exam<strong>in</strong>es the effect of medical debt<br />
on bankruptcy fil<strong>in</strong>gs (e.g., ???), although unlike these papers, I treat bankruptcy as<br />
threat-po<strong>in</strong>t, not a dependent variable to be expla<strong>in</strong>ed. In this, my approach more<br />
closely resembles the “<strong>in</strong>formal bankruptcy” viewpo<strong>in</strong>t advanced by ?, who show that<br />
credit card debt is charged off without a bankruptcy fil<strong>in</strong>g <strong>in</strong> the majority of cases.<br />
The rest of the paper proceeds as follows: Section 2 presents the <strong>in</strong>stitutional<br />
background <strong>and</strong> a simple model. Section 3 provides an overview of the data. Sections<br />
4 discusses the identification strategy. The ma<strong>in</strong> empirical results are presented <strong>in</strong><br />
Sections 5 <strong>and</strong> 6. The micro-simulation model is presented <strong>in</strong> Section 7. Section 8<br />
discusses puzzles <strong>in</strong> the literature <strong>and</strong> policy implications. Section 9 concludes.<br />
1.2 Bankruptcy as a Form of High-Deductible Health<br />
Insurance<br />
1.2.1 Institutional Background<br />
The implicit <strong>in</strong>surance from bankruptcy arises from the comb<strong>in</strong>ation of three <strong>in</strong>sti-<br />
tutional features: the fact that most medical care is provided on credit even when<br />
repayment is unlikely, the ability of households to discharge this debt <strong>in</strong> bankruptcy,<br />
<strong>and</strong> the <strong>in</strong>centive for households <strong>and</strong> creditors to come to a negotiated solution to<br />
avoid the deadweight loss from a formal bankruptcy fil<strong>in</strong>g.
CHAPTER 1. BANKRUPTCY 7<br />
The Emergency Medical Treatment <strong>and</strong> Active Labor Act (EMTALA) requires<br />
that hospitals treat patients with emergency medical conditions, <strong>and</strong> prohibits them<br />
from delay<strong>in</strong>g treatment to <strong>in</strong>quire about <strong>in</strong>surance status or means of payment. 7 As<br />
a matter of practice, most hospitals provide non-emergency medical care on credit as<br />
well. Hospitals generally lack the <strong>in</strong>frastructure to bill patients at the po<strong>in</strong>t of service<br />
(?) <strong>and</strong> rarely deny service when repayment is unlikely. 8<br />
Hav<strong>in</strong>g received medical care on credit, bankruptcy law allows households to write<br />
off this debt <strong>in</strong> exchange for assets or future earn<strong>in</strong>gs. Chapter 7 is the most pop-<br />
ular form of personal bankruptcy, account<strong>in</strong>g for about 70 percent of all fil<strong>in</strong>gs (?).<br />
Under Chapter 7, households can discharge most unsecured debt such as credit card<br />
debt, <strong>in</strong>stallment loans, <strong>and</strong> medical bills. In return, creditors can seize assets above<br />
exemption levels that vary by asset type <strong>and</strong> state of residence.<br />
Chapter 13 is the other bankruptcy option. Under Chapter 13, households dis-<br />
charge most unsecured debt <strong>in</strong> exchange for payments out of disposable <strong>in</strong>come over<br />
the follow<strong>in</strong>g 3 to 5 years. By statute, these payments must be of at least the value<br />
that creditors would receive <strong>in</strong> Chapter 7. They are rarely larger because, <strong>in</strong> the<br />
period I study, all households have the option to file for Chapter 7. 9 Follow<strong>in</strong>g ?, I<br />
use seizable assets under Chapter 7 to characterize payments under both chapters of<br />
the bankruptcy code.<br />
Households, however, do not have to formally declare bankruptcy to receive the<br />
implicit <strong>in</strong>surance it provides. Under the threat-po<strong>in</strong>t of bankruptcy, households <strong>and</strong><br />
medical providers often resolve payments without an actual bankruptcy fil<strong>in</strong>g. There<br />
are multiple junctures where this occurs. Discounts on the list price of treatment—<br />
known as charity care—are offered at the po<strong>in</strong>t of service to the obviously <strong>in</strong>digent. 10<br />
7 U.S.C. 42 §1395dd.<br />
8 In a survey of nonprofit hospitals, 90 percent reported never deny<strong>in</strong>g any medical services to<br />
patients with no <strong>in</strong>surance (?). For-profit hospitals seem to operate similarly. For example, ?<br />
rejects the hypothesis that for-profit hospitals have a lower preference for charity care. ? f<strong>in</strong>d that<br />
the majority of emergency departments offer preventative care to un<strong>in</strong>sured patients.<br />
9 The Bankruptcy Abuse Prevention <strong>and</strong> Consumer Protection Act (BAPCPA), effective <strong>in</strong> October<br />
2005, established a “means test” for Chapter 7. It restricted households earn<strong>in</strong>g more than<br />
the state median <strong>in</strong>come from fil<strong>in</strong>g under Chapter 7 <strong>in</strong> most circumstances. The households most<br />
effected by the reform are unlikely to be marg<strong>in</strong>al to the mechanism I analyze.<br />
10 Federal <strong>and</strong> state laws also <strong>in</strong>fluence charity care provision. Nonprofits use charity care to meet
CHAPTER 1. BANKRUPTCY 8<br />
After treatment, many hospitals encourage f<strong>in</strong>ancially strapped households to nego-<br />
tiate discounts, requir<strong>in</strong>g the submission of <strong>in</strong>formation on <strong>in</strong>come <strong>and</strong> assets (e.g.,<br />
W-2s <strong>and</strong> mortgage payments) as part of their charity care applications. 11 Even when<br />
charity care is not provided, the lion’s share of medical debt is charged off <strong>in</strong> the col-<br />
lection process. Despite contract<strong>in</strong>g with debt collection agencies, providers recover<br />
only about 10 to 20 percent of bills submitted to the un<strong>in</strong>sured (?).<br />
Overall, bad debt from the un<strong>in</strong>sured was estimated at about $16 to $18 billion<br />
<strong>in</strong> 2004 (LeCuyer <strong>and</strong> S<strong>in</strong>ghal, 2007). While the exact proportion of debt discharged<br />
without a bankruptcy fil<strong>in</strong>g is unknown, ? f<strong>in</strong>d that the ratio of “medical” to “non-<br />
medical” bankruptcies, accord<strong>in</strong>g to their def<strong>in</strong>ition, is the same for households with<br />
<strong>and</strong> without <strong>in</strong>surance coverage, suggest<strong>in</strong>g that a large portion of the un<strong>in</strong>sured’s<br />
medical debt is charged off outside of formal bankruptcy. This is not unique to<br />
medical debt. ? report that the majority of credit card debt is charged off <strong>in</strong> what<br />
they call “<strong>in</strong>formal bankruptcy.”<br />
1.2.2 A Model of Bankruptcy as High-Deductible Health In-<br />
surance<br />
To br<strong>in</strong>g together these <strong>in</strong>stitutional features, I build a stylized model of households,<br />
medical providers, <strong>and</strong> bankruptcy. Households have a representative agent with<br />
expected utility preferences over wealth w = w E +w S , composed of exempt assets w E<br />
(net wealth that cannot be seized <strong>in</strong> a bankruptcy fill<strong>in</strong>g) <strong>and</strong> seizable assets w S (net<br />
wealth that can be seized by creditors). 12 They face medical shocks with list price<br />
m drawn from distribution M <strong>and</strong> choose whether to purchase health <strong>in</strong>surance to<br />
protect aga<strong>in</strong>st this f<strong>in</strong>ancial risk.<br />
Medical providers are obligated to provide medical services m <strong>and</strong> then attempt<br />
their Community Benefit requirement. Some states subsidize care to the <strong>in</strong>digent through unpaid<br />
care pools. I account for these factors <strong>in</strong> the empirical analysis.<br />
11 When this <strong>in</strong>formation is not provided, hospitals run credit checks on <strong>in</strong>debted patients, fil<strong>in</strong>g<br />
suit if they f<strong>in</strong>d evidence of a mortgage or sav<strong>in</strong>gs that could be claimed (“In Their Debt,” Baltimore<br />
Sun, December 12, 2008 to December 24th, 2008).<br />
12 I discuss endogeniz<strong>in</strong>g wealth at the end of this section.
CHAPTER 1. BANKRUPTCY 11<br />
Prediction 2. Hold<strong>in</strong>g overall wealth constant, the will<strong>in</strong>gness to pay for <strong>in</strong>surance<br />
v—<strong>and</strong> therefore <strong>in</strong>surance coverage—is <strong>in</strong>creas<strong>in</strong>g <strong>in</strong> the level of seizable assets w S .<br />
Because the implicit <strong>in</strong>surance from bankruptcy is a substitute for conventional health<br />
<strong>in</strong>surance, households with more seizable assets have a higher will<strong>in</strong>gness to pay for<br />
<strong>in</strong>surance <strong>and</strong> are ceteris paribus more likely to be <strong>in</strong>sured. I derive the prediction <strong>in</strong><br />
Appendix Section 1.10.1.<br />
The prediction is robust to natural extensions of the model. For example, allow<strong>in</strong>g<br />
<strong>in</strong>sured households to receive more or better medical treatment (?) <strong>in</strong>creases the<br />
<strong>in</strong>centive to purchase coverage, but households with fewer seizable assets are still<br />
relatively less likely to <strong>in</strong>sure. Similarly, <strong>in</strong>creas<strong>in</strong>g the cost of bankruptcy to account<br />
for factors such as stigma (?) or reduced future access to credit (?) does not affect the<br />
basic prediction. And endogeniz<strong>in</strong>g the level of seizable <strong>and</strong> exempt assets actually<br />
strengthens the relationship between <strong>in</strong>surance coverage <strong>and</strong> seizable assets because<br />
households that choose to forgo coverage have an additional <strong>in</strong>centive to reduce their<br />
seizable asset hold<strong>in</strong>gs.<br />
A more subtle po<strong>in</strong>t relates to the <strong>in</strong>formation available to households. The model<br />
assumes that households know their level of seizable assets w S <strong>and</strong> their health risk.<br />
Obviously this is an exaggeration. What matters is that households have some knowl-<br />
edge of the f<strong>in</strong>ancial risk from forgo<strong>in</strong>g <strong>in</strong>surance. For example, if households learn<br />
from the news-media or peers that medical providers <strong>in</strong> their community frequently<br />
seize home equity, then homeowners may be more likely to purchase coverage—even<br />
if they know noth<strong>in</strong>g about the mechanism.<br />
1.3 Data Overview<br />
I use two ma<strong>in</strong> data sources to test the predictions of the model. I exam<strong>in</strong>e the effect<br />
of bankruptcy on medical costs us<strong>in</strong>g data from the 2000 to 2005 waves of the Medical<br />
Expenditure Panel Survey (MEPS). The survey has detailed <strong>in</strong>formation on medical<br />
costs <strong>and</strong> <strong>in</strong>surance coverage. At the Data Center, encrypted state identifiers <strong>and</strong>
CHAPTER 1. BANKRUPTCY 12<br />
newly edited asset <strong>and</strong> debt variables are also available. 17<br />
I exam<strong>in</strong>e the effects on coverage us<strong>in</strong>g the MEPS data <strong>and</strong> the 1999 to 2005 waves<br />
of the biennial Panel Survey of Income Dynamics (PSID). The survey has <strong>public</strong> use<br />
<strong>in</strong>formation on <strong>in</strong>surance coverage, assets <strong>and</strong> debts variables, <strong>and</strong> state identifiers.<br />
Because the state of residence is non-encrypted <strong>in</strong> the PSID, I use this dataset for<br />
the primary <strong>in</strong>surance coverage analysis <strong>and</strong> replicate the results <strong>in</strong> the MEPS.<br />
In both data sets, I aggregate the <strong>in</strong>dividual-level data to the household level <strong>and</strong><br />
<strong>in</strong>flation-adjust monetary variables to 2005 dollars us<strong>in</strong>g the CPI-U. I also exclude<br />
households with one or more members enrolled <strong>in</strong> <strong>public</strong> <strong>in</strong>surance or a head age 65<br />
or older due to their eligibility for <strong>public</strong> Medicare <strong>in</strong>surance. 18 This leaves me with<br />
34,841 observations <strong>in</strong> the MEPS <strong>and</strong> 22,844 observations <strong>in</strong> the PSID.<br />
1.3.1 Asset Exemptions<br />
I codify assets exemptions us<strong>in</strong>g ?, a do-it-yourself guide to personal bankruptcy.<br />
Table 1.1 shows these exemptions. Contemporaneous homestead exemptions exhibit<br />
substantial variation, rang<strong>in</strong>g from zero <strong>in</strong> seven states to unlimited <strong>in</strong> eight others;<br />
vehicle exemptions range from zero <strong>in</strong> 15 states to at least $10,000 <strong>in</strong> five others;<br />
<strong>and</strong> wildcard exemptions, which can be applied to any asset, show a similar degree<br />
of variation. California residents can file under two different exemption systems, <strong>and</strong><br />
residents of 14 states can file under the federal exemption system if they choose. The<br />
last column shows homestead exemptions <strong>in</strong> the year of 1920 from ?. 19<br />
1.3.2 Seizable Assets<br />
Let i denote households <strong>and</strong> j denoted states. Seizable assets are a function of house-<br />
hold assets <strong>and</strong> debts <strong>and</strong> exemptions laws, denoted by vectors wi <strong>and</strong> ej. Follow<strong>in</strong>g<br />
the general structure of ?, seizable asset can be decomposed <strong>in</strong>to assets that can be<br />
17 ? f<strong>in</strong>d that the estimates of net worth <strong>in</strong> the MEPS are comparable to those <strong>in</strong> the Survey of<br />
Income <strong>and</strong> Program Participation (SIPP). My analysis does not go back before 2000 because when<br />
I started the project only asset <strong>and</strong> debt data from the 2000 to 2005 period had been edited.<br />
18 In the MEPS, I also drop the 3.6 percent of households with miss<strong>in</strong>g wealth variables.<br />
19 States that did not exist <strong>and</strong> states that had only acre-based exemptions are denoted as miss<strong>in</strong>g.
CHAPTER 1. BANKRUPTCY 14<br />
Appendix Section 1.10.2.<br />
1.3.3 Medical Costs<br />
Medical costs variables are shown <strong>in</strong> Appendix Table 1.13. Annual medical charges,<br />
def<strong>in</strong>ed as the list price of medical services used that year, average $6,647 per house-<br />
hold. Total payments, def<strong>in</strong>ed as the sum of payments received, are less than charges<br />
due to both discounts negotiated by <strong>in</strong>surance providers <strong>and</strong> medical care provided<br />
as charity care or bad debt. For privately <strong>in</strong>sured households, total payments average<br />
$4,480 per household. N<strong>in</strong>ety-four percent of these payments are either out-of-pocket<br />
payments or payments made by private <strong>in</strong>surance providers. For the un<strong>in</strong>sured, to-<br />
tal payments average $1,267 per household. Fifty-two percent of these payments are<br />
out-of-pocket. Miscellaneous payments, such as payments from charity care pools,<br />
worker’s compensation, or automobile <strong>in</strong>surance, account for most of the rest.<br />
In the empirical analysis, I use the out-of-pocket payments variable to measure<br />
the f<strong>in</strong>ancial risk faced by the un<strong>in</strong>sured. While this variable will accurately capture<br />
f<strong>in</strong>ancial risk <strong>in</strong> most circumstances, it may <strong>in</strong>accurately measure f<strong>in</strong>ancial risk for<br />
two reasons. First, out-of-pocket payments overstate f<strong>in</strong>ancial risk when these pay-<br />
ments are put on a credit card that is ultimately discharged <strong>in</strong> bankruptcy. Because<br />
households with fewer seizable assets are more likely to discharge credit card debt,<br />
out-of-pocket payments may overestimate f<strong>in</strong>ancial risk for some low seizable assets<br />
households.<br />
Second, out-of-pocket payments may understate f<strong>in</strong>ancial risk when medical providers<br />
break up large bills <strong>in</strong>to <strong>in</strong>stallments s<strong>in</strong>ce households are not prompted to report pay-<br />
ments that extend beyond the survey period. Because households with more seizable<br />
assets are more likely to make these large, multi-<strong>in</strong>stallment payments, out-of-pocket<br />
payments may understate f<strong>in</strong>ancial risk for high seizable assets households. Both<br />
overstated f<strong>in</strong>ancial risk for low seizable assets households <strong>and</strong> understated f<strong>in</strong>ancial<br />
risk for households with high seizable assets may lead to attenuated estimates of the<br />
relationship between seizable assets <strong>and</strong> f<strong>in</strong>ancial risk as measured by out-of-pocket<br />
payments, suggest<strong>in</strong>g that the empirical estimates should be <strong>in</strong>terpreted as a lower
CHAPTER 1. BANKRUPTCY 15<br />
bound of the effect of bankruptcy <strong>in</strong>surance on household f<strong>in</strong>ancial risk.<br />
I use Relative Risk Scores to control for medical utilization. As I discuss <strong>in</strong> Section<br />
1.4, controll<strong>in</strong>g for utilization is important because the direction of the unconditional<br />
relationship between out-of-pocket payments <strong>and</strong> seizable assets <strong>in</strong> theoretically am-<br />
biguous. To control for utilization, I use the Relative Risk Score variable constructed<br />
us<strong>in</strong>g the RiskSmart Version 2.2 software created by DxCG Inc. 23 This software uses<br />
<strong>in</strong>formation on age, sex, <strong>and</strong> medical diagnoses to project expected medical utiliza-<br />
tion based on regression models developed by the company. Because the software<br />
does not use geographical <strong>in</strong>formation to project utilization, the Relative Risk Score<br />
is orthogonal to asset exemption laws <strong>and</strong> other state-level factors.<br />
1.3.4 Insurance Premiums<br />
I conduct additional analysis us<strong>in</strong>g data on health <strong>in</strong>surance premiums <strong>in</strong> the <strong>in</strong>di-<br />
vidual market. In particular, I use data on premium quotes <strong>in</strong> each state that are<br />
listed on eHealthInsurance, a website that aggregates premium quotes from most of<br />
the major <strong>in</strong>surance providers. The data I use were collected <strong>in</strong> November, 2010.<br />
I collect premiums <strong>in</strong> each state for a 30-year-old non-smok<strong>in</strong>g male. Because pre-<br />
miums quotes are zip code specific, the data are collected for a zip code r<strong>and</strong>omly<br />
selected from the 10 most populous zip codes <strong>in</strong> the state. Along with premiums,<br />
I collect data on the <strong>in</strong>surance provider, plan br<strong>and</strong> name, deductible, co<strong>in</strong>surance<br />
rate, <strong>and</strong> co-payment for a office visit. I def<strong>in</strong>e an <strong>in</strong>surance plan as all observations<br />
with the same <strong>in</strong>surer, br<strong>and</strong> name, deductible, co<strong>in</strong>surance <strong>and</strong> co-payment. Table<br />
1.14 shows basic summary statistics for the data. The data set covers 41 states <strong>and</strong><br />
1,891 plans. The mean premium is $103 per month, the mean deductible is $3,351,<br />
<strong>and</strong> the mean co<strong>in</strong>surance rate is 15 percent.<br />
23 See form HC-092 on the MEPS website for a full description of the construction of this variable.
CHAPTER 1. BANKRUPTCY 17<br />
levels of seizable assets. For small charges the three groups make similar out-of-pocket<br />
payments, consistent with most households be<strong>in</strong>g below the cap. For large charges,<br />
out-of-pocket payments sharply diverge. While households with less than $10,000 <strong>in</strong><br />
seizable assets have their out-of-pocket payments capped, households with more than<br />
$50,000 <strong>in</strong> seizable assets see their out-of-pocket payments cont<strong>in</strong>ue to scale up with<br />
charges, consistent with a cap that is <strong>in</strong>creas<strong>in</strong>g <strong>in</strong> seizable assets.<br />
To test the capp<strong>in</strong>g-of-cost prediction more rigorously, I estimate regression mod-<br />
els of out-of-pocket payments on seizable assets, condition<strong>in</strong>g on the level of medical<br />
care received. Controll<strong>in</strong>g for medical utilization is important because the sign of<br />
the unconditional effect of seizable assets on out-of-pocket payments is theoretically<br />
ambiguous. To see this, consider the effect of reduc<strong>in</strong>g a household’s level of seiz-<br />
able assets. Due to the mechanical effect of the implicit <strong>in</strong>surance from bankruptcy,<br />
out-of-pocket payments should decrease. On the other h<strong>and</strong>, due to moral hazard,<br />
households may <strong>in</strong>crease their medical utilization, rais<strong>in</strong>g out-of-pocket costs <strong>and</strong><br />
potentially offsett<strong>in</strong>g the mechanical effect <strong>in</strong> the opposite direction.<br />
For the analysis, I restrict the sample to un<strong>in</strong>sured households with positive med-<br />
ical utilization. Lett<strong>in</strong>g i denote households <strong>and</strong> j denote states, the second stage<br />
cost equation takes the form<br />
ln(OOPij +1)=βwln(w S ij)+f(RRSi; βRRS)+XijβX + νij, (1.4)<br />
where the dependent variable is log out-of-pocket payments, w S ij is seizable assets,<br />
f(RRSi; β k m) is a fourth-order polynomial <strong>in</strong> the Relative Risk Score, Xij is a vector<br />
of household <strong>and</strong> state characteristics, <strong>and</strong> νij is the error term. 26 The capp<strong>in</strong>g of<br />
cost prediction is supported by a positive coefficient on seizable assets (βw > 0).<br />
1.4.3 Cross-State Variation<br />
Consistently estimat<strong>in</strong>g the parameters of <strong>in</strong>terest poses four dist<strong>in</strong>ct identification<br />
problems. The first issue is omitted variables: that coverage or costs <strong>and</strong> seizable<br />
26 The depend variable is rarely zero. In the sample analyzed, less than 4 percent of households<br />
make zero out-of-pocket payments.
CHAPTER 1. BANKRUPTCY 20<br />
or coverage over the course of the twentieth century. Moreover, historical evidence<br />
shows that 1920 homestead exemption levels resulted from an idiosyncratic array<br />
of n<strong>in</strong>eteenth century historical circumstances. Describ<strong>in</strong>g the key factors driv<strong>in</strong>g<br />
the establishment of homestead exemptions <strong>in</strong> the n<strong>in</strong>eteenth century, ? cites no<br />
less diverse a list than “Texas colonizers <strong>and</strong> western developers, labor <strong>and</strong> l<strong>and</strong><br />
reformers, antimonopoly Jacksonian egalitarians, defenders of family security <strong>and</strong><br />
women’s property rights, southern planters <strong>and</strong> yeomen devastated by the Civil War.”<br />
These heterogenous causes reduce the concern that historical homestead exemptions<br />
merely proxy for a persistent state-level characteristic such as the strength of the<br />
pro-debtor political movement.<br />
Importantly, historical homestead exemptions are also a good predictor of con-<br />
temporaneous exemptions values. 27 Figure 1.7 shows this graphically, plott<strong>in</strong>g the<br />
average level of seizable home equity for a constant, nationally representative sample<br />
of households under 2005 homestead exemptions (y-axis) <strong>and</strong> <strong>in</strong>flation-adjusted 1920<br />
homestead exemptions (x-axis) <strong>in</strong> each state. The plot also shows the fitted l<strong>in</strong>e from<br />
a bivariate regression. As the slope coefficient <strong>in</strong>dicates, homestead exemptions have<br />
become less generous over time, with seizable assets <strong>in</strong>creas<strong>in</strong>g on average by 90 per-<br />
cent. The R-squared value is 0.43, with the New Engl<strong>and</strong> states <strong>in</strong> the lower right<br />
corner be<strong>in</strong>g the most prom<strong>in</strong>ent outliers. 28<br />
I take two further approaches to reduce the concern that unobservable factors<br />
are driv<strong>in</strong>g the effect. The first is to sequentially add controls for Census Regions<br />
(e.g., Northeast) <strong>and</strong> Census Divisions (e.g., New Engl<strong>and</strong>) to the ma<strong>in</strong> specification.<br />
Stable results across these specifications should address the concern that the effects<br />
are be<strong>in</strong>g driven by a spatially-correlated, unobserved factor.<br />
The second is to control for a rich set of state-level legislative covariates. In<br />
the coverage equations, I control for <strong>in</strong>surance market regulations (e.g., community<br />
rat<strong>in</strong>g requirements, coverage m<strong>and</strong>ates) that may affect premiums. 29 While the<br />
27Many of the changes s<strong>in</strong>ce 1920 have simply been <strong>in</strong>flation updates passed by <strong>in</strong>dividual state<br />
legislatures (?).<br />
28A keyword search of newspaper articles <strong>in</strong> a six-month w<strong>in</strong>dow around major changes <strong>in</strong> Massachusetts<br />
<strong>and</strong> Connecticut assets exemptions failed to reveal any <strong>in</strong>formation on the reasons for<br />
these <strong>in</strong>creases.<br />
29The data on these regulations was taken from a Blue Cross Blue Shield (2002) compilation
CHAPTER 1. BANKRUPTCY 21<br />
<strong>public</strong>ly <strong>in</strong>sured are excluded from the basel<strong>in</strong>e sample, correlation between asset<br />
exemption laws <strong>and</strong> eligibility thresholds for <strong>public</strong> <strong>in</strong>surance programs might bias the<br />
estimates though sample selection. To assuage this concern, I estimate the regression<br />
models on samples exclud<strong>in</strong>g <strong>and</strong> <strong>in</strong>clud<strong>in</strong>g the <strong>public</strong>ly <strong>in</strong>sured. I also control for<br />
the presence <strong>and</strong> generosity of Medicaid Medically Needy programs that provide an<br />
alternative form of safety net coverage. 30 In the cost equations, I control for hospital<br />
characteristics <strong>and</strong> other state-level factors (e.g., share of private hospitals) that may<br />
affect out-of-pocket payments. 31<br />
1.5 The Effect on Insurance Coverage<br />
1.5.1 First Stage Estimates<br />
Table 1.2 shows implied first stage regressions of log seizable assets on different <strong>in</strong>stru-<br />
mental variables. Column 1 shows estimates with the basel<strong>in</strong>e <strong>in</strong>strument. Columns<br />
2 <strong>and</strong> 3 show estimates with <strong>in</strong>struments that isolate the variation due to seizable<br />
homestead equity <strong>and</strong> seizable non-homestead assets. These <strong>in</strong>struments are calcu-<br />
lated by tak<strong>in</strong>g a constant, nationally representative of households <strong>and</strong> calculat<strong>in</strong>g<br />
their level of seizable homestead <strong>and</strong> non-homestead equity as though they lived <strong>in</strong><br />
each state. Column 4 shows estimates us<strong>in</strong>g the <strong>in</strong>strument that isolates variation<br />
due to 1920 homestead exemptions. All the specifications <strong>in</strong>clude demographic con-<br />
trols (age group, family structure, race, education, <strong>and</strong> <strong>in</strong>come), state controls (mean<br />
<strong>in</strong>come, percent unemployed, percent covered by Medicaid), <strong>and</strong> year fixed effects.<br />
St<strong>and</strong>ard errors <strong>in</strong> all specifications here <strong>and</strong> throughout the paper are clustered at<br />
the state level. The <strong>in</strong>struments have substantial power. The basel<strong>in</strong>e <strong>in</strong>strument<br />
has an F-statistic of nearly 500 (column 1) <strong>and</strong> even the <strong>in</strong>strument based on 1920<br />
gracious shared by Am<strong>and</strong>a Kowalski.<br />
30I thank Jay Bhattacharya for alternat<strong>in</strong>g me to the presence of these programs. I use 2003 data<br />
on these programs taken from ?.<br />
31In particular, I control for the share of private <strong>and</strong> nonprofit hospitals, taken from the Hospital<br />
Statistics 2005 published by the American Hospital Association. I control for Disproportionate Share<br />
Hospital payments per 1,000 residents taken from the Kaiser Family Foundation. I control for the<br />
number of Federally Qualified Health Centers per 100,000 residents.
CHAPTER 1. BANKRUPTCY 22<br />
homestead exemptions has an F-statistic over 20 (column 4).<br />
1.5.2 Basel<strong>in</strong>e Coverage Estimates<br />
Before turn<strong>in</strong>g to the regression estimates, Figure 1.8 presents visual evidence on the<br />
crowd-out prediction. Plots on the same row show the exact same data. Datapo<strong>in</strong>ts<br />
are <strong>in</strong>dicated with state abbreviations <strong>in</strong> the plots <strong>in</strong> the left column <strong>and</strong> with circles<br />
proportional to the number of observations <strong>in</strong> the plots <strong>in</strong> the right column. Panels<br />
A <strong>and</strong> B plot <strong>in</strong>surance coverage aga<strong>in</strong>st seizable assets averaged by state. In the raw<br />
data, there is a strong upward-slopp<strong>in</strong>g relationship. The relationship is consistent<br />
across most states, with the outliers ma<strong>in</strong>ly states with very few observations. Panels<br />
C <strong>and</strong> D plot <strong>in</strong>surance coverage aga<strong>in</strong>st the basel<strong>in</strong>e <strong>in</strong>strument averaged by state—<br />
the graphical analogue to a bivariate reduced form regressions. The upward-slop<strong>in</strong>g<br />
relationship is similar although the figures are more noisy. (I discuss Panels E <strong>and</strong> F<br />
below.)<br />
Table 1.3 presents the basel<strong>in</strong>e regression specifications. The first four columns<br />
show OLS estimates. Column 1 only has controls for year fixed effects. Column 2 adds<br />
the demographic <strong>and</strong> state controls. Columns 3 <strong>and</strong> 4 add controls for wealth <strong>and</strong><br />
premiums. 32 S<strong>in</strong>ce wealth <strong>and</strong> premiums may be directly affected by asset exemption<br />
laws, the parameter estimates with these controls are <strong>in</strong>appropriate for counterfac-<br />
tuals <strong>in</strong>volv<strong>in</strong>g changes <strong>in</strong> these exemptions. The coefficient on log seizable assets is<br />
6.13 <strong>in</strong> column 1 <strong>and</strong> about 2.5 across the other specifications. The similar estimates<br />
with the wealth <strong>and</strong> premium controls suggest that the effect is not mediated through<br />
these channels. 33<br />
Columns 5 to 8 show the 2SLS estimates with the basel<strong>in</strong>e <strong>in</strong>strument (column 1<br />
of Table 1.2), add<strong>in</strong>g controls <strong>in</strong> the same manner as columns 1 through 4. The co-<br />
efficient on log seizable assets is 5.43 <strong>in</strong> the preferred specification with demographic<br />
32The premium <strong>in</strong>dex is state fixed effects from a regression of log premiums on plan <strong>and</strong> state<br />
fixed effects.<br />
33The slight <strong>in</strong>crease <strong>in</strong> columns 4 <strong>and</strong> 8 is due to the fact that premium data is only available<br />
for a subset of states. When compared to a specification without this control estimated on the same<br />
sample, controll<strong>in</strong>g for premiums very slightly reduces the coefficient on log seizable assets. This is<br />
consistent with a negative correlation between asset exemption levels <strong>and</strong> <strong>in</strong>surance premiums.
CHAPTER 1. BANKRUPTCY 23<br />
<strong>and</strong> state controls (column 6) <strong>and</strong> significantly positive at more than the 0.1 percent<br />
level. Like the OLS specifications, add<strong>in</strong>g controls for wealth (column 7) <strong>and</strong> pre-<br />
miums (column 8) has little effect. The larger 2SLS estimates are consistent with<br />
measurement error attenuat<strong>in</strong>g the coefficient of <strong>in</strong>terest <strong>in</strong> the OLS specifications.<br />
The reduced form panels of Figure 1.8, discussed above, suggested that Texas<br />
might be important to the 2SLS results. Figure 1.9 exam<strong>in</strong>es the robustness of the<br />
estimates to this type of concern by show<strong>in</strong>g the coefficient on log seizable assets<br />
from 51 separate regressions of the preferred specification (column 6 of Table 1.3)<br />
that sequentially drop households from the <strong>in</strong>dicated state. The figure shows that<br />
neither Texas—nor any other state for that matter—is driv<strong>in</strong>g the results. Dropp<strong>in</strong>g<br />
Texas does slightly reduce the coefficient on log seizable assets. However, the estimate<br />
without Texas is significantly greater than zero <strong>and</strong> slightly larger than the estimate<br />
that excludes the state of Arizona.<br />
1.5.3 Sensitivity Analysis<br />
As discussed above, I address the potential bias from endogenous asset exemption laws<br />
with three strategies. I <strong>in</strong>clude controls for state legislative factors. I add controls<br />
for Census Regions (e.g., Northeast) <strong>and</strong> Census Divisions (e.g,. New Engl<strong>and</strong>). And<br />
I show specifications that use variation due to 1920 homestead exemptions as an<br />
<strong>in</strong>strument.<br />
Column 1 of Table 1.4 presents estimates of the preferred specification on a sam-<br />
ple that <strong>in</strong>cludes households with <strong>public</strong> <strong>in</strong>surance <strong>and</strong> column 2 adds controls for<br />
state <strong>in</strong>surance regulations <strong>and</strong> Medicaid Medically Needy programs to the preferred<br />
specification. Columns 3 <strong>and</strong> 4 sequentially add fixed effects for Census Regions <strong>and</strong><br />
Census Divisions. The coefficient on log seizable assets is stable across these speci-<br />
fications, reduc<strong>in</strong>g concerns that the estimates are driven by unobserved state level<br />
factors.<br />
Column 5 applies a variant of the basel<strong>in</strong>e <strong>in</strong>strument that is calculated by averag-<br />
<strong>in</strong>g log seizable assets for the constant, nationally representative sample of households<br />
by age-by-educations groups rather than at the population level. The coefficient on
CHAPTER 1. BANKRUPTCY 24<br />
log seizable assets is very similar with this <strong>in</strong>strument. Columns 6 <strong>and</strong> 7 apply <strong>in</strong>-<br />
struments that isolate the variation due to seizable homestead equity <strong>and</strong> seizable<br />
non-homestead assets respectively. The coefficient is a smaller 4.53 (p-value ¡ 0.05)<br />
when estimated us<strong>in</strong>g variation <strong>in</strong> homestead exemptions <strong>and</strong> a larger 6.16 (p-value<br />
¡0.01) when estimated us<strong>in</strong>g variation <strong>in</strong> non-homestead seizable assets.<br />
Before discuss<strong>in</strong>g the eighth column, I return to Panels E <strong>and</strong> F of Figure 1.9.<br />
These panels show reduced form plots of <strong>in</strong>surance coverage aga<strong>in</strong>st log seizable home<br />
equity under <strong>in</strong>flation-adjusted 1920 homestead exemption laws averaged by state.<br />
The historical <strong>in</strong>strument plots show a similar upward-slop<strong>in</strong>g relationship <strong>and</strong> are<br />
somewhat less noisy than the basel<strong>in</strong>e <strong>in</strong>strument plots <strong>in</strong> the row above. One reason<br />
for this is that the New Engl<strong>and</strong> states—which are outliers <strong>in</strong> Panels C <strong>and</strong> D—have<br />
greater seizable assets levels under the historical exemptions <strong>and</strong> are closer to the<br />
regression l<strong>in</strong>e.<br />
Column 8 of Table 1.4 shows the 2SLS estimates with the historical <strong>in</strong>strument.<br />
With the historical <strong>in</strong>strument, the coefficient on log seizable assets is slightly larger<br />
than the preferred estimate (6.39 versus 5.43). A Hausman test cannot reject the va-<br />
lidity of the basel<strong>in</strong>e <strong>in</strong>strument. Given that the historical <strong>in</strong>strument has lower power<br />
<strong>in</strong> the first stage, it seems prudent to ma<strong>in</strong>ta<strong>in</strong> the basel<strong>in</strong>e <strong>in</strong>strument parameter of<br />
5.43 as the preferred estimate.<br />
1.5.4 Heterogeneity <strong>in</strong> the Effect on Coverage<br />
Table 1.5 exam<strong>in</strong>es heterogeneity <strong>in</strong> the coefficient on log seizable assets by estimat<strong>in</strong>g<br />
the preferred specification on subsamples of the data. The 2SLS estimate is somewhat<br />
larger when estimated on the sample of younger versus older households (def<strong>in</strong>ed by<br />
a household head younger than 35) <strong>and</strong> the sample of renters versus homeowners.<br />
Households earn<strong>in</strong>g less than the median <strong>in</strong>come are significantly more responsive to<br />
seizable assets levels (8.18 <strong>in</strong> column 5 versus 2.79 <strong>in</strong> column 6). This is consistent with<br />
?, who f<strong>in</strong>d that low <strong>in</strong>come households are more premium elastic <strong>in</strong> their dem<strong>and</strong> for<br />
health <strong>in</strong>surance. The coefficient on log seizable assets is larger for households with<br />
seizable assets below the median value (8.24 <strong>in</strong> column 7 versus 4.87 <strong>in</strong> column 8).
CHAPTER 1. BANKRUPTCY 25<br />
1.5.5 The Effect on Wealth <strong>and</strong> Premiums<br />
Table 1.3 showed that <strong>in</strong>clud<strong>in</strong>g controls for wealth <strong>and</strong> premiums did not have much<br />
of an effect on the parameter of <strong>in</strong>terest. Tables 1.6 <strong>and</strong> 1.7 exam<strong>in</strong>e the effects of<br />
asset exemption law on these variables on their own.<br />
Table 1.6 shows OLS <strong>and</strong> 2SLS estimates of assets on state-level measures of asset<br />
exemption law (the <strong>in</strong>struments). Columns 1 to 6 exam<strong>in</strong>e the effects of exemptions<br />
for specific asset categories (home equity <strong>and</strong> vehicle equity). In these specifications,<br />
the dependent variable is an <strong>in</strong>dicator for positive assets or the log asset value. All<br />
specifications <strong>in</strong>clude households controls—<strong>in</strong>clud<strong>in</strong>g a fourth-order polynomial <strong>in</strong><br />
wealth—<strong>and</strong> the basel<strong>in</strong>e <strong>in</strong>strument to control for the overall generosity of asset<br />
exemption laws <strong>in</strong> that state. Columns 1 <strong>and</strong> 2 show estimates of home equity on the<br />
measure of homestead exemptions laws. Columns 3 <strong>and</strong> 4 show similar specifications<br />
with homestead exemptions from 1920 used as an <strong>in</strong>strument. Columns 5 <strong>and</strong> 6 show<br />
estimates with a measure of vehicle equity as the dependent variable. None of the<br />
specifications show a statistically significant relationship between asset exemption<br />
laws <strong>and</strong> assets.<br />
Columns 7 <strong>and</strong> 8 of Table 1.6 exam<strong>in</strong>e the effect of asset exemption law on overall<br />
wealth. In these specifications, the dependent variable is an <strong>in</strong>dicator for positive<br />
wealth or the log wealth value. The parameter on the basel<strong>in</strong>e <strong>in</strong>strument—which<br />
is simply a measure of the generosity of asset exemption laws—is the coefficient of<br />
<strong>in</strong>terest. The specifications <strong>in</strong>clude the st<strong>and</strong>ard set of demographic controls except<br />
for the polynomial <strong>in</strong> wealth. Like the estimates for specific asset categories, there is<br />
no evidence of relationship between asset exemption laws <strong>and</strong> wealth levels.<br />
Table 1.7 exam<strong>in</strong>es the effect of asset exemption law on health <strong>in</strong>surance premi-<br />
ums by show<strong>in</strong>g estimates of log premiums on the basel<strong>in</strong>e <strong>in</strong>strument. Observations<br />
<strong>in</strong> these regressions are monthly premiums for a 30-year-old non-smok<strong>in</strong>g male. All<br />
specifications <strong>in</strong>clude plan fixed effects so that the parameters are identified off dif-<br />
ferences <strong>in</strong> premiums for the same <strong>in</strong>surance product across different states.<br />
The estimates show a marg<strong>in</strong>ally significant, negative relationship between asset<br />
exemption laws <strong>and</strong> <strong>in</strong>surance premiums, consistent with higher asset exemptions<br />
<strong>in</strong>flat<strong>in</strong>g hospital costs. Columns 1 to 3 show OLS specifications of log premiums
CHAPTER 1. BANKRUPTCY 26<br />
on the basel<strong>in</strong>e <strong>in</strong>strument. Column 1 has only plan fixed effects. Column 2 adds<br />
state demographic factors. Column 3 <strong>in</strong>cludes controls for state <strong>in</strong>surance market<br />
regulations. The estimates are quite similar across specifications <strong>and</strong> <strong>in</strong>dicate that a<br />
log po<strong>in</strong>t <strong>in</strong>crease <strong>in</strong> seizable assets is associated with a 7 to 10 percent decrease <strong>in</strong><br />
premiums, although the effect is only significantly less than zero at the 10 percent<br />
level. Columns 4 to 6 shows estimates where seizable home equity under 1920 as-<br />
set exemption laws is used as an <strong>in</strong>strument for contemporaneous asset exemptions.<br />
Controls are added analogously to columns 1 to 3. The estimates are slightly more<br />
negative but less precisely estimated. Column 6, which <strong>in</strong>cludes the full set of con-<br />
trols, <strong>in</strong>dicates that a log po<strong>in</strong>t <strong>in</strong>crease <strong>in</strong> asset exemptions decreases premiums by<br />
16 percent <strong>and</strong> is significantly negative at the 5 percent level.<br />
1.5.6 Summary <strong>and</strong> Interpretation<br />
Support<strong>in</strong>g the crowd-out prediction, the estimates show a robust positive relationship<br />
between <strong>in</strong>surance coverage <strong>and</strong> seizable assets. The preferred estimate <strong>in</strong>dicates that<br />
a log po<strong>in</strong>t <strong>in</strong>crease <strong>in</strong> seizable assets raises <strong>in</strong>surance coverage by 5.43 percentage<br />
po<strong>in</strong>ts. The effect is similar when identified us<strong>in</strong>g variation due to 1920 homestead<br />
exemptions, <strong>and</strong> is stable to the <strong>in</strong>troduction of controls for Census Regions <strong>and</strong><br />
Census Divisions <strong>and</strong> well as relevant state level legislative factors.<br />
There are a number of ways to put this estimate <strong>in</strong> context. Perhaps the most<br />
natural way—given the source of identify<strong>in</strong>g variation—is to consider counterfactuals<br />
<strong>in</strong>volv<strong>in</strong>g cross-state variation <strong>in</strong> asset exemption law. In the sample, 77.3 percent<br />
of <strong>in</strong>dividuals are <strong>in</strong>sured (recall that the sample excludes the elderly <strong>and</strong> those with<br />
<strong>public</strong> <strong>in</strong>surance). If the exemption laws of the most debtor-friendly state (Texas)<br />
were applied nationally, the estimates suggest that the number of un<strong>in</strong>sured would<br />
be <strong>in</strong>creased by 30.0 percent (6.8 percentage po<strong>in</strong>ts). 34 If the exemptions laws of<br />
the least debtor-friendly state (Delaware) were applied nationally, 16.3 percent (3.7<br />
percentage po<strong>in</strong>ts) of the un<strong>in</strong>sured would take-up <strong>in</strong>surance. 35 Take-up under the<br />
34 Couples fil<strong>in</strong>g jo<strong>in</strong>tly <strong>in</strong> Texas can exempt all of their home equity, an unlimited amount of<br />
retirement assets, <strong>and</strong> up to $60,000 of any asset of their choose. See Table 1.1 for details.<br />
35 Delaware gives couples fil<strong>in</strong>g jo<strong>in</strong>tly an unlimited exemption for retirement assets <strong>and</strong> a $500
CHAPTER 1. BANKRUPTCY 27<br />
least debtor-friendly laws is smaller <strong>in</strong> magnitude because most un<strong>in</strong>sured household<br />
have little wealth <strong>and</strong> are thus not exposed to substantially more f<strong>in</strong>ancial risk under<br />
these laws.<br />
Achiev<strong>in</strong>g the same <strong>in</strong>crease <strong>in</strong> coverage as implied by the Delaware laws would<br />
require a large premium subsidy. Too see this, I use the -0.084 take-up premium<br />
semi-elasticity for the un<strong>in</strong>sured estimated by the ?. 36 With this elasticity, <strong>in</strong>duc<strong>in</strong>g<br />
a 3.7 percentage po<strong>in</strong>ts <strong>in</strong>crease <strong>in</strong> coverage requires a premium subsidy of 44.0 (=<br />
100 × (0.037/0.084)) percent.<br />
The effect does not seem to be driven by an endogenous response of either wealth<br />
or premiums. Assets are completely unresponsive to asset exemption laws. And while<br />
a log po<strong>in</strong>t <strong>in</strong>crease <strong>in</strong> asset exemptions is associated with a (marg<strong>in</strong>ally significant)<br />
7 percent decrease <strong>in</strong> <strong>in</strong>surance premiums, controll<strong>in</strong>g for premiums does not impact<br />
the coefficient on seizable assets. This is consistent with f<strong>in</strong>d<strong>in</strong>gs <strong>in</strong> the literature<br />
that the dem<strong>and</strong> for health <strong>in</strong>surance is highly premium <strong>in</strong>elastic. For example, with<br />
the -0.084 premium semi-elasticity discussed above, a 7 percent decrease <strong>in</strong> <strong>in</strong>surance<br />
premiums is projected to <strong>in</strong>crease <strong>in</strong>surance coverage by only 0.6 (=-0.084 × -0.07)<br />
percentage po<strong>in</strong>ts.<br />
1.6 The Effect on Costs<br />
1.6.1 Basel<strong>in</strong>e Estimates<br />
Table 1.8 presents the basel<strong>in</strong>e OLS <strong>and</strong> 2SLS estimates of the effect on costs. The<br />
sample <strong>in</strong>cludes all un<strong>in</strong>sured households with positive medical utilization. St<strong>and</strong>ard<br />
errors are clustered at the state level <strong>in</strong> all specifications. In the OLS specification<br />
without controls (column 1), the elasticity of out-of-pocket payments with respect<br />
wildcard exemption. Households cannot exempt any home or vehicle equity <strong>in</strong> excess of the $500<br />
wildcard. See Table 1.1 for details.<br />
36 The ? estimate of -0.084 is identified off premium variation due to state-level community rat<strong>in</strong>g<br />
<strong>and</strong> premium compression regulations. The estimate is central to the small number of estimates <strong>in</strong><br />
the literature. It is smaller than the estimate of ?, who use the <strong>in</strong>troduction of a tax subsidy for<br />
<strong>in</strong>surance purchases by the self-employed, <strong>and</strong> is larger than the estimate from ?. See? for a review<br />
of estimates <strong>in</strong> the literature.
CHAPTER 1. BANKRUPTCY 28<br />
to seizable assets is 0.56. Includ<strong>in</strong>g controls for household demographics <strong>and</strong> uti-<br />
lization reduces this coefficient to 0.22 (column 2). Controll<strong>in</strong>g for wealth has little<br />
<strong>in</strong>cremental effect (column 3).<br />
The 2SLS estimates us<strong>in</strong>g the basel<strong>in</strong>e <strong>in</strong>strument are shown <strong>in</strong> columns 4 to 6,<br />
add<strong>in</strong>g controls <strong>in</strong> the same manner as columns 1 to 3. With controls, the 2SLS<br />
po<strong>in</strong>t estimates are slightly larger than the correspond<strong>in</strong>g OLS specifications. In the<br />
preferred specification with controls for demographics <strong>and</strong> utilization (column 5), the<br />
elasticity of out-of-pocket payments with respect to seizable assets is 0.37. As before,<br />
controll<strong>in</strong>g for wealth (column 6) has little effect.<br />
1.6.2 Sensitivity Analysis<br />
Table 1.9 exam<strong>in</strong>es the sensitivity of these results to concerns about potential bias<br />
from endogenous asset exemptions. Column 1 adds controls for hospital <strong>and</strong> state<br />
characteristics. Because hospitals with different ownership structures may have differ-<br />
ent <strong>in</strong>centives <strong>and</strong> preferences for charity care, I <strong>in</strong>clude controls for the share of <strong>public</strong><br />
<strong>and</strong> nonprofit hospitals <strong>in</strong> each state (private hospitals are the omitted category). I<br />
control for Disproportionate Share Hospital (DSH) payments per 1,000 residents as<br />
these may affect <strong>in</strong>centives to provide unpaid care, the presence of state charity care<br />
pools, <strong>and</strong> the number of Federally Qualified Health Centers per 100,000 people. The<br />
parameters of <strong>in</strong>terest are virtually unchanged <strong>in</strong> these specifications <strong>and</strong> none of the<br />
covariates have coefficients that are statistically dist<strong>in</strong>guishable from zero.<br />
Column 2 uses the <strong>in</strong>strument that isolates variation due to contemporaneous<br />
homestead exemptions <strong>and</strong> column 3 uses the <strong>in</strong>strument that isolates variation due<br />
to exemptions for non-homestead assets with no effect on the parameter of <strong>in</strong>ter-<br />
est. Column 4 applies the <strong>in</strong>strument that isolates variation due to 1920 homestead<br />
exemptions. Us<strong>in</strong>g this source of variation addresses concerns that contemporane-<br />
ous asset exemptions may be <strong>in</strong>fluenced by unobserved state-level factors such as<br />
the strength of the hospital lobby or state-level attitudes towards charity care. The<br />
estimates are very similar with this <strong>in</strong>strument.
CHAPTER 1. BANKRUPTCY 29<br />
The f<strong>in</strong>al two columns of Table 1.9 exam<strong>in</strong>e heterogeneity <strong>in</strong> the effect by utiliza-<br />
tion level. The capp<strong>in</strong>g-of-cost prediction <strong>and</strong> the simple descriptive evidence shown<br />
<strong>in</strong> Figures 1.2 <strong>and</strong> 1.4 suggest that seizable assets should have a larger impact on<br />
households with high utilization, as the “deductible” of bankruptcy <strong>in</strong>surance is more<br />
likely to b<strong>in</strong>d for these households. Columns 5 <strong>and</strong> 6 allow the effect to vary by<br />
whether utilization—as measured by the Relative Risk Score—is above the 90th per-<br />
centile (high utilization) or below this level (low utilization). The elasticity parameter<br />
is 0.51 for high utilization <strong>and</strong> about 0.37 for low utilization, consistent with a high<br />
deductible structure for this <strong>in</strong>surance.<br />
1.6.3 Summary <strong>and</strong> Interpretation<br />
Figures 1.2 <strong>and</strong> 1.4 showed that un<strong>in</strong>sured households with fewer seizable assets have<br />
their out-of-pocket payments capped <strong>in</strong> the raw data. The regression estimates show<br />
that this effect is robust to demographic controls <strong>and</strong> variation <strong>in</strong> seizable assets<br />
solely due to cross-state differences <strong>in</strong> asset exemption law. The preferred estimate<br />
<strong>in</strong>dicates that a log po<strong>in</strong>t <strong>in</strong>crease <strong>in</strong> seizable assets raises out-of-pocket payments by<br />
37 percent, conditional on medical utilization. Consistent with the high-deductible<br />
structure of this <strong>in</strong>surance, the elasticity is larger for households with high utilization.<br />
1.7 Micro-Simulation Model<br />
To shed light on puzzles <strong>in</strong> the literature <strong>and</strong> exam<strong>in</strong>e policy implications, I calibrate<br />
a micro-simulation model of <strong>in</strong>surance choice. The model is based on the set of un<strong>in</strong>-<br />
sured <strong>and</strong> privately <strong>in</strong>sured households <strong>in</strong> the 2005 PSID. Households face household-<br />
specific medical cost distributions that depend on the age <strong>and</strong> sex of each household<br />
member. Premiums for conventional health <strong>in</strong>surance are based on these costs, scaled<br />
to account for moral hazard, adm<strong>in</strong>istrative costs, <strong>and</strong> the cross-subsidization of un-<br />
paid care to the un<strong>in</strong>sured. Follow<strong>in</strong>g closely the formulation <strong>in</strong> Section 1.2, the model<br />
uses an expected utility framework to implicitly def<strong>in</strong>e each household’s will<strong>in</strong>gness<br />
to pay for conventional health <strong>in</strong>surance. Households purchase coverage if <strong>and</strong> only
CHAPTER 1. BANKRUPTCY 31<br />
CARA utility. 39 I show results with risk aversion parameters of α =2.5 × 10 −5 (low<br />
risk aversion), α =5.0 × 10 −5 (moderate risk aversion), <strong>and</strong> α =7.5 × 10 −4 (high risk<br />
aversion). Divid<strong>in</strong>g by the median wealth level of $40,318, these parameters can be<br />
<strong>in</strong>terpreted as relative risk coefficients of γ = 1, 2, <strong>and</strong> 3. Household-level medical cost<br />
distributions are constructed us<strong>in</strong>g <strong>in</strong>dividual-level medical cost distributions for age-<br />
by-sex-by-<strong>in</strong>surance status cells <strong>in</strong> the 2005 MEPS. The markup for moral hazard<br />
is µ1 = 0.56, the value implied by the RAND health <strong>in</strong>surance experiment price<br />
elasticity of -0.22 <strong>in</strong> an arc elasticity framework (?). The adm<strong>in</strong>istrative load factor<br />
is set to µ2 =0.1 for households with employer sponsored <strong>in</strong>surance <strong>and</strong> µ2 =0.5<br />
for the un<strong>in</strong>sured <strong>and</strong> households <strong>in</strong> the non-group market. 40 I solve for the cross-<br />
subsidization parameter µ3 endogenously us<strong>in</strong>g the cost distributions of un<strong>in</strong>sured<br />
households <strong>in</strong> the model. In Appendix Section 1.10.3, I discuss the construction of<br />
the medical cost distributions <strong>and</strong> premiums <strong>in</strong> more detail. I show that the calibrated<br />
premiums closely match quoted premiums <strong>in</strong> the <strong>in</strong>dividual market.<br />
1.8 Puzzles <strong>and</strong> Policy<br />
In this section of the paper, I use the micro-simulation model to <strong>in</strong>vestigate how<br />
bankruptcy <strong>in</strong>surance sheds light on puzzles <strong>in</strong> the health policy literature <strong>and</strong> to<br />
exam<strong>in</strong>e implications of this mechanism for the design of health <strong>in</strong>surance policy.<br />
1.8.1 Puzzles<br />
Puzzle 1: Why are 47 million <strong>in</strong>dividuals un<strong>in</strong>sured?<br />
Expla<strong>in</strong><strong>in</strong>g why households are un<strong>in</strong>sured is a central puzzle for scholars of health<br />
<strong>in</strong>surance. In his review of the literature ? writes, “there are a variety of hypothe-<br />
ses for why so many <strong>in</strong>dividuals are un<strong>in</strong>sured, but no clear sense that this set of<br />
explanations can account for the 47 million <strong>in</strong>dividuals.” Bankruptcy <strong>in</strong>surance is a<br />
39 Us<strong>in</strong>g a CARA specification avoids the problems associated with non-positive wealth. Calibrations<br />
with CRRA utility <strong>and</strong> a consumption floor generate stronger results.<br />
40 This values are take from ?.
CHAPTER 1. BANKRUPTCY 32<br />
compell<strong>in</strong>g additional explanation. I used regression analysis to quantify the impor-<br />
tance of this mechanism <strong>in</strong> the sections above. I use the micro-simulation model for<br />
an alternative perspective.<br />
Table 1.10 shows the percent of <strong>in</strong>sured <strong>and</strong> un<strong>in</strong>sured <strong>in</strong>dividuals predicted by<br />
the model to purchase a $2,000 deductible plan. Column 1 shows these predicted<br />
values from a simulation where there is no bankruptcy <strong>in</strong>surance. Column 2 adds<br />
bankruptcy <strong>in</strong>surance to the model. Without bankruptcy, <strong>in</strong>sured <strong>and</strong> un<strong>in</strong>sured<br />
<strong>in</strong>dividuals are difficult to separate. Us<strong>in</strong>g the variation <strong>in</strong> health risk, adm<strong>in</strong>istrative<br />
costs, <strong>and</strong> tax preferences, the model can only separate predicted coverage by 8 to 28<br />
percentage po<strong>in</strong>ts. Because there are large differences <strong>in</strong> seizable assets between these<br />
groups, account<strong>in</strong>g for bankruptcy substantially improves the explanatory power of<br />
the model, exp<strong>and</strong><strong>in</strong>g the gap <strong>in</strong> predicted coverage to 43 to 54 percentage po<strong>in</strong>ts.<br />
Bankruptcy <strong>in</strong>surance thus to <strong>in</strong>creases the gap <strong>in</strong> predicted coverage by 14 to 46<br />
percentage po<strong>in</strong>ts.<br />
Puzzle 2: The low take-up of high deductible plans<br />
A second puzzle is the low take-up of high deductible health plans (HDHP) by the<br />
un<strong>in</strong>sured. HDHPs were <strong>in</strong>tended by their proponents to exp<strong>and</strong> <strong>in</strong>surance coverage,<br />
yet despite offer<strong>in</strong>g low premiums they have not been successful <strong>in</strong> this regard (?).<br />
Crowd-out from implicit <strong>in</strong>surance is an appeal<strong>in</strong>g explanation for this failure. Be-<br />
cause the median un<strong>in</strong>sured household has less than $5,000 <strong>in</strong> seizable assets, HDHPs<br />
are largely redundant to the implicit <strong>in</strong>surance they already hold. Figure 1.10 makes<br />
this case quantitatively, plott<strong>in</strong>g the percent of un<strong>in</strong>sured households projected to<br />
purchase <strong>in</strong>surance by deductible level. Without bankruptcy, the probability of pur-<br />
chase <strong>in</strong>creases sharply from 19 percent for a $1,000 deductible plan to 43 percent for<br />
a $5,000 deductible plan, as the concavity of utility makes <strong>in</strong>surance more valuable at<br />
higher deductible levels. With bankruptcy, this <strong>in</strong>crease is virtually elim<strong>in</strong>ated, ris<strong>in</strong>g<br />
from 2 percent for a $1,000 deductible plan to only 6 percent for a $5,000 deductible<br />
contract, a full 37 percentage po<strong>in</strong>ts below the projected value without bankruptcy.
CHAPTER 1. BANKRUPTCY 33<br />
Puzzle 3: Ris<strong>in</strong>g risk, fall<strong>in</strong>g coverage<br />
A third puzzle is the association between ris<strong>in</strong>g health <strong>in</strong>surance costs <strong>and</strong> fall<strong>in</strong>g cov-<br />
erage. ? show that more than half the decrease <strong>in</strong> <strong>in</strong>surance coverage over the 1990s<br />
can be expla<strong>in</strong>ed by ris<strong>in</strong>g premiums. Yet as the authors expla<strong>in</strong>, from the st<strong>and</strong>po<strong>in</strong>t<br />
of economic theory this is counter<strong>in</strong>tuitive. With st<strong>and</strong>ard risk preferences, ris<strong>in</strong>g un-<br />
derly<strong>in</strong>g costs should lead to <strong>in</strong>creased <strong>in</strong>surance coverage. Tak<strong>in</strong>g bankruptcy <strong>in</strong>to<br />
account, however, reverses this <strong>in</strong>tuition. The decrease <strong>in</strong> coverage can be expla<strong>in</strong>ed<br />
by households substitut<strong>in</strong>g conventional health <strong>in</strong>surance for bankruptcy <strong>in</strong>surance<br />
that is <strong>in</strong>creas<strong>in</strong>g <strong>in</strong> actuarial value without <strong>in</strong>creas<strong>in</strong>g <strong>in</strong> price.<br />
Puzzle 4: The <strong>in</strong>surance generosity gap<br />
A fourth puzzle is the <strong>in</strong>surance generosity gap. In his review of the literature, ? asks<br />
why most U.S. households appear to be either under-<strong>in</strong>sured or over-<strong>in</strong>sured but rarely<br />
<strong>in</strong>-between. Implicit <strong>in</strong>surance from bankruptcy can expla<strong>in</strong> this f<strong>in</strong>d<strong>in</strong>g. To illustrate<br />
how bankruptcy <strong>in</strong>surance creates a generosity gap, Figure 1.11 presents a stylized<br />
budget set. Without bankruptcy, households face the st<strong>and</strong>ard cont<strong>in</strong>uous tradeoff<br />
between <strong>in</strong>surance generosity (y-axis) <strong>and</strong> other goods (x-axis). Implicit <strong>in</strong>surance<br />
generates a notch: households receive some implicit <strong>in</strong>surance without giv<strong>in</strong>g up other<br />
goods. Convex preferences give rise to an <strong>in</strong>surance generosity gap, with households<br />
sort<strong>in</strong>g <strong>in</strong>to the more generous conventional health <strong>in</strong>surance <strong>and</strong> the less generous<br />
implicit <strong>in</strong>surance from bankruptcy.<br />
1.8.2 Policy Implications<br />
I use the micro-simulation model to exam<strong>in</strong>e welfare implications of this mechanism.<br />
From a policy design perspective, bankruptcy <strong>in</strong>surance has many potential draw-<br />
backs: It forces the un<strong>in</strong>sured to receive care <strong>in</strong> emergency rooms, which are often<br />
not the most appropriate sett<strong>in</strong>gs (?), <strong>and</strong> may lead to <strong>in</strong>efficiently low levels of<br />
preventative care, <strong>in</strong>flat<strong>in</strong>g overall costs (?). Negotiation under the threat-po<strong>in</strong>t of<br />
bankruptcy is probably not the most efficient way of process<strong>in</strong>g payments <strong>and</strong> there<br />
are deadweight costs <strong>and</strong> externalities to formal bankruptcy.
CHAPTER 1. BANKRUPTCY 34<br />
For this exercise, however, I focus on a s<strong>in</strong>gle problem: because bankruptcy <strong>in</strong>sur-<br />
ance has a price of zero but cross-subsidized costs, too many households choose to be<br />
un<strong>in</strong>sured. I f<strong>in</strong>d this <strong>in</strong>efficiency particularly <strong>in</strong>terest<strong>in</strong>g because it directly relates<br />
to the <strong>in</strong>surance coverage decision <strong>and</strong> because the problem is naturally addressed<br />
with penalties for be<strong>in</strong>g un<strong>in</strong>sured—a key <strong>and</strong> controversial element of PPACA. 41<br />
Of course, conventional health <strong>in</strong>surance has well-documented <strong>in</strong>efficiencies as<br />
well. Relative to bankruptcy <strong>in</strong>surance, moral hazard is particularly relavent. 42 With<br />
conventional health <strong>in</strong>surance it is well known that medical providers <strong>and</strong> patients<br />
often have <strong>in</strong>centives to supply <strong>and</strong> dem<strong>and</strong> excess medical care. With bankruptcy<br />
<strong>in</strong>surance, on the other h<strong>and</strong>, physicians are more likely to be exposed to the social<br />
cost of their decisions <strong>and</strong> patients have little leverage to dem<strong>and</strong> excess treatment.<br />
Thus, for the optimal “Pigovian penalties,” low moral hazard bankruptcy <strong>in</strong>surance<br />
may be the efficient choice for some households.<br />
Table 1.11 shows the welfare effects of different penalty systems. For each penalty<br />
system, I allow households to choose between conventional <strong>in</strong>surance at the cali-<br />
brated premiums <strong>and</strong> bankruptcy <strong>in</strong>surance at the cost of the penalty. 43 The results<br />
are shown relative to a basel<strong>in</strong>e <strong>in</strong> which households can choose bankruptcy at no<br />
cost. The first set of rows shows coverage <strong>and</strong> welfare under the optimal Pigovian<br />
penalties—def<strong>in</strong>ed as the household-specific actuarially fair cost of the implicit <strong>in</strong>sur-<br />
ance from bankruptcy. 44 The optimal penalties average $218 per person <strong>and</strong> <strong>in</strong>duce<br />
7 to 8 percent of the un<strong>in</strong>sured to take-up coverage. As <strong>in</strong>dicated by the higher<br />
will<strong>in</strong>gness to pay, these households purchase more generous coverage than they had<br />
from bankruptcy. Due to the <strong>in</strong>creased moral hazard, however, costs rise by almost<br />
as much. The net effect is an <strong>in</strong>crease <strong>in</strong> surplus of only $4 to $5 per person.<br />
The second panel shows the welfare effects of the PPACA penalty system. When<br />
fully implemented <strong>in</strong> 2016, these penalties will equal the greater of $625 or 2.5 percent<br />
41For this exercise, I assume that un<strong>in</strong>sured households are not already subsidized through the<br />
tax code or some other channel.<br />
42Empirically, adverse selection does not seem to affect <strong>in</strong>surance choice on the extensive marg<strong>in</strong><br />
(??).<br />
43Recall that calibrated premiums are calculated us<strong>in</strong>g costs scaled up for moral hazard, adm<strong>in</strong>istrative<br />
load<strong>in</strong>g, <strong>and</strong> the cross-subsidization of the un<strong>in</strong>sured.<br />
44That is, expected medical costs above seizable assets.
CHAPTER 1. BANKRUPTCY 35<br />
of <strong>in</strong>come per household, up to a maximum of $2,085. Under these penalties, deflated<br />
to 2005 levels assum<strong>in</strong>g trend <strong>in</strong>flation, take-up ranges from 50 to 64 percent. The<br />
generosity of coverage does not change much on average while—due to higher moral<br />
hazard—costs <strong>in</strong>crease by a significant amount. The net effect is a decrease <strong>in</strong> surplus<br />
of $9 to $13 per person.<br />
The third panel considers the welfare effects of prevent<strong>in</strong>g households from dis-<br />
charg<strong>in</strong>g medical debt <strong>in</strong> bankruptcy—the exact exercise performed <strong>in</strong> coverage <strong>and</strong><br />
cost counterfactuals. Without bankruptcy, all households purchase conventional <strong>in</strong>-<br />
surance. As <strong>in</strong>dicated by the lower will<strong>in</strong>gness to pay, these plans are less generous<br />
on average. Because of the higher moral hazard from conventional <strong>in</strong>surance, costs<br />
do not shr<strong>in</strong>k by a commensurate amount. The net effect is a reduction <strong>in</strong> surplus<br />
of $36 to $43 per person. While the exact numbers, of course, are a function of the<br />
particular calibration parameters, the takeaway po<strong>in</strong>t from this last exercise is a sim-<br />
ple message: despite <strong>in</strong>creas<strong>in</strong>g <strong>in</strong>surance coverage, elim<strong>in</strong>at<strong>in</strong>g bankruptcy <strong>in</strong>surance<br />
may not be socially desirable.<br />
1.9 Conclusion<br />
Underst<strong>and</strong><strong>in</strong>g why households choose health <strong>in</strong>surance is fundamental to positive <strong>and</strong><br />
normative analysis of health <strong>in</strong>surance policy—yet the <strong>in</strong>surance coverage decision is<br />
not well understood. The objective of this paper is to describe <strong>and</strong> evaluate how the<br />
implicit <strong>in</strong>surance from bankruptcy bears on this decision. In the first part of the<br />
paper, I argued that the fact that most medical care is provided on credit coupled<br />
with the ability of households to discharge this debt for seizable assets <strong>in</strong> bankruptcy<br />
provides households with a form of high-deducible health <strong>in</strong>surance.<br />
I next evaluated the quantitative significance of this bankruptcy <strong>in</strong>surance. Ex-<br />
ploit<strong>in</strong>g variation <strong>in</strong> seizable assets that is orthogonal to wealth <strong>and</strong> other household<br />
characteristics, I found that households with more seizable assets make higher out-<br />
of-pocket payments <strong>and</strong> are substantially more likely to be <strong>in</strong>sured. The estimates<br />
suggest that if medical debt could not be discharged <strong>in</strong> bankruptcy, 16.3 percent<br />
of the un<strong>in</strong>sured would purchase coverage. Achiev<strong>in</strong>g the same <strong>in</strong>crease <strong>in</strong> coverage
CHAPTER 1. BANKRUPTCY 38<br />
education loans. While the PSID <strong>and</strong> MEPS dist<strong>in</strong>guish rema<strong>in</strong><strong>in</strong>g hous<strong>in</strong>g <strong>and</strong> vehi-<br />
cle pr<strong>in</strong>ciple from other debts, these other debts are not further decomposed. 45 The<br />
lead<strong>in</strong>g potential concern is that these other debts <strong>in</strong>clude a substantial amount of<br />
education debt—particularly for households with recent college graduates who may<br />
also be on the marg<strong>in</strong> of <strong>in</strong>surance coverage. To overcome this issue, I estimate the<br />
share of education debt <strong>in</strong> total unsecured debt <strong>in</strong> the 2004 Survey of Consumer<br />
F<strong>in</strong>ances (SCF) <strong>and</strong> use this estimate to project education debt <strong>in</strong> the PSID <strong>and</strong><br />
MEPS. In particular, I use parameter estimates from a probit regression of the share<br />
of education debt <strong>in</strong> total unsecured debt on <strong>in</strong>come, family structure, age of the<br />
household head, <strong>and</strong> unsecured debt level fully <strong>in</strong>teracted with a categorical variable<br />
for educational atta<strong>in</strong>ment of the head to project education debt values.<br />
1.10.3 Micro-Simulation Details<br />
Medical Cost Distributions<br />
I construct the household-level medical cost distributions us<strong>in</strong>g <strong>in</strong>dividual-level med-<br />
ical cost data from the 2005 MEPS for age-by-sex-by-<strong>in</strong>surance status cells. 46 For<br />
<strong>in</strong>sured <strong>in</strong>dividuals, costs are def<strong>in</strong>ed as total payments. For un<strong>in</strong>sured <strong>in</strong>dividu-<br />
als, my measure of costs <strong>in</strong> construct <strong>in</strong> the follow<strong>in</strong>g way: I start with medical<br />
charges as this variable accounts for medical services written off as charity care or<br />
bad debt. I then scale down charges by the cost-charge ratio (CCR) for the privately<br />
<strong>in</strong>sured population to account for the discount typically extended to the un<strong>in</strong>sured. 47<br />
45 The debt variable <strong>in</strong> the PSID is based on the question “Aside from the debts that we have<br />
already talked about, like any mortgage on your ma<strong>in</strong> or vehicle loans – do [you/you or anyone <strong>in</strong><br />
your family] currently have any other debts such as credit card charges, student loans, medical or<br />
legal bills, or loans from relatives?” In the MEPS it is based on the question “Does anyone <strong>in</strong> the<br />
family have any debts that we haven’t asked about, such as credit card balances, medical debts,<br />
life <strong>in</strong>surance policy loans, loans from relatives, <strong>and</strong> so forth?” While both these variables may<br />
<strong>in</strong>clude additional non-dischargeable debt (e.g., department store loans collateralized by durable<br />
goods, unpaid taxes), the bias <strong>in</strong>troduced by this is likely to be small.<br />
46 The age-by-sex groups are no more than 18 years old, male age 19 to 34, female age 19 to 34,<br />
male age 35 to 64, <strong>and</strong> female age 35 to 64.<br />
47 Recall from Figure 1.2 that privately <strong>in</strong>sured <strong>and</strong> un<strong>in</strong>sured households make similar payments<br />
for low charges.
CHAPTER 1. BANKRUPTCY 40<br />
TaxSim program.<br />
Appendix Table 1.15 shows the result<strong>in</strong>g distributions of medical costs <strong>and</strong> pre-<br />
miums normalized by household size for comparison. Costs have a long right tail:<br />
while 91 percent of persons <strong>in</strong>cur less than $1,000 <strong>in</strong> a given year, about 1 <strong>in</strong> 36 <strong>in</strong>cur<br />
more than $5,000 <strong>and</strong> 1 <strong>in</strong> 69 <strong>in</strong>cur more than $10,000 <strong>in</strong> medical costs. Calibrated<br />
premiums drop steeply from $2,126 per person for a plan with first dollar coverage to<br />
$1,009 per person for a $10,000 deductible plan.<br />
As a reality check, Appendix Table 1.15 compares these premiums to quoted<br />
<strong>in</strong>dividual market premiums for the median un<strong>in</strong>sured <strong>in</strong>dividual <strong>in</strong> the data (a 32-<br />
year-old male), adjusted for <strong>in</strong>flation <strong>and</strong> cost-shar<strong>in</strong>g. 48 The calibrated <strong>and</strong> market<br />
premiums are quite similar. The calibrated premiums are slightly less expensive<br />
for low deductible levels <strong>and</strong> somewhat more expensive for high deductibles. This<br />
difference could be expla<strong>in</strong>ed by selection or by heterogeneity <strong>in</strong> the moral hazard<br />
parameter across the expenditure distribution.<br />
48 The <strong>in</strong>dividual market premiums, from www.ehealth<strong>in</strong>surance.com, are for a non-smok<strong>in</strong>g 32year-old<br />
male. The policies are issued by Aetna <strong>and</strong> start <strong>in</strong> May 2010. They are adjusted for<br />
<strong>in</strong>flation us<strong>in</strong>g the Medical Care component of the CPI-U.
Percent by <strong>in</strong>surance status<br />
CHAPTER 1. BANKRUPTCY 41<br />
70%<br />
60%<br />
50%<br />
40%<br />
30%<br />
20%<br />
10%<br />
0%<br />
Figure 1.1: Histogram of Seizable Assets by Insurance Status<br />
Payments<br />
CHAPTER 1. BANKRUPTCY 42<br />
$35,000<br />
$30,000<br />
$25,000<br />
$20,000<br />
$15,000<br />
$10,000<br />
$5,000<br />
$-<br />
Figure 1.2: Payments vs. Charges by Insurance Status<br />
Privately <strong>in</strong>sured<br />
Un<strong>in</strong>sured<br />
$- $10,000 $20,000 $30,000 $40,000 $50,000 $60,000 $70,000 $80,000<br />
Charges<br />
Notes: Payments versus medical charges for un<strong>in</strong>sured <strong>and</strong> private <strong>in</strong>sured households.<br />
Payments are the sum of out-of-pocket payments <strong>and</strong> payments from private <strong>in</strong>surers. The<br />
plot is created by averag<strong>in</strong>g payments <strong>and</strong> charges at 20ths of the charge distribution.<br />
Pooled 2000 to 2005 MEPS, <strong>in</strong>flation-adjusted to 2005 us<strong>in</strong>g the CPI-U. Excludes<br />
households with a head age 65 or older. Household-level estimates weighted by number of<br />
<strong>in</strong>dividuals per household for <strong>in</strong>terpretation at the <strong>in</strong>dividual level.
CHAPTER 1. BANKRUPTCY 43<br />
Figure 1.3: Provider Bill<strong>in</strong>g Decision: Submitted Bill vs. List Price of Medical Care<br />
s (submitted bill)<br />
w S (seizable assets)<br />
Formal bankruptcy<br />
m (list price of medical care)<br />
s = m<br />
s̅(ŵ S )<br />
Informal bankruptcy<br />
s̅(ŵ S )
OOP payments<br />
CHAPTER 1. BANKRUPTCY 44<br />
$9,000<br />
$8,000<br />
$7,000<br />
$6,000<br />
$5,000<br />
$4,000<br />
$3,000<br />
$2,000<br />
$1,000<br />
$-<br />
Figure 1.4: Payments vs. Charges by Seizable Assets<br />
wS ! $50,000<br />
$10,000 " wS < $50,000<br />
wS< $10,000<br />
$- $5,000 $10,000 $15,000 $20,000 $25,000 $30,000 $35,000 $40,000 $45,000<br />
Charges<br />
Notes: Out-of-pocket payments versus medical charges for un<strong>in</strong>sured households with low<br />
(¡ $10,000), moderate ($10,000 to $50,000), <strong>and</strong> high (≥ $50,000) levels of seizable assets.<br />
The plot is created by averag<strong>in</strong>g payments <strong>and</strong> charges at 20ths of the charge distribution.<br />
Pooled 2000 to 2005 MEPS, <strong>in</strong>flation-adjusted to 2005 us<strong>in</strong>g the CPI-U. Excludes<br />
households with a head age 65 or older. Household-level estimates weighted by number of<br />
<strong>in</strong>dividuals per household for <strong>in</strong>terpretation at the <strong>in</strong>dividual level.
Simulated <strong>in</strong>strument<br />
CHAPTER 1. BANKRUPTCY 45<br />
11<br />
10.5<br />
10<br />
9.5<br />
9<br />
8.5<br />
8<br />
Figure 1.5: Simulated Instrument by State<br />
DE PANJMIOHSCNELAALKYTNWYILMONCUTHIINNYMDGAORWAIDCACONMWIVAMANDCTARMTAKIAFLMNOKAZSDWVMSMEVTNHRINVKSDCTX<br />
States<br />
Notes: The simulated <strong>in</strong>strument is log seizable assets for a constant, nationally<br />
representative sample of households as though they lived <strong>in</strong> each state. The sample is<br />
made up of un<strong>in</strong>sured <strong>and</strong> privately <strong>in</strong>sured households, exclud<strong>in</strong>g households with a head<br />
age 65 or older. Pooled 1999 to 2005 PSID, <strong>in</strong>flation-adjusted to 2005 us<strong>in</strong>g the CPI-U.<br />
See text for details on the seizable assets calculation.
ln(seizable sssets)<br />
CHAPTER 1. BANKRUPTCY 46<br />
Figure 1.6: Log Seizable Assets Percentiles by State for a Constant Sample of Households<br />
11.5<br />
11<br />
10.5<br />
10<br />
9.5<br />
9<br />
8.5<br />
8<br />
7.5<br />
DE PANJMISCOHNEALTNKYINILMOMDWYNCLAGANYUTVAHIORWAWIIDNMCACONDAKWVCTIAMAARMTFLSDMNOKAZMSMEVTNHRINVKSDCTX<br />
Notes: Percentiles of log seizable assets for a constant, nationally representative sample of<br />
households as though they lived <strong>in</strong> each state. The sample is made up of un<strong>in</strong>sured <strong>and</strong><br />
privately <strong>in</strong>sured households, exclud<strong>in</strong>g households with a head age 65 or older. Pooled<br />
1999 to 2005 PSID, <strong>in</strong>flation-adjusted to 2005 us<strong>in</strong>g the CPI-U. See text for details on the<br />
seizable assets calculation.<br />
State<br />
50th<br />
40th<br />
30th<br />
20th
CHAPTER 1. BANKRUPTCY 47<br />
Figure 1.7: Seizable Homestead Equity <strong>in</strong> 2005 vs. 1920 for a Constant Sample of<br />
Households<br />
Seizable home equity <strong>in</strong> 2005<br />
0 20,000 40,000 60,000 80,000<br />
Slope = 1.96 (0.38)<br />
R-squared = 0.42<br />
ID<br />
CA<br />
NV<br />
TX<br />
AZ<br />
MS<br />
AR<br />
AL<br />
NE<br />
LA<br />
CO<br />
MI<br />
GA<br />
MO<br />
40,000 50,000 60,000 70,000 80,000<br />
Seizable home equity <strong>in</strong> 1920<br />
Notes: Mean seizable home equity for a constant, nationally representative sample of<br />
households under 2005 <strong>and</strong> <strong>in</strong>flation-adjusted 1920 homestead exemption laws as they<br />
lived <strong>in</strong> each state. The New Engl<strong>and</strong> states are shaded red. The sample is made up of<br />
un<strong>in</strong>sured <strong>and</strong> privately <strong>in</strong>sured households, exclud<strong>in</strong>g households with a head age 65 or<br />
older. Slope coefficient from a bivariate regression weighted by population, with a robust<br />
st<strong>and</strong>ard error <strong>in</strong> parentheses. Pooled 1999 to 2005 PSID, <strong>in</strong>flation-adjusted to 2005 us<strong>in</strong>g<br />
the CPI-U. States that did not exist or had acre-based homestead exemption laws are<br />
excluded. Household-level estimates weighted by number of <strong>in</strong>dividuals per household for<br />
<strong>in</strong>terpretation at the <strong>in</strong>dividual level.<br />
WY<br />
VA<br />
UT<br />
WV<br />
KY OH SC<br />
NC IL<br />
NY NJ<br />
TN<br />
WA<br />
NM<br />
CT<br />
MA<br />
IN<br />
ME<br />
VT<br />
NH<br />
PA<br />
MD DE<br />
RI
50 60 70 80 90 100<br />
Insurance coverage (percentage po<strong>in</strong>ts)<br />
Figure 1.8: Insurance Coverage vs. Seizable Assets<br />
ND<br />
ME<br />
MT<br />
VT<br />
MN<br />
CT<br />
MA<br />
AK<br />
AL<br />
AR AZ<br />
IA PA<br />
DC<br />
MS<br />
DE<br />
KS<br />
OK<br />
WI NY KY<br />
TN MO<br />
GA ID<br />
CO IL<br />
VA NMUT NV MD MI<br />
NH<br />
NC OH<br />
OR<br />
NE HI<br />
SC<br />
FL<br />
IN<br />
SD<br />
CA<br />
WA<br />
TX<br />
WY LA<br />
WV<br />
8 9 10 11 12<br />
ln(seizable assets)<br />
(a) Insurance Coverage vs. Seizable<br />
Assets<br />
60 70 80 90 100<br />
Insurance coverage (percentage po<strong>in</strong>ts)<br />
TX<br />
MT<br />
VT<br />
RI<br />
CT<br />
ND MA<br />
ME<br />
MN<br />
NJ<br />
IA<br />
DC<br />
NH<br />
NV<br />
KS<br />
NY TN<br />
WI WA ID<br />
KY<br />
AK<br />
GA IL<br />
NM<br />
CO MD<br />
MO<br />
VA OR UT<br />
OH<br />
NC NE<br />
FL<br />
HI<br />
IN<br />
SC<br />
AL<br />
AZ<br />
AR<br />
PA<br />
MI<br />
SD<br />
MS OK<br />
WV<br />
8.5 9 9.5 10 10.5<br />
ln(seizable assets) for constant sample<br />
(c) Insurance Coverage vs. Simulated<br />
Instrument<br />
60 70 80 90 100<br />
Insurance coverage (percentage po<strong>in</strong>ts)<br />
ID<br />
NV<br />
GAMI<br />
MO<br />
CO VAUT<br />
NE<br />
NY TN<br />
KY WA IL<br />
NM<br />
OH<br />
NC<br />
SC<br />
PA<br />
NH<br />
IN<br />
MD<br />
AZ<br />
AR<br />
AL<br />
CA<br />
TX<br />
MS<br />
WY<br />
8.6 8.8 9 9.2 9.4 9.6<br />
ln(seizable home equity) for constant sample<br />
(e) Insurance Coverage vs. Historical<br />
Instrument<br />
LA<br />
CA<br />
RI<br />
CT<br />
NJ<br />
WV<br />
MA<br />
WY<br />
NJ<br />
ME<br />
LA<br />
VT<br />
DE<br />
DE<br />
RI<br />
50 60 70 80 90 100<br />
Insurance coverage (percentage po<strong>in</strong>ts)<br />
8 9 10 11 12<br />
ln(seizable assets)<br />
(b) Insurance Coverage vs. Seizable<br />
Assets<br />
60 70 80 90 100<br />
Insurance coverage (percentage po<strong>in</strong>ts)<br />
8.5 9 9.5 10 10.5<br />
ln(seizable assets) for constant saple<br />
(d) Insurance Coverage vs. Simulated<br />
Instrument<br />
60 70 80 90 100<br />
Insurance coverage (percentage po<strong>in</strong>ts)<br />
8.6 8.8 9 9.2 9.4 9.6<br />
ln(seizable home equity) for constant sample<br />
(f) Insurance Coverage vs. Historical<br />
Instrument<br />
Notes: Figures on the same row show the exact same data. In the left column, data po<strong>in</strong>ts<br />
are <strong>in</strong>dicated with state abbreviations. In the right column, states are <strong>in</strong>dicated with<br />
circles proportional to the number of observations from that state. Panels A <strong>and</strong> B plot<br />
the raw data: <strong>in</strong>surance coverage aga<strong>in</strong>st log seizable assets averaged by state. Panels C<br />
<strong>and</strong> D plots the reduced form: <strong>in</strong>surance coverage aga<strong>in</strong>st the <strong>in</strong>strument by state. The<br />
simulated <strong>in</strong>strument is mean log seizable assets for a constant, nationally representative<br />
sample of households as though they lived <strong>in</strong> that state. Panels E <strong>and</strong> F plots the reduced<br />
form with the historical <strong>in</strong>strument. The historical <strong>in</strong>strument is constructed similarly to<br />
the simulated <strong>in</strong>strument us<strong>in</strong>g seizable home equity under <strong>in</strong>flation-adjusted 1920<br />
homestead exemption laws. The sample is made up of un<strong>in</strong>sured <strong>and</strong> privately <strong>in</strong>sured<br />
households, exclud<strong>in</strong>g households with a head age 65 or older. Pooled data from the 1999<br />
to 2005 PSID, <strong>in</strong>flation-adjusted to 2005 us<strong>in</strong>g the CPI-U.
CHAPTER 1. BANKRUPTCY 49<br />
Coefficient on ln(seizable assets)<br />
9<br />
8<br />
7<br />
6<br />
5<br />
4<br />
3<br />
2<br />
1<br />
0<br />
Figure 1.9: Effect on Coverage on Samples Exclud<strong>in</strong>g Each State<br />
AK ALARAZCACOCTDCDEFLGAHIIAIDILINKSKYLAMAMDMEMIMNMOMSMTNCNDNENHNJNMNVNYOHOKORPARISCSDTNTXUTVAVTWAWIWVWY<br />
Excluded state<br />
Notes: Figure shows the coefficients <strong>and</strong> 95 percent confidence <strong>in</strong>tervals from the basel<strong>in</strong>e<br />
2SLS specification exclud<strong>in</strong>g the <strong>in</strong>dicated state. The specification is the same as column<br />
6 of Table 1.3. The 95 percent confidence <strong>in</strong>tervals are constructed us<strong>in</strong>g robust st<strong>and</strong>ard<br />
errors clustered at the state level. See the Table 1.3 note for more details.
Pr(WTP > premium)<br />
CHAPTER 1. BANKRUPTCY 50<br />
Figure 1.10: Micro-Simulation Estimates of Percent Covered Without <strong>and</strong> With<br />
Bankruptcy Insurance by Deductible<br />
70%<br />
60%<br />
50%<br />
40%<br />
30%<br />
20%<br />
10%<br />
0%<br />
Pr(WTP without > premium)<br />
Pr(WTP with > premium)<br />
$- $1,000 $2,000 $3,000 $4,000 $5,000 $6,000 $7,000 $8,000<br />
Deductible<br />
Notes: Micro-simulation estimates of the probability the will<strong>in</strong>gness to pay (WTP)<br />
exceeds the premium without <strong>and</strong> with bankruptcy by deductible level. Will<strong>in</strong>gness to pay<br />
is calculated us<strong>in</strong>g CARA utility with parameter of 2.5 × 10 −5 . Premiums calculated as<br />
the expected value of medical costs above the deductible scaled up to account for moral<br />
hazard (elasticity of -0.22) <strong>and</strong> adm<strong>in</strong>istrative load<strong>in</strong>g (50 percent). Household-level<br />
estimates weighted by number of <strong>in</strong>dividuals per household for <strong>in</strong>terpretation at the<br />
<strong>in</strong>dividual level.
CHAPTER 1. BANKRUPTCY 51<br />
Insurance generosity<br />
Generosity gap<br />
Figure 1.11: The Insurance Generosity Gap<br />
Conventional <strong>in</strong>surance<br />
Other goods<br />
Implicit <strong>in</strong>surance
CHAPTER 1. BANKRUPTCY 52<br />
Table 1.1: Asset Exemption Laws by State<br />
Contemporaneous exemptions Homestead<br />
Other<br />
exemptions<br />
f<strong>in</strong>ancial<br />
Wildcard no Federal for town lots<br />
State Homestead Vehicle Retirement assets Wildcard homestead available <strong>in</strong> 1920<br />
Alabama 10,000 0 Unlimited 0 6,000 6,000 No 2,000<br />
Alaska 67,500 7,500 Unlimited 3,500 0 0 No n/a<br />
Arizona 150,000 10,000 Unlimited 300 0 0 No 4,000<br />
Arkansas Unlimited 2,400 40,000 0 500 500 Yes 2,500<br />
California--system 1 75,000 4,600 Unlimited 1,825 0 0 No 5,000<br />
California--system 2 0 2,975 Unlimited 0 19,675 19,675 No n/a<br />
Colorado 90,000 6,000 Unlimited 0 0 0 No 2,000<br />
Connecticut 150,000 3,000 Unlimited 0 2,000 2,000 Yes 1,000<br />
Delaware 0 0 Unlimited 0 500 500 No 0<br />
District of Columbia Unlimited 5,150 Unlimited 0 17,850 17,850 Yes n/a<br />
Florida Unlimited 2,000 Unlimited 0 2,000 2,000 No n/a<br />
Georgia 10,000 7,000 Unlimited 0 11,200 11,200 No 1,600<br />
Hawaii 40,000 5,150 Unlimited 0 0 0 Yes n/a<br />
Idaho 50,000 6,000 Unlimited 0 1,600 1,600 No 5,000<br />
Ill<strong>in</strong>ois 15,000 2,400 Unlimited 0 4,000 4,000 No 1,000<br />
Indiana 0 0 Unlimited 0 20,000 20,000 No 600<br />
Iowa Unlimited 1,000 Unlimited 0 200 200 No n/a<br />
Kansas Unlimited 40,000 Unlimited 0 0 0 No n/a<br />
Kentucky 10,000 5,000 Unlimited 0 2,000 2,000 No 1,000<br />
Louisiana 25,000 0 Unlimited 0 0 0 No 2,000<br />
Ma<strong>in</strong>e 70,000 10,000 Unlimited 0 12,800 12,800 No 500<br />
Maryl<strong>and</strong> 0 0 Unlimited 0 22,000 22,000 No 0<br />
Massachusetts 1,000,000 1,400 Unlimited 1,250 0 0 Yes 800<br />
Michigan 7,000 0 Unlimited 0 0 0 No 1,500<br />
M<strong>in</strong>nesota 200,000 7,600 Unlimited 0 0 0 Yes n/a<br />
Mississippi 150,000 0 Unlimited 0 10,000 10,000 No 3,000<br />
Missouri 15,000 6,000 Unlimited 0 1,250 1,250 No 1,500<br />
Montana 200,000 5,000 Unlimited 0 0 0 No n/a<br />
Nebraska 12,500 0 Unlimited 0 0 5,000 No 2,000<br />
Nevada 400,000 30,000 1,000,000 0 0 0 No 5,000<br />
New Hampshire 200,000 8,000 Unlimited 0 8,000 8,000 Yes 500<br />
New Jersey 0 0 Unlimited 0 2,000 2,000 Yes 1,000<br />
New Mexico 60,000 8,000 Unlimited 0 1,000 4,000 Yes 1,000<br />
New York 20,000 0 Unlimited 0 10,000 10,000 No 1,000<br />
North Carol<strong>in</strong>a 13,000 3,000 Unlimited 0 8,000 8,000 No 1,000<br />
North Dakota 80,000 2,400 200,000 0 0 15,000 No n/a<br />
Ohio 10,000 2,000 Unlimited 800 800 800 No 1,000<br />
Oklahoma Unlimited 6,000 Unlimited 0 0 0 No n/a<br />
Oregon 33,000 3,400 15,000 15,000 800 800 No n/a<br />
Pennsylvania 0 0 Unlimited 0 600 600 Yes 300<br />
Rhode Isl<strong>and</strong> 200,000 20,000 Unlimited 0 0 0 Yes 0<br />
South Carol<strong>in</strong>a 10,000 2,400 Unlimited 0 0 2,000 No 1,000<br />
South Dakota Unlimited 0 500,000 0 4,000 4,000 No n/a<br />
Tennessee 7,500 0 Unlimited 0 8,000 8,000 No 1,000<br />
Texas Unlimited 0 Unlimited 0 60,000 60,000 Yes 5,000<br />
Utah 40,000 5,000 Unlimited 0 0 0 No 2,000<br />
Vermont 150,000 5,000 Unlimited 1,400 8,400 8,400 Yes 2,000<br />
Virg<strong>in</strong>ia 0 4,000 35,000 0 32,000 32,000 No 500<br />
Wash<strong>in</strong>gton 40,000 5,000 Unlimited 0 4,000 4,000 Yes 1,000<br />
West Virg<strong>in</strong>ia 0 4,800 Unlimited 0 51,600 51,600 No 1,000<br />
Wiscons<strong>in</strong> 40,000 0 Unlimited 2,000 10,000 10,000 Yes n/a<br />
Wyom<strong>in</strong>g 20,000 4,800 Unlimited 0 0 0 No 2,500<br />
Federal 18,500 5,900 Unlimited 0 20,450 20,450 n/a n/a<br />
Averages* 58,821 4,884 298,333 501 6,592 7,073 27% 1,679<br />
Notes: Contemporaneous exemptions for couples fil<strong>in</strong>g jo<strong>in</strong>tly from Elias (2007) <strong>and</strong> historical exemptions for couples fil<strong>in</strong>g jo<strong>in</strong>tly<br />
from Goodman (1993). Under contemporaneous law, California residents can choose between system 1 <strong>and</strong> 2 <strong>and</strong> residents can<br />
choose federal exemptions <strong>in</strong> states where federal exemptions are available. Wildcard no homestead exemption is available to<br />
households which do not take the homestead exemption. For the historical exemptions, states that did not exist <strong>and</strong> states that had<br />
acre-based exemptions are denoted as n/a. States that did not have homestead exemptions are assigned a value of zero.<br />
*Excludes states with unlimited or n/a exemptions.
CHAPTER 1. BANKRUPTCY 53<br />
Table 1.2: Implied First Stage Estimates<br />
Dependent variable: ln(w<br />
(1) (2) (3) (4)<br />
Instruments<br />
Basel<strong>in</strong>e <strong>in</strong>strument 1.00<br />
(0.04)<br />
Homestead <strong>in</strong>strument 0.38<br />
(0.06)<br />
Non-homestead <strong>in</strong>strument 1.10<br />
(0.10)<br />
Historical homestead <strong>in</strong>strument<br />
Age group<br />
1.41<br />
(0.31)<br />
35-44 -0.14 -0.14 -0.13 -0.12<br />
(0.02) (0.02) (0.02) (0.02)<br />
45-54 -0.20 -0.20 -0.20 -0.20<br />
(0.03) (0.03) (0.03) (0.04)<br />
55-64 -0.23 -0.23 -0.21 -0.22<br />
Family structure<br />
(0.04) (0.04) (0.04) (0.04)<br />
Couple -0.21 -0.21 -0.20 -0.22<br />
(0.03) (0.03) (0.03) (0.03)<br />
S<strong>in</strong>gle parent -0.02 -0.02 -0.02 0.00<br />
(0.03) (0.03) (0.03) (0.03)<br />
Couple with children -0.27 -0.26 -0.26 -0.26<br />
Race<br />
(0.03) (0.03) (0.03) (0.04)<br />
Non-white 0.00 -0.01 0.01 0.02<br />
Education<br />
(0.02) (0.03) (0.02) (0.02)<br />
High school to some college 0.02 0.03 0.03 0.05<br />
(0.03) (0.03) (0.03) (0.02)<br />
College or greater -0.01 0.00 0.01 0.01<br />
(0.03) (0.03) (0.03) (0.03)<br />
S )<br />
Income polynomial (4th order) Yes Yes Yes Yes<br />
State controls Yes Yes Yes Yes<br />
R-squared 0.89 0.88 0.88 0.89<br />
N 20,265 20,265 20,265 17,418<br />
F-statistic on excluded <strong>in</strong>strument 494.06 36.16 116.93 20.95<br />
Prob > F 0.00 0.00 0.00 0.00<br />
Notes: The basel<strong>in</strong>e <strong>in</strong>strument is mean log seizable assets for a constant, nationally representative sample of households<br />
as though they lived <strong>in</strong> that state. The homestead, non-homestead, <strong>and</strong> historical homestead <strong>in</strong>struments are constructed<br />
similarly us<strong>in</strong>g seizable home equity, non-homestead seizable assets, <strong>and</strong> seizable home equity under <strong>in</strong>flation-adjusted<br />
1920 homestead exemption laws. The excluded demographic groups are age 18-34, s<strong>in</strong>gle, less than high school, <strong>and</strong><br />
white. The state controls are mean <strong>in</strong>come, percent unemployed, <strong>and</strong> percent covered by Medicaid. Robust st<strong>and</strong>ard errors<br />
clustered at the state level <strong>in</strong> parentheses. Pooled 1999 to 2005 PSID exclud<strong>in</strong>g households with <strong>public</strong> <strong>in</strong>surance or a head<br />
age 65 or older. Inflation-adjusted to 2005 us<strong>in</strong>g the CPI-U.
CHAPTER 1. BANKRUPTCY 54<br />
Table 1.3: Basel<strong>in</strong>e Coverage Estimates<br />
(1) (2) (3) (4) (5) (6) (7) (8)<br />
ln(w S Dependent variable: Percent of household <strong>in</strong>sured<br />
OLS 2SLS<br />
) 6.13 2.45 2.36 2.63 8.80 5.43 5.38 6.15<br />
Age group<br />
(0.37) (0.20) (0.53) (0.25) (2.27) (1.16) (0.95) (0.87)<br />
35-44 4.48 3.98 4.48 3.49 4.28 3.32<br />
(0.97) (0.93) (1.13) (1.10) (0.88) (1.60)<br />
45-54 3.08 2.48 2.91 1.22 2.94 0.78<br />
(1.04) (1.02) (1.34) (1.27) (0.98) (2.39)<br />
55-64 8.03 7.44 8.48 4.53 7.79 4.42<br />
Family structure<br />
(1.08) (1.05) (1.44) (1.81) (1.01) (3.63)<br />
Couple 5.38 4.76 6.55 5.05 5.29 5.89<br />
(1.55) (1.43) (1.92) (1.48) (1.42) (1.95)<br />
S<strong>in</strong>gle parent 3.04 3.00 4.72 3.29 3.13 5.01<br />
(1.55) (1.51) (1.92) (1.52) (1.48) (2.00)<br />
Couple with children 11.41 10.51 12.53 11.16 11.14 12.25<br />
Race<br />
(1.51) (1.38) (1.96) (1.41) (1.37) (1.83)<br />
Non-white -6.34 -6.06 -5.85 -4.95 -5.86 -4.10<br />
Education<br />
(0.76) (0.77) (1.01) (0.88) (0.82) (1.26)<br />
High school to some college 9.86 9.83 8.65 9.12 9.51 7.77<br />
(1.88) (1.86) (2.37) (1.81) (1.83) (1.70)<br />
College or greater 11.79 12.03 10.76 10.35 11.94 8.91<br />
(1.77) (1.73) (2.10) (1.68) (1.70) (1.86)<br />
Income polynomial (4th order) Yes Yes Yes Yes Yes Yes<br />
Wealth polynomial (4th order) Yes Yes<br />
State controls Yes Yes Yes Yes Yes Yes<br />
State premium <strong>in</strong>dex (log po<strong>in</strong>ts) -1.15 -1.47<br />
(0.91) (0.93)<br />
Instrument<br />
Basel<strong>in</strong>e <strong>in</strong>strument Yes Yes Yes Yes<br />
R-squared 0.13 0.27 0.28 0.27 - - - -<br />
N 20,265 20,265 20,265 14,510 20,265 20,265 20,265 14,510<br />
Notes: The dependent variable is the percent of household member-months <strong>in</strong>sured. In the 2SLS specifications, the<br />
basel<strong>in</strong>e <strong>in</strong>strument is mean log seizable assets for a constant, nationally representative sample of households as though<br />
they lived <strong>in</strong> that state. The excluded demographic groups are age 18-34, s<strong>in</strong>gle, less than high school, <strong>and</strong> white. The<br />
state controls are mean <strong>in</strong>come, percent unemployed, <strong>and</strong> percent covered by Medicaid. See text for a description of the<br />
premium <strong>in</strong>dex. All specifications <strong>in</strong>clude an <strong>in</strong>dicator for the bottom-cod<strong>in</strong>g of seizable assets. Robust st<strong>and</strong>ard errors<br />
clustered at the state level <strong>in</strong> parentheses. Pooled 1999 to 2005 PSID exclud<strong>in</strong>g households with <strong>public</strong> <strong>in</strong>surance or a<br />
head age 65 or older. Inflation-adjusted to 2005 us<strong>in</strong>g the CPI-U.
CHAPTER 1. BANKRUPTCY 55<br />
Table 1.4: Sensitivity Analysis of the Effect on Coverage<br />
Basel<strong>in</strong>e IV by<br />
Non-homestead<br />
Publicly <strong>in</strong>sured Ins mkt regs Census region Census division group Homestead IV IV Historical IV<br />
(1) (2) (3) (4) (5) (6) (7) (8)<br />
ln(w S Dependent variable: Percent of household <strong>in</strong>sured<br />
) 6.30 4.71 5.01 4.23 4.71 4.53 6.16 6.39<br />
(1.11) (1.23) (1.83) (0.72) (1.17) (1.29) (0.67) (0.82)<br />
Demographic controls Yes Yes Yes Yes Yes Yes Yes Yes<br />
State controls Yes Yes Yes Yes Yes Yes Yes Yes<br />
Additional state legislative factors Yes<br />
Census region FE Yes<br />
Census division FE Yes<br />
Instrument<br />
Basel<strong>in</strong>e <strong>in</strong>strument Yes Yes Yes Yes<br />
Basel<strong>in</strong>e <strong>in</strong>strument by group Yes<br />
Homestead <strong>in</strong>strument Yes<br />
Non-homestead <strong>in</strong>strument Yes<br />
Historical <strong>in</strong>strument Yes<br />
N 25,450 20,265 20,265 20,265 20,265 20,265 20,265 17,418<br />
Notes: The dependent variable is the percent of household member-months <strong>in</strong>sured. The basel<strong>in</strong>e <strong>in</strong>strument is mean log seizable assets for a constant, nationally<br />
representative sample of households as though they lived <strong>in</strong> that state. Means are taken by age <strong>and</strong> education group for the basel<strong>in</strong>e <strong>in</strong>strument by group variable. The<br />
homestead, non-homestead <strong>and</strong> historical homestead <strong>in</strong>struments are constructed similarly to the basel<strong>in</strong>e <strong>in</strong>strument us<strong>in</strong>g seizable home equity, non-homestead seizable<br />
assets, <strong>and</strong> seizable home equity under <strong>in</strong>flation-adjusted 1920 homestead exemption laws. Demographic controls are dummies for age group, family structure, race, <strong>and</strong><br />
education. State controls are mean <strong>in</strong>come, percent unemployed, <strong>and</strong> percent covered by Medicaid. Additional state legislative factors are <strong>in</strong>surance market regulations <strong>and</strong><br />
Medicaid Medically Needy program parameters. See text for details. All specifications <strong>in</strong>clude an <strong>in</strong>dicator for the bottom-cod<strong>in</strong>g of seizable assets. Robust st<strong>and</strong>ard errors<br />
clustered at the state level <strong>in</strong> parentheses. Pooled 1999 to 2005 PSID exclud<strong>in</strong>g households with <strong>public</strong> <strong>in</strong>surance or a head age 65 or older. Inflation-adjusted to 2005<br />
us<strong>in</strong>g the CPI-U.
CHAPTER 1. BANKRUPTCY 56<br />
Table 1.5: Heterogeneity <strong>in</strong> the Effect on Coverage<br />
Younger household Older household<br />
Lower <strong>in</strong>come (< Higher <strong>in</strong>come (≥ Lower seizable Higher seizable<br />
head(
CHAPTER 1. BANKRUPTCY 57<br />
Table 1.6: The Effect on Assets<br />
Dependent variable:<br />
1(homeowner) ln(home equity) 1(homeowner) ln(home equity) 1(vehicle owner) ln(vehicle equity) 1(wealth > 0) ln(wealth)<br />
OLS 2SLS OLS<br />
(1) (2) (3) (4) (5) (6) (7) (8)<br />
Homestead <strong>in</strong>strument -0.009 0.000 0.066 0.000<br />
(0.014) (0.000) (0.149) (0.000)<br />
Vehicle <strong>in</strong>strument -0.088 -0.288<br />
(0.064) (0.149)<br />
Basel<strong>in</strong>e <strong>in</strong>strument 0.008 0.000 -0.078 0.000 -0.025 -0.047 -0.011 -0.011<br />
(0.031) (0.000) (0.178) (0.000) (0.014) (0.041) (0.007) (0.083)<br />
Demographic controls Yes Yes Yes Yes Yes Yes Yes Yes<br />
Income polynomial (4th order) Yes Yes Yes Yes Yes Yes Yes Yes<br />
Wealth polynomial (4th order) Yes Yes Yes Yes Yes Yes<br />
Instrument<br />
Historical homestead Yes Yes<br />
R-squared 0.47 1.00 - - 0.16 0.35 0.10 0.43<br />
N 20,265 17,293 17,418 14,836 20,265 17,699 20,265 17,293<br />
Notes: Regressions of assets (<strong>in</strong>dicators for positive values <strong>and</strong> log values) on state-level measures of asset exemption law (the <strong>in</strong>struments). The homestead <strong>in</strong>strument is mean log seizable<br />
home equity for a constant, nationally representative sample of households as though they lived <strong>in</strong> each state. The historical, vehicle, <strong>and</strong> basel<strong>in</strong>e <strong>in</strong>struments are constructed similarly us<strong>in</strong>g<br />
seizable home equity under <strong>in</strong>flation-adjusted 1920 homestead exemption laws, seizable vehicle equity, <strong>and</strong> seizable assets. Demographic controls are dummies for age group, family<br />
structure, race, <strong>and</strong> education. Pooled 1999 to 2005 PSID exclud<strong>in</strong>g households with <strong>public</strong> <strong>in</strong>surance or a head age 65 or older. Inflation-adjusted to 2005 us<strong>in</strong>g the CPI-U. Homeownership<br />
rate is 72.5 percent. Vehicle ownership rate is 91.0 percent.
CHAPTER 1. BANKRUPTCY 58<br />
Table 1.7: The Effect on Premiums<br />
Dependent variable: ln(premium)<br />
OLS 2SLS<br />
(1) (2) (3) (4) (5) (6)<br />
Basel<strong>in</strong>e <strong>in</strong>strument -0.098 -0.098 -0.074 -0.093 -0.111 -0.160<br />
State controls<br />
(0.059) (0.053) (0.038) (0.106) (0.107) (0.076)<br />
Medicaid (share) -0.354 -0.289 -0.417 -0.617<br />
(1.004) (0.794) (1.185) (0.890)<br />
Income 0.000 0.000 0.000 0.000<br />
(0.000) (0.000) (0.000) (0.000)<br />
Unemployment (share) -0.065 -0.540 1.120 -0.333<br />
Additional state legislative factors<br />
(2.675) (1.850) (3.060) (1.767)<br />
Coverage m<strong>and</strong>ates 0.003 0.000<br />
(0.004) (0.004)<br />
Any will<strong>in</strong>g phamacists 0.058 0.031<br />
(0.051) (0.058)<br />
Any will<strong>in</strong>g provider -0.026 -0.050<br />
(0.060) (0.067)<br />
Community rat<strong>in</strong>g 0.532 -0.016<br />
(0.089) (0.070)<br />
Guarenteed issue -0.450 0.091<br />
(0.133) (0.135)<br />
Charity care pool -0.140 -0.207<br />
(0.060) (0.064)<br />
Plan FE Yes Yes Yes Yes Yes Yes<br />
Instrument<br />
Historical <strong>in</strong>strument Yes Yes Yes<br />
R-squared 0.833 0.834 0.865 - - -<br />
N 1,891 1,891 1,891 1,420 1,420 1,420<br />
Notes: The dependent variable is the monthly premium for a 30-year-old non-smok<strong>in</strong>g male<br />
for plans offered by eHealthInsurance. The basel<strong>in</strong>e <strong>in</strong>strument is mean log seizable assets<br />
for a constant, nationally representative sample of households as though they lived <strong>in</strong> that<br />
state. In the 2SLS specifications, mean seizable home equity under 1920 homestead<br />
exemptions laws is used as an <strong>in</strong>strument. Robust st<strong>and</strong>ard errors clustered at the state<br />
level <strong>in</strong> parentheses. Premiums <strong>in</strong>flation-adjusted to 2005 us<strong>in</strong>g the CPI-U.
CHAPTER 1. BANKRUPTCY 59<br />
Table 1.8: Basel<strong>in</strong>e Costs Estimates<br />
(1) (2) (3) (4) (5) (6)<br />
ln(w S Dependent variable: ln(out-of-pocket+1)<br />
OLS 2SLS<br />
) 0.561 0.219 0.271 0.572 0.368 0.350<br />
Age group<br />
(0.014) (0.030) (0.037) (0.014) (0.040) (0.038)<br />
35-44 -0.118 -0.092 -0.237 -0.123<br />
(0.119) (0.120) (0.128) (0.121)<br />
45-54 -0.136 -0.075 -0.255 -0.088<br />
(0.136) (0.133) (0.146) (0.134)<br />
55-64 -0.210 -0.088 -0.368 -0.089<br />
Family structure<br />
(0.155) (0.158) (0.162) (0.158)<br />
Couple 0.226 0.298 0.245 0.335<br />
(0.103) (0.096) (0.107) (0.098)<br />
S<strong>in</strong>gle parent 0.009 0.026 0.083 0.065<br />
(0.121) (0.121) (0.113) (0.119)<br />
Couple with children -0.217 -0.135 -0.138 -0.072<br />
Race<br />
(0.146) (0.141) (0.153) (0.139)<br />
Non-white 0.011 -0.043 -0.049 -0.092<br />
Education<br />
(0.119) (0.122) (0.122) (0.120)<br />
High school to some college 0.205 0.173 -0.001 0.095<br />
(0.113) (0.116) (0.115) (0.115)<br />
College or greater 0.219 0.254 -0.067 0.167<br />
(0.175) (0.176) (0.179) (0.181)<br />
Income polynomial (4th order) Yes Yes Yes Yes<br />
Wealth polynomial (4th order) Yes Yes<br />
Relative Risk Score polynomial (4th order) Yes Yes Yes Yes<br />
Instrument<br />
Basel<strong>in</strong>e <strong>in</strong>strument Yes Yes Yes<br />
R-squared 0.856 0.881 0.882 - - -<br />
N 3,401 3,401 3,401 3,401 3,401 3,401<br />
Notes: Regressions of out-of-pocket payments on seizable assets <strong>in</strong> the sample of un<strong>in</strong>sured households with positive<br />
medical utilization. In the 2SLS specifications, the basel<strong>in</strong>e <strong>in</strong>strument is mean log seizable assets for a constant,<br />
nationally representative sample of households as though they lived <strong>in</strong> state state. The excluded demographic groups are<br />
age 18-34, s<strong>in</strong>gle, less than high school, <strong>and</strong> white. The Relative Risk Score is a measure of utilization based on medical<br />
diagnoses. All specifications <strong>in</strong>clude an <strong>in</strong>dicator for the bottom-cod<strong>in</strong>g of seizable assets. Robust st<strong>and</strong>ard errors<br />
clustered at the state level <strong>in</strong> parentheses. Pooled 2000 to 2005 MEPS <strong>in</strong>flation-adjusted to 2005 us<strong>in</strong>g the CPI-U.
CHAPTER 1. BANKRUPTCY 60<br />
Table 1.9: Sensitivity Analysis of the Effect on Costs<br />
Nonhomestead<br />
IV Historical IV<br />
High/low<br />
utilization,<br />
basel<strong>in</strong>e IV<br />
High/low<br />
utilization,<br />
historical IV<br />
Hospital Homestead<br />
controls IV<br />
(1) (2) (3) (4) (5) (5)<br />
ln(w S ) 0.380 0.377 0.382 0.375<br />
(0.066) (0.038) (0.039) (0.039)<br />
ln(w S ) X 1(High utilization) 0.512 0.506<br />
(0.044) (0.046)<br />
ln(w S ) X 1(Low utilization) 0.360 0.372<br />
(0.038) (0.037)<br />
Demographic controls Yes Yes Yes Yes Yes Yes<br />
State hospital factors<br />
Nonprofit share -0.123<br />
(0.439)<br />
Forprofit share 0.913<br />
(0.694)<br />
DSH payments per 1,000 0.000<br />
(0.000)<br />
Charity care pool <strong>in</strong>dicator 0.009<br />
(0.214)<br />
FQHC per 100,000 0.043<br />
(0.043)<br />
Dependent variable: ln(out-of-pocket+1)<br />
Relative Risk Score polynomial (4th order) Yes Yes Yes Yes Yes Yes<br />
Instrument<br />
Basel<strong>in</strong>e <strong>in</strong>strument Yes Yes<br />
Homestead <strong>in</strong>strument Yes<br />
Non-homestead <strong>in</strong>strument Yes<br />
Historical <strong>in</strong>strument Yes Yes<br />
N 3,401 3,401 3,401 2,839 3,401 2,839<br />
Notes: Regressions of out-of-pocket payments on seizable assets <strong>in</strong> the sample of un<strong>in</strong>sured households with positive medical<br />
utilization. The basel<strong>in</strong>e <strong>in</strong>strument is mean log seizable assets by state for a constant, nationally representative sample of households<br />
as though they lived <strong>in</strong> each state. The homestead, non-homestead, <strong>and</strong> historical homestead <strong>in</strong>struments are constructed similarly<br />
us<strong>in</strong>g variation <strong>in</strong> seizable home equity, non-homestead seizable assets, <strong>and</strong> seizable home equity under <strong>in</strong>flation-adjusted 1920<br />
homestead exemption laws. Demographic controls are dummies for age group, family structure, race, <strong>and</strong> education. The Relative Risk<br />
Score is a measure of utilization based on medical diagnoses. All specifications <strong>in</strong>clude an <strong>in</strong>dicator for the bottom-cod<strong>in</strong>g of seizable<br />
assets. Robust st<strong>and</strong>ard errors clustered at the state level <strong>in</strong> parentheses. Pooled 2000 to 2005 MEPS <strong>in</strong>flation-adjusted to 2005 us<strong>in</strong>g<br />
the CPI-U.
CHAPTER 1. BANKRUPTCY 61<br />
Table 1.10: Micro-Simulation Estimates of Percent Covered Without <strong>and</strong> With<br />
Bankruptcy Insurance by Insurance Status<br />
Without bankruptcy<br />
Pr(WTP > premium)<br />
With bankruptcy Difference<br />
Low risk aversion<br />
Un<strong>in</strong>sured 34% 5% 28%<br />
Insured 62% 48% 14%<br />
Difference<br />
Moderate risk aversion<br />
-28% -43% 14%*<br />
Un<strong>in</strong>sured 71% 11% 60%<br />
Insured 85% 62% 23%<br />
Difference<br />
High risk aversion<br />
-14% -51% 37%*<br />
Un<strong>in</strong>sured 91% 21% 70%<br />
Insured 99% 75% 24%<br />
Difference -8% -54% 46%*<br />
Notes: Micro-simulation estimates of the probability that the will<strong>in</strong>gness to pay exceeds the<br />
premium by <strong>in</strong>surance status for a $2,000 deductible plan. Will<strong>in</strong>gness to pay (WTP) calculated<br />
us<strong>in</strong>g CARA utility with parameters of 2.5 x 10^-5 (low risk aversion), 5.0 x 10^-5 (moderate<br />
risk aversion), <strong>and</strong> 7.5 X 10^-5 (high risk aversion). Premiums calculated as the expected<br />
value of medical costs above the deductible scaled up to account for moral hazard (elasticity of<br />
-0.22), adm<strong>in</strong>istrative load<strong>in</strong>g (10 <strong>and</strong> 50 percent), <strong>and</strong> the cross subsidization of unpaid care<br />
(endogenously determ<strong>in</strong>ed). Household-level estimates weighted by the number of <strong>in</strong>dividuals<br />
per household for <strong>in</strong>terpretation at the <strong>in</strong>dividual level.<br />
*Difference-<strong>in</strong>-differences.
CHAPTER 1. BANKRUPTCY 62<br />
Table 1.11: Policy Counterfactuals<br />
Penalty Take-up ! WTP ! Cost ! Surplus<br />
Pigovian penalty<br />
Low risk aversion $218.21 7.5% $10.02 $6.06 $3.95<br />
Moderate risk aversion $218.21 7.3% $11.02 $6.83 $4.19<br />
High risk aversion<br />
PPACA penalty<br />
$218.21 7.9% $12.30 $7.72 $4.58<br />
Low risk aversion $481.43 63.6% -$7.38 $5.34 -$12.72<br />
Moderate risk aversion $481.43 56.4% $3.02 $14.40 -$11.38<br />
High risk aversion<br />
Medical debt non-dischargeable<br />
$481.43 49.7% $11.09 $20.24 -$9.15<br />
Low risk aversion n/a 100.0% -$48.26 -$12.12 -$36.14<br />
Moderate risk aversion n/a 100.0% -$28.18 $12.75 -$40.93<br />
High risk aversion n/a 100.0% -$6.43 $36.74 -$43.17<br />
Notes: Micro-simulation estimates of <strong>in</strong>surance take-up, will<strong>in</strong>gness to pay, costs, <strong>and</strong> social surplus<br />
from different penalty systems relative to a basel<strong>in</strong>e <strong>in</strong> which households can choose bankruptcy at no<br />
cost. The Pigovian penalty is the household-specific social cost of the implicit <strong>in</strong>surance from bankruptcy.<br />
PPACA is the <strong>in</strong>flation-adjusted, fully phased-<strong>in</strong> penalty under this legislation, def<strong>in</strong>ed as the greater of<br />
$625 or 2.5 percent of <strong>in</strong>come, up to a maximum of $2,085 per household. Medical debt nondischargeable<br />
exposes households to the full f<strong>in</strong>ancial risk when un<strong>in</strong>sured. Take-up is the percent of<br />
un<strong>in</strong>sured <strong>in</strong>dividuals that take up coverage. Will<strong>in</strong>gness to pay (WTP) is calculated us<strong>in</strong>g CARA utility<br />
with parameters of 2.5 x 10^-5 (low risk aversion), 5.0 x 10^-5 (moderate risk aversion), <strong>and</strong> 7.5 x<br />
10^-5 (high risk aversion). Household-level estimates weighted by number of <strong>in</strong>dividuals per household<br />
for <strong>in</strong>terpretation at the <strong>in</strong>dividual level.
CHAPTER 1. BANKRUPTCY 63<br />
Table 1.12: Summary Statistics: Seizable Assets by Insurance Status<br />
Mean Std. Dev. 5th 25th<br />
Percentile<br />
50th 75th 95th<br />
Panel A: All (n = 22,844)<br />
Seizable assets $216,943 $1,105,939 -$10,148 $2,344 $34,328 $155,219 $796,366<br />
Gross seizable assets $221,434 $1,106,139 $0 $3,500 $38,400 $158,699 $797,593<br />
Seizable home equity (70.4% homeownership) $52,487 $134,239 $0 $0 $0 $53,760 $249,205<br />
Other seizable assets $168,948 $1,054,503 $0 $1,010 $18,000 $81,180 $616,680<br />
Dischargeable debt $6,659 $15,675 $0 $0 $1,024 $7,305 $28,678<br />
Fil<strong>in</strong>g costs<br />
Panel B: Privately Insured (n = 20,197)<br />
$2,000 $0 $2,000 $2,000 $2,000 $2,000 $2,000<br />
Seizable assets $233,241 $1,153,125 -$9,880 $3,423 $42,809 $168,945 $849,475<br />
Gross seizable assets $237,970 $1,153,292 $0 $5,860 $46,294 $172,870 $852,001<br />
Seizable home equity (73.3% homeownership) $56,416 $139,145 $0 $0 $0 $60,320 $260,000<br />
Other seizable assets $181,554 $1,099,952 $0 $2,308 $21,500 $90,649 $654,239<br />
Dischargeable debt $6,898 $15,979 $0 $0 $1,297 $7,684 $29,511<br />
Fil<strong>in</strong>g costs<br />
Panel C: Un<strong>in</strong>sured (n = 2,647)<br />
$2,000 $0 $2,000 $2,000 $2,000 $2,000 $2,000<br />
Seizable assets $41,658 $207,719 -$13,305 $2,000 $2,344 $14,937 $212,000<br />
Gross seizable assets $43,586 $207,884 $0 $0 $637 $16,000 $218,871<br />
Seizable home equity (38.5% homeownership) $10,222 $40,560 $0 $0 $0 $0 $75,000<br />
Other seizable assets $33,364 $197,079 $0 $0 $0 $9,348 $158,089<br />
Dischargeable debt $4,095 $11,615 $0 $0 $0 $3,264 $20,304<br />
Fil<strong>in</strong>g costs $2,000 $0 $2,000 $2,000 $2,000 $2,000 $2,000<br />
Notes: Household-level statistics from the pooled 1999 to 2005 PSID <strong>in</strong>flation-adjusted to 2005 us<strong>in</strong>g the CPI-U. Excludes households with a head age 65 or<br />
older <strong>and</strong> those with <strong>public</strong> <strong>in</strong>surance. See text seizable assets calculation for details.<br />
Table 1.13: Summary Statistics: Medical Costs by Insurance Status<br />
Mean Std. Dev. 5th 25th<br />
Percentile<br />
50th 75th 95th<br />
Panel A: All (n = 34,841)<br />
Charges $6,647 $17,781 $0 $420 $1,836 $6,099 $27,029<br />
Total payments $4,085 $9,704 $0 $309 $1,388 $4,275 $15,859<br />
Private payments $2,974 $8,648 $0 $37 $661 $2,701 $12,448<br />
Public payments $85 $1,094 $0 $0 $0 $0 $78<br />
Misc payments $243 $2,102 $0 $0 $0 $0 $715<br />
Out-of-pocket payments<br />
Panel B: Insured (n = 31,753)<br />
$783 $1,565 $0 $76 $340 $907 $2,934<br />
Charges $7,201 $18,437 $0 $590 $2,201 $6,860 $28,617<br />
Total payments $4,480 $10,131 $0 $455 $1,678 $4,786 $16,767<br />
Private payments $3,391 $9,157 $0 $190 $954 $3,209 $13,528<br />
Public payments $83 $1,078 $0 $0 $0 $0 $89<br />
Misc payments $204 $1,673 $0 $0 $0 $0 $567<br />
Out-of-pocket payments<br />
Panel C: Un<strong>in</strong>sured (n = 3,088)<br />
$801 $1,474 $0 $103 $375 $943 $2,939<br />
Charges $2,691 $11,353 $0 $0 $232 $1,429 $11,022<br />
Total payments $1,267 $4,966 $0 $0 $136 $858 $5,231<br />
Private payments $0 $0 $0 $0 $0 $0 $0<br />
Public payments $96 $1,200 $0 $0 $0 $0 $0<br />
Misc payments $515 $3,989 $0 $0 $0 $40 $1,678<br />
Out-of-pocket payments $657 $2,097 $0 $0 $80 $543 $2,906<br />
Notes: Household-level statistics from the pooled 2000 to 2005 MEPS <strong>in</strong>flation-adjusted to 2005 us<strong>in</strong>g the CPI-U. Charges<br />
are the list price of medical care received. Private payments are from private <strong>in</strong>surance providers (e.g., employer sponsored<br />
<strong>in</strong>surance, <strong>in</strong>dividual market <strong>in</strong>surance), <strong>public</strong> payments are from <strong>public</strong> <strong>in</strong>surance (e.g, Medicaid), miscellaneous<br />
payments are from other sources (e.g. Workman's Compensation), <strong>and</strong> out-of-pocket payments are from households. Total<br />
payments are the sum of the payment values.
CHAPTER 1. BANKRUPTCY 64<br />
Table 1.14: Summary Statistics: Premiums<br />
N Mean Std. Dev. M<strong>in</strong>. Max.<br />
Premium (per month) 1,891 103 46 30 503<br />
Deductible 1,891 3,351 2,229 0 8,963<br />
Co<strong>in</strong>surance (%) 1,891 15 12 0 50<br />
Notes: Plan characteristics for all plans offered to a 30-year-old non-smok<strong>in</strong>g male <strong>in</strong> each state<br />
on eHealthInsurance. Values <strong>in</strong>flation-adjusted to 2005 us<strong>in</strong>g the CPI-U.<br />
Table 1.15: Costs <strong>and</strong> Premiums by Deductible Level<br />
Deductible Pr(Costs > deductible) Calibrated<br />
Premium<br />
Individual Market<br />
$0 46.0% $2,126 $2,140<br />
$1,000 10.4% $1,735 $2,061<br />
$2,000 5.9% $1,546 N/A<br />
$5,000 2.7% $1,259 $874<br />
$10,000 1.5% $1,009 N/A<br />
Notes: Costs are medical costs when un<strong>in</strong>sured. Calibrated premiums are calculated as the<br />
expected value of medical costs above the deductible scaled up to account for moral hazard<br />
(elasticity of -0.22) <strong>and</strong> adm<strong>in</strong>istrative load<strong>in</strong>g (50 percent). Individual market premiums are for<br />
policies start<strong>in</strong>g <strong>in</strong> May 2010 issued by Aetna. They are adjusted for <strong>in</strong>flation us<strong>in</strong>g the Medical<br />
Care component of the CPI-U. See Appendix D for details.
Chapter 2<br />
Pric<strong>in</strong>g <strong>and</strong> Welfare <strong>in</strong> Health Plan<br />
Choice<br />
with M. Kate Bundorf <strong>and</strong> Jonathan Lev<strong>in</strong><br />
2.1 Introduction<br />
Whether competition <strong>in</strong> health <strong>in</strong>surance markets leads to efficient outcomes is a<br />
central question for health policy. Markets are effective when prices direct consumers<br />
<strong>and</strong> firms to behave efficiently. But <strong>in</strong> health <strong>in</strong>surance markets, prices often do<br />
not reflect the different costs of coverage for different enrollees. This generates two<br />
concerns. If <strong>in</strong>surers receive premiums that do not reflect enrollee risk, they have an<br />
<strong>in</strong>centive to engage <strong>in</strong> risk selection through plan design (??). Similarly, if consumers<br />
face prices that do not reflect cost differences across plans, they may select coverage<br />
<strong>in</strong>efficiently (??). While it is widely recognized that these problems may impair the<br />
efficiency of competitive health <strong>in</strong>surance markets, evidence on their quantitative<br />
importance for social welfare is limited.<br />
In the U.S. private market, employers generally contract with <strong>in</strong>surers to cre-<br />
ate a menu of plans from which employees select coverage. The government or a<br />
quasi-<strong>public</strong> <strong>organization</strong> plays a similar role <strong>in</strong> the U.S. Medicare program <strong>and</strong> the<br />
national systems of Germany <strong>and</strong> the Netherl<strong>and</strong>s. To address <strong>in</strong>centive problems <strong>in</strong><br />
65
CHAPTER 2. HEALTH PLAN CHOICE 66<br />
plan design, these <strong>in</strong>termediaries have begun to “risk-adjust” payments to plans (?).<br />
Consumer prices, however, are typically not adjusted for <strong>in</strong>dividual risk. Indeed, <strong>in</strong><br />
the U.S., regulations prohibit employers <strong>and</strong> <strong>public</strong> programs from charg<strong>in</strong>g enrollees<br />
different amounts based on nearly all health-related factors. 1 Moreover, even with<strong>in</strong><br />
<strong>in</strong>stitutional limitations, contributions set by employers or <strong>in</strong> regulated markets may<br />
not be welfare-maximiz<strong>in</strong>g given the complexities of self-selection.<br />
In this paper, we analyze the effect of plan pric<strong>in</strong>g on allocative efficiency. We<br />
beg<strong>in</strong> by mak<strong>in</strong>g a basic theoretical po<strong>in</strong>t regard<strong>in</strong>g plan prices <strong>and</strong> efficient match-<br />
<strong>in</strong>g. Exist<strong>in</strong>g work suggests that, while poorly chosen contribution policies may lead<br />
to <strong>in</strong>efficient outcomes, the problem can be solved by choos<strong>in</strong>g an optimal uniform<br />
contribution (e.g., ????). These analyses, however, make strong assumptions regard-<br />
<strong>in</strong>g the correlation between enrollee risk <strong>and</strong> preferences for coverage. We show that<br />
if these assumptions are violated, no uniform contribution policy leads to efficient<br />
consumer choices.<br />
The ma<strong>in</strong> part of the paper exam<strong>in</strong>es the implications of this observation. We<br />
develop a simple econometric model of health plan dem<strong>and</strong> <strong>and</strong> costs, estimate the<br />
model on a novel dataset of small employers, <strong>and</strong> then use the parameter estimates to<br />
simulate the welfare implications of alternative pric<strong>in</strong>g policies. Based on our simula-<br />
tions, we estimate that, <strong>in</strong> this sett<strong>in</strong>g, observed pric<strong>in</strong>g policies are less efficient than<br />
what could be achieved with risk-rated plan contributions. The shortfall is between<br />
$60 <strong>and</strong> $325 annually per enrollee, or 2-11% of coverage costs. Employers could<br />
realize approximately 1/4 of this surplus by adjust<strong>in</strong>g their non-risk-rated contribu-<br />
tions, but captur<strong>in</strong>g the rema<strong>in</strong>der would require sett<strong>in</strong>g different prices for people <strong>in</strong><br />
the same firm. We also f<strong>in</strong>d that employees select plans based on private <strong>in</strong>formation<br />
about their health status. A hypothetical social planner who <strong>in</strong>corporated this private<br />
<strong>in</strong>formation <strong>in</strong>to prices could <strong>in</strong>crease welfare by an additional $35-$100 annually per<br />
enrollee. Despite these <strong>in</strong>efficiencies, the observed plan offer<strong>in</strong>gs generate substantial<br />
benefits over any s<strong>in</strong>gle-plan offer<strong>in</strong>g.<br />
An underst<strong>and</strong><strong>in</strong>g of the characteristics of the two <strong>in</strong>surers we study is important<br />
for <strong>in</strong>terpret<strong>in</strong>g these results. One has a fairly broad network of providers <strong>and</strong> relies<br />
1 Premiums vary by smok<strong>in</strong>g status <strong>in</strong> some markets.
CHAPTER 2. HEALTH PLAN CHOICE 67<br />
on patient cost shar<strong>in</strong>g <strong>and</strong> primary care gatekeepers to control utilization. The other<br />
has an <strong>in</strong>tegrated <strong>and</strong> closed delivery system <strong>and</strong> requires little patient cost shar<strong>in</strong>g.<br />
Estimates from the cost model suggest that these <strong>in</strong>surers have very different cost<br />
structures. While costs are similar between the two for <strong>in</strong>dividuals of average health,<br />
the <strong>in</strong>tegrated delivery system has higher costs for healthy enrollees <strong>and</strong> significantly<br />
lower costs for less healthy enrollees.<br />
On the dem<strong>and</strong> side, we f<strong>in</strong>d that household preferences as well as health status<br />
affect plan choice. However, <strong>in</strong> contrast to other studies, we f<strong>in</strong>d that no s<strong>in</strong>gle plan<br />
experiences systematic adverse selection. An <strong>in</strong>terest<strong>in</strong>g feature of our data is that<br />
the plans are not obviously ordered by coverage level: consumers have to trade off<br />
provider network restrictions aga<strong>in</strong>st the level of cost shar<strong>in</strong>g. This “horizontal” plan<br />
differentiation differentiates our sett<strong>in</strong>g from that of many other studies <strong>in</strong> which<br />
consumers choose how much coverage to purchase. We view this horizontal plan<br />
differentiation as the likely explanation for the lack of systematic adverse selection.<br />
This type of differentiation seems particularly salient given the recent evolution<br />
of the health <strong>in</strong>surance market. In 1987, approximately three-quarters of people with<br />
employer-sponsored health <strong>in</strong>surance had conventional coverage, under which plans<br />
differed primarily <strong>in</strong> their cost shar<strong>in</strong>g. By 2007, <strong>in</strong> contrast, the market was domi-<br />
nated by managed care plans which use a mix of supply-side <strong>and</strong> dem<strong>and</strong>-side utiliza-<br />
tion management (?). This evolution suggests that classic <strong>in</strong>sights based on purely<br />
risk-based sort<strong>in</strong>g may not adequately capture the dynamics of today’s market. 2<br />
The <strong>in</strong>efficiencies we f<strong>in</strong>d are driven by two forces: heterogeneity <strong>in</strong> household<br />
preferences <strong>and</strong> the significant cost advantage of the <strong>in</strong>tegrated system for <strong>in</strong>dividuals<br />
<strong>in</strong> worse health. Our estimates suggest that although high risk households have some<br />
preference for flexibility, a large fraction would choose the <strong>in</strong>tegrated delivery system<br />
if they had to <strong>in</strong>ternalize the relevant cost differential between plans. Achiev<strong>in</strong>g this<br />
with a uniform contribution policy, however, would require that all households face a<br />
steep premium for the more flexible <strong>in</strong>surer. This <strong>in</strong> turn would create a welfare loss<br />
for lower risk households who value flexibility. While the exact numbers we estimate<br />
2 ? stress that a broad view of heterogeneity <strong>in</strong> preferences is important for underst<strong>and</strong><strong>in</strong>g many<br />
aspects of <strong>in</strong>surance markets.
CHAPTER 2. HEALTH PLAN CHOICE 68<br />
are of course specific to our sett<strong>in</strong>g, the trade-off we identify may be relevant more<br />
broadly. This analysis suggests that the <strong>in</strong>tegrated model of health care delivery faces<br />
an important challenge <strong>in</strong> health <strong>in</strong>surance markets: current pric<strong>in</strong>g policies make it<br />
difficult to target households for whom the <strong>in</strong>tegrated model promises substantial cost<br />
sav<strong>in</strong>gs.<br />
Our analysis ties <strong>in</strong> to past work on health plan choice <strong>and</strong> the efficiency of<br />
health <strong>in</strong>surance markets. We draw on this work on health plan choice, which is<br />
summarized by ? <strong>and</strong> ?, <strong>in</strong> model<strong>in</strong>g how employee dem<strong>and</strong> varies with observed <strong>and</strong><br />
unobserved risk <strong>and</strong> preference characteristics. Our paper is more directly related to<br />
recent work that uses econometric methods to quantify the efficiency implications of<br />
adverse selection <strong>in</strong> health <strong>in</strong>surance markets (?????). Our paper po<strong>in</strong>ts out that<br />
uniform pric<strong>in</strong>g, as is commonly observed, may lead to <strong>in</strong>efficiency when enrollees of<br />
similar risk have different preferences for coverage. These other papers, <strong>in</strong> contrast,<br />
analyze alternative <strong>in</strong>stitutional features of health <strong>in</strong>surance markets that contribute<br />
to adverse selection. We relate both our empirical approach <strong>and</strong> our f<strong>in</strong>d<strong>in</strong>gs to these<br />
papers <strong>in</strong> Section 2.5.4.<br />
We emphasize that our analysis has some important limitations. First, it is based<br />
on a particular, <strong>and</strong> only moderately-sized, sample of workers <strong>and</strong> firms. To address<br />
this, we perform a variety of sensitivity analyses on our key parameter estimates,<br />
which we discuss <strong>in</strong> the last section. Second, we take plan offer<strong>in</strong>gs as given. This<br />
seems reasonable given that we are look<strong>in</strong>g at small to medium size employers, but a<br />
broader analysis of pric<strong>in</strong>g ideally would <strong>in</strong>corporate plan design. Third, we do not<br />
address issues of utilization behavior, or try to assess the relative social efficiency of<br />
health care utilization under the different plans <strong>in</strong> our data. F<strong>in</strong>ally, our analysis is<br />
based on a static model, so we abstract from issues of dynamic <strong>in</strong>surance. We discuss<br />
this issue <strong>in</strong> the conclusion.<br />
2.2 Health Plan Pric<strong>in</strong>g <strong>and</strong> Market Efficiency<br />
We illustrate the relationship between pric<strong>in</strong>g <strong>and</strong> market efficiency by adapt<strong>in</strong>g the<br />
model of ?. In their model, consumers are dist<strong>in</strong>guished by their forecastable health
CHAPTER 2. HEALTH PLAN CHOICE 69<br />
risk, denoted θ. Each consumer chooses between a high-cost plan A <strong>and</strong> a low cost<br />
plan B. We can th<strong>in</strong>k of the plans, for now, as vertically differentiated. The plans’<br />
expected costs of cover<strong>in</strong>g a type-θ consumer are cA(θ) <strong>and</strong> cB(θ). Let ∆c (θ) =<br />
cA(θ) − cB(θ) denote the cost differential. We assume ∆c is strictly positive <strong>and</strong><br />
<strong>in</strong>creas<strong>in</strong>g <strong>in</strong> θ.<br />
Let vA(θ) <strong>and</strong> vB(θ) denote a type θ’s expected (dollar) value from be<strong>in</strong>g covered<br />
by each of the plans. For the moment, the benefits of coverage are determ<strong>in</strong>ed only<br />
by forecastable health risk. We assume that contributions vary across plans but<br />
not across consumers. A consumer who makes a contribution pj to enroll <strong>in</strong> plan<br />
j ∈{A, B} gets a net benefit vj(θ) − pj. 3 Def<strong>in</strong>e ∆v(θ) =vA(θ) − vB(θ) to be the<br />
additional amount a type-θ consumer would pay for the high-cost plan.<br />
The efficient assignment places a type-θ consumer <strong>in</strong> plan A if <strong>and</strong> only if<br />
∆v(θ) − ∆c(θ) ≥ 0.<br />
At the same time, a type-θ consumer will select plan A if <strong>and</strong> only if<br />
∆v(θ) − ∆p ≥ 0,<br />
where ∆p = pA − pB is the <strong>in</strong>cremental contribution for plan A.<br />
Are there prices that lead to an efficient outcome? Assume that ∆v(θ) is <strong>in</strong>creas<strong>in</strong>g<br />
<strong>in</strong> θ, which seems appropriate if plan A simply offers more generous coverage or easier<br />
access to care. Then for any <strong>in</strong>cremental contribution ∆p, atype-θ consumer will<br />
choose A if <strong>and</strong> only if θ ≥ θ(∆p), where θ (∆p) is a threshold that can be varied<br />
arbitrarily with ∆p. 4 Therefore it is possible to achieve efficient sort<strong>in</strong>g if <strong>and</strong> only if<br />
the efficient assignment also <strong>in</strong>volves a threshold rule, i.e. if ∆v(θ)−∆c(θ) is negative<br />
up to some θ ∗ <strong>and</strong> positive above it. Intuitively, the requirement for efficiency is that<br />
will<strong>in</strong>gness to pay <strong>in</strong>creases more quickly with risk than the cost differential between<br />
3 Here we make the simplify<strong>in</strong>g assumption, which we ma<strong>in</strong>ta<strong>in</strong> <strong>in</strong> our econometric model, that<br />
plan preferences are additively separable <strong>in</strong> the plan premium. See ? for an extensive discussion of<br />
this assumption.<br />
4 An empirical prediction of this model is that plan A will experience unfavorable selection, <strong>and</strong><br />
its risk composition will be worse the larger is ∆p.
CHAPTER 2. HEALTH PLAN CHOICE 70<br />
plans.<br />
Exist<strong>in</strong>g analyses assume that the surplus function has the requisite s<strong>in</strong>gle cross<strong>in</strong>g<br />
property (e.g., ????). In this case, shown <strong>in</strong> Figure 1:(a), efficiency can be achieved<br />
by sett<strong>in</strong>g ∆p =∆c(θ ∗ ). The problem emphasized <strong>in</strong> the literature is that purchasers<br />
may not choose the correct premium differential. If ∆p is too high, plan A attracts<br />
only very high risks, <strong>and</strong> if prices are based on past outcomes, one can even end<br />
up with an adverse selection “death spiral” (?). Alternatively, if ∆p is too low, too<br />
many people select plan A, <strong>in</strong>clud<strong>in</strong>g some for whom the benefits are less than the<br />
<strong>in</strong>cremental social costs.<br />
This familiar analysis can be questioned on several levels. First, even if we cont<strong>in</strong>ue<br />
to assume that consumers differ only <strong>in</strong> their health status <strong>and</strong> plans differ ma<strong>in</strong>ly <strong>in</strong><br />
the amount of coverage they offer, it could be socially efficient for high risks to be <strong>in</strong> a<br />
cost-conscious plan. 5 For <strong>in</strong>stance, the cost sav<strong>in</strong>gs from a plan that actively manages<br />
utilization might more than compensate these consumers for the loss of flexibility. In<br />
this case, shown <strong>in</strong> Figure 1(b), uniform pric<strong>in</strong>g is <strong>in</strong>herently <strong>in</strong>efficient because there<br />
is no way to <strong>in</strong>duce only the relatively high risks to self-select <strong>in</strong>to plan B when<br />
everyone faces the same prices.<br />
Perhaps the more obvious issue from an empirical perspective is that health plans<br />
often differ well beyond offer<strong>in</strong>g “more” or “less” coverage, <strong>and</strong> consumers differ on<br />
dimensions other than health risk. It is <strong>in</strong>creas<strong>in</strong>gly common for firms to offer em-<br />
ployees an HMO option that places greater restriction on provider choice, <strong>and</strong> a PPO<br />
option that allows broader access to provider, with greater cost shar<strong>in</strong>g. Consumers<br />
<strong>in</strong> relatively poor health may value flexibility but be wary of <strong>in</strong>creased cost-shar<strong>in</strong>g.<br />
As a result, heterogeneity <strong>in</strong> tastes or <strong>in</strong>come may be at least as important as health<br />
status <strong>in</strong> driv<strong>in</strong>g choice.<br />
To capture this, th<strong>in</strong>k of plan A as a PPO <strong>and</strong> plan B as an HMO, <strong>and</strong> suppose<br />
that consumers vary <strong>in</strong> both forecastable risk <strong>and</strong> taste. Specifically, let ε denote a<br />
5 Arguably the benefits of deliver<strong>in</strong>g care efficiently may be largest for the chronically ill. The<br />
most detailed analysis of differences <strong>in</strong> utilization between traditional Medicare coverage <strong>and</strong> Medicare<br />
managed care plans found that the reductions <strong>in</strong> utilization generated by managed care plans<br />
were concentrated among high risk beneficiaries <strong>and</strong> that these reductions <strong>in</strong> utilization were not<br />
associated with differences <strong>in</strong> short term health outcomes (?).
CHAPTER 2. HEALTH PLAN CHOICE 72<br />
risk/preference space.<br />
Each of these issues is <strong>in</strong>herently quantitative <strong>in</strong> nature. We use the econometric<br />
model developed <strong>in</strong> the next section to identify the relevant cost <strong>and</strong> dem<strong>and</strong> param-<br />
eters <strong>and</strong> then evaluate empirically how various pric<strong>in</strong>g arrangements affect social<br />
welfare.<br />
2.3 Data <strong>and</strong> Environment<br />
2.3.1 Institutional Sett<strong>in</strong>g<br />
Our analysis is based on data from a private firm that helps small <strong>and</strong> mid-sized<br />
employers manage health benefits. This firm, who we refer to as the <strong>in</strong>termediary,<br />
obta<strong>in</strong>s agreements from <strong>in</strong>surers to offer plans to small employers, signs up employ-<br />
ers, <strong>and</strong> adm<strong>in</strong>isters their health benefit. We exam<strong>in</strong>e data from 11 employers who<br />
purchased coverage from the <strong>in</strong>termediary <strong>in</strong> a s<strong>in</strong>gle metropolitan area <strong>in</strong> the western<br />
United States dur<strong>in</strong>g 2004 <strong>and</strong> 2005.<br />
In this market, the <strong>in</strong>termediary works with two <strong>in</strong>surers. One <strong>in</strong>surer contracts<br />
non-exclusively with a relatively broad set of providers. It offers an HMO plan (net-<br />
work HMO) that requires enrollees to choose a primary care physician <strong>and</strong> obta<strong>in</strong> a<br />
referral for specialist visits, <strong>and</strong> does not cover care from out-of-network providers.<br />
It also offers a PPO plan (network PPO) that does not require referrals <strong>and</strong> covers<br />
providers outside the plan’s network at an <strong>in</strong>creased cost-share. 7 The second <strong>in</strong>surer<br />
has an <strong>in</strong>tegrated <strong>and</strong> closed delivery system that facilitates supply side utilization<br />
management. It offers a st<strong>and</strong>ard HMO (<strong>in</strong>tegrated HMO) <strong>and</strong> a po<strong>in</strong>t-of-service op-<br />
tion (<strong>in</strong>tegrated POS) that allows enrollees to seek care outside the <strong>in</strong>tegrated system<br />
at a higher cost.<br />
The employers that hire the <strong>in</strong>termediary beg<strong>in</strong> by choos<strong>in</strong>g which plans to offer<br />
their employees. Employers may customize the basic plans to a limited degree by<br />
7 This <strong>in</strong>surer also offers a po<strong>in</strong>t-of-service (POS) plan that is the HMO with the option to go outof-network<br />
at higher cost. Unfortunately we are not able to dist<strong>in</strong>guish between network POS <strong>and</strong><br />
HMO enrollees. As a result, we simplify our analysis by dropp<strong>in</strong>g the three employer-years where<br />
the network POS was offered. Our results are not sensitive to alternative approaches to h<strong>and</strong>l<strong>in</strong>g<br />
this issue.
CHAPTER 2. HEALTH PLAN CHOICE 73<br />
vary<strong>in</strong>g characteristics such as the deductible <strong>and</strong> the level of co<strong>in</strong>surance, but most<br />
dimensions are fixed. Employers typically offer four coverage tiers: employee only,<br />
employee plus spouse, employee plus children, <strong>and</strong> employee plus family. 8 The level<br />
of cost shar<strong>in</strong>g varies across coverage tiers. The employers do not offer any health<br />
<strong>in</strong>surance plans beyond those offered by the <strong>in</strong>termediary.<br />
The <strong>in</strong>surers then provide bids for each of the selected plans, rely<strong>in</strong>g on <strong>in</strong>forma-<br />
tion from the <strong>in</strong>termediary. In an employer’s first year with the <strong>in</strong>termediary, this<br />
<strong>in</strong>formation is just the distribution of employees by age <strong>and</strong> sex. In subsequent years,<br />
the <strong>in</strong>surers receive additional <strong>in</strong>formation on the health status of the workers, <strong>in</strong> the<br />
form of a risk score described below. The <strong>in</strong>termediary <strong>in</strong>structs the <strong>in</strong>surers to bid<br />
as if they were cover<strong>in</strong>g all workers with<strong>in</strong> the firm. While the <strong>in</strong>surers provide bids<br />
for each tier, the bids for tiers other than employee-only are simply scaled from the<br />
employee-only bids by a constant that is very similar across employers <strong>and</strong> plans.<br />
After the bids are received, the employer sets the employee contribution for each<br />
plan <strong>and</strong> coverage tier. The employees then make their choices, <strong>and</strong> the plans are<br />
required to accept all employees who choose to enroll. The last step is a series of<br />
payments. For each employee that enrolls <strong>in</strong> a plan, the employer pays the <strong>in</strong>ter-<br />
mediary the <strong>in</strong>surer’s bid. The <strong>in</strong>termediary passes these payments to the <strong>in</strong>surers,<br />
implement<strong>in</strong>g transfers between <strong>in</strong>surers if there is variation <strong>in</strong> the health status of<br />
employees <strong>and</strong> dependents enrolled <strong>in</strong> the different plans.<br />
The <strong>in</strong>termediary uses a st<strong>and</strong>ard methodology for measur<strong>in</strong>g enrollee health sta-<br />
tus, the RxGroup model developed by DxCG, Inc. The model produces risk scores<br />
based on a person’s age, sex, <strong>and</strong> health status, where health status is <strong>in</strong>ferred from<br />
prior use of prescription drugs, reported by the <strong>in</strong>surers. 910 In order to persuade the<br />
8Two firms def<strong>in</strong>e coverage tiers based on employee only, employee plus one dependent, <strong>and</strong><br />
employee plus two or more dependents.<br />
9In our analysis, we use the term “risk score” to refer to the DxCG prediction of an <strong>in</strong>dividual’s<br />
health expenditures relative to the mean of the much larger base sample on which DxCG calibrates<br />
their model. We note that our use of the term risk refers only to the level <strong>and</strong> not to the variance<br />
of the expected expenditure, although we might naturally expect a relationship between the two.<br />
10DxCG uses an <strong>in</strong>ternally-developed algorithm to <strong>in</strong>fer the presence <strong>and</strong> severity of chronic conditions<br />
from prescription drug use. The health expenditure model is estimated on a very large sample<br />
(1,000,000+) of people under 65 with private health <strong>in</strong>surance. Us<strong>in</strong>g the estimated model, the<br />
software predicts covered health expenditures for a given <strong>in</strong>dividual. A score of 1 corresponds to a
CHAPTER 2. HEALTH PLAN CHOICE 74<br />
<strong>in</strong>surers to participate, the <strong>in</strong>termediary had to conv<strong>in</strong>ce them they would be ade-<br />
quately compensated for any unfavorable selection they might experience. Because<br />
the risk score is crucial <strong>in</strong> this regard, the <strong>in</strong>termediary worked with the <strong>in</strong>surers to<br />
ensure that the risk scores reflected the underly<strong>in</strong>g health status of enrollees rather<br />
than the methods used by the plans to manage utilization. For <strong>in</strong>stance, one concern<br />
was that the <strong>in</strong>tegrated <strong>in</strong>surer might substitute low-priced drugs more aggressively,<br />
lead<strong>in</strong>g the algorithm to under-estimate the severity of chronic illness for its enrollees.<br />
Discussions between the <strong>in</strong>termediary <strong>and</strong> the plans along these l<strong>in</strong>es led to a num-<br />
ber of m<strong>in</strong>or adjustments <strong>in</strong> the risk scor<strong>in</strong>g algorithm. Our discussions with the<br />
participants suggest that both the <strong>in</strong>termediary <strong>and</strong> the <strong>in</strong>surers had strong <strong>in</strong>cen-<br />
tives to ensure that health status was measured accurately, <strong>and</strong> that it is reasonable<br />
to assume that the risk score accurately captures employee health status <strong>and</strong> is not<br />
contam<strong>in</strong>ated by differences <strong>in</strong> the plans.<br />
In addition to prescription drug utilization, each <strong>in</strong>surer also provides the <strong>in</strong>ter-<br />
mediary with the realized costs for each employer group. The network <strong>in</strong>surer reports<br />
average claims per member per month for enrollees covered by either of the <strong>in</strong>surer’s<br />
products. The <strong>in</strong>tegrated <strong>in</strong>surer reports similar <strong>in</strong>formation developed from an <strong>in</strong>-<br />
ternal cost account<strong>in</strong>g system. Neither <strong>in</strong>surer dist<strong>in</strong>guishes between its plans when<br />
report<strong>in</strong>g this <strong>in</strong>formation.<br />
2.3.2 Data <strong>and</strong> Descriptive Statistics<br />
Our data <strong>in</strong>cludes all of the <strong>in</strong>formation discussed above: the plan offer<strong>in</strong>gs <strong>and</strong><br />
contribution policies of each employer, the risk scores <strong>and</strong> plan choices of employees<br />
<strong>and</strong> their dependents, <strong>and</strong> the bids <strong>and</strong> reported costs of each <strong>in</strong>surer. A primary<br />
strength of the data is that it <strong>in</strong>cludes both dem<strong>and</strong>-side <strong>in</strong>formation on employees<br />
<strong>and</strong> their choice behavior <strong>and</strong> supply-side <strong>in</strong>formation on <strong>in</strong>surer costs <strong>and</strong> bids <strong>in</strong> a<br />
sett<strong>in</strong>g with two very different types of <strong>in</strong>surers. In addition, many of the employers<br />
we observe offer nearly identical plans but have different risk profiles <strong>and</strong> contribution<br />
policies which provides useful variation to identify dem<strong>and</strong> <strong>and</strong> costs.<br />
mean prediction from the orig<strong>in</strong>al estimation sample. See ? for more detail.
CHAPTER 2. HEALTH PLAN CHOICE 75<br />
Another useful feature of the data is that we observe each employer dur<strong>in</strong>g their<br />
first year of participation <strong>in</strong> the program. Insurers have little <strong>in</strong>formation on firm<br />
characteristics beyond that provided by the <strong>in</strong>termediary dur<strong>in</strong>g the first year, allow-<br />
<strong>in</strong>g us to observe how plans bid when they have similar <strong>in</strong>formation on the likely risk of<br />
a group. 11 On the dem<strong>and</strong> side, a large literature documents that health plan choices<br />
are highly persistent (e.g., ?), so observ<strong>in</strong>g choice behavior <strong>in</strong> the first year likely<br />
provides a good <strong>in</strong>dication of steady-state dem<strong>and</strong> <strong>and</strong> allows us to observe the plan<br />
characteristics <strong>and</strong> prices at the time of <strong>in</strong>itial choice. The data’s ma<strong>in</strong> limitations are<br />
the fairly small number of observations <strong>and</strong> restricted set of employee characteristics<br />
relative to, say, the HR records of a large employer, <strong>and</strong> also the aggregated report<strong>in</strong>g<br />
of realized costs.<br />
The 11 firms have 2,044 covered employees <strong>and</strong> 4,652 enrollees (employees <strong>and</strong><br />
their dependents). We observe five of the employers for two years, creat<strong>in</strong>g a total of<br />
3,683 employee-years <strong>and</strong> 6,603 enrollee-years. Table 2.1 provides summary statistics<br />
on the covered employees, the enrollees, <strong>and</strong> the firms. Sixty-two percent of employees<br />
are female; the average age is just over forty. Fifty-eight percent of enrollees are female<br />
<strong>and</strong> enrollees are younger on average than employees, driven primarily by covered<br />
children. Twenty-eight percent of employees enroll <strong>in</strong> a plan that covers their spouse<br />
<strong>and</strong> 27 percent enroll <strong>in</strong> a plan that covers at least one child.<br />
Table 2.1 also presents risk scores at the employee, enrollee, <strong>and</strong> employer lev-<br />
els. A score of one represents an average <strong>in</strong>dividual <strong>in</strong> a nationally representative<br />
sample, <strong>and</strong> a score of two <strong>in</strong>dicates that an <strong>in</strong>dividual’s expected health costs are<br />
twice the average. The average risk scores of employees <strong>and</strong> enrollees are 1.25 <strong>and</strong><br />
1.01, respectively. The difference reflects the lower expected expenditures for covered<br />
children. Average risk ranges widely across employers, from 0.63 to 1.91. One reason<br />
for the degree of variation is the small number of enrollees at some of the firms <strong>in</strong> our<br />
data. This variation plays a key role <strong>in</strong> our analysis. We use <strong>in</strong>formation on <strong>in</strong>surer<br />
bids <strong>and</strong> realized costs to estimate models of the relationship between costs <strong>and</strong> risk.<br />
11 In a few cases, an employer had a prior contract with one of the <strong>in</strong>surers. We have exam<strong>in</strong>ed<br />
whether <strong>in</strong>corporat<strong>in</strong>g this <strong>in</strong>to our employee dem<strong>and</strong> model affects our estimates <strong>and</strong> found it did<br />
not. One concern is that this situation could result <strong>in</strong> asymmetric <strong>in</strong>formation between the plans <strong>in</strong><br />
the bidd<strong>in</strong>g, but we th<strong>in</strong>k this is unlikely to be an important problem.
CHAPTER 2. HEALTH PLAN CHOICE 76<br />
Because <strong>in</strong>surers report both bids <strong>and</strong> costs at the employer level, variation across<br />
employers <strong>in</strong> average risk is necessary to identify these relationships.<br />
Table 2.2 provides <strong>in</strong>formation on the plans offered by the employers <strong>in</strong> our sam-<br />
ple. Most employers offer all four plans, <strong>and</strong> all offer both HMOs <strong>and</strong> at least one<br />
other plan. On average, the <strong>in</strong>tegrated HMO is the least expensive plan <strong>and</strong> has<br />
the lowest enrollee contribution. This plan features high rates of co<strong>in</strong>surance, a low<br />
deductible, <strong>and</strong> a low out-of-pocket maximum. The network PPO is on average the<br />
most expensive plan <strong>and</strong> has the highest employee contribution. It features lower<br />
co<strong>in</strong>surance rates, higher deductibles <strong>and</strong> higher maximum expenditures. Roughly<br />
speak<strong>in</strong>g, the other two plans fall between these extremes. While bids for each plan<br />
vary substantially across tiers, reflect<strong>in</strong>g differences <strong>in</strong> expected expenditures, the<br />
bids for tiers other than employee only are simply scaled by a factor that is very<br />
similar across both plans <strong>and</strong> employers. Employee contributions also vary across<br />
tiers, with employees typically fac<strong>in</strong>g a greater fraction of the plan bid for dependent<br />
coverage. Variation <strong>in</strong> these contributions is important for the identification of our<br />
dem<strong>and</strong> model. We discuss contributions <strong>in</strong> detail <strong>in</strong> the identification section.<br />
We summarize enrollment patterns <strong>in</strong> Table 2.3. The <strong>in</strong>tegrated HMO attracts<br />
by far the most enrollees with a 59% market share among employees <strong>and</strong> 60% market<br />
share among enrollees. We also f<strong>in</strong>d little evidence of extensive risk selection across<br />
the plans. The <strong>in</strong>tegrated HMO attracts a slightly younger population, <strong>and</strong> women,<br />
particularly women employees, disproportionately choose the network <strong>and</strong> <strong>in</strong>tegrated<br />
HMOs. But the differences across the plans <strong>in</strong> both average age <strong>and</strong> average risk<br />
score are small. This lack of sort<strong>in</strong>g is not driven by heterogeneity across firms <strong>in</strong> the<br />
choice sets. If we condition on employers that offer both the PPO <strong>and</strong> the <strong>in</strong>tegrated<br />
HMO, for example, the average enrollee risk is 1.04 <strong>in</strong> both plans.
CHAPTER 2. HEALTH PLAN CHOICE 77<br />
2.4 Econometric Model<br />
2.4.1 Consumer Preferences, Plan Costs <strong>and</strong> Market Behav-<br />
ior<br />
In this section, we develop an econometric model that allows us to jo<strong>in</strong>tly estimate<br />
consumer preferences <strong>and</strong> health plan costs. It should be noted that by costs we mean<br />
overall costs to the <strong>in</strong>surer for a given enrollee <strong>in</strong> a given plan. Although we discuss<br />
factors that may expla<strong>in</strong> the variation <strong>in</strong> costs below, overall cost is sufficient for<br />
welfare analysis <strong>and</strong> it is not necessary to decompose whether these cost differences<br />
arise from, for example, moral hazard or physician reimbursement rates or some other<br />
factor (c.f., ?).<br />
In contrast to the simple theoretical model discussed above, the econometric model<br />
allows for multiple plans, vary<strong>in</strong>g plan characteristics, <strong>and</strong> both observable <strong>and</strong> pri-<br />
vately known dimensions of health risk <strong>and</strong> consumer tastes. Nevertheless, we aim<br />
for the most parsimonious model that permits a credible assessment of market ef-<br />
ficiency. In what follows, we describe the key components of the model: consumer<br />
choice, health plan costs, health plan bidd<strong>in</strong>g, <strong>and</strong> employer contribution sett<strong>in</strong>g, <strong>and</strong><br />
identify the stochastic assumptions on the unobservables that permit estimation.<br />
Consumer Choice<br />
We use a st<strong>and</strong>ard latent utility model to describe household choice behavior,<br />
where a household’s (money-metric) utility from choos<strong>in</strong>g a plan depends on a com-<br />
b<strong>in</strong>ation of household <strong>and</strong> plan characteristics. Specifically, household h’s utility from<br />
choos<strong>in</strong>g plan j is:<br />
uhj = φjαφ + xhαxj + ψ(rh + µh; αrj) − pj + σεεhj. (2.1)<br />
In this representation, household utility depends on observable plan characteris-<br />
tics φj, the monthly plan contribution pj, 12 observable household demographics xh,<br />
12 We convert employee contributions, which are made with pre-tax dollars, to post-tax dollars by<br />
adjust<strong>in</strong>g them by the marg<strong>in</strong>al tax rate (see Footnote 13 for discussion). For a given household
CHAPTER 2. HEALTH PLAN CHOICE 78<br />
an idiosyncratic preference εhj, <strong>and</strong> household health risk. Our measure of household<br />
health risk is aggregated from the <strong>in</strong>dividual level. For each <strong>in</strong>dividual i, we decom-<br />
pose health risk <strong>in</strong>to the observable risk score ri <strong>and</strong> additional privately known health<br />
factors µi. The µis capture <strong>in</strong>formation about health status that may affect choice<br />
behavior, but is not subject to risk adjustment. Equivalently, we can <strong>in</strong>terpret µi as<br />
measurement error <strong>in</strong> the risk score. We assume that each µi is an i.i.d. draw from a<br />
normal distribution with mean zero <strong>and</strong> variance σ 2 µ, <strong>and</strong> that the idiosyncratic tastes<br />
εhj are i.i.d. type I extreme value r<strong>and</strong>om variables (i.e. logit errors).<br />
We h<strong>and</strong>le heterogeneity <strong>in</strong> household size <strong>and</strong> composition by assum<strong>in</strong>g that,<br />
apart from the treatment of health risk, each household behaves as if it had a rep-<br />
resentative member with characteristics equal to the average of those of household<br />
members. 13 We parameterize household risk us<strong>in</strong>g two variables: the average risk<br />
of household members (i.e. the average of the ri + µi) <strong>and</strong> an <strong>in</strong>dicator of whether<br />
the household <strong>in</strong>cludes a high risk member. We def<strong>in</strong>e high risk as be<strong>in</strong>g above 2.25,<br />
which corresponds to the 90% percentile of the observed risk score distribution. The<br />
other household characteristics <strong>in</strong> the model are the averages of age <strong>and</strong> the male<br />
<strong>in</strong>dicator among covered household members as well as imputed household <strong>in</strong>come. 14<br />
In addition to the employee contribution, plan characteristics φj <strong>in</strong>clude a dummy<br />
variable for plan (the network HMO <strong>and</strong> PPO <strong>and</strong> the <strong>in</strong>tegrated HMO <strong>and</strong> POS),<br />
the relevant co<strong>in</strong>surance rate <strong>and</strong> deductible for the given employee, <strong>and</strong> an <strong>in</strong>dicator<br />
of non-st<strong>and</strong>ard drug coverage. 15 To be consistent with our approach to household<br />
h, letρh be the nom<strong>in</strong>al contribution <strong>and</strong> τh the household’s marg<strong>in</strong>al tax rate. The tax adjusted<br />
contribution is ph =(1− τh)ρh.<br />
13 We experimented with estimat<strong>in</strong>g different weights for household members, <strong>and</strong> also with restrict<strong>in</strong>g<br />
the sample to <strong>in</strong>dividual enrollees. Neither has much effect on our results. The Appendix<br />
<strong>in</strong>cludes <strong>in</strong>dividual enrollee estimates.<br />
14 We impute taxable <strong>in</strong>come for each household <strong>in</strong> our sample by estimat<strong>in</strong>g a model of household<br />
<strong>in</strong>come as a function of worker age, sex, family structure, firm size <strong>and</strong> <strong>in</strong>dustry us<strong>in</strong>g data from the<br />
Current Population Survey for 2004 <strong>and</strong> 2005 on workers with employer-sponsored health <strong>in</strong>surance<br />
<strong>in</strong> the correspond<strong>in</strong>g state. We then use the model to impute household <strong>in</strong>come for each employee <strong>in</strong><br />
our data <strong>in</strong>corporat<strong>in</strong>g r<strong>and</strong>om draws from the posterior distributions of the regression coefficients<br />
<strong>and</strong> the st<strong>and</strong>ard deviation of the residuals. Based on these predictions, we use Taxsim to calculate<br />
marg<strong>in</strong>al tax rates based on federal, state, <strong>and</strong> FICA taxes mak<strong>in</strong>g some assumptions on the correlation<br />
of coverage tier with fil<strong>in</strong>g status <strong>and</strong> number of dependents. The average taxable family<br />
<strong>in</strong>come <strong>and</strong> marg<strong>in</strong>al tax rate for workers <strong>in</strong> our sample are about $73,00 <strong>and</strong> 41%, respectively.<br />
15 While the prescription drug coverage for each plan is complicated, comprised of both formulate
CHAPTER 2. HEALTH PLAN CHOICE 81<br />
where rf denotes the average risk of employees <strong>in</strong> firm f, which the <strong>in</strong>surer forecasts<br />
us<strong>in</strong>g the available demographic <strong>in</strong>formation, xf. 16<br />
We model expected plan bids as a mark-up over expected cost. So plan j’s bid<br />
for firm f is:<br />
Bjf = δj · (aj + bj · (E[rf|xf] − 1)) + νjf, (2.7)<br />
where νjf is an <strong>in</strong>dependent mean zero r<strong>and</strong>om variable. The new parameter <strong>in</strong>tro-<br />
duced <strong>in</strong> the bid model is the mark-up, δj. We constra<strong>in</strong> the mark-up to be constant<br />
across the plans offered by a particular <strong>in</strong>surer. Although <strong>in</strong> theory an <strong>in</strong>surer could<br />
vary the mark-up across its different plans, because the cost data are at the <strong>in</strong>surer-<br />
firm level, we are unable to identify separately the mark-up <strong>and</strong> the fixed costs for<br />
each plan offered by an <strong>in</strong>surer. Naturally we expect the mark-up parameters to be<br />
larger than one.<br />
Employer Contribution Sett<strong>in</strong>g<br />
The last part of our model specifies how employers set required plan contributions.<br />
We adopt a simple model <strong>in</strong> which employers pass on a fraction of their cost for the<br />
lowest cost plan, <strong>and</strong> then a fraction of the <strong>in</strong>cremental cost for higher cost plans. We<br />
allow these fractions, denoted β <strong>and</strong> γ, to vary across firm-years <strong>and</strong> coverage tiers.<br />
Let B lf denote the m<strong>in</strong>imum bid received for coverage tier l <strong>in</strong> firm-year f, de-<br />
note plan j’s bid for coverage tier l <strong>in</strong> firm-year f as Bjlf. We model the required<br />
contribution as:<br />
pjlf = βlf · B lf + γlf · (Bjlf − B lf)+ξjlf. (2.8)<br />
This model describes employer behavior <strong>in</strong> our data remarkably well. The resid-<br />
uals from the l<strong>in</strong>ear regression (2.8) have a st<strong>and</strong>ard deviation of 7.64, <strong>and</strong> the R-<br />
squared is 0.99. As noted above, approximately half of the firms <strong>in</strong> our data choose<br />
a ”proportional pass-through” strategy where β = γ. The others choose an ”<strong>in</strong>cre-<br />
mental pass-through” strategy <strong>in</strong> which β
CHAPTER 2. HEALTH PLAN CHOICE 82<br />
2.4.2 Discussion of Model <strong>and</strong> Identification<br />
The key quantities <strong>in</strong> our model are plan costs <strong>and</strong> plan dem<strong>and</strong> as functions of fore-<br />
castable risk, <strong>and</strong> the price elasticity of dem<strong>and</strong>. The former determ<strong>in</strong>e the efficient<br />
allocation of households to plans, while the latter determ<strong>in</strong>es how price changes affect<br />
self-selection. We now discuss the variation <strong>in</strong> the data that identifies each of these<br />
quantities <strong>in</strong> estimation.<br />
Identify<strong>in</strong>g plan costs is straightforward. The effect of forecastable risk on plan<br />
costs is identified by variation across firms <strong>in</strong> the average risk scores of workers <strong>and</strong><br />
dependents, <strong>and</strong> how it affects <strong>in</strong>surer bids <strong>and</strong> realized costs. We identify the mark-<br />
up parameters, δj, by the difference between the plan bids <strong>and</strong> reported costs. A<br />
ma<strong>in</strong>ta<strong>in</strong>ed assumption <strong>in</strong> estimat<strong>in</strong>g mark-ups is that <strong>in</strong>surers base their bids on<br />
only the <strong>in</strong>formation about employees that is provided by the <strong>in</strong>termediary. We<br />
discuss this assumption more below, but we believe it is reasonable given the small<br />
size of the contracts <strong>and</strong> the fact that we consider only the first year of plan bids.<br />
The effect of household risk on choice behavior (i.e. the coefficients αrj <strong>in</strong> the<br />
dem<strong>and</strong> equation) is identified by variation <strong>in</strong> observable risk across households. Our<br />
model also allows private <strong>in</strong>formation about health status to affect choice. The key<br />
parameter here is the variance of the private <strong>in</strong>formation, σ 2 µ, which is identified by<br />
the correlation between consumers’ enrollment decisions <strong>and</strong> plans’ realized costs.<br />
This identification is aided by cross-firm variation <strong>in</strong> contribution policies <strong>and</strong> demo-<br />
graphics that, conditional on observable health risk, affect enrollment but not realized<br />
costs. 17 The identification obta<strong>in</strong>ed from price variation is similar to the identifica-<br />
tion <strong>in</strong> st<strong>and</strong>ard selection models, <strong>and</strong> relies on the exclusion restriction if a given<br />
<strong>in</strong>dividual i is enrolled <strong>in</strong> a given plan j, his or her utilization does not depend on<br />
the per-month premium (although of course it may depend on other elements of the<br />
plan such as the copayment rate).<br />
The most subtle identification issues arise <strong>in</strong> estimat<strong>in</strong>g the effect of plan con-<br />
tributions on dem<strong>and</strong>. Plan contributions are the result of plan bids <strong>and</strong> employer<br />
pass-through decisions. Our model allows four sources of variation <strong>in</strong> contributions:<br />
17 Our dem<strong>and</strong> model also <strong>in</strong>cludes plan characteristics such as co<strong>in</strong>surance <strong>and</strong> deductible. Their<br />
coefficients are identified off cross-firm <strong>and</strong> cross-tier variation <strong>in</strong> the characteristics.
CHAPTER 2. HEALTH PLAN CHOICE 83<br />
cross-firm variation <strong>in</strong> demographics (xf) that leads plans to submit different bids,<br />
idiosyncratic variation <strong>in</strong> plan bids (νjf), cross-firm <strong>and</strong> cross-tier variation <strong>in</strong> em-<br />
ployer pass-through rates (γjlf), <strong>and</strong> idiosyncratic variation <strong>in</strong> the plan contributions<br />
(ξjlf). 18<br />
Figure 2.4 demonstrates this variation by plott<strong>in</strong>g the <strong>in</strong>cremental contributions<br />
aga<strong>in</strong>st the <strong>in</strong>cremental bids for each plan relative to the <strong>in</strong>tegrated HMO, which is<br />
usually the plan requir<strong>in</strong>g the lowest employee contribution. We plot contribution<br />
rates for two tiers, employee only <strong>and</strong> employee plus spouse, to demonstrate how<br />
contributions vary across tier. For the employee plus spouse data, we divide both<br />
the contributions <strong>and</strong> the bids by two to obta<strong>in</strong> per-enrollee prices. Difference <strong>in</strong><br />
the bids for the <strong>in</strong>tegrated HMO <strong>and</strong> the network PPO ranges from $50 to $150<br />
per month, with a large fraction due to cross-firm variation <strong>in</strong> demographic risk.<br />
Comb<strong>in</strong>ations of <strong>in</strong>cremental contributions <strong>and</strong> bids that lie along the 45 degree l<strong>in</strong>e<br />
<strong>in</strong> Figure 2.3 represent employers who pass on the full marg<strong>in</strong>al cost of higher plan<br />
bids to employees. A subset of employers adopt this approach. Another subset of<br />
employers fully subsidize the higher cost plans, sett<strong>in</strong>g <strong>in</strong>cremental contributions of<br />
zero. Between these two extremes are employers who partially subsidize higher cost<br />
plans through contribution policies. In general, employers tend to pass on a greater<br />
portion of <strong>in</strong>cremental costs for plans with dependent coverage.<br />
The availability of multiple sources of variation permits some flexibility <strong>in</strong> estimat-<br />
<strong>in</strong>g price elasticities. Recall that accurate identification requires us<strong>in</strong>g price variation<br />
that is not correlated with idiosyncratic household tastes εhj or privately known health<br />
risk µh. Our basel<strong>in</strong>e estimates use all four sources of variation. We also employ <strong>in</strong>-<br />
strumental variables to isolate different sources of variation. The <strong>in</strong>struments are<br />
predicted plan contributions based on alternative covariates. The bottom l<strong>in</strong>e from<br />
these specifications is that our price elasticity estimates are quite robust to focus<strong>in</strong>g<br />
on different sources of variation <strong>in</strong> contributions. This robustness, despite our rela-<br />
tively small sample, suggests that endogeneity may not be an important concern, at<br />
least <strong>in</strong> this sett<strong>in</strong>g. Nevertheless, we now discuss the issues <strong>in</strong> detail.<br />
18 We also <strong>in</strong>troduce variation <strong>in</strong> employee contributions through the imputed marg<strong>in</strong>al tax rates,<br />
but we control for imputed <strong>in</strong>come <strong>and</strong> relevant household demographics <strong>in</strong> the dem<strong>and</strong> equation.
CHAPTER 2. HEALTH PLAN CHOICE 84<br />
Perhaps the most obvious identification concern is that employers believe their<br />
employees will prefer a particular plan <strong>and</strong> price accord<strong>in</strong>gly. This could mean cater-<br />
<strong>in</strong>g to employees with a low contribution, or sett<strong>in</strong>g a high contribution to pass on<br />
costs. Either would generate a correlation between the idiosyncratic part of the con-<br />
tribution ξjlf <strong>and</strong> household preferences εhj. To mitigate this concern, we <strong>in</strong>strument<br />
for the actual plan contribution us<strong>in</strong>g the predicted value (ˆpjlf) from the contribution<br />
model (2.8). We take this as our preferred specification <strong>in</strong> perform<strong>in</strong>g welfare analysis<br />
although the results are similar to the basel<strong>in</strong>e case with no <strong>in</strong>struments.<br />
A second concern is that plan bids are correlated with unobserved household<br />
tastes. This could happen if an <strong>in</strong>surer believed its plan was attractive due to, say,<br />
a nearby cl<strong>in</strong>ic location. It would generate a correlation between the idiosyncratic<br />
bid component, νjf, <strong>and</strong> household preferences εhj. We view this problem as most<br />
likely of marg<strong>in</strong>al importance given the limited <strong>in</strong>formation on the part of <strong>in</strong>surers.<br />
Nevertheless, we check our estimates by <strong>in</strong>strument<strong>in</strong>g for plan contribution with a<br />
predicted value that is constructed by plugg<strong>in</strong>g the predicted bid ˆ Bjf from (2.7) <strong>in</strong>to<br />
the contribution model (2.8). This specification purges the variation <strong>in</strong> both νjf <strong>and</strong><br />
ξjlf. The results are similar to our preferred specification.<br />
A third issue for identification is that employer pass-through rates might be sys-<br />
tematically <strong>in</strong>fluenced by employee preferences. This also seems unlikely, ma<strong>in</strong>ly<br />
because pass-through rates <strong>in</strong> our data are uncorrelated with observable differences<br />
across firms. Figure 2.4 plots employer pass-through rates aga<strong>in</strong>st employee health<br />
status, dependent health status, worker <strong>in</strong>come <strong>and</strong> firm size. There is no correla-<br />
tion, suggest<strong>in</strong>g that cross-firm differences <strong>in</strong> contribution policies may be due more<br />
to idiosyncratic factors, such as management philosophy, than employee tastes. Nev-<br />
ertheless, we aga<strong>in</strong> use an IV strategy to verify that our results are not driven by a<br />
correlation between the pass-through coefficients γjlf <strong>and</strong> unobserved preferences εhj.<br />
To this end, we <strong>in</strong>strument for plan contribution us<strong>in</strong>g predicted values from a variant<br />
of the contribution model (2.8) <strong>in</strong> which pass-through coefficients are restricted to be<br />
identical across firms. This purges cross-firm variation <strong>in</strong> γjlf as well as the variation<br />
<strong>in</strong> ξjlf. The results are aga<strong>in</strong> similar although with large st<strong>and</strong>ard errors. 19<br />
19 A f<strong>in</strong>al identification concern is that household choices may be <strong>in</strong>fluenced by the health status of
CHAPTER 2. HEALTH PLAN CHOICE 88<br />
2.5.1 Model Estimates<br />
Table 2.4 presents parameter estimates from three different specifications of the de-<br />
m<strong>and</strong> model. 21 The first column is a basel<strong>in</strong>e model where we do not <strong>in</strong>strument<br />
for plan contributions, <strong>and</strong> do not allow for private <strong>in</strong>formation about household<br />
risk. The second <strong>and</strong> third columns <strong>in</strong>strument for plan contributions us<strong>in</strong>g the pre-<br />
dicted values from the contribution model (2.8). The third column, which is our<br />
preferred specification, allows for private <strong>in</strong>formation about risk. To scale the utility<br />
to money-metric form, we divide each coefficient by the coefficient on the monthly<br />
contribution <strong>and</strong> adjust the st<strong>and</strong>ard errors accord<strong>in</strong>gly. We report the price effects<br />
as semi-elasticities at the bottom of the table.<br />
Effect of Demographics <strong>and</strong> Risk on Choice<br />
The dem<strong>and</strong> estimates <strong>in</strong>dicate that overall sort<strong>in</strong>g on the basis of risk is rather<br />
modest, but that different plans experience unfavorable selection across differ<strong>in</strong>g com-<br />
ponents risk. Older employees, who on average cost more to <strong>in</strong>sure, prefer the network<br />
HMO <strong>and</strong> the <strong>in</strong>tegrated POS plan to the <strong>in</strong>tegrated HMO. An additional year of age<br />
is associated with an <strong>in</strong>crease <strong>in</strong> the will<strong>in</strong>gness to pay for the network HMO relative<br />
to the <strong>in</strong>tegrated HMO of $1.75 per month (Column 1). Because older people are<br />
often <strong>in</strong> worse health, they are likely to place a higher value on the broader provider<br />
network of the network plan which would give them greater freedom <strong>in</strong> choos<strong>in</strong>g<br />
among providers. Women, who at the age of workers <strong>in</strong> our data typically cost more<br />
to <strong>in</strong>sure than men, prefer the <strong>in</strong>tegrated HMO to either the <strong>in</strong>tegrated POS plan<br />
or the network PPO. Women are will<strong>in</strong>g to pay $35 per month less than men for<br />
the network PPO relative to the <strong>in</strong>tegrated HMO (Column 1). Women may have<br />
stronger preferences for the <strong>in</strong>tegrated HMO if they perceive that it is more effective<br />
<strong>in</strong> provid<strong>in</strong>g preventive cares s<strong>in</strong>ce, <strong>in</strong> this age group, more preventive services are<br />
recommended for women than for men. The effects of age <strong>and</strong> sex are not particularly<br />
sensitive to the use of <strong>in</strong>struments for the employee contribution (Column 2) or the<br />
21 The Table does not report every parameter. The parameters not reported are the plan fixed<br />
effects, <strong>and</strong> the coefficients on imputed household <strong>in</strong>come <strong>and</strong> an <strong>in</strong>dicator for non-st<strong>and</strong>ard drug<br />
coverage.
CHAPTER 2. HEALTH PLAN CHOICE 89<br />
<strong>in</strong>corporation of unobserved risk (Column 3).<br />
We f<strong>in</strong>d some sort<strong>in</strong>g on the basis of health status conditional on age <strong>and</strong> sex,<br />
driven primarily by hav<strong>in</strong>g a very high risk household member. The effects of the<br />
l<strong>in</strong>ear risk score on plan choice are generally small <strong>and</strong> imprecise. Households with<br />
a high risk member, however, are less likely to enroll <strong>in</strong> the network HMO <strong>and</strong> more<br />
likely to enroll <strong>in</strong> the network PPO than the <strong>in</strong>tegrated HMO. In our preferred spec-<br />
ification (Column 3), an employee with a high risk family member is will<strong>in</strong>g to pay<br />
$28 per month more than an employee without a high risk family member to enroll<br />
<strong>in</strong> the network PPO relative to the <strong>in</strong>tegrated HMO. This is consistent, once aga<strong>in</strong>,<br />
with those who are more likely to use care plac<strong>in</strong>g a greater value on less restrictive<br />
provider networks.<br />
Our results also suggest that private <strong>in</strong>formation about health risk plays a role<br />
<strong>in</strong> plan choice, although the estimate is not precise. We estimate that the st<strong>and</strong>ard<br />
deviation of private risk <strong>in</strong>formation σµ is 0.68, which is substantial relative to the<br />
st<strong>and</strong>ard deviation of the observed risk scores (1.56 <strong>in</strong> Table 2.1). Roughly speak<strong>in</strong>g,<br />
observed risk scores appear to pick up just over 2/3 of the health status <strong>in</strong>formation<br />
that factors <strong>in</strong>to plan choice.<br />
While our f<strong>in</strong>d<strong>in</strong>gs with respect to risk selection are not <strong>in</strong>consistent with exist<strong>in</strong>g<br />
research, they suggest a relatively complex pattern of sort<strong>in</strong>g. Much of the exist<strong>in</strong>g<br />
literature f<strong>in</strong>ds evidence of unfavorable selection <strong>in</strong>to more generous plans (??) .<br />
We also f<strong>in</strong>d that the highest risk enrollees favor the most flexible plan, the network<br />
PPO. Overall, however, the average risk across plans is quite similar due to off-<br />
sett<strong>in</strong>g selection along different demographic dimensions, <strong>in</strong>clud<strong>in</strong>g age <strong>and</strong> gender,<br />
that are correlated with risk. This f<strong>in</strong>d<strong>in</strong>g is consistent with the idea that the plans<br />
cater to <strong>in</strong>dividuals with different tastes for health care, rather than offer<strong>in</strong>g different<br />
quantities of care, or target<strong>in</strong>g <strong>in</strong>dividuals of different health status.<br />
Effect of Plan Prices on Choice<br />
In the bottom panel, we present price semi-elasticities of dem<strong>and</strong>, def<strong>in</strong>ed as the<br />
percentage decrease <strong>in</strong> market share result<strong>in</strong>g from a $100 <strong>in</strong>crease <strong>in</strong> the annual<br />
enrollee contribution, evaluated at the mean choice probability for each plan. On
CHAPTER 2. HEALTH PLAN CHOICE 90<br />
average, a $100 dollar <strong>in</strong>crease <strong>in</strong> the annual enrollee contribution decreases market<br />
share by 7 to 9 percent. While <strong>in</strong>strument<strong>in</strong>g for the contribution reduces the precision<br />
of the estimate, it has relatively little effect on its magnitude. These estimates suggest<br />
that dem<strong>and</strong> is relatively <strong>in</strong>elastic, <strong>in</strong> l<strong>in</strong>e with other estimates <strong>in</strong> the literature. In<br />
the Onl<strong>in</strong>e Appendix, we discuss studies <strong>in</strong> sett<strong>in</strong>gs similar to ours. Across these<br />
studies, a $100 <strong>in</strong>crease <strong>in</strong> the annual contribution reduces market share by 1.6 to 9.6<br />
percent, plac<strong>in</strong>g our estimate <strong>in</strong> the middle to high end of the range.<br />
The results <strong>in</strong> Table 2.4 also <strong>in</strong>clude the estimated value of plan characteristics<br />
other than price, such as co<strong>in</strong>surance rate <strong>and</strong> deductible. Enrollees appear to be<br />
moderately sensitive to both. We estimate that a 10 percentage po<strong>in</strong>t <strong>in</strong>crease <strong>in</strong><br />
the co<strong>in</strong>surance rate is valued at approximately $276 dollars annually, which is about<br />
10 percent of the annual cost per enrollee reported by the <strong>in</strong>surers. Our estimates<br />
<strong>in</strong>dicate that enrollees are not particularly sensitive to the deductible when choos<strong>in</strong>g<br />
among plans. 22<br />
Because the estimates of risk <strong>and</strong> price elasticity are the key parameters for our<br />
welfare calculations, we have exam<strong>in</strong>ed the sensitivity of these estimates to a variety<br />
of issues. In the Onl<strong>in</strong>e Appendix we present estimates where we vary the sample of<br />
households <strong>and</strong> use different <strong>in</strong>struments (discussed <strong>in</strong> Section 2.5.4) for the employee<br />
contributions. We also discuss specifications with alternative sets of controls. The<br />
bottom l<strong>in</strong>e is that the estimates are robust to variation across these dimensions.<br />
Structure of Plan Costs<br />
The difference <strong>in</strong> cost structures for the <strong>in</strong>tegrated <strong>and</strong> network <strong>in</strong>surer can be<br />
seen <strong>in</strong> the raw data depicted <strong>in</strong> Figures 2.5 <strong>and</strong> 2.6. Figure 2.5 is a scatterplot of<br />
enrollee risk scores aga<strong>in</strong>st realized costs. Each po<strong>in</strong>t corresponds to an <strong>in</strong>surer-firm-<br />
year. The x-axis is the average risk of the <strong>in</strong>surer’s enrollees; the y-axis is the reported<br />
costs per enrollee. The l<strong>in</strong>es represent the model’s prediction (shown <strong>in</strong> Table 2.5)<br />
of expected costs for the network PPO <strong>and</strong> the <strong>in</strong>tegrated HMO. Figure 2.6 displays<br />
correspond<strong>in</strong>g <strong>in</strong>formation for bids. It shows the average risk of a firm’s employees<br />
plotted aga<strong>in</strong>st plan bids, with each observation at the plan-firm-year level.<br />
22 The results are unchanged when out-of-pocket maximum are <strong>in</strong>cluded as plan characteristics.
CHAPTER 2. HEALTH PLAN CHOICE 91<br />
As the figures illustrate, the plans seem to have similar costs for enrollees with<br />
average health risk <strong>and</strong> divergent costs for enrollees <strong>in</strong> good <strong>and</strong> poor health. The<br />
expected monthly cost for an enrollee with a risk score of 1 is $235 for the <strong>in</strong>tegrated<br />
HMO, $236 for the <strong>in</strong>tegrated POS, $218 for the network HMO, <strong>and</strong> $238 for the<br />
network PPO. For less healthy enrollees, the <strong>in</strong>tegrated <strong>in</strong>surer has a substantial cost<br />
advantage. The expected monthly cost of an enrollee with a risk score of two is $309<br />
for the <strong>in</strong>tegrated HMO, compared to $507 for the network HMO <strong>and</strong> $413 for the<br />
network PPO. Network plans do relatively well for low risks. The expected monthly<br />
cost for an enrollee with a risk score of 0.5 is $198 for the <strong>in</strong>tegrated HMO, as opposed<br />
to $151 for the network PPO <strong>and</strong> $74 for the network HMO.<br />
The structure of plan costs we estimate is consistent with the basic idea that<br />
patient cost shar<strong>in</strong>g may be effective at limit<strong>in</strong>g provider visits while supply-side<br />
mechanisms may be more effective at limit<strong>in</strong>g costs conditional on receiv<strong>in</strong>g services<br />
(see, e.g., ?). While we do not have visit-level data to support the claim, the steep cost<br />
curves for the network plans are consistent with cost shar<strong>in</strong>g limit<strong>in</strong>g visits, particu-<br />
larly for low risks, but hav<strong>in</strong>g little effect on the high risks who consume healthcare<br />
on the <strong>in</strong>tensive marg<strong>in</strong>. In contrast, the <strong>in</strong>tegrated plans with their relatively low<br />
cost shar<strong>in</strong>g but stronger supply side utilization controls may be less effective at lim-<br />
it<strong>in</strong>g provider visits for low risks but more effective at manag<strong>in</strong>g costs conditional on<br />
provider visits for the high risks. Another factor expla<strong>in</strong><strong>in</strong>g the relatively high costs<br />
for low risks <strong>in</strong> the <strong>in</strong>tegrated plan may be the greater use of preventive services.<br />
These results also provide some <strong>in</strong>sight on the potential issue of bias <strong>in</strong> the risk<br />
scores. As discussed earlier, if the risk scores are biased, the likely direction of the bias<br />
is an underestimate of the risk of the enrollees <strong>in</strong> the <strong>in</strong>tegrated plan. This is because<br />
the risk scores are based on the types of prescription drugs used by plan enrollees<br />
<strong>and</strong> the <strong>in</strong>tegrated plan more aggressively manages drug utilization. If this were the<br />
case, then we would underestimate the cost sav<strong>in</strong>gs for high risks generated by the<br />
<strong>in</strong>tegrated plan. In other words, the large cost differentials we observe would be even<br />
larger. While we believe that bias <strong>in</strong> the risk scores is unlikely to be an important<br />
issue, if it exists, it would likely re<strong>in</strong>force our key f<strong>in</strong>d<strong>in</strong>gs.<br />
We note that our estimates for the <strong>in</strong>tegrated POS plan <strong>in</strong>dicate that its costs
CHAPTER 2. HEALTH PLAN CHOICE 92<br />
are more similar to those of the network plans than the <strong>in</strong>tegrated HMO. While<br />
these results are not consistent with our f<strong>in</strong>d<strong>in</strong>gs regard<strong>in</strong>g the differences <strong>in</strong> the cost<br />
structures between the two types of <strong>in</strong>surers, s<strong>in</strong>ce the bulk of the <strong>in</strong>tegrated plans<br />
enrollment is concentrated <strong>in</strong> the HMO product, it is likely that the HMO is more<br />
representative of its cost structure, particularly given our small sample size. The<br />
other <strong>in</strong>consistency <strong>in</strong> our cost results is the difference <strong>in</strong> the cost structure between<br />
the plans offered by the network <strong>in</strong>surer. In particular, <strong>in</strong> contrast to the <strong>in</strong>tegrated<br />
HMO, the network HMO appears not to generate cost sav<strong>in</strong>gs for high risks relative<br />
to the PPO product. While it is difficult to underst<strong>and</strong> why this plan would be<br />
more expensive for unhealthy enrollees relative to its less managed PPO counterpart,<br />
we note that our estimates are based on relatively small samples, mak<strong>in</strong>g it not<br />
particularly surpris<strong>in</strong>g that some of our results may not have an <strong>in</strong>tuitive explanation.<br />
In addition, many factors are likely to affect health plans costs so our estimates may<br />
also reflect true differences <strong>in</strong> cost structures for which we did not have strong prior<br />
beliefs. The estimated mark-up varies across <strong>in</strong>surers. We estimate that the network<br />
<strong>in</strong>surer <strong>and</strong> the <strong>in</strong>tegrated <strong>in</strong>surer bid 24 percent <strong>and</strong> 8 percent over expected costs,<br />
respectively.<br />
The sensitivity of cost differentials as a function of enrollee risk, compared to the<br />
relatively modest effect of risk on plan preferences, has an important implication.<br />
It <strong>in</strong>dicates that as consumer risk varies, changes <strong>in</strong> relative plan costs rather than<br />
changes <strong>in</strong> preferences will drive the efficient allocation. As our simple theory model<br />
illustrated, this will not happen under self-selection without a mechanism that allows<br />
different risk groups to face different premium differentials. In our sett<strong>in</strong>g, prices do<br />
not have this feature, suggest<strong>in</strong>g the potential for <strong>in</strong>efficiency. We return to this po<strong>in</strong>t<br />
<strong>in</strong> the next section, when we quantify social welfare.<br />
A factor to keep <strong>in</strong> m<strong>in</strong>d when evaluat<strong>in</strong>g our estimates of plan costs is that we<br />
observe the <strong>in</strong>surers’ costs of coverage, not the overall dollars spent on care. The<br />
dist<strong>in</strong>ction is important because, <strong>in</strong> plans with copayments <strong>and</strong> deductibles, enrollees<br />
bear a share of the cost of care that we do not capture <strong>in</strong> our data. These payments<br />
are largest at the network PPO <strong>and</strong> smallest at the <strong>in</strong>tegrated HMO. While our<br />
model assumes that these payments will be <strong>in</strong>ternalized <strong>in</strong> mak<strong>in</strong>g plan choices, they
CHAPTER 2. HEALTH PLAN CHOICE 93<br />
do affect the <strong>in</strong>terpretation of the effects of the different plan types on utilization of<br />
care. In particular,the reduction <strong>in</strong> <strong>in</strong>sured costs for low risks <strong>in</strong> the network plan may<br />
represent, at least <strong>in</strong> part, a shift from <strong>in</strong>sured to un<strong>in</strong>sured payments, rather than a<br />
reduction <strong>in</strong> utilization or prices. For high risks, <strong>in</strong> contrast, the difference <strong>in</strong> <strong>in</strong>sured<br />
costs between the plans likely underestimates the extent to which the <strong>in</strong>tegrated plan<br />
reduces total costs.<br />
2.5.2 Quantify<strong>in</strong>g Social Welfare Inefficiencies<br />
In this section, we use the estimated dem<strong>and</strong> <strong>and</strong> cost model to compute the <strong>in</strong>effi-<br />
ciency associated with observed contribution policies relative to alternative efficient<br />
benchmarks. We also compare welfare between the observed policies <strong>and</strong> alterna-<br />
tive uniform contribution policies to demonstrate the extent to which the <strong>in</strong>efficiency<br />
associated with a uniform contribution could be reduced with<strong>in</strong> the current <strong>in</strong>stitu-<br />
tional constra<strong>in</strong>ts. Table 2.6 presents the results of these simulations. The left-h<strong>and</strong><br />
panels present the market share, average enrollee risk, <strong>and</strong> the average <strong>in</strong>cremental<br />
contribution for each plan under five different pric<strong>in</strong>g scenarios. The <strong>in</strong>cremental con-<br />
tribution represents the monthly contribution per enrollee relative to the <strong>in</strong>tegrated<br />
HMO averaged across all households. The right-h<strong>and</strong> panels present <strong>in</strong>formation on<br />
the change <strong>in</strong> surplus relative to the observed allocation for each scenario.<br />
The Welfare Cost of Observed Prices<br />
In the top panels, we calculate the <strong>in</strong>efficiency of observed pric<strong>in</strong>g policies rela-<br />
tive to two risk-rated benchmarks. The first is <strong>in</strong>dividual risk rat<strong>in</strong>g based on the<br />
observed risk scores. This pric<strong>in</strong>g policy, which we refer to as “feasible risk-rated<br />
contributions”, maximizes social welfare conditional on knowledge of the risk scores,<br />
but not each household’s private <strong>in</strong>formation. The third panel of the Table reports<br />
outcomes when prices are first-best, i.e. risk-rated based on both <strong>public</strong> <strong>and</strong> private<br />
<strong>in</strong>formation.<br />
Overall, under risk-rated contributions, high-risk households face higher premiums<br />
<strong>and</strong> low-risk households face somewhat lower premiums for the network plans relative<br />
to observed contribution policies. In both the feasible <strong>and</strong> first-best scenarios, this
CHAPTER 2. HEALTH PLAN CHOICE 94<br />
leads to a substantial re-allocation of enrollees across plans, although overall market<br />
shares change modestly. With feasible risk-rat<strong>in</strong>g, the average enrollee risk at the<br />
<strong>in</strong>tegrated HMO <strong>in</strong>creases from its observed level of 0.99 to 1.49, <strong>and</strong> the network<br />
HMO experiences a decl<strong>in</strong>e <strong>in</strong> average enrollee risk from 1.03 to 0.58. This reallocation<br />
of households across plans substantially reduces overall <strong>in</strong>surer costs, by $44 per<br />
enrollee-month, <strong>and</strong> <strong>in</strong>creases total social surplus by just over $27 per enrollee-month.<br />
The <strong>in</strong>crease <strong>in</strong> social welfare represents approximately 11% of average <strong>in</strong>surer costs<br />
<strong>in</strong> our sample.<br />
A substantial fraction of the welfare ga<strong>in</strong> is due to the highest <strong>and</strong> lowest risk<br />
households mak<strong>in</strong>g more efficient plan choices. Table 2.7 decomposes the welfare<br />
calculation by household risk qu<strong>in</strong>tiles. The lowest <strong>and</strong> highest risk qu<strong>in</strong>tiles (aver-<br />
age household risk below 0.36 <strong>and</strong> above 1.33) generate about three-quarters of the<br />
welfare effect. This raises a concern that our calculation might be driven <strong>in</strong> part by<br />
extrapolat<strong>in</strong>g plan costs out of sample. As Figures 2.5 <strong>and</strong> 2.6 illustrate, we observe<br />
plan bids <strong>and</strong> costs only for average risk scores between 0.75 <strong>and</strong> 2.0. In contrast,<br />
household risk ranges from 0.16 to 30.1. To address this, we truncate the cost differ-<br />
entials between plans at their 0.75 <strong>and</strong> 2.0 levels <strong>and</strong> re-calculate the welfare numbers.<br />
These calculations appear <strong>in</strong> the f<strong>in</strong>al columns of Tables 2.6 <strong>and</strong> 2.7. We view the<br />
numbers based on truncated cost differences as a lower bound on welfare differences,<br />
<strong>and</strong> the basel<strong>in</strong>e numbers based on straight-l<strong>in</strong>e extrapolation as closer to an upper<br />
bound. Truncat<strong>in</strong>g the cost differentials has little effect on the result<strong>in</strong>g assignment<br />
of households to plans, but as one might expect, it reduces the welfare cost of ob-<br />
served pric<strong>in</strong>g to $5 per enrollee-month, or 2% of <strong>in</strong>surer costs, relative to the feasible<br />
optimum.<br />
It is also <strong>in</strong>terest<strong>in</strong>g to compare what is possible us<strong>in</strong>g prices based on observed<br />
risk scores to what <strong>in</strong> pr<strong>in</strong>ciple could be achieved us<strong>in</strong>g both observed risk scores <strong>and</strong><br />
households’ private <strong>in</strong>formation. This calculation captures the extent to which private<br />
<strong>in</strong>formation on risk constra<strong>in</strong>s the efficiency of feasible relative to optimal risk-rated<br />
pric<strong>in</strong>g. Chang<strong>in</strong>g from feasible risk rated contributions to the first-best scenario<br />
<strong>in</strong>creases social surplus by between $2 <strong>and</strong> $8 per enrollee-month, depend<strong>in</strong>g on the<br />
treatment of costs for extreme risk, or roughly 1-3 percent of <strong>in</strong>surer costs. One way
CHAPTER 2. HEALTH PLAN CHOICE 95<br />
to <strong>in</strong>terpret this is that, <strong>in</strong> our sample, a social planner could achieve approximately<br />
70% of the potential welfare ga<strong>in</strong>s associated with <strong>in</strong>dividualized pric<strong>in</strong>g us<strong>in</strong>g only<br />
observable <strong>in</strong>formation on risk.<br />
Social Welfare without Risk-Rat<strong>in</strong>g<br />
The calculations above <strong>in</strong>dicate that the observed prices fall well short of the effi-<br />
cient benchmark. A natural question is whether efficiency ga<strong>in</strong>s could be realized even<br />
without risk-rated contributions. That is, to the extent re-allocat<strong>in</strong>g high <strong>and</strong> low<br />
risk households would <strong>in</strong>crease social welfare, is it possible to <strong>in</strong>duce this reallocation<br />
given current <strong>in</strong>stitutional pric<strong>in</strong>g constra<strong>in</strong>ts? At first glance, the answer is unclear.<br />
After all, current <strong>in</strong>stitutions require uniform pric<strong>in</strong>g with<strong>in</strong> firm-tiers, but this still<br />
allows a fair amount of pric<strong>in</strong>g flexibility with<strong>in</strong> our sample. For example, average<br />
risk varies substantially across the firms <strong>in</strong> our data, suggest<strong>in</strong>g that cross-firm vari-<br />
ation <strong>in</strong> contribution policies could alleviate some of the <strong>in</strong>efficiency associated with<br />
uniform contributions.<br />
The next scenario <strong>in</strong> Table 2.6 addresses the question of what is possible with-<br />
out <strong>in</strong>dividualized pric<strong>in</strong>g by consider<strong>in</strong>g contributions that maximize social welfare<br />
subject to be<strong>in</strong>g uniform with<strong>in</strong> each firm-tier. As <strong>in</strong> the case of fully risk-rated<br />
prices, optimiz<strong>in</strong>g uniform with<strong>in</strong> firm contributions leads to a reallocation of high-<br />
risk households <strong>in</strong>to the <strong>in</strong>tegrated plans <strong>and</strong> away from the network plans, particu-<br />
larly the PPO. The shift is much less dramatic, however, than under full risk rat<strong>in</strong>g.<br />
Overall social surplus is $1.40-6.70 higher per enrollee-month than under the observed<br />
policies, but still $3.60-20.40 below the efficient level. This <strong>in</strong>dicates that about 3/4<br />
of the observed <strong>in</strong>efficiency is due to the requirement of nondiscrim<strong>in</strong>atory pric<strong>in</strong>g<br />
with<strong>in</strong> firms. Nevertheless, it appears that employers could <strong>in</strong>crease social surplus<br />
by around 1-3% of average <strong>in</strong>surer costs simply by adjust<strong>in</strong>g their contributions to<br />
better reflect differences <strong>in</strong> underly<strong>in</strong>g plan costs.<br />
One difficulty for employers, of course, is that match<strong>in</strong>g contributions to plan<br />
costs may be a fairly complex exercise. Many benefits consultants, <strong>in</strong>clud<strong>in</strong>g the<br />
<strong>in</strong>termediary <strong>in</strong> our data, suggest a simpler approach, which is to pass on the full<br />
<strong>in</strong>cremental premium for all but the lowest priced plan. We refer to this as the
CHAPTER 2. HEALTH PLAN CHOICE 96<br />
“Enthoven Rule”(?). About 1/2 of the firms <strong>in</strong> our sample use this approach for at<br />
least some workers. The last entry <strong>in</strong> Table 2.6 considers the effect of mov<strong>in</strong>g all<br />
the firms to an Enthoven-style approach. Perhaps surpris<strong>in</strong>gly, this has relatively<br />
little effect on overall welfare, or on household choices. The reason is that dem<strong>and</strong><br />
is not very price elastic <strong>and</strong> from a practical st<strong>and</strong>po<strong>in</strong>t most firms already pass<br />
through a substantial fraction of the premium differentials. So relative to the price<br />
changes needed to move substantial numbers of households across plans, a change to<br />
an Enthoven policy has only a modest effect on choices.<br />
This last observation raises an important po<strong>in</strong>t for our pric<strong>in</strong>g experiments. The<br />
relatively low elasticity of dem<strong>and</strong> means that the contribution differentials needed<br />
to re-allocate households <strong>in</strong> the direction of efficiency are sizeable. For <strong>in</strong>stance,<br />
maximiz<strong>in</strong>g welfare while keep<strong>in</strong>g contributions uniform with<strong>in</strong> firm-tiers would lead<br />
to some households see<strong>in</strong>g a $87 per-enrollee monthly premium for the network PPO<br />
relative to the <strong>in</strong>tegrated HMO. A move to efficient risk-rated prices would <strong>in</strong>crease<br />
this differential even more for some high-cost households. For <strong>in</strong>stance, an <strong>in</strong>dividual<br />
employee with a risk score of 3 would face a monthly premium differential of between<br />
$101 <strong>and</strong> $202 depend<strong>in</strong>g on our cost extrapolation. These large price differentials<br />
<strong>in</strong>dicate that achiev<strong>in</strong>g efficient allocations may raise issues of fairness or affordability<br />
of coverage for particular subgroups.<br />
2.5.3 The Value of Plan Choice<br />
By choos<strong>in</strong>g to offer benefits through the <strong>in</strong>termediary, each of the firms <strong>in</strong> our sample<br />
moved from offer<strong>in</strong>g a s<strong>in</strong>gle health plan to offer<strong>in</strong>g multiple plans from two carri-<br />
ers. A clear benefit of plan choice is that households with different preferences can<br />
select their preferred plan. Our estimates <strong>in</strong>dicate a substantial amount of preference<br />
heterogeneity, <strong>and</strong> hence suggest substantial welfare ga<strong>in</strong>s from giv<strong>in</strong>g households<br />
multiple plan options.<br />
To illustrate this, Table 2.8 compares aggregate surplus under the observed of-<br />
fer<strong>in</strong>gs to the surplus that would be obta<strong>in</strong>ed if all the households <strong>in</strong> our sample<br />
were enrolled <strong>in</strong> one of the four plans. The most natural benchmark is the <strong>in</strong>tegrated
CHAPTER 2. HEALTH PLAN CHOICE 97<br />
HMO, as it would be the most efficient s<strong>in</strong>gle-plan offer<strong>in</strong>g for every firm <strong>in</strong> our data.<br />
Relative to the <strong>in</strong>tegrated HMO benchmark, the observed plan offer<strong>in</strong>gs <strong>in</strong>crease so-<br />
cial welfare by almost $70 per enrollee-month for the firms <strong>in</strong> our data. Virtually<br />
all of this is due to an <strong>in</strong>crease <strong>in</strong> consumer surplus (gross of plan contributions)<br />
rather than to a reduction <strong>in</strong> <strong>in</strong>surer costs. Indeed, <strong>in</strong>surer costs would be lowest if<br />
all households were enrolled <strong>in</strong> the network HMO but the reduction <strong>in</strong> social surplus<br />
would be large due the reduction <strong>in</strong> consumer surplus.<br />
One caveat to this calculation is the logit dem<strong>and</strong> specification is notorious for<br />
generat<strong>in</strong>g large “new product” welfare ga<strong>in</strong>s. Roughly speak<strong>in</strong>g, the problem is that<br />
each new product adds a new preference dimension, <strong>and</strong> some households <strong>in</strong>variably<br />
enjoy a large welfare ga<strong>in</strong> from this addition due to the logit distributional assump-<br />
tion. So while we th<strong>in</strong>k the benefits of plan choice are real, we urge some caution <strong>in</strong><br />
<strong>in</strong>terpret<strong>in</strong>g the magnitude of the measured effect.<br />
2.5.4 Discussion <strong>and</strong> Sensitivity Analysis<br />
Our estimates of market <strong>in</strong>efficiencies are based on a particular set of employers <strong>in</strong> a<br />
particular geographic area. One way to address external validity is to compare our<br />
estimates with some other studies of specific environments, such as ?, ?, <strong>and</strong> ?. These<br />
studies all rely on data from <strong>in</strong>dividual large employers, <strong>and</strong> <strong>in</strong> each case, the plans<br />
are plausibly dist<strong>in</strong>guished by their level of generosity, mak<strong>in</strong>g the environments a<br />
bit different from ours. All three studies f<strong>in</strong>d evidence that more generous plans are<br />
adversely selected. Cutler <strong>and</strong> Reber document this by us<strong>in</strong>g enrollee age as a proxy<br />
for risk. The latter two studies, like ours, use data on realized costs.<br />
Despite the difference <strong>in</strong> <strong>in</strong>stitutional sett<strong>in</strong>gs, the bottom l<strong>in</strong>e welfare estimates<br />
from these studies are fairly similar, <strong>and</strong> also similar to our estimates. Cutler <strong>and</strong><br />
Reber estimate that observed prices at Harvard University reduce welfare by around<br />
2-4% of coverage costs relative to optimal uniform prices. E<strong>in</strong>av, F<strong>in</strong>kelste<strong>in</strong> <strong>and</strong><br />
Cullen estimate that <strong>in</strong> their sett<strong>in</strong>g average cost pric<strong>in</strong>g has a welfare cost of roughly<br />
2% relative to optimal uniform pric<strong>in</strong>g. Carl<strong>in</strong> <strong>and</strong> Town f<strong>in</strong>d much smaller welfare
CHAPTER 2. HEALTH PLAN CHOICE 98<br />
effects, due to very low dem<strong>and</strong> elasticity. 23 Note that these papers all focus on<br />
uniform pric<strong>in</strong>g, which we have noted is generally <strong>in</strong>efficient except <strong>in</strong> special cases.<br />
When we use our estimates to compare observed pric<strong>in</strong>g to optimal uniform prices,<br />
we f<strong>in</strong>d welfare costs of approximately 1-3% of coverage costs. In this sense, there<br />
appears to be a fair amount of agreement between studies.<br />
As a group, these studies also re<strong>in</strong>force our earlier observation that <strong>in</strong>efficiencies<br />
from pric<strong>in</strong>g can be driven both by the nature of sort<strong>in</strong>g <strong>and</strong> risk-selection, <strong>and</strong> by the<br />
price elasticity of dem<strong>and</strong>, which determ<strong>in</strong>es the extent to which implicit subsidies or<br />
taxes affect choices. To guage the sensitivity of our own estimates to these factors,<br />
we recalculated the surplus difference between the observed <strong>and</strong> the feasible efficient<br />
allocation assum<strong>in</strong>g that dem<strong>and</strong> was twice as sensitive to price as we have estimated,<br />
<strong>and</strong> half as sensitive. We performed a similar analysis vary<strong>in</strong>g the risk sensitivity of<br />
dem<strong>and</strong>. These analyses <strong>in</strong>crease the range of the welfare ga<strong>in</strong>s to 1-13% of total<br />
coverage costs. Given the range of dem<strong>and</strong> estimates <strong>in</strong> the literature, one may want<br />
to assign a correspond<strong>in</strong>g range of uncerta<strong>in</strong>ty to the potential welfare costs of price<br />
distortions.<br />
2.6 Conclusion<br />
Economists have long understood that competition <strong>in</strong> health <strong>in</strong>surance markets is no<br />
guarantee of efficiency. This paper contributes to a nascent literature that attempts<br />
to quantify market <strong>in</strong>efficiencies <strong>and</strong> identify their sources. A ma<strong>in</strong> f<strong>in</strong>d<strong>in</strong>g is that<br />
observed contribution policies distort enrollment decisions from their efficient level.<br />
We calculate that the welfare loss is on the order of 2-11% of the total cost of coverage.<br />
Captur<strong>in</strong>g these ga<strong>in</strong>s <strong>in</strong> full would require the use of risk-rated contribution policies.<br />
Absent such policies, optimally set employee contributions might <strong>in</strong>crease welfare by<br />
1-3% of coverage costs. A key po<strong>in</strong>t to emphasize is that despite these distortions,<br />
there appear to be substantial ga<strong>in</strong>s <strong>in</strong> our context from <strong>in</strong>troduc<strong>in</strong>g plan choice at<br />
23 One explanation for their <strong>in</strong>elastic dem<strong>and</strong> estimate is that they rely on time-series variation<br />
<strong>in</strong> contributions. As discussed above, employees appear to be more price sensitive <strong>in</strong> mak<strong>in</strong>g <strong>in</strong>itial<br />
choices than <strong>in</strong> mak<strong>in</strong>g changes once they are enrolled.
CHAPTER 2. HEALTH PLAN CHOICE 99<br />
the employer level.<br />
One important po<strong>in</strong>t about risk-rated premiums <strong>in</strong> health <strong>in</strong>surance is that cov-<br />
erage is typically purchased on an annual basis. While risk-rat<strong>in</strong>g might <strong>in</strong>crease<br />
static efficiency, it exposes households to reclassification risk as their health status<br />
changes over time. This is one argument for community rat<strong>in</strong>g or nondiscrimation <strong>in</strong><br />
employer contributions. It is <strong>in</strong>terest<strong>in</strong>g to ask whether there are ways to promote<br />
static efficiency that nevertheless mitigate reclassification risk. One natural approach<br />
is to ensure that consumers have a basel<strong>in</strong>e option whose price is <strong>in</strong>dependent of risk,<br />
<strong>and</strong> allow <strong>in</strong>cremental purchases to be risk-rated. An alternative would be longer-<br />
term contracts that provide dynamic <strong>in</strong>surance (e.g., ?). Underst<strong>and</strong><strong>in</strong>g the dynamic<br />
aspects of health <strong>in</strong>surance are an important challenge for future empirical work.<br />
2.7 Appendix to Chapter 2<br />
This Appendix compares our dem<strong>and</strong> estimates to the broader literature on health<br />
plan choice, <strong>and</strong> discusses alternative specifications of the consumer dem<strong>and</strong> model.<br />
2.7.1 Comparison to Dem<strong>and</strong> Estimates <strong>in</strong> Other Sett<strong>in</strong>gs<br />
As reported <strong>in</strong> Table 2.4, we f<strong>in</strong>d that on average a $100 dollar <strong>in</strong>crease <strong>in</strong> the annual<br />
enrollee contribution decreases market share by 7 to 9 percent. Studies <strong>in</strong> similar<br />
sett<strong>in</strong>gs vary <strong>in</strong> their method for report<strong>in</strong>g elasticities. Follow<strong>in</strong>g ?, we reconcile<br />
the estimates of price elasticity across studies by convert<strong>in</strong>g them to semi-elastiticies<br />
(calculated as the percent change <strong>in</strong> market share <strong>in</strong> response to a $100 <strong>in</strong>crease<br />
<strong>in</strong> the employee contribution). Because the studies cover different time periods, we<br />
adjust the change <strong>in</strong> the employee contribution for <strong>in</strong>flation. After mak<strong>in</strong>g these<br />
adjustments, most studies estimate semi- elastiticities <strong>in</strong> the range of -1.5 to -4.3 for<br />
the ma<strong>in</strong> specifications (????). Estimates from ? are somewhat higher (rang<strong>in</strong>g from<br />
-2.4 to -9.6 for the ma<strong>in</strong> specifications <strong>and</strong> even higher <strong>in</strong> some models). Similarly, ?,<br />
who analyze retirees, have an estimate that implies a semi-elastiticity of -9.0.
CHAPTER 2. HEALTH PLAN CHOICE 100<br />
2.7.2 Alternative Specifications of the Dem<strong>and</strong> Model<br />
As we discuss <strong>in</strong> the ma<strong>in</strong> text, the estimates of risk <strong>and</strong> price elasticity are key<br />
parameters for our welfare calculations. Table 2.9 provides additional estimates of<br />
the dem<strong>and</strong> model to exam<strong>in</strong>e the sensitivity of these parameters. Column 1 of Ta-<br />
ble 2.9 repeats the first specification of Table 2.4 as a basel<strong>in</strong>e. This specification<br />
pools data from enrollees <strong>in</strong> different coverage tiers, <strong>in</strong>creas<strong>in</strong>g the sample size <strong>and</strong><br />
generat<strong>in</strong>g additional variation <strong>in</strong> plan contributions due to the higher pass-through<br />
rates for dependent coverage tiers, but requires us to make assumptions about house-<br />
hold aggregation. We test whether our estimates are sensitive to these assumptions<br />
by restrict<strong>in</strong>g the sample to employees purchas<strong>in</strong>g employee-only coverage (Column<br />
2) <strong>and</strong> controll<strong>in</strong>g for family structure <strong>in</strong> our full-sample specification (Column 3).<br />
Both variations give similar results to our basel<strong>in</strong>e specification (Column 1). While<br />
the estimate of the effect of be<strong>in</strong>g a high risk <strong>in</strong>dividual on dem<strong>and</strong> for the network<br />
PPO is smaller <strong>in</strong> magnitude <strong>in</strong> the employee-only sample (Column 2), the estimate<br />
is imprecise likely because it is identified by a small number of enrollees. When we <strong>in</strong>-<br />
clude controls for family structure <strong>in</strong> the model estimated on the full sample, we f<strong>in</strong>d<br />
that families with a spouse seem to prefer the PPO relative to the average household<br />
(Column 3).<br />
As discussed <strong>in</strong> Section 2.5.4 , we develop different <strong>in</strong>struments to exploit alterna-<br />
tive sources of identify<strong>in</strong>g variation <strong>in</strong> employee contributions, test<strong>in</strong>g the robustness<br />
of our estimates to different exogeneity assumptions. In Table 2.4 (Columns 2 <strong>and</strong> 3),<br />
we demonstrate that the estimate of the price effect is robust to across-plan variation<br />
<strong>in</strong> employer contribution sett<strong>in</strong>g. To address the concern that employer contribution<br />
rates may be systematically <strong>in</strong>fluenced by employee preferences, we <strong>in</strong>strument for<br />
employee contributions with a variant of the contribution model <strong>in</strong> which the the<br />
pass-through coefficients are restricted to be identical across firms (Column 4). We<br />
view the concern that <strong>in</strong>surer bids are correlated with household preferences to be<br />
dim<strong>in</strong>ished because of the limited <strong>in</strong>formation <strong>in</strong>surers have <strong>in</strong> our sett<strong>in</strong>g. Never-<br />
theless we report a specification where we <strong>in</strong>clude predicted bids <strong>in</strong> the contribution<br />
model (Column 5). The results from both these specifications are similar although<br />
with larger st<strong>and</strong>ard errors.
CHAPTER 2. HEALTH PLAN CHOICE 101<br />
F<strong>in</strong>ally, we note that while we present results from a relatively parsimonious<br />
model, our estimates do not appear to be sensitive to <strong>in</strong>clud<strong>in</strong>g a variety of different<br />
control variables. Additional controls we have explored <strong>in</strong>clude plan out-of-pocket<br />
maximums, whether a plan was offered to an employer group prior to the employer<br />
hir<strong>in</strong>g the <strong>in</strong>termediary, <strong>and</strong> the average health status of an employee’s co-workers.
CHAPTER 2. HEALTH PLAN CHOICE 102<br />
Table 2.1: Risk <strong>and</strong> Demographics<br />
Mean Sd. M<strong>in</strong>. Max.<br />
Employees (N = 3683)<br />
Risk Score 1.21 1.56 0.18 30.06<br />
Age 40.56 12.01 18.00 72.00<br />
Female 0.62 0.48 - -<br />
Spouse 0.28 0.45 - -<br />
Child 0.27 0.44 - -<br />
Enrollees (N = 6603)<br />
Risk Score 1.01 1.45 0.14 30.06<br />
Age 32.13 17.67 0.00 72.00<br />
Female 0.58 0.49 - -<br />
Spouse 0.19 0.39 - -<br />
Child 0.26 0.44 - -<br />
Firm-Years (N = 16)<br />
Risk Score 0.97 0.31 0.63 1.91<br />
Age 31.67 4.63 25.71 46.09<br />
Female 0.53 0.12 0.30 0.70<br />
Spouse 0.19 0.07 0.08 0.27<br />
Child 0.26 0.08 0.06 0.39<br />
Employees 230.19 241.51 28.00 838.00<br />
Dependents 182.50 117.51 9.00 331.00<br />
Notes: In the first panel, spouse <strong>and</strong> child refer to the fraction of employees who enroll<br />
with a spouse or at least one child. In the second <strong>and</strong> third panels, these entries are the<br />
fraction of spouses <strong>and</strong> children <strong>in</strong> the set of enrollees. The first <strong>and</strong> second panels pool<br />
observations across firms <strong>and</strong> years. The third panel shows statistics of firm-year level<br />
averages, taken across all enrollees.
CHAPTER 2. HEALTH PLAN CHOICE 103<br />
Table 2.2: Plan Characteristics<br />
Network Integrated<br />
HMO PPO HMO POS All<br />
Offer<strong>in</strong>g Plan<br />
Firms 11 10 11 9 -<br />
Firm-Years 16 14 16 13 -<br />
Bid (Monthly)<br />
Employee 307 332 260 276 294<br />
(64) (59) (30) (26) (54)<br />
Employee plus spouse 645 689 544 579 616<br />
(154) (123) (61) (54) (120)<br />
Employee plus child(ren) 591 632 498 532 565<br />
(143) (115) (58) (53) (111)<br />
Employee plus family 918 989 779 832 882<br />
(200) (176) (87) (76) (164)<br />
Contribution (Monthly)<br />
Employee 45 73 38 58 53<br />
(34) (54) (32) (40) (41)<br />
Employee plus spouse 252 303 203 255 253<br />
(120) (103) (77) (75) (100)<br />
Employee plus child(ren) 221 265 177 223 222<br />
(97) (86) (62) (55) (81)<br />
Employee plus family 418 495 342 415 418<br />
(213) (182) (144) (140) (176)<br />
Co<strong>in</strong>surance (%)<br />
Employee 87 86 97 78 87<br />
(6) (5) (7) (2) (9)<br />
Deductible (Annual)<br />
Employee 387 440 69 336 304<br />
(264) (306) (163) (94) (262)<br />
Out-of-Pocket Max (Annual)<br />
Employee 2818 2850 1591 2686 2468<br />
(462) (474) (625) (731) (775)<br />
Notes: Mean plan characteristics, with st<strong>and</strong>ard deviations <strong>in</strong> parentheses. Plan<br />
characteristics are pooled across years. Co<strong>in</strong>surance, deductible, <strong>and</strong> out-of-pocket<br />
maximum are <strong>in</strong>-network values <strong>and</strong> are highly correlated (ρ >.9) with out-of-network<br />
co<strong>in</strong>surance, deductible <strong>and</strong> out-of-pocket maximum. Coverage tiers based on employee<br />
plus one dependent <strong>and</strong> employee plus two or more dependents are used at two firms.<br />
Bids <strong>and</strong> costs for these coverage tiers are not shown.
CHAPTER 2. HEALTH PLAN CHOICE 104<br />
Table 2.3: Risk <strong>and</strong> Demographics by Plan<br />
Network Integrated<br />
HMO PPO HMO POS All<br />
Employees (N=3683)<br />
Risk Score 1.19 1.22 1.22 1.19 1.21<br />
Age 42.17 40.79 39.73 41.35 40.56<br />
Female 0.62 0.52 0.65 0.56 0.62<br />
Market Share (%) 22.94 7.38 58.72 10.96 100<br />
Enrollees (N=6603)<br />
Risk Score 1.02 1.04 0.99 1.05 1.01<br />
Age 34.19 33.06 30.94 34.12 32.15<br />
Female 0.58 0.54 0.59 0.55 0.58<br />
Market Share (%) 21.24 7.84 60.35 10.57 100<br />
Notes: Employees <strong>and</strong> enrollees are pooled across firms <strong>and</strong> years.
CHAPTER 2. HEALTH PLAN CHOICE 106<br />
Table 2.5: Costs <strong>and</strong> Bids<br />
(1) (2)<br />
Network Insurer Markup 1.29 (0.12) 1.27 (0.07)<br />
Integrated Insurer Markup 1.08 (0.04) 1.07 (0.03)<br />
NHMO 218.42 (21.35) 195.08 (9.48)<br />
X (Risk Score - 1) 288.25 (86.05) 265.36 (30.29)<br />
X Co<strong>in</strong>surance 0.32 (0.94)<br />
NPPO 238.32 (22.65) 204.59 (9.82)<br />
X (Risk Score - 1) 174.92 (34.34) 167.73 (23.81)<br />
X Co<strong>in</strong>surance 1.23 (1.29)<br />
IHMO 234.86 (9.71) 228.77 (6.77)<br />
X (Risk Score - 1) 73.67 (22.01) 104.80 (18.80)<br />
X Co<strong>in</strong>surance 0.38 (0.54)<br />
IPOS 236.37 (14.13) 216.74 (13.01)<br />
X (Risk Score - 1) 188.64 (69.35) 200.78 (61.60)<br />
X Co<strong>in</strong>surance 1.65 (0.82)<br />
N 91 91<br />
Notes: GMM estimates of cost parameters. See text for details. Co<strong>in</strong>surance is de-meaned<br />
at the plan level.
CHAPTER 2. HEALTH PLAN CHOICE 107<br />
Table 2.6: Match<strong>in</strong>g <strong>and</strong> Welfare under Alternative Contribution Policies<br />
Match<strong>in</strong>g Welfare † Truncated<br />
Gross Insurer Social Social<br />
NHMO NPPO IHMO IPOS Surplus ‡ Costs ‡ Surplus ‡ Surplus ‡<br />
Observed<br />
Market Shares 0.25 0.09 0.54 0.12 0.00 0.00 0.00 0.00<br />
Risk Score 1.03 1.07 0.99 1.02<br />
Incremental Contribution † 9.30 23.70 0.00 5.00<br />
Feasible Risk Rated Contributions<br />
Market Shares 0.37 0.09 0.43 0.11 -16.60 -43.70 27.10 5.00<br />
Risk Score 0.58 0.78 1.49 0.74<br />
Incremental Contribution -14.70 11.80 0.00 -1.30<br />
Optimal Risk Rated Contributions<br />
Market Shares 0.38 0.08 0.44 0.10 -22.10 -57.50 35.50 7.80<br />
Risk Score 0.60 0.79 1.46 0.76<br />
Incremental Contribution -14.90 11.80 0.00 -1.60<br />
Uniform by Tier with<strong>in</strong> Firms<br />
Market Shares 0.31 0.09 0.49 0.12 -6.10 -12.80 6.70 1.40<br />
Risk Score 0.86 1.02 1.11 0.97<br />
Incremental Contribution -16.50 8.90 0.00 -1.10<br />
Enthoven Rule<br />
Market Shares 0.22 0.08 0.58 0.13 -1.10 -0.80 -0.30 -0.50<br />
Risk Score 1.01 1.05 1.00 1.02<br />
Incremental Contribution 28.70 39.90 0.00 10.80<br />
Notes: Feasible Risk Rated Contributions implements efficient match<strong>in</strong>g by sett<strong>in</strong>g<br />
<strong>in</strong>cremental contributions equal to <strong>in</strong>cremental costs, conditional on observable risk but<br />
not privately known risk. Optimal Risk Rated Contributions sets <strong>in</strong>cremental<br />
contributions equal to <strong>in</strong>cremental costs, conditional on both observable <strong>and</strong> privately<br />
known risk. Uniform by Tier with<strong>in</strong> Firms maximizes social surplus subject to the<br />
constra<strong>in</strong>t that contributions vary only by coverage tier <strong>and</strong> by firm, but not by <strong>in</strong>dividual<br />
risk. Enthoven Rule is implemented by sett<strong>in</strong>g <strong>in</strong>cremental contributions equal to<br />
<strong>in</strong>cremental bids. Reported risk score is conditional on plan choice. The truncated results<br />
holds cost differentials between plans for risk scores lower than 0.75 <strong>and</strong> higher than 2.0 at<br />
these boundary levels.<br />
† Incremental contribution, gross surplus, <strong>in</strong>surer costs, <strong>and</strong> social surplus are averaged<br />
across enrollees <strong>and</strong> denom<strong>in</strong>ated <strong>in</strong> $ per month.<br />
‡ Gross surplus, <strong>in</strong>surer costs <strong>and</strong> social surplus are normalized to zero under the<br />
observed allocation. Other scenarios show gross surplus as social surplus relative to the<br />
observed allocation. Under the observed allocation, costs average $241.70 per enrollee per<br />
month. Gross <strong>and</strong> social surplus are not p<strong>in</strong>ned down.
CHAPTER 2. HEALTH PLAN CHOICE 108<br />
Table 2.7: Match<strong>in</strong>g <strong>and</strong> Welfare by Risk Score Qu<strong>in</strong>tile<br />
Feasible Risk Rated Contributions versus Observed<br />
Match<strong>in</strong>g Welfare Truncated<br />
∆Gross ∆Insurer ∆Social ∆Social<br />
Qu<strong>in</strong>tile (Risk Score range) NHMO NPPO IHMO IPOS Surplus Costs Surplus Surplus<br />
Qu<strong>in</strong>tile 1 (¡0.36)<br />
∆MarketShare 0.332 0.000 -0.330 -0.002 -27.2 -56.9 29.8 4.3<br />
∆IncrementalContribution -179.4 -93.4 0.0 -86.6<br />
Qu<strong>in</strong>tile 2 (0.36, 0.54)<br />
∆MarketShare 0.265 0.003 -0.266 -0.001 -16.6 -35.6 18.9 3.4<br />
∆IncrementalContribution -141.6 -75.9 0.0 -65.6<br />
Qu<strong>in</strong>tile 3 (0.54, 0.79)<br />
∆MarketShare 0.181 0.006 -0.189 0.002 -7.7 -17.1 9.3 1.3<br />
∆IncrementalContribution -99.1 -53.4 0.0 -44.6<br />
Qu<strong>in</strong>tile 4 (0.79, 1.33)<br />
∆MarketShare 0.040 0.004 -0.037 -0.007 -0.8 -2.4 1.6 0.4<br />
∆IncrementalContribution -21.0 -19.7 0.0 -1.2<br />
Qu<strong>in</strong>tile 5 (¿1.33)<br />
∆MarketShare -0.184 -0.047 0.299 -0.069 -30.3 -105.9 75.6 15.4<br />
∆IncrementalContribution 324.8 154.5 0.0 179.3<br />
Total<br />
∆MarketShare 0.128 -0.007 -0.106 -0.015 -16.6 -43.8 27.1 5.0<br />
∆IncrementalContribution -23.9 -11.9 0.0 -6.3<br />
Notes: ∆ Market Share, ∆ Incremental Contribution, ∆ Gross Surplus, ∆ Insurer Costs<br />
<strong>and</strong> ∆ Social Surplus are calculated as the difference between the feasible risk rated <strong>and</strong><br />
observed values of these variables. Truncated fixes cost differentials between plans for risk<br />
scores outside of 0.75 <strong>and</strong> 2.0. Values averaged across enrollees with<strong>in</strong> each qu<strong>in</strong>tile <strong>and</strong><br />
denom<strong>in</strong>ated <strong>in</strong> $ per month. (Total values are averaged across all enrollees.)
CHAPTER 2. HEALTH PLAN CHOICE 109<br />
Table 2.8: The Value of Plan Choice<br />
Gross Surplus ‡<br />
Welfare †<br />
Insurer Costs ‡<br />
Social Surplus ‡<br />
Observed 0.0 0.0 0.0<br />
All enrolled <strong>in</strong>:<br />
NHMO -148.8 -9.2 -139.7<br />
NPPO -216.9 5.8 -222.7<br />
IHMO -71.4 -2.1 -69.4<br />
IPOS -180.7 4.5 -185.2<br />
Notes: † Gross surplus, <strong>in</strong>surer costs, <strong>and</strong> social surplus are averaged across enrollees<br />
<strong>and</strong> denom<strong>in</strong>ated <strong>in</strong> $ per month.<br />
‡ Gross surplus, <strong>in</strong>surer costs <strong>and</strong> social surplus are normalized to zero under the<br />
observed allocation. Other scenarios show gross surplus as social surplus relative to the<br />
observed allocation. Under the observed allocation, costs average $241.70 per enrollee per<br />
month. Gross <strong>and</strong> social surplus are not p<strong>in</strong>ned down.
CHAPTER 2. HEALTH PLAN CHOICE 110<br />
Figure 1B: Efficient allocation assigns high-risk to plan A<br />
Figure 1A: Efficient allocation assigns high-risk to plan B<br />
"c(!)<br />
Value ($)<br />
Value ($)<br />
High-risk always self-select <strong>in</strong>to plan B.<br />
Choice of "p can <strong>in</strong>duce any threshold.<br />
No uniform "p generates the efficient<br />
allocation.<br />
"v(!)<br />
"v(!)<br />
High-risks always self-select <strong>in</strong>to plan B.<br />
Choice of "p can <strong>in</strong>duce any threshold.<br />
"p= "c( !* ) generates the efficient<br />
allocation of risk types across plans.<br />
"c(!)<br />
"p<br />
"p<br />
Efficiently <strong>in</strong> A<br />
Efficiently <strong>in</strong> B<br />
Efficiently <strong>in</strong> B<br />
Efficiently <strong>in</strong> A<br />
Consumer Risk (!)<br />
!*<br />
Consumer Risk (!)<br />
!*<br />
Figure 2B: Efficient allocation assigns high-risk to HMO<br />
Figure 2A: Efficient allocation assigns high-risk to PPO<br />
High (!,") always self-select to PPO.<br />
Choice of #p can <strong>in</strong>duce any translation<br />
of the pictured <strong>in</strong>difference curve.<br />
#v(!,")=#p<br />
Consumer<br />
taste for<br />
choice(")<br />
High (!,") always self-select to PPO.<br />
Choice of #p can <strong>in</strong>duce any translation<br />
of the pictured <strong>in</strong>difference curve.<br />
#v(!,")=#p<br />
Consumer<br />
taste for<br />
choice(")<br />
#v(!,")=#c(!)<br />
Choose<br />
HMO<br />
<strong>in</strong>efficiently<br />
Choose PPO<br />
efficiently<br />
Choose PPO efficiently<br />
Choose HMO,<br />
Efficiently <strong>in</strong> PPO<br />
Choose PPO<br />
Efficiently <strong>in</strong> HMO<br />
Choose HMO<br />
efficiently<br />
Choose HMO<br />
efficiently<br />
Consumer risk (!)<br />
!’ s.t<br />
#c(!’)=#p<br />
Choose #v(!,")=#c(!)<br />
PPO<br />
<strong>in</strong>efficiently<br />
Consumer risk (!)<br />
!’ s.t<br />
#c(!’)=#p
CHAPTER 2. HEALTH PLAN CHOICE 111<br />
Figure 2.3: Contributions <strong>and</strong> Bids Relative to Integrated HMO<br />
-50 0050<br />
50 100 150 200 -50 0050<br />
50 100 150 200 -50 0050<br />
50 100 150 200 -50 0050<br />
50 100 150 200 Employee Plus Spouse<br />
Employee Plus Spouse<br />
Network PPO<br />
Network PPO<br />
Network HMO<br />
Network HMO<br />
Integrated POS<br />
Integrated POS<br />
Y=X Incremental Contribution<br />
Incremental Contribution<br />
Incremental Bid<br />
Incremental Bid<br />
-50<br />
200<br />
Incremental Contribution<br />
0 50 100 150<br />
-50<br />
-50<br />
0<br />
Employee<br />
50<br />
100<br />
Network PPO<br />
Integrated POS<br />
200<br />
150<br />
100<br />
50<br />
0<br />
-50<br />
150 200 -50<br />
Incremental Bid<br />
Employee Plus Spouse<br />
0<br />
50<br />
Network HMO<br />
Notes: Incremental Contribution <strong>and</strong> Incremental Bid are relative to Integrated HMO. In<br />
Employee Plus Spouse, numbers are divided by two for comparability.<br />
Y=X<br />
100<br />
150<br />
200
CHAPTER 2. HEALTH PLAN CHOICE 112<br />
Figure 2.4: Employer Contributions <strong>and</strong> Employee Characteristics<br />
0.2 .2 .4 .6 .8 .81 1Beta Beta .5 .51 1.5 2 2.5 Risk Score Risk Score 00.2<br />
.2 .4 .6 .8 .81 1Beta Beta 1 1.2 1.4 1.6 1.8 22Risk<br />
Score Risk Score 00.2<br />
.2 .4 .6 .8 .81 1Beta Beta 20000 30000 40000 50000 60000 Income 00.2<br />
.2 .4 .6 .8 .81 1Beta Beta 00500<br />
500 1000 1500 Number of employees<br />
Number of employees<br />
0<br />
Beta Risk Employee versus Score<br />
versus Employee Risk Beta Score<br />
Employee Plus Spouse Beta<br />
Employee Plus Spouse Beta<br />
versus Risk Score<br />
versus Risk Score<br />
Mean Beta versus Income<br />
Mean Beta versus Income<br />
Mean Beta versus Firm Size<br />
Mean Beta versus Firm Size<br />
Employee Beta versus Risk Score<br />
1<br />
.8<br />
Beta<br />
.4 .6<br />
.2<br />
0<br />
1<br />
.8<br />
Beta<br />
.4 .6<br />
.2<br />
0<br />
.5<br />
20000<br />
1<br />
30000<br />
1.5<br />
Risk Score<br />
Mean Beta versus Income<br />
40000<br />
Income<br />
2<br />
50000<br />
2.5<br />
60000<br />
1<br />
.8<br />
Beta<br />
.4 .6<br />
.2<br />
0<br />
1<br />
.8<br />
Beta<br />
.4 .6<br />
.2<br />
0<br />
0<br />
Employee Plus Spouse Beta<br />
versus Risk Score<br />
1<br />
1.2<br />
1.4 1.6<br />
Risk Score<br />
500 1000<br />
Number of employees<br />
1.8<br />
Mean Beta versus Firm Size<br />
Notes: Each po<strong>in</strong>t represents a firm-year. Beta is the <strong>in</strong>cremental pass-through of bids.<br />
See text for details. Employee Beta versus Risk Score plots the beta for employees aga<strong>in</strong>st<br />
their mean risk score. Employee Plus Spouse Beta versus Risk Score plots the beta for<br />
those <strong>in</strong> employee plus spouse plans aga<strong>in</strong>st their risk score. Scatter plots for other<br />
coverage tiers look similar. Mean Beta versus Income plots the mean beta (across<br />
coverage tiers) aga<strong>in</strong>st mean employee <strong>in</strong>come. Mean Beta versus Firm Size plots the<br />
mean beta (across coverage tiers) aga<strong>in</strong>st the number of employees <strong>in</strong> the firm. Separate<br />
scatter plots by coverage tier look similar.<br />
2<br />
1500
CHAPTER 2. HEALTH PLAN CHOICE 113<br />
600<br />
100 200 300 400 500 600 Costs .5 .51 1.5 2 2.5 Risk Score Risk Score Nework Insurer<br />
Nework Insurer<br />
Intergrated Insurer<br />
Intergrated Insurer<br />
100<br />
500<br />
400<br />
Costs<br />
300<br />
200<br />
100<br />
.5<br />
Nework Insurer<br />
Figure 2.5: Costs by Risk Score<br />
1<br />
Intergrated Insurer<br />
1.5<br />
Risk Score<br />
Notes: Each po<strong>in</strong>t represents a <strong>in</strong>surer-employer-year. Risk Score is the average enrollee<br />
risk score <strong>in</strong> a firm-year. Costs is the <strong>in</strong>surers’ monthly cost per enrollee. Fitted l<strong>in</strong>es<br />
represent the Network HMO <strong>and</strong> Integrated HMO.<br />
2<br />
2.5
CHAPTER 2. HEALTH PLAN CHOICE 114<br />
100 200 300 400 500 600 Bids .5 .51 1.5 22E[Risk<br />
Score|Age,Male]<br />
E[Risk Score|Age,Male]<br />
Network HMO<br />
Network HMO<br />
Network PPO<br />
Network PPO<br />
Intergrated HMO<br />
Intergrated HMO<br />
Integrated POS<br />
Integrated POS<br />
100<br />
600<br />
500<br />
Bids<br />
400<br />
300<br />
200<br />
100<br />
.5<br />
Network HMO<br />
Intergrated HMO<br />
Figure 2.6: Bids by Risk Score<br />
Network PPO<br />
Integrated POS<br />
1<br />
1.5<br />
E[Risk Score|Age,Male]<br />
Notes: Each po<strong>in</strong>t represents a plan-employer-year. E[Risk—Age, Male] is the average<br />
predicted risk score of potential enrollees <strong>in</strong> a firm-year. Bids is the per-month bid. Fitted<br />
l<strong>in</strong>es represent the Network HMO <strong>and</strong> Integrated HMO.<br />
2
Chapter 3<br />
Private Coverage <strong>and</strong> Public<br />
Costs: Identify<strong>in</strong>g the Effect of<br />
Private Supplemental Insurance on<br />
Medicare Spend<strong>in</strong>g<br />
with Marika Cabral<br />
3.1 Introduction<br />
In many <strong>in</strong>surance contexts, <strong>in</strong>dividuals hold primary <strong>and</strong> supplemental <strong>in</strong>surance<br />
policies for the same basic outcome. A common example is the purchase of pri-<br />
vate supplemental health <strong>in</strong>surance to top up <strong>public</strong>ly provided primary coverage. 1<br />
However, unlike st<strong>and</strong>ard markets where complementarities between products are<br />
typically isolated to the dem<strong>and</strong>-side, complementarities between primary <strong>and</strong> sup-<br />
plemental <strong>in</strong>surance can affect costs as well. A supplemental <strong>in</strong>surance policy that<br />
reduces the consumer cost-shar<strong>in</strong>g requirements (e.g., deductibles <strong>and</strong> co-payments)<br />
1 Other examples <strong>in</strong>clude disability <strong>in</strong>surance, where private coverage can supplement <strong>public</strong> disability<br />
<strong>in</strong>surance (SSDI) <strong>and</strong> rental car <strong>in</strong>surance, where coverage from a credit card can top-up the<br />
primary rental car <strong>in</strong>surance provided by a rental company.<br />
117
CHAPTER 3. MEDIGAP 118<br />
of a primary <strong>in</strong>surance policy can affect total utilization, impos<strong>in</strong>g a fiscal externality<br />
on the provider of the primary <strong>in</strong>surance product.<br />
This paper estimates the effect of private supplemental Medigap <strong>in</strong>surance on pub-<br />
lic Medicare spend<strong>in</strong>g. Medicare is the primary <strong>in</strong>surer of elderly Americans, cover<strong>in</strong>g<br />
about 70 percent of the care they receive. To <strong>in</strong>sure aga<strong>in</strong>st the rema<strong>in</strong><strong>in</strong>g costs, most<br />
Medicare beneficiaries (91 percent) hold some form of supplemental <strong>in</strong>surance, which<br />
reduces out-of-pocket f<strong>in</strong>ancial risk but also blunts the cost-shar<strong>in</strong>g requirements of<br />
Medicare that are designed to limit utilization.<br />
To identify the effect of Medigap supplemental <strong>in</strong>surance on total utilization, we<br />
construct an <strong>in</strong>strument that leverages discont<strong>in</strong>uities <strong>in</strong> Medigap premiums at state<br />
boundaries. Medical costs, <strong>and</strong> thus the costs f<strong>in</strong>anced through supplemental Medi-<br />
gap <strong>in</strong>surance, exhibit considerable with<strong>in</strong>-state variation due to factors rang<strong>in</strong>g from<br />
household <strong>in</strong>comes to local physician practice styles to the supply of medical resources.<br />
Yet despite this local variation, with<strong>in</strong>-state variation <strong>in</strong> Medigap premiums is very<br />
limited (?). This means that on opposite sides of state boundaries, otherwise identical<br />
<strong>in</strong>dividuals who belong to the same hospital catchment area can face very different<br />
Medigap premiums solely due to Medicare costs <strong>in</strong> other parts of their states. Our<br />
empirical strategy uses this exogenous premium variation—<strong>and</strong> the correspond<strong>in</strong>g<br />
exogenous shift of <strong>in</strong>dividuals <strong>in</strong> <strong>and</strong> out of Medigap coverage—to identify the pre-<br />
mium elasticity of Medigap dem<strong>and</strong> <strong>and</strong> the causal effect of Medigap on total medical<br />
spend<strong>in</strong>g.<br />
The effect we identify is of substantial policy relevance. The high level <strong>and</strong> growth<br />
rate of Medicare spend<strong>in</strong>g makes it a primary concern of long-run US federal budget<br />
experts (?). Because of this, mechanisms that <strong>in</strong>crease costs <strong>and</strong> the correspond<strong>in</strong>g<br />
policies that could be used to address these ris<strong>in</strong>g costs are attract<strong>in</strong>g close atten-<br />
tion. However, as we discuss below, the current estimates of the effect of Medigap are<br />
biased due to selection on unobservables. This bias means that previous estimates<br />
may not reflect the behavioral responses to typical policies, such as a tax on Medi-<br />
gap premiums. Our approach overcomes these shortcom<strong>in</strong>gs allow<strong>in</strong>g us to explore<br />
economically mean<strong>in</strong>gful policy counterfactuals.<br />
In addition to the direct relevance for Medicare <strong>and</strong> Medigap policy, our study may
CHAPTER 3. MEDIGAP 119<br />
be of <strong>in</strong>terest more generally. Sett<strong>in</strong>gs with private supplemental coverage <strong>and</strong> <strong>public</strong><br />
primary <strong>in</strong>surance are widespread. For example, <strong>in</strong> France more than 92 percent of<br />
the population holds private supplemental <strong>in</strong>surance to protect aga<strong>in</strong>st the substantial<br />
co<strong>in</strong>surance payments (10 to 40 percent) of the universal <strong>public</strong> health <strong>in</strong>surance<br />
system, <strong>and</strong> private supplemental policies that top up <strong>public</strong> health <strong>in</strong>surance benefits<br />
are popular <strong>in</strong> Austria, Belgium, <strong>and</strong> Denmark as well. 2 Yet despite economists’<br />
<strong>in</strong>terest <strong>in</strong> the <strong>in</strong>teraction of <strong>public</strong> <strong>and</strong> private <strong>in</strong>surance as demonstrated by the<br />
large literature on crowd-out (cf. ??), there is relatively little work on the fiscal<br />
externalities associated with jo<strong>in</strong>tly held private <strong>and</strong> <strong>public</strong> <strong>in</strong>surance. 3<br />
At a conceptual level, supplemental Medigap coverage could impose a negative or<br />
positive fiscal externality on Medicare. Because Medigap policies reduce the marg<strong>in</strong>al<br />
price for care, it is straightforward to see that overall utilization <strong>and</strong> Medicare costs<br />
could <strong>in</strong>crease. Medigap policies could alternatively decrease costs if, for example,<br />
reduced cost-shar<strong>in</strong>g improved treatment adherence, reduc<strong>in</strong>g the risk of hospitaliza-<br />
tion. In the context of Medicare supplemental <strong>in</strong>surance available to retired California<br />
state employees, ? f<strong>in</strong>d there is a significant positive cost externality of supplemental<br />
<strong>in</strong>surance, what they call an “offset effect,” for those that are chronically ill. 4 For<br />
the overall retiree population, however, the externality is negative with supplemental<br />
<strong>in</strong>surance <strong>in</strong>creas<strong>in</strong>g Medicare costs.<br />
The conventional wisdom is that supplemental <strong>in</strong>surance on net imposes a negative<br />
fiscal externality on Medicare, although the magnitude of the effect is contended. The<br />
central estimate, used by the Congressional Budget Office (CBO) for official budget<br />
2 Statistics are from the follow<strong>in</strong>g sources: ? “Selected European Countries Health Systems,” <strong>and</strong><br />
? “The Grass Is Not Always Greener: A Look at National Health Care Systems Around the World”<br />
Cato Policy Analysis no. 613. In Austria, about a third of the population has a supplemental private<br />
<strong>in</strong>surance plan that covers additional charges not covered under the basic benefits, <strong>and</strong> about 30<br />
percent of Belgians carry private supplemental policies. Approximately 30 percent of population<br />
<strong>in</strong> Denmark purchases Voluntary Health Insurance (VHI) <strong>in</strong> order to cover the costs of statutory<br />
copayments of the universal coverage package.<br />
3 Exceptions <strong>in</strong>clude ?, ?, <strong>and</strong> ? on supplemental <strong>in</strong>surance.<br />
4 The authors argue that drug coverage is probably the s<strong>in</strong>gle most important channel through<br />
with the offset operates for the chronically ill. Even though the time period we look at is before<br />
Medicare Part D drug coverage was added, Medigap options that had drug coverage were very<br />
unpopular as their premiums were very high. Thus, we may expect the even among the chronically<br />
ill, Medigap private supplemental <strong>in</strong>surance may not have the offset effect the authors observe <strong>in</strong><br />
the CALPERs population.
CHAPTER 3. MEDIGAP 120<br />
estimates, is that “Medigap policyholders use at least 25 percent more services than<br />
Medicare enrollees who have no supplemental coverage.” Yet, as po<strong>in</strong>ted out by ?, the<br />
research CBO draws upon does not account for selection <strong>in</strong> Medigap enrollment <strong>and</strong><br />
may be biased. In particular, Lemiuex et al. argue that selection is probably adverse,<br />
lead<strong>in</strong>g these studies to overestimate the impact of Medigap on costs. Recent work,<br />
however, f<strong>in</strong>ds little or no evidence of adverse selection <strong>in</strong>to Medigap. 5 In fact, Fang<br />
et al. (2008) f<strong>in</strong>d evidence of significant advantageous selection <strong>in</strong>to Medigap. The<br />
authors argue that this advantageous selection is driven by cognitive ability, which is<br />
correlated with better health.<br />
We view our study as a natural next step <strong>in</strong> this literature. Our approach addresses<br />
the issue of selection by us<strong>in</strong>g plausibly exogenous variation <strong>in</strong> Medigap premiums.<br />
In particular, we take advantage of the fact that there is substantial with<strong>in</strong>-state<br />
variation costs but limited with<strong>in</strong>-state variation <strong>in</strong> premiums <strong>in</strong> the Medigap market.<br />
Us<strong>in</strong>g local medical costs as a control, we f<strong>in</strong>d that with<strong>in</strong>-state but out of locality<br />
medical costs are a powerful predictor of premiums, provid<strong>in</strong>g us with a source of<br />
premium variation that is unrelated to unobserved determ<strong>in</strong>ants of <strong>in</strong>dividual medical<br />
costs. We estimate a Medigap selection equation us<strong>in</strong>g this variation to identify the<br />
premium elasticity of Medigap dem<strong>and</strong>. Us<strong>in</strong>g the predicted values from this selection<br />
equation to <strong>in</strong>strument for Medigap enrollment, we consistently estimate the impact<br />
of Medigap on medical utilization.<br />
Our ma<strong>in</strong> f<strong>in</strong>d<strong>in</strong>g is Medigap coverage <strong>in</strong>creases overall medical costs by 57 percent<br />
on the marg<strong>in</strong>. The effect is about 40 percent larger than comparable OLS estimates,<br />
consistent with advantageous selection. We also f<strong>in</strong>d that Medigap dem<strong>and</strong> is mod-<br />
erately elastic, with our preferred estimate <strong>in</strong>dicat<strong>in</strong>g that a 10 percent <strong>in</strong>crease <strong>in</strong><br />
premiums reduces take-up by 4.7 percentage po<strong>in</strong>ts. Both the Medigap effect <strong>and</strong><br />
dem<strong>and</strong> semi-elasticity are precisely estimated <strong>and</strong> robust to alternative functional<br />
forms <strong>and</strong> sets of controls.<br />
In addition to <strong>in</strong>dicat<strong>in</strong>g advantageous selection, the 40 percent larger <strong>in</strong>strumen-<br />
tal variables estimates for the effect of Medigap are consistent with heightened moral<br />
5 These papers <strong>in</strong>clude ?, ?, ?, <strong>and</strong> ?.
CHAPTER 3. MEDIGAP 121<br />
hazard for <strong>in</strong>dividuals on the marg<strong>in</strong> of Medigap choice. S<strong>in</strong>ce these marg<strong>in</strong>al <strong>in</strong>di-<br />
viduals are more price sensitive when it comes to premiums, it is natural for them to<br />
be more sensitive to the cost-shar<strong>in</strong>g <strong>in</strong>struments of Medigap, suggest<strong>in</strong>g larger ef-<br />
fects for these marg<strong>in</strong>al <strong>in</strong>dividuals than the average Medigap policyholder. 6 For our<br />
purposes, this feature is a strength of our empirical strategy rather than a weakness.<br />
The effect for <strong>in</strong>dividuals marg<strong>in</strong>al to a change <strong>in</strong> premiums is the effect of <strong>in</strong>terest<br />
for counterfactuals <strong>in</strong>volv<strong>in</strong>g a premium tax or subsidy, as the <strong>in</strong>dividuals identify<strong>in</strong>g<br />
the estimated effect are exactly those who would be affected by such a policy.<br />
We construct back-of-the-envelope policy counterfactuals of taxes <strong>and</strong> subsidies to<br />
Medigap premiums us<strong>in</strong>g our estimated parameters. A 20 percent tax would reduce<br />
coverage by 9.5 percentage po<strong>in</strong>ts, imply<strong>in</strong>g a total budgetary sav<strong>in</strong>gs of 6.2 percent of<br />
Medicare costs of the sample. About 1/4 of the sav<strong>in</strong>gs would come from tax revenues<br />
<strong>and</strong> 3/4 from reduced medical costs. A 40 percent tax would save 11.4 percent <strong>and</strong><br />
a Pigovian tax that fully accounts for the negative fiscal externality would generate<br />
sav<strong>in</strong>gs of 18.1 percent. We note that our parameters are estimated us<strong>in</strong>g modest<br />
premium variation, <strong>and</strong> thus our policy counterfactuals, which <strong>in</strong>volve large taxes<br />
<strong>and</strong> subsidies, should be viewed with some caution. As we discuss below, a more<br />
nuanced tax policy that accounts for offsets might be able to generate both larger<br />
costs sav<strong>in</strong>gs <strong>and</strong> better health.<br />
The rema<strong>in</strong>der of the paper proceeds as follows. Section 3.2 describes background<br />
<strong>in</strong>formation about Medicare <strong>and</strong> the supplemental <strong>in</strong>surance market. Section 3.3<br />
briefly describes the data used <strong>in</strong> our empirical analysis. Section 3.4 presents our<br />
empirical model. In Section 3.5 , we discuss our identification strategy. Results are<br />
presented <strong>in</strong> Section 3.6. Section 3.7 conta<strong>in</strong>s the policy counterfactuals. Section 3.8<br />
concludes.<br />
6 The <strong>in</strong>tuition for this relationship that suggests <strong>in</strong>dividuals at the marg<strong>in</strong> of select<strong>in</strong>g <strong>in</strong>to<br />
<strong>in</strong>surance based on premiums will be more sensitive to reductions <strong>in</strong> co<strong>in</strong>surance. A similar <strong>in</strong>tution<br />
is formalized <strong>in</strong> a model <strong>in</strong> a work<strong>in</strong>g paper by ?. In their paper, the authors refer to this sort of<br />
selection as “selection on moral hazard.”
CHAPTER 3. MEDIGAP 122<br />
3.2 Background<br />
Medicare is the government health <strong>in</strong>surance program for elderly, cover<strong>in</strong>g nearly<br />
all <strong>in</strong>dividuals 65 <strong>and</strong> older <strong>in</strong> the United States. In 2008, Medicare covered 44<br />
million beneficiaries at a cost of $468 billion, account<strong>in</strong>g for about 13 percent of<br />
government expenditure <strong>and</strong> 3.2 percent of GDP. By 2025, Medicare is projected to<br />
cover 71 million beneficiaries, consum<strong>in</strong>g 20 percent of government expenditure <strong>and</strong><br />
5.5 percent of GDP. 7<br />
Medicare beneficiaries choose between <strong>public</strong>ly adm<strong>in</strong>istered traditional fee-for-<br />
service (FFS) <strong>in</strong>surance coverage <strong>and</strong> Medicare Advantage policies, private <strong>in</strong>surance<br />
plans with premiums subsidized by Medicare. 8 The majority (88 percent) of Medicare<br />
beneficiaries have FFS coverage. This coverage allows beneficiaries their choice of<br />
doctor <strong>and</strong> the ability to see a specialist without a referral. To control costs, fee-<br />
for-service Medicare uses cost-shar<strong>in</strong>g, expos<strong>in</strong>g beneficiaries to about 20 percent of<br />
the cost of care received on the marg<strong>in</strong>. A key feature of traditional Medicare is<br />
that there is no annual or lifetime out-of-pocket maximum amount that a beneficiary<br />
will pay, so <strong>in</strong>dividuals are exposed to significant f<strong>in</strong>ancial risk. Traditional Medicare<br />
cost-shar<strong>in</strong>g <strong>in</strong>struments for 2000 <strong>and</strong> 2005 (the first <strong>and</strong> last years of our sample)<br />
are shown <strong>in</strong> Table 3.1.<br />
To protect aga<strong>in</strong>st the f<strong>in</strong>ancial risk of traditional Medicare, the vast majority<br />
(91 percent) of beneficiaries carry supplemental <strong>in</strong>surance. Some beneficiaries obta<strong>in</strong><br />
supplemental <strong>in</strong>surance through a former employer, poor beneficiaries may qualify for<br />
supplemental <strong>in</strong>surance through the government sponsored Medicaid program, <strong>and</strong><br />
everyone else has the option to purchase private Medigap coverage. Medicare Ad-<br />
vantage is an alternative to traditional fee-for-service Medicare that is available to<br />
many beneficiaries. 9 These private <strong>in</strong>surance policies are subsidized by Medicare, <strong>and</strong><br />
7 Percent of GDP numbers are gross of premiums; percent of budget numbers are net (Annual<br />
Report of the Boards of Trustees of the Federal Hospital Insurance <strong>and</strong> Federal Supplementary<br />
Medical Insurance Trust Funds, 2010; Analytical Perspectives: Budget of the U.S. Government,<br />
2010).<br />
8 We will refer to Medicare Part C as Medicare Advantage <strong>in</strong> this paper, although this option was<br />
called Medicare + Choice dur<strong>in</strong>g the beg<strong>in</strong>n<strong>in</strong>g of the period we analyze.<br />
9 Although Medicare Advantage beneficiaries can technically sign up for Medigap policies, this is<br />
discouraged by Medicare <strong>and</strong> <strong>in</strong>dividuals do not seem to do this <strong>in</strong> practice. Medicare Advantage
CHAPTER 3. MEDIGAP 123<br />
have benefits that generally cover more services than offered <strong>in</strong> traditional Medicare<br />
<strong>in</strong> exchange for beneficiaries accept<strong>in</strong>g a more restricted network of providers. In our<br />
analysis, we model an <strong>in</strong>dividual’s choice between three options: Medicare Advan-<br />
tage, traditional Medicare without supplemental <strong>in</strong>surance, <strong>and</strong> traditional Medicare<br />
with supplemental <strong>in</strong>surance. In do<strong>in</strong>g this, we exclude <strong>in</strong>dividuals on Medicaid or<br />
employer provided supplemental health <strong>in</strong>surance. 10 Table 3.2 shows the breakdown<br />
of <strong>in</strong>dividuals on the various types of supplemental <strong>in</strong>surance.<br />
Medigap beneficiaries typically purchase Medigap <strong>in</strong>surance with<strong>in</strong> six months of<br />
turn<strong>in</strong>g 65 years old <strong>and</strong> sign<strong>in</strong>g up for traditional Medicare, dur<strong>in</strong>g what is called the<br />
open enrollment period. Dur<strong>in</strong>g this period, <strong>in</strong>dividuals cannot be denied coverage for<br />
any reason, <strong>and</strong> pric<strong>in</strong>g is limited to a small set of characteristics: gender, location,<br />
<strong>and</strong> smok<strong>in</strong>g status of the enrollee. 11 Plans bought dur<strong>in</strong>g this period are guaranteed<br />
renewable as long as Medigap enrollees pay plan premiums every year. The federal<br />
government regulates how Medigap policy prices can evolve. In particular, when<br />
an <strong>in</strong>dividual enrolls <strong>in</strong> a Medigap plan, he is choos<strong>in</strong>g an age-price profile that<br />
may be adjusted with medical <strong>in</strong>flation but may not be cont<strong>in</strong>gent on his current<br />
or future health status. Thus, along with the contemporaneous benefits, Medigap<br />
coverage provides <strong>in</strong>surance aga<strong>in</strong>st reclassification risk <strong>in</strong> future periods. In practice,<br />
<strong>in</strong>dividuals tend to sign up for a Medigap plan at age 65, <strong>and</strong> renew their policy each<br />
year. 12<br />
In most states, <strong>in</strong>dividuals choose from a menu of st<strong>and</strong>ardized plans labeled<br />
A through J, adm<strong>in</strong>istered by compet<strong>in</strong>g private <strong>in</strong>surance providers. 13 Table 3.3<br />
plans tend to give more f<strong>in</strong>ancial coverage <strong>in</strong> exchange for a more restricted network of providers relative<br />
to traditional Medicare. S<strong>in</strong>ce Medigap polices are tailored to fill gaps of traditional Medicare,<br />
the supplemental coverage that these policies would provide an <strong>in</strong>dividual on Medicare Advantage<br />
would be largely redundant.<br />
10 Those who are offered employer sponsored <strong>in</strong>surance typically receive a good deal <strong>in</strong> that the<br />
plans are very subsidized. S<strong>in</strong>ce we believe that the take-up rate on employer-sponsored <strong>in</strong>surance<br />
is nearly 100 percent, we excluded those on employer sponsored supplemental <strong>in</strong>surance from our<br />
analysis as they have a different <strong>and</strong> unobserved choice set.<br />
11 In practice, smok<strong>in</strong>g status is not priced dur<strong>in</strong>g the sample period we exam<strong>in</strong>e.<br />
12 This is the pattern we observe <strong>in</strong> the 3-year panel of the Medicare Current Beneficiary Survey.<br />
13 Massachusetts, Wiscons<strong>in</strong>, <strong>and</strong> M<strong>in</strong>nesota st<strong>and</strong>ardized their plans prior to federal regulation<br />
<strong>and</strong> have cont<strong>in</strong>ued their own offer<strong>in</strong>gs. The Medicare Prescription Drug, Improvement, <strong>and</strong> Modernization<br />
Act of 2003 (MMA) <strong>in</strong>troduced plans K <strong>and</strong> L <strong>and</strong> elim<strong>in</strong>ated the sale of Medigap plans<br />
with drug benefits (H, I <strong>and</strong> J).
CHAPTER 3. MEDIGAP 124<br />
shows enrollment by plan letter <strong>and</strong> the plan characteristics <strong>in</strong> detail. Plans C <strong>and</strong><br />
F, by far the most popular plans, are chosen by over 60 percent of the Medigap<br />
beneficiaries. Both of these plans cover the hospital <strong>and</strong> physician deductibles, <strong>in</strong><br />
addition to the basic benefits common to all lettered plans. In this paper, we estimate<br />
the marg<strong>in</strong>al effect of the average Medigap plan, rather than the effect of particular<br />
plan characteristics. We believe the marg<strong>in</strong>al effect of Medigap captures a mean<strong>in</strong>gful<br />
concept <strong>in</strong> this context because the vast majority of <strong>in</strong>dividuals are on Medigap plans<br />
face the same marg<strong>in</strong>al change <strong>in</strong> prices. The common characteristics of the Medigap<br />
plans are ex ante the most likely drivers of moral hazard accord<strong>in</strong>g to previous studies.<br />
In particular, it is thought that physician expenditures is one of the most flexible<br />
marg<strong>in</strong>s of health care spend<strong>in</strong>g, <strong>and</strong> all plans <strong>in</strong>clude Part B physician co<strong>in</strong>surance<br />
coverage. 14 Part B co<strong>in</strong>surance coverage it thought to generate substantial moral<br />
hazard because it dramatically reduces the marg<strong>in</strong>al prices for doctor visits for almost<br />
everyone with physician expenditures. 15<br />
3.3 Data<br />
We have two primary sources of data. We use the Medicare Current Beneficiary<br />
Survey (MCBS) for 2000 through 2005. This data is a short panel of Medicare<br />
beneficiaries that <strong>in</strong>cludes cost data by payer, <strong>in</strong>formation about Medigap coverage,<br />
demographic <strong>in</strong>formation <strong>and</strong> relatively detailed health status variables. In addition<br />
to the MCBS data, we use Medigap premium data from Weiss Rat<strong>in</strong>gs. We obta<strong>in</strong>ed<br />
premium data for 2000 <strong>and</strong> 2003. This data <strong>in</strong>cludes premiums charged dur<strong>in</strong>g the<br />
open enrollment period by company, state, <strong>and</strong> plan letter. We use these two years<br />
of premiums to extrapolate premiums from 2000 to 2005. 16<br />
We match MCBS beneficiaries with the prices they faced at age 65. This means<br />
14 See ?.<br />
15 Because elderly <strong>in</strong>dividuals almost all have physician expenditures of 135 dollars annually (the<br />
approximate value of the Part B deductible), the marg<strong>in</strong>al price they face is the 20 percent co<strong>in</strong>surance<br />
rate for physician visits beyond this deductible. Compare this to the Part B deductible<br />
coverage, which is not available <strong>in</strong> all plans, but probably doesnt lead to much moral hazard.<br />
16 In particular, we assume a l<strong>in</strong>ear trend <strong>in</strong> premiums between 2000 <strong>and</strong> 2003. Us<strong>in</strong>g this trend,<br />
we project forward until 2005.
CHAPTER 3. MEDIGAP 128<br />
We use bootstrap methods to calculate the st<strong>and</strong>ard errors, loop<strong>in</strong>g over the entire<br />
set of equations for consistency.<br />
3.5 Identify<strong>in</strong>g Variation<br />
3.5.1 The Instrument<br />
The basic concept<br />
The idea of the empirical model is to use plausibly exogenous variation <strong>in</strong> Medigap<br />
premiums (equation 3.3) to shift <strong>in</strong>dividuals <strong>in</strong> <strong>and</strong> out of Medigap coverage (equation<br />
3.2) <strong>and</strong> thereby identify the impact of Medigap on total medical spend<strong>in</strong>g (equation<br />
3.1).<br />
To isolate variation <strong>in</strong> premiums that is conditionally r<strong>and</strong>om, we leverage discon-<br />
t<strong>in</strong>uities <strong>in</strong> Medigap premiums at state boundaries. As we discuss below, Medicare<br />
costs, <strong>and</strong> thus the costs of supplemental Medigap <strong>in</strong>surance, exhibit considerable<br />
with<strong>in</strong>-state variation due to factors rang<strong>in</strong>g from household <strong>in</strong>comes to local physi-<br />
cian practice styles to the supply of medical resources (????). Yet despite this local<br />
variation, with<strong>in</strong>-state variation <strong>in</strong> Medigap premiums is highly limited (?). 17 This<br />
means that otherwise identical <strong>in</strong>dividuals <strong>in</strong> otherwise identical localities can face<br />
sharply different Medigap premiums solely due to Medicare costs <strong>in</strong> other parts of<br />
their state.<br />
We put this idea <strong>in</strong>to practice by us<strong>in</strong>g state-level costs as an <strong>in</strong>strument while<br />
simultaneously controll<strong>in</strong>g for local costs <strong>in</strong> all the equations of our model. Controll<strong>in</strong>g<br />
for local costs is of course essential, as factors that <strong>in</strong>fluence costs <strong>in</strong> a local area almost<br />
surely <strong>in</strong>fluence costs <strong>and</strong> enrollment decisions <strong>in</strong> our sample, even conditional on the<br />
set of <strong>in</strong>dividual controls we <strong>in</strong>clude.<br />
17 Although firms are allowed to vary premiums at the zip code level, ? f<strong>in</strong>d there is very little<br />
with<strong>in</strong>-state variation <strong>in</strong> the Medigap premiums for a given plan offered by given <strong>in</strong>surance company.<br />
Maestas et al. cite state-level report<strong>in</strong>g requirements <strong>and</strong> regulations as a potential explanation.<br />
Because plans must, for example, meet loss ratio requirements at state level, vary<strong>in</strong>g premiums<br />
more locally may be adm<strong>in</strong>istratively burdensome. We discuss a robustness check to this issue <strong>in</strong><br />
Section 3.6.
CHAPTER 3. MEDIGAP 129<br />
Data <strong>and</strong> summary statistics<br />
In particular, we implement our <strong>in</strong>strumental variables strategy by us<strong>in</strong>g fee-for-<br />
service (FFS) Medicare cost data aggregated by the Dartmouth Atlas to the state<br />
<strong>and</strong> local level. 18 The advantage of the Dartmouth Atlas data is the methodology used<br />
to def<strong>in</strong>e local geographic areas. The Atlas aggregates cost data at two levels: the<br />
Hospital Service Area (HSA) <strong>and</strong> Hospital Referral Region (HRR). The HSA is the<br />
smaller unit of aggregation, def<strong>in</strong>ed by the set of zip codes whose residents receive most<br />
of their hospitalizations from hospitals <strong>in</strong> that area. These areas are approximately<br />
the size of a county: there are 3,446 HSA <strong>and</strong> 3,140 counties <strong>in</strong> the United States. To<br />
capture larger areas, the Dartmouth Atlas looks at where <strong>in</strong>dividuals are referred to<br />
for major cardiovascular surgery <strong>and</strong> neurosurgery. These are typically big, superstar<br />
hospitals, usual <strong>in</strong> big cities. HRR are more than 10 times as large as HSA: there are<br />
306 HRR <strong>in</strong> the country. To be conservative, we use larger HRR as the local area <strong>in</strong><br />
our analysis.<br />
Table 3.4 summarizes HRR level data on per capita FFS Medicare costs, tabulated<br />
by state. For expositional purposes, we show data from 2003, a year <strong>in</strong> the middle<br />
of our sample period. As is well known, there are vast differences <strong>in</strong> per capita costs<br />
between regions: costs range from $4,520 per capita <strong>in</strong> Salem, OR to $11,588 per<br />
capita <strong>in</strong> the Bronx, NY. A fact that is less well known, but vital to the power of<br />
our <strong>in</strong>strument, is that there are substantial differences <strong>in</strong> HRR-level costs with<strong>in</strong><br />
states as well. Per capita costs vary by more than $5,000 with<strong>in</strong> NY, FL <strong>and</strong> TX,<br />
<strong>and</strong> by more than $1,000 with<strong>in</strong> 25 states that are home to most (75 percent) of the<br />
<strong>in</strong>dividuals <strong>in</strong> our sample.<br />
Instrument power<br />
Figure 3.1 provides a concrete example of the power of our <strong>in</strong>strument. Panel A shows<br />
a map of per capita FFS Medicare costs <strong>in</strong> the states of New York <strong>and</strong> Vermont by<br />
HRR. Panel B shows Medigap premiums <strong>in</strong> the same two states. A comparison<br />
18 The Dartmouth Atlas data is constructed from the Cont<strong>in</strong>uous Medicare History Sample<br />
(CMHS), a 5 percent sample of the bill<strong>in</strong>g records Medicare beneficiaries collected by the Centers<br />
for Medicare <strong>and</strong> Medicaid Services (CMS).
CHAPTER 3. MEDIGAP 130<br />
of costs <strong>and</strong> premiums <strong>in</strong> upstate New York <strong>and</strong> Vermont is illum<strong>in</strong>at<strong>in</strong>g. While<br />
these two regions have almost identical per capita Medicare costs (no more than<br />
$5,864 <strong>in</strong> upstate NY versus $5,858 <strong>in</strong> VT), Medigap premiums <strong>in</strong> these regions are<br />
significant different ($1,677 versus $1,320). The bottom of the map shows why. New<br />
York state of course has New York City to its south, a region with substantially<br />
higher Medicare costs than the northern part of the state (maximums of $11,588 <strong>in</strong><br />
the south versus $5,864 upstate). It is the high cost metropolitan south, comb<strong>in</strong>ed<br />
with the limited with<strong>in</strong>-state variation <strong>in</strong> premiums, which <strong>in</strong>flates Medigap prices <strong>in</strong><br />
upstate New York without impact<strong>in</strong>g premiums <strong>in</strong> Vermont, creat<strong>in</strong>g an exogenous<br />
source of premium variation.<br />
To exam<strong>in</strong>e the statistical power of the <strong>in</strong>strument, we calculate the <strong>in</strong>cremental<br />
power – or partial R-squared – from add<strong>in</strong>g state costs to a regression of premiums<br />
on HRR-level costs. That is, we calculate the additional power from state costs<br />
when already controll<strong>in</strong>g for costs <strong>in</strong> the local region. As shown <strong>in</strong> Appendix Table<br />
3.10, a regression of log premiums on log HRR costs has an R-squared of 0.19 <strong>and</strong> a<br />
coefficient on HRR costs of 0.35. Add<strong>in</strong>g state costs reduces the coefficient on HRR<br />
costs to zero <strong>and</strong> <strong>in</strong>creases the R-squared from 0.19 to 0.50, imply<strong>in</strong>g that 62 percent<br />
(=0.31/0.50) of the explanatory power of Medicare costs comes from outside local<br />
HRRs. The F-statistic on state costs is highly significant.<br />
3.5.2 Potential Concerns<br />
The exclusion restriction<br />
A key assumption underp<strong>in</strong>n<strong>in</strong>g our empirical strategy is that, conditional on HRR-<br />
level costs, state-level Medicare costs only impact utilization through premiums <strong>and</strong><br />
Medicare choice. One potential concern with this assumption is that because <strong>in</strong>di-<br />
viduals travel outside their local HRRs to receive health care, the local measure of<br />
utilization is an <strong>in</strong>sufficient control. This concern is misplaced, however, s<strong>in</strong>ce local<br />
HRR costs account for travel, measur<strong>in</strong>g total utilization based on the location of<br />
residence regardless of where care is received. Thus, if <strong>in</strong>dividuals <strong>in</strong> one area tend<br />
to travel to another place with higher (or lower) costs for medical services, utilization
CHAPTER 3. MEDIGAP 131<br />
levels <strong>in</strong> this area will reflect this propensity to travel.<br />
A more subtle concern is that the marg<strong>in</strong>al Medigap purchaser may be more<br />
(or less) likely to travel than the average Medicare beneficiary. If this is the case,<br />
then controll<strong>in</strong>g for average local health care utilization is <strong>in</strong>sufficient. Basically, the<br />
concern is that although most people <strong>in</strong> Rural Town, USA don’t travel to Big City,<br />
USA for care, the marg<strong>in</strong>al Medigap purchaser might be the type of person who does.<br />
A solution to this problem is to def<strong>in</strong>e markets broadly enough so that the closest Big<br />
City, USA is <strong>in</strong>cluded <strong>in</strong> the market. We believe this is exactly what we do by us<strong>in</strong>g<br />
HRRs to def<strong>in</strong>e local markets as these regions are constructed explicitly to <strong>in</strong>clude<br />
the major hospital where people travel for sophisticated treatment.<br />
Therefore, given a measure of costs which accounts for travel <strong>and</strong> the conservative<br />
def<strong>in</strong>ition of local areas, we th<strong>in</strong>k that the exclusion restriction is unlikely to be<br />
violated. Nevertheless, if out-of-HRR costs did somehow directly impact costs, it<br />
would downwardly bias the coefficient on Medgap, work<strong>in</strong>g aga<strong>in</strong>st the direction of<br />
our ma<strong>in</strong> f<strong>in</strong>d<strong>in</strong>g. 19<br />
Medicare Advantage<br />
A f<strong>in</strong>al topic of discussion is our proxy for the price <strong>and</strong> availability of Medicare<br />
Advantage coverage: a second order polynomial <strong>in</strong> the county-year specific MA pene-<br />
tration rate. Controll<strong>in</strong>g for the determ<strong>in</strong>ants of MA enrollment is important for the<br />
explanatory power of the choice equation as Medigap <strong>and</strong> MA are typically viewed as<br />
substitutes. For similar reasons, it is important to <strong>in</strong>clude MA enrollees <strong>in</strong> the sample<br />
of analysis as MA coverage is the relevant alternative coverage to consider for changes<br />
<strong>in</strong> the Medigap market. Yet for valid estimation, the variables used to predict MA<br />
choice must satisfy the exclusion restriction of not directly impact<strong>in</strong>g medical costs.<br />
Below we make the case for their exclusion.<br />
The ma<strong>in</strong> argument for the exogeneity of county-year specific MA penetration<br />
rates is that much of the variation <strong>in</strong> this variable is driven by high-level federal<br />
19 In this case, higher out-of-HRR costs would simultaneously shift <strong>in</strong>dividuals out of coverage<br />
through the <strong>in</strong>strument <strong>and</strong> <strong>in</strong>crease costs directly, creat<strong>in</strong>g a negative correlation between coverage<br />
<strong>and</strong> unobserved costs <strong>and</strong> bias<strong>in</strong>g downwards the coefficient on Medigap.
CHAPTER 3. MEDIGAP 132<br />
policy that could not plausibly be related to the unobserved component <strong>in</strong> the cost<br />
equation. MA penetration decreased over the first part of our sample as policies<br />
enacted <strong>in</strong> the 1997 Balanced Budget Act to reduce Medicare payments <strong>in</strong>directly<br />
decreased MA payments. MA penetration rates were then buoyed up by <strong>in</strong>creased<br />
payments written <strong>in</strong>to the 2003 Medicare Modernization Act (?). In addition, the<br />
goal of MA has shifted over time from promot<strong>in</strong>g efficiency to <strong>in</strong>clud<strong>in</strong>g broader goals<br />
such as regional equity mean<strong>in</strong>g there has been great variation over time <strong>in</strong> federal<br />
policy toward regional MA <strong>in</strong>surers. Thus most of the changes over time were driven<br />
by federal policy that could not be related to the unobservable.<br />
We also statistically test for the endogeneity of the county-year specific MA pen-<br />
etration variables. Recall that our Meidgap choice equation has two sets of excluded<br />
variables: Medigap premiums <strong>and</strong> the MA penetration variables. Based on the as-<br />
sumption that estimates from a specification us<strong>in</strong>g only Medigap premiums are con-<br />
sistent but <strong>in</strong>efficient <strong>and</strong> estimates from a specification with both Medigap premiums<br />
<strong>and</strong> MA penetration are efficient, a Durb<strong>in</strong>-Wu-Hausman test squarely fails to reject<br />
the exogeneity of MA penetration variables. In our preferred specification with the<br />
full set of controls, the test returns a Chi-squared value of 0.8, a value vastly below<br />
the 5 percent critical value of 37.6. Tests run on other specifications yield similar<br />
results.<br />
3.6 Results <strong>and</strong> Discussion<br />
3.6.1 Premiums <strong>and</strong> Medigap Choice<br />
Below we present basel<strong>in</strong>e estimates of the empirical model, start<strong>in</strong>g with the first<br />
stage premium equation, followed by the mult<strong>in</strong>omial logit second stage Medigap<br />
choice equations, <strong>and</strong> f<strong>in</strong>ally the third stage cost equation. Each of these equations is<br />
estimated with three different sets of control variables. A sparse specification <strong>in</strong>clud-<br />
<strong>in</strong>g only key variables <strong>and</strong> year fixed effects; a specification that <strong>in</strong>cludes the demo-<br />
graphic controls; <strong>and</strong> a full specification that, along with the demographic variables,
CHAPTER 3. MEDIGAP 133<br />
<strong>in</strong>cludes controls for health status such as self-reported health, <strong>in</strong>dicators of medi-<br />
cal diagnoses <strong>and</strong> treatment, <strong>and</strong> widely-used measures of functional status known as<br />
Activities of Daily Liv<strong>in</strong>g (ADL) <strong>and</strong> Instrumental Activities of Daily Liv<strong>in</strong>g (IADL).<br />
Table 3.5 presents the first stage estimates of Medigap premiums on state- <strong>and</strong><br />
HRR-level costs <strong>in</strong> logs. St<strong>and</strong>ard errors are clustered at the year-by-state level<br />
because premiums vary at this level. For the purposes of our policy counterfactual<br />
analysis, the most important parameter <strong>in</strong> this table is the elasticity of premiums<br />
with respect to state-level costs. This coefficient is 0.50 to 0.60 across specifications<br />
<strong>and</strong> precisely estimated (st<strong>and</strong>ard error of less than 0.07). As discussed <strong>in</strong> Section 3.5,<br />
most of the identify<strong>in</strong>g variation comes from state-level FFS Medicare costs. With<br />
HRR costs <strong>and</strong> the other controls, the F-statistic on state-level costs is well above<br />
the critical level.<br />
Marg<strong>in</strong>al effects for the probability of Medigap choice from the mult<strong>in</strong>omial logit<br />
choice equation are presented <strong>in</strong> Table 3.6. Recall that residuals from the premium<br />
equation are <strong>in</strong>cluded as a control variable to overcome the potential endogeneity of<br />
premiums. The premium semi-elasticity of dem<strong>and</strong> is estimated at a stable 0.38 to<br />
0.48 across specifications. (The premium elasticity estimates from the l<strong>in</strong>ear proba-<br />
bility model, shown <strong>in</strong> Appendix Table 3.11, are similar, rang<strong>in</strong>g from 0.31 to 0.42.)<br />
With demographic controls, the po<strong>in</strong>t estimates are significantly negative at more<br />
than the 5 percent level. In the preferred logit specification with the full set of con-<br />
trols (specification 3), we can reject any premium elasticity outside of 0.13 to 0.82<br />
with a 95 percent confidence <strong>in</strong>terval.<br />
Males are 3 percent less likely to purchase Medigap coverage. Medigap partici-<br />
pation is 20 percent lower for blacks, <strong>in</strong>creas<strong>in</strong>g <strong>in</strong> education, but flat with respect<br />
to <strong>in</strong>come. Although we do not show it here, it should be noted that there is strong<br />
evidence that Medigap choice is <strong>in</strong>creas<strong>in</strong>g <strong>in</strong> factors the proxy for good health as<br />
well as underly<strong>in</strong>g cognitive ability (?).
CHAPTER 3. MEDIGAP 134<br />
3.6.2 Causal Impact on Costs<br />
Table 3.7 presents estimates of the third stage cost model. Columns 1 through 3<br />
show simple OLS estimates of log total costs on a Medigap <strong>in</strong>dicator <strong>and</strong> controls;<br />
columns 4 through 6 show estimates with predicted values from the mult<strong>in</strong>omial logit<br />
model for Medigap choice used as <strong>in</strong>struments. The models are estimated with block<br />
bootstrap st<strong>and</strong>ard errors clustered by <strong>in</strong>dividual for consistency (?). In Appendix<br />
Table 3.12, we show 3SLS estimates of the cost equation us<strong>in</strong>g the l<strong>in</strong>ear probability<br />
model for Medigap choice.<br />
The OLS estimates (columns 1 to 3) suggest that Medigap coverage <strong>in</strong>creases total<br />
medical utilization by about 40 percent. With the full set of controls, a 95 percent<br />
confidence <strong>in</strong>terval allows us to reject any effect outside of 16 to 65 percent. The<br />
<strong>in</strong>strumental variables estimates are about 40 percent larger than the OLS estimate<br />
across the different sets of controls. With demographic controls (columns 5 <strong>and</strong> 6), the<br />
po<strong>in</strong>t estimates suggest that Medigap coverage causally <strong>in</strong>creases medical utilization<br />
by 57 to 61 percent for <strong>in</strong>dividuals local to the <strong>in</strong>strument. (The 3SLS estimates,<br />
shown <strong>in</strong> Appendix Table 3.12 <strong>in</strong>dicate effects of 59 to 63 percent.) In the logit<br />
specification with the full set of controls, a 95 percent confidence <strong>in</strong>terval allows us<br />
to reject any effect outside of 30 to 84 percent.<br />
We f<strong>in</strong>d it sensible that the <strong>in</strong>strumental variables estimates are larger than those<br />
estimated by OLS (57 versus 41 percent). As we have discussed elsewhere, Fang et<br />
al. (2009) f<strong>in</strong>d strong evidence of advantageous selection on a set of health variables<br />
that <strong>in</strong>crease the R-squared of their basel<strong>in</strong>e model from 7 to 21 percent. Yet even<br />
these controls leave 79 percent of the variation unexpla<strong>in</strong>ed. If there is some selection<br />
on <strong>in</strong>formation conta<strong>in</strong>ed <strong>in</strong> this rema<strong>in</strong><strong>in</strong>g variation, <strong>and</strong> this selection cont<strong>in</strong>ues to<br />
be advantageous, then we would expect an approach that overcomes this selection<br />
problem to yield larger estimates than a simple OLS approach.<br />
A second reason for why our basel<strong>in</strong>e estimates may be larger than the OLS<br />
estimates is because we are measur<strong>in</strong>g the moral hazard effect for <strong>in</strong>dividuals on the<br />
marg<strong>in</strong> of select<strong>in</strong>g <strong>in</strong>to Medigap. The OLS estimates are best <strong>in</strong>terpreted as an<br />
average moral hazard effect that may be biased due to selection. In contrast, our<br />
basel<strong>in</strong>e logit estimates capture the moral hazard effect for <strong>in</strong>dividuals on the marg<strong>in</strong>
CHAPTER 3. MEDIGAP 135<br />
of select<strong>in</strong>g <strong>in</strong>to Medigap at current prices. These <strong>in</strong>dividuals are probably more<br />
price sensitive when it comes to premiums. If they are more sensitive to the price of<br />
medical care <strong>in</strong>duced by Medigap coverage as well, then the moral hazard effect for<br />
these marg<strong>in</strong>al <strong>in</strong>dividuals will be larger than the average Medigap policyholders.<br />
We th<strong>in</strong>k it is important to emphasize two th<strong>in</strong>gs about the <strong>in</strong>terpretation of our<br />
basel<strong>in</strong>e estimates related to this po<strong>in</strong>t. First, the moral hazard of <strong>in</strong>dividuals on the<br />
marg<strong>in</strong> is the precisely the parameter of <strong>in</strong>terested for policy analysis of a Medigap<br />
premium tax or subsidy. If the government levies, for example, a tax on Medigap<br />
premiums, then the <strong>in</strong>dividuals that select out of Medigap as a result of the tax will<br />
reduce their total medical care utilization by our basel<strong>in</strong>e estimate. (In Section 3.7,<br />
we demonstrate the effect of tax <strong>and</strong> subsidy policies us<strong>in</strong>g our estimates.) Second,<br />
one can view our moral hazard estimate as an upper bound on the moral hazard of<br />
all Medigap beneficiaries follow<strong>in</strong>g the logic that marg<strong>in</strong>al beneficiaries have greater<br />
moral hazard than <strong>in</strong>fra-marg<strong>in</strong>al beneficiaries.<br />
To compare our effect to other estimates <strong>in</strong> the literature, we follow st<strong>and</strong>ard<br />
practices (??) <strong>and</strong> convert our po<strong>in</strong>t estimates to arc elasticities of expenditure (?).<br />
Under the approximation that Medigap decreases the “price” of a unit of healthcare<br />
from 0.2 to 0.0, the preferred po<strong>in</strong>t estimate of 0.57 implies an arc elasticity of 1.13.<br />
We th<strong>in</strong>k there are good reasons for why our estimate is higher than the “gold stan-<br />
dard” estimate of 0.22 from the RAND Health Insurance Experiment (HIE) of the<br />
1970s. The first is that our estimate may be local to particularly price sensitive <strong>in</strong>di-<br />
viduals, as discussed above. A second reason is that elderly people may be more price<br />
sensitive than the non-elderly population. While the RAND HIE only <strong>in</strong>cluded people<br />
under age 65, the Medicare population we study <strong>in</strong>cludes those over age 65. Other<br />
papers have found results consistent with the idea that the elderly have higher price<br />
elasticities of medical utilization (??). A third reason is that <strong>in</strong>novation <strong>in</strong> medic<strong>in</strong>e<br />
over the last 40 years could have affected price elasticities, perhaps by creat<strong>in</strong>g more<br />
treatment options for a given diagnosis that <strong>in</strong>dividuals can substitute among.
CHAPTER 3. MEDIGAP 136<br />
3.6.3 Robustness <strong>and</strong> External Validity<br />
Our first robustness check addresses the log specification for medical costs. It is well<br />
known that the skewed distribution of outcomes can create difficulties <strong>in</strong> estimat<strong>in</strong>g<br />
models with medical costs as a dependent variable. To show that the log specification<br />
we use is not driv<strong>in</strong>g the results, Panel A of Table 3.8 shows po<strong>in</strong>t estimates from the<br />
three-stage model where costs <strong>in</strong> the cost equation are denoted <strong>in</strong> levels. For com-<br />
parison, we divide these estimates by the average costs <strong>in</strong> our sample <strong>and</strong> display the<br />
results next to the estimated parameters from the model <strong>in</strong> logs. With demographic<br />
controls, the implied percentage changes <strong>in</strong> medical spend<strong>in</strong>g are similar to the log<br />
estimates (0.68 versus 0.57) although less precisely estimated.<br />
As discussed earlier, our analysis is restricted to the sample of <strong>in</strong>dividuals who<br />
turned 65 no sooner than the year 2000, as price data is not available from before<br />
this year. This, of course, means that the estimates should not be directly ascribed<br />
to the entire Medigap population because of potential age or cohort heterogeneity<br />
<strong>in</strong> the ma<strong>in</strong> effect. However, we can shed some light on the external validity of the<br />
ma<strong>in</strong> results with simple OLS regressions. Panel B of Table 3.8 shows OLS estimates<br />
of log medical costs on a Medigap <strong>in</strong>dicator for the entire sample of <strong>in</strong>dividuals <strong>in</strong><br />
the data <strong>and</strong> the subset we analyze. With demographic <strong>and</strong> health controls, the OLS<br />
estimates from the samples are very similar (0.34 versus 0.37). However, to the extent<br />
that health controls may be partly determ<strong>in</strong>ed by Medigap choice, the specification<br />
with only demographic covariates may be more appropriate. The lower OLS estimate<br />
for the full sample <strong>in</strong> this specification (0.11 versus 0.35) suggests some caution when<br />
extrapolat<strong>in</strong>g the estimates of out sample.<br />
F<strong>in</strong>ally, as discussed Section 3.5, there is <strong>in</strong> fact a small degree of with<strong>in</strong>-state<br />
variation <strong>in</strong> Medigap premiums <strong>in</strong> a small number of states <strong>in</strong> the sample (Maestas et<br />
al., 2009). To exam<strong>in</strong>e the robustness of the ma<strong>in</strong> results to this variation, we restrict<br />
our sample to states where, accord<strong>in</strong>g to Maestas et al., the with<strong>in</strong>-state coefficient of<br />
variation of Medigap premiums is less than 0.10 for 75 percent of firms issu<strong>in</strong>g plans. 20<br />
Panel C of Table 3.8 shows full model <strong>and</strong> OLS estimates of the coefficient on Medigap<br />
20 The states we drop are CA, FL, IL, LA, MI, NV, NY, PA, TX, <strong>and</strong> VA.
CHAPTER 3. MEDIGAP 137<br />
from the basel<strong>in</strong>e <strong>and</strong> restricted sample. The full model estimates are slightly lower<br />
(0.42 versus 0.57) although the ratio of full model to OLS estimates is very similar<br />
across samples. A possible reason for the lower full model <strong>and</strong> OLS estimates <strong>in</strong> the<br />
restricted sample is that the excluded states, which <strong>in</strong>clude states notorious for the<br />
overuse of medical services (e.g., FL, NY, TX), are states where the moral hazard from<br />
Medigap may be the largest. The premium semi-elasticity of dem<strong>and</strong> estimates, which<br />
are potentially the most susceptible to bias from with<strong>in</strong>-state premium variation, are<br />
almost identical across the restricted <strong>and</strong> basel<strong>in</strong>e samples.<br />
3.7 Policy Counterfactuals<br />
Hav<strong>in</strong>g shown that private Medigap <strong>in</strong>surance <strong>in</strong>creases medical utilization <strong>and</strong> thus<br />
Medicare costs by as much as 59 percent, it seems natural to consider corrective tax-<br />
ation. The idea of tax<strong>in</strong>g Medigap premiums is not new. In their 2008 chapter of<br />
budget options for health care, for example, the Congressional Budget Office (CBO)<br />
considered a 5 percent excise tax on Medigap premiums as a mechanism to reduce<br />
the federal deficit. As CBO po<strong>in</strong>ted out, an excise tax on Medigap premiums would<br />
reduce the deficit through two ma<strong>in</strong> channels: it would raise revenue from <strong>in</strong>dividuals<br />
who ma<strong>in</strong>ta<strong>in</strong>ed their Medigap coverage while at the same time reduc<strong>in</strong>g Medicare<br />
costs for people who gave up their supplemental coverage. Yet the CBOs calculations<br />
of the taxs impact were likely based on a downwardly biased estimate of the effect of<br />
supplemental <strong>in</strong>surance, mean<strong>in</strong>g the CBO estimates of the effect of a tax underes-<br />
timated the budgetary effect. 21 In contrast, the parameters estimated here leverage<br />
plausibly exogenous variation <strong>in</strong> Medigap premiums, account<strong>in</strong>g not only for selec-<br />
tion on unobservable <strong>in</strong>dividual characteristics but also for potentially heterogeneous<br />
effects on <strong>in</strong>dividuals marg<strong>in</strong>al to a change <strong>in</strong> premiums.<br />
We use our estimated parameters to produce back-of-the-envelope budgetary es-<br />
timates <strong>in</strong> the follow manner: First, the counterfactual Medigap market share is<br />
21 In another section (Option 82) of their budget options volume, CBO cites studies claim<strong>in</strong>g<br />
that Medigap coverage <strong>in</strong>creases utilization by 25 percent versus <strong>in</strong>dividuals without supplementary<br />
coverage.
CHAPTER 3. MEDIGAP 138<br />
calculated us<strong>in</strong>g the estimated premium semi-elasticity of 0.475 (Table 3.6, specifi-<br />
cation 3) <strong>and</strong> assum<strong>in</strong>g the tax is completely passed through. 22 We estimate tax<br />
revenue by apply<strong>in</strong>g the tax to the result<strong>in</strong>g share of Medigap enrollees. Cost sav<strong>in</strong>gs<br />
are calculated by reduc<strong>in</strong>g the medical costs of <strong>in</strong>dividuals who drop Medigap by 37.0<br />
percent, the factor implied by the estimated parameter of 0.588 from specification 3<br />
of Table 3.7 (i.e., 0.370 =1 - 1/(1+0.588)). Medicare cost sav<strong>in</strong>gs are calculated as<br />
80 percent of this number, consistent with Medicare pay<strong>in</strong>g 80 percent on the marg<strong>in</strong><br />
for care. The total budgetary impact is simply the sum of the revenue <strong>and</strong> Medicare<br />
cost effects.<br />
Table 3.9 presents these revenue <strong>and</strong> cost sav<strong>in</strong>gs estimates for the st<strong>and</strong>ard sample<br />
of <strong>in</strong>dividuals with Medigap, Medicare Advantage, <strong>and</strong> no supplemental coverage.<br />
In this sample, basel<strong>in</strong>e Medigap enrollment is 35.8 percent, <strong>and</strong> Medical costs for<br />
Medigap enrollees average $9,458 per annum. The results are presented <strong>in</strong> dollars per<br />
beneficiary <strong>and</strong> as a percentage of Medicare costs of $5,528 on average for Medigap<br />
enrollees. A 20 percent tax would reduce coverage by 9.5 percentage po<strong>in</strong>ts, rais<strong>in</strong>g<br />
1.3 percent of total costs <strong>in</strong> tax revenue <strong>and</strong> 4.8 percent of total costs <strong>in</strong> reduced<br />
utilization for total budgetary sav<strong>in</strong>gs of 6.2 percent. A 40 <strong>and</strong> 60 percent tax would<br />
save 11.4 <strong>and</strong> 15.6 percent of costs respectively. And a 76 percent tax, which would<br />
<strong>in</strong>crease Medigap premiums by $1,076, or about 50.5 percent of the implied $2,802<br />
(=0.37 x $9,458 x 0.80) externality, would reduce the Medigap market share to zero,<br />
sav<strong>in</strong>gs 18.1 percent of total costs entirely though decreased medical utilization.<br />
Although it probably goes without say<strong>in</strong>g, we rem<strong>in</strong>d the reader that the pa-<br />
rameters are estimated us<strong>in</strong>g local variation <strong>in</strong> premiums <strong>and</strong> the projected effects of<br />
larger taxes <strong>and</strong> subsidies should be viewed with appropriate caution. It is also worth<br />
mention<strong>in</strong>g that under current law, a tax that reduces FFS Medicare costs would re-<br />
duce m<strong>in</strong>imum Medicare Advantage payments. This is not taken <strong>in</strong>to account <strong>in</strong> our<br />
simple tax simulations, <strong>and</strong> would likely magnify the cost sav<strong>in</strong>gs we project.<br />
Last but not least, we want to call attention to the fact that a uniform tax on<br />
Medigap premiums is a blunt <strong>in</strong>strument for controll<strong>in</strong>g Medicare costs. Ch<strong>and</strong>ra et<br />
22 In the case of <strong>in</strong>complete pass-though, one can th<strong>in</strong>k of the counterfactuals as simulat<strong>in</strong>g the<br />
impact of tax that would give rise to the full pass-though <strong>in</strong>crease <strong>in</strong> premiums.
CHAPTER 3. MEDIGAP 139<br />
al. (2007) have shown that targeted subsidies that reduce cost-shar<strong>in</strong>g for chronically<br />
ill Medicare beneficiaries can generate cost sav<strong>in</strong>gs or “offsets” by encourag<strong>in</strong>g treat-<br />
ment adherence <strong>and</strong> reduc<strong>in</strong>g unnecessary hospitalizations. A nuanced tax policy,<br />
which takes such effects <strong>in</strong>to account, might be able to generate both <strong>in</strong>creased costs<br />
sav<strong>in</strong>gs <strong>and</strong> improved health.<br />
3.8 Conclusion<br />
In this paper, we have estimated the fiscal externality imposed by private Medigap<br />
supplemental <strong>in</strong>surance on <strong>public</strong> Medicare costs. We f<strong>in</strong>d that the moral hazard<br />
<strong>in</strong>duced by Medigap is substantial, with marg<strong>in</strong>al enrollees us<strong>in</strong>g 57 percent more<br />
services due to Medigap coverage. Policy counterfactuals reveal that a uniform tax<br />
that raises premiums 20 percent could generate comb<strong>in</strong>ed revenue <strong>and</strong> sav<strong>in</strong>gs of 6.2 of<br />
basel<strong>in</strong>e costs; a Pigovian tax that fully accounts for the fiscal externality could yield<br />
sav<strong>in</strong>gs of 18.1 percent. We are <strong>in</strong>terested <strong>in</strong> the effects of more nuanced tax policies<br />
that account for “offsets” from treatment adherence or other positive externalities. In<br />
light of the f<strong>in</strong>d<strong>in</strong>gs of significant moral hazard from Medigap, we view the study of<br />
supplemental <strong>in</strong>surance around the world as an important topic for future research.
CHAPTER 3. MEDIGAP 140<br />
(10000,11000]<br />
(9000,10000]<br />
(8000,9000]<br />
(7000,8000]<br />
(6000,7000]<br />
[5000,6000]<br />
Figure 3.1: Example of Identify<strong>in</strong>g Variation<br />
(10000,11000]<br />
(9000,10000]<br />
(8000,9000]<br />
(7000,8000]<br />
(6000,7000]<br />
[5000,6000] <strong>and</strong> VT per capita Medicare costs<br />
(10000,11000] NY (9000,10000] <strong>and</strong> (8000,9000] VT (7000,8000] per (6000,7000] capita [5000,6000]<br />
Medicare NY costs<br />
NY <strong>and</strong> VT per capita Medicare costs<br />
(1600,1700]<br />
(1500,1600]<br />
(1400,1500]<br />
[1300,1400] NY <strong>and</strong> VT premiums<br />
(1600,1700]<br />
NY <strong>and</strong> VT premiums<br />
(1600,1700]<br />
(1500,1600]<br />
(1400,1500]<br />
[1300,1400]<br />
NY <strong>and</strong> VT premiums
CHAPTER 3. MEDIGAP 141<br />
Table 3.1: Medicare Cost-Shar<strong>in</strong>g<br />
Part A: Hospital Expenditures Part B: Physician Expenditures<br />
Per day copayment<br />
SNF<br />
Per day copayment<br />
Year Deductible Days 61-90 Days 91-150* Deductible Co<strong>in</strong>surance Days 21-100<br />
2000 $776 $194 $388 $100 20% $97<br />
2005 $912 $228 $456 $110 20% $114<br />
*Medicare only pays for part A hosptializations <strong>in</strong> excess of 90 days through the drawdown of lifetime reserve days, of<br />
which beneficiaries have 60 over their lifetime.
CHAPTER 3. MEDIGAP 142<br />
Table 3.2: Summary Statistics by Supplemental Insurance Type<br />
ESI Medicaid Medigap MA None Total<br />
Enrollment<br />
N 16,477 6,594 11,830 13,045 4,475 52,420<br />
Percent<br />
Demographics<br />
31% 13% 23% 25% 9% 100%<br />
Male 43.1% 27.5% 33.8% 39.4% 42.0% 38.0%<br />
(0.4%) (0.5%) (0.4%) (0.4%) (0.7%) (0.2%)<br />
Age 75.0 77.9 76.4 74.7 77.1 75.8<br />
(0.1) (0.1) (0.1) (0.1) (0.1) (0.0)<br />
White 91.8% 71.1% 93.9% 85.5% 81.3% 87.2%<br />
(0.2%) (0.5%) (0.2%) (0.3%) (0.6%) (0.1%)<br />
College or more 22.6% 3.8% 15.1% 15.1% 10.5% 15.7%<br />
(0.3%) (0.2%) (0.3%) (0.3%) (0.4%) (0.2%)<br />
Income 37,818 10,839 30,215 30,691 20,887 29,490<br />
Utilization<br />
(371) (309) (439) (641) (431) (227)<br />
Total 11,008 25,456 10,152 8,062 18,169 12,511<br />
(147) (366) (167) (141) (409) (95)<br />
Medicare* 5,971 11,871 6,214 2,111 8,107 5,990<br />
(106) (258) (124) (79) (260) (64)<br />
ESI 2,907 113 14 182 132 988<br />
(37) (13) (2) (12) (17) (13)<br />
Medicaid 1 9,359 0 405 8 1,279<br />
(1) (176) (0) (36) (3) (29)<br />
Medigap 103 149 1,472 73 89 409<br />
(9) (11) (40) (5) (11) (10)<br />
Out-of-pocket 1,669 3,164 1,967 1,691 8,425 2,506<br />
(40) (72) (28) (43) (227) (29)<br />
Other 357 800 485 3,600 1,407 1,338<br />
(16) (51) (20) (85) (77) (24)<br />
Source: 2000-2005 MCBS<br />
Notes: ESI is employer sponsored supplmental <strong>in</strong>surance. MA is Medicare Advantage or Medicare+Choice as it<br />
was called <strong>in</strong> the beg<strong>in</strong><strong>in</strong>g of the sample period. Sample excludes <strong>in</strong>dividuals with positive VA expenditure <strong>and</strong><br />
<strong>in</strong>dividuals resid<strong>in</strong>g <strong>in</strong> states without Medigap coverage (MA, MN, WI). Payments from unexpected sources (e.g,<br />
Medigaid payments for ESI enrollees) due to switches <strong>in</strong> supplemental <strong>in</strong>surance type dur<strong>in</strong>g the year. St<strong>and</strong>ard<br />
errors <strong>in</strong> parentheses.<br />
*Does not account for Medicare capation payments to private plans
CHAPTER 3. MEDIGAP 143<br />
Table 3.3: Medigap Enrollment <strong>and</strong> Plan Characteristics by Letter<br />
Medigap Letter<br />
A B C D E F* G H I J*<br />
Enrollment<br />
N 238 291 894 293 133 1,757 154 117 112 348<br />
Percent 6% 7% 21% 7% 3% 41% 4% 3% 3% 8%<br />
Plan Characteristics<br />
X X X X X X X X X X<br />
Basic benefits (Part A con<strong>in</strong>surance,<br />
Part B co<strong>in</strong>surance, blood, additional<br />
lifetime hospital days)<br />
Skilled nurs<strong>in</strong>g facility co<strong>in</strong>surance X X X X X X X X X<br />
Part A deductible X X X X X X X X X X<br />
Part B deductible X X X<br />
Part B excess charges X 80%<br />
Foreign Travel Emergecy X X X X X X X X X<br />
Home health care X X X X<br />
Prescription drugs X X X<br />
Preventive medical care X X<br />
Source: 2000-2005 MCBS<br />
Notes: Sample excludes <strong>in</strong>dividuals with positive VA expenditure <strong>and</strong> <strong>in</strong>dividuals resid<strong>in</strong>g <strong>in</strong> states without Medigap coverage (MA, MN, WI).<br />
*Plans F <strong>and</strong> J have a high-deductible option that requires beneficiaries to pay $1,580 before receiv<strong>in</strong>g Medigap coverage. These plans are rarely offered an<br />
have only 14 enrollees <strong>in</strong> the sample.
CHAPTER 3. MEDIGAP 144<br />
Table 3.4: HRR Level-Medicare Costs Tabulated By State<br />
State N Mean Std. Dev. M<strong>in</strong>. Max.<br />
NY 10 7,453 2,389 5,235 11,588<br />
NV 2 7,283 1,730 6,059 8,506<br />
FL 18 7,271 1,209 6,091 11,422<br />
MS 6 6,796 1,176 5,310 8,473<br />
CA 24 7,094 1,080 5,432 10,290<br />
TX 22 7,258 1,078 5,472 10,626<br />
IL 13 6,817 1,061 5,097 8,496<br />
MI 15 7,033 1,059 5,683 8,636<br />
PA 14 6,707 1,044 5,017 8,845<br />
LA 10 7,922 772 7,148 9,020<br />
UT 3 5,990 747 5,335 6,804<br />
IN 9 6,198 728 5,123 7,458<br />
NJ 7 8,626 713 7,216 9,426<br />
AR 5 6,162 691 4,983 6,709<br />
MA 3 8,123 621 7,431 8,631<br />
WI 8 5,355 620 4,587 6,304<br />
GA 7 6,160 534 5,068 6,852<br />
CO 7 6,112 516 5,211 6,767<br />
NH 2 5,785 487 5,440 6,129<br />
OR 5 5,049 466 4,520 5,547<br />
NC 9 6,165 456 5,563 6,955<br />
OH 10 6,932 455 6,385 7,674<br />
MO 6 6,106 445 5,541 6,634<br />
IA 8 5,063 420 4,690 5,814<br />
SC 5 6,430 404 6,043 7,065<br />
VA 8 5,770 374 5,275 6,312<br />
CT 3 7,864 369 7,527 8,258<br />
ND 4 5,315 356 4,902 5,677<br />
WA 6 5,716 352 5,170 6,108<br />
KY 5 6,439 331 6,088 6,807<br />
TN 7 6,641 307 6,374 7,026<br />
MN 5 5,544 298 5,193 5,876<br />
AZ 4 6,410 280 6,036 6,714<br />
AL 6 6,710 278 6,390 6,994<br />
ID 2 5,114 269 4,924 5,304<br />
MD 3 7,734 261 7,568 8,034<br />
ME 2 5,990 260 5,806 6,173<br />
OK 3 6,932 193 6,770 7,146<br />
SD 2 5,243 153 5,134 5,351<br />
KS 2 6,294 125 6,205 6,382<br />
NE 2 5,721 87 5,659 5,782<br />
MT 3 5,383 69 5,304 5,430<br />
WV 3 6,475 58 6,433 6,541<br />
AK 1 6,719 6,719 6,719<br />
DC 1 6,961 6,961 6,961<br />
DE 1 6,850 6,850 6,850<br />
HI 1 4,778 4,778 4,778<br />
NM 1 5,254 5,254 5,254<br />
RI 1 7,187 7,187 7,187<br />
VT 1 5,858 5,858 5,858<br />
WY 1 5,989 5,989 5,989<br />
Total 306 6,635 1,163 4,520 11,588<br />
Source: Dartmouth Atlas of Health Care<br />
Notes: Each observation is per capita FFS Medicare costs <strong>in</strong> a given HRR <strong>in</strong> 2003. The<br />
data is sorted by st<strong>and</strong>ard deviation.
CHAPTER 3. MEDIGAP 145<br />
Table 3.5: First Stage: OLS Estimates of Premiums on State <strong>and</strong> HRR Costs<br />
(1)<br />
Dependent Variable: Log Premium<br />
(2) (3)<br />
Est. SE Est. SE Est. SE<br />
Log state cost 0.5020 (0.0583) 0.5990 (0.0686) 0.5950 (0.0678)<br />
MA penetration rate (%) 0.0004 (0.0018) 0.0000 (0.0018) 0.0001 (0.0019)<br />
MA penetration rate squared (%) 0.0000 (0.0000) 0.0000 (0.0000) 0.0000 (0.0000)<br />
Log HRR cost -0.0005 (0.0266) -0.0020 (0.0252) 0.0003 (0.0248)<br />
Male -0.0052 (0.0060) 0.0045 (0.0084)<br />
Age-65 0.0259 (0.0115) 0.0231 (0.0112)<br />
(Age-65)^2 0.0037 (0.0062) 0.0053 (0.0061)<br />
(Age-65)^3 -0.0007 (0.0009) -0.0010 (0.0008)<br />
Disabled eligibility<br />
Race group<br />
-0.0016 (0.0058) -0.0004 (0.0059)<br />
Asian 0.0149 (0.0141) 0.0121 (0.0134)<br />
Black 0.0097 (0.0118) 0.0114 (0.0116)<br />
Other<br />
Education group<br />
-0.0238 (0.0154) -0.0249 (0.0148)<br />
High school -0.0243 (0.0051) -0.0243 (0.0050)<br />
Less than college -0.0052 (0.0065) -0.0048 (0.0063)<br />
College or more -0.0094 (0.0073) -0.0113 (0.0070)<br />
Log <strong>in</strong>come 0.0007 (0.0028) 0.0011 (0.0028)<br />
Married -0.0104 (0.0050) -0.0102 (0.0053)<br />
Work<strong>in</strong>g 0.0026 (0.0054) 0.0015 (0.0055)<br />
Served <strong>in</strong> armed forces<br />
Self-reported health<br />
0.0136 (0.0064) 0.0122 (0.0067)<br />
Excellent -0.0042 (0.0119)<br />
Very good -0.0134 (0.0110)<br />
Good -0.0137 (0.0095)<br />
Fair -0.0209 (0.0096)<br />
Diagnosis <strong>in</strong>dicators Yes<br />
Treatment <strong>in</strong>dicators Yes<br />
IADL <strong>in</strong>dicators Yes<br />
Year FE Yes Yes Yes<br />
R-squared 25.1% 31.2% 33.2%<br />
N 3,556 3,556 3,556<br />
F-statistic on log state cost 74.01 76.15 76.88<br />
Prob > F 0.00 0.00 0.00<br />
Source: 2000-2005 MCBS; Weiss Rat<strong>in</strong>gs; Dartmouth Atlas<br />
Notes: Parameter estimates from OLS regressions of premiums on state- <strong>and</strong> HRR-level Medicare costs. The sample is restricted to<br />
<strong>in</strong>dividuals with Medigap, HMO or no supplemental coverage who turned 65 <strong>in</strong> 2000 or later. The omitted race group is white,<br />
education group is less than high school, <strong>and</strong> health status is poor. St<strong>and</strong>ard errors clustered at the state-by-year level <strong>in</strong> parentheses.
CHAPTER 3. MEDIGAP 146<br />
Table 3.6: Second Stage: Marg<strong>in</strong>al Effects for Medigap Choice from Mult<strong>in</strong>omial<br />
Logit Model<br />
(1)<br />
Marg<strong>in</strong>al Effects for Medigap Choice<br />
(2) (3)<br />
Mfx. SE Mfx. SE Mfx. SE<br />
Log premium -0.375 (0.214) -0.477 (0.220) -0.473 (0.221)<br />
Log premium residuals 0.398 (0.227) 0.532 (0.242) 0.531 (0.242)<br />
County MA penetration (%) -0.026 (0.002) -0.029 (0.002) -0.029 (0.002)<br />
County MA penetration squared (%) 0.000 (0.000) 0.000 (0.000) 0.000 (0.000)<br />
Log HRR cost -0.049 (0.079) 0.002 (0.090) -0.009 (0.092)<br />
Male -0.028 (0.028) -0.029 (0.028)<br />
Age-65 0.061 (0.043) 0.065 (0.043)<br />
(Age-65)^2 -0.037 (0.020) -0.040 (0.020)<br />
(Age-65)^3 0.005 (0.003) 0.006 (0.003)<br />
Disabled eligibility<br />
Race group<br />
-0.010 (0.028) -0.013 (0.028)<br />
Asian 0.000 (0.068) 0.000 (0.069)<br />
Black -0.263 (0.026) -0.265 (0.026)<br />
Other<br />
Education group<br />
-0.025 (0.058) -0.028 (0.059)<br />
High school 0.038 (0.030) 0.041 (0.031)<br />
Less than college 0.126 (0.036) 0.130 (0.036)<br />
College or more 0.185 (0.039) 0.188 (0.042)<br />
Log <strong>in</strong>come -0.018 (0.013) -0.017 (0.014)<br />
Married 0.004 (0.024) 0.004 (0.024)<br />
Work<strong>in</strong>g 0.014 (0.029) 0.020 (0.029)<br />
Served <strong>in</strong> armed forces<br />
Self-reported health<br />
-0.033 (0.037) -0.033 (0.038)<br />
Excellent -0.094 (0.051)<br />
Very good -0.113 (0.052)<br />
Good -0.097 (0.049)<br />
Fair<br />
Diagnosis <strong>in</strong>dicators<br />
Treatment <strong>in</strong>dicators<br />
IADL <strong>in</strong>dicators<br />
-0.097 (0.050)<br />
Year FE Yes Yes<br />
Log pseudolikelihood/R-squared -15,205,864 -13,914,167 -13,853,769<br />
N 3,556 3,556 3,556<br />
Source: 2000-2005 MCBS; Weiss Rat<strong>in</strong>gs; Dartmouth Atlas, CMS State-County Penetration Files<br />
Notes: Marg<strong>in</strong>al effects for Medigap choice from mult<strong>in</strong>omial logit model with control function for Medigap premiums.<br />
Marg<strong>in</strong>al effects calculated at the sample means. For b<strong>in</strong>ary variables the predicted change <strong>in</strong> probability for a 0 to 1 change<br />
is shown. The sample is restricted to <strong>in</strong>dividuals with Medigap, HMO or no supplemental coverage who turned 65 <strong>in</strong> 2000 or<br />
later. The omitted race group is white, education group is less than high school, <strong>and</strong> health status is poor. Bootstrap<br />
st<strong>and</strong>ard errors looped over the premium <strong>and</strong> choice equations <strong>and</strong> clustered by <strong>in</strong>dividual <strong>in</strong> parentheses.
CHAPTER 3. MEDIGAP 147<br />
Table 3.7: Third Stage: OLS <strong>and</strong> Full Model Estimates of Medical Costs on Medigap<br />
Indicator<br />
Dependent Variable: Log Medical Costs<br />
OLS Full Model<br />
(1) (2) (3) (4) (5) (6)<br />
Medigap 0.397 0.401 0.407 0.515 0.614 0.567<br />
(0.054) (0.053) (0.049) (0.158) (0.146) (0.138)<br />
Log HRR cost -0.071 0.233 0.128 -0.018 0.325 0.192<br />
(0.132) (0.134) (0.124) (0.173) (0.171) (0.164)<br />
Male -0.055 -0.071 -0.043 -0.049<br />
(0.061) (0.079) (0.074) (0.070)<br />
Age-65 0.817 0.904 0.810 0.890<br />
(0.125) (0.114) (0.126) (0.118)<br />
(Age-65)^2 -0.241 -0.299 -0.235 -0.278<br />
(0.072) (0.065) (0.065) (0.061)<br />
(Age-65)^3 0.023 0.030 0.021 0.027<br />
(0.010) (0.009) (0.009) (0.008)<br />
Disabled eligibility 0.419 0.333 0.422 0.346<br />
Race group<br />
(0.073) (0.067) (0.078) (0.074)<br />
Asian -0.314 -0.206 -0.292 -0.289<br />
(0.143) (0.136) (0.176) (0.175)<br />
Black -0.098 -0.171 -0.050 -0.148<br />
(0.090) (0.083) (0.113) (0.107)<br />
Other 0.034 -0.079 0.052 -0.057<br />
Education group<br />
(0.141) (0.131) (0.168) (0.155)<br />
High school 0.114 0.199 0.098 0.216<br />
(0.069) (0.064) (0.086) (0.082)<br />
Less than college 0.118 0.236 0.096 0.292<br />
(0.073) (0.069) (0.087) (0.083)<br />
College or more 0.184 0.347 0.156 0.404<br />
(0.084) (0.080) (0.103) (0.098)<br />
Log <strong>in</strong>come -0.012 0.049 -0.006 0.037<br />
(0.029) (0.026) (0.033) (0.030)<br />
Married -0.081 0.011 -0.090 -0.058<br />
(0.055) (0.052) (0.067) (0.062)<br />
Work<strong>in</strong>g -0.429 -0.168 -0.434 -0.250<br />
(0.060) (0.056) (0.066) (0.066)<br />
Served <strong>in</strong> armed forces -0.107 -0.119 -0.101 -0.091<br />
Self-reported health<br />
(0.080) (0.073) (0.098) (0.091)<br />
Excellent -0.616 -1.920<br />
(0.140) (0.138)<br />
Very good -0.565 -1.769<br />
(0.136) (0.133)<br />
Good -0.407 -1.392<br />
(0.129) (0.130)<br />
Fair -0.211 -0.890<br />
(0.130) (0.140)<br />
Diagnosis <strong>in</strong>dicators Yes Yes<br />
Treatment <strong>in</strong>dicators Yes Yes<br />
IADL <strong>in</strong>dicators Yes Yes<br />
Year FE Yes Yes Yes Yes Yes Yes<br />
R-squared 3.6% 11.1% 29.4% 3.4% 10.7% 19.9%<br />
N 3,415 3,415 3,415 3,415 3,415 3,415<br />
Source: 2000-2005 MCBS; Weiss Rat<strong>in</strong>gs<br />
Notes: Parameter estimates from cost equation. Specifications 1-3 show OLS estimates; 4-6 show estimates where<br />
Medigap choice is <strong>in</strong>strumed for with the predicted values from the logit cost equation. See text for details. Sample of<br />
<strong>in</strong>dividuals with HMO, Medigap or no supplemental coverage who turned 65 no sooner than 2000 exclud<strong>in</strong>g those with<br />
positive VA expenditure. Omitted race group is white, education group is less than high school, <strong>and</strong> health status is poor.<br />
Bootstrap st<strong>and</strong>ard errors looped over the premium <strong>and</strong> choice equations <strong>and</strong> clustered by <strong>in</strong>dividual <strong>in</strong> parentheses.<br />
H_diabetes
CHAPTER 3. MEDIGAP 148<br />
Table 3.8: Robustness Checks: Key Parameter Estimates from Alternative Specifications<br />
<strong>and</strong> Samples<br />
(1)<br />
Panel A: Levels versus Logs<br />
(2) (3)<br />
Levels<br />
Full model coefficient on Medigap <strong>in</strong> Levels 5829 (1850) 5705 (1572) 4978 (1520)<br />
Average Cost 7344 (339) 7344 (339) 7344 (339)<br />
Percent <strong>in</strong>crease<br />
Logs<br />
0.794 (0.252) 0.777 (0.214) 0.678 (0.207)<br />
Full model coefficient on Medigap <strong>in</strong> Logs 0.515 (0.158) 0.614 (0.146) 0.567 (0.138)<br />
Panel B: Full versus Analyzed Sample<br />
Full sample<br />
OLS Coefficient on Medigap<br />
Analyzed sample<br />
0.138 (0.020) 0.105 (0.020) 0.343 (0.017)<br />
OLS Coefficent on Medigap 0.379 (0.059) 0.352 (0.058) 0.370 (0.053)<br />
Panel C: Lowest Premium Variation States versus Analyzed Sample<br />
Lowest premums variation states<br />
Full model coeffiicient on Medigap 0.418 (0.207) 0.436 (0.188) 0.424 (0.178)<br />
OLS coefficient on Medigap 0.319 (0.083) 0.310 (0.081) 0.294 (0.077)<br />
Premium semi-elasticity of dem<strong>and</strong> for Medigap<br />
Analyzed sample<br />
-0.759 (1.589) -0.441 (1.546) -0.591 (1.525)<br />
Full model coeffiicient on Medigap 0.515 (0.158) 0.614 (0.146) 0.567 (0.138)<br />
OLS coefficient on Medigap 0.379 (0.059) 0.352 (0.058) 0.370 (0.053)<br />
Premium semi-elasticity of Medigap dem<strong>and</strong><br />
Controls <strong>in</strong> all panels<br />
-0.314 (0.187) -0.406 (0.160) -0.419 (0.158)<br />
Basic Controls Yes Yes Yes Yes Yes Yes<br />
Demographic Controls Yes Yes Yes Yes<br />
Health Controls Yes Yes<br />
Source: 2000-2005 MCBS; Weiss Rat<strong>in</strong>gs<br />
Notes: Parameter estimates for the effect of Medigap on total medical costs from various specifications. The sample is restricted to<br />
<strong>in</strong>dividuals with Medigap, HMO. Panel A shows estimates from the full three-stage model with total medical costs <strong>in</strong> levels <strong>and</strong> logs.Panel B<br />
presents estimates on the full sample <strong>and</strong> the analyzed sample restricted to <strong>in</strong>dividuals that turn 65 <strong>in</strong> 2000 or later. Panel C shows<br />
estimates of the ma<strong>in</strong> paramters from the subset of states with virtually no with<strong>in</strong>-state premium varition <strong>and</strong> the analyzed sample. The<br />
controls <strong>in</strong> specifications 1-3 match those <strong>in</strong> Table 7 specifications 1-3. Bootstrap st<strong>and</strong>ard errors clustered by <strong>in</strong>dividual <strong>in</strong> parentheses.
CHAPTER 3. MEDIGAP 149<br />
Table 3.9: Policy Counterfactuals for Subsidies <strong>and</strong> Taxes of Medigap Premiums<br />
Tax Delta Medigap<br />
Tax revenue<br />
Medicare costs Budgetary effect<br />
% of premiums ppt $/beneficiary % $/beneficiary % $/beneficiary %<br />
-40.0% 19.0% -310 -5.6% 532 9.6% 843 15.2%<br />
-20.0% 9.5% -128 -2.3% 266 4.8% 394 7.1%<br />
20.0% -9.5% 74 1.3% -266 -4.8% -341 -6.2%<br />
40.0% -19.0% 95 1.7% -532 -9.6% -627 -11.4%<br />
60.0% -28.5% 62 1.1% -798 -14.4% -860 -15.6%<br />
76.0% -35.8% 0 0.0% -1,003 -18.1% -1,003 -18.1%<br />
Notes: Counterfactuals based on premium elasticity <strong>and</strong> Medigap estimates with full set of controls from Tables 6 <strong>and</strong> 7. Basel<strong>in</strong>e Medigap market<br />
share is 35.8 percent <strong>and</strong> premiums average $1,443. Medicare costs effect caculated assum<strong>in</strong>g Medicare pays for 80 percent of care on the marg<strong>in</strong>.<br />
Percent of costs calculated us<strong>in</strong>g Medicare costs of $5,528 on average for Medigap enrollees.
CHAPTER 3. MEDIGAP 150<br />
Table 3.10: OLS Estimates of Premiums on HRR- <strong>and</strong> State-Level Medicare Costs<br />
Dependent Variable: Log premium<br />
(1) (2)<br />
Log HHR costs 0.35 0.00<br />
(0.04) (0.04)<br />
Log state costs 0.63<br />
(0.04)<br />
R-squared 0.19 0.50<br />
N 290 290<br />
F-stat on log state costs 251.04<br />
Prob > F 0.00<br />
Source: Premium data from Weiss Rat<strong>in</strong>gs; Medicare expenditure data<br />
from the Dartmouth Atlas of Health Care<br />
Notes: There is one observation for each HRR where Medigap is offered.<br />
All variables are de-meaned. Premiums are for Plan C <strong>and</strong> all data is from<br />
2003. Robust st<strong>and</strong>ard errors are parentheses.
CHAPTER 3. MEDIGAP 151<br />
Table 3.11: Second Stage: 2SLS Estimates of Medigap Choice from L<strong>in</strong>ear Probability<br />
Model<br />
(1)<br />
Dependard Variable: Medigap Indicator<br />
(2) (3)<br />
Est. SE Est. SE Est. SE<br />
Log premium -0.314 (0.187) -0.406 (0.160) -0.419 (0.158)<br />
County MA penetration (%) -0.026 (0.002) -0.028 (0.002) -0.028 (0.002)<br />
County MA penetration squared (%) 0.000 (0.000) 0.000 (0.000) 0.000 (0.000)<br />
Log HRR cost 0.002 (0.063) 0.104 (0.064) 0.091 (0.063)<br />
Male -0.050 (0.021) -0.117 (0.030)<br />
Age-65 0.046 (0.033) 0.038 (0.033)<br />
(Age-65)^2 -0.025 (0.017) -0.023 (0.017)<br />
(Age-65)^3 0.004 (0.002) 0.004 (0.002)<br />
Disabled eligibility<br />
Race group<br />
-0.019 (0.021) -0.015 (0.021)<br />
Asian 0.012 (0.049) 0.031 (0.053)<br />
Black -0.208 (0.024) -0.204 (0.025)<br />
Other<br />
Education group<br />
-0.040 (0.044) -0.050 (0.044)<br />
High school 0.083 (0.024) 0.085 (0.024)<br />
Less than college 0.153 (0.025) 0.156 (0.026)<br />
College or more 0.192 (0.032) 0.190 (0.033)<br />
Log <strong>in</strong>come 0.003 (0.010) 0.001 (0.009)<br />
Married 0.023 (0.019) 0.023 (0.020)<br />
Work<strong>in</strong>g 0.011 (0.022) 0.013 (0.022)<br />
Served <strong>in</strong> armed forces<br />
Self-reported health<br />
-0.020 (0.029) -0.022 (0.029)<br />
Excellent -0.108 (0.049)<br />
Very good -0.141 (0.048)<br />
Good -0.112 (0.045)<br />
Fair -0.092 (0.046)<br />
Diagnosis <strong>in</strong>dicators Yes<br />
Treatment <strong>in</strong>dicators Yes<br />
IADL <strong>in</strong>dicators Yes<br />
Year FE Yes Yes Yes<br />
Log pseudolikelihood/R-squared 16.4% 20.4% 22.8%<br />
N 3,556 3,556 3,556<br />
Source: 2000-2005 MCBS; Weiss Rat<strong>in</strong>gs<br />
Notes: Parameter estimates from 2SLS estimates of l<strong>in</strong>ear probability model for Medigap choice. The sample is restricted<br />
to <strong>in</strong>dividuals with Medigap, HMO or no supplemental coverage who turned 65 <strong>in</strong> 2000 or later. The omitted race group is<br />
white, education group is less than high school, <strong>and</strong> health status is poor. Bootstrap st<strong>and</strong>ard errors clustered by<br />
<strong>in</strong>dividual <strong>in</strong> parentheses.
CHAPTER 3. MEDIGAP 152<br />
Table 3.12: Third Stage: 3SLS Estimates of Medical Costs on Medigap Indicator<br />
(1)<br />
Dependent Variable: Log Medical Costs<br />
(2) (3)<br />
Log premium 0.495 (0.152) 0.630 (0.151) 0.585 (0.152)<br />
Log HRR cost -0.019 (0.165) 0.348 (0.171) 0.207 (0.171)<br />
Male -0.044 (0.072) -0.057 (0.102)<br />
Age-65 0.806 (0.128) 0.899 (0.114)<br />
(Age-65)^2 -0.232 (0.065) -0.295 (0.059)<br />
(Age-65)^3 0.021 (0.009) 0.029 (0.008)<br />
Disabled eligibility<br />
Race group<br />
0.421 (0.075) 0.325 (0.062)<br />
Asian -0.276 (0.175) -0.183 (0.179)<br />
Black -0.044 (0.110) -0.129 (0.110)<br />
Other<br />
Education group<br />
0.043 (0.182) -0.061 (0.152)<br />
High school 0.100 (0.079) 0.192 (0.076)<br />
Less than college 0.089 (0.080) 0.221 (0.075)<br />
College or more 0.159 (0.095) 0.335 (0.092)<br />
Log <strong>in</strong>come -0.007 (0.034) 0.054 (0.030)<br />
Married -0.081 (0.069) 0.016 (0.065)<br />
Work<strong>in</strong>g -0.435 (0.062) -0.172 (0.059)<br />
Served <strong>in</strong> armed forces<br />
Self-reported health<br />
-0.100 (0.100) -0.114 (0.096)<br />
Excellent -0.596 (0.137)<br />
Very good -0.547 (0.145)<br />
Good -0.387 (0.137)<br />
Fair -0.192 (0.129)<br />
Diagnosis <strong>in</strong>dicators Yes<br />
Treatment <strong>in</strong>dicators Yes<br />
IADL <strong>in</strong>dicators Yes<br />
Year FE Yes Yes Yes<br />
R-squared 3.5% 10.6% 29.1%<br />
N 3,415 3,415 3,415<br />
Source: 2000-2005 MCBS; Weiss Rat<strong>in</strong>gs<br />
Notes: Parameter estimates from 3SLS cost equation. Sample of <strong>in</strong>dividuals with HMO, Medigap or no supplemental<br />
coverage who turned 65 no sooner than 2000 exclud<strong>in</strong>g those with positive VA expenditure. Omitted race group is white,<br />
education group is less than high school, <strong>and</strong> health status is poor. Bootstrap st<strong>and</strong>ard errors clustered by <strong>in</strong>dividual <strong>in</strong><br />
parentheses.
Chapter 4<br />
Do Expir<strong>in</strong>g Budgets Lead to<br />
Wasteful Year-End Spend<strong>in</strong>g?<br />
Evidence from Federal<br />
Procurement<br />
with Jeffrey Liebman<br />
4.1 Introduction<br />
Many <strong>organization</strong>s have budgets that expire at year’s end. In the U.S., most budget<br />
authority provided to federal government agencies for discretionary spend<strong>in</strong>g requires<br />
the agencies to obligate funds by the end of the fiscal year or return them to the Trea-<br />
sury general fund, <strong>and</strong> state <strong>and</strong> municipal agencies typically face similar constra<strong>in</strong>ts<br />
(???). 1<br />
For these <strong>organization</strong>s, unspent fund<strong>in</strong>g may not only represent a lost opportunity<br />
1 At the end of the federal fiscal year, unobligated balances cease to be available for the purpose<br />
of <strong>in</strong>curr<strong>in</strong>g new obligations. They sit <strong>in</strong> an expired account for 5 years <strong>in</strong> case adjustments are<br />
needed <strong>in</strong> order to accurately account for the cost of obligations <strong>in</strong>curred dur<strong>in</strong>g the fiscal year for<br />
which the funds were appropriated. At the end of the 5 years, the funds revert to the Treasury<br />
general fund.<br />
153
CHAPTER 4. YEAR-END SPENDING 154<br />
but can also signal a lack of need to budget-setters, decreas<strong>in</strong>g fund<strong>in</strong>g <strong>in</strong> future<br />
budget cycles (??). When current spend<strong>in</strong>g is explicitly used as the basel<strong>in</strong>e <strong>in</strong> sett<strong>in</strong>g<br />
budgets for the follow<strong>in</strong>g year, this signal<strong>in</strong>g effect is magnified.<br />
This “use it or lose it” feature of time-limited budget authority has the potential to<br />
result <strong>in</strong> low value spend<strong>in</strong>g—s<strong>in</strong>ce the opportunity cost to <strong>organization</strong>s of spend<strong>in</strong>g<br />
about-to-expire funds is effectively zero. Exacerbat<strong>in</strong>g this problem is the <strong>in</strong>centive to<br />
build up a ra<strong>in</strong>y day fund over the front end of the budget cycle. Most <strong>organization</strong>s<br />
are de facto liquidity constra<strong>in</strong>ed, fac<strong>in</strong>g at the very least a high cost to acquir<strong>in</strong>g mid-<br />
cycle budget authority. When future spend<strong>in</strong>g dem<strong>and</strong>s are uncerta<strong>in</strong>, <strong>organization</strong>s<br />
have an <strong>in</strong>centive to hoard. Thus, at the end of the budget cycle, <strong>organization</strong>s often<br />
have a buffer-stock of fund<strong>in</strong>g which they must rush to spend through.<br />
To illustrate the key mechanisms which create the potential for wasteful year-end<br />
spend<strong>in</strong>g, we build a simple model of spend<strong>in</strong>g with fixed budgets <strong>and</strong> expir<strong>in</strong>g funds.<br />
Annual budget cycles are divided <strong>in</strong>to two six-month periods. Spend<strong>in</strong>g exhibits de-<br />
creas<strong>in</strong>g returns with<strong>in</strong> each period <strong>and</strong> is scaled by a productivity parameter that<br />
is unknown <strong>in</strong> advance. In the face of this uncerta<strong>in</strong>ty, <strong>organization</strong>s engage <strong>in</strong> pre-<br />
cautionary sav<strong>in</strong>gs <strong>in</strong> the first period. In the second period, the prospect of expir<strong>in</strong>g<br />
funds leads to a rush to spend. As a result, average spend<strong>in</strong>g is higher <strong>and</strong> average<br />
quality is lower at the end of the year.<br />
There is some suggestive evidence consistent with these predictions for the U.S.<br />
federal government. A Department of Defense employee <strong>in</strong>terviewed by ? describes<br />
“merchants <strong>and</strong> contractors camped outside contract<strong>in</strong>g offices on September 30th<br />
(the close of the fiscal year) just <strong>in</strong> case money came through to fund their contracts.”<br />
A 1980 report by the Senate Subcommittee on Oversight of Government Management<br />
on “Hurry-Up Spend<strong>in</strong>g” found that the rush to spend led to poorly def<strong>in</strong>ed contracts,<br />
limited competition <strong>and</strong> <strong>in</strong>flated prices, <strong>and</strong> the procurement of goods <strong>and</strong> services<br />
for which there was no current need (?). At a congressional hear<strong>in</strong>g <strong>in</strong> 2006, agency<br />
representatives admitted to a “use-it-or-lose-it” mentality <strong>and</strong> a “rush to obligate”<br />
at year’s end (?). The 2007 Federal Acquisition Advisory Panel concluded that “the<br />
large volume of procurement execution be<strong>in</strong>g effected late <strong>in</strong> the year is hav<strong>in</strong>g a<br />
negative effect on the contract<strong>in</strong>g process <strong>and</strong> is a significant motivator for many of
CHAPTER 4. YEAR-END SPENDING 155<br />
the issues we have noted with respect to, among another th<strong>in</strong>gs, lack of competition<br />
<strong>and</strong> poor management of <strong>in</strong>teragency contracts” (?).<br />
Yet despite these accounts, there is no hard evidence on whether year-end spend-<br />
<strong>in</strong>g surges are currently occurr<strong>in</strong>g <strong>in</strong> the U.S. federal government or whether year-end<br />
spend<strong>in</strong>g is lower-value than spend<strong>in</strong>g dur<strong>in</strong>g the rest of the year. Government Ac-<br />
countability Office (GAO) reports <strong>in</strong> 1980 <strong>and</strong> 1985 documented that fourth quarter<br />
spend<strong>in</strong>g was somewhat higher than spend<strong>in</strong>g dur<strong>in</strong>g the rest of year us<strong>in</strong>g aggregate<br />
agency spend<strong>in</strong>g data. In a follow-up report, ? stated that because “substantial<br />
reforms <strong>in</strong> procurement plann<strong>in</strong>g <strong>and</strong> competition requirements have changed the<br />
environment . . . year-end spend<strong>in</strong>g is unlikely to present the same magnitude of<br />
problems <strong>and</strong> issues as before.” However, this later report was unable to exam<strong>in</strong>e<br />
quarterly agency spend<strong>in</strong>g patterns for 1997 because agency compliance with quar-<br />
terly report<strong>in</strong>g requirements was <strong>in</strong>complete. Other government professionals cite<br />
similar constra<strong>in</strong>ts to empirical analysis. In summariz<strong>in</strong>g the response of Department<br />
of Defense contract<strong>in</strong>g officers to the <strong>in</strong>terview question, “How would you measure<br />
the quality of year-end spend<strong>in</strong>g?” ? writes, “Absent flagrant abuse that no one<br />
could miss, there is no practical way of weigh<strong>in</strong>g year-end spend<strong>in</strong>g.”<br />
This paper address the empirical shortfall. It beg<strong>in</strong>s by document<strong>in</strong>g the with<strong>in</strong>-<br />
year pattern of federal spend<strong>in</strong>g on government contracts, the ma<strong>in</strong> category of spend-<br />
<strong>in</strong>g where significant discretion about tim<strong>in</strong>g exists. The analysis demonstrates that<br />
there is a large surge <strong>in</strong> the 52nd week of the year that is concentrated <strong>in</strong> procure-<br />
ments for construction-related goods <strong>and</strong> services, furnish<strong>in</strong>gs <strong>and</strong> office equipment,<br />
<strong>and</strong> I.T. services <strong>and</strong> equipment. It also shows that the date of completion of annual<br />
appropriations legislation has a noticeable effect on the tim<strong>in</strong>g of federal contract<strong>in</strong>g,<br />
consistent with anecdotal claims that part of the difficulties agencies face <strong>in</strong> effective<br />
management of acquisitions comes from tardy enactment of appropriations legisla-<br />
tion. The estimates show that a delay of ten weeks, roughly the average over this<br />
time period, raises raises the share of spend<strong>in</strong>g <strong>in</strong> the last month by 1 percentage<br />
po<strong>in</strong>t, from a base of about 15 percent.<br />
We then analyze the impact of the end-of-year spend<strong>in</strong>g surge on spend<strong>in</strong>g quality<br />
us<strong>in</strong>g a newly available dataset on the status of the federal government’s 686 major
CHAPTER 4. YEAR-END SPENDING 156<br />
<strong>in</strong>formation technology projects—a total of $130 billion <strong>in</strong> spend<strong>in</strong>g. Consistent with<br />
the model, spend<strong>in</strong>g on these I.T. projects spikes <strong>in</strong> the last week of the fiscal year,<br />
<strong>in</strong>creas<strong>in</strong>g to 7.2 times the rest-of-year weekly average. Moreover, the spike is not<br />
isolated to a small set of agencies or a subset of years, but rather a persistent feature<br />
both across agencies <strong>and</strong> over time. In t<strong>and</strong>em with the spend<strong>in</strong>g <strong>in</strong>crease, there<br />
is a sharp drop-off <strong>in</strong> <strong>in</strong>vestment quality. Based on a categorical <strong>in</strong>dex of overall<br />
<strong>in</strong>vestment performance, which comb<strong>in</strong>es assessments from agency <strong>in</strong>formation officers<br />
with data on cost <strong>and</strong> timel<strong>in</strong>ess, we f<strong>in</strong>d that projects that orig<strong>in</strong>ate <strong>in</strong> the last week<br />
of the fiscal year have 2.2 to 5.6 times higher odds of hav<strong>in</strong>g a lower quality score.<br />
Ordered logit <strong>and</strong> OLS regressions show that this effect is robust to agency <strong>and</strong> year<br />
specific factors as well as to a rich set of project characteristic controls.<br />
Our f<strong>in</strong>d<strong>in</strong>gs suggest that the various safeguard measures put <strong>in</strong>to place <strong>in</strong> re-<br />
sponse to the 1980 Senate Subcommittee report (?) <strong>and</strong> to broader concerns about<br />
acquisition plann<strong>in</strong>g have not been fully successful <strong>in</strong> elim<strong>in</strong>at<strong>in</strong>g the end-of-year<br />
rush-to-spend <strong>in</strong>efficiency. An alternative solution is to give agencies the ability to<br />
roll over some of their unused fund<strong>in</strong>g for an additional year. Provisions of this nature<br />
have been applied with apparent success <strong>in</strong> the states of Oklahoma <strong>and</strong> Wash<strong>in</strong>gton,<br />
as well as <strong>in</strong> the UK (??). With<strong>in</strong> the U.S. federal government, the Department of<br />
Justice (DOJ) has obta<strong>in</strong>ed special authority to roll over unused budget authority<br />
<strong>in</strong>to a fund that can be used <strong>in</strong> the follow<strong>in</strong>g year.<br />
We extend the model to allow for rollover <strong>and</strong> show that, <strong>in</strong> the context of the<br />
model, the efficiency ga<strong>in</strong>s from this ability are unequivocally positive. To test this<br />
prediction, we study I.T. contracts at the Department of Justice which has special<br />
rollover authority. We show that there is only a small end-of-year I.T. spend<strong>in</strong>g spike<br />
at DOJ <strong>and</strong> that the one major I.T. contract DOJ issued <strong>in</strong> the 52nd week of the<br />
year has a quality rat<strong>in</strong>g that is well above average.<br />
The rest of the paper proceeds as follows. Section 4.2 presents a model of wasteful<br />
year-end spend<strong>in</strong>g <strong>and</strong> discusses the mechanisms that could potentially lead to end-<br />
of-year spend<strong>in</strong>g be<strong>in</strong>g of lower quality than spend<strong>in</strong>g dur<strong>in</strong>g the rest of the year.<br />
Section 4.3 exam<strong>in</strong>es the surge <strong>in</strong> year-end spend<strong>in</strong>g us<strong>in</strong>g a comprehensive dataset<br />
on federal procurement. Section 4.4 tests for a year-end drop-off <strong>in</strong> quality us<strong>in</strong>g data
CHAPTER 4. YEAR-END SPENDING 157<br />
on I.T. <strong>in</strong>vestments. Section 4.5 analyzes the Department of Justice experience with<br />
rollover authority. Section 4.6 concludes.<br />
4.2 A Model of Wasteful Year-End Spend<strong>in</strong>g<br />
In this section, we present a simple model to illustrate how expir<strong>in</strong>g budgets can give<br />
rise to wasteful year-end spend<strong>in</strong>g. The model has three key features. First, there<br />
are decreas<strong>in</strong>g returns to spend<strong>in</strong>g with<strong>in</strong> each sub-year period. Decreas<strong>in</strong>g returns<br />
could be motivated by a priority-based budget<strong>in</strong>g rule, where dur<strong>in</strong>g a given period<br />
<strong>organization</strong>s allocate resources to projects accord<strong>in</strong>g to the surplus they provide.<br />
Alternatively, decreas<strong>in</strong>g returns could be motivated by short-run rigidities <strong>in</strong> the<br />
production function. For example, federal agencies with a fixed staff of contract<strong>in</strong>g<br />
specialists might have less time to devote to each contract <strong>in</strong> a period with abnormally<br />
high spend<strong>in</strong>g.<br />
Second, there is uncerta<strong>in</strong>ty over the value of future budget resources. One can<br />
th<strong>in</strong>k of uncerta<strong>in</strong>ty aris<strong>in</strong>g from either dem<strong>and</strong> or supply factors. Shifts <strong>in</strong> mili-<br />
tary strategy or an <strong>in</strong>fluenza outbreak, for example, could generate an unanticipated<br />
change <strong>in</strong> dem<strong>and</strong> for budget resources. On the supply side, uncerta<strong>in</strong>ty could be<br />
driven by variation <strong>in</strong> the price or quality of desired goods <strong>and</strong> services. 2<br />
Third, resources that are not spent by the end of the year cannot be rolled over<br />
to produce value <strong>in</strong> subsequent periods. At the end of this section, we also analyze<br />
what happens when this constra<strong>in</strong>t is relaxed.<br />
4.2.1 The Basel<strong>in</strong>e Model<br />
Consider an annual model of budget<strong>in</strong>g where an <strong>organization</strong> chooses how to appor-<br />
tion budget authority, B, normalized to 1, over two six-month periods, denoted by<br />
t = {1, 2}, to maximize a Cobb-Douglas objective. Denote spend<strong>in</strong>g <strong>in</strong> each period by<br />
xt, <strong>and</strong> normalize its price to 1. To model decreas<strong>in</strong>g returns <strong>and</strong> uncerta<strong>in</strong>ty, assume<br />
that the Cobb-Douglas elasticity parameters αt are stochastic i.i.d. draws from the<br />
2 As an example of supply side uncerta<strong>in</strong>ty, dur<strong>in</strong>g the recent recession many agencies have experienced<br />
construction costs for Recovery Act projects that were below projections.
CHAPTER 4. YEAR-END SPENDING 159<br />
To summarize, decreas<strong>in</strong>g returns <strong>and</strong> uncerta<strong>in</strong>ty create an <strong>in</strong>centive for organi-<br />
zations to build up a ra<strong>in</strong>y day fund <strong>in</strong> the first period, spend<strong>in</strong>g less than half of their<br />
budget on average. At the end of the year, spend<strong>in</strong>g <strong>in</strong>creases <strong>and</strong> average quality is<br />
lower than <strong>in</strong> the earlier part of the year.<br />
It is worth emphasiz<strong>in</strong>g that this model illustrates two different channels through<br />
which end-of-year spend<strong>in</strong>g can be of lower quality. The first channel comes from<br />
the uncerta<strong>in</strong>ty about future needs <strong>and</strong> will lead <strong>organization</strong>s to engage <strong>in</strong> only<br />
high value projects early <strong>in</strong> the year <strong>and</strong> wait to undertake lower value projects once<br />
the full-year dem<strong>and</strong>s on resources are clearer. The second channel comes from the<br />
production-function rigidities motivated by the anecdotal evidence that contract<strong>in</strong>g<br />
officers become over-extended <strong>in</strong> the end-of-year rush to get money out the door.<br />
Either channel by itself would be enough to produce an empirical relationship <strong>in</strong><br />
which end-of-year spend<strong>in</strong>g is of lower value.<br />
4.2.2 Rollover Budget Authority<br />
The <strong>in</strong>efficiency from wasteful year-end spend<strong>in</strong>g raises the question of whether any-<br />
th<strong>in</strong>g can be done to reduce it. Reduc<strong>in</strong>g uncerta<strong>in</strong>ty would be helpful, but is <strong>in</strong>fea-<br />
sible <strong>in</strong> practice for many <strong>organization</strong>s due to the <strong>in</strong>herently unpredictable nature<br />
of some types of shocks. Another way to potentially <strong>in</strong>crease efficiency would be to<br />
allow <strong>organization</strong>s to roll over budget authority across fiscal years. Under such a<br />
system, budget<strong>in</strong>g would still occur on an annual basis, but rather than expir<strong>in</strong>g at<br />
year’s end, unused funds would be added to the newly granted budget authority <strong>in</strong><br />
the next year.<br />
The idea that budget authority should last for longer than one year is not new. As<br />
? has po<strong>in</strong>ted out, grant<strong>in</strong>g Congress the power to collect taxes, Article 1, Section 8<br />
of the U.S. Constitution gave Congress the power of taxation, “To raise <strong>and</strong> support<br />
armies, but no appropriation of money to that use shall be for longer term than two<br />
years.” Not only does this suggest that the Found<strong>in</strong>g Fathers though that two-year<br />
limits were reasonable <strong>in</strong> some <strong>in</strong>stances, but by fail<strong>in</strong>g to attach this clause to other<br />
forms of federal expenditure, they implied by omission that periods longer than two
CHAPTER 4. YEAR-END SPENDING 162<br />
For many types of government spend<strong>in</strong>g, there is little potential for a spike <strong>in</strong><br />
spend<strong>in</strong>g at the end of the year. The 65 percent of spend<strong>in</strong>g that is made up of<br />
m<strong>and</strong>atory programs <strong>and</strong> <strong>in</strong>terest on the debt is not subject to the tim<strong>in</strong>g limitations<br />
associated with annual appropriations. The 13 percent of spend<strong>in</strong>g that pays for<br />
compensation for federal employees is unlikely to exhibit an end-of-year surge s<strong>in</strong>ce<br />
new hires br<strong>in</strong>g ongo<strong>in</strong>g costs. This leaves procurement of goods <strong>and</strong> services from<br />
the private sector as the ma<strong>in</strong> category of government spend<strong>in</strong>g where an end-of-year<br />
spend<strong>in</strong>g surge could potentially occur. We therefore focus our empirical work on the<br />
procurement of goods <strong>and</strong> services that accounted for $538 billion or 15.3 percent of<br />
government spend<strong>in</strong>g <strong>in</strong> 2009 (up from $165 billion dollars <strong>and</strong> 9.2 percent <strong>in</strong> 2000).<br />
It is worth not<strong>in</strong>g that even with<strong>in</strong> procurement spend<strong>in</strong>g, there are some cat-<br />
egories of spend<strong>in</strong>g for which it would be unlikely to observe an end-of-year spike.<br />
Some types of appropriated spend<strong>in</strong>g, such as military construction, come with longer<br />
spend<strong>in</strong>g horizons to provide greater flexibility to agencies. Moreover, there are limits<br />
to what k<strong>in</strong>ds of purchases can be made at year end.<br />
In particular, Federal law provides that appropriations are available only to “meet<br />
the bona fide needs of the fiscal year for which they are appropriated.” Balances<br />
rema<strong>in</strong><strong>in</strong>g at the end of the year cannot generally be used to prepay for the next<br />
year’s needs. A classic example of an improper obligation was an order for gasol<strong>in</strong>e<br />
placed 3 days before the end of the fiscal year to be delivered <strong>in</strong> monthly <strong>in</strong>stallments<br />
throughout the follow<strong>in</strong>g fiscal year (?). That said, when there is an ongo<strong>in</strong>g need<br />
<strong>and</strong> it is impossible to separate the purchase <strong>in</strong>to components performed <strong>in</strong> different<br />
fiscal years, it can be appropriate to enter <strong>in</strong>to a contract <strong>in</strong> one fiscal year even<br />
though a significant portion of the performance is <strong>in</strong> the subsequent fiscal year. In<br />
contrast, contracts that are readily severable generally may not cross fiscal years<br />
(unless specifically authorized by statute). 6<br />
6 Over the past two decades, Congress has significantly exp<strong>and</strong>ed multi-year contract<strong>in</strong>g authorities.<br />
For example, the General Services Adm<strong>in</strong>istration can enter <strong>in</strong>to leases for periods of up to 20<br />
years, <strong>and</strong> agencies can contract for services from utilities for periods of up to 10 years.
CHAPTER 4. YEAR-END SPENDING 163<br />
4.3.1 The Federal Procurement Data System<br />
Fall<strong>in</strong>g technology costs <strong>and</strong> the government transparency movement have comb<strong>in</strong>ed<br />
to produce an extraord<strong>in</strong>ary <strong>in</strong>crease <strong>in</strong> the amount of government data available on<br />
the web (??). As of October 2010, Data.gov had 2,936 U.S. federal executive branch<br />
datasets available. The Federal Fund<strong>in</strong>g Accountability <strong>and</strong> Transparency Act of<br />
2006, sponsored by Senators Coburn, Obama, Carper, <strong>and</strong> McCa<strong>in</strong>, required OMB<br />
to create a <strong>public</strong> website, show<strong>in</strong>g every federal award, <strong>in</strong>clud<strong>in</strong>g the name of the<br />
entity receiv<strong>in</strong>g the award <strong>and</strong> the amount of the award, among other <strong>in</strong>formation.<br />
USAspend<strong>in</strong>g.gov was launched <strong>in</strong> December 2007 <strong>and</strong> now conta<strong>in</strong>s data on federal<br />
contracts, grants, direct payments, <strong>and</strong> loans.<br />
The data currently available on USAspend<strong>in</strong>g.gov <strong>in</strong>clude the full Federal Procure-<br />
ment Data System (FPDS) from 2000 through 2009. FPDS is the data system that<br />
tracks all federal contracts. Every new contract awarded as well as every follow-on<br />
contract<strong>in</strong>g action such as a contract renewal or modification results <strong>in</strong> an observation<br />
<strong>in</strong> FPDS. Up to 176 pieces of <strong>in</strong>formation are available for each contract <strong>in</strong>clud<strong>in</strong>g<br />
dollar value, a four digit code describ<strong>in</strong>g the product or service be<strong>in</strong>g purchased, the<br />
component of the agency mak<strong>in</strong>g the purchase, the identity of the provider, the type<br />
of contract be<strong>in</strong>g used (fixed price, cost-type, time <strong>and</strong> materials, etc.), <strong>and</strong> the type<br />
of bidd<strong>in</strong>g mechanism used. While FPDS was orig<strong>in</strong>ally created <strong>in</strong> 1978, agency re-<br />
port<strong>in</strong>g was <strong>in</strong>complete for many years, <strong>and</strong> we are told that it would be difficult to<br />
assemble comprehensive data for years before 2000. Moreover, while FPDS is thought<br />
to conta<strong>in</strong> all government contracts from 2000 on, data quality for many fields was<br />
uneven before the 2003 FPDS modernization. Therefore, for most of the FPDS-based<br />
analyses <strong>in</strong> this paper, we limit ourselves to data from fiscal years 2004 through 2009. 7<br />
Table 4.1 shows selected characteristics of the FPDS 2004 to 2009 sample. There<br />
were 14.6 million contracts dur<strong>in</strong>g this period or an average of 2.4 million a year.<br />
The distribution of contract size is highly skewed. N<strong>in</strong>ety-five percent of contracts<br />
were for dollar amounts below $100,000, while 78 percent of contract spend<strong>in</strong>g is<br />
7 FPDS excludes classified contracts. Data are made available <strong>in</strong> FPDS soon after an award,<br />
except dur<strong>in</strong>g wartime the Department of Defense is permitted a 90 day delay to m<strong>in</strong>imize the<br />
potential for disclosure of mission critical <strong>in</strong>formation.
CHAPTER 4. YEAR-END SPENDING 164<br />
accounted for by contracts of more than $1 million. Seventy percent of contract<br />
spend<strong>in</strong>g is by the Department of Defense. The Department of Energy <strong>and</strong> NASA,<br />
which rely on contractors to run large labs <strong>and</strong> production facilities, <strong>and</strong> the General<br />
Services Adm<strong>in</strong>istration, which enters <strong>in</strong>to government-wide contracts <strong>and</strong> contracts<br />
on behalf of other agencies, are the next largest agencies <strong>in</strong> terms of spend<strong>in</strong>g on<br />
contracts. Twenty-n<strong>in</strong>e percent of contracts were non-competitive, 20 percent were<br />
competitive but received only a s<strong>in</strong>gle bid, <strong>and</strong> 51 percent received more than 1 bid.<br />
Sixty-five percent were fixed price, 30 percent were cost-reimbursement, <strong>and</strong> 6 percent<br />
were on a time <strong>and</strong> materials or labor hours basis.<br />
4.3.2 The With<strong>in</strong>-Year Pattern of Government Procurement<br />
Spend<strong>in</strong>g<br />
Figure 4.1 shows contract spend<strong>in</strong>g by week, pool<strong>in</strong>g data from 2004 through 2009.<br />
There is a clear spike <strong>in</strong> spend<strong>in</strong>g at the end of the year with 16.5 percent of all<br />
spend<strong>in</strong>g occurr<strong>in</strong>g <strong>in</strong> the last month of the year <strong>and</strong> 8.7 percent occurr<strong>in</strong>g <strong>in</strong> the<br />
last week. The bottom panel shows that when measured by the number of contracts<br />
rather than the dollar value, there is also clear evidence of an end-of-the-year spike,<br />
with 12.0 percent occurr<strong>in</strong>g <strong>in</strong> the last month <strong>and</strong> 5.6 percent occurr<strong>in</strong>g <strong>in</strong> the last<br />
week.<br />
Table 4.1 shows that the end of the year spend<strong>in</strong>g surge occurs <strong>in</strong> all major<br />
government agencies. If spend<strong>in</strong>g were distributed uniformly throughout the year,<br />
we would expect to see 1.9 percent <strong>in</strong> the f<strong>in</strong>al week of the year. The lowest agency<br />
percentage is 3.6 percent.<br />
Table 4.3 shows the percent of spend<strong>in</strong>g on different types of goods <strong>and</strong> services<br />
that occurs at the end of the year. The table shows some of the largest spend-<br />
<strong>in</strong>g categories along with some selected smaller categories that are very similar to<br />
the large categories. Construction-related goods <strong>and</strong> services, furnish<strong>in</strong>gs <strong>and</strong> office<br />
equipment, <strong>and</strong> I.T. services <strong>and</strong> equipment all have end-of-year spend<strong>in</strong>g rates that<br />
are significantly higher than the average. These categories of spend<strong>in</strong>g often represent<br />
areas where there is significant flexibility about tim<strong>in</strong>g for perform<strong>in</strong>g ma<strong>in</strong>tenance
CHAPTER 4. YEAR-END SPENDING 165<br />
or upgrad<strong>in</strong>g facilities <strong>and</strong> equipment, <strong>and</strong> which, because they represent on-go<strong>in</strong>g<br />
needs, have a reasonable chance of satisfy<strong>in</strong>g the bona fide needs requirement even if<br />
spend<strong>in</strong>g happens at the end of the year.<br />
The categories of spend<strong>in</strong>g under the “Services” head<strong>in</strong>g have end-of-year spend<strong>in</strong>g<br />
rates that are near the average. For these k<strong>in</strong>ds of services it will often be difficult to<br />
meet the bona fide needs requirements unless the services are <strong>in</strong>separable from larger<br />
purchases, the services are necessary to provide cont<strong>in</strong>uity <strong>in</strong>to the beg<strong>in</strong>n<strong>in</strong>g of the<br />
next fiscal year, or the services are covered by special multiyear contract<strong>in</strong>g authori-<br />
ties. Thus it is not surpris<strong>in</strong>g that their rate of end of year spend<strong>in</strong>g is lower than that<br />
for construction, for example. There are two categories of spend<strong>in</strong>g where there is<br />
very little year-end surge. The first is ongo<strong>in</strong>g expenses such as fuels where attempts<br />
to spend at the end of the year would represent a blatant violation of prohibitions<br />
aga<strong>in</strong>st pay<strong>in</strong>g for the follow<strong>in</strong>g year’s expenses with current year appropriations.<br />
The second is military weapons systems where because of long plann<strong>in</strong>g horizons <strong>and</strong><br />
the flexibility provided by special appropriations authorities, one would not expect<br />
to see a concentration of spend<strong>in</strong>g at the end of the year.<br />
Figure 4.1 also shows a spike <strong>in</strong> spend<strong>in</strong>g <strong>in</strong> the first week of the year, along with<br />
smaller spikes at the beg<strong>in</strong>n<strong>in</strong>g of each quarter. The spend<strong>in</strong>g patterns for these<br />
beg<strong>in</strong>n<strong>in</strong>g of period contracts are very different from those at the end of the year.<br />
Appendix Table 4.9 shows that leases <strong>and</strong> service contracts are responsible for most<br />
of the beg<strong>in</strong>n<strong>in</strong>g-of-period spikes.<br />
4.3.3 The Impact of Appropriations Tim<strong>in</strong>g on the With<strong>in</strong>-<br />
Year Pattern of Government Procurement Spend<strong>in</strong>g<br />
It is the exception rather than the rule for Congress to pass annual appropriations<br />
bills before the beg<strong>in</strong>n<strong>in</strong>g of the fiscal year. Over the 10 years from 2000 to 2009, the<br />
full annual appropriations process was never completed on time. Although defense<br />
appropriations bills were enacted before the start of the fiscal year 4 times, <strong>in</strong> 8 of<br />
the ten years, appropriations for all or nearly all of the civilian agencies were enacted<br />
<strong>in</strong> a s<strong>in</strong>gle consolidated appropriations act well after the start of the fiscal year.
CHAPTER 4. YEAR-END SPENDING 166<br />
Analysts have attributed some of the challenges fac<strong>in</strong>g federal acquisition to the<br />
tard<strong>in</strong>ess of the appropriations process, s<strong>in</strong>ce the delays <strong>in</strong>troduce uncerta<strong>in</strong>ty <strong>and</strong><br />
compress the time available to plan <strong>and</strong> implement a successful acquisition strategy<br />
(?). In this subsection we analyze the relationship between the tim<strong>in</strong>g of the annual<br />
appropriations acts <strong>and</strong> the with<strong>in</strong>-year pattern of government contract spend<strong>in</strong>g.<br />
For this analysis, we use the full 2000 to 2009 FPDS data, even though the data prior<br />
to 2004 are of lower quality. In particular, <strong>in</strong> these earlier years it appears that for<br />
some agencies, contracts are all assigned dates <strong>in</strong> the middle of the month <strong>and</strong> the<br />
with<strong>in</strong>-month weekly pattern is therefore not fully available.<br />
Figure 4.2 shows results from regress<strong>in</strong>g measures of end-of-year spend<strong>in</strong>g on the<br />
tim<strong>in</strong>g of annual appropriations. This analysis has two data po<strong>in</strong>ts for each year,<br />
one represent<strong>in</strong>g defense spend<strong>in</strong>g <strong>and</strong> the other represent<strong>in</strong>g non-defense spend<strong>in</strong>g.<br />
For each observation we measure the share of annual contract spend<strong>in</strong>g occurr<strong>in</strong>g <strong>in</strong><br />
the last quarter, month, <strong>and</strong> week of the year <strong>and</strong> the “weeks late” of the enactment<br />
of annual appropriations legislation (enactment is def<strong>in</strong>ed by the date the President<br />
signs the legislation). “Weeks late” measures time relative to October 1 <strong>and</strong> takes on<br />
negative values when appropriations were enacted prior to the start of the fiscal year.<br />
For defense spend<strong>in</strong>g, “weeks late” measures the date that the defense appropriations<br />
bill was enacted. For non-defense spend<strong>in</strong>g the date is assigned from the date of<br />
the consolidated appropriations act, or, <strong>in</strong> the case of the two years <strong>in</strong> which there<br />
was not a consolidated act, a date that is the midpo<strong>in</strong>t of the <strong>in</strong>dividual non-defense<br />
appropriations acts. 8<br />
There is a clear pattern <strong>in</strong> the data <strong>in</strong> which later appropriation dates result<br />
<strong>in</strong> a greater fraction of end-of-year spend<strong>in</strong>g. In the plots, we show the separate<br />
slopes of the defense <strong>and</strong> non-defense observations. Defense spend<strong>in</strong>g tends to be<br />
appropriated earlier <strong>and</strong> to have less end of year spend<strong>in</strong>g, but the slopes for the two<br />
types of spend<strong>in</strong>g are similar. The labels show the regression coefficients, <strong>in</strong>clud<strong>in</strong>g<br />
the coefficients from a pooled regression <strong>in</strong> which defense <strong>and</strong> non-defense spend<strong>in</strong>g<br />
8 We aggregate all non-defense spend<strong>in</strong>g because it facilitates communication of the pattern of<br />
results while captur<strong>in</strong>g nearly all of the available variation. We have also run analysis <strong>in</strong> which we<br />
assign each non-defense agency the date of its <strong>in</strong>dividual appropriations act <strong>and</strong> obta<strong>in</strong> very similar<br />
results.
CHAPTER 4. YEAR-END SPENDING 167<br />
have different <strong>in</strong>tercepts but are constra<strong>in</strong>ed to have the same slope. The estimates<br />
show that a delay of ten weeks, roughly the average over this time period, raises the<br />
share of spend<strong>in</strong>g <strong>in</strong> the last quarter by 2 percentage po<strong>in</strong>ts from a base of about<br />
27 percent. A ten-week delay raises the share of spend<strong>in</strong>g <strong>in</strong> the last month by 1<br />
percentage po<strong>in</strong>t, from a base of about 15 percent. Both coefficients are statistically<br />
significant at the 1 percent level. As we mentioned above, we do not have reliable<br />
with<strong>in</strong>-month data on tim<strong>in</strong>g for the years before 2004, so we exclude years before<br />
2004 for the analysis of spend<strong>in</strong>g dur<strong>in</strong>g the last week of the year. The estimates<br />
<strong>in</strong>dicate that a 10 week delay raises the share of spend<strong>in</strong>g by 1 percentage po<strong>in</strong>t on<br />
a base of 9 percent. Due to the smaller sample, the estimate is less precise, with a<br />
p-value of .07.<br />
Overall, the analysis <strong>in</strong> this section shows that the end-of-year spend<strong>in</strong>g surge<br />
is alive <strong>and</strong> well, thirty years after Congress <strong>and</strong> GAO focused significant attention<br />
on the problem <strong>and</strong> despite reforms designed to limit it. Moreover, claims that late<br />
appropriations <strong>in</strong>crease the end-of-year volume of contract<strong>in</strong>g activity are accurate,<br />
suggest<strong>in</strong>g that late appropriations may be exacerbat<strong>in</strong>g the already adverse effects<br />
of hav<strong>in</strong>g an acquisition workforce operat<strong>in</strong>g beyond capacity at the end of the year.<br />
A surge <strong>in</strong> end-of-year spend<strong>in</strong>g does not necessarily imply bad outcomes. Agency<br />
acquisition staff can plan ahead for the possibility that extra funds will be available.<br />
Indeed, for large contracts weeks <strong>and</strong> even months of lead time are generally necessary.<br />
The next section of the paper therefore analyzes the relative quality of end-of-year<br />
contract spend<strong>in</strong>g to explore whether there are any adverse effects of the end-of-year<br />
spend<strong>in</strong>g surge.<br />
4.4 Is End of Year Spend<strong>in</strong>g of Lower Quality?<br />
Our model predicts that end-of-year spend<strong>in</strong>g will be of lower quality both because<br />
agencies will spend money at the end of the year on low value projects <strong>and</strong> because<br />
the <strong>in</strong>creased volume of contract<strong>in</strong>g at the end of the year will lead to less effective<br />
management of those acquisitions. As the discussion <strong>in</strong> the <strong>in</strong>troduction <strong>in</strong>dicated,<br />
it has been challeng<strong>in</strong>g historically to study contract quality because of the limited
CHAPTER 4. YEAR-END SPENDING 168<br />
availability of data. In this section of the paper, we use a new dataset that <strong>in</strong>cludes<br />
quality <strong>in</strong>formation on 686 of the most important federal I.T. procurements to study<br />
whether end-of-the year procurements are of lower quality.<br />
4.4.1 I.T. Dashboard<br />
Our data come from the federal I.T. Dashboard (www.itdashboard.gov) which tracks<br />
the performance of the most important federal I.T. projects. The I.T. Dashboard<br />
came onl<strong>in</strong>e <strong>in</strong> beta form <strong>in</strong> June, 2009 <strong>and</strong> provides the <strong>public</strong> with measures of<br />
the overall performance of major I.T. projects. Like the USAspend<strong>in</strong>g.gov data dis-<br />
cussed earlier, the I.T. Dashboard is part of the trend toward “open government”<br />
<strong>and</strong> part of a shift <strong>in</strong> federal management philosophy toward monitor<strong>in</strong>g performance<br />
trends rather than tak<strong>in</strong>g static snapshots of performance <strong>and</strong> of mak<strong>in</strong>g the trends<br />
<strong>public</strong> both for the sake of transparency <strong>and</strong> to motivate agencies to achieve high<br />
performance (?). 9<br />
Along with provid<strong>in</strong>g credible performance data for a portion of contract spend<strong>in</strong>g,<br />
study<strong>in</strong>g federal I.T. projects has two other advantages. The first is the ubiquity of<br />
I.T. spend<strong>in</strong>g. Major <strong>in</strong>formation technology projects are carried out by nearly all<br />
components of the U.S. federal government. Compared to an analysis of, say, the<br />
purchase of military or medical equipment, an analysis of I.T. spend<strong>in</strong>g sh<strong>in</strong>es a much<br />
broader light on the work<strong>in</strong>gs of government, allow<strong>in</strong>g us to test our hypotheses across<br />
agencies with a wide range of missions <strong>and</strong> <strong>organization</strong>al cultures.<br />
The second advantage is that federal I.T. spend<strong>in</strong>g is an important <strong>and</strong> grow<strong>in</strong>g<br />
federal activity. Federal I.T. expenditure was $81.9 billion <strong>in</strong> 2010, <strong>and</strong> has been grow-<br />
<strong>in</strong>g at an <strong>in</strong>flation-adjusted rate of 3.8 percent over the past 5 years. 10 . Moreover,<br />
these expenditure levels do not account for the social surplus from these projects.<br />
9 The legislative foundation for the I.T. Dashboard was laid by the Cl<strong>in</strong>ger-Cohen Act of 1996,<br />
which established Chief Information Officers at 27 major federal agencies <strong>and</strong> called on them to<br />
“monitor the performance of the <strong>in</strong>formation technology programs of the agency, [<strong>and</strong>] evaluate<br />
the performance of those programs on the basis of applicable performance measurements.” The<br />
E-Government Act of 2002 built upon this by requir<strong>in</strong>g the <strong>public</strong> display of these data.<br />
10 Analytical Perspectives: Budget of the U.S. Government, 2010
CHAPTER 4. YEAR-END SPENDING 169<br />
It is reasonable to th<strong>in</strong>k that <strong>in</strong>formation systems used to monitor terrorist activi-<br />
ties, adm<strong>in</strong>ister Social Security payments, <strong>and</strong> coord<strong>in</strong>ate the health care of military<br />
veterans could have welfare impacts that far exceed their dollar costs.<br />
F<strong>in</strong>ally, it should be noted that while we are duly cautious about external validly,<br />
the widespread nature of I.T. <strong>in</strong>vestment across all types of <strong>organization</strong>s, <strong>in</strong>clud<strong>in</strong>g<br />
private sector ones, makes a study of I.T. purchases more broadly relevant than would<br />
be certa<strong>in</strong> other categories of spend<strong>in</strong>g where the federal government may be the only<br />
purchaser. Not only do non-federal <strong>organization</strong>s buy similar products under similar<br />
budget structures, but they often purchase these products from the same firms that<br />
sell to U.S. federal agencies. These firms know the end-of-year budget<strong>in</strong>g game, <strong>and</strong> if<br />
they play it at the U.S. federal level, there may be reason to believe that they operate<br />
similarly elsewhere. 11<br />
4.4.2 Data <strong>and</strong> Summary Statistics<br />
The I.T. Dashboard displays <strong>in</strong>formation on major, ongo<strong>in</strong>g projects made by 27 of<br />
the largest agencies of the federal government. The <strong>in</strong>formation is gleaned from the<br />
Exhibit 53 <strong>and</strong> Exhibit 300 forms that agencies are required to submit to the Office<br />
of Management <strong>and</strong> Budget <strong>and</strong> is constantly updated on the Dashboard website,<br />
allow<strong>in</strong>g users to view <strong>and</strong> conduct simple analysis of the data. The data we use was<br />
downloaded <strong>in</strong> March, 2010 at which time there were 761 projects be<strong>in</strong>g tracked.<br />
For the analysis, we drop the 73 observations that are miss<strong>in</strong>g the quality mea-<br />
sures, date of award, or cost variables. We also drop two enormous projects because<br />
their size would cause them to dom<strong>in</strong>ate all of the weighted regression results <strong>and</strong><br />
because they are too high-profile to be <strong>in</strong>dicative of normal budget<strong>in</strong>g practices. 12<br />
This leaves us with a basel<strong>in</strong>e sample of 686 projects <strong>and</strong> $130 billion <strong>in</strong> planned<br />
total spend<strong>in</strong>g.<br />
11 See ? for a discussion of the <strong>in</strong>centives fac<strong>in</strong>g government contractors.<br />
12 These projects are a $45.5 billion project at the Department of Defense <strong>and</strong> a $19.5 billion<br />
project at the Department of Homel<strong>and</strong> Security; the next largest project is $3.9 billion <strong>and</strong> the<br />
average of the rema<strong>in</strong><strong>in</strong>g observations is $219 million. Because the dropped observations have above<br />
average overall rat<strong>in</strong>gs <strong>and</strong> are not from the last week of the year, omitt<strong>in</strong>g the observations works<br />
aga<strong>in</strong>st us f<strong>in</strong>d<strong>in</strong>g the effect predicted my our model.
CHAPTER 4. YEAR-END SPENDING 170<br />
Table 4.4 shows the year of orig<strong>in</strong>ation of these projects <strong>and</strong> the agencies at which<br />
they occurred. Almost two-thirds of these projects (64.6 percent) <strong>and</strong> half of the<br />
spend<strong>in</strong>g (50.3 percent) orig<strong>in</strong>ated <strong>in</strong> 2005 or later, although there are some ongo<strong>in</strong>g<br />
projects that orig<strong>in</strong>ated more than 20 years ago. 13 The projects are distributed<br />
fairly broadly across agencies. Although the Department of Defense, Department of<br />
Transportation, <strong>and</strong> Department of Veteran’s Affairs have higher levels of spend<strong>in</strong>g,<br />
the vast majority of the agencies have at least 10 projects (21 of 27) <strong>and</strong> at least $1<br />
billion <strong>in</strong> aggregate spend<strong>in</strong>g (20 of 27).<br />
The ma<strong>in</strong> performance measure tracked on the I.T. dashboard is the overall rat<strong>in</strong>g<br />
for the project. The canonical approach to track<strong>in</strong>g acquisitions is to measure cost,<br />
schedule, <strong>and</strong> performance. The overall rat<strong>in</strong>g therefore comb<strong>in</strong>es three sub<strong>in</strong>dexes.<br />
The cost rat<strong>in</strong>g sub<strong>in</strong>dex is based on the absolute percent deviation between the<br />
planned <strong>and</strong> actual cost of the project. Projects that are on average with<strong>in</strong> 5 percent<br />
of the scheduled cost receive a score of 10, projects that are with<strong>in</strong> 5 percent to 10<br />
percent on average receive a score of 9, <strong>and</strong> so on down to zero. Because the symmetric<br />
treatment of under <strong>and</strong> over-cost projects is somewhat unnatural, <strong>in</strong> our analysis we<br />
also construct an alternative <strong>in</strong>dex, “cost overrun” which gives under-cost projects<br />
the highest scores <strong>and</strong> over-cost projects the lowest. In this <strong>in</strong>dex, projects that are<br />
at least 45 percent under-cost receive a score of 10, projects that are 35 percent to 45<br />
percent under-cost receive a score of 9, <strong>and</strong> so on.<br />
The schedule rat<strong>in</strong>g sub<strong>in</strong>dex is based on the average tard<strong>in</strong>ess of the project<br />
across milestones. Projects that are no more than 30 days overdue on average receive<br />
a score of 10, projects that are no more than 90 days overdue on average receive a<br />
score of 5, <strong>and</strong> projects that are more than 90 days overdue on average receive a score<br />
of 0.<br />
The third sub<strong>in</strong>dex is a subjective CIO evaluation score, designed to <strong>in</strong>corporate<br />
an assessment of contract performance. The rat<strong>in</strong>g is <strong>in</strong>tended to reflect the CIO’s<br />
“assessment of the risk of the <strong>in</strong>vestment’s ability to accomplish its goals,” with<br />
the CIO <strong>in</strong>structed to “consult appropriate stakeholders <strong>in</strong> mak<strong>in</strong>g their evaluation,<br />
13 We address sample selection issues <strong>in</strong> the sensitivity section below.
CHAPTER 4. YEAR-END SPENDING 172<br />
To classify year-end projects, we use the date the first contract of the projects was<br />
signed, creat<strong>in</strong>g an <strong>in</strong>dicator variable for projects that orig<strong>in</strong>ated <strong>in</strong> the last seven<br />
days of September, the end of the fiscal year. Most I.T. projects are comprised of a<br />
series of contracts that are renewed <strong>and</strong> altered as milestones are met <strong>and</strong> the nature<br />
of the project evolves. We th<strong>in</strong>k that us<strong>in</strong>g the date the first contract was signed<br />
to classify the start date of the project is the best approach as the key structure of<br />
the project is most likely determ<strong>in</strong>ed at its onset. While future contract awards may<br />
affect the quality of the project, we only observe outcomes at the project level. We<br />
view any potential measurement error from our approach as <strong>in</strong>troduc<strong>in</strong>g downward<br />
bias <strong>in</strong> our coefficient of <strong>in</strong>terest as contracts <strong>in</strong>itially awarded before the last week<br />
of the year may be contam<strong>in</strong>ated by modifications made <strong>in</strong> the last week of a later<br />
year <strong>and</strong> contracts <strong>in</strong>itially awarded at the rush of year’s end may be rectified at a<br />
later po<strong>in</strong>t.<br />
Figure 4.3 shows the weekly pattern of spend<strong>in</strong>g <strong>in</strong> the I.T. Dashboard sample.<br />
As <strong>in</strong> the broader FPDS sample, there is a spike <strong>in</strong> spend<strong>in</strong>g <strong>in</strong> the 52nd week of the<br />
year. Spend<strong>in</strong>g <strong>and</strong> the number of projects <strong>in</strong> the last week <strong>in</strong>crease to 7.2 <strong>and</strong> 8.3<br />
times their rest-of-year weekly averages respectively. Alternatively put, while only<br />
account<strong>in</strong>g for 1.9 percent of the days of the year, the last week accounts for 12.3<br />
percent of spend<strong>in</strong>g <strong>and</strong> 14.0 percent of the number of projects. Activity is tilted<br />
even more strongly towards the last week if the sample of projects is restricted to<br />
the 65.1 percent of contracts that are for less than $100 million. Given the longer<br />
plann<strong>in</strong>g horizon for larger acquisitions, it is not surpris<strong>in</strong>g that we see more of a<br />
year-end spike for the smaller contracts. 17<br />
4.4.3 The Relative Quality of Year-End I.T. Contracts<br />
Figure 4.4 shows the distributions of the overall rat<strong>in</strong>g <strong>in</strong>dex for last-week-of-the-year<br />
projects <strong>and</strong> projects from the rest of the year. In these histograms, the rat<strong>in</strong>gs on<br />
the 0 to 10 scale are b<strong>in</strong>ned <strong>in</strong>to 5 categories with the lowest category represent<strong>in</strong>g<br />
overall rat<strong>in</strong>gs less than 2, the second lowest represent<strong>in</strong>g overall rat<strong>in</strong>gs between 2<br />
17 As <strong>in</strong> the broader FPDS sample, the end of year spike <strong>in</strong> the I.T. data is a broad phenomenon,<br />
not limited to a few agencies.
CHAPTER 4. YEAR-END SPENDING 174<br />
Recall that odds ratios capture the proportional change <strong>in</strong> the odds of a higher cat-<br />
egorical value associated with a unit <strong>in</strong>crease <strong>in</strong> the dependent variable, so that an<br />
odds ratio of 1/2 <strong>in</strong>dicates that the odds of a higher categorical value are 50 percent<br />
lower, or reciprocally that the odds of a lower categorical variable are 2 times as great.<br />
The results <strong>in</strong> this table are weighted by <strong>in</strong>flation-adjusted spend<strong>in</strong>g.<br />
The first column of the table shows the impact of a last week contract on the<br />
rat<strong>in</strong>g <strong>in</strong> a regression with no covariates. Columns 2 through 4 sequentially add <strong>in</strong><br />
fixed effects for year, agency, <strong>and</strong> project characteristics. In all of the specifications,<br />
the odds ratios are well below one—rang<strong>in</strong>g from 0.18 to 0.46—imply<strong>in</strong>g that last week<br />
spend<strong>in</strong>g is of significantly lower quality than spend<strong>in</strong>g <strong>in</strong> the rest of the year (the<br />
p-values are less than 0.01 <strong>in</strong> all specifications). The estimates imply that spend<strong>in</strong>g<br />
that orig<strong>in</strong>ates <strong>in</strong> the last week of the fiscal year has 2.2 to 5.6 times higher odds of<br />
hav<strong>in</strong>g a lower quality score.<br />
4.4.4 Sensitivity Analysis<br />
This subsection explores the robustness of the basic estimates. It shows how the<br />
results vary with different treatment of large contracts, with different functional form<br />
assumptions, <strong>and</strong> when selection <strong>in</strong>to the sample is modeled.<br />
Figure 4.4 showed that the f<strong>in</strong>d<strong>in</strong>g that year end projects are of lower quality was<br />
more pronounced <strong>in</strong> the dollar weighted analysis than <strong>in</strong> the unweighted analysis,<br />
suggest<strong>in</strong>g that a few large poor perform<strong>in</strong>g contracts may be heavily affect<strong>in</strong>g the<br />
results. The first four columns of Table 4.7 analyze this issue. The first two columns<br />
split the sample at the median contract size of $62 million. In the sample of smaller<br />
contracts, the coefficient of 0.60 is substantially below one but is less precisely es-<br />
timated (p-value of 0.17). The po<strong>in</strong>t estimate <strong>in</strong> column (3) from an unweighted<br />
regression is quite similar to the estimate <strong>in</strong> column (1) for the smaller contracts,<br />
but with added precision from doubl<strong>in</strong>g the sample size by <strong>in</strong>clud<strong>in</strong>g the full sam-<br />
ple (p-value of .02). Results <strong>in</strong> which we W<strong>in</strong>sorize the weights, assign<strong>in</strong>g a weight<br />
of $1 billion to the 4 percent of projects that are larger than $1 billion, are about<br />
half way between the full sample weighted <strong>and</strong> unweighted results (p-value less than
CHAPTER 4. YEAR-END SPENDING 175<br />
0.01). Overall, it is clear that the pattern of lower rat<strong>in</strong>g for end of year contracts is<br />
a broad phenomenon. It is also clear that the sample conta<strong>in</strong>s several very large low<br />
rated projects that were orig<strong>in</strong>ated <strong>in</strong> the last week of the year—possibly provid<strong>in</strong>g<br />
evidence that it is particularly risky to rush very large contracts out the door as the<br />
fiscal year deadl<strong>in</strong>e approaches.<br />
Column (5) of Table 4.4 shows results from an ord<strong>in</strong>ary least squares (OLS) model<br />
<strong>in</strong> which the raw overall rat<strong>in</strong>g is regressed on an <strong>in</strong>dicator for the contract orig<strong>in</strong>at<strong>in</strong>g<br />
<strong>in</strong> the last week of the year <strong>and</strong> on controls. The regression coefficient of -1.00 shows<br />
that I.T. spend<strong>in</strong>g contracted <strong>in</strong> the last week of the year receives rat<strong>in</strong>gs that are<br />
on average a full po<strong>in</strong>t lower on the 0 to 10 rat<strong>in</strong>g scale. This estimate also confirms<br />
that the f<strong>in</strong>d<strong>in</strong>g of lower quality year end spend<strong>in</strong>g is not limited to the ordered logit<br />
functional form.<br />
An important feature of our sample is that it reflects only active I.T. projects.<br />
Projects that have already been completed or projects that were term<strong>in</strong>ated without<br />
reach<strong>in</strong>g completion are not <strong>in</strong> our sample. Unfortunately, because the I.T. dashboard<br />
<strong>and</strong> the CIO rat<strong>in</strong>gs are br<strong>and</strong> new, it is not possible to acquire rat<strong>in</strong>g <strong>in</strong>formation<br />
on the major I.T. projects that are no longer ongo<strong>in</strong>g.<br />
Ideally, one would want a sample of all major I.T. projects that orig<strong>in</strong>ated <strong>in</strong> a<br />
particular period <strong>in</strong> time. The bias <strong>in</strong>troduced by the way <strong>in</strong> which our sample was<br />
constructed most likely leads us to underestimate the end-of-year-effect. In particular,<br />
very bad contracts begun <strong>in</strong> the last week of the year are likely to be canceled <strong>and</strong><br />
would not appear <strong>in</strong> our data set. Similarly, very well executed contracts from earlier<br />
<strong>in</strong> the year are likely to be completed ahead of schedule <strong>and</strong> also not appear <strong>in</strong> our<br />
data set. Thus, our estimates likely understate the gap <strong>in</strong> quality that we would f<strong>in</strong>d<br />
if we could compare all contracts from the last week of the year with all contracts<br />
from the rest of the year.<br />
To explore the extent of bias that a selection mechanism like the one just described<br />
might <strong>in</strong>troduce <strong>in</strong>to our estimates, we assembled a data set of all 3,859 major I.T.<br />
projects that orig<strong>in</strong>ated between 2002 <strong>and</strong> 2010. We were able to assemble this data<br />
set us<strong>in</strong>g the annual Exhibit 53 reports that allow the Office of Management <strong>and</strong><br />
Budget to track I.T. projects across the major federal agencies. These data show that
CHAPTER 4. YEAR-END SPENDING 176<br />
more recently orig<strong>in</strong>ated projects are significantly more likely to be <strong>in</strong> our sample.<br />
Our sample conta<strong>in</strong>s 85 percent of the total spend<strong>in</strong>g on projects that orig<strong>in</strong>ated <strong>in</strong><br />
2007 or later <strong>and</strong> only 28 percent of the spend<strong>in</strong>g on projects that orig<strong>in</strong>ated before<br />
this date.<br />
A simple way to assess whether there is selection is to estimate the model on<br />
samples split <strong>in</strong>to earlier <strong>and</strong> later years. A difference <strong>in</strong> the coefficient of <strong>in</strong>terest<br />
across samples, given the assumption that there is no time trend <strong>in</strong> the effect, would<br />
be <strong>in</strong>dicative of selection bias. Given this assumption, however, we can estimate the<br />
parameter of <strong>in</strong>terest exactly by us<strong>in</strong>g the date of project orig<strong>in</strong>ation to identify a<br />
selection correction term. Column (6) implements this strategy, show<strong>in</strong>g estimates<br />
from a st<strong>and</strong>ard Heckman selection model where the year or orig<strong>in</strong>ation is excluded<br />
from the second stage. The results show a larger effect than the correspond<strong>in</strong>g OLS<br />
estimate, but the lack of precession means that we cannot rule out that the effects<br />
are the same. 19 The negative coefficient on the selection term, although statistically<br />
<strong>in</strong>dist<strong>in</strong>guishable from zero, suggests that lower quality projects are on net more likely<br />
to rema<strong>in</strong> <strong>in</strong> the sample over time.<br />
4.4.5 Why Are Year End Contracts of Lower Quality?<br />
The results from the I.T. dashboard show that, consistent with the predictions of our<br />
model, year end spend<strong>in</strong>g is of lower quality than spend<strong>in</strong>g obligated earlier <strong>in</strong> the<br />
year. 20 Our model posited two channels: agencies may save low priority projects for<br />
the end of the year <strong>and</strong> undertake them only if they have no better uses for the funds,<br />
<strong>and</strong> the high volume of contract<strong>in</strong>g activity at the end of the year might allow for<br />
less management attention per project.<br />
19 Consistent with this f<strong>in</strong>d<strong>in</strong>g, OLS estimates on a sample split <strong>in</strong> 2007 show a larger po<strong>in</strong>t estimate<br />
<strong>in</strong> the later years, but we cannot reject the hypothesis that the coefficients are the same.<br />
20 In addition to the last week of the year results described above, we have also exam<strong>in</strong>ed last<br />
month of the year spend<strong>in</strong>g <strong>and</strong> f<strong>in</strong>d that spend<strong>in</strong>g <strong>in</strong> the balance of the last month of the year is of<br />
moderately lower quality than that <strong>in</strong> the first 11 months of the year. We have also exam<strong>in</strong>ed the<br />
quality of first week of the year spend<strong>in</strong>g (which also spikes). The po<strong>in</strong>t estimate for the first week<br />
of the year suggests somewhat higher spend<strong>in</strong>g quality, but the odds ratio difference from 1.0 is not<br />
statistically significant.
CHAPTER 4. YEAR-END SPENDING 177<br />
There is also a third possible mechanism. 21 Some program managers or contract<strong>in</strong>g<br />
officers may be <strong>in</strong>cl<strong>in</strong>ed toward procrast<strong>in</strong>ation <strong>and</strong> these same employees may be ones<br />
who do a worse job of plann<strong>in</strong>g for, writ<strong>in</strong>g, or manag<strong>in</strong>g contracts. Thus, we may<br />
see a surge of low quality contracts at the end of the year because that is when the<br />
least effective acquisition professionals get their contracts out the door. This third<br />
mechanism could have different policy implications from the first two because allow<strong>in</strong>g<br />
agencies to roll over funds would not necessarily improve outcomes—the least effective<br />
acquisition professionals would simply issue their contracts at a different time of year.<br />
One way to evaluate this procrast<strong>in</strong>ation hypothesis would be to exam<strong>in</strong>e the with<strong>in</strong>-<br />
year calendar distribution of other contracts issued by the contract<strong>in</strong>g officers who<br />
issued last week contracts <strong>in</strong> our I.T. Dashboard sample to see if they appear to be<br />
procrast<strong>in</strong>ators who consistently complete their contracts late <strong>in</strong> the year. While we<br />
can identify the contract<strong>in</strong>g officers responsible for a contract <strong>in</strong> both the dashboard<br />
sample <strong>and</strong> the FPDS sample, we cannot currently l<strong>in</strong>k the two samples, though we<br />
believe we ultimately will be able to do so.<br />
Another way to explore the possible mechanisms beh<strong>in</strong>d the poor outcomes for<br />
end-of-year contracts is to exam<strong>in</strong>e the subcomponents of the overall rat<strong>in</strong>g to see<br />
which sub<strong>in</strong>dices are responsible for the result. Appendix Table 4.10 repeats our ma<strong>in</strong><br />
ordered logit analysis with each sub<strong>in</strong>dex as the dependent variable. The results show<br />
clearly that it is the evaluation by the agency CIO that is responsible for the ma<strong>in</strong><br />
f<strong>in</strong>d<strong>in</strong>g. Neither the cost rat<strong>in</strong>g nor the schedule rat<strong>in</strong>g has an odds ratio that is<br />
significantly different from 1. The CIO evaluation shows that the odds of hav<strong>in</strong>g a<br />
higher rat<strong>in</strong>g are one-sixth as high for last-week-of-the-year contracts. The coefficient<br />
<strong>in</strong> the CIO regression is <strong>in</strong>sensitive to add<strong>in</strong>g the cost rat<strong>in</strong>g <strong>and</strong> schedul<strong>in</strong>g rat<strong>in</strong>g<br />
<strong>in</strong>to the regression, suggest<strong>in</strong>g that it is <strong>in</strong>formation <strong>in</strong> the CIO rat<strong>in</strong>g that is not<br />
<strong>in</strong>corporated <strong>in</strong> the other rat<strong>in</strong>g that is responsible for the result. This f<strong>in</strong>d<strong>in</strong>g is not<br />
all that surpris<strong>in</strong>g. As we mentioned above, the I.T. dashboard explicitly places more<br />
faith <strong>in</strong> the CIO’s assessment than <strong>in</strong> the other components by allow<strong>in</strong>g the CIO<br />
assessment to override the other components if it is lower than the other components.<br />
Moreover, the ability to reset milestone targets makes it difficult to assess the cost<br />
21 We thank Steve Kelman for suggest<strong>in</strong>g this third mechanism.
CHAPTER 4. YEAR-END SPENDING 178<br />
<strong>and</strong> schedule rat<strong>in</strong>gs. But while not surpris<strong>in</strong>g, the fact that it is the CIO evaluation<br />
that is driv<strong>in</strong>g the result means that we cannot learn much about the mechanism<br />
from the sub<strong>in</strong>dices, s<strong>in</strong>ce the CIO evaluation <strong>in</strong> a comprehensive measure of the I.T.<br />
project’s performance.<br />
Another way to explore possible mechanisms is to exam<strong>in</strong>e whether other ob-<br />
servable features of end-of-year contracts are different from those earlier <strong>in</strong> the year.<br />
Specifically, we exam<strong>in</strong>e whether features that policymakers often def<strong>in</strong>e as high risk—<br />
such as lack of competitive bidd<strong>in</strong>g or use of cost-reimbursement rather than fixed<br />
cost pric<strong>in</strong>g—are more prevalent <strong>in</strong> end of year contracts. For this analysis we return<br />
to the FPDS sample of all contracts from 2004 to 2009. To facilitate the analysis,<br />
we aggregate the 14.6 million observations up to the level of the covariates. We then<br />
estimate l<strong>in</strong>ear probability models with <strong>in</strong>dicators for contract characteristics (e.g.,<br />
a non-competitively sourced <strong>in</strong>dicator) as the dependent variable on an <strong>in</strong>dicator<br />
for last week of the fiscal year <strong>and</strong> controls. The regressions are weighted by total<br />
spend<strong>in</strong>g <strong>in</strong> each cell.<br />
The first three columns of Appendix Table 4.11 exam<strong>in</strong>e shifts <strong>in</strong> the degree of<br />
competitive sourc<strong>in</strong>g at the end of the year. The use of non-competitive contracts<br />
shows little change. However, contracts that are competitively sourced are signifi-<br />
cantly more likely to receive only one bid perhaps because the end of year rush leaves<br />
less time to allow bidd<strong>in</strong>g to take place. The estimates <strong>in</strong>dicate that there is almost<br />
a 10 percent <strong>in</strong>crease <strong>in</strong> the percent of contracts receiv<strong>in</strong>g only a s<strong>in</strong>gle bid—a 1.7<br />
percentage po<strong>in</strong>t <strong>in</strong>crease one a base of 20 percent. On net, then, there is a modest<br />
<strong>in</strong>crease <strong>in</strong> “risky” non-competitive <strong>and</strong> one bid contracts at the end of the year.<br />
Column (3) shows that spend<strong>in</strong>g on contracts that are either non-competitive or one<br />
bid contract <strong>in</strong>creases by 1 percentage po<strong>in</strong>ts on a base of 49 percent <strong>in</strong> the last week<br />
of the year.<br />
The second three columns <strong>in</strong>vestigate the type of contract used. Contracts that<br />
provide for cost reimbursement rather than specify<strong>in</strong>g a fixed price are often seen as<br />
high risk because they have significant potential for cost overruns. Time <strong>and</strong> material<br />
or labor hours (T&M/LH) contracts raise similar concerns because they <strong>in</strong>volve open<br />
ended commitments to pay for whatever quantity of labor <strong>and</strong> materials are used
CHAPTER 4. YEAR-END SPENDING 179<br />
to accomplish the task specified <strong>in</strong> the contract. Column (4) shows that end-of-year<br />
contracts are less likely to <strong>in</strong>clude these sorts of high risk contract terms possibly<br />
because agencies are attempt<strong>in</strong>g to use up def<strong>in</strong>ed sums of excess funds <strong>and</strong> therefore<br />
limit contracts to the available amounts. The use of T&M/LH contracts <strong>in</strong>creases<br />
by 0.4 percentage po<strong>in</strong>ts, which is substantial compared to a base of 5.5 percent.<br />
Because T&M/LH contracts are <strong>in</strong>frequent, column (6) shows a net decrease <strong>in</strong> the<br />
comb<strong>in</strong>ed use of risky cost-reimbursement <strong>and</strong> T&M/LH contract spend<strong>in</strong>g of about<br />
3 percentage po<strong>in</strong>ts on a base of 36 percent.<br />
Overall, the analysis <strong>in</strong> this section does not offer any clear <strong>in</strong>sights <strong>in</strong>to what<br />
might be caus<strong>in</strong>g lower performance among end-of-year contracts. There is no reason<br />
to expect it to be a s<strong>in</strong>gle mechanism. All three mechanisms could be re<strong>in</strong>forc<strong>in</strong>g<br />
each other <strong>in</strong> contribut<strong>in</strong>g to the phenomenon.<br />
4.5 Do Rollover Provisions Raise Spend<strong>in</strong>g Qual-<br />
ity?<br />
The third prediction of the model is that allow<strong>in</strong>g for the rollover of unused fund<strong>in</strong>g<br />
unambiguously improves quality, both overall <strong>and</strong> at year’s end. Intuitively, organi-<br />
zations are less likely to engage <strong>in</strong> wasteful year-end spend<strong>in</strong>g when the fund<strong>in</strong>g could<br />
be used for higher value projects <strong>in</strong> the next budget period.<br />
As we noted <strong>in</strong> the <strong>in</strong>troduction, s<strong>in</strong>ce 1992 the Department of Justice (DOJ)<br />
has had special authority to roll over unused funds to pay for future I.T. needs. In<br />
this section of the paper, we exam<strong>in</strong>e whether the quality of DOJ’s end-of-year I.T.<br />
spend<strong>in</strong>g is higher than that of federal agencies who lack rollover authority.<br />
4.5.1 The DOJ’s Rollover Authority<br />
The DOJ authority provides that “unobligated balances of appropriations available<br />
to the Department of Justice dur<strong>in</strong>g such fiscal year may be transferred <strong>in</strong>to the<br />
capital account of the Work<strong>in</strong>g Capital Fund to be available for the department-wide<br />
acquisition of capital equipment, development <strong>and</strong> implementation of law enforcement
CHAPTER 4. YEAR-END SPENDING 180<br />
or litigation related automated data process<strong>in</strong>g systems, <strong>and</strong> for the improvement<br />
<strong>and</strong> implementation of the Department’s f<strong>in</strong>ancial management <strong>and</strong> payroll/personnel<br />
systems.” 22 While other agencies have ongo<strong>in</strong>g work<strong>in</strong>g capital funds, appropriated<br />
funds contributed to those funds reta<strong>in</strong> their fiscal year restrictions.<br />
Between 1992 <strong>and</strong> 2006, approximately $1.8 billion <strong>in</strong> annual appropriations were<br />
transferred to the DOJ work<strong>in</strong>g capital fund from unused appropriations balances (?).<br />
Nonetheless, Table 4.2 shows that DOJ has an end-of-year spend<strong>in</strong>g surge comparable<br />
to that of other agencies when all spend<strong>in</strong>g is taken <strong>in</strong>to account, with 9.4 percent of<br />
its spend<strong>in</strong>g occurr<strong>in</strong>g <strong>in</strong> the last week of the year.<br />
Even with rollover authority, there rema<strong>in</strong> <strong>in</strong>centives for agencies to use up their<br />
full allocation of fund<strong>in</strong>g. Large balances carried over from one period to another<br />
are likely to be <strong>in</strong>terpreted by OMB <strong>and</strong> Congressional appropriators as a signal that<br />
budget resources are excessive <strong>and</strong> lead to reduced budgets <strong>in</strong> subsequent periods.<br />
For example, Senator Coburn issued a report <strong>in</strong> 2008 entitled “Justice Denied: Waste<br />
<strong>and</strong> Management at the Department of Justice” <strong>in</strong> which he stated: “Every year<br />
Congress appropriates more than $20 billion for the Department of Justice to carry<br />
out its mission, <strong>and</strong> every year the Department ends the year with billions of unspent<br />
dollars. But <strong>in</strong>stead of return<strong>in</strong>g this unneeded <strong>and</strong> unspent money to the taxpayers,<br />
DOJ rolls it over year to year, essentially ma<strong>in</strong>ta<strong>in</strong><strong>in</strong>g a billion dollar bank account<br />
that it can dip <strong>in</strong>to for projects for which the money was not orig<strong>in</strong>ally <strong>in</strong>tended”<br />
(?). 23<br />
It is not just external pressure that may lead an <strong>organization</strong> to spend all of its<br />
resources even <strong>in</strong> the presence of rollover authority. Components of an agency may<br />
not be will<strong>in</strong>g to return resources to the center if they are not ensured of be<strong>in</strong>g able<br />
to spend those resources after they are rolled over. Indeed, <strong>in</strong> 2006 testimony before<br />
22 Public Law 102-104: 28 USC 527 note.<br />
23 Coburn goes on to add “. . . perhaps without the pressure to rush to spend funds before<br />
they are canceled, DOJ may, <strong>in</strong> fact, make more prudent spend<strong>in</strong>g decisions with unobligated funds.<br />
This has not been studied <strong>and</strong> warrants exam<strong>in</strong>ation for potential cost sav<strong>in</strong>gs across the federal<br />
government. As long as DOJ is bank<strong>in</strong>g billions of dollars from year to year that the Department<br />
has some discretion to spend on its priorities as they arise, however, Congress should more carefully<br />
review how much it is appropriat<strong>in</strong>g for DOJ programs. If a particular <strong>in</strong>itiative or office does not<br />
need or spend as much as Congress has appropriated, then Congress should consider appropriat<strong>in</strong>g<br />
less for that particular office <strong>and</strong> the Department overall.”
CHAPTER 4. YEAR-END SPENDING 181<br />
Congress, Deputy Assistant Attorney General Lee Lofthus may have been suggest<strong>in</strong>g<br />
a connection between those components of DOJ that contribute expir<strong>in</strong>g funds to the<br />
work<strong>in</strong>g capital fund <strong>and</strong> those that benefit from it when he noted that the FBI was<br />
both the biggest contributor to the fund <strong>and</strong> the biggest beneficiary (?).<br />
Given that rollover funds at DOJ are used for I.T. purposes, one might expect<br />
to see little or no end-of-year spike <strong>in</strong> I.T. spend<strong>in</strong>g, because I.T. components of<br />
the agency will know that they will directly benefit from funds contributed to the<br />
work<strong>in</strong>g capital fund. This is <strong>in</strong>deed the case. In the comprehensive FPDS data,<br />
only 3.4 percent of DOJ’s I.T. spend<strong>in</strong>g occurs <strong>in</strong> the last week—the 19th lowest of<br />
the 21 major agencies. In the I.T. dashboard data, DOJ has 16 ongo<strong>in</strong>g major I.T.<br />
<strong>in</strong>vestments, with planned total spend<strong>in</strong>g of just over $5.1 billion. Only 1 <strong>in</strong>vestment<br />
occurred <strong>in</strong> the last week of year. The $99 million cost of this <strong>in</strong>vestment implies<br />
that 1.9 percent of DOJ spend<strong>in</strong>g occurred <strong>in</strong> the last week compared to an average<br />
of 11.0 percent across the other agencies.<br />
Although the sample size of <strong>in</strong>vestments is very small, the quality of DOJ’s year-<br />
end spend<strong>in</strong>g is high. The one last week <strong>in</strong>vestment has the highest quality score of all<br />
major I.T. <strong>in</strong>vestments at the agency. Difference-<strong>in</strong>-differences estimates presented <strong>in</strong><br />
Table 4.8, which allow us to control for <strong>in</strong>vestment characteristics, suggest that the<br />
DOJ pattern is sufficiently unusual to be statistically significant—even though it is<br />
identified off of a s<strong>in</strong>gle end-of-year observation. Specifically, difference-<strong>in</strong>-differences<br />
po<strong>in</strong>t estimates <strong>in</strong>dicate that year-end quality <strong>in</strong>creases by 2.0 to 3.5 categorical levels<br />
at DOJ relative to other agencies. The p-values on the <strong>in</strong>teraction variable are less<br />
than 0.01 <strong>in</strong> all specifications.<br />
Despite the statistical <strong>and</strong> economic significance of the estimates, we are hesitant<br />
to draw strong conclusions from the estimated effect. In addition to the small number<br />
of projects, the fact that the effect is identified off of a s<strong>in</strong>gle agency raises the potential<br />
for bias from unobserved <strong>organization</strong>al characteristics. Nevertheless, the evidence<br />
that exists appears consistent with the prediction that rollover <strong>in</strong>creases year-end<br />
quality.
CHAPTER 4. YEAR-END SPENDING 182<br />
4.6 Conclusion<br />
Our model of an <strong>organization</strong> fac<strong>in</strong>g a fixed period <strong>in</strong> which it must spend its budget<br />
resources made three predictions. We have confirmed all three of them us<strong>in</strong>g data on<br />
U.S. federal contract<strong>in</strong>g. First, there is a surge of spend<strong>in</strong>g at the end of the year.<br />
Second, end of year spend<strong>in</strong>g is of lower quality. Third, permitt<strong>in</strong>g the rollover of<br />
spend<strong>in</strong>g <strong>in</strong>to subsequent periods leads to higher quality.<br />
Because we cannot identify the exact mechanism produc<strong>in</strong>g the decl<strong>in</strong>e <strong>in</strong> spend<strong>in</strong>g<br />
quality at the end of the year, it is difficult to draw firm policy conclusions from the<br />
result. If the low spend<strong>in</strong>g quality comes from agencies squ<strong>and</strong>er<strong>in</strong>g end of year<br />
resources on low priority projects, possibly compounded by <strong>in</strong>sufficient management<br />
attention dur<strong>in</strong>g the end of year spend<strong>in</strong>g rush, then allow<strong>in</strong>g for rollover of unused<br />
balances or switch<strong>in</strong>g to two-year budget<strong>in</strong>g might improve spend<strong>in</strong>g quality. But as<br />
long as future agency budgets are based <strong>in</strong> part on whether agencies exhausted their<br />
resources <strong>in</strong> the current period, there will still be an <strong>in</strong>centive for year-end spend<strong>in</strong>g<br />
surges. And unless the rollover balances stay with the same part of the <strong>organization</strong><br />
that managed to save them, agency subcomponents will still have an <strong>in</strong>centive to<br />
use up their allocations. An alternative approach would be to apply greater scrut<strong>in</strong>y<br />
to end-of-the-year spend<strong>in</strong>g with the presumption that any spend<strong>in</strong>g above levels<br />
occurr<strong>in</strong>g earlier <strong>in</strong> the year was unwarranted. This latter approach could also be the<br />
proper management prescription if low quality end-of-year spend<strong>in</strong>g results from a<br />
correlation between acquisition officer skill <strong>and</strong> a tendency to procrast<strong>in</strong>ate—agencies<br />
would want to give greater attention to end of year contracts because the acquisition<br />
officials responsible for them need more oversight.<br />
In evaluat<strong>in</strong>g possible policy reforms one should not lose sight of the potential<br />
benefits of one-year budget periods. The annual appropriations cycle may provide<br />
benefits from greater Congressional control over executive branch operations. More-<br />
over, the use-it-or-lose it feature of appropriated funds may push projects out the<br />
door that would otherwise languish due to bureaucratic delays.
CHAPTER 4. YEAR-END SPENDING 183<br />
Figure 4.1: Federal Contract<strong>in</strong>g by Week, Pooled 2004 to 2009 FPDS<br />
Spend<strong>in</strong>g (billions)<br />
Number of contracts (thous<strong>and</strong>s)<br />
$250<br />
$200<br />
$150<br />
$100<br />
$50<br />
$-<br />
800<br />
600<br />
400<br />
200<br />
1 2 3 4 5 6 7 8 910111213141516171819202122232425262728293031323334353637383940414243444546474849505152<br />
Week<br />
(a) Spend<strong>in</strong>g<br />
1 2 3 4 5 6 7 8 910111213141516171819202122232425262728293031323334353637383940414243444546474849505152<br />
Week<br />
(b) Number of Contracts<br />
Source: Federal Procurement Data System, accessed October, 2010 via<br />
www.usaspend<strong>in</strong>g.gov.<br />
Note: Total spend<strong>in</strong>g <strong>and</strong> number of contracts by week of the fiscal year. Spend<strong>in</strong>g values<br />
<strong>in</strong>flation-adjusted to 2009 dollars us<strong>in</strong>g the CPI-U.
CHAPTER 4. YEAR-END SPENDING 184<br />
20 25 30 35<br />
Last quarter spend<strong>in</strong>g (percent)<br />
Non-defense<br />
2005<br />
2001<br />
2007<br />
2004<br />
Figure 4.2: Year-End Spend<strong>in</strong>g by Appropriations Date<br />
2003<br />
2006<br />
2008<br />
2002<br />
2000<br />
2000<br />
2005<br />
2008<br />
2001<br />
2006<br />
2002<br />
2004<br />
2007<br />
2003<br />
2009<br />
Defense<br />
-10 0 10 20 30<br />
Weeks late<br />
2009<br />
Non-defense slope = .193 (.094)<br />
Defense slope = .180 (.053)<br />
Pooled slope = .183 (.044)<br />
(a) Last Quarter Spend<strong>in</strong>g<br />
6 8 10 12 14 16<br />
Last week spend<strong>in</strong>g (percent)<br />
Non-defense slope = .112 (.146)<br />
Defense slope = .081 (.045)<br />
Pooled slope = .089 (.047)<br />
2005<br />
2007<br />
Defense<br />
2004<br />
2006<br />
2008<br />
2005<br />
10 15 20<br />
Last month spend<strong>in</strong>g (percent)<br />
2008<br />
2006<br />
2001<br />
2005<br />
Non-defense<br />
2006<br />
2003<br />
2007<br />
2004<br />
2008<br />
2002<br />
2008<br />
2000<br />
20052001<br />
2006<br />
2000<br />
2002<br />
2004<br />
2003<br />
2007<br />
2009<br />
Defense<br />
-10 0 10 20 30<br />
Weeks late<br />
2004<br />
2007<br />
2009<br />
Non-defense slope = .130 (.078)<br />
Defense slope = .089 (.030)<br />
Pooled slope = .100 (.029)<br />
(b) Last Month Spend<strong>in</strong>g<br />
Non-defense<br />
-10 0 10 20 30<br />
Weeks late<br />
(c) Last Week Spend<strong>in</strong>g<br />
Source: Federal Procurement Data System, accessed October, 2010 via<br />
www.usaspend<strong>in</strong>g.gov <strong>and</strong> Library of Congress.<br />
Note: Vertical axes show the percent of annual spend<strong>in</strong>g occur<strong>in</strong>g <strong>in</strong> the last quarter,<br />
month, <strong>and</strong> week of the fiscal year. Horizontal axes shows the passage dates for the<br />
non-defense <strong>and</strong> defense appropriation bills, relative to the first day of the fiscal year <strong>in</strong><br />
weeks. For defense spend<strong>in</strong>g, weeks late measures the date that the defense appropriations<br />
bill was enacted. For non-defense spend<strong>in</strong>g the date is assigned from the date of the<br />
consolidated appropriations act, or, <strong>in</strong> the case of the two years <strong>in</strong> which there was not a<br />
consolidated act, a date that is the midpo<strong>in</strong>t of the <strong>in</strong>dividual non-defense appropriations<br />
acts. Plots show fitted l<strong>in</strong>es <strong>and</strong> slope coefficients from bivariate regressions on defense<br />
<strong>and</strong> non-defense spend<strong>in</strong>g. Pooled coefficients from a regression where defense <strong>and</strong><br />
non-defense spend<strong>in</strong>g have different <strong>in</strong>tercepts but are constra<strong>in</strong>ed to have the same slope.<br />
Robust st<strong>and</strong>ard errors <strong>in</strong> parentheses.<br />
2009<br />
2009
CHAPTER 4. YEAR-END SPENDING 185<br />
Spend<strong>in</strong>g (millions)<br />
Number of projects<br />
$16,000<br />
$12,000<br />
100<br />
$8,000<br />
$4,000<br />
80<br />
60<br />
40<br />
20<br />
0<br />
$0<br />
Figure 4.3: I.T. Contract<strong>in</strong>g by Week<br />
1 2 3 4 5 6 7 8 910111213141516171819202122232425262728293031323334353637383940414243444546474849505152<br />
1 2 3 4 5 6 7<br />
8 9101112131415<br />
16 171819202122<br />
Weeks<br />
(a) Spend<strong>in</strong>g<br />
23 24252627282930<br />
Weeks<br />
31 323334353637<br />
(b) Number of Projects<br />
38 39404142434445<br />
46 474849505152<br />
Source: I.T. Dashboard data, accessed March, 2010 via http://it.usaspend<strong>in</strong>g.gov.<br />
Note: Total spend<strong>in</strong>g <strong>and</strong> number of I.T projects by week of the fiscal year. Spend<strong>in</strong>g<br />
values <strong>in</strong>flation-adjusted to 2009 dollars us<strong>in</strong>g the CPI-U.
CHAPTER 4. YEAR-END SPENDING 186<br />
Figure 4.4: Year-End <strong>and</strong> Rest-of-Year Overall Rat<strong>in</strong>gs<br />
40.0%<br />
30.0%<br />
20.0%<br />
10.0%<br />
0.0%<br />
45.0%<br />
30.0%<br />
15.0%<br />
0.0%<br />
25.8%<br />
8.5%<br />
Last week<br />
Rest of year<br />
3.0%<br />
22.9%<br />
5.5%<br />
11.0%<br />
32.4%<br />
15.5%<br />
36.5%<br />
24.8%<br />
1 2 3 4 5<br />
Last week<br />
Rest of year<br />
2.9%<br />
2.1%<br />
2.8%<br />
Overall Rat<strong>in</strong>g<br />
(a) Spend<strong>in</strong>g<br />
25.5%<br />
24.4%<br />
37.2%<br />
40.6%<br />
26.6%<br />
1 2 3 4 5<br />
Overall Rat<strong>in</strong>g<br />
(b) Number of projects<br />
Source: I.T. Dashboard data, accessed March, 2010 via http://it.usaspend<strong>in</strong>g.gov.<br />
Note: Overall rat<strong>in</strong>g histograms for I.T. projects orig<strong>in</strong>at<strong>in</strong>g <strong>in</strong> the last week <strong>and</strong> rest of<br />
the year. To construct this figure, rat<strong>in</strong>gs are b<strong>in</strong>ned <strong>in</strong>to 5 categories with the lowest<br />
category represent<strong>in</strong>g overall rat<strong>in</strong>gs less than 2, the second lowest represent<strong>in</strong>g overall<br />
rat<strong>in</strong>gs between 2 <strong>and</strong> 4, <strong>and</strong> so on. See text for details on the overall rat<strong>in</strong>g <strong>in</strong>dex. Panel<br />
A weights projects by <strong>in</strong>flation-adjusted spend<strong>in</strong>g. Panel B shows unweighted values.<br />
22.6%<br />
29.3%
CHAPTER 4. YEAR-END SPENDING 187<br />
Table 4.1: Summary Statistics: Federal Contract<strong>in</strong>g, Pooled 2004 to 2009 FPDS<br />
Spend<strong>in</strong>g<br />
Contracts<br />
Billions Percent Count Percent<br />
Totals $2,597 100.0% 14,600,000 100.0%<br />
Year<br />
2004 $304 11.7% 1,413,316 9.7%<br />
2005 $355 13.7% 1,857,960 12.7%<br />
2006 $405 15.6% 2,719,479 18.6%<br />
2007 $452 17.4% 2,977,426 20.4%<br />
2008 $542 20.9% 3,292,063 22.5%<br />
2009 $538 20.7% 2,307,904 15.8%<br />
Contract size<br />
Less than $100K $166 6.4% 13,800,000 94.5%<br />
$100K to $1M $398 15.3% 626,134 4.3%<br />
At least $1M $2,033 78.3% 98,001 0.7%<br />
Agency<br />
Agriculture $25 1.0% 241,626 1.7%<br />
Commerce $13 0.5% 112,756 0.8%<br />
Defense $1,824 70.2% 3,536,530 24.2%<br />
Education $8 0.3% 12,806 0.1%<br />
Energy $142 5.5% 37,756 0.3%<br />
Environmental Protection Agency $8 0.3% 62,713 0.4%<br />
General Services Adm<strong>in</strong>istration $82 3.2% 4,830,748 33.1%<br />
Health <strong>and</strong> Human Services $76 2.9% 249,907 1.7%<br />
Homel<strong>and</strong> Security $74 2.8% 255,461 1.7%<br />
Hous<strong>in</strong>g <strong>and</strong> Urban Development $6 0.2% 15,666 0.1%<br />
Interior $25 1.0% 377,743 2.6%<br />
Justice $33 1.3% 420,379 2.9%<br />
Labor $13 0.5% 41,229 0.3%<br />
National Aeronautics <strong>and</strong> Space Adm<strong>in</strong>istration $83 3.2% 81,211 0.6%<br />
National Science Foundation $2 0.1% 4,201 0.0%<br />
Other $37 1.4% 179,283 1.2%<br />
Small Bus<strong>in</strong>ess Adm<strong>in</strong>istration $0 0.0% 3,361 0.0%<br />
State $34 1.3% 239,019 1.6%<br />
Transportation $21 0.8% 57,235 0.4%<br />
Treasury $25 1.0% 177,662 1.2%<br />
Veterans Affairs $67 2.6% 3,630,856 24.9%<br />
Competition type<br />
Non-competitive $745 28.7% 3,553,453 24.3%<br />
Competitive with one bid $521 20.0% 3,883,273 26.6%<br />
Competitive with more than one bid $1,332 51.3% 7,131,422 48.8%<br />
Contract type<br />
Fixed price $1,675 64.5% 14,200,000 97.3%<br />
Cost-reimbursement $780 30.0% 151,362 1.0%<br />
Time <strong>and</strong> materials/labor hours $142 5.5% 249,705 1.7%<br />
Source: Federal Procurement Data System, accessed October, 2010 via www.usaspend<strong>in</strong>g.gov<br />
Note: Contract spend<strong>in</strong>g <strong>in</strong>flation adjusted to 2009 dollars us<strong>in</strong>g the CPI-U.
CHAPTER 4. YEAR-END SPENDING 188<br />
Table 4.2: Year-End Contract Spend<strong>in</strong>g by Agency, Pooled 2004 to 2009 FPDS<br />
Spend<strong>in</strong>g<br />
Percent of spend<strong>in</strong>g<br />
(billions) Last month Last week<br />
Agriculture $24.8 17.0% 6.2%<br />
Commerce $13.4 21.4% 5.6%<br />
Defense $1,820.0 16.0% 8.6%<br />
Education $8.2 18.6% 11.2%<br />
Energy $142.0 6.6% 4.0%<br />
Environmental Protection Agency $8.1 22.3% 10.4%<br />
General Services Adm<strong>in</strong>istration $82.0 12.9% 7.0%<br />
Health <strong>and</strong> Human Services $76.4 25.5% 12.2%<br />
Homel<strong>and</strong> Security $73.6 22.7% 9.4%<br />
Hous<strong>in</strong>g <strong>and</strong> Urban Development $5.7 18.5% 11.7%<br />
Interior $25.3 23.2% 7.6%<br />
Justice $32.6 17.9% 9.4%<br />
Labor $12.7 12.9% 5.9%<br />
National Aeronautics <strong>and</strong> Space Adm<strong>in</strong>istration $82.7 16.9% 11.0%<br />
National Science Foundation $2.0 27.7% 11.5%<br />
Small Bus<strong>in</strong>ess Adm<strong>in</strong>istration $0.4 31.9% 16.3%<br />
State $33.5 34.9% 20.4%<br />
Transportation $20.5 17.6% 3.6%<br />
Treasury $24.9 15.3% 9.6%<br />
Veterans Affairs $66.9 18.2% 9.5%<br />
Other $37.4 28.6% 18.9%<br />
Total $2,600.0 16.5% 8.7%<br />
Source: Federal Procurement Data System, accessed October, 2010 via www.usaspend<strong>in</strong>g.gov.<br />
Note: Contract spend<strong>in</strong>g <strong>in</strong>flation adjusted to 2009 dollars us<strong>in</strong>g the CPI-U.
CHAPTER 4. YEAR-END SPENDING 189<br />
Table 4.3: Year-End Contract Spend<strong>in</strong>g By Selected Product or Service Code, Pooled<br />
2004 to 2009 FPDS<br />
Spend<strong>in</strong>g<br />
Percent of spend<strong>in</strong>g<br />
(billions) Last month Last week<br />
Construction-related<br />
Construction of structures <strong>and</strong> facilities $136.0 40.9% 28.6%<br />
Ma<strong>in</strong>tenance, repair, or alteration of real property $72.5 34.8% 20.1%<br />
Architect <strong>and</strong> eng<strong>in</strong>eer<strong>in</strong>g services $32.8 26.1% 13.8%<br />
Installation of equipment $4.0 33.9% 20.4%<br />
Prefabricated structures <strong>and</strong> scaffold<strong>in</strong>g $3.7 34.9% 18.4%<br />
Furnish<strong>in</strong>gs <strong>and</strong> office equipment<br />
Furniture $8.0 37.3% 18.4%<br />
Office supplies <strong>and</strong> devices $4.0 24.9% 16.6%<br />
Household <strong>and</strong> commercial furnish<strong>in</strong>gs <strong>and</strong> appliances $1.2 37.8% 20.7%<br />
Office mach<strong>in</strong>es, text process<strong>in</strong>g systems <strong>and</strong> equipment $1.1 33.5% 17.0%<br />
I.T. services <strong>and</strong> equipment<br />
Automatic data process<strong>in</strong>g <strong>and</strong> telecom. services $145.0 21.0% 12.3%<br />
Automatic data process<strong>in</strong>g equipment $53.7 29.2% 14.9%<br />
Services<br />
Professional, adm<strong>in</strong>, <strong>and</strong> management support services $336.0 19.1% 9.9%<br />
Research <strong>and</strong> development $309.0 11.3% 5.3%<br />
Utilities <strong>and</strong> housekeep<strong>in</strong>g services $73.7 15.6% 9.1%<br />
Ongo<strong>in</strong>g<br />
Fuels, lubricants, oils <strong>and</strong> waxes $72.7 13.2% 0.7%<br />
Medical services $68.8 4.9% 1.7%<br />
Chemicals <strong>and</strong> chemical products $6.2 3.3% 1.3%<br />
Tires <strong>and</strong> tubes $1.0 8.7% 2.7%<br />
Toiletries $0.3 12.2% 3.0%<br />
Military weapons systems<br />
Aircraft <strong>and</strong> airframe structural components $141.0 5.7% 2.9%<br />
Ships, small craft, pontoons, <strong>and</strong> float<strong>in</strong>g docks $48.5 7.5% 2.1%<br />
Guided missiles $38.0 8.1% 3.5%<br />
Other $1,111.6 13.6% 6.8%<br />
Total $2,600.0 16.5% 8.7%<br />
Source: Federal Procurement Data System, accessed October, 2010 via www.usaspend<strong>in</strong>g.gov.<br />
Note: Contract spend<strong>in</strong>g <strong>in</strong> the last month <strong>and</strong> week of the fiscal year by selected 2-digit product or service code, <strong>in</strong>flation<br />
adjusted to 2009 dollars us<strong>in</strong>g the CPI-U. Categories jo<strong>in</strong>tly account for 57.2 percent of total spend<strong>in</strong>g.
CHAPTER 4. YEAR-END SPENDING 190<br />
Table 4.4: Summary Statistics: Major I.T. Projects as of March, 2010<br />
IT spend<strong>in</strong>g<br />
IT projects<br />
Millions Percent Count Percent<br />
Total $129,729 100.0% 686 100.0%<br />
Agency<br />
Agency for International Development $265 0.2% 3 0.4%<br />
Agriculture $1,864 1.4% 33 4.8%<br />
Commerce $11,042 8.5% 46 6.7%<br />
Corps of Eng<strong>in</strong>eers $4,012 3.1% 11 1.6%<br />
Defense $14,889 11.5% 46 6.7%<br />
Education $1,407 1.1% 25 3.6%<br />
Energy $4,914 3.8% 26 3.8%<br />
Environmental Protection Agency $3,166 2.4% 20 2.9%<br />
General Services Adm<strong>in</strong>istration $2,162 1.7% 25 3.6%<br />
Health <strong>and</strong> Human Services $8,990 6.9% 64 9.3%<br />
Homel<strong>and</strong> Security $13,068 10.1% 70 10.2%<br />
Hous<strong>in</strong>g <strong>and</strong> Urban Development $1,605 1.2% 10 1.5%<br />
Interior $4,557 3.5% 39 5.7%<br />
Justice $4,376 3.4% 15 2.2%<br />
Labor $2,434 1.9% 34 5.0%<br />
National Aeronautics <strong>and</strong> Space Adm<strong>in</strong>istration $9,722 7.5% 22 3.2%<br />
National Archives <strong>and</strong> Records Adm<strong>in</strong>istration $649 0.5% 8 1.2%<br />
National Science Foundation $374 0.3% 6 0.9%<br />
Nuclear Regulatory Commission $515 0.4% 16 2.3%<br />
Office of Personnel Management $497 0.4% 7 1.0%<br />
Small Bus<strong>in</strong>ess Adm<strong>in</strong>istration $269 0.2% 9 1.3%<br />
Smithsonian Institution $58 0.0% 9 1.3%<br />
Social Security Adm<strong>in</strong>istration $1,236 1.0% 13 1.9%<br />
State $3,705 2.9% 13 1.9%<br />
Transportation $12,514 9.6% 42 6.1%<br />
Treasury $4,921 3.8% 41 6.0%<br />
Veterans Affairs $16,521 12.7% 33 4.8%<br />
Year of orig<strong>in</strong>ation<br />
1981 $2,706 2.1% 1 0.1%<br />
1991 $61 0.0% 1 0.1%<br />
1992 $322 0.2% 1 0.1%<br />
1993 $409 0.3% 2 0.3%<br />
1994 $155 0.1% 2 0.3%<br />
1996 $3,050 2.4% 7 1.0%<br />
1997 $1,430 1.1% 3 0.4%<br />
1998 $2,891 2.2% 5 0.7%<br />
1999 $2,814 2.2% 10 1.5%<br />
2000 $2,855 2.2% 15 2.2%<br />
2001 $8,463 6.5% 17 2.5%<br />
2002 $12,577 9.7% 32 4.7%<br />
2003 $13,860 10.7% 60 8.7%<br />
2004 $12,818 9.9% 87 12.7%<br />
2005 $13,529 10.4% 95 13.8%<br />
2006 $16,169 12.5% 126 18.4%<br />
2007 $17,935 13.8% 121 17.6%<br />
2008 $14,176 10.9% 75 10.9%<br />
2009 $3,508 2.7% 26 3.8%<br />
Source: I.T. Dashboard data, accessed March, 2010 via http://it.usaspend<strong>in</strong>g.gov<br />
Notes: Major I.T. <strong>in</strong>vestments by federal agency <strong>and</strong> year of orig<strong>in</strong>ation, <strong>in</strong>flation adjusted to 2009 dollars us<strong>in</strong>g the CPI-U.
CHAPTER 4. YEAR-END SPENDING 191<br />
Table 4.5: Summary Statistics: Quality Indexes <strong>and</strong> Project Characteristics for Major<br />
I.T. Projects<br />
Mean Std. Dev. M<strong>in</strong> Max<br />
Planned cost (millions) 189.11 447.06 0.10 4770.89<br />
Overall rat<strong>in</strong>g 7.07 2.30 0.00 10.00<br />
Rat<strong>in</strong>g sub<strong>in</strong>dexes<br />
CIO evaluation 3.95 0.94 1.00 5.00<br />
Cost rat<strong>in</strong>g 8.72 2.52 0.00 10.00<br />
Cost overrun 5.25 1.49 0.00 10.00<br />
Schedule rat<strong>in</strong>g 8.43 3.09 0.00 10.00<br />
Count Percent<br />
Investment phase<br />
Full-Acquisition 59 8.6%<br />
Mixed Life Cycle 304 44.3%<br />
Multi-Agency Collaboration 29 4.2%<br />
Operations <strong>and</strong> Ma<strong>in</strong>tenance 278 40.5%<br />
Plann<strong>in</strong>g 16 2.3%<br />
Service group<br />
Management of Government Resources 124 18.1%<br />
Miss<strong>in</strong>g 2 0.3%<br />
Service Types <strong>and</strong> Components 125 18.2%<br />
Services for Citizens 344 50.2%<br />
Support Delivery of Services to Citizen 91 13.3%<br />
L<strong>in</strong>e of bus<strong>in</strong>ess<br />
Adm<strong>in</strong>istrative Management 15 2.2%<br />
Controls <strong>and</strong> Oversight 12 1.8%<br />
Defense <strong>and</strong> National Security 30 4.4%<br />
Disaster Management 20 2.9%<br />
Economic Development 9 1.3%<br />
Education 16 2.3%<br />
Energy 5 0.7%<br />
Environmental Management 32 4.7%<br />
F<strong>in</strong>ancial Management 81 11.8%<br />
General Government [CA] 45 6.6%<br />
General Science <strong>and</strong> Innovation 22 3.2%<br />
Health 55 8.0%<br />
Homel<strong>and</strong> Security 40 5.8%<br />
Human Resource Management 24 3.5%<br />
Income Security 17 2.5%<br />
Information <strong>and</strong> Technology Management 85 12.4%<br />
International Affairs <strong>and</strong> Commerce 7 1.0%<br />
Law Enforcement 12 1.8%<br />
Natural Resources 16 2.3%<br />
Plann<strong>in</strong>g <strong>and</strong> Budget<strong>in</strong>g 8 1.2%<br />
Public Affairs 13 1.9%<br />
Revenue Collection 8 1.2%<br />
Supply Cha<strong>in</strong> Management 25 3.6%<br />
Transportation 45 6.6%<br />
Workforce Management 5 0.7%<br />
Other 39 5.7%<br />
Total 686 100.0%<br />
Source: I.T. Dashboard data, accessed March, 2010 via http://it.usaspend<strong>in</strong>g.gov<br />
Notes: Planned total cost is <strong>in</strong>flation-adjusted to 2009 dollars us<strong>in</strong>g the CPI-U. Overall rat<strong>in</strong>g is a quality <strong>in</strong>dex that comb<strong>in</strong>es that<br />
CIO evaluation, cost rat<strong>in</strong>g, <strong>and</strong> schedul<strong>in</strong>g rat<strong>in</strong>g sub<strong>in</strong>dexes (see text for details). It takes values from 0 to 10, with 10 be<strong>in</strong>g the<br />
best score. The CIO evaluation is the agency CIO's assessment of project quality. It takes values from 1 to 5, with 5 be<strong>in</strong>g the best.<br />
The cost rat<strong>in</strong>g is based on the absolute percent deviation between the planned <strong>and</strong> actual cost of the project. The cost overrun is a<br />
non-absolute measure that assigns over-cost projects the lowest scores. The schedule rat<strong>in</strong>g is based on the average tard<strong>in</strong>ess of<br />
the project. The cost <strong>and</strong> schedule <strong>in</strong>dexes takes values from 0 to 10, with 10 be<strong>in</strong>g the best. The l<strong>in</strong>e of bus<strong>in</strong>ess other category<br />
comb<strong>in</strong>es all categories with 4 or fewer projects.
CHAPTER 4. YEAR-END SPENDING 192<br />
Table 4.6: Ordered Logit Regressions of Overall Rat<strong>in</strong>g on Last Week <strong>and</strong> Controls<br />
Odds ratio of higher overall rat<strong>in</strong>g<br />
(1) (2) (3) (4)<br />
Last week 0.26 0.46 0.30 0.18<br />
(0.07) (0.14) (0.10) (0.07)<br />
Year FE X X X<br />
Agency FE X X<br />
Project characteristics FE X<br />
N 671 671 671 671<br />
Source: I.T. Dashboard data, accessed March, 2010 via http://it.usaspend<strong>in</strong>g.gov<br />
Notes: Odds ratios from ordered logit regressions. Overall rat<strong>in</strong>g is a quality <strong>in</strong>dex that<br />
comb<strong>in</strong>es that CIO evaluation, cost rat<strong>in</strong>g, <strong>and</strong> schedul<strong>in</strong>g rat<strong>in</strong>g sub<strong>in</strong>dexes (see text for<br />
details). It takes values from 0 to 10, with 10 be<strong>in</strong>g the best score. Project characteristics are<br />
fixed effects for <strong>in</strong>vestment phase, service group, <strong>and</strong> l<strong>in</strong>e of bus<strong>in</strong>ess (see Table 5).<br />
Observations weighted by <strong>in</strong>flation-adjusted spend<strong>in</strong>g. St<strong>and</strong>ard errors <strong>in</strong> parentheses.
CHAPTER 4. YEAR-END SPENDING 193<br />
Table 4.7: Alternative Overall Rat<strong>in</strong>gs Specifications<br />
Odds ratio of higher overall rat<strong>in</strong>g from ordered logit Coefficients from l<strong>in</strong>ear model<br />
W<strong>in</strong>sorized<br />
Heckman<br />
Contracts < $62M Contracts ≥ $62M Unweighted weights OLS selection model<br />
(1) (2) (3) (4) (5) (6)<br />
Last week 0.60 0.18 0.56 0.37 -1.00 -1.57<br />
(0.23) (0.11) (0.14) (0.12) (0.39) (0.64)<br />
Year FE X X X X X<br />
Agency FE X X X X X X<br />
Project characteristics X X X X X X<br />
Weighted by spend<strong>in</strong>g X X X X X<br />
! -0.87<br />
(0.85)<br />
R-squared 0.69<br />
N 335 336 671 671 671 3,803<br />
Source: I.T. Dashboard data, accessed March, 2010 via http://it.usaspend<strong>in</strong>g.gov <strong>and</strong> 2003 to 2010 Exhibit 53 reports, available<br />
at http://www.whitehouse.gov/omb/e-gov/docs/.<br />
Notes: Columns (1) to (4) show odds ratios from ordered logit regressions. Columns (1) <strong>and</strong> (2) split the sample at the median<br />
value. Column (3) shows odds ratios from an unweighted regression. Column (4) W<strong>in</strong>sorizes the spend<strong>in</strong>g weight at $1 billion<br />
(96th percentile).Columns (5) <strong>and</strong> (6) show regression coefficients from l<strong>in</strong>ear regressions. Column (5) reports coefficients from<br />
a simple OLS regression. Column (6) reports coefficient from a Heckman selection model with a l<strong>in</strong>ear second stage. In this<br />
regression the sample is all major I.T. projects recorded <strong>in</strong> the Exhibit 53 reports. The excluded variable <strong>in</strong> this selection model<br />
is the year of project orig<strong>in</strong>ation. The coefficient ! is the implied coefficient on the <strong>in</strong>verse Mill's ratio selection term. Overall<br />
rat<strong>in</strong>g is a quality <strong>in</strong>dex that comb<strong>in</strong>es that CIO evaluation, cost rat<strong>in</strong>g, <strong>and</strong> schedul<strong>in</strong>g rat<strong>in</strong>g sub<strong>in</strong>dexes (see text for details).<br />
It takes values from 0 to 10, with 10 be<strong>in</strong>g the best score. Project characteristics are fixed effects for <strong>in</strong>vestment phase, service<br />
group, <strong>and</strong> l<strong>in</strong>e of bus<strong>in</strong>ess (see Table 5). Observations weighted by <strong>in</strong>flation-adjusted spend<strong>in</strong>g unless otherwise mentioned.<br />
Robust st<strong>and</strong>ard errors <strong>in</strong> parentheses.
CHAPTER 4. YEAR-END SPENDING 194<br />
Table 4.8: Difference-<strong>in</strong>-Differences Estimates of Overall Rat<strong>in</strong>g on Justice <strong>and</strong> Last<br />
Week<br />
OLS Estimates<br />
(1) (2) (3) (4) (5) (6)<br />
Justice X last week 3.54 2.29 2.85 2.36 2.251 2.49<br />
(1.19) (1.16) (0.75) (0.65) 0.593 0.898<br />
Last week -1.91 -1.06 -0.93 -0.99 -0.814 -0.468<br />
(1.10) (0.82) (0.48) (0.39) 0.391 0.238<br />
Justice 0.06 -0.59 -3.33 -3.88 -4.022 -2.028<br />
(0.51) (0.49) (0.47) (0.59) 0.578 1.01<br />
Year FE X X X X X<br />
Agency FE X X X X<br />
Project characteristics X X X<br />
Weighted by spend<strong>in</strong>g X X X X W<strong>in</strong>sorized*<br />
R-squared 0.06 0.22 0.58 0.68 0.60 0.48<br />
N 686 686 686 686 686 686<br />
Source: I.T. Dashboard data, accessed March, 2010 via http://it.usaspend<strong>in</strong>g.gov<br />
Notes: Coefficients from OLS regressions of overall rat<strong>in</strong>g on fully <strong>in</strong>teracted Justice <strong>and</strong> last week<br />
<strong>in</strong>dicators <strong>and</strong> controls. Overall rat<strong>in</strong>g is a quality <strong>in</strong>dex that comb<strong>in</strong>es that CIO evaluation, cost rat<strong>in</strong>g,<br />
<strong>and</strong> schedul<strong>in</strong>g rat<strong>in</strong>g sub<strong>in</strong>dexes (see text for details). It takes values from 0 to 10, with 10 be<strong>in</strong>g the<br />
best score. Project characteristics are fixed effects for <strong>in</strong>vestment phase, service group, <strong>and</strong> l<strong>in</strong>e of<br />
bus<strong>in</strong>ess (see Table 5). Robust st<strong>and</strong>ard errors <strong>in</strong> parentheses.<br />
*Spend<strong>in</strong>g weight W<strong>in</strong>sorized at $1 billion (96th percentile).
CHAPTER 4. YEAR-END SPENDING 195<br />
Table 4.9: First Week Contract Spend<strong>in</strong>g for Selected Product or Service Codes,<br />
Pooled 2004 to 2009 FPDS<br />
Spend<strong>in</strong>g First week<br />
(billions) (percent)<br />
Leases<br />
Lease or rental of facilities $29.2 26.2%<br />
Lease or rental of equipment $5.4 13.1%<br />
Service contracts<br />
Utilities <strong>and</strong> housekeep<strong>in</strong>g services $73.7 11.1%<br />
Medical services $68.8 11.3%<br />
Transportation, travel <strong>and</strong> relocation services $39.3 15.5%<br />
Social services $5.5 9.3%<br />
Other $2,378.1 3.1%<br />
Total $2,600.0 4.0%<br />
Source: Federal Procurement Data System, accessed October, 2010 via www.usaspend<strong>in</strong>g.gov<br />
Note: Percent of contract spend<strong>in</strong>g <strong>in</strong> the last month <strong>and</strong> week of the fiscal year by selected 2digit<br />
product or service code, <strong>in</strong>flation adjusted to 2009 dollars us<strong>in</strong>g the CPI-U. Categories<br />
account for about 8.5% of spend<strong>in</strong>g but 29.7% of first week spend<strong>in</strong>g.
CHAPTER 4. YEAR-END SPENDING 196<br />
Table 4.10: Ordered Logit Regressions of Sub<strong>in</strong>dexes on Last Week <strong>and</strong> Controls<br />
Odds ratio from ordered logit<br />
Evalutation by Agency Cost Cost Schedule<br />
CIO<br />
rat<strong>in</strong>g overrun rat<strong>in</strong>g<br />
(1) (2) (3) (4) (5)<br />
Last week of September 0.14 0.16 0.80 0.74 1.15<br />
(0.06) (0.07) (0.36) (0.30) (0.66)<br />
Cost <strong>and</strong> schedule rat<strong>in</strong>g X<br />
Agency FE X X X X X<br />
Year FE X X X X X<br />
Project characteristics X X X X X<br />
Weighted by spend<strong>in</strong>g X X X X X<br />
N 671 671 671 671 671<br />
Source: I.T. Dashboard data, accessed March, 2010 via http://it.usaspend<strong>in</strong>g.gov<br />
Notes: Odds ratios from ordered logit regressions. The CIO evaluation is the agency CIO's<br />
assessment of project quality. It takes values from 1 to 5, with 5 be<strong>in</strong>g the best. The cost<br />
rat<strong>in</strong>g is based on the absolute percent deviation between the planned <strong>and</strong> actual cost of<br />
the project. The cost overrun is a non-absolute measure that assigns over-cost projects the<br />
lowest scores. The schedule rat<strong>in</strong>g is based on the average tard<strong>in</strong>ess of the project. The<br />
cost <strong>and</strong> schedule <strong>in</strong>dexes takes values from 0 to 10, with 10 be<strong>in</strong>g the best. Project<br />
characteristics are fixed effects for <strong>in</strong>vestment phase, service group, <strong>and</strong> l<strong>in</strong>e of bus<strong>in</strong>ess<br />
(see Table 5). St<strong>and</strong>ard errors <strong>in</strong> parentheses.
CHAPTER 4. YEAR-END SPENDING 197<br />
Table 4.11: Year-End Contract Characteristics Regressions<br />
Dependent Variable:<br />
Cost-<br />
Non-competitive Costreimbursement<br />
or<br />
Non-competitive One bid<br />
or one bid reimbursement T&M/LH<br />
T&M/LH<br />
(1) (2) (3) (4) (5) (6)<br />
Last week -0.002 0.017 0.010 -0.032 0.004 -0.028<br />
(0.002) (0.002) (0.002) (0.002) (0.001) (0.002)<br />
Year FE X X X X X X<br />
Agency FE X X X X X X<br />
Product or service code FE X X X X X X<br />
R-squared 0.41 0.31 0.30 0.52 0.21 0.51<br />
N 402,400 402,400 402,400 402,400 402,400 402,400<br />
Mean of dependent variable 28.7% 20.0% 48.7% 30.0% 5.5% 35.5%<br />
Source: Federal Procurement Data System, accessed October, 2010 via www.usaspend<strong>in</strong>g.gov<br />
Note: Table shows coefficients from l<strong>in</strong>ear probability model regressions of contract type <strong>and</strong> competition type <strong>in</strong>dicators on last week<br />
<strong>and</strong> controls. To facility the analysis, the data is aggregated to the level of the covariates <strong>and</strong> the regressions are weighted by<br />
<strong>in</strong>flation-adjusted spend<strong>in</strong>g <strong>in</strong> each cell.
CHAPTER 4. YEAR-END SPENDING 198<br />
Table 4.12: Percent of Projects <strong>in</strong> I.T. Dashboard Data<br />
Spend<strong>in</strong>g Projects<br />
All I.T Dashboard<br />
% <strong>in</strong> I.T.<br />
Dashboard All I.T Dashboard<br />
% <strong>in</strong> I.T.<br />
Dashboard<br />
Year of orig<strong>in</strong><br />
≤ 2001 $68,460 $14,538 21.2% 813 48 5.9%<br />
2002 $114,668 $12,848 11.2% 1,018 61 6.0%<br />
2003 $115,286 $51,004 44.2% 653 113 17.3%<br />
2004 $53,151 $10,309 19.4% 467 71 15.2%<br />
2005 $35,027 $16,456 47.0% 250 56 22.4%<br />
2006 $13,023 $5,172 39.7% 191 77 40.3%<br />
2007 $61,953 $55,665 89.8% 248 183 73.8%<br />
2008 $19,864 $19,752 99.4% 135 127 94.1%<br />
2009 $498 $491 98.7% 16 13 81.3%<br />
2010 $285 $273 95.5% 13 10 76.9%<br />
Total $482,215 $186,509 38.7% 3,803 759 20.0%<br />
Source: I.T. Dashboard data, accessed March, 2010 via http://it.usaspend<strong>in</strong>g.gov <strong>and</strong> 2003 to 2010<br />
Exhibit 53 reports, available at http://www.whitehouse.gov/omb/e-gov/docs/.<br />
Notes: All spend<strong>in</strong>g <strong>and</strong> projects are totals from agency Exhibit 53 reports. I.T. Dashboard spend<strong>in</strong>g <strong>and</strong><br />
projects are totals <strong>in</strong> the I.T. Dashboard dataset (<strong>in</strong>clud<strong>in</strong>g projects dropped from the basel<strong>in</strong>e sample due<br />
to miss<strong>in</strong>g values). Spend<strong>in</strong>g values <strong>in</strong>flation-adjusted us<strong>in</strong>g the CPI-U.