essays in public finance and industrial organization a dissertation ...

ESSAYS IN PUBLIC FINANCE
AND
INDUSTRIAL ORGANIZATION
A DISSERTATION
SUBMITTED TO THE DEPARTMENT OF ECONOMICS
AND THE COMMITTEE ON GRADUATE STUDIES
OF STANFORD UNIVERSITY
IN PARTIAL FULFILLMENT OF THE REQUIREMENTS
FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY
Neale Mahoney
May 2011

© 2011 by Neale Ashok Mahoney. All Rights Reserved.
Re-distributed by Stanford University under license with the author.
This work is licensed under a Creative Commons Attribution-
Noncommercial 3.0 United States License.
http://creativecommons.org/licenses/by-nc/3.0/us/
This dissertation is online at: http://purl.stanford.edu/fx992pz6459
ii

I certify that I have read this dissertation and that, in my opinion, it is fully adequate
in scope and quality as a dissertation for the degree of Doctor of Philosophy.
Jonathan Levin, Primary Adviser
I certify that I have read this dissertation and that, in my opinion, it is fully adequate
in scope and quality as a dissertation for the degree of Doctor of Philosophy.
Liran Einav
I certify that I have read this dissertation and that, in my opinion, it is fully adequate
in scope and quality as a dissertation for the degree of Doctor of Philosophy.
Approved for the Stanford University Committee on Graduate Studies.
Caroline Hoxby
Patricia J. Gumport, Vice Provost Graduate Education
This signature page was generated electronically upon submission of this dissertation in
electronic format. An original signed hard copy of the signature page is on file in
University Archives.
iii

Preface
This dissertation has four chapters. The first three chapters examine health insurance
markets in the U.S., focusing in particular on contexts where there are important
interactions between health insurance plans. The fourth chapter is on the U.S. budget,
examining the implications of annual budget cycles on the quantity and quality of
end-of-year spending.
Chapter 1, entitled “Bankruptcy as Implicit Health Insurance” examines the inter-
action between health insurance and the implicit insurance that people have because
they can file (or threaten to file) for bankruptcy. With a simple model that captures
key institutional features, I demonstrate that the financial risk from medical shocks
is capped by the assets that could be seized in bankruptcy. For households with mod-
est seizable assets, this implicit “bankruptcy insurance” can crowd out conventional
health insurance. I test these predictions using variation in the state laws that specify
the type and level of assets that can be seized in bankruptcy. Because of the differing
laws, people who have the same assets and receive the same medical care face different
losses in bankruptcy. Exploiting the variation in seizable assets that is orthogonal to
wealth and other household characteristics, I show that households with fewer seiz-
able assets are more likely to be uninsured. This finding is consistent with another:
uninsured households with fewer seizable assets end up making lower out-of-pocket
medical payments. The estimates suggest that if the laws of the least debtor-friendly
state of Delaware were applied nationally, 16.3 percent of the uninsured would buy
health insurance. Achieving the same increase in coverage would require a premium
subsidy of approximately 44.0 percent. To shed light on puzzles in the literature and
examine policy counterfactuals, I calibrate a utility-based, micro-simulation model of
iv

insurance choice. Among other things, simulations show that “bankruptcy insurance”
explains the low take-up of high-deductible health insurance.
Chapter 2, entitled “Pricing and Welfare in Health Plan Choice”, is coauthored
with M. Kate Bundorf and Jonathan Levin. The starting point for the paper is the
simple observation that when insurance premiums do not reflect individual differences
in expected costs, consumers may choose plans inefficiently. We study this problem in
health insurance markets, a setting in which prices often do not incorporate observable
differences in expected costs. We develop a simple model and estimate it using data
on small employers. In this setting, the welfare loss compared to the feasible risk-
rated benchmark is around 2-11% of coverage costs. Three-quarters of this is due
to restrictions on risk-rating employee contributions; the rest is due to inefficient
contribution choices. Despite the inefficiency, the benefits from plan choice relative
to each of the single-plan options are substantial.
Chapter 3, entitled “The Private Coverage and Public Costs: Identifying the Effect
of Private Supplemental Insurance on Medicare Spending,” is coauthored with Marika
Cabral. While most elderly Americans have health insurance coverage through Medi-
care, traditional Medicare policies leave individuals exposed to significant financial
risk. Private supplemental insurance to “fill the gaps” of Medicare, known as Medi-
gap, is very popular. In this Chapter, we estimate the impact of this supplemental
insurance on total medical spending using an instrumental variables strategy that
leverages discontinuities in Medigap premiums at state boundaries. Our estimates
suggest that Medigap increases medical spending by 57 percent—or about 40 percent
more than previous estimates. Back-of-the-envelope calculations indicate that a 20
percent tax on premiums would generate combined revenue and savings of 6.2 percent
of baseline costs; a Pigovian tax that fully accounts for the fiscal externality would
yield savings of 18.1 percent.
Chapter 4, entitled “Do Expiring Budgets Lead to Wasteful Year-End Spending?
Evidence from Federal Procurement,” is coauthored with Jeffrey Liebman. Many
organizations fund their spending out of a fixed budget that expires at year’s end.
Faced with uncertainty over future spending demands, these organizations have an
incentive to build a buffer stock of funds over the front end of the budget cycle.
v

When demand does not materialize, they then rush to spend these funds on lower
quality projects at the end of the year. We test these predictions using data on
procurement spending by the U.S. federal government. Using data on all federal
contracts from 2004 through 2009, we document that spending spikes in all major
federal agencies during the 52nd week of the year as the agencies rush to exhaust
expiring budget authority. Spending in the last week of the year is 4.9 times higher
than the rest-of-the-year weekly average. We examine the relative quality of year-end
spending using a newly available dataset that tracks the quality of $130 billion in
information technology (I.T.) projects made by federal agencies. Consistent with the
model, average project quality falls at the end of the year. Quality scores in the last
week of the year are 2.2 to 5.6 times more likely to be below the central value. To
explore the impact of allowing agencies to roll unused spending over into subsequent
fiscal years, we study the I.T. contracts of an agency with special authority to roll
over unused funding. We show that there is only a small end-of-year I.T. spending
spike in this agency and that the one major I.T. contract this agency issued in the
52nd week of the year has a quality rating that is well above average.
vi

Acknowledgements
I am deeply thankful to my dissertation committee for their advice and support. I
thank Jonathan Levin for his mentorship. I could not ask for a better adviser or role
model. I thank Caroline Hoxby and Liran Einav for their exceptional guidance. They
say that your graduate advisers are your advisers for life. Having these advisers for
my years at Stanford is already more than I deserve.
I am also grateful to the many faculty members who provided valuable comments
and suggestions for my research during my graduate studies. Among them are Ran
Abramitzky, Doug Bernheim, Jay Bhattacharya, Nick Bloom, Tim Bresnahan, Kate
Bundorf, Jacub Kastl, and Jeffrey Leibman. I thank Kyna Fong for joining my oral
exam committee.
I acknowledge the Ric Weiland Graduate Fellowship, Kapnick Fellowship, and
Shultz Scholarship for financial support.
My time at Stanford has been a case study in the virtuous cycle of positive peer
effects. I thank Albert Bollard and Max Floetotto for their friendship. I am lucky to
have had Aaron Bodoh-Creed, Marika Cabral, Alex Hirsch, Carlos Lever, Marcello
Miccoli, Izi Sin, and Gui Woolston as colleagues and friends. I thank Alessandra
Voena most of all.
I am also grateful to my advisers at Brown University for introducing me to the
practice of economic research. I thank Herschel Grossman for countless hours of
debate and discussion and Roberto Serrano for his wisdom in urging me to pursue
graduate studies.
Finally, I am immensely thankful to my family, my parents Raymond Mahoney
and Nalini Shah-Mahoney and my brother Colin, for their unconditional love and
vii

support and the abundant happiness they bring to my life.
viii

Contents
Preface iv
Acknowledgements vii
1 Bankruptcy as Implicit Health Insurance 1
1.1 Introduction ................................ 1
1.2 Bankruptcy as a Form of High-Deductible Health Insurance ..... 6
1.2.1 Institutional Background ..................... 6
1.2.2 A Model of Bankruptcy as High-Deductible Health Insurance . 8
1.3 Data Overview .............................. 11
1.3.1 Asset Exemptions ......................... 12
1.3.2 Seizable Assets .......................... 12
1.3.3 Medical Costs ........................... 14
1.3.4 Insurance Premiums ....................... 15
1.4 Empirical Strategy ............................ 16
1.4.1 Coverage Equation ........................ 16
1.4.2 Costs Equation .......................... 16
1.4.3 Cross-State Variation ....................... 17
1.4.4 Historical Homestead Exemptions ................ 19
1.5 The Effect on Insurance Coverage .................... 21
1.5.1 First Stage Estimates ....................... 21
1.5.2 Baseline Coverage Estimates ................... 22
1.5.3 Sensitivity Analysis ........................ 23
ix

1.5.4 Heterogeneity in the Effect on Coverage ............ 24
1.5.5 The Effect on Wealth and Premiums .............. 25
1.5.6 Summary and Interpretation ................... 26
1.6 The Effect on Costs ............................ 27
1.6.1 Baseline Estimates ........................ 27
1.6.2 Sensitivity Analysis ........................ 28
1.6.3 Summary and Interpretation ................... 29
1.7 Micro-Simulation Model ......................... 29
1.8 Puzzles and Policy ............................ 31
1.8.1 Puzzles ............................... 31
1.8.2 Policy Implications ........................ 33
1.9 Conclusion ................................. 35
1.10 Appendix to Chapter 1 .......................... 36
1.10.1 Model Predictions ......................... 36
1.10.2 Seizable Assets Calculation Details ............... 37
1.10.3 Micro-Simulation Details ..................... 38
2 Pricing and Welfare in Health Plan Choice 65
2.1 Introduction ................................ 65
2.2 Health Plan Pricing and Market Efficiency ............... 68
2.3 Data and Environment .......................... 72
2.3.1 Institutional Setting ....................... 72
2.3.2 Data and Descriptive Statistics ................. 74
2.4 Econometric Model ............................ 77
2.4.1 Consumer Preferences, Plan Costs and Market Behavior ... 77
2.4.2 Discussion of Model and Identification ............. 82
2.4.3 Estimation Strategy ....................... 85
2.4.4 Welfare Measurement ....................... 86
2.5 Empirical Results ............................. 87
2.5.1 Model Estimates ......................... 88
2.5.2 Quantifying Social Welfare Inefficiencies ............ 93
x

2.5.3 The Value of Plan Choice .................... 96
2.5.4 Discussion and Sensitivity Analysis ............... 97
2.6 Conclusion ................................. 98
2.7 Appendix to Chapter 2 .......................... 99
2.7.1 Comparison to Demand Estimates in Other Settings ..... 99
2.7.2 Alternative Specifications of the Demand Model ........ 100
3 Private Coverage and Public Costs: Identifying the Effect of Private
Supplemental Insurance on Medicare Spending 117
3.1 Introduction ................................ 117
3.2 Background ................................ 122
3.3 Data .................................... 124
3.4 Empirical Model ............................. 125
3.4.1 Estimating equations ....................... 125
3.4.2 Estimation approach ....................... 127
3.5 Identifying Variation ........................... 128
3.5.1 The Instrument .......................... 128
3.5.2 Potential Concerns ........................ 130
3.6 Results and Discussion .......................... 132
3.6.1 Premiums and Medigap Choice ................. 132
3.6.2 Causal Impact on Costs ..................... 134
3.6.3 Robustness and External Validity ................ 136
3.7 Policy Counterfactuals .......................... 137
3.8 Conclusion ................................. 139
4 Do Expiring Budgets Lead to Wasteful Year-End Spending? Evi-
dence from Federal Procurement 153
4.1 Introduction ................................ 153
4.2 A Model of Wasteful Year-End Spending ................ 157
4.2.1 The Baseline Model ........................ 157
4.2.2 Rollover Budget Authority .................... 159
4.2.3 Extending the Model To Allow For Rollover .......... 160
xi

4.3 Does Spending Spike at the End of the Year? ............. 161
4.3.1 The Federal Procurement Data System ............. 163
4.3.2 The Within-Year Pattern of Government Procurement Spending 164
4.3.3 The Impact of Appropriations Timing on the Within-Year Pat-
tern of Government Procurement Spending ........... 165
4.4 Is End of Year Spending of Lower Quality? ............... 167
4.4.1 I.T. Dashboard .......................... 168
4.4.2 Data and Summary Statistics .................. 169
4.4.3 The Relative Quality of Year-End I.T. Contracts ....... 172
4.4.4 Sensitivity Analysis ........................ 174
4.4.5 Why Are Year End Contracts of Lower Quality? ........ 176
4.5 Do Rollover Provisions Raise Spending Quality? ............ 179
4.5.1 The DOJ’s Rollover Authority .................. 179
4.6 Conclusion ................................. 182
xii

List of Tables
1.1 Asset Exemption Laws by State ..................... 52
1.2 Implied First Stage Estimates ...................... 53
1.3 Baseline Coverage Estimates ....................... 54
1.4 Sensitivity Analysis of the Effect on Coverage ............. 55
1.5 Heterogeneity in the Effect on Coverage ................. 56
1.6 The Effect on Assets ........................... 57
1.7 The Effect on Premiums ......................... 58
1.8 Baseline Costs Estimates ......................... 59
1.9 Sensitivity Analysis of the Effect on Costs ............... 60
1.10 Micro-Simulation Estimates of Percent Covered Without and With
Bankruptcy Insurance by Insurance Status ............... 61
1.11 Policy Counterfactuals .......................... 62
1.12 Summary Statistics: Seizable Assets by Insurance Status ....... 63
1.13 Summary Statistics: Medical Costs by Insurance Status ........ 63
1.14 Summary Statistics: Premiums ..................... 64
1.15 Costs and Premiums by Deductible Level ................ 64
2.1 Risk and Demographics .......................... 102
2.2 Plan Characteristics ........................... 103
2.3 Risk and Demographics by Plan ..................... 104
2.4 Demand Model .............................. 105
2.5 Costs and Bids .............................. 106
2.6 Matching and Welfare under Alternative Contribution Policies .... 107
2.7 Matching and Welfare by Risk Score Quintile ............. 108
xiii

2.8 The Value of Plan Choice ........................ 109
2.9 Alternative Demand Model Specifications ................ 115
3.1 Medicare Cost-Sharing .......................... 141
3.2 Summary Statistics by Supplemental Insurance Type ......... 142
3.3 Medigap Enrollment and Plan Characteristics by Letter ........ 143
3.4 HRR Level-Medicare Costs Tabulated By State ............ 144
3.5 First Stage: OLS Estimates of Premiums on State and HRR Costs .. 145
3.6 Second Stage: Marginal Effects for Medigap Choice from Multinomial
Logit Model ................................ 146
3.7 Third Stage: OLS and Full Model Estimates of Medical Costs on Medi-
gap Indicator ............................... 147
3.8 Robustness Checks: Key Parameter Estimates from Alternative Spec-
ifications and Samples .......................... 148
3.9 Policy Counterfactuals for Subsidies and Taxes of Medigap Premiums 149
3.10 OLS Estimates of Premiums on HRR- and State-Level Medicare Costs 150
3.11 Second Stage: 2SLS Estimates of Medigap Choice from Linear Proba-
bility Model ................................ 151
3.12 Third Stage: 3SLS Estimates of Medical Costs on Medigap Indicator 152
4.1 Summary Statistics: Federal Contracting, Pooled 2004 to 2009 FPDS 187
4.2 Year-End Contract Spending by Agency, Pooled 2004 to 2009 FPDS . 188
4.3 Year-End Contract Spending By Selected Product or Service Code,
Pooled 2004 to 2009 FPDS ........................ 189
4.4 Summary Statistics: Major I.T. Projects as of March, 2010 ...... 190
4.5 Summary Statistics: Quality Indexes and Project Characteristics for
Major I.T. Projects ............................ 191
4.6 Ordered Logit Regressions of Overall Rating on Last Week and Controls192
4.7 Alternative Overall Ratings Specifications ............... 193
4.8 Difference-in-Differences Estimates of Overall Rating on Justice and
Last Week ................................. 194
xiv

4.9 First Week Contract Spending for Selected Product or Service Codes,
Pooled 2004 to 2009 FPDS ........................ 195
4.10 Ordered Logit Regressions of Subindexes on Last Week and Controls 196
4.11 Year-End Contract Characteristics Regressions ............. 197
4.12 Percent of Projects in I.T. Dashboard Data ............... 198
xv

List of Figures
1.1 Histogram of Seizable Assets by Insurance Status ........... 41
1.2 Payments vs. Charges by Insurance Status ............... 42
1.3 Provider Billing Decision: Submitted Bill vs. List Price of Medical Care 43
1.4 Payments vs. Charges by Seizable Assets ................ 44
1.5 Simulated Instrument by State ...................... 45
1.6 Log Seizable Assets Percentiles by State for a Constant Sample of
Households ................................ 46
1.7 Seizable Homestead Equity in 2005 vs. 1920 for a Constant Sample of
Households ................................ 47
1.8 Insurance Coverage vs. Seizable Assets ................. 48
1.9 Effect on Coverage on Samples Excluding Each State ......... 49
1.10 Micro-Simulation Estimates of Percent Covered Without and With
Bankruptcy Insurance by Deductible .................. 50
1.11 The Insurance Generosity Gap ...................... 51
2.3 Contributions and Bids Relative to Integrated HMO .......... 111
2.4 Employer Contributions and Employee Characteristics ........ 112
2.5 Costs by Risk Score ............................ 113
2.6 Bids by Risk Score ............................ 114
3.1 Example of Identifying Variation .................... 140
4.1 Federal Contracting by Week, Pooled 2004 to 2009 FPDS ....... 183
4.2 Year-End Spending by Appropriations Date .............. 184
xvi

4.3 I.T. Contracting by Week ........................ 185
4.4 Year-End and Rest-of-Year Overall Ratings ............... 186
xvii

Chapter 1
Bankruptcy as Implicit Health
Insurance
1.1 Introduction
There is a large literature in economics evaluating the effects of government policy
on health insurance coverage in the United States. 1 The question of why households
choose to be uninsured is less well understood. 2 To better understand the insurance
coverage decision, this paper examines a mechanism that has received little attention:
implicit insurance from the threat-point of personal bankruptcy.
The implicit insurance from bankruptcy arises from the confluence of three factors.
First, due to federal law, hospitals are required to provide emergency treatment on
credit—and typically provide non-emergency care without any upfront payment as
well. Second, under Chapter 7 of the U.S. bankruptcy code, households can discharge
medical debt, giving up assets above asset exemption limits in return. 3 Third, because
1 See ? for review of the take-up and crowd-out effects of public insurance expansions. See ? for
a review of the impact of tax subsidies on the employer provision of insurance. See ? for a review
of the effects of tax policy on insurance take-up in the non-group market.
2 In a review of the literature ? concludes, “there are a variety of hypotheses for why so many
individuals are uninsured, but no clear sense that this set of explanations can account for the 47
million individuals.”
3 The Bankruptcy Abuse Prevention and Consumer Protection Act (BAPCPA) of 2005 was implemented
after the period I analyze. It prevents households with more than the state median income
1

CHAPTER 1. BANKRUPTCY 2
of the deadweight cost of the bankruptcy process, households and creditors have an
incentive to negotiate payments without a formal bankruptcy filing.
Bankruptcy, as a result, provides households with a form of high-deductible health
insurance. Households are exposed to the financial risk from medical shocks up to
the level of assets that can be seized in bankruptcy and insured against financial risk
above this level.
Summary data on the uninsured suggest that this mechanism could be important.
Figure 1.1 shows that uninsured households have vastly fewer seizable assets than
households with private insurance. Sixty-three percent of the uninsured would give up
less than $5,000 in a bankruptcy filing, compared to only 28 percent of the privately
insured. Figure 1.2 shows that payments by the uninsured are substantially lower
when receiving a high volume of medical care. While payments by the privately
insured scale up proportionally with medical charges, payments by the uninsured are
capped on average at just over $5,000.
To more rigorously examine the mechanism, I construct a simple model of bankruptcy,
medical billing, and insurance choice. The model predicts that, conditional on wealth,
out-of-pocket medical payments should be decreasing in the level of seizable assets for
a given volume of medical care received. Holding wealth constant, households with
fewer seizable assets should be less likely to purchase conventional coverage.
I test these predictions using variation in the state-level asset exemption laws that
specify the type and level of assets that can be seized in bankruptcy and detailed
asset data from the restricted access Medical Expenditure Panel Survey (MEPS) and
the Panel Survey of Income Dynamics (PSID). The degree of cross-state variation
in the asset exemption laws in substantial. Kansas, for example, allows households
to exempt an unlimited amount of home equity and up to $40,000 in vehicle equity.
Neighboring Nebraska allows households to keep no more than $12,500 in home equity
or a $5,000 wildcard of any type of asset. Both states allow households to keep
retirement assets. I construct a simulated instrument that isolates the variation in
seizable assets solely due to these laws, mechanically purging variation due to wealth
from filing under Chapter 7 in most circumstances. The households most affected by the reform are
unlikely to be marginal to the mechanism I analyze.

CHAPTER 1. BANKRUPTCY 3
and other household characteristics.
Using this source of variation and cost data from the MEPS, I find that uninsured
households with fewer seizable assets make lower out-of-pocket payments for a given
level of medical care received. My preferred estimate indicates that a log point drop
in seizable assets reduces out-of-pocket payments by 37 percent. Consistent with the
high-deductible nature of this insurance, the drop is larger for households that utilize
more medical services as the “deductible” of this implicit insurance is more likely to
bind.
Using the same source of variation and MEPS and PSID data, I find that house-
holds with fewer seizable assets are less likely to have insurance. The estimates suggest
that if the bankruptcy laws of the least debtor-friendly state of Delaware were applied
nationally, 16.3 percent of the uninsured would buy health insurance. With a take-up
semi-elasticity of -0.084 from the literature, achieving the same increase in take-up
would require a premium subsidy of 44.0 percent. 4
I use three strategies to address the concern that asset exemption laws may be
correlated with unobserved state-level factors. The first strategy is to use variation
due to 1920 homestead exemptions as an instrument. By using homestead exemptions
from before the era of widespread health insurance, the instrument alleviates potential
bias from factors that might have simultaneously caused changes in asset exemption
law and the dependent variables over the course of the twentieth century. Moreover,
historical evidence (e.g., ?) showing that 1920 homestead exemption levels resulted
from an idiosyncratic array of nineteenth century historical circumstances diminishes
the concern that this variable merely proxies for a persistent state characteristic (such
as the strength of a pro-debtor political movement).
The second strategy is to sequentially add fixed effects for Census Regions (e.g.,
Northeast) and Census Divisions (e.g., New England) to the main specification. If a
spatially correlated unobserved factor is driving the findings, the results should change
with the inclusion of these covariates. Stable estimates across these specifications
mitigate this concern. The third strategy is to add controls for a rich set of potentially
4 The -0.084 estimates is taken from ? and is based on premium variation due to state-level
community rating and premium compression regulations. As I discuss below, this estimate is in the
center of the range in the literature.

CHAPTER 1. BANKRUPTCY 4
relevant legislative factors. The fact that the estimates are uncharged by the inclusion
of these covariates provides further support for the case that the identifying variation
is uncorrelated with unobserved state-level determinants of costs and coverage.
I take the analysis a step further by calibrating a utility-based, micro-simulation
model of insurance choice. The model is based on a nationally representative sample
of households. Households face household-specific medical shock distributions that
depend on the age and sex of each household member. To maximize their expected
utility over wealth, households choose to either purchase conventional insurance at
market premiums or rely on the high-deductible insurance from bankruptcy.
I use this model to shed light on puzzles in the health policy literature. One of the
puzzles I examine is the low take-up of high-deductible health plans (HDHP) by the
uninsured (?). Proponents of these plans have argued that by offering lower premi-
ums, HDHPs would expand coverage among the uninsured. But because the implicit
insurance from bankruptcy often resembles a high-deductible policy, HDHPs are rel-
atively more likely to be crowded out by this mechanism. In the micro-simulation
model, accounting for bankruptcy reduces the percentage of uninsured households
projected to purchase a $1,000 deductible plan by 13 percentage points. For a $5,000
deductible plan, bankruptcy reduces demand by 37 percentage points, or from 43 to
6percent.
A second puzzle I examine is heterogeneity in the demand for coverage. Without
bankruptcy, households with and without insurance coverage are difficult to separate.
Using variation in medical risk, tax exemptions, and administrative costs, the model
can separate the predicted coverage levels of uninsured and insured households by only
14 percentage points. Because there are large differences in seizable assets across these
groups, accounting for bankruptcy has a substantial incremental effect, widening the
gap in predicted coverage from 14 to 51 percentage points.
The mechanism I study may be relevant to policy design. On the one hand, the
implicit insurance from bankruptcy has obvious inefficiencies. Uninsured households
receive a substantial amount of non-emergency medical care in emergency rooms,
which are obviously not optimized for this purpose (?). At the same time, these
households receive less preventative care than they would with conventional health

CHAPTER 1. BANKRUPTCY 5
insurance (?). There are deadweight costs to negotiation and collections under the
threat-point of bankruptcy. And the fact that uninsured households are not exposed
to the social cost of this implicit insurance means that too many households choose
to be uninsured.
Yet conventional health insurance has inefficiencies of its own. A particularly
interesting inefficiency relative to bankruptcy insurance is moral hazard. With con-
ventional insurance, medical providers and patients often have incentives to supply
and demand excess medical care. Under the implicit insurance from bankruptcy,
however, physicians are more likely to be exposed to the social cost of their deci-
sions and patients have little leverage to demand excess treatment. The result is that
bankruptcy may be a lower moral hazard form of social insurance. 5
While a comprehensive analysis of the costs and benefits of bankruptcy is overly
ambitious, I can examine one key tension between these forms of insurance with the
micro-simulation model. In particular, the model allows me to trade off the benefit of
bankruptcy as a lower moral hazard form of insurance against the inefficiency due to
households not facing the full social cost from being uninsured. This tradeoff suggests
a corrective system of “Pigovian penalties” that expose households to the full social
cost of the implicit insurance they receive. 6
With the optimal penalty of $218 per person on average, about 7 percent of
the uninsured take up coverage and aggregate surplus increases by a small $4 to $5
per person. Analyzed in this framework, the penalties under the Patient Projection
and Affordable Care Act (PPACA), which average $418 per person, are too large,
decreasing aggregate surplus by $9 to $13 per person on average. By effectively
eliminating the low moral hazard option, the counterfactual of making medical debt
non-dischargeable in bankruptcy reduces surplus by an average of $36 to $43 per
person. While the exact welfare numbers should be viewed with some caution, the
exercise suggests that dramatically reducing bankruptcy insurance—while having the
5 Estimates of the moral hazard effects from conventional health insurance, when identified off
large medical costs, are incremental to any moral hazard effects from the implicit insurance from
bankruptcy.
6 For this exercise, I assume that uninsured households are not already subsidized through the
tax code or some other channel.

CHAPTER 1. BANKRUPTCY 6
superficial benefit of increasing conventional coverage—may not be socially desirable.
By studying the interaction between implicit and conventional insurance, this pa-
per is closely related to the literature on long-term care insurance and the implicit
insurance from spending down assets and qualifying for Medicare, such as ?. Like
them, I find that implicit insurance can cause substantial crowd-out. It is more gen-
erally related to a literature in macroeconomics that assesses the equilibrium effects
of bankruptcy as a form of consumption insurance against a range of different shocks,
including those related to earnings, divorce, childbearing, lawsuits, and medical bills
(??). By examining unpaid care, this paper is also related to ? and ?, who find a
negative association between measures of charity care and insurance coverage. And
this paper shares similarities with a literature that examines the effect of medical debt
on bankruptcy filings (e.g., ???), although unlike these papers, I treat bankruptcy as
threat-point, not a dependent variable to be explained. In this, my approach more
closely resembles the “informal bankruptcy” viewpoint advanced by ?, who show that
credit card debt is charged off without a bankruptcy filing in the majority of cases.
The rest of the paper proceeds as follows: Section 2 presents the institutional
background and a simple model. Section 3 provides an overview of the data. Sections
4 discusses the identification strategy. The main empirical results are presented in
Sections 5 and 6. The micro-simulation model is presented in Section 7. Section 8
discusses puzzles in the literature and policy implications. Section 9 concludes.
1.2 Bankruptcy as a Form of High-Deductible Health
Insurance
1.2.1 Institutional Background
The implicit insurance from bankruptcy arises from the combination of three insti-
tutional features: the fact that most medical care is provided on credit even when
repayment is unlikely, the ability of households to discharge this debt in bankruptcy,
and the incentive for households and creditors to come to a negotiated solution to
avoid the deadweight loss from a formal bankruptcy filing.

CHAPTER 1. BANKRUPTCY 7
The Emergency Medical Treatment and Active Labor Act (EMTALA) requires
that hospitals treat patients with emergency medical conditions, and prohibits them
from delaying treatment to inquire about insurance status or means of payment. 7 As
a matter of practice, most hospitals provide non-emergency medical care on credit as
well. Hospitals generally lack the infrastructure to bill patients at the point of service
(?) and rarely deny service when repayment is unlikely. 8
Having received medical care on credit, bankruptcy law allows households to write
off this debt in exchange for assets or future earnings. Chapter 7 is the most pop-
ular form of personal bankruptcy, accounting for about 70 percent of all filings (?).
Under Chapter 7, households can discharge most unsecured debt such as credit card
debt, installment loans, and medical bills. In return, creditors can seize assets above
exemption levels that vary by asset type and state of residence.
Chapter 13 is the other bankruptcy option. Under Chapter 13, households dis-
charge most unsecured debt in exchange for payments out of disposable income over
the following 3 to 5 years. By statute, these payments must be of at least the value
that creditors would receive in Chapter 7. They are rarely larger because, in the
period I study, all households have the option to file for Chapter 7. 9 Following ?, I
use seizable assets under Chapter 7 to characterize payments under both chapters of
the bankruptcy code.
Households, however, do not have to formally declare bankruptcy to receive the
implicit insurance it provides. Under the threat-point of bankruptcy, households and
medical providers often resolve payments without an actual bankruptcy filing. There
are multiple junctures where this occurs. Discounts on the list price of treatment—
known as charity care—are offered at the point of service to the obviously indigent. 10
7 U.S.C. 42 §1395dd.
8 In a survey of nonprofit hospitals, 90 percent reported never denying any medical services to
patients with no insurance (?). For-profit hospitals seem to operate similarly. For example, ?
rejects the hypothesis that for-profit hospitals have a lower preference for charity care. ? find that
the majority of emergency departments offer preventative care to uninsured patients.
9 The Bankruptcy Abuse Prevention and Consumer Protection Act (BAPCPA), effective in October
2005, established a “means test” for Chapter 7. It restricted households earning more than
the state median income from filing under Chapter 7 in most circumstances. The households most
effected by the reform are unlikely to be marginal to the mechanism I analyze.
10 Federal and state laws also influence charity care provision. Nonprofits use charity care to meet

CHAPTER 1. BANKRUPTCY 8
After treatment, many hospitals encourage financially strapped households to nego-
tiate discounts, requiring the submission of information on income and assets (e.g.,
W-2s and mortgage payments) as part of their charity care applications. 11 Even when
charity care is not provided, the lion’s share of medical debt is charged off in the col-
lection process. Despite contracting with debt collection agencies, providers recover
only about 10 to 20 percent of bills submitted to the uninsured (?).
Overall, bad debt from the uninsured was estimated at about $16 to $18 billion
in 2004 (LeCuyer and Singhal, 2007). While the exact proportion of debt discharged
without a bankruptcy filing is unknown, ? find that the ratio of “medical” to “non-
medical” bankruptcies, according to their definition, is the same for households with
and without insurance coverage, suggesting that a large portion of the uninsured’s
medical debt is charged off outside of formal bankruptcy. This is not unique to
medical debt. ? report that the majority of credit card debt is charged off in what
they call “informal bankruptcy.”
1.2.2 A Model of Bankruptcy as High-Deductible Health In-
surance
To bring together these institutional features, I build a stylized model of households,
medical providers, and bankruptcy. Households have a representative agent with
expected utility preferences over wealth w = w E +w S , composed of exempt assets w E
(net wealth that cannot be seized in a bankruptcy filling) and seizable assets w S (net
wealth that can be seized by creditors). 12 They face medical shocks with list price
m drawn from distribution M and choose whether to purchase health insurance to
protect against this financial risk.
Medical providers are obligated to provide medical services m and then attempt
their Community Benefit requirement. Some states subsidize care to the indigent through unpaid
care pools. I account for these factors in the empirical analysis.
11 When this information is not provided, hospitals run credit checks on indebted patients, filing
suit if they find evidence of a mortgage or savings that could be claimed (“In Their Debt,” Baltimore
Sun, December 12, 2008 to December 24th, 2008).
12 I discuss endogenizing wealth at the end of this section.

CHAPTER 1. BANKRUPTCY 11
Prediction 2. Holding overall wealth constant, the willingness to pay for insurance
v—and therefore insurance coverage—is increasing in the level of seizable assets w S .
Because the implicit insurance from bankruptcy is a substitute for conventional health
insurance, households with more seizable assets have a higher willingness to pay for
insurance and are ceteris paribus more likely to be insured. I derive the prediction in
Appendix Section 1.10.1.
The prediction is robust to natural extensions of the model. For example, allowing
insured households to receive more or better medical treatment (?) increases the
incentive to purchase coverage, but households with fewer seizable assets are still
relatively less likely to insure. Similarly, increasing the cost of bankruptcy to account
for factors such as stigma (?) or reduced future access to credit (?) does not affect the
basic prediction. And endogenizing the level of seizable and exempt assets actually
strengthens the relationship between insurance coverage and seizable assets because
households that choose to forgo coverage have an additional incentive to reduce their
seizable asset holdings.
A more subtle point relates to the information available to households. The model
assumes that households know their level of seizable assets w S and their health risk.
Obviously this is an exaggeration. What matters is that households have some knowl-
edge of the financial risk from forgoing insurance. For example, if households learn
from the news-media or peers that medical providers in their community frequently
seize home equity, then homeowners may be more likely to purchase coverage—even
if they know nothing about the mechanism.
1.3 Data Overview
I use two main data sources to test the predictions of the model. I examine the effect
of bankruptcy on medical costs using data from the 2000 to 2005 waves of the Medical
Expenditure Panel Survey (MEPS). The survey has detailed information on medical
costs and insurance coverage. At the Data Center, encrypted state identifiers and

CHAPTER 1. BANKRUPTCY 12
newly edited asset and debt variables are also available. 17
I examine the effects on coverage using the MEPS data and the 1999 to 2005 waves
of the biennial Panel Survey of Income Dynamics (PSID). The survey has public use
information on insurance coverage, assets and debts variables, and state identifiers.
Because the state of residence is non-encrypted in the PSID, I use this dataset for
the primary insurance coverage analysis and replicate the results in the MEPS.
In both data sets, I aggregate the individual-level data to the household level and
inflation-adjust monetary variables to 2005 dollars using the CPI-U. I also exclude
households with one or more members enrolled in public insurance or a head age 65
or older due to their eligibility for public Medicare insurance. 18 This leaves me with
34,841 observations in the MEPS and 22,844 observations in the PSID.
1.3.1 Asset Exemptions
I codify assets exemptions using ?, a do-it-yourself guide to personal bankruptcy.
Table 1.1 shows these exemptions. Contemporaneous homestead exemptions exhibit
substantial variation, ranging from zero in seven states to unlimited in eight others;
vehicle exemptions range from zero in 15 states to at least $10,000 in five others;
and wildcard exemptions, which can be applied to any asset, show a similar degree
of variation. California residents can file under two different exemption systems, and
residents of 14 states can file under the federal exemption system if they choose. The
last column shows homestead exemptions in the year of 1920 from ?. 19
1.3.2 Seizable Assets
Let i denote households and j denoted states. Seizable assets are a function of house-
hold assets and debts and exemptions laws, denoted by vectors wi and ej. Following
the general structure of ?, seizable asset can be decomposed into assets that can be
17 ? find that the estimates of net worth in the MEPS are comparable to those in the Survey of
Income and Program Participation (SIPP). My analysis does not go back before 2000 because when
I started the project only asset and debt data from the 2000 to 2005 period had been edited.
18 In the MEPS, I also drop the 3.6 percent of households with missing wealth variables.
19 States that did not exist and states that had only acre-based exemptions are denoted as missing.

CHAPTER 1. BANKRUPTCY 14
Appendix Section 1.10.2.
1.3.3 Medical Costs
Medical costs variables are shown in Appendix Table 1.13. Annual medical charges,
defined as the list price of medical services used that year, average $6,647 per house-
hold. Total payments, defined as the sum of payments received, are less than charges
due to both discounts negotiated by insurance providers and medical care provided
as charity care or bad debt. For privately insured households, total payments average
$4,480 per household. Ninety-four percent of these payments are either out-of-pocket
payments or payments made by private insurance providers. For the uninsured, to-
tal payments average $1,267 per household. Fifty-two percent of these payments are
out-of-pocket. Miscellaneous payments, such as payments from charity care pools,
worker’s compensation, or automobile insurance, account for most of the rest.
In the empirical analysis, I use the out-of-pocket payments variable to measure
the financial risk faced by the uninsured. While this variable will accurately capture
financial risk in most circumstances, it may inaccurately measure financial risk for
two reasons. First, out-of-pocket payments overstate financial risk when these pay-
ments are put on a credit card that is ultimately discharged in bankruptcy. Because
households with fewer seizable assets are more likely to discharge credit card debt,
out-of-pocket payments may overestimate financial risk for some low seizable assets
households.
Second, out-of-pocket payments may understate financial risk when medical providers
break up large bills into installments since households are not prompted to report pay-
ments that extend beyond the survey period. Because households with more seizable
assets are more likely to make these large, multi-installment payments, out-of-pocket
payments may understate financial risk for high seizable assets households. Both
overstated financial risk for low seizable assets households and understated financial
risk for households with high seizable assets may lead to attenuated estimates of the
relationship between seizable assets and financial risk as measured by out-of-pocket
payments, suggesting that the empirical estimates should be interpreted as a lower

CHAPTER 1. BANKRUPTCY 15
bound of the effect of bankruptcy insurance on household financial risk.
I use Relative Risk Scores to control for medical utilization. As I discuss in Section
1.4, controlling for utilization is important because the direction of the unconditional
relationship between out-of-pocket payments and seizable assets in theoretically am-
biguous. To control for utilization, I use the Relative Risk Score variable constructed
using the RiskSmart Version 2.2 software created by DxCG Inc. 23 This software uses
information on age, sex, and medical diagnoses to project expected medical utiliza-
tion based on regression models developed by the company. Because the software
does not use geographical information to project utilization, the Relative Risk Score
is orthogonal to asset exemption laws and other state-level factors.
1.3.4 Insurance Premiums
I conduct additional analysis using data on health insurance premiums in the indi-
vidual market. In particular, I use data on premium quotes in each state that are
listed on eHealthInsurance, a website that aggregates premium quotes from most of
the major insurance providers. The data I use were collected in November, 2010.
I collect premiums in each state for a 30-year-old non-smoking male. Because pre-
miums quotes are zip code specific, the data are collected for a zip code randomly
selected from the 10 most populous zip codes in the state. Along with premiums,
I collect data on the insurance provider, plan brand name, deductible, coinsurance
rate, and co-payment for a office visit. I define an insurance plan as all observations
with the same insurer, brand name, deductible, coinsurance and co-payment. Table
1.14 shows basic summary statistics for the data. The data set covers 41 states and
1,891 plans. The mean premium is $103 per month, the mean deductible is $3,351,
and the mean coinsurance rate is 15 percent.
23 See form HC-092 on the MEPS website for a full description of the construction of this variable.

CHAPTER 1. BANKRUPTCY 17
levels of seizable assets. For small charges the three groups make similar out-of-pocket
payments, consistent with most households being below the cap. For large charges,
out-of-pocket payments sharply diverge. While households with less than $10,000 in
seizable assets have their out-of-pocket payments capped, households with more than
$50,000 in seizable assets see their out-of-pocket payments continue to scale up with
charges, consistent with a cap that is increasing in seizable assets.
To test the capping-of-cost prediction more rigorously, I estimate regression mod-
els of out-of-pocket payments on seizable assets, conditioning on the level of medical
care received. Controlling for medical utilization is important because the sign of
the unconditional effect of seizable assets on out-of-pocket payments is theoretically
ambiguous. To see this, consider the effect of reducing a household’s level of seiz-
able assets. Due to the mechanical effect of the implicit insurance from bankruptcy,
out-of-pocket payments should decrease. On the other hand, due to moral hazard,
households may increase their medical utilization, raising out-of-pocket costs and
potentially offsetting the mechanical effect in the opposite direction.
For the analysis, I restrict the sample to uninsured households with positive med-
ical utilization. Letting i denote households and j denote states, the second stage
cost equation takes the form
ln(OOPij +1)=βwln(w S ij)+f(RRSi; βRRS)+XijβX + νij, (1.4)
where the dependent variable is log out-of-pocket payments, w S ij is seizable assets,
f(RRSi; β k m) is a fourth-order polynomial in the Relative Risk Score, Xij is a vector
of household and state characteristics, and νij is the error term. 26 The capping of
cost prediction is supported by a positive coefficient on seizable assets (βw > 0).
1.4.3 Cross-State Variation
Consistently estimating the parameters of interest poses four distinct identification
problems. The first issue is omitted variables: that coverage or costs and seizable
26 The depend variable is rarely zero. In the sample analyzed, less than 4 percent of households
make zero out-of-pocket payments.

CHAPTER 1. BANKRUPTCY 20
or coverage over the course of the twentieth century. Moreover, historical evidence
shows that 1920 homestead exemption levels resulted from an idiosyncratic array
of nineteenth century historical circumstances. Describing the key factors driving
the establishment of homestead exemptions in the nineteenth century, ? cites no
less diverse a list than “Texas colonizers and western developers, labor and land
reformers, antimonopoly Jacksonian egalitarians, defenders of family security and
women’s property rights, southern planters and yeomen devastated by the Civil War.”
These heterogenous causes reduce the concern that historical homestead exemptions
merely proxy for a persistent state-level characteristic such as the strength of the
pro-debtor political movement.
Importantly, historical homestead exemptions are also a good predictor of con-
temporaneous exemptions values. 27 Figure 1.7 shows this graphically, plotting the
average level of seizable home equity for a constant, nationally representative sample
of households under 2005 homestead exemptions (y-axis) and inflation-adjusted 1920
homestead exemptions (x-axis) in each state. The plot also shows the fitted line from
a bivariate regression. As the slope coefficient indicates, homestead exemptions have
become less generous over time, with seizable assets increasing on average by 90 per-
cent. The R-squared value is 0.43, with the New England states in the lower right
corner being the most prominent outliers. 28
I take two further approaches to reduce the concern that unobservable factors
are driving the effect. The first is to sequentially add controls for Census Regions
(e.g., Northeast) and Census Divisions (e.g., New England) to the main specification.
Stable results across these specifications should address the concern that the effects
are being driven by a spatially-correlated, unobserved factor.
The second is to control for a rich set of state-level legislative covariates. In
the coverage equations, I control for insurance market regulations (e.g., community
rating requirements, coverage mandates) that may affect premiums. 29 While the
27Many of the changes since 1920 have simply been inflation updates passed by individual state
legislatures (?).
28A keyword search of newspaper articles in a six-month window around major changes in Massachusetts
and Connecticut assets exemptions failed to reveal any information on the reasons for
these increases.
29The data on these regulations was taken from a Blue Cross Blue Shield (2002) compilation

CHAPTER 1. BANKRUPTCY 21
publicly insured are excluded from the baseline sample, correlation between asset
exemption laws and eligibility thresholds for public insurance programs might bias the
estimates though sample selection. To assuage this concern, I estimate the regression
models on samples excluding and including the publicly insured. I also control for
the presence and generosity of Medicaid Medically Needy programs that provide an
alternative form of safety net coverage. 30 In the cost equations, I control for hospital
characteristics and other state-level factors (e.g., share of private hospitals) that may
affect out-of-pocket payments. 31
1.5 The Effect on Insurance Coverage
1.5.1 First Stage Estimates
Table 1.2 shows implied first stage regressions of log seizable assets on different instru-
mental variables. Column 1 shows estimates with the baseline instrument. Columns
2 and 3 show estimates with instruments that isolate the variation due to seizable
homestead equity and seizable non-homestead assets. These instruments are calcu-
lated by taking a constant, nationally representative of households and calculating
their level of seizable homestead and non-homestead equity as though they lived in
each state. Column 4 shows estimates using the instrument that isolates variation
due to 1920 homestead exemptions. All the specifications include demographic con-
trols (age group, family structure, race, education, and income), state controls (mean
income, percent unemployed, percent covered by Medicaid), and year fixed effects.
Standard errors in all specifications here and throughout the paper are clustered at
the state level. The instruments have substantial power. The baseline instrument
has an F-statistic of nearly 500 (column 1) and even the instrument based on 1920
gracious shared by Amanda Kowalski.
30I thank Jay Bhattacharya for alternating me to the presence of these programs. I use 2003 data
on these programs taken from ?.
31In particular, I control for the share of private and nonprofit hospitals, taken from the Hospital
Statistics 2005 published by the American Hospital Association. I control for Disproportionate Share
Hospital payments per 1,000 residents taken from the Kaiser Family Foundation. I control for the
number of Federally Qualified Health Centers per 100,000 residents.

CHAPTER 1. BANKRUPTCY 22
homestead exemptions has an F-statistic over 20 (column 4).
1.5.2 Baseline Coverage Estimates
Before turning to the regression estimates, Figure 1.8 presents visual evidence on the
crowd-out prediction. Plots on the same row show the exact same data. Datapoints
are indicated with state abbreviations in the plots in the left column and with circles
proportional to the number of observations in the plots in the right column. Panels
A and B plot insurance coverage against seizable assets averaged by state. In the raw
data, there is a strong upward-slopping relationship. The relationship is consistent
across most states, with the outliers mainly states with very few observations. Panels
C and D plot insurance coverage against the baseline instrument averaged by state—
the graphical analogue to a bivariate reduced form regressions. The upward-sloping
relationship is similar although the figures are more noisy. (I discuss Panels E and F
below.)
Table 1.3 presents the baseline regression specifications. The first four columns
show OLS estimates. Column 1 only has controls for year fixed effects. Column 2 adds
the demographic and state controls. Columns 3 and 4 add controls for wealth and
premiums. 32 Since wealth and premiums may be directly affected by asset exemption
laws, the parameter estimates with these controls are inappropriate for counterfac-
tuals involving changes in these exemptions. The coefficient on log seizable assets is
6.13 in column 1 and about 2.5 across the other specifications. The similar estimates
with the wealth and premium controls suggest that the effect is not mediated through
these channels. 33
Columns 5 to 8 show the 2SLS estimates with the baseline instrument (column 1
of Table 1.2), adding controls in the same manner as columns 1 through 4. The co-
efficient on log seizable assets is 5.43 in the preferred specification with demographic
32The premium index is state fixed effects from a regression of log premiums on plan and state
fixed effects.
33The slight increase in columns 4 and 8 is due to the fact that premium data is only available
for a subset of states. When compared to a specification without this control estimated on the same
sample, controlling for premiums very slightly reduces the coefficient on log seizable assets. This is
consistent with a negative correlation between asset exemption levels and insurance premiums.

CHAPTER 1. BANKRUPTCY 23
and state controls (column 6) and significantly positive at more than the 0.1 percent
level. Like the OLS specifications, adding controls for wealth (column 7) and pre-
miums (column 8) has little effect. The larger 2SLS estimates are consistent with
measurement error attenuating the coefficient of interest in the OLS specifications.
The reduced form panels of Figure 1.8, discussed above, suggested that Texas
might be important to the 2SLS results. Figure 1.9 examines the robustness of the
estimates to this type of concern by showing the coefficient on log seizable assets
from 51 separate regressions of the preferred specification (column 6 of Table 1.3)
that sequentially drop households from the indicated state. The figure shows that
neither Texas—nor any other state for that matter—is driving the results. Dropping
Texas does slightly reduce the coefficient on log seizable assets. However, the estimate
without Texas is significantly greater than zero and slightly larger than the estimate
that excludes the state of Arizona.
1.5.3 Sensitivity Analysis
As discussed above, I address the potential bias from endogenous asset exemption laws
with three strategies. I include controls for state legislative factors. I add controls
for Census Regions (e.g., Northeast) and Census Divisions (e.g,. New England). And
I show specifications that use variation due to 1920 homestead exemptions as an
instrument.
Column 1 of Table 1.4 presents estimates of the preferred specification on a sam-
ple that includes households with public insurance and column 2 adds controls for
state insurance regulations and Medicaid Medically Needy programs to the preferred
specification. Columns 3 and 4 sequentially add fixed effects for Census Regions and
Census Divisions. The coefficient on log seizable assets is stable across these speci-
fications, reducing concerns that the estimates are driven by unobserved state level
factors.
Column 5 applies a variant of the baseline instrument that is calculated by averag-
ing log seizable assets for the constant, nationally representative sample of households
by age-by-educations groups rather than at the population level. The coefficient on

CHAPTER 1. BANKRUPTCY 24
log seizable assets is very similar with this instrument. Columns 6 and 7 apply in-
struments that isolate the variation due to seizable homestead equity and seizable
non-homestead assets respectively. The coefficient is a smaller 4.53 (p-value ¡ 0.05)
when estimated using variation in homestead exemptions and a larger 6.16 (p-value
¡0.01) when estimated using variation in non-homestead seizable assets.
Before discussing the eighth column, I return to Panels E and F of Figure 1.9.
These panels show reduced form plots of insurance coverage against log seizable home
equity under inflation-adjusted 1920 homestead exemption laws averaged by state.
The historical instrument plots show a similar upward-sloping relationship and are
somewhat less noisy than the baseline instrument plots in the row above. One reason
for this is that the New England states—which are outliers in Panels C and D—have
greater seizable assets levels under the historical exemptions and are closer to the
regression line.
Column 8 of Table 1.4 shows the 2SLS estimates with the historical instrument.
With the historical instrument, the coefficient on log seizable assets is slightly larger
than the preferred estimate (6.39 versus 5.43). A Hausman test cannot reject the va-
lidity of the baseline instrument. Given that the historical instrument has lower power
in the first stage, it seems prudent to maintain the baseline instrument parameter of
5.43 as the preferred estimate.
1.5.4 Heterogeneity in the Effect on Coverage
Table 1.5 examines heterogeneity in the coefficient on log seizable assets by estimating
the preferred specification on subsamples of the data. The 2SLS estimate is somewhat
larger when estimated on the sample of younger versus older households (defined by
a household head younger than 35) and the sample of renters versus homeowners.
Households earning less than the median income are significantly more responsive to
seizable assets levels (8.18 in column 5 versus 2.79 in column 6). This is consistent with
?, who find that low income households are more premium elastic in their demand for
health insurance. The coefficient on log seizable assets is larger for households with
seizable assets below the median value (8.24 in column 7 versus 4.87 in column 8).

CHAPTER 1. BANKRUPTCY 25
1.5.5 The Effect on Wealth and Premiums
Table 1.3 showed that including controls for wealth and premiums did not have much
of an effect on the parameter of interest. Tables 1.6 and 1.7 examine the effects of
asset exemption law on these variables on their own.
Table 1.6 shows OLS and 2SLS estimates of assets on state-level measures of asset
exemption law (the instruments). Columns 1 to 6 examine the effects of exemptions
for specific asset categories (home equity and vehicle equity). In these specifications,
the dependent variable is an indicator for positive assets or the log asset value. All
specifications include households controls—including a fourth-order polynomial in
wealth—and the baseline instrument to control for the overall generosity of asset
exemption laws in that state. Columns 1 and 2 show estimates of home equity on the
measure of homestead exemptions laws. Columns 3 and 4 show similar specifications
with homestead exemptions from 1920 used as an instrument. Columns 5 and 6 show
estimates with a measure of vehicle equity as the dependent variable. None of the
specifications show a statistically significant relationship between asset exemption
laws and assets.
Columns 7 and 8 of Table 1.6 examine the effect of asset exemption law on overall
wealth. In these specifications, the dependent variable is an indicator for positive
wealth or the log wealth value. The parameter on the baseline instrument—which
is simply a measure of the generosity of asset exemption laws—is the coefficient of
interest. The specifications include the standard set of demographic controls except
for the polynomial in wealth. Like the estimates for specific asset categories, there is
no evidence of relationship between asset exemption laws and wealth levels.
Table 1.7 examines the effect of asset exemption law on health insurance premi-
ums by showing estimates of log premiums on the baseline instrument. Observations
in these regressions are monthly premiums for a 30-year-old non-smoking male. All
specifications include plan fixed effects so that the parameters are identified off dif-
ferences in premiums for the same insurance product across different states.
The estimates show a marginally significant, negative relationship between asset
exemption laws and insurance premiums, consistent with higher asset exemptions
inflating hospital costs. Columns 1 to 3 show OLS specifications of log premiums

CHAPTER 1. BANKRUPTCY 26
on the baseline instrument. Column 1 has only plan fixed effects. Column 2 adds
state demographic factors. Column 3 includes controls for state insurance market
regulations. The estimates are quite similar across specifications and indicate that a
log point increase in seizable assets is associated with a 7 to 10 percent decrease in
premiums, although the effect is only significantly less than zero at the 10 percent
level. Columns 4 to 6 shows estimates where seizable home equity under 1920 as-
set exemption laws is used as an instrument for contemporaneous asset exemptions.
Controls are added analogously to columns 1 to 3. The estimates are slightly more
negative but less precisely estimated. Column 6, which includes the full set of con-
trols, indicates that a log point increase in asset exemptions decreases premiums by
16 percent and is significantly negative at the 5 percent level.
1.5.6 Summary and Interpretation
Supporting the crowd-out prediction, the estimates show a robust positive relationship
between insurance coverage and seizable assets. The preferred estimate indicates that
a log point increase in seizable assets raises insurance coverage by 5.43 percentage
points. The effect is similar when identified using variation due to 1920 homestead
exemptions, and is stable to the introduction of controls for Census Regions and
Census Divisions and well as relevant state level legislative factors.
There are a number of ways to put this estimate in context. Perhaps the most
natural way—given the source of identifying variation—is to consider counterfactuals
involving cross-state variation in asset exemption law. In the sample, 77.3 percent
of individuals are insured (recall that the sample excludes the elderly and those with
public insurance). If the exemption laws of the most debtor-friendly state (Texas)
were applied nationally, the estimates suggest that the number of uninsured would
be increased by 30.0 percent (6.8 percentage points). 34 If the exemptions laws of
the least debtor-friendly state (Delaware) were applied nationally, 16.3 percent (3.7
percentage points) of the uninsured would take-up insurance. 35 Take-up under the
34 Couples filing jointly in Texas can exempt all of their home equity, an unlimited amount of
retirement assets, and up to $60,000 of any asset of their choose. See Table 1.1 for details.
35 Delaware gives couples filing jointly an unlimited exemption for retirement assets and a $500

CHAPTER 1. BANKRUPTCY 27
least debtor-friendly laws is smaller in magnitude because most uninsured household
have little wealth and are thus not exposed to substantially more financial risk under
these laws.
Achieving the same increase in coverage as implied by the Delaware laws would
require a large premium subsidy. Too see this, I use the -0.084 take-up premium
semi-elasticity for the uninsured estimated by the ?. 36 With this elasticity, inducing
a 3.7 percentage points increase in coverage requires a premium subsidy of 44.0 (=
100 × (0.037/0.084)) percent.
The effect does not seem to be driven by an endogenous response of either wealth
or premiums. Assets are completely unresponsive to asset exemption laws. And while
a log point increase in asset exemptions is associated with a (marginally significant)
7 percent decrease in insurance premiums, controlling for premiums does not impact
the coefficient on seizable assets. This is consistent with findings in the literature
that the demand for health insurance is highly premium inelastic. For example, with
the -0.084 premium semi-elasticity discussed above, a 7 percent decrease in insurance
premiums is projected to increase insurance coverage by only 0.6 (=-0.084 × -0.07)
percentage points.
1.6 The Effect on Costs
1.6.1 Baseline Estimates
Table 1.8 presents the baseline OLS and 2SLS estimates of the effect on costs. The
sample includes all uninsured households with positive medical utilization. Standard
errors are clustered at the state level in all specifications. In the OLS specification
without controls (column 1), the elasticity of out-of-pocket payments with respect
wildcard exemption. Households cannot exempt any home or vehicle equity in excess of the $500
wildcard. See Table 1.1 for details.
36 The ? estimate of -0.084 is identified off premium variation due to state-level community rating
and premium compression regulations. The estimate is central to the small number of estimates in
the literature. It is smaller than the estimate of ?, who use the introduction of a tax subsidy for
insurance purchases by the self-employed, and is larger than the estimate from ?. See? for a review
of estimates in the literature.

CHAPTER 1. BANKRUPTCY 28
to seizable assets is 0.56. Including controls for household demographics and uti-
lization reduces this coefficient to 0.22 (column 2). Controlling for wealth has little
incremental effect (column 3).
The 2SLS estimates using the baseline instrument are shown in columns 4 to 6,
adding controls in the same manner as columns 1 to 3. With controls, the 2SLS
point estimates are slightly larger than the corresponding OLS specifications. In the
preferred specification with controls for demographics and utilization (column 5), the
elasticity of out-of-pocket payments with respect to seizable assets is 0.37. As before,
controlling for wealth (column 6) has little effect.
1.6.2 Sensitivity Analysis
Table 1.9 examines the sensitivity of these results to concerns about potential bias
from endogenous asset exemptions. Column 1 adds controls for hospital and state
characteristics. Because hospitals with different ownership structures may have differ-
ent incentives and preferences for charity care, I include controls for the share of public
and nonprofit hospitals in each state (private hospitals are the omitted category). I
control for Disproportionate Share Hospital (DSH) payments per 1,000 residents as
these may affect incentives to provide unpaid care, the presence of state charity care
pools, and the number of Federally Qualified Health Centers per 100,000 people. The
parameters of interest are virtually unchanged in these specifications and none of the
covariates have coefficients that are statistically distinguishable from zero.
Column 2 uses the instrument that isolates variation due to contemporaneous
homestead exemptions and column 3 uses the instrument that isolates variation due
to exemptions for non-homestead assets with no effect on the parameter of inter-
est. Column 4 applies the instrument that isolates variation due to 1920 homestead
exemptions. Using this source of variation addresses concerns that contemporane-
ous asset exemptions may be influenced by unobserved state-level factors such as
the strength of the hospital lobby or state-level attitudes towards charity care. The
estimates are very similar with this instrument.

CHAPTER 1. BANKRUPTCY 29
The final two columns of Table 1.9 examine heterogeneity in the effect by utiliza-
tion level. The capping-of-cost prediction and the simple descriptive evidence shown
in Figures 1.2 and 1.4 suggest that seizable assets should have a larger impact on
households with high utilization, as the “deductible” of bankruptcy insurance is more
likely to bind for these households. Columns 5 and 6 allow the effect to vary by
whether utilization—as measured by the Relative Risk Score—is above the 90th per-
centile (high utilization) or below this level (low utilization). The elasticity parameter
is 0.51 for high utilization and about 0.37 for low utilization, consistent with a high
deductible structure for this insurance.
1.6.3 Summary and Interpretation
Figures 1.2 and 1.4 showed that uninsured households with fewer seizable assets have
their out-of-pocket payments capped in the raw data. The regression estimates show
that this effect is robust to demographic controls and variation in seizable assets
solely due to cross-state differences in asset exemption law. The preferred estimate
indicates that a log point increase in seizable assets raises out-of-pocket payments by
37 percent, conditional on medical utilization. Consistent with the high-deductible
structure of this insurance, the elasticity is larger for households with high utilization.
1.7 Micro-Simulation Model
To shed light on puzzles in the literature and examine policy implications, I calibrate
a micro-simulation model of insurance choice. The model is based on the set of unin-
sured and privately insured households in the 2005 PSID. Households face household-
specific medical cost distributions that depend on the age and sex of each household
member. Premiums for conventional health insurance are based on these costs, scaled
to account for moral hazard, administrative costs, and the cross-subsidization of un-
paid care to the uninsured. Following closely the formulation in Section 1.2, the model
uses an expected utility framework to implicitly define each household’s willingness
to pay for conventional health insurance. Households purchase coverage if and only

CHAPTER 1. BANKRUPTCY 31
CARA utility. 39 I show results with risk aversion parameters of α =2.5 × 10 −5 (low
risk aversion), α =5.0 × 10 −5 (moderate risk aversion), and α =7.5 × 10 −4 (high risk
aversion). Dividing by the median wealth level of $40,318, these parameters can be
interpreted as relative risk coefficients of γ = 1, 2, and 3. Household-level medical cost
distributions are constructed using individual-level medical cost distributions for age-
by-sex-by-insurance status cells in the 2005 MEPS. The markup for moral hazard
is µ1 = 0.56, the value implied by the RAND health insurance experiment price
elasticity of -0.22 in an arc elasticity framework (?). The administrative load factor
is set to µ2 =0.1 for households with employer sponsored insurance and µ2 =0.5
for the uninsured and households in the non-group market. 40 I solve for the cross-
subsidization parameter µ3 endogenously using the cost distributions of uninsured
households in the model. In Appendix Section 1.10.3, I discuss the construction of
the medical cost distributions and premiums in more detail. I show that the calibrated
premiums closely match quoted premiums in the individual market.
1.8 Puzzles and Policy
In this section of the paper, I use the micro-simulation model to investigate how
bankruptcy insurance sheds light on puzzles in the health policy literature and to
examine implications of this mechanism for the design of health insurance policy.
1.8.1 Puzzles
Puzzle 1: Why are 47 million individuals uninsured?
Explaining why households are uninsured is a central puzzle for scholars of health
insurance. In his review of the literature ? writes, “there are a variety of hypothe-
ses for why so many individuals are uninsured, but no clear sense that this set of
explanations can account for the 47 million individuals.” Bankruptcy insurance is a
39 Using a CARA specification avoids the problems associated with non-positive wealth. Calibrations
with CRRA utility and a consumption floor generate stronger results.
40 This values are take from ?.

CHAPTER 1. BANKRUPTCY 32
compelling additional explanation. I used regression analysis to quantify the impor-
tance of this mechanism in the sections above. I use the micro-simulation model for
an alternative perspective.
Table 1.10 shows the percent of insured and uninsured individuals predicted by
the model to purchase a $2,000 deductible plan. Column 1 shows these predicted
values from a simulation where there is no bankruptcy insurance. Column 2 adds
bankruptcy insurance to the model. Without bankruptcy, insured and uninsured
individuals are difficult to separate. Using the variation in health risk, administrative
costs, and tax preferences, the model can only separate predicted coverage by 8 to 28
percentage points. Because there are large differences in seizable assets between these
groups, accounting for bankruptcy substantially improves the explanatory power of
the model, expanding the gap in predicted coverage to 43 to 54 percentage points.
Bankruptcy insurance thus to increases the gap in predicted coverage by 14 to 46
percentage points.
Puzzle 2: The low take-up of high deductible plans
A second puzzle is the low take-up of high deductible health plans (HDHP) by the
uninsured. HDHPs were intended by their proponents to expand insurance coverage,
yet despite offering low premiums they have not been successful in this regard (?).
Crowd-out from implicit insurance is an appealing explanation for this failure. Be-
cause the median uninsured household has less than $5,000 in seizable assets, HDHPs
are largely redundant to the implicit insurance they already hold. Figure 1.10 makes
this case quantitatively, plotting the percent of uninsured households projected to
purchase insurance by deductible level. Without bankruptcy, the probability of pur-
chase increases sharply from 19 percent for a $1,000 deductible plan to 43 percent for
a $5,000 deductible plan, as the concavity of utility makes insurance more valuable at
higher deductible levels. With bankruptcy, this increase is virtually eliminated, rising
from 2 percent for a $1,000 deductible plan to only 6 percent for a $5,000 deductible
contract, a full 37 percentage points below the projected value without bankruptcy.

CHAPTER 1. BANKRUPTCY 33
Puzzle 3: Rising risk, falling coverage
A third puzzle is the association between rising health insurance costs and falling cov-
erage. ? show that more than half the decrease in insurance coverage over the 1990s
can be explained by rising premiums. Yet as the authors explain, from the standpoint
of economic theory this is counterintuitive. With standard risk preferences, rising un-
derlying costs should lead to increased insurance coverage. Taking bankruptcy into
account, however, reverses this intuition. The decrease in coverage can be explained
by households substituting conventional health insurance for bankruptcy insurance
that is increasing in actuarial value without increasing in price.
Puzzle 4: The insurance generosity gap
A fourth puzzle is the insurance generosity gap. In his review of the literature, ? asks
why most U.S. households appear to be either under-insured or over-insured but rarely
in-between. Implicit insurance from bankruptcy can explain this finding. To illustrate
how bankruptcy insurance creates a generosity gap, Figure 1.11 presents a stylized
budget set. Without bankruptcy, households face the standard continuous tradeoff
between insurance generosity (y-axis) and other goods (x-axis). Implicit insurance
generates a notch: households receive some implicit insurance without giving up other
goods. Convex preferences give rise to an insurance generosity gap, with households
sorting into the more generous conventional health insurance and the less generous
implicit insurance from bankruptcy.
1.8.2 Policy Implications
I use the micro-simulation model to examine welfare implications of this mechanism.
From a policy design perspective, bankruptcy insurance has many potential draw-
backs: It forces the uninsured to receive care in emergency rooms, which are often
not the most appropriate settings (?), and may lead to inefficiently low levels of
preventative care, inflating overall costs (?). Negotiation under the threat-point of
bankruptcy is probably not the most efficient way of processing payments and there
are deadweight costs and externalities to formal bankruptcy.

CHAPTER 1. BANKRUPTCY 34
For this exercise, however, I focus on a single problem: because bankruptcy insur-
ance has a price of zero but cross-subsidized costs, too many households choose to be
uninsured. I find this inefficiency particularly interesting because it directly relates
to the insurance coverage decision and because the problem is naturally addressed
with penalties for being uninsured—a key and controversial element of PPACA. 41
Of course, conventional health insurance has well-documented inefficiencies as
well. Relative to bankruptcy insurance, moral hazard is particularly relavent. 42 With
conventional health insurance it is well known that medical providers and patients
often have incentives to supply and demand excess medical care. With bankruptcy
insurance, on the other hand, physicians are more likely to be exposed to the social
cost of their decisions and patients have little leverage to demand excess treatment.
Thus, for the optimal “Pigovian penalties,” low moral hazard bankruptcy insurance
may be the efficient choice for some households.
Table 1.11 shows the welfare effects of different penalty systems. For each penalty
system, I allow households to choose between conventional insurance at the cali-
brated premiums and bankruptcy insurance at the cost of the penalty. 43 The results
are shown relative to a baseline in which households can choose bankruptcy at no
cost. The first set of rows shows coverage and welfare under the optimal Pigovian
penalties—defined as the household-specific actuarially fair cost of the implicit insur-
ance from bankruptcy. 44 The optimal penalties average $218 per person and induce
7 to 8 percent of the uninsured to take-up coverage. As indicated by the higher
willingness to pay, these households purchase more generous coverage than they had
from bankruptcy. Due to the increased moral hazard, however, costs rise by almost
as much. The net effect is an increase in surplus of only $4 to $5 per person.
The second panel shows the welfare effects of the PPACA penalty system. When
fully implemented in 2016, these penalties will equal the greater of $625 or 2.5 percent
41For this exercise, I assume that uninsured households are not already subsidized through the
tax code or some other channel.
42Empirically, adverse selection does not seem to affect insurance choice on the extensive margin
(??).
43Recall that calibrated premiums are calculated using costs scaled up for moral hazard, administrative
loading, and the cross-subsidization of the uninsured.
44That is, expected medical costs above seizable assets.

CHAPTER 1. BANKRUPTCY 35
of income per household, up to a maximum of $2,085. Under these penalties, deflated
to 2005 levels assuming trend inflation, take-up ranges from 50 to 64 percent. The
generosity of coverage does not change much on average while—due to higher moral
hazard—costs increase by a significant amount. The net effect is a decrease in surplus
of $9 to $13 per person.
The third panel considers the welfare effects of preventing households from dis-
charging medical debt in bankruptcy—the exact exercise performed in coverage and
cost counterfactuals. Without bankruptcy, all households purchase conventional in-
surance. As indicated by the lower willingness to pay, these plans are less generous
on average. Because of the higher moral hazard from conventional insurance, costs
do not shrink by a commensurate amount. The net effect is a reduction in surplus
of $36 to $43 per person. While the exact numbers, of course, are a function of the
particular calibration parameters, the takeaway point from this last exercise is a sim-
ple message: despite increasing insurance coverage, eliminating bankruptcy insurance
may not be socially desirable.
1.9 Conclusion
Understanding why households choose health insurance is fundamental to positive and
normative analysis of health insurance policy—yet the insurance coverage decision is
not well understood. The objective of this paper is to describe and evaluate how the
implicit insurance from bankruptcy bears on this decision. In the first part of the
paper, I argued that the fact that most medical care is provided on credit coupled
with the ability of households to discharge this debt for seizable assets in bankruptcy
provides households with a form of high-deducible health insurance.
I next evaluated the quantitative significance of this bankruptcy insurance. Ex-
ploiting variation in seizable assets that is orthogonal to wealth and other household
characteristics, I found that households with more seizable assets make higher out-
of-pocket payments and are substantially more likely to be insured. The estimates
suggest that if medical debt could not be discharged in bankruptcy, 16.3 percent
of the uninsured would purchase coverage. Achieving the same increase in coverage

CHAPTER 1. BANKRUPTCY 38
education loans. While the PSID and MEPS distinguish remaining housing and vehi-
cle principle from other debts, these other debts are not further decomposed. 45 The
leading potential concern is that these other debts include a substantial amount of
education debt—particularly for households with recent college graduates who may
also be on the margin of insurance coverage. To overcome this issue, I estimate the
share of education debt in total unsecured debt in the 2004 Survey of Consumer
Finances (SCF) and use this estimate to project education debt in the PSID and
MEPS. In particular, I use parameter estimates from a probit regression of the share
of education debt in total unsecured debt on income, family structure, age of the
household head, and unsecured debt level fully interacted with a categorical variable
for educational attainment of the head to project education debt values.
1.10.3 Micro-Simulation Details
Medical Cost Distributions
I construct the household-level medical cost distributions using individual-level med-
ical cost data from the 2005 MEPS for age-by-sex-by-insurance status cells. 46 For
insured individuals, costs are defined as total payments. For uninsured individu-
als, my measure of costs in construct in the following way: I start with medical
charges as this variable accounts for medical services written off as charity care or
bad debt. I then scale down charges by the cost-charge ratio (CCR) for the privately
insured population to account for the discount typically extended to the uninsured. 47
45 The debt variable in the PSID is based on the question “Aside from the debts that we have
already talked about, like any mortgage on your main or vehicle loans – do [you/you or anyone in
your family] currently have any other debts such as credit card charges, student loans, medical or
legal bills, or loans from relatives?” In the MEPS it is based on the question “Does anyone in the
family have any debts that we haven’t asked about, such as credit card balances, medical debts,
life insurance policy loans, loans from relatives, and so forth?” While both these variables may
include additional non-dischargeable debt (e.g., department store loans collateralized by durable
goods, unpaid taxes), the bias introduced by this is likely to be small.
46 The age-by-sex groups are no more than 18 years old, male age 19 to 34, female age 19 to 34,
male age 35 to 64, and female age 35 to 64.
47 Recall from Figure 1.2 that privately insured and uninsured households make similar payments
for low charges.

CHAPTER 1. BANKRUPTCY 40
TaxSim program.
Appendix Table 1.15 shows the resulting distributions of medical costs and pre-
miums normalized by household size for comparison. Costs have a long right tail:
while 91 percent of persons incur less than $1,000 in a given year, about 1 in 36 incur
more than $5,000 and 1 in 69 incur more than $10,000 in medical costs. Calibrated
premiums drop steeply from $2,126 per person for a plan with first dollar coverage to
$1,009 per person for a $10,000 deductible plan.
As a reality check, Appendix Table 1.15 compares these premiums to quoted
individual market premiums for the median uninsured individual in the data (a 32-
year-old male), adjusted for inflation and cost-sharing. 48 The calibrated and market
premiums are quite similar. The calibrated premiums are slightly less expensive
for low deductible levels and somewhat more expensive for high deductibles. This
difference could be explained by selection or by heterogeneity in the moral hazard
parameter across the expenditure distribution.
48 The individual market premiums, from www.ehealthinsurance.com, are for a non-smoking 32year-old
male. The policies are issued by Aetna and start in May 2010. They are adjusted for
inflation using the Medical Care component of the CPI-U.

Percent by insurance status
CHAPTER 1. BANKRUPTCY 41
70%
60%
50%
40%
30%
20%
10%
0%
Figure 1.1: Histogram of Seizable Assets by Insurance Status

Payments
CHAPTER 1. BANKRUPTCY 42
$35,000
$30,000
$25,000
$20,000
$15,000
$10,000
$5,000
$-
Figure 1.2: Payments vs. Charges by Insurance Status
Privately insured
Uninsured
$- $10,000 $20,000 $30,000 $40,000 $50,000 $60,000 $70,000 $80,000
Charges
Notes: Payments versus medical charges for uninsured and private insured households.
Payments are the sum of out-of-pocket payments and payments from private insurers. The
plot is created by averaging payments and charges at 20ths of the charge distribution.
Pooled 2000 to 2005 MEPS, inflation-adjusted to 2005 using the CPI-U. Excludes
households with a head age 65 or older. Household-level estimates weighted by number of
individuals per household for interpretation at the individual level.

CHAPTER 1. BANKRUPTCY 43
Figure 1.3: Provider Billing Decision: Submitted Bill vs. List Price of Medical Care
s (submitted bill)
w S (seizable assets)
Formal bankruptcy
m (list price of medical care)
s = m
s̅(ŵ S )
Informal bankruptcy
s̅(ŵ S )

OOP payments
CHAPTER 1. BANKRUPTCY 44
$9,000
$8,000
$7,000
$6,000
$5,000
$4,000
$3,000
$2,000
$1,000
$-
Figure 1.4: Payments vs. Charges by Seizable Assets
wS ! $50,000
$10,000 " wS < $50,000
wS< $10,000
$- $5,000 $10,000 $15,000 $20,000 $25,000 $30,000 $35,000 $40,000 $45,000
Charges
Notes: Out-of-pocket payments versus medical charges for uninsured households with low
(¡ $10,000), moderate ($10,000 to $50,000), and high (≥ $50,000) levels of seizable assets.
The plot is created by averaging payments and charges at 20ths of the charge distribution.
Pooled 2000 to 2005 MEPS, inflation-adjusted to 2005 using the CPI-U. Excludes
households with a head age 65 or older. Household-level estimates weighted by number of
individuals per household for interpretation at the individual level.

Simulated instrument
CHAPTER 1. BANKRUPTCY 45
11
10.5
10
9.5
9
8.5
8
Figure 1.5: Simulated Instrument by State
DE PANJMIOHSCNELAALKYTNWYILMONCUTHIINNYMDGAORWAIDCACONMWIVAMANDCTARMTAKIAFLMNOKAZSDWVMSMEVTNHRINVKSDCTX
States
Notes: The simulated instrument is log seizable assets for a constant, nationally
representative sample of households as though they lived in each state. The sample is
made up of uninsured and privately insured households, excluding households with a head
age 65 or older. Pooled 1999 to 2005 PSID, inflation-adjusted to 2005 using the CPI-U.
See text for details on the seizable assets calculation.

ln(seizable sssets)
CHAPTER 1. BANKRUPTCY 46
Figure 1.6: Log Seizable Assets Percentiles by State for a Constant Sample of Households
11.5
11
10.5
10
9.5
9
8.5
8
7.5
DE PANJMISCOHNEALTNKYINILMOMDWYNCLAGANYUTVAHIORWAWIIDNMCACONDAKWVCTIAMAARMTFLSDMNOKAZMSMEVTNHRINVKSDCTX
Notes: Percentiles of log seizable assets for a constant, nationally representative sample of
households as though they lived in each state. The sample is made up of uninsured and
privately insured households, excluding households with a head age 65 or older. Pooled
1999 to 2005 PSID, inflation-adjusted to 2005 using the CPI-U. See text for details on the
seizable assets calculation.
State
50th
40th
30th
20th

CHAPTER 1. BANKRUPTCY 47
Figure 1.7: Seizable Homestead Equity in 2005 vs. 1920 for a Constant Sample of
Households
Seizable home equity in 2005
0 20,000 40,000 60,000 80,000
Slope = 1.96 (0.38)
R-squared = 0.42
ID
CA
NV
TX
AZ
MS
AR
AL
NE
LA
CO
MI
GA
MO
40,000 50,000 60,000 70,000 80,000
Seizable home equity in 1920
Notes: Mean seizable home equity for a constant, nationally representative sample of
households under 2005 and inflation-adjusted 1920 homestead exemption laws as they
lived in each state. The New England states are shaded red. The sample is made up of
uninsured and privately insured households, excluding households with a head age 65 or
older. Slope coefficient from a bivariate regression weighted by population, with a robust
standard error in parentheses. Pooled 1999 to 2005 PSID, inflation-adjusted to 2005 using
the CPI-U. States that did not exist or had acre-based homestead exemption laws are
excluded. Household-level estimates weighted by number of individuals per household for
interpretation at the individual level.
WY
VA
UT
WV
KY OH SC
NC IL
NY NJ
TN
WA
NM
CT
MA
IN
ME
VT
NH
PA
MD DE
RI

50 60 70 80 90 100
Insurance coverage (percentage points)
Figure 1.8: Insurance Coverage vs. Seizable Assets
ND
ME
MT
VT
MN
CT
MA
AK
AL
AR AZ
IA PA
DC
MS
DE
KS
OK
WI NY KY
TN MO
GA ID
CO IL
VA NMUT NV MD MI
NH
NC OH
OR
NE HI
SC
FL
IN
SD
CA
WA
TX
WY LA
WV
8 9 10 11 12
ln(seizable assets)
(a) Insurance Coverage vs. Seizable
Assets
60 70 80 90 100
Insurance coverage (percentage points)
TX
MT
VT
RI
CT
ND MA
ME
MN
NJ
IA
DC
NH
NV
KS
NY TN
WI WA ID
KY
AK
GA IL
NM
CO MD
MO
VA OR UT
OH
NC NE
FL
HI
IN
SC
AL
AZ
AR
PA
MI
SD
MS OK
WV
8.5 9 9.5 10 10.5
ln(seizable assets) for constant sample
(c) Insurance Coverage vs. Simulated
Instrument
60 70 80 90 100
Insurance coverage (percentage points)
ID
NV
GAMI
MO
CO VAUT
NE
NY TN
KY WA IL
NM
OH
NC
SC
PA
NH
IN
MD
AZ
AR
AL
CA
TX
MS
WY
8.6 8.8 9 9.2 9.4 9.6
ln(seizable home equity) for constant sample
(e) Insurance Coverage vs. Historical
Instrument
LA
CA
RI
CT
NJ
WV
MA
WY
NJ
ME
LA
VT
DE
DE
RI
50 60 70 80 90 100
Insurance coverage (percentage points)
8 9 10 11 12
ln(seizable assets)
(b) Insurance Coverage vs. Seizable
Assets
60 70 80 90 100
Insurance coverage (percentage points)
8.5 9 9.5 10 10.5
ln(seizable assets) for constant saple
(d) Insurance Coverage vs. Simulated
Instrument
60 70 80 90 100
Insurance coverage (percentage points)
8.6 8.8 9 9.2 9.4 9.6
ln(seizable home equity) for constant sample
(f) Insurance Coverage vs. Historical
Instrument
Notes: Figures on the same row show the exact same data. In the left column, data points
are indicated with state abbreviations. In the right column, states are indicated with
circles proportional to the number of observations from that state. Panels A and B plot
the raw data: insurance coverage against log seizable assets averaged by state. Panels C
and D plots the reduced form: insurance coverage against the instrument by state. The
simulated instrument is mean log seizable assets for a constant, nationally representative
sample of households as though they lived in that state. Panels E and F plots the reduced
form with the historical instrument. The historical instrument is constructed similarly to
the simulated instrument using seizable home equity under inflation-adjusted 1920
homestead exemption laws. The sample is made up of uninsured and privately insured
households, excluding households with a head age 65 or older. Pooled data from the 1999
to 2005 PSID, inflation-adjusted to 2005 using the CPI-U.

CHAPTER 1. BANKRUPTCY 49
Coefficient on ln(seizable assets)
9
8
7
6
5
4
3
2
1
0
Figure 1.9: Effect on Coverage on Samples Excluding Each State
AK ALARAZCACOCTDCDEFLGAHIIAIDILINKSKYLAMAMDMEMIMNMOMSMTNCNDNENHNJNMNVNYOHOKORPARISCSDTNTXUTVAVTWAWIWVWY
Excluded state
Notes: Figure shows the coefficients and 95 percent confidence intervals from the baseline
2SLS specification excluding the indicated state. The specification is the same as column
6 of Table 1.3. The 95 percent confidence intervals are constructed using robust standard
errors clustered at the state level. See the Table 1.3 note for more details.

Pr(WTP > premium)
CHAPTER 1. BANKRUPTCY 50
Figure 1.10: Micro-Simulation Estimates of Percent Covered Without and With
Bankruptcy Insurance by Deductible
70%
60%
50%
40%
30%
20%
10%
0%
Pr(WTP without > premium)
Pr(WTP with > premium)
$- $1,000 $2,000 $3,000 $4,000 $5,000 $6,000 $7,000 $8,000
Deductible
Notes: Micro-simulation estimates of the probability the willingness to pay (WTP)
exceeds the premium without and with bankruptcy by deductible level. Willingness to pay
is calculated using CARA utility with parameter of 2.5 × 10 −5 . Premiums calculated as
the expected value of medical costs above the deductible scaled up to account for moral
hazard (elasticity of -0.22) and administrative loading (50 percent). Household-level
estimates weighted by number of individuals per household for interpretation at the
individual level.

CHAPTER 1. BANKRUPTCY 51
Insurance generosity
Generosity gap
Figure 1.11: The Insurance Generosity Gap
Conventional insurance
Other goods
Implicit insurance

CHAPTER 1. BANKRUPTCY 52
Table 1.1: Asset Exemption Laws by State
Contemporaneous exemptions Homestead
Other
exemptions
financial
Wildcard no Federal for town lots
State Homestead Vehicle Retirement assets Wildcard homestead available in 1920
Alabama 10,000 0 Unlimited 0 6,000 6,000 No 2,000
Alaska 67,500 7,500 Unlimited 3,500 0 0 No n/a
Arizona 150,000 10,000 Unlimited 300 0 0 No 4,000
Arkansas Unlimited 2,400 40,000 0 500 500 Yes 2,500
California--system 1 75,000 4,600 Unlimited 1,825 0 0 No 5,000
California--system 2 0 2,975 Unlimited 0 19,675 19,675 No n/a
Colorado 90,000 6,000 Unlimited 0 0 0 No 2,000
Connecticut 150,000 3,000 Unlimited 0 2,000 2,000 Yes 1,000
Delaware 0 0 Unlimited 0 500 500 No 0
District of Columbia Unlimited 5,150 Unlimited 0 17,850 17,850 Yes n/a
Florida Unlimited 2,000 Unlimited 0 2,000 2,000 No n/a
Georgia 10,000 7,000 Unlimited 0 11,200 11,200 No 1,600
Hawaii 40,000 5,150 Unlimited 0 0 0 Yes n/a
Idaho 50,000 6,000 Unlimited 0 1,600 1,600 No 5,000
Illinois 15,000 2,400 Unlimited 0 4,000 4,000 No 1,000
Indiana 0 0 Unlimited 0 20,000 20,000 No 600
Iowa Unlimited 1,000 Unlimited 0 200 200 No n/a
Kansas Unlimited 40,000 Unlimited 0 0 0 No n/a
Kentucky 10,000 5,000 Unlimited 0 2,000 2,000 No 1,000
Louisiana 25,000 0 Unlimited 0 0 0 No 2,000
Maine 70,000 10,000 Unlimited 0 12,800 12,800 No 500
Maryland 0 0 Unlimited 0 22,000 22,000 No 0
Massachusetts 1,000,000 1,400 Unlimited 1,250 0 0 Yes 800
Michigan 7,000 0 Unlimited 0 0 0 No 1,500
Minnesota 200,000 7,600 Unlimited 0 0 0 Yes n/a
Mississippi 150,000 0 Unlimited 0 10,000 10,000 No 3,000
Missouri 15,000 6,000 Unlimited 0 1,250 1,250 No 1,500
Montana 200,000 5,000 Unlimited 0 0 0 No n/a
Nebraska 12,500 0 Unlimited 0 0 5,000 No 2,000
Nevada 400,000 30,000 1,000,000 0 0 0 No 5,000
New Hampshire 200,000 8,000 Unlimited 0 8,000 8,000 Yes 500
New Jersey 0 0 Unlimited 0 2,000 2,000 Yes 1,000
New Mexico 60,000 8,000 Unlimited 0 1,000 4,000 Yes 1,000
New York 20,000 0 Unlimited 0 10,000 10,000 No 1,000
North Carolina 13,000 3,000 Unlimited 0 8,000 8,000 No 1,000
North Dakota 80,000 2,400 200,000 0 0 15,000 No n/a
Ohio 10,000 2,000 Unlimited 800 800 800 No 1,000
Oklahoma Unlimited 6,000 Unlimited 0 0 0 No n/a
Oregon 33,000 3,400 15,000 15,000 800 800 No n/a
Pennsylvania 0 0 Unlimited 0 600 600 Yes 300
Rhode Island 200,000 20,000 Unlimited 0 0 0 Yes 0
South Carolina 10,000 2,400 Unlimited 0 0 2,000 No 1,000
South Dakota Unlimited 0 500,000 0 4,000 4,000 No n/a
Tennessee 7,500 0 Unlimited 0 8,000 8,000 No 1,000
Texas Unlimited 0 Unlimited 0 60,000 60,000 Yes 5,000
Utah 40,000 5,000 Unlimited 0 0 0 No 2,000
Vermont 150,000 5,000 Unlimited 1,400 8,400 8,400 Yes 2,000
Virginia 0 4,000 35,000 0 32,000 32,000 No 500
Washington 40,000 5,000 Unlimited 0 4,000 4,000 Yes 1,000
West Virginia 0 4,800 Unlimited 0 51,600 51,600 No 1,000
Wisconsin 40,000 0 Unlimited 2,000 10,000 10,000 Yes n/a
Wyoming 20,000 4,800 Unlimited 0 0 0 No 2,500
Federal 18,500 5,900 Unlimited 0 20,450 20,450 n/a n/a
Averages* 58,821 4,884 298,333 501 6,592 7,073 27% 1,679
Notes: Contemporaneous exemptions for couples filing jointly from Elias (2007) and historical exemptions for couples filing jointly
from Goodman (1993). Under contemporaneous law, California residents can choose between system 1 and 2 and residents can
choose federal exemptions in states where federal exemptions are available. Wildcard no homestead exemption is available to
households which do not take the homestead exemption. For the historical exemptions, states that did not exist and states that had
acre-based exemptions are denoted as n/a. States that did not have homestead exemptions are assigned a value of zero.
*Excludes states with unlimited or n/a exemptions.

CHAPTER 1. BANKRUPTCY 53
Table 1.2: Implied First Stage Estimates
Dependent variable: ln(w
(1) (2) (3) (4)
Instruments
Baseline instrument 1.00
(0.04)
Homestead instrument 0.38
(0.06)
Non-homestead instrument 1.10
(0.10)
Historical homestead instrument
Age group
1.41
(0.31)
35-44 -0.14 -0.14 -0.13 -0.12
(0.02) (0.02) (0.02) (0.02)
45-54 -0.20 -0.20 -0.20 -0.20
(0.03) (0.03) (0.03) (0.04)
55-64 -0.23 -0.23 -0.21 -0.22
Family structure
(0.04) (0.04) (0.04) (0.04)
Couple -0.21 -0.21 -0.20 -0.22
(0.03) (0.03) (0.03) (0.03)
Single parent -0.02 -0.02 -0.02 0.00
(0.03) (0.03) (0.03) (0.03)
Couple with children -0.27 -0.26 -0.26 -0.26
Race
(0.03) (0.03) (0.03) (0.04)
Non-white 0.00 -0.01 0.01 0.02
Education
(0.02) (0.03) (0.02) (0.02)
High school to some college 0.02 0.03 0.03 0.05
(0.03) (0.03) (0.03) (0.02)
College or greater -0.01 0.00 0.01 0.01
(0.03) (0.03) (0.03) (0.03)
S )
Income polynomial (4th order) Yes Yes Yes Yes
State controls Yes Yes Yes Yes
R-squared 0.89 0.88 0.88 0.89
N 20,265 20,265 20,265 17,418
F-statistic on excluded instrument 494.06 36.16 116.93 20.95
Prob > F 0.00 0.00 0.00 0.00
Notes: The baseline instrument is mean log seizable assets for a constant, nationally representative sample of households
as though they lived in that state. The homestead, non-homestead, and historical homestead instruments are constructed
similarly using seizable home equity, non-homestead seizable assets, and seizable home equity under inflation-adjusted
1920 homestead exemption laws. The excluded demographic groups are age 18-34, single, less than high school, and
white. The state controls are mean income, percent unemployed, and percent covered by Medicaid. Robust standard errors
clustered at the state level in parentheses. Pooled 1999 to 2005 PSID excluding households with public insurance or a head
age 65 or older. Inflation-adjusted to 2005 using the CPI-U.

CHAPTER 1. BANKRUPTCY 54
Table 1.3: Baseline Coverage Estimates
(1) (2) (3) (4) (5) (6) (7) (8)
ln(w S Dependent variable: Percent of household insured
OLS 2SLS
) 6.13 2.45 2.36 2.63 8.80 5.43 5.38 6.15
Age group
(0.37) (0.20) (0.53) (0.25) (2.27) (1.16) (0.95) (0.87)
35-44 4.48 3.98 4.48 3.49 4.28 3.32
(0.97) (0.93) (1.13) (1.10) (0.88) (1.60)
45-54 3.08 2.48 2.91 1.22 2.94 0.78
(1.04) (1.02) (1.34) (1.27) (0.98) (2.39)
55-64 8.03 7.44 8.48 4.53 7.79 4.42
Family structure
(1.08) (1.05) (1.44) (1.81) (1.01) (3.63)
Couple 5.38 4.76 6.55 5.05 5.29 5.89
(1.55) (1.43) (1.92) (1.48) (1.42) (1.95)
Single parent 3.04 3.00 4.72 3.29 3.13 5.01
(1.55) (1.51) (1.92) (1.52) (1.48) (2.00)
Couple with children 11.41 10.51 12.53 11.16 11.14 12.25
Race
(1.51) (1.38) (1.96) (1.41) (1.37) (1.83)
Non-white -6.34 -6.06 -5.85 -4.95 -5.86 -4.10
Education
(0.76) (0.77) (1.01) (0.88) (0.82) (1.26)
High school to some college 9.86 9.83 8.65 9.12 9.51 7.77
(1.88) (1.86) (2.37) (1.81) (1.83) (1.70)
College or greater 11.79 12.03 10.76 10.35 11.94 8.91
(1.77) (1.73) (2.10) (1.68) (1.70) (1.86)
Income polynomial (4th order) Yes Yes Yes Yes Yes Yes
Wealth polynomial (4th order) Yes Yes
State controls Yes Yes Yes Yes Yes Yes
State premium index (log points) -1.15 -1.47
(0.91) (0.93)
Instrument
Baseline instrument Yes Yes Yes Yes
R-squared 0.13 0.27 0.28 0.27 - - - -
N 20,265 20,265 20,265 14,510 20,265 20,265 20,265 14,510
Notes: The dependent variable is the percent of household member-months insured. In the 2SLS specifications, the
baseline instrument is mean log seizable assets for a constant, nationally representative sample of households as though
they lived in that state. The excluded demographic groups are age 18-34, single, less than high school, and white. The
state controls are mean income, percent unemployed, and percent covered by Medicaid. See text for a description of the
premium index. All specifications include an indicator for the bottom-coding of seizable assets. Robust standard errors
clustered at the state level in parentheses. Pooled 1999 to 2005 PSID excluding households with public insurance or a
head age 65 or older. Inflation-adjusted to 2005 using the CPI-U.

CHAPTER 1. BANKRUPTCY 55
Table 1.4: Sensitivity Analysis of the Effect on Coverage
Baseline IV by
Non-homestead
Publicly insured Ins mkt regs Census region Census division group Homestead IV IV Historical IV
(1) (2) (3) (4) (5) (6) (7) (8)
ln(w S Dependent variable: Percent of household insured
) 6.30 4.71 5.01 4.23 4.71 4.53 6.16 6.39
(1.11) (1.23) (1.83) (0.72) (1.17) (1.29) (0.67) (0.82)
Demographic controls Yes Yes Yes Yes Yes Yes Yes Yes
State controls Yes Yes Yes Yes Yes Yes Yes Yes
Additional state legislative factors Yes
Census region FE Yes
Census division FE Yes
Instrument
Baseline instrument Yes Yes Yes Yes
Baseline instrument by group Yes
Homestead instrument Yes
Non-homestead instrument Yes
Historical instrument Yes
N 25,450 20,265 20,265 20,265 20,265 20,265 20,265 17,418
Notes: The dependent variable is the percent of household member-months insured. The baseline instrument is mean log seizable assets for a constant, nationally
representative sample of households as though they lived in that state. Means are taken by age and education group for the baseline instrument by group variable. The
homestead, non-homestead and historical homestead instruments are constructed similarly to the baseline instrument using seizable home equity, non-homestead seizable
assets, and seizable home equity under inflation-adjusted 1920 homestead exemption laws. Demographic controls are dummies for age group, family structure, race, and
education. State controls are mean income, percent unemployed, and percent covered by Medicaid. Additional state legislative factors are insurance market regulations and
Medicaid Medically Needy program parameters. See text for details. All specifications include an indicator for the bottom-coding of seizable assets. Robust standard errors
clustered at the state level in parentheses. Pooled 1999 to 2005 PSID excluding households with public insurance or a head age 65 or older. Inflation-adjusted to 2005
using the CPI-U.

CHAPTER 1. BANKRUPTCY 56
Table 1.5: Heterogeneity in the Effect on Coverage
Younger household Older household
Lower income (< Higher income (≥ Lower seizable Higher seizable
head(

CHAPTER 1. BANKRUPTCY 57
Table 1.6: The Effect on Assets
Dependent variable:
1(homeowner) ln(home equity) 1(homeowner) ln(home equity) 1(vehicle owner) ln(vehicle equity) 1(wealth > 0) ln(wealth)
OLS 2SLS OLS
(1) (2) (3) (4) (5) (6) (7) (8)
Homestead instrument -0.009 0.000 0.066 0.000
(0.014) (0.000) (0.149) (0.000)
Vehicle instrument -0.088 -0.288
(0.064) (0.149)
Baseline instrument 0.008 0.000 -0.078 0.000 -0.025 -0.047 -0.011 -0.011
(0.031) (0.000) (0.178) (0.000) (0.014) (0.041) (0.007) (0.083)
Demographic controls Yes Yes Yes Yes Yes Yes Yes Yes
Income polynomial (4th order) Yes Yes Yes Yes Yes Yes Yes Yes
Wealth polynomial (4th order) Yes Yes Yes Yes Yes Yes
Instrument
Historical homestead Yes Yes
R-squared 0.47 1.00 - - 0.16 0.35 0.10 0.43
N 20,265 17,293 17,418 14,836 20,265 17,699 20,265 17,293
Notes: Regressions of assets (indicators for positive values and log values) on state-level measures of asset exemption law (the instruments). The homestead instrument is mean log seizable
home equity for a constant, nationally representative sample of households as though they lived in each state. The historical, vehicle, and baseline instruments are constructed similarly using
seizable home equity under inflation-adjusted 1920 homestead exemption laws, seizable vehicle equity, and seizable assets. Demographic controls are dummies for age group, family
structure, race, and education. Pooled 1999 to 2005 PSID excluding households with public insurance or a head age 65 or older. Inflation-adjusted to 2005 using the CPI-U. Homeownership
rate is 72.5 percent. Vehicle ownership rate is 91.0 percent.

CHAPTER 1. BANKRUPTCY 58
Table 1.7: The Effect on Premiums
Dependent variable: ln(premium)
OLS 2SLS
(1) (2) (3) (4) (5) (6)
Baseline instrument -0.098 -0.098 -0.074 -0.093 -0.111 -0.160
State controls
(0.059) (0.053) (0.038) (0.106) (0.107) (0.076)
Medicaid (share) -0.354 -0.289 -0.417 -0.617
(1.004) (0.794) (1.185) (0.890)
Income 0.000 0.000 0.000 0.000
(0.000) (0.000) (0.000) (0.000)
Unemployment (share) -0.065 -0.540 1.120 -0.333
Additional state legislative factors
(2.675) (1.850) (3.060) (1.767)
Coverage mandates 0.003 0.000
(0.004) (0.004)
Any willing phamacists 0.058 0.031
(0.051) (0.058)
Any willing provider -0.026 -0.050
(0.060) (0.067)
Community rating 0.532 -0.016
(0.089) (0.070)
Guarenteed issue -0.450 0.091
(0.133) (0.135)
Charity care pool -0.140 -0.207
(0.060) (0.064)
Plan FE Yes Yes Yes Yes Yes Yes
Instrument
Historical instrument Yes Yes Yes
R-squared 0.833 0.834 0.865 - - -
N 1,891 1,891 1,891 1,420 1,420 1,420
Notes: The dependent variable is the monthly premium for a 30-year-old non-smoking male
for plans offered by eHealthInsurance. The baseline instrument is mean log seizable assets
for a constant, nationally representative sample of households as though they lived in that
state. In the 2SLS specifications, mean seizable home equity under 1920 homestead
exemptions laws is used as an instrument. Robust standard errors clustered at the state
level in parentheses. Premiums inflation-adjusted to 2005 using the CPI-U.

CHAPTER 1. BANKRUPTCY 59
Table 1.8: Baseline Costs Estimates
(1) (2) (3) (4) (5) (6)
ln(w S Dependent variable: ln(out-of-pocket+1)
OLS 2SLS
) 0.561 0.219 0.271 0.572 0.368 0.350
Age group
(0.014) (0.030) (0.037) (0.014) (0.040) (0.038)
35-44 -0.118 -0.092 -0.237 -0.123
(0.119) (0.120) (0.128) (0.121)
45-54 -0.136 -0.075 -0.255 -0.088
(0.136) (0.133) (0.146) (0.134)
55-64 -0.210 -0.088 -0.368 -0.089
Family structure
(0.155) (0.158) (0.162) (0.158)
Couple 0.226 0.298 0.245 0.335
(0.103) (0.096) (0.107) (0.098)
Single parent 0.009 0.026 0.083 0.065
(0.121) (0.121) (0.113) (0.119)
Couple with children -0.217 -0.135 -0.138 -0.072
Race
(0.146) (0.141) (0.153) (0.139)
Non-white 0.011 -0.043 -0.049 -0.092
Education
(0.119) (0.122) (0.122) (0.120)
High school to some college 0.205 0.173 -0.001 0.095
(0.113) (0.116) (0.115) (0.115)
College or greater 0.219 0.254 -0.067 0.167
(0.175) (0.176) (0.179) (0.181)
Income polynomial (4th order) Yes Yes Yes Yes
Wealth polynomial (4th order) Yes Yes
Relative Risk Score polynomial (4th order) Yes Yes Yes Yes
Instrument
Baseline instrument Yes Yes Yes
R-squared 0.856 0.881 0.882 - - -
N 3,401 3,401 3,401 3,401 3,401 3,401
Notes: Regressions of out-of-pocket payments on seizable assets in the sample of uninsured households with positive
medical utilization. In the 2SLS specifications, the baseline instrument is mean log seizable assets for a constant,
nationally representative sample of households as though they lived in state state. The excluded demographic groups are
age 18-34, single, less than high school, and white. The Relative Risk Score is a measure of utilization based on medical
diagnoses. All specifications include an indicator for the bottom-coding of seizable assets. Robust standard errors
clustered at the state level in parentheses. Pooled 2000 to 2005 MEPS inflation-adjusted to 2005 using the CPI-U.

CHAPTER 1. BANKRUPTCY 60
Table 1.9: Sensitivity Analysis of the Effect on Costs
Nonhomestead
IV Historical IV
High/low
utilization,
baseline IV
High/low
utilization,
historical IV
Hospital Homestead
controls IV
(1) (2) (3) (4) (5) (5)
ln(w S ) 0.380 0.377 0.382 0.375
(0.066) (0.038) (0.039) (0.039)
ln(w S ) X 1(High utilization) 0.512 0.506
(0.044) (0.046)
ln(w S ) X 1(Low utilization) 0.360 0.372
(0.038) (0.037)
Demographic controls Yes Yes Yes Yes Yes Yes
State hospital factors
Nonprofit share -0.123
(0.439)
Forprofit share 0.913
(0.694)
DSH payments per 1,000 0.000
(0.000)
Charity care pool indicator 0.009
(0.214)
FQHC per 100,000 0.043
(0.043)
Dependent variable: ln(out-of-pocket+1)
Relative Risk Score polynomial (4th order) Yes Yes Yes Yes Yes Yes
Instrument
Baseline instrument Yes Yes
Homestead instrument Yes
Non-homestead instrument Yes
Historical instrument Yes Yes
N 3,401 3,401 3,401 2,839 3,401 2,839
Notes: Regressions of out-of-pocket payments on seizable assets in the sample of uninsured households with positive medical
utilization. The baseline instrument is mean log seizable assets by state for a constant, nationally representative sample of households
as though they lived in each state. The homestead, non-homestead, and historical homestead instruments are constructed similarly
using variation in seizable home equity, non-homestead seizable assets, and seizable home equity under inflation-adjusted 1920
homestead exemption laws. Demographic controls are dummies for age group, family structure, race, and education. The Relative Risk
Score is a measure of utilization based on medical diagnoses. All specifications include an indicator for the bottom-coding of seizable
assets. Robust standard errors clustered at the state level in parentheses. Pooled 2000 to 2005 MEPS inflation-adjusted to 2005 using
the CPI-U.

CHAPTER 1. BANKRUPTCY 61
Table 1.10: Micro-Simulation Estimates of Percent Covered Without and With
Bankruptcy Insurance by Insurance Status
Without bankruptcy
Pr(WTP > premium)
With bankruptcy Difference
Low risk aversion
Uninsured 34% 5% 28%
Insured 62% 48% 14%
Difference
Moderate risk aversion
-28% -43% 14%*
Uninsured 71% 11% 60%
Insured 85% 62% 23%
Difference
High risk aversion
-14% -51% 37%*
Uninsured 91% 21% 70%
Insured 99% 75% 24%
Difference -8% -54% 46%*
Notes: Micro-simulation estimates of the probability that the willingness to pay exceeds the
premium by insurance status for a $2,000 deductible plan. Willingness to pay (WTP) calculated
using CARA utility with parameters of 2.5 x 10^-5 (low risk aversion), 5.0 x 10^-5 (moderate
risk aversion), and 7.5 X 10^-5 (high risk aversion). Premiums calculated as the expected
value of medical costs above the deductible scaled up to account for moral hazard (elasticity of
-0.22), administrative loading (10 and 50 percent), and the cross subsidization of unpaid care
(endogenously determined). Household-level estimates weighted by the number of individuals
per household for interpretation at the individual level.
*Difference-in-differences.

CHAPTER 1. BANKRUPTCY 62
Table 1.11: Policy Counterfactuals
Penalty Take-up ! WTP ! Cost ! Surplus
Pigovian penalty
Low risk aversion $218.21 7.5% $10.02 $6.06 $3.95
Moderate risk aversion $218.21 7.3% $11.02 $6.83 $4.19
High risk aversion
PPACA penalty
$218.21 7.9% $12.30 $7.72 $4.58
Low risk aversion $481.43 63.6% -$7.38 $5.34 -$12.72
Moderate risk aversion $481.43 56.4% $3.02 $14.40 -$11.38
High risk aversion
Medical debt non-dischargeable
$481.43 49.7% $11.09 $20.24 -$9.15
Low risk aversion n/a 100.0% -$48.26 -$12.12 -$36.14
Moderate risk aversion n/a 100.0% -$28.18 $12.75 -$40.93
High risk aversion n/a 100.0% -$6.43 $36.74 -$43.17
Notes: Micro-simulation estimates of insurance take-up, willingness to pay, costs, and social surplus
from different penalty systems relative to a baseline in which households can choose bankruptcy at no
cost. The Pigovian penalty is the household-specific social cost of the implicit insurance from bankruptcy.
PPACA is the inflation-adjusted, fully phased-in penalty under this legislation, defined as the greater of
$625 or 2.5 percent of income, up to a maximum of $2,085 per household. Medical debt nondischargeable
exposes households to the full financial risk when uninsured. Take-up is the percent of
uninsured individuals that take up coverage. Willingness to pay (WTP) is calculated using CARA utility
with parameters of 2.5 x 10^-5 (low risk aversion), 5.0 x 10^-5 (moderate risk aversion), and 7.5 x
10^-5 (high risk aversion). Household-level estimates weighted by number of individuals per household
for interpretation at the individual level.

CHAPTER 1. BANKRUPTCY 63
Table 1.12: Summary Statistics: Seizable Assets by Insurance Status
Mean Std. Dev. 5th 25th
Percentile
50th 75th 95th
Panel A: All (n = 22,844)
Seizable assets $216,943 $1,105,939 -$10,148 $2,344 $34,328 $155,219 $796,366
Gross seizable assets $221,434 $1,106,139 $0 $3,500 $38,400 $158,699 $797,593
Seizable home equity (70.4% homeownership) $52,487 $134,239 $0 $0 $0 $53,760 $249,205
Other seizable assets $168,948 $1,054,503 $0 $1,010 $18,000 $81,180 $616,680
Dischargeable debt $6,659 $15,675 $0 $0 $1,024 $7,305 $28,678
Filing costs
Panel B: Privately Insured (n = 20,197)
$2,000 $0 $2,000 $2,000 $2,000 $2,000 $2,000
Seizable assets $233,241 $1,153,125 -$9,880 $3,423 $42,809 $168,945 $849,475
Gross seizable assets $237,970 $1,153,292 $0 $5,860 $46,294 $172,870 $852,001
Seizable home equity (73.3% homeownership) $56,416 $139,145 $0 $0 $0 $60,320 $260,000
Other seizable assets $181,554 $1,099,952 $0 $2,308 $21,500 $90,649 $654,239
Dischargeable debt $6,898 $15,979 $0 $0 $1,297 $7,684 $29,511
Filing costs
Panel C: Uninsured (n = 2,647)
$2,000 $0 $2,000 $2,000 $2,000 $2,000 $2,000
Seizable assets $41,658 $207,719 -$13,305 $2,000 $2,344 $14,937 $212,000
Gross seizable assets $43,586 $207,884 $0 $0 $637 $16,000 $218,871
Seizable home equity (38.5% homeownership) $10,222 $40,560 $0 $0 $0 $0 $75,000
Other seizable assets $33,364 $197,079 $0 $0 $0 $9,348 $158,089
Dischargeable debt $4,095 $11,615 $0 $0 $0 $3,264 $20,304
Filing costs $2,000 $0 $2,000 $2,000 $2,000 $2,000 $2,000
Notes: Household-level statistics from the pooled 1999 to 2005 PSID inflation-adjusted to 2005 using the CPI-U. Excludes households with a head age 65 or
older and those with public insurance. See text seizable assets calculation for details.
Table 1.13: Summary Statistics: Medical Costs by Insurance Status
Mean Std. Dev. 5th 25th
Percentile
50th 75th 95th
Panel A: All (n = 34,841)
Charges $6,647 $17,781 $0 $420 $1,836 $6,099 $27,029
Total payments $4,085 $9,704 $0 $309 $1,388 $4,275 $15,859
Private payments $2,974 $8,648 $0 $37 $661 $2,701 $12,448
Public payments $85 $1,094 $0 $0 $0 $0 $78
Misc payments $243 $2,102 $0 $0 $0 $0 $715
Out-of-pocket payments
Panel B: Insured (n = 31,753)
$783 $1,565 $0 $76 $340 $907 $2,934
Charges $7,201 $18,437 $0 $590 $2,201 $6,860 $28,617
Total payments $4,480 $10,131 $0 $455 $1,678 $4,786 $16,767
Private payments $3,391 $9,157 $0 $190 $954 $3,209 $13,528
Public payments $83 $1,078 $0 $0 $0 $0 $89
Misc payments $204 $1,673 $0 $0 $0 $0 $567
Out-of-pocket payments
Panel C: Uninsured (n = 3,088)
$801 $1,474 $0 $103 $375 $943 $2,939
Charges $2,691 $11,353 $0 $0 $232 $1,429 $11,022
Total payments $1,267 $4,966 $0 $0 $136 $858 $5,231
Private payments $0 $0 $0 $0 $0 $0 $0
Public payments $96 $1,200 $0 $0 $0 $0 $0
Misc payments $515 $3,989 $0 $0 $0 $40 $1,678
Out-of-pocket payments $657 $2,097 $0 $0 $80 $543 $2,906
Notes: Household-level statistics from the pooled 2000 to 2005 MEPS inflation-adjusted to 2005 using the CPI-U. Charges
are the list price of medical care received. Private payments are from private insurance providers (e.g., employer sponsored
insurance, individual market insurance), public payments are from public insurance (e.g, Medicaid), miscellaneous
payments are from other sources (e.g. Workman's Compensation), and out-of-pocket payments are from households. Total
payments are the sum of the payment values.

CHAPTER 1. BANKRUPTCY 64
Table 1.14: Summary Statistics: Premiums
N Mean Std. Dev. Min. Max.
Premium (per month) 1,891 103 46 30 503
Deductible 1,891 3,351 2,229 0 8,963
Coinsurance (%) 1,891 15 12 0 50
Notes: Plan characteristics for all plans offered to a 30-year-old non-smoking male in each state
on eHealthInsurance. Values inflation-adjusted to 2005 using the CPI-U.
Table 1.15: Costs and Premiums by Deductible Level
Deductible Pr(Costs > deductible) Calibrated
Premium
Individual Market
$0 46.0% $2,126 $2,140
$1,000 10.4% $1,735 $2,061
$2,000 5.9% $1,546 N/A
$5,000 2.7% $1,259 $874
$10,000 1.5% $1,009 N/A
Notes: Costs are medical costs when uninsured. Calibrated premiums are calculated as the
expected value of medical costs above the deductible scaled up to account for moral hazard
(elasticity of -0.22) and administrative loading (50 percent). Individual market premiums are for
policies starting in May 2010 issued by Aetna. They are adjusted for inflation using the Medical
Care component of the CPI-U. See Appendix D for details.

Chapter 2
Pricing and Welfare in Health Plan
Choice
with M. Kate Bundorf and Jonathan Levin
2.1 Introduction
Whether competition in health insurance markets leads to efficient outcomes is a
central question for health policy. Markets are effective when prices direct consumers
and firms to behave efficiently. But in health insurance markets, prices often do
not reflect the different costs of coverage for different enrollees. This generates two
concerns. If insurers receive premiums that do not reflect enrollee risk, they have an
incentive to engage in risk selection through plan design (??). Similarly, if consumers
face prices that do not reflect cost differences across plans, they may select coverage
inefficiently (??). While it is widely recognized that these problems may impair the
efficiency of competitive health insurance markets, evidence on their quantitative
importance for social welfare is limited.
In the U.S. private market, employers generally contract with insurers to cre-
ate a menu of plans from which employees select coverage. The government or a
quasi-public organization plays a similar role in the U.S. Medicare program and the
national systems of Germany and the Netherlands. To address incentive problems in
65

CHAPTER 2. HEALTH PLAN CHOICE 66
plan design, these intermediaries have begun to “risk-adjust” payments to plans (?).
Consumer prices, however, are typically not adjusted for individual risk. Indeed, in
the U.S., regulations prohibit employers and public programs from charging enrollees
different amounts based on nearly all health-related factors. 1 Moreover, even within
institutional limitations, contributions set by employers or in regulated markets may
not be welfare-maximizing given the complexities of self-selection.
In this paper, we analyze the effect of plan pricing on allocative efficiency. We
begin by making a basic theoretical point regarding plan prices and efficient match-
ing. Existing work suggests that, while poorly chosen contribution policies may lead
to inefficient outcomes, the problem can be solved by choosing an optimal uniform
contribution (e.g., ????). These analyses, however, make strong assumptions regard-
ing the correlation between enrollee risk and preferences for coverage. We show that
if these assumptions are violated, no uniform contribution policy leads to efficient
consumer choices.
The main part of the paper examines the implications of this observation. We
develop a simple econometric model of health plan demand and costs, estimate the
model on a novel dataset of small employers, and then use the parameter estimates to
simulate the welfare implications of alternative pricing policies. Based on our simula-
tions, we estimate that, in this setting, observed pricing policies are less efficient than
what could be achieved with risk-rated plan contributions. The shortfall is between
$60 and $325 annually per enrollee, or 2-11% of coverage costs. Employers could
realize approximately 1/4 of this surplus by adjusting their non-risk-rated contribu-
tions, but capturing the remainder would require setting different prices for people in
the same firm. We also find that employees select plans based on private information
about their health status. A hypothetical social planner who incorporated this private
information into prices could increase welfare by an additional $35-$100 annually per
enrollee. Despite these inefficiencies, the observed plan offerings generate substantial
benefits over any single-plan offering.
An understanding of the characteristics of the two insurers we study is important
for interpreting these results. One has a fairly broad network of providers and relies
1 Premiums vary by smoking status in some markets.

CHAPTER 2. HEALTH PLAN CHOICE 67
on patient cost sharing and primary care gatekeepers to control utilization. The other
has an integrated and closed delivery system and requires little patient cost sharing.
Estimates from the cost model suggest that these insurers have very different cost
structures. While costs are similar between the two for individuals of average health,
the integrated delivery system has higher costs for healthy enrollees and significantly
lower costs for less healthy enrollees.
On the demand side, we find that household preferences as well as health status
affect plan choice. However, in contrast to other studies, we find that no single plan
experiences systematic adverse selection. An interesting feature of our data is that
the plans are not obviously ordered by coverage level: consumers have to trade off
provider network restrictions against the level of cost sharing. This “horizontal” plan
differentiation differentiates our setting from that of many other studies in which
consumers choose how much coverage to purchase. We view this horizontal plan
differentiation as the likely explanation for the lack of systematic adverse selection.
This type of differentiation seems particularly salient given the recent evolution
of the health insurance market. In 1987, approximately three-quarters of people with
employer-sponsored health insurance had conventional coverage, under which plans
differed primarily in their cost sharing. By 2007, in contrast, the market was domi-
nated by managed care plans which use a mix of supply-side and demand-side utiliza-
tion management (?). This evolution suggests that classic insights based on purely
risk-based sorting may not adequately capture the dynamics of today’s market. 2
The inefficiencies we find are driven by two forces: heterogeneity in household
preferences and the significant cost advantage of the integrated system for individuals
in worse health. Our estimates suggest that although high risk households have some
preference for flexibility, a large fraction would choose the integrated delivery system
if they had to internalize the relevant cost differential between plans. Achieving this
with a uniform contribution policy, however, would require that all households face a
steep premium for the more flexible insurer. This in turn would create a welfare loss
for lower risk households who value flexibility. While the exact numbers we estimate
2 ? stress that a broad view of heterogeneity in preferences is important for understanding many
aspects of insurance markets.

CHAPTER 2. HEALTH PLAN CHOICE 68
are of course specific to our setting, the trade-off we identify may be relevant more
broadly. This analysis suggests that the integrated model of health care delivery faces
an important challenge in health insurance markets: current pricing policies make it
difficult to target households for whom the integrated model promises substantial cost
savings.
Our analysis ties in to past work on health plan choice and the efficiency of
health insurance markets. We draw on this work on health plan choice, which is
summarized by ? and ?, in modeling how employee demand varies with observed and
unobserved risk and preference characteristics. Our paper is more directly related to
recent work that uses econometric methods to quantify the efficiency implications of
adverse selection in health insurance markets (?????). Our paper points out that
uniform pricing, as is commonly observed, may lead to inefficiency when enrollees of
similar risk have different preferences for coverage. These other papers, in contrast,
analyze alternative institutional features of health insurance markets that contribute
to adverse selection. We relate both our empirical approach and our findings to these
papers in Section 2.5.4.
We emphasize that our analysis has some important limitations. First, it is based
on a particular, and only moderately-sized, sample of workers and firms. To address
this, we perform a variety of sensitivity analyses on our key parameter estimates,
which we discuss in the last section. Second, we take plan offerings as given. This
seems reasonable given that we are looking at small to medium size employers, but a
broader analysis of pricing ideally would incorporate plan design. Third, we do not
address issues of utilization behavior, or try to assess the relative social efficiency of
health care utilization under the different plans in our data. Finally, our analysis is
based on a static model, so we abstract from issues of dynamic insurance. We discuss
this issue in the conclusion.
2.2 Health Plan Pricing and Market Efficiency
We illustrate the relationship between pricing and market efficiency by adapting the
model of ?. In their model, consumers are distinguished by their forecastable health

CHAPTER 2. HEALTH PLAN CHOICE 69
risk, denoted θ. Each consumer chooses between a high-cost plan A and a low cost
plan B. We can think of the plans, for now, as vertically differentiated. The plans’
expected costs of covering a type-θ consumer are cA(θ) and cB(θ). Let ∆c (θ) =
cA(θ) − cB(θ) denote the cost differential. We assume ∆c is strictly positive and
increasing in θ.
Let vA(θ) and vB(θ) denote a type θ’s expected (dollar) value from being covered
by each of the plans. For the moment, the benefits of coverage are determined only
by forecastable health risk. We assume that contributions vary across plans but
not across consumers. A consumer who makes a contribution pj to enroll in plan
j ∈{A, B} gets a net benefit vj(θ) − pj. 3 Define ∆v(θ) =vA(θ) − vB(θ) to be the
additional amount a type-θ consumer would pay for the high-cost plan.
The efficient assignment places a type-θ consumer in plan A if and only if
∆v(θ) − ∆c(θ) ≥ 0.
At the same time, a type-θ consumer will select plan A if and only if
∆v(θ) − ∆p ≥ 0,
where ∆p = pA − pB is the incremental contribution for plan A.
Are there prices that lead to an efficient outcome? Assume that ∆v(θ) is increasing
in θ, which seems appropriate if plan A simply offers more generous coverage or easier
access to care. Then for any incremental contribution ∆p, atype-θ consumer will
choose A if and only if θ ≥ θ(∆p), where θ (∆p) is a threshold that can be varied
arbitrarily with ∆p. 4 Therefore it is possible to achieve efficient sorting if and only if
the efficient assignment also involves a threshold rule, i.e. if ∆v(θ)−∆c(θ) is negative
up to some θ ∗ and positive above it. Intuitively, the requirement for efficiency is that
willingness to pay increases more quickly with risk than the cost differential between
3 Here we make the simplifying assumption, which we maintain in our econometric model, that
plan preferences are additively separable in the plan premium. See ? for an extensive discussion of
this assumption.
4 An empirical prediction of this model is that plan A will experience unfavorable selection, and
its risk composition will be worse the larger is ∆p.

CHAPTER 2. HEALTH PLAN CHOICE 70
plans.
Existing analyses assume that the surplus function has the requisite single crossing
property (e.g., ????). In this case, shown in Figure 1:(a), efficiency can be achieved
by setting ∆p =∆c(θ ∗ ). The problem emphasized in the literature is that purchasers
may not choose the correct premium differential. If ∆p is too high, plan A attracts
only very high risks, and if prices are based on past outcomes, one can even end
up with an adverse selection “death spiral” (?). Alternatively, if ∆p is too low, too
many people select plan A, including some for whom the benefits are less than the
incremental social costs.
This familiar analysis can be questioned on several levels. First, even if we continue
to assume that consumers differ only in their health status and plans differ mainly in
the amount of coverage they offer, it could be socially efficient for high risks to be in a
cost-conscious plan. 5 For instance, the cost savings from a plan that actively manages
utilization might more than compensate these consumers for the loss of flexibility. In
this case, shown in Figure 1(b), uniform pricing is inherently inefficient because there
is no way to induce only the relatively high risks to self-select into plan B when
everyone faces the same prices.
Perhaps the more obvious issue from an empirical perspective is that health plans
often differ well beyond offering “more” or “less” coverage, and consumers differ on
dimensions other than health risk. It is increasingly common for firms to offer em-
ployees an HMO option that places greater restriction on provider choice, and a PPO
option that allows broader access to provider, with greater cost sharing. Consumers
in relatively poor health may value flexibility but be wary of increased cost-sharing.
As a result, heterogeneity in tastes or income may be at least as important as health
status in driving choice.
To capture this, think of plan A as a PPO and plan B as an HMO, and suppose
that consumers vary in both forecastable risk and taste. Specifically, let ε denote a
5 Arguably the benefits of delivering care efficiently may be largest for the chronically ill. The
most detailed analysis of differences in utilization between traditional Medicare coverage and Medicare
managed care plans found that the reductions in utilization generated by managed care plans
were concentrated among high risk beneficiaries and that these reductions in utilization were not
associated with differences in short term health outcomes (?).

CHAPTER 2. HEALTH PLAN CHOICE 72
risk/preference space.
Each of these issues is inherently quantitative in nature. We use the econometric
model developed in the next section to identify the relevant cost and demand param-
eters and then evaluate empirically how various pricing arrangements affect social
welfare.
2.3 Data and Environment
2.3.1 Institutional Setting
Our analysis is based on data from a private firm that helps small and mid-sized
employers manage health benefits. This firm, who we refer to as the intermediary,
obtains agreements from insurers to offer plans to small employers, signs up employ-
ers, and administers their health benefit. We examine data from 11 employers who
purchased coverage from the intermediary in a single metropolitan area in the western
United States during 2004 and 2005.
In this market, the intermediary works with two insurers. One insurer contracts
non-exclusively with a relatively broad set of providers. It offers an HMO plan (net-
work HMO) that requires enrollees to choose a primary care physician and obtain a
referral for specialist visits, and does not cover care from out-of-network providers.
It also offers a PPO plan (network PPO) that does not require referrals and covers
providers outside the plan’s network at an increased cost-share. 7 The second insurer
has an integrated and closed delivery system that facilitates supply side utilization
management. It offers a standard HMO (integrated HMO) and a point-of-service op-
tion (integrated POS) that allows enrollees to seek care outside the integrated system
at a higher cost.
The employers that hire the intermediary begin by choosing which plans to offer
their employees. Employers may customize the basic plans to a limited degree by
7 This insurer also offers a point-of-service (POS) plan that is the HMO with the option to go outof-network
at higher cost. Unfortunately we are not able to distinguish between network POS and
HMO enrollees. As a result, we simplify our analysis by dropping the three employer-years where
the network POS was offered. Our results are not sensitive to alternative approaches to handling
this issue.

CHAPTER 2. HEALTH PLAN CHOICE 73
varying characteristics such as the deductible and the level of coinsurance, but most
dimensions are fixed. Employers typically offer four coverage tiers: employee only,
employee plus spouse, employee plus children, and employee plus family. 8 The level
of cost sharing varies across coverage tiers. The employers do not offer any health
insurance plans beyond those offered by the intermediary.
The insurers then provide bids for each of the selected plans, relying on informa-
tion from the intermediary. In an employer’s first year with the intermediary, this
information is just the distribution of employees by age and sex. In subsequent years,
the insurers receive additional information on the health status of the workers, in the
form of a risk score described below. The intermediary instructs the insurers to bid
as if they were covering all workers within the firm. While the insurers provide bids
for each tier, the bids for tiers other than employee-only are simply scaled from the
employee-only bids by a constant that is very similar across employers and plans.
After the bids are received, the employer sets the employee contribution for each
plan and coverage tier. The employees then make their choices, and the plans are
required to accept all employees who choose to enroll. The last step is a series of
payments. For each employee that enrolls in a plan, the employer pays the inter-
mediary the insurer’s bid. The intermediary passes these payments to the insurers,
implementing transfers between insurers if there is variation in the health status of
employees and dependents enrolled in the different plans.
The intermediary uses a standard methodology for measuring enrollee health sta-
tus, the RxGroup model developed by DxCG, Inc. The model produces risk scores
based on a person’s age, sex, and health status, where health status is inferred from
prior use of prescription drugs, reported by the insurers. 910 In order to persuade the
8Two firms define coverage tiers based on employee only, employee plus one dependent, and
employee plus two or more dependents.
9In our analysis, we use the term “risk score” to refer to the DxCG prediction of an individual’s
health expenditures relative to the mean of the much larger base sample on which DxCG calibrates
their model. We note that our use of the term risk refers only to the level and not to the variance
of the expected expenditure, although we might naturally expect a relationship between the two.
10DxCG uses an internally-developed algorithm to infer the presence and severity of chronic conditions
from prescription drug use. The health expenditure model is estimated on a very large sample
(1,000,000+) of people under 65 with private health insurance. Using the estimated model, the
software predicts covered health expenditures for a given individual. A score of 1 corresponds to a

CHAPTER 2. HEALTH PLAN CHOICE 74
insurers to participate, the intermediary had to convince them they would be ade-
quately compensated for any unfavorable selection they might experience. Because
the risk score is crucial in this regard, the intermediary worked with the insurers to
ensure that the risk scores reflected the underlying health status of enrollees rather
than the methods used by the plans to manage utilization. For instance, one concern
was that the integrated insurer might substitute low-priced drugs more aggressively,
leading the algorithm to under-estimate the severity of chronic illness for its enrollees.
Discussions between the intermediary and the plans along these lines led to a num-
ber of minor adjustments in the risk scoring algorithm. Our discussions with the
participants suggest that both the intermediary and the insurers had strong incen-
tives to ensure that health status was measured accurately, and that it is reasonable
to assume that the risk score accurately captures employee health status and is not
contaminated by differences in the plans.
In addition to prescription drug utilization, each insurer also provides the inter-
mediary with the realized costs for each employer group. The network insurer reports
average claims per member per month for enrollees covered by either of the insurer’s
products. The integrated insurer reports similar information developed from an in-
ternal cost accounting system. Neither insurer distinguishes between its plans when
reporting this information.
2.3.2 Data and Descriptive Statistics
Our data includes all of the information discussed above: the plan offerings and
contribution policies of each employer, the risk scores and plan choices of employees
and their dependents, and the bids and reported costs of each insurer. A primary
strength of the data is that it includes both demand-side information on employees
and their choice behavior and supply-side information on insurer costs and bids in a
setting with two very different types of insurers. In addition, many of the employers
we observe offer nearly identical plans but have different risk profiles and contribution
policies which provides useful variation to identify demand and costs.
mean prediction from the original estimation sample. See ? for more detail.

CHAPTER 2. HEALTH PLAN CHOICE 75
Another useful feature of the data is that we observe each employer during their
first year of participation in the program. Insurers have little information on firm
characteristics beyond that provided by the intermediary during the first year, allow-
ing us to observe how plans bid when they have similar information on the likely risk of
a group. 11 On the demand side, a large literature documents that health plan choices
are highly persistent (e.g., ?), so observing choice behavior in the first year likely
provides a good indication of steady-state demand and allows us to observe the plan
characteristics and prices at the time of initial choice. The data’s main limitations are
the fairly small number of observations and restricted set of employee characteristics
relative to, say, the HR records of a large employer, and also the aggregated reporting
of realized costs.
The 11 firms have 2,044 covered employees and 4,652 enrollees (employees and
their dependents). We observe five of the employers for two years, creating a total of
3,683 employee-years and 6,603 enrollee-years. Table 2.1 provides summary statistics
on the covered employees, the enrollees, and the firms. Sixty-two percent of employees
are female; the average age is just over forty. Fifty-eight percent of enrollees are female
and enrollees are younger on average than employees, driven primarily by covered
children. Twenty-eight percent of employees enroll in a plan that covers their spouse
and 27 percent enroll in a plan that covers at least one child.
Table 2.1 also presents risk scores at the employee, enrollee, and employer lev-
els. A score of one represents an average individual in a nationally representative
sample, and a score of two indicates that an individual’s expected health costs are
twice the average. The average risk scores of employees and enrollees are 1.25 and
1.01, respectively. The difference reflects the lower expected expenditures for covered
children. Average risk ranges widely across employers, from 0.63 to 1.91. One reason
for the degree of variation is the small number of enrollees at some of the firms in our
data. This variation plays a key role in our analysis. We use information on insurer
bids and realized costs to estimate models of the relationship between costs and risk.
11 In a few cases, an employer had a prior contract with one of the insurers. We have examined
whether incorporating this into our employee demand model affects our estimates and found it did
not. One concern is that this situation could result in asymmetric information between the plans in
the bidding, but we think this is unlikely to be an important problem.

CHAPTER 2. HEALTH PLAN CHOICE 76
Because insurers report both bids and costs at the employer level, variation across
employers in average risk is necessary to identify these relationships.
Table 2.2 provides information on the plans offered by the employers in our sam-
ple. Most employers offer all four plans, and all offer both HMOs and at least one
other plan. On average, the integrated HMO is the least expensive plan and has
the lowest enrollee contribution. This plan features high rates of coinsurance, a low
deductible, and a low out-of-pocket maximum. The network PPO is on average the
most expensive plan and has the highest employee contribution. It features lower
coinsurance rates, higher deductibles and higher maximum expenditures. Roughly
speaking, the other two plans fall between these extremes. While bids for each plan
vary substantially across tiers, reflecting differences in expected expenditures, the
bids for tiers other than employee only are simply scaled by a factor that is very
similar across both plans and employers. Employee contributions also vary across
tiers, with employees typically facing a greater fraction of the plan bid for dependent
coverage. Variation in these contributions is important for the identification of our
demand model. We discuss contributions in detail in the identification section.
We summarize enrollment patterns in Table 2.3. The integrated HMO attracts
by far the most enrollees with a 59% market share among employees and 60% market
share among enrollees. We also find little evidence of extensive risk selection across
the plans. The integrated HMO attracts a slightly younger population, and women,
particularly women employees, disproportionately choose the network and integrated
HMOs. But the differences across the plans in both average age and average risk
score are small. This lack of sorting is not driven by heterogeneity across firms in the
choice sets. If we condition on employers that offer both the PPO and the integrated
HMO, for example, the average enrollee risk is 1.04 in both plans.

CHAPTER 2. HEALTH PLAN CHOICE 77
2.4 Econometric Model
2.4.1 Consumer Preferences, Plan Costs and Market Behav-
ior
In this section, we develop an econometric model that allows us to jointly estimate
consumer preferences and health plan costs. It should be noted that by costs we mean
overall costs to the insurer for a given enrollee in a given plan. Although we discuss
factors that may explain the variation in costs below, overall cost is sufficient for
welfare analysis and it is not necessary to decompose whether these cost differences
arise from, for example, moral hazard or physician reimbursement rates or some other
factor (c.f., ?).
In contrast to the simple theoretical model discussed above, the econometric model
allows for multiple plans, varying plan characteristics, and both observable and pri-
vately known dimensions of health risk and consumer tastes. Nevertheless, we aim
for the most parsimonious model that permits a credible assessment of market ef-
ficiency. In what follows, we describe the key components of the model: consumer
choice, health plan costs, health plan bidding, and employer contribution setting, and
identify the stochastic assumptions on the unobservables that permit estimation.
Consumer Choice
We use a standard latent utility model to describe household choice behavior,
where a household’s (money-metric) utility from choosing a plan depends on a com-
bination of household and plan characteristics. Specifically, household h’s utility from
choosing plan j is:
uhj = φjαφ + xhαxj + ψ(rh + µh; αrj) − pj + σεεhj. (2.1)
In this representation, household utility depends on observable plan characteris-
tics φj, the monthly plan contribution pj, 12 observable household demographics xh,
12 We convert employee contributions, which are made with pre-tax dollars, to post-tax dollars by
adjusting them by the marginal tax rate (see Footnote 13 for discussion). For a given household

CHAPTER 2. HEALTH PLAN CHOICE 78
an idiosyncratic preference εhj, and household health risk. Our measure of household
health risk is aggregated from the individual level. For each individual i, we decom-
pose health risk into the observable risk score ri and additional privately known health
factors µi. The µis capture information about health status that may affect choice
behavior, but is not subject to risk adjustment. Equivalently, we can interpret µi as
measurement error in the risk score. We assume that each µi is an i.i.d. draw from a
normal distribution with mean zero and variance σ 2 µ, and that the idiosyncratic tastes
εhj are i.i.d. type I extreme value random variables (i.e. logit errors).
We handle heterogeneity in household size and composition by assuming that,
apart from the treatment of health risk, each household behaves as if it had a rep-
resentative member with characteristics equal to the average of those of household
members. 13 We parameterize household risk using two variables: the average risk
of household members (i.e. the average of the ri + µi) and an indicator of whether
the household includes a high risk member. We define high risk as being above 2.25,
which corresponds to the 90% percentile of the observed risk score distribution. The
other household characteristics in the model are the averages of age and the male
indicator among covered household members as well as imputed household income. 14
In addition to the employee contribution, plan characteristics φj include a dummy
variable for plan (the network HMO and PPO and the integrated HMO and POS),
the relevant coinsurance rate and deductible for the given employee, and an indicator
of non-standard drug coverage. 15 To be consistent with our approach to household
h, letρh be the nominal contribution and τh the household’s marginal tax rate. The tax adjusted
contribution is ph =(1− τh)ρh.
13 We experimented with estimating different weights for household members, and also with restricting
the sample to individual enrollees. Neither has much effect on our results. The Appendix
includes individual enrollee estimates.
14 We impute taxable income for each household in our sample by estimating a model of household
income as a function of worker age, sex, family structure, firm size and industry using data from the
Current Population Survey for 2004 and 2005 on workers with employer-sponsored health insurance
in the corresponding state. We then use the model to impute household income for each employee in
our data incorporating random draws from the posterior distributions of the regression coefficients
and the standard deviation of the residuals. Based on these predictions, we use Taxsim to calculate
marginal tax rates based on federal, state, and FICA taxes making some assumptions on the correlation
of coverage tier with filing status and number of dependents. The average taxable family
income and marginal tax rate for workers in our sample are about $73,00 and 41%, respectively.
15 While the prescription drug coverage for each plan is complicated, comprised of both formulate

CHAPTER 2. HEALTH PLAN CHOICE 81
where rf denotes the average risk of employees in firm f, which the insurer forecasts
using the available demographic information, xf. 16
We model expected plan bids as a mark-up over expected cost. So plan j’s bid
for firm f is:
Bjf = δj · (aj + bj · (E[rf|xf] − 1)) + νjf, (2.7)
where νjf is an independent mean zero random variable. The new parameter intro-
duced in the bid model is the mark-up, δj. We constrain the mark-up to be constant
across the plans offered by a particular insurer. Although in theory an insurer could
vary the mark-up across its different plans, because the cost data are at the insurer-
firm level, we are unable to identify separately the mark-up and the fixed costs for
each plan offered by an insurer. Naturally we expect the mark-up parameters to be
larger than one.
Employer Contribution Setting
The last part of our model specifies how employers set required plan contributions.
We adopt a simple model in which employers pass on a fraction of their cost for the
lowest cost plan, and then a fraction of the incremental cost for higher cost plans. We
allow these fractions, denoted β and γ, to vary across firm-years and coverage tiers.
Let B lf denote the minimum bid received for coverage tier l in firm-year f, de-
note plan j’s bid for coverage tier l in firm-year f as Bjlf. We model the required
contribution as:
pjlf = βlf · B lf + γlf · (Bjlf − B lf)+ξjlf. (2.8)
This model describes employer behavior in our data remarkably well. The resid-
uals from the linear regression (2.8) have a standard deviation of 7.64, and the R-
squared is 0.99. As noted above, approximately half of the firms in our data choose
a ”proportional pass-through” strategy where β = γ. The others choose an ”incre-
mental pass-through” strategy in which β

CHAPTER 2. HEALTH PLAN CHOICE 82
2.4.2 Discussion of Model and Identification
The key quantities in our model are plan costs and plan demand as functions of fore-
castable risk, and the price elasticity of demand. The former determine the efficient
allocation of households to plans, while the latter determines how price changes affect
self-selection. We now discuss the variation in the data that identifies each of these
quantities in estimation.
Identifying plan costs is straightforward. The effect of forecastable risk on plan
costs is identified by variation across firms in the average risk scores of workers and
dependents, and how it affects insurer bids and realized costs. We identify the mark-
up parameters, δj, by the difference between the plan bids and reported costs. A
maintained assumption in estimating mark-ups is that insurers base their bids on
only the information about employees that is provided by the intermediary. We
discuss this assumption more below, but we believe it is reasonable given the small
size of the contracts and the fact that we consider only the first year of plan bids.
The effect of household risk on choice behavior (i.e. the coefficients αrj in the
demand equation) is identified by variation in observable risk across households. Our
model also allows private information about health status to affect choice. The key
parameter here is the variance of the private information, σ 2 µ, which is identified by
the correlation between consumers’ enrollment decisions and plans’ realized costs.
This identification is aided by cross-firm variation in contribution policies and demo-
graphics that, conditional on observable health risk, affect enrollment but not realized
costs. 17 The identification obtained from price variation is similar to the identifica-
tion in standard selection models, and relies on the exclusion restriction if a given
individual i is enrolled in a given plan j, his or her utilization does not depend on
the per-month premium (although of course it may depend on other elements of the
plan such as the copayment rate).
The most subtle identification issues arise in estimating the effect of plan con-
tributions on demand. Plan contributions are the result of plan bids and employer
pass-through decisions. Our model allows four sources of variation in contributions:
17 Our demand model also includes plan characteristics such as coinsurance and deductible. Their
coefficients are identified off cross-firm and cross-tier variation in the characteristics.

CHAPTER 2. HEALTH PLAN CHOICE 83
cross-firm variation in demographics (xf) that leads plans to submit different bids,
idiosyncratic variation in plan bids (νjf), cross-firm and cross-tier variation in em-
ployer pass-through rates (γjlf), and idiosyncratic variation in the plan contributions
(ξjlf). 18
Figure 2.4 demonstrates this variation by plotting the incremental contributions
against the incremental bids for each plan relative to the integrated HMO, which is
usually the plan requiring the lowest employee contribution. We plot contribution
rates for two tiers, employee only and employee plus spouse, to demonstrate how
contributions vary across tier. For the employee plus spouse data, we divide both
the contributions and the bids by two to obtain per-enrollee prices. Difference in
the bids for the integrated HMO and the network PPO ranges from $50 to $150
per month, with a large fraction due to cross-firm variation in demographic risk.
Combinations of incremental contributions and bids that lie along the 45 degree line
in Figure 2.3 represent employers who pass on the full marginal cost of higher plan
bids to employees. A subset of employers adopt this approach. Another subset of
employers fully subsidize the higher cost plans, setting incremental contributions of
zero. Between these two extremes are employers who partially subsidize higher cost
plans through contribution policies. In general, employers tend to pass on a greater
portion of incremental costs for plans with dependent coverage.
The availability of multiple sources of variation permits some flexibility in estimat-
ing price elasticities. Recall that accurate identification requires using price variation
that is not correlated with idiosyncratic household tastes εhj or privately known health
risk µh. Our baseline estimates use all four sources of variation. We also employ in-
strumental variables to isolate different sources of variation. The instruments are
predicted plan contributions based on alternative covariates. The bottom line from
these specifications is that our price elasticity estimates are quite robust to focusing
on different sources of variation in contributions. This robustness, despite our rela-
tively small sample, suggests that endogeneity may not be an important concern, at
least in this setting. Nevertheless, we now discuss the issues in detail.
18 We also introduce variation in employee contributions through the imputed marginal tax rates,
but we control for imputed income and relevant household demographics in the demand equation.

CHAPTER 2. HEALTH PLAN CHOICE 84
Perhaps the most obvious identification concern is that employers believe their
employees will prefer a particular plan and price accordingly. This could mean cater-
ing to employees with a low contribution, or setting a high contribution to pass on
costs. Either would generate a correlation between the idiosyncratic part of the con-
tribution ξjlf and household preferences εhj. To mitigate this concern, we instrument
for the actual plan contribution using the predicted value (ˆpjlf) from the contribution
model (2.8). We take this as our preferred specification in performing welfare analysis
although the results are similar to the baseline case with no instruments.
A second concern is that plan bids are correlated with unobserved household
tastes. This could happen if an insurer believed its plan was attractive due to, say,
a nearby clinic location. It would generate a correlation between the idiosyncratic
bid component, νjf, and household preferences εhj. We view this problem as most
likely of marginal importance given the limited information on the part of insurers.
Nevertheless, we check our estimates by instrumenting for plan contribution with a
predicted value that is constructed by plugging the predicted bid ˆ Bjf from (2.7) into
the contribution model (2.8). This specification purges the variation in both νjf and
ξjlf. The results are similar to our preferred specification.
A third issue for identification is that employer pass-through rates might be sys-
tematically influenced by employee preferences. This also seems unlikely, mainly
because pass-through rates in our data are uncorrelated with observable differences
across firms. Figure 2.4 plots employer pass-through rates against employee health
status, dependent health status, worker income and firm size. There is no correla-
tion, suggesting that cross-firm differences in contribution policies may be due more
to idiosyncratic factors, such as management philosophy, than employee tastes. Nev-
ertheless, we again use an IV strategy to verify that our results are not driven by a
correlation between the pass-through coefficients γjlf and unobserved preferences εhj.
To this end, we instrument for plan contribution using predicted values from a variant
of the contribution model (2.8) in which pass-through coefficients are restricted to be
identical across firms. This purges cross-firm variation in γjlf as well as the variation
in ξjlf. The results are again similar although with large standard errors. 19
19 A final identification concern is that household choices may be influenced by the health status of

CHAPTER 2. HEALTH PLAN CHOICE 88
2.5.1 Model Estimates
Table 2.4 presents parameter estimates from three different specifications of the de-
mand model. 21 The first column is a baseline model where we do not instrument
for plan contributions, and do not allow for private information about household
risk. The second and third columns instrument for plan contributions using the pre-
dicted values from the contribution model (2.8). The third column, which is our
preferred specification, allows for private information about risk. To scale the utility
to money-metric form, we divide each coefficient by the coefficient on the monthly
contribution and adjust the standard errors accordingly. We report the price effects
as semi-elasticities at the bottom of the table.
Effect of Demographics and Risk on Choice
The demand estimates indicate that overall sorting on the basis of risk is rather
modest, but that different plans experience unfavorable selection across differing com-
ponents risk. Older employees, who on average cost more to insure, prefer the network
HMO and the integrated POS plan to the integrated HMO. An additional year of age
is associated with an increase in the willingness to pay for the network HMO relative
to the integrated HMO of $1.75 per month (Column 1). Because older people are
often in worse health, they are likely to place a higher value on the broader provider
network of the network plan which would give them greater freedom in choosing
among providers. Women, who at the age of workers in our data typically cost more
to insure than men, prefer the integrated HMO to either the integrated POS plan
or the network PPO. Women are willing to pay $35 per month less than men for
the network PPO relative to the integrated HMO (Column 1). Women may have
stronger preferences for the integrated HMO if they perceive that it is more effective
in providing preventive cares since, in this age group, more preventive services are
recommended for women than for men. The effects of age and sex are not particularly
sensitive to the use of instruments for the employee contribution (Column 2) or the
21 The Table does not report every parameter. The parameters not reported are the plan fixed
effects, and the coefficients on imputed household income and an indicator for non-standard drug
coverage.

CHAPTER 2. HEALTH PLAN CHOICE 89
incorporation of unobserved risk (Column 3).
We find some sorting on the basis of health status conditional on age and sex,
driven primarily by having a very high risk household member. The effects of the
linear risk score on plan choice are generally small and imprecise. Households with
a high risk member, however, are less likely to enroll in the network HMO and more
likely to enroll in the network PPO than the integrated HMO. In our preferred spec-
ification (Column 3), an employee with a high risk family member is willing to pay
$28 per month more than an employee without a high risk family member to enroll
in the network PPO relative to the integrated HMO. This is consistent, once again,
with those who are more likely to use care placing a greater value on less restrictive
provider networks.
Our results also suggest that private information about health risk plays a role
in plan choice, although the estimate is not precise. We estimate that the standard
deviation of private risk information σµ is 0.68, which is substantial relative to the
standard deviation of the observed risk scores (1.56 in Table 2.1). Roughly speaking,
observed risk scores appear to pick up just over 2/3 of the health status information
that factors into plan choice.
While our findings with respect to risk selection are not inconsistent with existing
research, they suggest a relatively complex pattern of sorting. Much of the existing
literature finds evidence of unfavorable selection into more generous plans (??) .
We also find that the highest risk enrollees favor the most flexible plan, the network
PPO. Overall, however, the average risk across plans is quite similar due to off-
setting selection along different demographic dimensions, including age and gender,
that are correlated with risk. This finding is consistent with the idea that the plans
cater to individuals with different tastes for health care, rather than offering different
quantities of care, or targeting individuals of different health status.
Effect of Plan Prices on Choice
In the bottom panel, we present price semi-elasticities of demand, defined as the
percentage decrease in market share resulting from a $100 increase in the annual
enrollee contribution, evaluated at the mean choice probability for each plan. On

CHAPTER 2. HEALTH PLAN CHOICE 90
average, a $100 dollar increase in the annual enrollee contribution decreases market
share by 7 to 9 percent. While instrumenting for the contribution reduces the precision
of the estimate, it has relatively little effect on its magnitude. These estimates suggest
that demand is relatively inelastic, in line with other estimates in the literature. In
the Online Appendix, we discuss studies in settings similar to ours. Across these
studies, a $100 increase in the annual contribution reduces market share by 1.6 to 9.6
percent, placing our estimate in the middle to high end of the range.
The results in Table 2.4 also include the estimated value of plan characteristics
other than price, such as coinsurance rate and deductible. Enrollees appear to be
moderately sensitive to both. We estimate that a 10 percentage point increase in
the coinsurance rate is valued at approximately $276 dollars annually, which is about
10 percent of the annual cost per enrollee reported by the insurers. Our estimates
indicate that enrollees are not particularly sensitive to the deductible when choosing
among plans. 22
Because the estimates of risk and price elasticity are the key parameters for our
welfare calculations, we have examined the sensitivity of these estimates to a variety
of issues. In the Online Appendix we present estimates where we vary the sample of
households and use different instruments (discussed in Section 2.5.4) for the employee
contributions. We also discuss specifications with alternative sets of controls. The
bottom line is that the estimates are robust to variation across these dimensions.
Structure of Plan Costs
The difference in cost structures for the integrated and network insurer can be
seen in the raw data depicted in Figures 2.5 and 2.6. Figure 2.5 is a scatterplot of
enrollee risk scores against realized costs. Each point corresponds to an insurer-firm-
year. The x-axis is the average risk of the insurer’s enrollees; the y-axis is the reported
costs per enrollee. The lines represent the model’s prediction (shown in Table 2.5)
of expected costs for the network PPO and the integrated HMO. Figure 2.6 displays
corresponding information for bids. It shows the average risk of a firm’s employees
plotted against plan bids, with each observation at the plan-firm-year level.
22 The results are unchanged when out-of-pocket maximum are included as plan characteristics.

CHAPTER 2. HEALTH PLAN CHOICE 91
As the figures illustrate, the plans seem to have similar costs for enrollees with
average health risk and divergent costs for enrollees in good and poor health. The
expected monthly cost for an enrollee with a risk score of 1 is $235 for the integrated
HMO, $236 for the integrated POS, $218 for the network HMO, and $238 for the
network PPO. For less healthy enrollees, the integrated insurer has a substantial cost
advantage. The expected monthly cost of an enrollee with a risk score of two is $309
for the integrated HMO, compared to $507 for the network HMO and $413 for the
network PPO. Network plans do relatively well for low risks. The expected monthly
cost for an enrollee with a risk score of 0.5 is $198 for the integrated HMO, as opposed
to $151 for the network PPO and $74 for the network HMO.
The structure of plan costs we estimate is consistent with the basic idea that
patient cost sharing may be effective at limiting provider visits while supply-side
mechanisms may be more effective at limiting costs conditional on receiving services
(see, e.g., ?). While we do not have visit-level data to support the claim, the steep cost
curves for the network plans are consistent with cost sharing limiting visits, particu-
larly for low risks, but having little effect on the high risks who consume healthcare
on the intensive margin. In contrast, the integrated plans with their relatively low
cost sharing but stronger supply side utilization controls may be less effective at lim-
iting provider visits for low risks but more effective at managing costs conditional on
provider visits for the high risks. Another factor explaining the relatively high costs
for low risks in the integrated plan may be the greater use of preventive services.
These results also provide some insight on the potential issue of bias in the risk
scores. As discussed earlier, if the risk scores are biased, the likely direction of the bias
is an underestimate of the risk of the enrollees in the integrated plan. This is because
the risk scores are based on the types of prescription drugs used by plan enrollees
and the integrated plan more aggressively manages drug utilization. If this were the
case, then we would underestimate the cost savings for high risks generated by the
integrated plan. In other words, the large cost differentials we observe would be even
larger. While we believe that bias in the risk scores is unlikely to be an important
issue, if it exists, it would likely reinforce our key findings.
We note that our estimates for the integrated POS plan indicate that its costs

CHAPTER 2. HEALTH PLAN CHOICE 92
are more similar to those of the network plans than the integrated HMO. While
these results are not consistent with our findings regarding the differences in the cost
structures between the two types of insurers, since the bulk of the integrated plans
enrollment is concentrated in the HMO product, it is likely that the HMO is more
representative of its cost structure, particularly given our small sample size. The
other inconsistency in our cost results is the difference in the cost structure between
the plans offered by the network insurer. In particular, in contrast to the integrated
HMO, the network HMO appears not to generate cost savings for high risks relative
to the PPO product. While it is difficult to understand why this plan would be
more expensive for unhealthy enrollees relative to its less managed PPO counterpart,
we note that our estimates are based on relatively small samples, making it not
particularly surprising that some of our results may not have an intuitive explanation.
In addition, many factors are likely to affect health plans costs so our estimates may
also reflect true differences in cost structures for which we did not have strong prior
beliefs. The estimated mark-up varies across insurers. We estimate that the network
insurer and the integrated insurer bid 24 percent and 8 percent over expected costs,
respectively.
The sensitivity of cost differentials as a function of enrollee risk, compared to the
relatively modest effect of risk on plan preferences, has an important implication.
It indicates that as consumer risk varies, changes in relative plan costs rather than
changes in preferences will drive the efficient allocation. As our simple theory model
illustrated, this will not happen under self-selection without a mechanism that allows
different risk groups to face different premium differentials. In our setting, prices do
not have this feature, suggesting the potential for inefficiency. We return to this point
in the next section, when we quantify social welfare.
A factor to keep in mind when evaluating our estimates of plan costs is that we
observe the insurers’ costs of coverage, not the overall dollars spent on care. The
distinction is important because, in plans with copayments and deductibles, enrollees
bear a share of the cost of care that we do not capture in our data. These payments
are largest at the network PPO and smallest at the integrated HMO. While our
model assumes that these payments will be internalized in making plan choices, they

CHAPTER 2. HEALTH PLAN CHOICE 93
do affect the interpretation of the effects of the different plan types on utilization of
care. In particular,the reduction in insured costs for low risks in the network plan may
represent, at least in part, a shift from insured to uninsured payments, rather than a
reduction in utilization or prices. For high risks, in contrast, the difference in insured
costs between the plans likely underestimates the extent to which the integrated plan
reduces total costs.
2.5.2 Quantifying Social Welfare Inefficiencies
In this section, we use the estimated demand and cost model to compute the ineffi-
ciency associated with observed contribution policies relative to alternative efficient
benchmarks. We also compare welfare between the observed policies and alterna-
tive uniform contribution policies to demonstrate the extent to which the inefficiency
associated with a uniform contribution could be reduced within the current institu-
tional constraints. Table 2.6 presents the results of these simulations. The left-hand
panels present the market share, average enrollee risk, and the average incremental
contribution for each plan under five different pricing scenarios. The incremental con-
tribution represents the monthly contribution per enrollee relative to the integrated
HMO averaged across all households. The right-hand panels present information on
the change in surplus relative to the observed allocation for each scenario.
The Welfare Cost of Observed Prices
In the top panels, we calculate the inefficiency of observed pricing policies rela-
tive to two risk-rated benchmarks. The first is individual risk rating based on the
observed risk scores. This pricing policy, which we refer to as “feasible risk-rated
contributions”, maximizes social welfare conditional on knowledge of the risk scores,
but not each household’s private information. The third panel of the Table reports
outcomes when prices are first-best, i.e. risk-rated based on both public and private
information.
Overall, under risk-rated contributions, high-risk households face higher premiums
and low-risk households face somewhat lower premiums for the network plans relative
to observed contribution policies. In both the feasible and first-best scenarios, this

CHAPTER 2. HEALTH PLAN CHOICE 94
leads to a substantial re-allocation of enrollees across plans, although overall market
shares change modestly. With feasible risk-rating, the average enrollee risk at the
integrated HMO increases from its observed level of 0.99 to 1.49, and the network
HMO experiences a decline in average enrollee risk from 1.03 to 0.58. This reallocation
of households across plans substantially reduces overall insurer costs, by $44 per
enrollee-month, and increases total social surplus by just over $27 per enrollee-month.
The increase in social welfare represents approximately 11% of average insurer costs
in our sample.
A substantial fraction of the welfare gain is due to the highest and lowest risk
households making more efficient plan choices. Table 2.7 decomposes the welfare
calculation by household risk quintiles. The lowest and highest risk quintiles (aver-
age household risk below 0.36 and above 1.33) generate about three-quarters of the
welfare effect. This raises a concern that our calculation might be driven in part by
extrapolating plan costs out of sample. As Figures 2.5 and 2.6 illustrate, we observe
plan bids and costs only for average risk scores between 0.75 and 2.0. In contrast,
household risk ranges from 0.16 to 30.1. To address this, we truncate the cost differ-
entials between plans at their 0.75 and 2.0 levels and re-calculate the welfare numbers.
These calculations appear in the final columns of Tables 2.6 and 2.7. We view the
numbers based on truncated cost differences as a lower bound on welfare differences,
and the baseline numbers based on straight-line extrapolation as closer to an upper
bound. Truncating the cost differentials has little effect on the resulting assignment
of households to plans, but as one might expect, it reduces the welfare cost of ob-
served pricing to $5 per enrollee-month, or 2% of insurer costs, relative to the feasible
optimum.
It is also interesting to compare what is possible using prices based on observed
risk scores to what in principle could be achieved using both observed risk scores and
households’ private information. This calculation captures the extent to which private
information on risk constrains the efficiency of feasible relative to optimal risk-rated
pricing. Changing from feasible risk rated contributions to the first-best scenario
increases social surplus by between $2 and $8 per enrollee-month, depending on the
treatment of costs for extreme risk, or roughly 1-3 percent of insurer costs. One way

CHAPTER 2. HEALTH PLAN CHOICE 95
to interpret this is that, in our sample, a social planner could achieve approximately
70% of the potential welfare gains associated with individualized pricing using only
observable information on risk.
Social Welfare without Risk-Rating
The calculations above indicate that the observed prices fall well short of the effi-
cient benchmark. A natural question is whether efficiency gains could be realized even
without risk-rated contributions. That is, to the extent re-allocating high and low
risk households would increase social welfare, is it possible to induce this reallocation
given current institutional pricing constraints? At first glance, the answer is unclear.
After all, current institutions require uniform pricing within firm-tiers, but this still
allows a fair amount of pricing flexibility within our sample. For example, average
risk varies substantially across the firms in our data, suggesting that cross-firm vari-
ation in contribution policies could alleviate some of the inefficiency associated with
uniform contributions.
The next scenario in Table 2.6 addresses the question of what is possible with-
out individualized pricing by considering contributions that maximize social welfare
subject to being uniform within each firm-tier. As in the case of fully risk-rated
prices, optimizing uniform within firm contributions leads to a reallocation of high-
risk households into the integrated plans and away from the network plans, particu-
larly the PPO. The shift is much less dramatic, however, than under full risk rating.
Overall social surplus is $1.40-6.70 higher per enrollee-month than under the observed
policies, but still $3.60-20.40 below the efficient level. This indicates that about 3/4
of the observed inefficiency is due to the requirement of nondiscriminatory pricing
within firms. Nevertheless, it appears that employers could increase social surplus
by around 1-3% of average insurer costs simply by adjusting their contributions to
better reflect differences in underlying plan costs.
One difficulty for employers, of course, is that matching contributions to plan
costs may be a fairly complex exercise. Many benefits consultants, including the
intermediary in our data, suggest a simpler approach, which is to pass on the full
incremental premium for all but the lowest priced plan. We refer to this as the

CHAPTER 2. HEALTH PLAN CHOICE 96
“Enthoven Rule”(?). About 1/2 of the firms in our sample use this approach for at
least some workers. The last entry in Table 2.6 considers the effect of moving all
the firms to an Enthoven-style approach. Perhaps surprisingly, this has relatively
little effect on overall welfare, or on household choices. The reason is that demand
is not very price elastic and from a practical standpoint most firms already pass
through a substantial fraction of the premium differentials. So relative to the price
changes needed to move substantial numbers of households across plans, a change to
an Enthoven policy has only a modest effect on choices.
This last observation raises an important point for our pricing experiments. The
relatively low elasticity of demand means that the contribution differentials needed
to re-allocate households in the direction of efficiency are sizeable. For instance,
maximizing welfare while keeping contributions uniform within firm-tiers would lead
to some households seeing a $87 per-enrollee monthly premium for the network PPO
relative to the integrated HMO. A move to efficient risk-rated prices would increase
this differential even more for some high-cost households. For instance, an individual
employee with a risk score of 3 would face a monthly premium differential of between
$101 and $202 depending on our cost extrapolation. These large price differentials
indicate that achieving efficient allocations may raise issues of fairness or affordability
of coverage for particular subgroups.
2.5.3 The Value of Plan Choice
By choosing to offer benefits through the intermediary, each of the firms in our sample
moved from offering a single health plan to offering multiple plans from two carri-
ers. A clear benefit of plan choice is that households with different preferences can
select their preferred plan. Our estimates indicate a substantial amount of preference
heterogeneity, and hence suggest substantial welfare gains from giving households
multiple plan options.
To illustrate this, Table 2.8 compares aggregate surplus under the observed of-
ferings to the surplus that would be obtained if all the households in our sample
were enrolled in one of the four plans. The most natural benchmark is the integrated

CHAPTER 2. HEALTH PLAN CHOICE 97
HMO, as it would be the most efficient single-plan offering for every firm in our data.
Relative to the integrated HMO benchmark, the observed plan offerings increase so-
cial welfare by almost $70 per enrollee-month for the firms in our data. Virtually
all of this is due to an increase in consumer surplus (gross of plan contributions)
rather than to a reduction in insurer costs. Indeed, insurer costs would be lowest if
all households were enrolled in the network HMO but the reduction in social surplus
would be large due the reduction in consumer surplus.
One caveat to this calculation is the logit demand specification is notorious for
generating large “new product” welfare gains. Roughly speaking, the problem is that
each new product adds a new preference dimension, and some households invariably
enjoy a large welfare gain from this addition due to the logit distributional assump-
tion. So while we think the benefits of plan choice are real, we urge some caution in
interpreting the magnitude of the measured effect.
2.5.4 Discussion and Sensitivity Analysis
Our estimates of market inefficiencies are based on a particular set of employers in a
particular geographic area. One way to address external validity is to compare our
estimates with some other studies of specific environments, such as ?, ?, and ?. These
studies all rely on data from individual large employers, and in each case, the plans
are plausibly distinguished by their level of generosity, making the environments a
bit different from ours. All three studies find evidence that more generous plans are
adversely selected. Cutler and Reber document this by using enrollee age as a proxy
for risk. The latter two studies, like ours, use data on realized costs.
Despite the difference in institutional settings, the bottom line welfare estimates
from these studies are fairly similar, and also similar to our estimates. Cutler and
Reber estimate that observed prices at Harvard University reduce welfare by around
2-4% of coverage costs relative to optimal uniform prices. Einav, Finkelstein and
Cullen estimate that in their setting average cost pricing has a welfare cost of roughly
2% relative to optimal uniform pricing. Carlin and Town find much smaller welfare

CHAPTER 2. HEALTH PLAN CHOICE 98
effects, due to very low demand elasticity. 23 Note that these papers all focus on
uniform pricing, which we have noted is generally inefficient except in special cases.
When we use our estimates to compare observed pricing to optimal uniform prices,
we find welfare costs of approximately 1-3% of coverage costs. In this sense, there
appears to be a fair amount of agreement between studies.
As a group, these studies also reinforce our earlier observation that inefficiencies
from pricing can be driven both by the nature of sorting and risk-selection, and by the
price elasticity of demand, which determines the extent to which implicit subsidies or
taxes affect choices. To guage the sensitivity of our own estimates to these factors,
we recalculated the surplus difference between the observed and the feasible efficient
allocation assuming that demand was twice as sensitive to price as we have estimated,
and half as sensitive. We performed a similar analysis varying the risk sensitivity of
demand. These analyses increase the range of the welfare gains to 1-13% of total
coverage costs. Given the range of demand estimates in the literature, one may want
to assign a corresponding range of uncertainty to the potential welfare costs of price
distortions.
2.6 Conclusion
Economists have long understood that competition in health insurance markets is no
guarantee of efficiency. This paper contributes to a nascent literature that attempts
to quantify market inefficiencies and identify their sources. A main finding is that
observed contribution policies distort enrollment decisions from their efficient level.
We calculate that the welfare loss is on the order of 2-11% of the total cost of coverage.
Capturing these gains in full would require the use of risk-rated contribution policies.
Absent such policies, optimally set employee contributions might increase welfare by
1-3% of coverage costs. A key point to emphasize is that despite these distortions,
there appear to be substantial gains in our context from introducing plan choice at
23 One explanation for their inelastic demand estimate is that they rely on time-series variation
in contributions. As discussed above, employees appear to be more price sensitive in making initial
choices than in making changes once they are enrolled.

CHAPTER 2. HEALTH PLAN CHOICE 99
the employer level.
One important point about risk-rated premiums in health insurance is that cov-
erage is typically purchased on an annual basis. While risk-rating might increase
static efficiency, it exposes households to reclassification risk as their health status
changes over time. This is one argument for community rating or nondiscrimation in
employer contributions. It is interesting to ask whether there are ways to promote
static efficiency that nevertheless mitigate reclassification risk. One natural approach
is to ensure that consumers have a baseline option whose price is independent of risk,
and allow incremental purchases to be risk-rated. An alternative would be longer-
term contracts that provide dynamic insurance (e.g., ?). Understanding the dynamic
aspects of health insurance are an important challenge for future empirical work.
2.7 Appendix to Chapter 2
This Appendix compares our demand estimates to the broader literature on health
plan choice, and discusses alternative specifications of the consumer demand model.
2.7.1 Comparison to Demand Estimates in Other Settings
As reported in Table 2.4, we find that on average a $100 dollar increase in the annual
enrollee contribution decreases market share by 7 to 9 percent. Studies in similar
settings vary in their method for reporting elasticities. Following ?, we reconcile
the estimates of price elasticity across studies by converting them to semi-elastiticies
(calculated as the percent change in market share in response to a $100 increase
in the employee contribution). Because the studies cover different time periods, we
adjust the change in the employee contribution for inflation. After making these
adjustments, most studies estimate semi- elastiticities in the range of -1.5 to -4.3 for
the main specifications (????). Estimates from ? are somewhat higher (ranging from
-2.4 to -9.6 for the main specifications and even higher in some models). Similarly, ?,
who analyze retirees, have an estimate that implies a semi-elastiticity of -9.0.

CHAPTER 2. HEALTH PLAN CHOICE 100
2.7.2 Alternative Specifications of the Demand Model
As we discuss in the main text, the estimates of risk and price elasticity are key
parameters for our welfare calculations. Table 2.9 provides additional estimates of
the demand model to examine the sensitivity of these parameters. Column 1 of Ta-
ble 2.9 repeats the first specification of Table 2.4 as a baseline. This specification
pools data from enrollees in different coverage tiers, increasing the sample size and
generating additional variation in plan contributions due to the higher pass-through
rates for dependent coverage tiers, but requires us to make assumptions about house-
hold aggregation. We test whether our estimates are sensitive to these assumptions
by restricting the sample to employees purchasing employee-only coverage (Column
2) and controlling for family structure in our full-sample specification (Column 3).
Both variations give similar results to our baseline specification (Column 1). While
the estimate of the effect of being a high risk individual on demand for the network
PPO is smaller in magnitude in the employee-only sample (Column 2), the estimate
is imprecise likely because it is identified by a small number of enrollees. When we in-
clude controls for family structure in the model estimated on the full sample, we find
that families with a spouse seem to prefer the PPO relative to the average household
(Column 3).
As discussed in Section 2.5.4 , we develop different instruments to exploit alterna-
tive sources of identifying variation in employee contributions, testing the robustness
of our estimates to different exogeneity assumptions. In Table 2.4 (Columns 2 and 3),
we demonstrate that the estimate of the price effect is robust to across-plan variation
in employer contribution setting. To address the concern that employer contribution
rates may be systematically influenced by employee preferences, we instrument for
employee contributions with a variant of the contribution model in which the the
pass-through coefficients are restricted to be identical across firms (Column 4). We
view the concern that insurer bids are correlated with household preferences to be
diminished because of the limited information insurers have in our setting. Never-
theless we report a specification where we include predicted bids in the contribution
model (Column 5). The results from both these specifications are similar although
with larger standard errors.

CHAPTER 2. HEALTH PLAN CHOICE 101
Finally, we note that while we present results from a relatively parsimonious
model, our estimates do not appear to be sensitive to including a variety of different
control variables. Additional controls we have explored include plan out-of-pocket
maximums, whether a plan was offered to an employer group prior to the employer
hiring the intermediary, and the average health status of an employee’s co-workers.

CHAPTER 2. HEALTH PLAN CHOICE 102
Table 2.1: Risk and Demographics
Mean Sd. Min. Max.
Employees (N = 3683)
Risk Score 1.21 1.56 0.18 30.06
Age 40.56 12.01 18.00 72.00
Female 0.62 0.48 - -
Spouse 0.28 0.45 - -
Child 0.27 0.44 - -
Enrollees (N = 6603)
Risk Score 1.01 1.45 0.14 30.06
Age 32.13 17.67 0.00 72.00
Female 0.58 0.49 - -
Spouse 0.19 0.39 - -
Child 0.26 0.44 - -
Firm-Years (N = 16)
Risk Score 0.97 0.31 0.63 1.91
Age 31.67 4.63 25.71 46.09
Female 0.53 0.12 0.30 0.70
Spouse 0.19 0.07 0.08 0.27
Child 0.26 0.08 0.06 0.39
Employees 230.19 241.51 28.00 838.00
Dependents 182.50 117.51 9.00 331.00
Notes: In the first panel, spouse and child refer to the fraction of employees who enroll
with a spouse or at least one child. In the second and third panels, these entries are the
fraction of spouses and children in the set of enrollees. The first and second panels pool
observations across firms and years. The third panel shows statistics of firm-year level
averages, taken across all enrollees.

CHAPTER 2. HEALTH PLAN CHOICE 103
Table 2.2: Plan Characteristics
Network Integrated
HMO PPO HMO POS All
Offering Plan
Firms 11 10 11 9 -
Firm-Years 16 14 16 13 -
Bid (Monthly)
Employee 307 332 260 276 294
(64) (59) (30) (26) (54)
Employee plus spouse 645 689 544 579 616
(154) (123) (61) (54) (120)
Employee plus child(ren) 591 632 498 532 565
(143) (115) (58) (53) (111)
Employee plus family 918 989 779 832 882
(200) (176) (87) (76) (164)
Contribution (Monthly)
Employee 45 73 38 58 53
(34) (54) (32) (40) (41)
Employee plus spouse 252 303 203 255 253
(120) (103) (77) (75) (100)
Employee plus child(ren) 221 265 177 223 222
(97) (86) (62) (55) (81)
Employee plus family 418 495 342 415 418
(213) (182) (144) (140) (176)
Coinsurance (%)
Employee 87 86 97 78 87
(6) (5) (7) (2) (9)
Deductible (Annual)
Employee 387 440 69 336 304
(264) (306) (163) (94) (262)
Out-of-Pocket Max (Annual)
Employee 2818 2850 1591 2686 2468
(462) (474) (625) (731) (775)
Notes: Mean plan characteristics, with standard deviations in parentheses. Plan
characteristics are pooled across years. Coinsurance, deductible, and out-of-pocket
maximum are in-network values and are highly correlated (ρ >.9) with out-of-network
coinsurance, deductible and out-of-pocket maximum. Coverage tiers based on employee
plus one dependent and employee plus two or more dependents are used at two firms.
Bids and costs for these coverage tiers are not shown.

CHAPTER 2. HEALTH PLAN CHOICE 104
Table 2.3: Risk and Demographics by Plan
Network Integrated
HMO PPO HMO POS All
Employees (N=3683)
Risk Score 1.19 1.22 1.22 1.19 1.21
Age 42.17 40.79 39.73 41.35 40.56
Female 0.62 0.52 0.65 0.56 0.62
Market Share (%) 22.94 7.38 58.72 10.96 100
Enrollees (N=6603)
Risk Score 1.02 1.04 0.99 1.05 1.01
Age 34.19 33.06 30.94 34.12 32.15
Female 0.58 0.54 0.59 0.55 0.58
Market Share (%) 21.24 7.84 60.35 10.57 100
Notes: Employees and enrollees are pooled across firms and years.

CHAPTER 2. HEALTH PLAN CHOICE 106
Table 2.5: Costs and Bids
(1) (2)
Network Insurer Markup 1.29 (0.12) 1.27 (0.07)
Integrated Insurer Markup 1.08 (0.04) 1.07 (0.03)
NHMO 218.42 (21.35) 195.08 (9.48)
X (Risk Score - 1) 288.25 (86.05) 265.36 (30.29)
X Coinsurance 0.32 (0.94)
NPPO 238.32 (22.65) 204.59 (9.82)
X (Risk Score - 1) 174.92 (34.34) 167.73 (23.81)
X Coinsurance 1.23 (1.29)
IHMO 234.86 (9.71) 228.77 (6.77)
X (Risk Score - 1) 73.67 (22.01) 104.80 (18.80)
X Coinsurance 0.38 (0.54)
IPOS 236.37 (14.13) 216.74 (13.01)
X (Risk Score - 1) 188.64 (69.35) 200.78 (61.60)
X Coinsurance 1.65 (0.82)
N 91 91
Notes: GMM estimates of cost parameters. See text for details. Coinsurance is de-meaned
at the plan level.

CHAPTER 2. HEALTH PLAN CHOICE 107
Table 2.6: Matching and Welfare under Alternative Contribution Policies
Matching Welfare † Truncated
Gross Insurer Social Social
NHMO NPPO IHMO IPOS Surplus ‡ Costs ‡ Surplus ‡ Surplus ‡
Observed
Market Shares 0.25 0.09 0.54 0.12 0.00 0.00 0.00 0.00
Risk Score 1.03 1.07 0.99 1.02
Incremental Contribution † 9.30 23.70 0.00 5.00
Feasible Risk Rated Contributions
Market Shares 0.37 0.09 0.43 0.11 -16.60 -43.70 27.10 5.00
Risk Score 0.58 0.78 1.49 0.74
Incremental Contribution -14.70 11.80 0.00 -1.30
Optimal Risk Rated Contributions
Market Shares 0.38 0.08 0.44 0.10 -22.10 -57.50 35.50 7.80
Risk Score 0.60 0.79 1.46 0.76
Incremental Contribution -14.90 11.80 0.00 -1.60
Uniform by Tier within Firms
Market Shares 0.31 0.09 0.49 0.12 -6.10 -12.80 6.70 1.40
Risk Score 0.86 1.02 1.11 0.97
Incremental Contribution -16.50 8.90 0.00 -1.10
Enthoven Rule
Market Shares 0.22 0.08 0.58 0.13 -1.10 -0.80 -0.30 -0.50
Risk Score 1.01 1.05 1.00 1.02
Incremental Contribution 28.70 39.90 0.00 10.80
Notes: Feasible Risk Rated Contributions implements efficient matching by setting
incremental contributions equal to incremental costs, conditional on observable risk but
not privately known risk. Optimal Risk Rated Contributions sets incremental
contributions equal to incremental costs, conditional on both observable and privately
known risk. Uniform by Tier within Firms maximizes social surplus subject to the
constraint that contributions vary only by coverage tier and by firm, but not by individual
risk. Enthoven Rule is implemented by setting incremental contributions equal to
incremental bids. Reported risk score is conditional on plan choice. The truncated results
holds cost differentials between plans for risk scores lower than 0.75 and higher than 2.0 at
these boundary levels.
† Incremental contribution, gross surplus, insurer costs, and social surplus are averaged
across enrollees and denominated in $ per month.
‡ Gross surplus, insurer costs and social surplus are normalized to zero under the
observed allocation. Other scenarios show gross surplus as social surplus relative to the
observed allocation. Under the observed allocation, costs average $241.70 per enrollee per
month. Gross and social surplus are not pinned down.

CHAPTER 2. HEALTH PLAN CHOICE 108
Table 2.7: Matching and Welfare by Risk Score Quintile
Feasible Risk Rated Contributions versus Observed
Matching Welfare Truncated
∆Gross ∆Insurer ∆Social ∆Social
Quintile (Risk Score range) NHMO NPPO IHMO IPOS Surplus Costs Surplus Surplus
Quintile 1 (¡0.36)
∆MarketShare 0.332 0.000 -0.330 -0.002 -27.2 -56.9 29.8 4.3
∆IncrementalContribution -179.4 -93.4 0.0 -86.6
Quintile 2 (0.36, 0.54)
∆MarketShare 0.265 0.003 -0.266 -0.001 -16.6 -35.6 18.9 3.4
∆IncrementalContribution -141.6 -75.9 0.0 -65.6
Quintile 3 (0.54, 0.79)
∆MarketShare 0.181 0.006 -0.189 0.002 -7.7 -17.1 9.3 1.3
∆IncrementalContribution -99.1 -53.4 0.0 -44.6
Quintile 4 (0.79, 1.33)
∆MarketShare 0.040 0.004 -0.037 -0.007 -0.8 -2.4 1.6 0.4
∆IncrementalContribution -21.0 -19.7 0.0 -1.2
Quintile 5 (¿1.33)
∆MarketShare -0.184 -0.047 0.299 -0.069 -30.3 -105.9 75.6 15.4
∆IncrementalContribution 324.8 154.5 0.0 179.3
Total
∆MarketShare 0.128 -0.007 -0.106 -0.015 -16.6 -43.8 27.1 5.0
∆IncrementalContribution -23.9 -11.9 0.0 -6.3
Notes: ∆ Market Share, ∆ Incremental Contribution, ∆ Gross Surplus, ∆ Insurer Costs
and ∆ Social Surplus are calculated as the difference between the feasible risk rated and
observed values of these variables. Truncated fixes cost differentials between plans for risk
scores outside of 0.75 and 2.0. Values averaged across enrollees within each quintile and
denominated in $ per month. (Total values are averaged across all enrollees.)

CHAPTER 2. HEALTH PLAN CHOICE 109
Table 2.8: The Value of Plan Choice
Gross Surplus ‡
Welfare †
Insurer Costs ‡
Social Surplus ‡
Observed 0.0 0.0 0.0
All enrolled in:
NHMO -148.8 -9.2 -139.7
NPPO -216.9 5.8 -222.7
IHMO -71.4 -2.1 -69.4
IPOS -180.7 4.5 -185.2
Notes: † Gross surplus, insurer costs, and social surplus are averaged across enrollees
and denominated in $ per month.
‡ Gross surplus, insurer costs and social surplus are normalized to zero under the
observed allocation. Other scenarios show gross surplus as social surplus relative to the
observed allocation. Under the observed allocation, costs average $241.70 per enrollee per
month. Gross and social surplus are not pinned down.

CHAPTER 2. HEALTH PLAN CHOICE 110
Figure 1B: Efficient allocation assigns high-risk to plan A
Figure 1A: Efficient allocation assigns high-risk to plan B
"c(!)
Value ($)
Value ($)
High-risk always self-select into plan B.
Choice of "p can induce any threshold.
No uniform "p generates the efficient
allocation.
"v(!)
"v(!)
High-risks always self-select into plan B.
Choice of "p can induce any threshold.
"p= "c( !* ) generates the efficient
allocation of risk types across plans.
"c(!)
"p
"p
Efficiently in A
Efficiently in B
Efficiently in B
Efficiently in A
Consumer Risk (!)
!*
Consumer Risk (!)
!*
Figure 2B: Efficient allocation assigns high-risk to HMO
Figure 2A: Efficient allocation assigns high-risk to PPO
High (!,") always self-select to PPO.
Choice of #p can induce any translation
of the pictured indifference curve.
#v(!,")=#p
Consumer
taste for
choice(")
High (!,") always self-select to PPO.
Choice of #p can induce any translation
of the pictured indifference curve.
#v(!,")=#p
Consumer
taste for
choice(")
#v(!,")=#c(!)
Choose
HMO
inefficiently
Choose PPO
efficiently
Choose PPO efficiently
Choose HMO,
Efficiently in PPO
Choose PPO
Efficiently in HMO
Choose HMO
efficiently
Choose HMO
efficiently
Consumer risk (!)
!’ s.t
#c(!’)=#p
Choose #v(!,")=#c(!)
PPO
inefficiently
Consumer risk (!)
!’ s.t
#c(!’)=#p

CHAPTER 2. HEALTH PLAN CHOICE 111
Figure 2.3: Contributions and Bids Relative to Integrated HMO
-50 0050
50 100 150 200 -50 0050
50 100 150 200 -50 0050
50 100 150 200 -50 0050
50 100 150 200 Employee Plus Spouse
Employee Plus Spouse
Network PPO
Network PPO
Network HMO
Network HMO
Integrated POS
Integrated POS
Y=X Incremental Contribution
Incremental Contribution
Incremental Bid
Incremental Bid
-50
200
Incremental Contribution
0 50 100 150
-50
-50
0
Employee
50
100
Network PPO
Integrated POS
200
150
100
50
0
-50
150 200 -50
Incremental Bid
Employee Plus Spouse
0
50
Network HMO
Notes: Incremental Contribution and Incremental Bid are relative to Integrated HMO. In
Employee Plus Spouse, numbers are divided by two for comparability.
Y=X
100
150
200

CHAPTER 2. HEALTH PLAN CHOICE 112
Figure 2.4: Employer Contributions and Employee Characteristics
0.2 .2 .4 .6 .8 .81 1Beta Beta .5 .51 1.5 2 2.5 Risk Score Risk Score 00.2
.2 .4 .6 .8 .81 1Beta Beta 1 1.2 1.4 1.6 1.8 22Risk
Score Risk Score 00.2
.2 .4 .6 .8 .81 1Beta Beta 20000 30000 40000 50000 60000 Income 00.2
.2 .4 .6 .8 .81 1Beta Beta 00500
500 1000 1500 Number of employees
Number of employees
0
Beta Risk Employee versus Score
versus Employee Risk Beta Score
Employee Plus Spouse Beta
Employee Plus Spouse Beta
versus Risk Score
versus Risk Score
Mean Beta versus Income
Mean Beta versus Income
Mean Beta versus Firm Size
Mean Beta versus Firm Size
Employee Beta versus Risk Score
1
.8
Beta
.4 .6
.2
0
1
.8
Beta
.4 .6
.2
0
.5
20000
1
30000
1.5
Risk Score
Mean Beta versus Income
40000
Income
2
50000
2.5
60000
1
.8
Beta
.4 .6
.2
0
1
.8
Beta
.4 .6
.2
0
0
Employee Plus Spouse Beta
versus Risk Score
1
1.2
1.4 1.6
Risk Score
500 1000
Number of employees
1.8
Mean Beta versus Firm Size
Notes: Each point represents a firm-year. Beta is the incremental pass-through of bids.
See text for details. Employee Beta versus Risk Score plots the beta for employees against
their mean risk score. Employee Plus Spouse Beta versus Risk Score plots the beta for
those in employee plus spouse plans against their risk score. Scatter plots for other
coverage tiers look similar. Mean Beta versus Income plots the mean beta (across
coverage tiers) against mean employee income. Mean Beta versus Firm Size plots the
mean beta (across coverage tiers) against the number of employees in the firm. Separate
scatter plots by coverage tier look similar.
2
1500

CHAPTER 2. HEALTH PLAN CHOICE 113
600
100 200 300 400 500 600 Costs .5 .51 1.5 2 2.5 Risk Score Risk Score Nework Insurer
Nework Insurer
Intergrated Insurer
Intergrated Insurer
100
500
400
Costs
300
200
100
.5
Nework Insurer
Figure 2.5: Costs by Risk Score
1
Intergrated Insurer
1.5
Risk Score
Notes: Each point represents a insurer-employer-year. Risk Score is the average enrollee
risk score in a firm-year. Costs is the insurers’ monthly cost per enrollee. Fitted lines
represent the Network HMO and Integrated HMO.
2
2.5

CHAPTER 2. HEALTH PLAN CHOICE 114
100 200 300 400 500 600 Bids .5 .51 1.5 22E[Risk
Score|Age,Male]
E[Risk Score|Age,Male]
Network HMO
Network HMO
Network PPO
Network PPO
Intergrated HMO
Intergrated HMO
Integrated POS
Integrated POS
100
600
500
Bids
400
300
200
100
.5
Network HMO
Intergrated HMO
Figure 2.6: Bids by Risk Score
Network PPO
Integrated POS
1
1.5
E[Risk Score|Age,Male]
Notes: Each point represents a plan-employer-year. E[Risk—Age, Male] is the average
predicted risk score of potential enrollees in a firm-year. Bids is the per-month bid. Fitted
lines represent the Network HMO and Integrated HMO.
2

Chapter 3
Private Coverage and Public
Costs: Identifying the Effect of
Private Supplemental Insurance on
Medicare Spending
with Marika Cabral
3.1 Introduction
In many insurance contexts, individuals hold primary and supplemental insurance
policies for the same basic outcome. A common example is the purchase of pri-
vate supplemental health insurance to top up publicly provided primary coverage. 1
However, unlike standard markets where complementarities between products are
typically isolated to the demand-side, complementarities between primary and sup-
plemental insurance can affect costs as well. A supplemental insurance policy that
reduces the consumer cost-sharing requirements (e.g., deductibles and co-payments)
1 Other examples include disability insurance, where private coverage can supplement public disability
insurance (SSDI) and rental car insurance, where coverage from a credit card can top-up the
primary rental car insurance provided by a rental company.
117

CHAPTER 3. MEDIGAP 118
of a primary insurance policy can affect total utilization, imposing a fiscal externality
on the provider of the primary insurance product.
This paper estimates the effect of private supplemental Medigap insurance on pub-
lic Medicare spending. Medicare is the primary insurer of elderly Americans, covering
about 70 percent of the care they receive. To insure against the remaining costs, most
Medicare beneficiaries (91 percent) hold some form of supplemental insurance, which
reduces out-of-pocket financial risk but also blunts the cost-sharing requirements of
Medicare that are designed to limit utilization.
To identify the effect of Medigap supplemental insurance on total utilization, we
construct an instrument that leverages discontinuities in Medigap premiums at state
boundaries. Medical costs, and thus the costs financed through supplemental Medi-
gap insurance, exhibit considerable within-state variation due to factors ranging from
household incomes to local physician practice styles to the supply of medical resources.
Yet despite this local variation, within-state variation in Medigap premiums is very
limited (?). This means that on opposite sides of state boundaries, otherwise identical
individuals who belong to the same hospital catchment area can face very different
Medigap premiums solely due to Medicare costs in other parts of their states. Our
empirical strategy uses this exogenous premium variation—and the corresponding
exogenous shift of individuals in and out of Medigap coverage—to identify the pre-
mium elasticity of Medigap demand and the causal effect of Medigap on total medical
spending.
The effect we identify is of substantial policy relevance. The high level and growth
rate of Medicare spending makes it a primary concern of long-run US federal budget
experts (?). Because of this, mechanisms that increase costs and the corresponding
policies that could be used to address these rising costs are attracting close atten-
tion. However, as we discuss below, the current estimates of the effect of Medigap are
biased due to selection on unobservables. This bias means that previous estimates
may not reflect the behavioral responses to typical policies, such as a tax on Medi-
gap premiums. Our approach overcomes these shortcomings allowing us to explore
economically meaningful policy counterfactuals.
In addition to the direct relevance for Medicare and Medigap policy, our study may

CHAPTER 3. MEDIGAP 119
be of interest more generally. Settings with private supplemental coverage and public
primary insurance are widespread. For example, in France more than 92 percent of
the population holds private supplemental insurance to protect against the substantial
coinsurance payments (10 to 40 percent) of the universal public health insurance
system, and private supplemental policies that top up public health insurance benefits
are popular in Austria, Belgium, and Denmark as well. 2 Yet despite economists’
interest in the interaction of public and private insurance as demonstrated by the
large literature on crowd-out (cf. ??), there is relatively little work on the fiscal
externalities associated with jointly held private and public insurance. 3
At a conceptual level, supplemental Medigap coverage could impose a negative or
positive fiscal externality on Medicare. Because Medigap policies reduce the marginal
price for care, it is straightforward to see that overall utilization and Medicare costs
could increase. Medigap policies could alternatively decrease costs if, for example,
reduced cost-sharing improved treatment adherence, reducing the risk of hospitaliza-
tion. In the context of Medicare supplemental insurance available to retired California
state employees, ? find there is a significant positive cost externality of supplemental
insurance, what they call an “offset effect,” for those that are chronically ill. 4 For
the overall retiree population, however, the externality is negative with supplemental
insurance increasing Medicare costs.
The conventional wisdom is that supplemental insurance on net imposes a negative
fiscal externality on Medicare, although the magnitude of the effect is contended. The
central estimate, used by the Congressional Budget Office (CBO) for official budget
2 Statistics are from the following sources: ? “Selected European Countries Health Systems,” and
? “The Grass Is Not Always Greener: A Look at National Health Care Systems Around the World”
Cato Policy Analysis no. 613. In Austria, about a third of the population has a supplemental private
insurance plan that covers additional charges not covered under the basic benefits, and about 30
percent of Belgians carry private supplemental policies. Approximately 30 percent of population
in Denmark purchases Voluntary Health Insurance (VHI) in order to cover the costs of statutory
copayments of the universal coverage package.
3 Exceptions include ?, ?, and ? on supplemental insurance.
4 The authors argue that drug coverage is probably the single most important channel through
with the offset operates for the chronically ill. Even though the time period we look at is before
Medicare Part D drug coverage was added, Medigap options that had drug coverage were very
unpopular as their premiums were very high. Thus, we may expect the even among the chronically
ill, Medigap private supplemental insurance may not have the offset effect the authors observe in
the CALPERs population.

CHAPTER 3. MEDIGAP 120
estimates, is that “Medigap policyholders use at least 25 percent more services than
Medicare enrollees who have no supplemental coverage.” Yet, as pointed out by ?, the
research CBO draws upon does not account for selection in Medigap enrollment and
may be biased. In particular, Lemiuex et al. argue that selection is probably adverse,
leading these studies to overestimate the impact of Medigap on costs. Recent work,
however, finds little or no evidence of adverse selection into Medigap. 5 In fact, Fang
et al. (2008) find evidence of significant advantageous selection into Medigap. The
authors argue that this advantageous selection is driven by cognitive ability, which is
correlated with better health.
We view our study as a natural next step in this literature. Our approach addresses
the issue of selection by using plausibly exogenous variation in Medigap premiums.
In particular, we take advantage of the fact that there is substantial within-state
variation costs but limited within-state variation in premiums in the Medigap market.
Using local medical costs as a control, we find that within-state but out of locality
medical costs are a powerful predictor of premiums, providing us with a source of
premium variation that is unrelated to unobserved determinants of individual medical
costs. We estimate a Medigap selection equation using this variation to identify the
premium elasticity of Medigap demand. Using the predicted values from this selection
equation to instrument for Medigap enrollment, we consistently estimate the impact
of Medigap on medical utilization.
Our main finding is Medigap coverage increases overall medical costs by 57 percent
on the margin. The effect is about 40 percent larger than comparable OLS estimates,
consistent with advantageous selection. We also find that Medigap demand is mod-
erately elastic, with our preferred estimate indicating that a 10 percent increase in
premiums reduces take-up by 4.7 percentage points. Both the Medigap effect and
demand semi-elasticity are precisely estimated and robust to alternative functional
forms and sets of controls.
In addition to indicating advantageous selection, the 40 percent larger instrumen-
tal variables estimates for the effect of Medigap are consistent with heightened moral
5 These papers include ?, ?, ?, and ?.

CHAPTER 3. MEDIGAP 121
hazard for individuals on the margin of Medigap choice. Since these marginal indi-
viduals are more price sensitive when it comes to premiums, it is natural for them to
be more sensitive to the cost-sharing instruments of Medigap, suggesting larger ef-
fects for these marginal individuals than the average Medigap policyholder. 6 For our
purposes, this feature is a strength of our empirical strategy rather than a weakness.
The effect for individuals marginal to a change in premiums is the effect of interest
for counterfactuals involving a premium tax or subsidy, as the individuals identifying
the estimated effect are exactly those who would be affected by such a policy.
We construct back-of-the-envelope policy counterfactuals of taxes and subsidies to
Medigap premiums using our estimated parameters. A 20 percent tax would reduce
coverage by 9.5 percentage points, implying a total budgetary savings of 6.2 percent of
Medicare costs of the sample. About 1/4 of the savings would come from tax revenues
and 3/4 from reduced medical costs. A 40 percent tax would save 11.4 percent and
a Pigovian tax that fully accounts for the negative fiscal externality would generate
savings of 18.1 percent. We note that our parameters are estimated using modest
premium variation, and thus our policy counterfactuals, which involve large taxes
and subsidies, should be viewed with some caution. As we discuss below, a more
nuanced tax policy that accounts for offsets might be able to generate both larger
costs savings and better health.
The remainder of the paper proceeds as follows. Section 3.2 describes background
information about Medicare and the supplemental insurance market. Section 3.3
briefly describes the data used in our empirical analysis. Section 3.4 presents our
empirical model. In Section 3.5 , we discuss our identification strategy. Results are
presented in Section 3.6. Section 3.7 contains the policy counterfactuals. Section 3.8
concludes.
6 The intuition for this relationship that suggests individuals at the margin of selecting into
insurance based on premiums will be more sensitive to reductions in coinsurance. A similar intution
is formalized in a model in a working paper by ?. In their paper, the authors refer to this sort of
selection as “selection on moral hazard.”

CHAPTER 3. MEDIGAP 122
3.2 Background
Medicare is the government health insurance program for elderly, covering nearly
all individuals 65 and older in the United States. In 2008, Medicare covered 44
million beneficiaries at a cost of $468 billion, accounting for about 13 percent of
government expenditure and 3.2 percent of GDP. By 2025, Medicare is projected to
cover 71 million beneficiaries, consuming 20 percent of government expenditure and
5.5 percent of GDP. 7
Medicare beneficiaries choose between publicly administered traditional fee-for-
service (FFS) insurance coverage and Medicare Advantage policies, private insurance
plans with premiums subsidized by Medicare. 8 The majority (88 percent) of Medicare
beneficiaries have FFS coverage. This coverage allows beneficiaries their choice of
doctor and the ability to see a specialist without a referral. To control costs, fee-
for-service Medicare uses cost-sharing, exposing beneficiaries to about 20 percent of
the cost of care received on the margin. A key feature of traditional Medicare is
that there is no annual or lifetime out-of-pocket maximum amount that a beneficiary
will pay, so individuals are exposed to significant financial risk. Traditional Medicare
cost-sharing instruments for 2000 and 2005 (the first and last years of our sample)
are shown in Table 3.1.
To protect against the financial risk of traditional Medicare, the vast majority
(91 percent) of beneficiaries carry supplemental insurance. Some beneficiaries obtain
supplemental insurance through a former employer, poor beneficiaries may qualify for
supplemental insurance through the government sponsored Medicaid program, and
everyone else has the option to purchase private Medigap coverage. Medicare Ad-
vantage is an alternative to traditional fee-for-service Medicare that is available to
many beneficiaries. 9 These private insurance policies are subsidized by Medicare, and
7 Percent of GDP numbers are gross of premiums; percent of budget numbers are net (Annual
Report of the Boards of Trustees of the Federal Hospital Insurance and Federal Supplementary
Medical Insurance Trust Funds, 2010; Analytical Perspectives: Budget of the U.S. Government,
2010).
8 We will refer to Medicare Part C as Medicare Advantage in this paper, although this option was
called Medicare + Choice during the beginning of the period we analyze.
9 Although Medicare Advantage beneficiaries can technically sign up for Medigap policies, this is
discouraged by Medicare and individuals do not seem to do this in practice. Medicare Advantage

CHAPTER 3. MEDIGAP 123
have benefits that generally cover more services than offered in traditional Medicare
in exchange for beneficiaries accepting a more restricted network of providers. In our
analysis, we model an individual’s choice between three options: Medicare Advan-
tage, traditional Medicare without supplemental insurance, and traditional Medicare
with supplemental insurance. In doing this, we exclude individuals on Medicaid or
employer provided supplemental health insurance. 10 Table 3.2 shows the breakdown
of individuals on the various types of supplemental insurance.
Medigap beneficiaries typically purchase Medigap insurance within six months of
turning 65 years old and signing up for traditional Medicare, during what is called the
open enrollment period. During this period, individuals cannot be denied coverage for
any reason, and pricing is limited to a small set of characteristics: gender, location,
and smoking status of the enrollee. 11 Plans bought during this period are guaranteed
renewable as long as Medigap enrollees pay plan premiums every year. The federal
government regulates how Medigap policy prices can evolve. In particular, when
an individual enrolls in a Medigap plan, he is choosing an age-price profile that
may be adjusted with medical inflation but may not be contingent on his current
or future health status. Thus, along with the contemporaneous benefits, Medigap
coverage provides insurance against reclassification risk in future periods. In practice,
individuals tend to sign up for a Medigap plan at age 65, and renew their policy each
year. 12
In most states, individuals choose from a menu of standardized plans labeled
A through J, administered by competing private insurance providers. 13 Table 3.3
plans tend to give more financial coverage in exchange for a more restricted network of providers relative
to traditional Medicare. Since Medigap polices are tailored to fill gaps of traditional Medicare,
the supplemental coverage that these policies would provide an individual on Medicare Advantage
would be largely redundant.
10 Those who are offered employer sponsored insurance typically receive a good deal in that the
plans are very subsidized. Since we believe that the take-up rate on employer-sponsored insurance
is nearly 100 percent, we excluded those on employer sponsored supplemental insurance from our
analysis as they have a different and unobserved choice set.
11 In practice, smoking status is not priced during the sample period we examine.
12 This is the pattern we observe in the 3-year panel of the Medicare Current Beneficiary Survey.
13 Massachusetts, Wisconsin, and Minnesota standardized their plans prior to federal regulation
and have continued their own offerings. The Medicare Prescription Drug, Improvement, and Modernization
Act of 2003 (MMA) introduced plans K and L and eliminated the sale of Medigap plans
with drug benefits (H, I and J).

CHAPTER 3. MEDIGAP 124
shows enrollment by plan letter and the plan characteristics in detail. Plans C and
F, by far the most popular plans, are chosen by over 60 percent of the Medigap
beneficiaries. Both of these plans cover the hospital and physician deductibles, in
addition to the basic benefits common to all lettered plans. In this paper, we estimate
the marginal effect of the average Medigap plan, rather than the effect of particular
plan characteristics. We believe the marginal effect of Medigap captures a meaningful
concept in this context because the vast majority of individuals are on Medigap plans
face the same marginal change in prices. The common characteristics of the Medigap
plans are ex ante the most likely drivers of moral hazard according to previous studies.
In particular, it is thought that physician expenditures is one of the most flexible
margins of health care spending, and all plans include Part B physician coinsurance
coverage. 14 Part B coinsurance coverage it thought to generate substantial moral
hazard because it dramatically reduces the marginal prices for doctor visits for almost
everyone with physician expenditures. 15
3.3 Data
We have two primary sources of data. We use the Medicare Current Beneficiary
Survey (MCBS) for 2000 through 2005. This data is a short panel of Medicare
beneficiaries that includes cost data by payer, information about Medigap coverage,
demographic information and relatively detailed health status variables. In addition
to the MCBS data, we use Medigap premium data from Weiss Ratings. We obtained
premium data for 2000 and 2003. This data includes premiums charged during the
open enrollment period by company, state, and plan letter. We use these two years
of premiums to extrapolate premiums from 2000 to 2005. 16
We match MCBS beneficiaries with the prices they faced at age 65. This means
14 See ?.
15 Because elderly individuals almost all have physician expenditures of 135 dollars annually (the
approximate value of the Part B deductible), the marginal price they face is the 20 percent coinsurance
rate for physician visits beyond this deductible. Compare this to the Part B deductible
coverage, which is not available in all plans, but probably doesnt lead to much moral hazard.
16 In particular, we assume a linear trend in premiums between 2000 and 2003. Using this trend,
we project forward until 2005.

CHAPTER 3. MEDIGAP 128
We use bootstrap methods to calculate the standard errors, looping over the entire
set of equations for consistency.
3.5 Identifying Variation
3.5.1 The Instrument
The basic concept
The idea of the empirical model is to use plausibly exogenous variation in Medigap
premiums (equation 3.3) to shift individuals in and out of Medigap coverage (equation
3.2) and thereby identify the impact of Medigap on total medical spending (equation
3.1).
To isolate variation in premiums that is conditionally random, we leverage discon-
tinuities in Medigap premiums at state boundaries. As we discuss below, Medicare
costs, and thus the costs of supplemental Medigap insurance, exhibit considerable
within-state variation due to factors ranging from household incomes to local physi-
cian practice styles to the supply of medical resources (????). Yet despite this local
variation, within-state variation in Medigap premiums is highly limited (?). 17 This
means that otherwise identical individuals in otherwise identical localities can face
sharply different Medigap premiums solely due to Medicare costs in other parts of
their state.
We put this idea into practice by using state-level costs as an instrument while
simultaneously controlling for local costs in all the equations of our model. Controlling
for local costs is of course essential, as factors that influence costs in a local area almost
surely influence costs and enrollment decisions in our sample, even conditional on the
set of individual controls we include.
17 Although firms are allowed to vary premiums at the zip code level, ? find there is very little
within-state variation in the Medigap premiums for a given plan offered by given insurance company.
Maestas et al. cite state-level reporting requirements and regulations as a potential explanation.
Because plans must, for example, meet loss ratio requirements at state level, varying premiums
more locally may be administratively burdensome. We discuss a robustness check to this issue in
Section 3.6.

CHAPTER 3. MEDIGAP 129
Data and summary statistics
In particular, we implement our instrumental variables strategy by using fee-for-
service (FFS) Medicare cost data aggregated by the Dartmouth Atlas to the state
and local level. 18 The advantage of the Dartmouth Atlas data is the methodology used
to define local geographic areas. The Atlas aggregates cost data at two levels: the
Hospital Service Area (HSA) and Hospital Referral Region (HRR). The HSA is the
smaller unit of aggregation, defined by the set of zip codes whose residents receive most
of their hospitalizations from hospitals in that area. These areas are approximately
the size of a county: there are 3,446 HSA and 3,140 counties in the United States. To
capture larger areas, the Dartmouth Atlas looks at where individuals are referred to
for major cardiovascular surgery and neurosurgery. These are typically big, superstar
hospitals, usual in big cities. HRR are more than 10 times as large as HSA: there are
306 HRR in the country. To be conservative, we use larger HRR as the local area in
our analysis.
Table 3.4 summarizes HRR level data on per capita FFS Medicare costs, tabulated
by state. For expositional purposes, we show data from 2003, a year in the middle
of our sample period. As is well known, there are vast differences in per capita costs
between regions: costs range from $4,520 per capita in Salem, OR to $11,588 per
capita in the Bronx, NY. A fact that is less well known, but vital to the power of
our instrument, is that there are substantial differences in HRR-level costs within
states as well. Per capita costs vary by more than $5,000 within NY, FL and TX,
and by more than $1,000 within 25 states that are home to most (75 percent) of the
individuals in our sample.
Instrument power
Figure 3.1 provides a concrete example of the power of our instrument. Panel A shows
a map of per capita FFS Medicare costs in the states of New York and Vermont by
HRR. Panel B shows Medigap premiums in the same two states. A comparison
18 The Dartmouth Atlas data is constructed from the Continuous Medicare History Sample
(CMHS), a 5 percent sample of the billing records Medicare beneficiaries collected by the Centers
for Medicare and Medicaid Services (CMS).

CHAPTER 3. MEDIGAP 130
of costs and premiums in upstate New York and Vermont is illuminating. While
these two regions have almost identical per capita Medicare costs (no more than
$5,864 in upstate NY versus $5,858 in VT), Medigap premiums in these regions are
significant different ($1,677 versus $1,320). The bottom of the map shows why. New
York state of course has New York City to its south, a region with substantially
higher Medicare costs than the northern part of the state (maximums of $11,588 in
the south versus $5,864 upstate). It is the high cost metropolitan south, combined
with the limited within-state variation in premiums, which inflates Medigap prices in
upstate New York without impacting premiums in Vermont, creating an exogenous
source of premium variation.
To examine the statistical power of the instrument, we calculate the incremental
power – or partial R-squared – from adding state costs to a regression of premiums
on HRR-level costs. That is, we calculate the additional power from state costs
when already controlling for costs in the local region. As shown in Appendix Table
3.10, a regression of log premiums on log HRR costs has an R-squared of 0.19 and a
coefficient on HRR costs of 0.35. Adding state costs reduces the coefficient on HRR
costs to zero and increases the R-squared from 0.19 to 0.50, implying that 62 percent
(=0.31/0.50) of the explanatory power of Medicare costs comes from outside local
HRRs. The F-statistic on state costs is highly significant.
3.5.2 Potential Concerns
The exclusion restriction
A key assumption underpinning our empirical strategy is that, conditional on HRR-
level costs, state-level Medicare costs only impact utilization through premiums and
Medicare choice. One potential concern with this assumption is that because indi-
viduals travel outside their local HRRs to receive health care, the local measure of
utilization is an insufficient control. This concern is misplaced, however, since local
HRR costs account for travel, measuring total utilization based on the location of
residence regardless of where care is received. Thus, if individuals in one area tend
to travel to another place with higher (or lower) costs for medical services, utilization

CHAPTER 3. MEDIGAP 131
levels in this area will reflect this propensity to travel.
A more subtle concern is that the marginal Medigap purchaser may be more
(or less) likely to travel than the average Medicare beneficiary. If this is the case,
then controlling for average local health care utilization is insufficient. Basically, the
concern is that although most people in Rural Town, USA don’t travel to Big City,
USA for care, the marginal Medigap purchaser might be the type of person who does.
A solution to this problem is to define markets broadly enough so that the closest Big
City, USA is included in the market. We believe this is exactly what we do by using
HRRs to define local markets as these regions are constructed explicitly to include
the major hospital where people travel for sophisticated treatment.
Therefore, given a measure of costs which accounts for travel and the conservative
definition of local areas, we think that the exclusion restriction is unlikely to be
violated. Nevertheless, if out-of-HRR costs did somehow directly impact costs, it
would downwardly bias the coefficient on Medgap, working against the direction of
our main finding. 19
Medicare Advantage
A final topic of discussion is our proxy for the price and availability of Medicare
Advantage coverage: a second order polynomial in the county-year specific MA pene-
tration rate. Controlling for the determinants of MA enrollment is important for the
explanatory power of the choice equation as Medigap and MA are typically viewed as
substitutes. For similar reasons, it is important to include MA enrollees in the sample
of analysis as MA coverage is the relevant alternative coverage to consider for changes
in the Medigap market. Yet for valid estimation, the variables used to predict MA
choice must satisfy the exclusion restriction of not directly impacting medical costs.
Below we make the case for their exclusion.
The main argument for the exogeneity of county-year specific MA penetration
rates is that much of the variation in this variable is driven by high-level federal
19 In this case, higher out-of-HRR costs would simultaneously shift individuals out of coverage
through the instrument and increase costs directly, creating a negative correlation between coverage
and unobserved costs and biasing downwards the coefficient on Medigap.

CHAPTER 3. MEDIGAP 132
policy that could not plausibly be related to the unobserved component in the cost
equation. MA penetration decreased over the first part of our sample as policies
enacted in the 1997 Balanced Budget Act to reduce Medicare payments indirectly
decreased MA payments. MA penetration rates were then buoyed up by increased
payments written into the 2003 Medicare Modernization Act (?). In addition, the
goal of MA has shifted over time from promoting efficiency to including broader goals
such as regional equity meaning there has been great variation over time in federal
policy toward regional MA insurers. Thus most of the changes over time were driven
by federal policy that could not be related to the unobservable.
We also statistically test for the endogeneity of the county-year specific MA pen-
etration variables. Recall that our Meidgap choice equation has two sets of excluded
variables: Medigap premiums and the MA penetration variables. Based on the as-
sumption that estimates from a specification using only Medigap premiums are con-
sistent but inefficient and estimates from a specification with both Medigap premiums
and MA penetration are efficient, a Durbin-Wu-Hausman test squarely fails to reject
the exogeneity of MA penetration variables. In our preferred specification with the
full set of controls, the test returns a Chi-squared value of 0.8, a value vastly below
the 5 percent critical value of 37.6. Tests run on other specifications yield similar
results.
3.6 Results and Discussion
3.6.1 Premiums and Medigap Choice
Below we present baseline estimates of the empirical model, starting with the first
stage premium equation, followed by the multinomial logit second stage Medigap
choice equations, and finally the third stage cost equation. Each of these equations is
estimated with three different sets of control variables. A sparse specification includ-
ing only key variables and year fixed effects; a specification that includes the demo-
graphic controls; and a full specification that, along with the demographic variables,

CHAPTER 3. MEDIGAP 133
includes controls for health status such as self-reported health, indicators of medi-
cal diagnoses and treatment, and widely-used measures of functional status known as
Activities of Daily Living (ADL) and Instrumental Activities of Daily Living (IADL).
Table 3.5 presents the first stage estimates of Medigap premiums on state- and
HRR-level costs in logs. Standard errors are clustered at the year-by-state level
because premiums vary at this level. For the purposes of our policy counterfactual
analysis, the most important parameter in this table is the elasticity of premiums
with respect to state-level costs. This coefficient is 0.50 to 0.60 across specifications
and precisely estimated (standard error of less than 0.07). As discussed in Section 3.5,
most of the identifying variation comes from state-level FFS Medicare costs. With
HRR costs and the other controls, the F-statistic on state-level costs is well above
the critical level.
Marginal effects for the probability of Medigap choice from the multinomial logit
choice equation are presented in Table 3.6. Recall that residuals from the premium
equation are included as a control variable to overcome the potential endogeneity of
premiums. The premium semi-elasticity of demand is estimated at a stable 0.38 to
0.48 across specifications. (The premium elasticity estimates from the linear proba-
bility model, shown in Appendix Table 3.11, are similar, ranging from 0.31 to 0.42.)
With demographic controls, the point estimates are significantly negative at more
than the 5 percent level. In the preferred logit specification with the full set of con-
trols (specification 3), we can reject any premium elasticity outside of 0.13 to 0.82
with a 95 percent confidence interval.
Males are 3 percent less likely to purchase Medigap coverage. Medigap partici-
pation is 20 percent lower for blacks, increasing in education, but flat with respect
to income. Although we do not show it here, it should be noted that there is strong
evidence that Medigap choice is increasing in factors the proxy for good health as
well as underlying cognitive ability (?).

CHAPTER 3. MEDIGAP 134
3.6.2 Causal Impact on Costs
Table 3.7 presents estimates of the third stage cost model. Columns 1 through 3
show simple OLS estimates of log total costs on a Medigap indicator and controls;
columns 4 through 6 show estimates with predicted values from the multinomial logit
model for Medigap choice used as instruments. The models are estimated with block
bootstrap standard errors clustered by individual for consistency (?). In Appendix
Table 3.12, we show 3SLS estimates of the cost equation using the linear probability
model for Medigap choice.
The OLS estimates (columns 1 to 3) suggest that Medigap coverage increases total
medical utilization by about 40 percent. With the full set of controls, a 95 percent
confidence interval allows us to reject any effect outside of 16 to 65 percent. The
instrumental variables estimates are about 40 percent larger than the OLS estimate
across the different sets of controls. With demographic controls (columns 5 and 6), the
point estimates suggest that Medigap coverage causally increases medical utilization
by 57 to 61 percent for individuals local to the instrument. (The 3SLS estimates,
shown in Appendix Table 3.12 indicate effects of 59 to 63 percent.) In the logit
specification with the full set of controls, a 95 percent confidence interval allows us
to reject any effect outside of 30 to 84 percent.
We find it sensible that the instrumental variables estimates are larger than those
estimated by OLS (57 versus 41 percent). As we have discussed elsewhere, Fang et
al. (2009) find strong evidence of advantageous selection on a set of health variables
that increase the R-squared of their baseline model from 7 to 21 percent. Yet even
these controls leave 79 percent of the variation unexplained. If there is some selection
on information contained in this remaining variation, and this selection continues to
be advantageous, then we would expect an approach that overcomes this selection
problem to yield larger estimates than a simple OLS approach.
A second reason for why our baseline estimates may be larger than the OLS
estimates is because we are measuring the moral hazard effect for individuals on the
margin of selecting into Medigap. The OLS estimates are best interpreted as an
average moral hazard effect that may be biased due to selection. In contrast, our
baseline logit estimates capture the moral hazard effect for individuals on the margin

CHAPTER 3. MEDIGAP 135
of selecting into Medigap at current prices. These individuals are probably more
price sensitive when it comes to premiums. If they are more sensitive to the price of
medical care induced by Medigap coverage as well, then the moral hazard effect for
these marginal individuals will be larger than the average Medigap policyholders.
We think it is important to emphasize two things about the interpretation of our
baseline estimates related to this point. First, the moral hazard of individuals on the
margin is the precisely the parameter of interested for policy analysis of a Medigap
premium tax or subsidy. If the government levies, for example, a tax on Medigap
premiums, then the individuals that select out of Medigap as a result of the tax will
reduce their total medical care utilization by our baseline estimate. (In Section 3.7,
we demonstrate the effect of tax and subsidy policies using our estimates.) Second,
one can view our moral hazard estimate as an upper bound on the moral hazard of
all Medigap beneficiaries following the logic that marginal beneficiaries have greater
moral hazard than infra-marginal beneficiaries.
To compare our effect to other estimates in the literature, we follow standard
practices (??) and convert our point estimates to arc elasticities of expenditure (?).
Under the approximation that Medigap decreases the “price” of a unit of healthcare
from 0.2 to 0.0, the preferred point estimate of 0.57 implies an arc elasticity of 1.13.
We think there are good reasons for why our estimate is higher than the “gold stan-
dard” estimate of 0.22 from the RAND Health Insurance Experiment (HIE) of the
1970s. The first is that our estimate may be local to particularly price sensitive indi-
viduals, as discussed above. A second reason is that elderly people may be more price
sensitive than the non-elderly population. While the RAND HIE only included people
under age 65, the Medicare population we study includes those over age 65. Other
papers have found results consistent with the idea that the elderly have higher price
elasticities of medical utilization (??). A third reason is that innovation in medicine
over the last 40 years could have affected price elasticities, perhaps by creating more
treatment options for a given diagnosis that individuals can substitute among.

CHAPTER 3. MEDIGAP 136
3.6.3 Robustness and External Validity
Our first robustness check addresses the log specification for medical costs. It is well
known that the skewed distribution of outcomes can create difficulties in estimating
models with medical costs as a dependent variable. To show that the log specification
we use is not driving the results, Panel A of Table 3.8 shows point estimates from the
three-stage model where costs in the cost equation are denoted in levels. For com-
parison, we divide these estimates by the average costs in our sample and display the
results next to the estimated parameters from the model in logs. With demographic
controls, the implied percentage changes in medical spending are similar to the log
estimates (0.68 versus 0.57) although less precisely estimated.
As discussed earlier, our analysis is restricted to the sample of individuals who
turned 65 no sooner than the year 2000, as price data is not available from before
this year. This, of course, means that the estimates should not be directly ascribed
to the entire Medigap population because of potential age or cohort heterogeneity
in the main effect. However, we can shed some light on the external validity of the
main results with simple OLS regressions. Panel B of Table 3.8 shows OLS estimates
of log medical costs on a Medigap indicator for the entire sample of individuals in
the data and the subset we analyze. With demographic and health controls, the OLS
estimates from the samples are very similar (0.34 versus 0.37). However, to the extent
that health controls may be partly determined by Medigap choice, the specification
with only demographic covariates may be more appropriate. The lower OLS estimate
for the full sample in this specification (0.11 versus 0.35) suggests some caution when
extrapolating the estimates of out sample.
Finally, as discussed Section 3.5, there is in fact a small degree of within-state
variation in Medigap premiums in a small number of states in the sample (Maestas et
al., 2009). To examine the robustness of the main results to this variation, we restrict
our sample to states where, according to Maestas et al., the within-state coefficient of
variation of Medigap premiums is less than 0.10 for 75 percent of firms issuing plans. 20
Panel C of Table 3.8 shows full model and OLS estimates of the coefficient on Medigap
20 The states we drop are CA, FL, IL, LA, MI, NV, NY, PA, TX, and VA.

CHAPTER 3. MEDIGAP 137
from the baseline and restricted sample. The full model estimates are slightly lower
(0.42 versus 0.57) although the ratio of full model to OLS estimates is very similar
across samples. A possible reason for the lower full model and OLS estimates in the
restricted sample is that the excluded states, which include states notorious for the
overuse of medical services (e.g., FL, NY, TX), are states where the moral hazard from
Medigap may be the largest. The premium semi-elasticity of demand estimates, which
are potentially the most susceptible to bias from within-state premium variation, are
almost identical across the restricted and baseline samples.
3.7 Policy Counterfactuals
Having shown that private Medigap insurance increases medical utilization and thus
Medicare costs by as much as 59 percent, it seems natural to consider corrective tax-
ation. The idea of taxing Medigap premiums is not new. In their 2008 chapter of
budget options for health care, for example, the Congressional Budget Office (CBO)
considered a 5 percent excise tax on Medigap premiums as a mechanism to reduce
the federal deficit. As CBO pointed out, an excise tax on Medigap premiums would
reduce the deficit through two main channels: it would raise revenue from individuals
who maintained their Medigap coverage while at the same time reducing Medicare
costs for people who gave up their supplemental coverage. Yet the CBOs calculations
of the taxs impact were likely based on a downwardly biased estimate of the effect of
supplemental insurance, meaning the CBO estimates of the effect of a tax underes-
timated the budgetary effect. 21 In contrast, the parameters estimated here leverage
plausibly exogenous variation in Medigap premiums, accounting not only for selec-
tion on unobservable individual characteristics but also for potentially heterogeneous
effects on individuals marginal to a change in premiums.
We use our estimated parameters to produce back-of-the-envelope budgetary es-
timates in the follow manner: First, the counterfactual Medigap market share is
21 In another section (Option 82) of their budget options volume, CBO cites studies claiming
that Medigap coverage increases utilization by 25 percent versus individuals without supplementary
coverage.

CHAPTER 3. MEDIGAP 138
calculated using the estimated premium semi-elasticity of 0.475 (Table 3.6, specifi-
cation 3) and assuming the tax is completely passed through. 22 We estimate tax
revenue by applying the tax to the resulting share of Medigap enrollees. Cost savings
are calculated by reducing the medical costs of individuals who drop Medigap by 37.0
percent, the factor implied by the estimated parameter of 0.588 from specification 3
of Table 3.7 (i.e., 0.370 =1 - 1/(1+0.588)). Medicare cost savings are calculated as
80 percent of this number, consistent with Medicare paying 80 percent on the margin
for care. The total budgetary impact is simply the sum of the revenue and Medicare
cost effects.
Table 3.9 presents these revenue and cost savings estimates for the standard sample
of individuals with Medigap, Medicare Advantage, and no supplemental coverage.
In this sample, baseline Medigap enrollment is 35.8 percent, and Medical costs for
Medigap enrollees average $9,458 per annum. The results are presented in dollars per
beneficiary and as a percentage of Medicare costs of $5,528 on average for Medigap
enrollees. A 20 percent tax would reduce coverage by 9.5 percentage points, raising
1.3 percent of total costs in tax revenue and 4.8 percent of total costs in reduced
utilization for total budgetary savings of 6.2 percent. A 40 and 60 percent tax would
save 11.4 and 15.6 percent of costs respectively. And a 76 percent tax, which would
increase Medigap premiums by $1,076, or about 50.5 percent of the implied $2,802
(=0.37 x $9,458 x 0.80) externality, would reduce the Medigap market share to zero,
savings 18.1 percent of total costs entirely though decreased medical utilization.
Although it probably goes without saying, we remind the reader that the pa-
rameters are estimated using local variation in premiums and the projected effects of
larger taxes and subsidies should be viewed with appropriate caution. It is also worth
mentioning that under current law, a tax that reduces FFS Medicare costs would re-
duce minimum Medicare Advantage payments. This is not taken into account in our
simple tax simulations, and would likely magnify the cost savings we project.
Last but not least, we want to call attention to the fact that a uniform tax on
Medigap premiums is a blunt instrument for controlling Medicare costs. Chandra et
22 In the case of incomplete pass-though, one can think of the counterfactuals as simulating the
impact of tax that would give rise to the full pass-though increase in premiums.

CHAPTER 3. MEDIGAP 139
al. (2007) have shown that targeted subsidies that reduce cost-sharing for chronically
ill Medicare beneficiaries can generate cost savings or “offsets” by encouraging treat-
ment adherence and reducing unnecessary hospitalizations. A nuanced tax policy,
which takes such effects into account, might be able to generate both increased costs
savings and improved health.
3.8 Conclusion
In this paper, we have estimated the fiscal externality imposed by private Medigap
supplemental insurance on public Medicare costs. We find that the moral hazard
induced by Medigap is substantial, with marginal enrollees using 57 percent more
services due to Medigap coverage. Policy counterfactuals reveal that a uniform tax
that raises premiums 20 percent could generate combined revenue and savings of 6.2 of
baseline costs; a Pigovian tax that fully accounts for the fiscal externality could yield
savings of 18.1 percent. We are interested in the effects of more nuanced tax policies
that account for “offsets” from treatment adherence or other positive externalities. In
light of the findings of significant moral hazard from Medigap, we view the study of
supplemental insurance around the world as an important topic for future research.

CHAPTER 3. MEDIGAP 140
(10000,11000]
(9000,10000]
(8000,9000]
(7000,8000]
(6000,7000]
[5000,6000]
Figure 3.1: Example of Identifying Variation
(10000,11000]
(9000,10000]
(8000,9000]
(7000,8000]
(6000,7000]
[5000,6000] and VT per capita Medicare costs
(10000,11000] NY (9000,10000] and (8000,9000] VT (7000,8000] per (6000,7000] capita [5000,6000]
Medicare NY costs
NY and VT per capita Medicare costs
(1600,1700]
(1500,1600]
(1400,1500]
[1300,1400] NY and VT premiums
(1600,1700]
NY and VT premiums
(1600,1700]
(1500,1600]
(1400,1500]
[1300,1400]
NY and VT premiums

CHAPTER 3. MEDIGAP 141
Table 3.1: Medicare Cost-Sharing
Part A: Hospital Expenditures Part B: Physician Expenditures
Per day copayment
SNF
Per day copayment
Year Deductible Days 61-90 Days 91-150* Deductible Coinsurance Days 21-100
2000 $776 $194 $388 $100 20% $97
2005 $912 $228 $456 $110 20% $114
*Medicare only pays for part A hosptializations in excess of 90 days through the drawdown of lifetime reserve days, of
which beneficiaries have 60 over their lifetime.

CHAPTER 3. MEDIGAP 142
Table 3.2: Summary Statistics by Supplemental Insurance Type
ESI Medicaid Medigap MA None Total
Enrollment
N 16,477 6,594 11,830 13,045 4,475 52,420
Percent
Demographics
31% 13% 23% 25% 9% 100%
Male 43.1% 27.5% 33.8% 39.4% 42.0% 38.0%
(0.4%) (0.5%) (0.4%) (0.4%) (0.7%) (0.2%)
Age 75.0 77.9 76.4 74.7 77.1 75.8
(0.1) (0.1) (0.1) (0.1) (0.1) (0.0)
White 91.8% 71.1% 93.9% 85.5% 81.3% 87.2%
(0.2%) (0.5%) (0.2%) (0.3%) (0.6%) (0.1%)
College or more 22.6% 3.8% 15.1% 15.1% 10.5% 15.7%
(0.3%) (0.2%) (0.3%) (0.3%) (0.4%) (0.2%)
Income 37,818 10,839 30,215 30,691 20,887 29,490
Utilization
(371) (309) (439) (641) (431) (227)
Total 11,008 25,456 10,152 8,062 18,169 12,511
(147) (366) (167) (141) (409) (95)
Medicare* 5,971 11,871 6,214 2,111 8,107 5,990
(106) (258) (124) (79) (260) (64)
ESI 2,907 113 14 182 132 988
(37) (13) (2) (12) (17) (13)
Medicaid 1 9,359 0 405 8 1,279
(1) (176) (0) (36) (3) (29)
Medigap 103 149 1,472 73 89 409
(9) (11) (40) (5) (11) (10)
Out-of-pocket 1,669 3,164 1,967 1,691 8,425 2,506
(40) (72) (28) (43) (227) (29)
Other 357 800 485 3,600 1,407 1,338
(16) (51) (20) (85) (77) (24)
Source: 2000-2005 MCBS
Notes: ESI is employer sponsored supplmental insurance. MA is Medicare Advantage or Medicare+Choice as it
was called in the begining of the sample period. Sample excludes individuals with positive VA expenditure and
individuals residing in states without Medigap coverage (MA, MN, WI). Payments from unexpected sources (e.g,
Medigaid payments for ESI enrollees) due to switches in supplemental insurance type during the year. Standard
errors in parentheses.
*Does not account for Medicare capation payments to private plans

CHAPTER 3. MEDIGAP 143
Table 3.3: Medigap Enrollment and Plan Characteristics by Letter
Medigap Letter
A B C D E F* G H I J*
Enrollment
N 238 291 894 293 133 1,757 154 117 112 348
Percent 6% 7% 21% 7% 3% 41% 4% 3% 3% 8%
Plan Characteristics
X X X X X X X X X X
Basic benefits (Part A coninsurance,
Part B coinsurance, blood, additional
lifetime hospital days)
Skilled nursing facility coinsurance X X X X X X X X X
Part A deductible X X X X X X X X X X
Part B deductible X X X
Part B excess charges X 80%
Foreign Travel Emergecy X X X X X X X X X
Home health care X X X X
Prescription drugs X X X
Preventive medical care X X
Source: 2000-2005 MCBS
Notes: Sample excludes individuals with positive VA expenditure and individuals residing in states without Medigap coverage (MA, MN, WI).
*Plans F and J have a high-deductible option that requires beneficiaries to pay $1,580 before receiving Medigap coverage. These plans are rarely offered an
have only 14 enrollees in the sample.

CHAPTER 3. MEDIGAP 144
Table 3.4: HRR Level-Medicare Costs Tabulated By State
State N Mean Std. Dev. Min. Max.
NY 10 7,453 2,389 5,235 11,588
NV 2 7,283 1,730 6,059 8,506
FL 18 7,271 1,209 6,091 11,422
MS 6 6,796 1,176 5,310 8,473
CA 24 7,094 1,080 5,432 10,290
TX 22 7,258 1,078 5,472 10,626
IL 13 6,817 1,061 5,097 8,496
MI 15 7,033 1,059 5,683 8,636
PA 14 6,707 1,044 5,017 8,845
LA 10 7,922 772 7,148 9,020
UT 3 5,990 747 5,335 6,804
IN 9 6,198 728 5,123 7,458
NJ 7 8,626 713 7,216 9,426
AR 5 6,162 691 4,983 6,709
MA 3 8,123 621 7,431 8,631
WI 8 5,355 620 4,587 6,304
GA 7 6,160 534 5,068 6,852
CO 7 6,112 516 5,211 6,767
NH 2 5,785 487 5,440 6,129
OR 5 5,049 466 4,520 5,547
NC 9 6,165 456 5,563 6,955
OH 10 6,932 455 6,385 7,674
MO 6 6,106 445 5,541 6,634
IA 8 5,063 420 4,690 5,814
SC 5 6,430 404 6,043 7,065
VA 8 5,770 374 5,275 6,312
CT 3 7,864 369 7,527 8,258
ND 4 5,315 356 4,902 5,677
WA 6 5,716 352 5,170 6,108
KY 5 6,439 331 6,088 6,807
TN 7 6,641 307 6,374 7,026
MN 5 5,544 298 5,193 5,876
AZ 4 6,410 280 6,036 6,714
AL 6 6,710 278 6,390 6,994
ID 2 5,114 269 4,924 5,304
MD 3 7,734 261 7,568 8,034
ME 2 5,990 260 5,806 6,173
OK 3 6,932 193 6,770 7,146
SD 2 5,243 153 5,134 5,351
KS 2 6,294 125 6,205 6,382
NE 2 5,721 87 5,659 5,782
MT 3 5,383 69 5,304 5,430
WV 3 6,475 58 6,433 6,541
AK 1 6,719 6,719 6,719
DC 1 6,961 6,961 6,961
DE 1 6,850 6,850 6,850
HI 1 4,778 4,778 4,778
NM 1 5,254 5,254 5,254
RI 1 7,187 7,187 7,187
VT 1 5,858 5,858 5,858
WY 1 5,989 5,989 5,989
Total 306 6,635 1,163 4,520 11,588
Source: Dartmouth Atlas of Health Care
Notes: Each observation is per capita FFS Medicare costs in a given HRR in 2003. The
data is sorted by standard deviation.

CHAPTER 3. MEDIGAP 145
Table 3.5: First Stage: OLS Estimates of Premiums on State and HRR Costs
(1)
Dependent Variable: Log Premium
(2) (3)
Est. SE Est. SE Est. SE
Log state cost 0.5020 (0.0583) 0.5990 (0.0686) 0.5950 (0.0678)
MA penetration rate (%) 0.0004 (0.0018) 0.0000 (0.0018) 0.0001 (0.0019)
MA penetration rate squared (%) 0.0000 (0.0000) 0.0000 (0.0000) 0.0000 (0.0000)
Log HRR cost -0.0005 (0.0266) -0.0020 (0.0252) 0.0003 (0.0248)
Male -0.0052 (0.0060) 0.0045 (0.0084)
Age-65 0.0259 (0.0115) 0.0231 (0.0112)
(Age-65)^2 0.0037 (0.0062) 0.0053 (0.0061)
(Age-65)^3 -0.0007 (0.0009) -0.0010 (0.0008)
Disabled eligibility
Race group
-0.0016 (0.0058) -0.0004 (0.0059)
Asian 0.0149 (0.0141) 0.0121 (0.0134)
Black 0.0097 (0.0118) 0.0114 (0.0116)
Other
Education group
-0.0238 (0.0154) -0.0249 (0.0148)
High school -0.0243 (0.0051) -0.0243 (0.0050)
Less than college -0.0052 (0.0065) -0.0048 (0.0063)
College or more -0.0094 (0.0073) -0.0113 (0.0070)
Log income 0.0007 (0.0028) 0.0011 (0.0028)
Married -0.0104 (0.0050) -0.0102 (0.0053)
Working 0.0026 (0.0054) 0.0015 (0.0055)
Served in armed forces
Self-reported health
0.0136 (0.0064) 0.0122 (0.0067)
Excellent -0.0042 (0.0119)
Very good -0.0134 (0.0110)
Good -0.0137 (0.0095)
Fair -0.0209 (0.0096)
Diagnosis indicators Yes
Treatment indicators Yes
IADL indicators Yes
Year FE Yes Yes Yes
R-squared 25.1% 31.2% 33.2%
N 3,556 3,556 3,556
F-statistic on log state cost 74.01 76.15 76.88
Prob > F 0.00 0.00 0.00
Source: 2000-2005 MCBS; Weiss Ratings; Dartmouth Atlas
Notes: Parameter estimates from OLS regressions of premiums on state- and HRR-level Medicare costs. The sample is restricted to
individuals with Medigap, HMO or no supplemental coverage who turned 65 in 2000 or later. The omitted race group is white,
education group is less than high school, and health status is poor. Standard errors clustered at the state-by-year level in parentheses.

CHAPTER 3. MEDIGAP 146
Table 3.6: Second Stage: Marginal Effects for Medigap Choice from Multinomial
Logit Model
(1)
Marginal Effects for Medigap Choice
(2) (3)
Mfx. SE Mfx. SE Mfx. SE
Log premium -0.375 (0.214) -0.477 (0.220) -0.473 (0.221)
Log premium residuals 0.398 (0.227) 0.532 (0.242) 0.531 (0.242)
County MA penetration (%) -0.026 (0.002) -0.029 (0.002) -0.029 (0.002)
County MA penetration squared (%) 0.000 (0.000) 0.000 (0.000) 0.000 (0.000)
Log HRR cost -0.049 (0.079) 0.002 (0.090) -0.009 (0.092)
Male -0.028 (0.028) -0.029 (0.028)
Age-65 0.061 (0.043) 0.065 (0.043)
(Age-65)^2 -0.037 (0.020) -0.040 (0.020)
(Age-65)^3 0.005 (0.003) 0.006 (0.003)
Disabled eligibility
Race group
-0.010 (0.028) -0.013 (0.028)
Asian 0.000 (0.068) 0.000 (0.069)
Black -0.263 (0.026) -0.265 (0.026)
Other
Education group
-0.025 (0.058) -0.028 (0.059)
High school 0.038 (0.030) 0.041 (0.031)
Less than college 0.126 (0.036) 0.130 (0.036)
College or more 0.185 (0.039) 0.188 (0.042)
Log income -0.018 (0.013) -0.017 (0.014)
Married 0.004 (0.024) 0.004 (0.024)
Working 0.014 (0.029) 0.020 (0.029)
Served in armed forces
Self-reported health
-0.033 (0.037) -0.033 (0.038)
Excellent -0.094 (0.051)
Very good -0.113 (0.052)
Good -0.097 (0.049)
Fair
Diagnosis indicators
Treatment indicators
IADL indicators
-0.097 (0.050)
Year FE Yes Yes
Log pseudolikelihood/R-squared -15,205,864 -13,914,167 -13,853,769
N 3,556 3,556 3,556
Source: 2000-2005 MCBS; Weiss Ratings; Dartmouth Atlas, CMS State-County Penetration Files
Notes: Marginal effects for Medigap choice from multinomial logit model with control function for Medigap premiums.
Marginal effects calculated at the sample means. For binary variables the predicted change in probability for a 0 to 1 change
is shown. The sample is restricted to individuals with Medigap, HMO or no supplemental coverage who turned 65 in 2000 or
later. The omitted race group is white, education group is less than high school, and health status is poor. Bootstrap
standard errors looped over the premium and choice equations and clustered by individual in parentheses.

CHAPTER 3. MEDIGAP 147
Table 3.7: Third Stage: OLS and Full Model Estimates of Medical Costs on Medigap
Indicator
Dependent Variable: Log Medical Costs
OLS Full Model
(1) (2) (3) (4) (5) (6)
Medigap 0.397 0.401 0.407 0.515 0.614 0.567
(0.054) (0.053) (0.049) (0.158) (0.146) (0.138)
Log HRR cost -0.071 0.233 0.128 -0.018 0.325 0.192
(0.132) (0.134) (0.124) (0.173) (0.171) (0.164)
Male -0.055 -0.071 -0.043 -0.049
(0.061) (0.079) (0.074) (0.070)
Age-65 0.817 0.904 0.810 0.890
(0.125) (0.114) (0.126) (0.118)
(Age-65)^2 -0.241 -0.299 -0.235 -0.278
(0.072) (0.065) (0.065) (0.061)
(Age-65)^3 0.023 0.030 0.021 0.027
(0.010) (0.009) (0.009) (0.008)
Disabled eligibility 0.419 0.333 0.422 0.346
Race group
(0.073) (0.067) (0.078) (0.074)
Asian -0.314 -0.206 -0.292 -0.289
(0.143) (0.136) (0.176) (0.175)
Black -0.098 -0.171 -0.050 -0.148
(0.090) (0.083) (0.113) (0.107)
Other 0.034 -0.079 0.052 -0.057
Education group
(0.141) (0.131) (0.168) (0.155)
High school 0.114 0.199 0.098 0.216
(0.069) (0.064) (0.086) (0.082)
Less than college 0.118 0.236 0.096 0.292
(0.073) (0.069) (0.087) (0.083)
College or more 0.184 0.347 0.156 0.404
(0.084) (0.080) (0.103) (0.098)
Log income -0.012 0.049 -0.006 0.037
(0.029) (0.026) (0.033) (0.030)
Married -0.081 0.011 -0.090 -0.058
(0.055) (0.052) (0.067) (0.062)
Working -0.429 -0.168 -0.434 -0.250
(0.060) (0.056) (0.066) (0.066)
Served in armed forces -0.107 -0.119 -0.101 -0.091
Self-reported health
(0.080) (0.073) (0.098) (0.091)
Excellent -0.616 -1.920
(0.140) (0.138)
Very good -0.565 -1.769
(0.136) (0.133)
Good -0.407 -1.392
(0.129) (0.130)
Fair -0.211 -0.890
(0.130) (0.140)
Diagnosis indicators Yes Yes
Treatment indicators Yes Yes
IADL indicators Yes Yes
Year FE Yes Yes Yes Yes Yes Yes
R-squared 3.6% 11.1% 29.4% 3.4% 10.7% 19.9%
N 3,415 3,415 3,415 3,415 3,415 3,415
Source: 2000-2005 MCBS; Weiss Ratings
Notes: Parameter estimates from cost equation. Specifications 1-3 show OLS estimates; 4-6 show estimates where
Medigap choice is instrumed for with the predicted values from the logit cost equation. See text for details. Sample of
individuals with HMO, Medigap or no supplemental coverage who turned 65 no sooner than 2000 excluding those with
positive VA expenditure. Omitted race group is white, education group is less than high school, and health status is poor.
Bootstrap standard errors looped over the premium and choice equations and clustered by individual in parentheses.
H_diabetes

CHAPTER 3. MEDIGAP 148
Table 3.8: Robustness Checks: Key Parameter Estimates from Alternative Specifications
and Samples
(1)
Panel A: Levels versus Logs
(2) (3)
Levels
Full model coefficient on Medigap in Levels 5829 (1850) 5705 (1572) 4978 (1520)
Average Cost 7344 (339) 7344 (339) 7344 (339)
Percent increase
Logs
0.794 (0.252) 0.777 (0.214) 0.678 (0.207)
Full model coefficient on Medigap in Logs 0.515 (0.158) 0.614 (0.146) 0.567 (0.138)
Panel B: Full versus Analyzed Sample
Full sample
OLS Coefficient on Medigap
Analyzed sample
0.138 (0.020) 0.105 (0.020) 0.343 (0.017)
OLS Coefficent on Medigap 0.379 (0.059) 0.352 (0.058) 0.370 (0.053)
Panel C: Lowest Premium Variation States versus Analyzed Sample
Lowest premums variation states
Full model coeffiicient on Medigap 0.418 (0.207) 0.436 (0.188) 0.424 (0.178)
OLS coefficient on Medigap 0.319 (0.083) 0.310 (0.081) 0.294 (0.077)
Premium semi-elasticity of demand for Medigap
Analyzed sample
-0.759 (1.589) -0.441 (1.546) -0.591 (1.525)
Full model coeffiicient on Medigap 0.515 (0.158) 0.614 (0.146) 0.567 (0.138)
OLS coefficient on Medigap 0.379 (0.059) 0.352 (0.058) 0.370 (0.053)
Premium semi-elasticity of Medigap demand
Controls in all panels
-0.314 (0.187) -0.406 (0.160) -0.419 (0.158)
Basic Controls Yes Yes Yes Yes Yes Yes
Demographic Controls Yes Yes Yes Yes
Health Controls Yes Yes
Source: 2000-2005 MCBS; Weiss Ratings
Notes: Parameter estimates for the effect of Medigap on total medical costs from various specifications. The sample is restricted to
individuals with Medigap, HMO. Panel A shows estimates from the full three-stage model with total medical costs in levels and logs.Panel B
presents estimates on the full sample and the analyzed sample restricted to individuals that turn 65 in 2000 or later. Panel C shows
estimates of the main paramters from the subset of states with virtually no within-state premium varition and the analyzed sample. The
controls in specifications 1-3 match those in Table 7 specifications 1-3. Bootstrap standard errors clustered by individual in parentheses.

CHAPTER 3. MEDIGAP 149
Table 3.9: Policy Counterfactuals for Subsidies and Taxes of Medigap Premiums
Tax Delta Medigap
Tax revenue
Medicare costs Budgetary effect
% of premiums ppt $/beneficiary % $/beneficiary % $/beneficiary %
-40.0% 19.0% -310 -5.6% 532 9.6% 843 15.2%
-20.0% 9.5% -128 -2.3% 266 4.8% 394 7.1%
20.0% -9.5% 74 1.3% -266 -4.8% -341 -6.2%
40.0% -19.0% 95 1.7% -532 -9.6% -627 -11.4%
60.0% -28.5% 62 1.1% -798 -14.4% -860 -15.6%
76.0% -35.8% 0 0.0% -1,003 -18.1% -1,003 -18.1%
Notes: Counterfactuals based on premium elasticity and Medigap estimates with full set of controls from Tables 6 and 7. Baseline Medigap market
share is 35.8 percent and premiums average $1,443. Medicare costs effect caculated assuming Medicare pays for 80 percent of care on the margin.
Percent of costs calculated using Medicare costs of $5,528 on average for Medigap enrollees.

CHAPTER 3. MEDIGAP 150
Table 3.10: OLS Estimates of Premiums on HRR- and State-Level Medicare Costs
Dependent Variable: Log premium
(1) (2)
Log HHR costs 0.35 0.00
(0.04) (0.04)
Log state costs 0.63
(0.04)
R-squared 0.19 0.50
N 290 290
F-stat on log state costs 251.04
Prob > F 0.00
Source: Premium data from Weiss Ratings; Medicare expenditure data
from the Dartmouth Atlas of Health Care
Notes: There is one observation for each HRR where Medigap is offered.
All variables are de-meaned. Premiums are for Plan C and all data is from
2003. Robust standard errors are parentheses.

CHAPTER 3. MEDIGAP 151
Table 3.11: Second Stage: 2SLS Estimates of Medigap Choice from Linear Probability
Model
(1)
Dependard Variable: Medigap Indicator
(2) (3)
Est. SE Est. SE Est. SE
Log premium -0.314 (0.187) -0.406 (0.160) -0.419 (0.158)
County MA penetration (%) -0.026 (0.002) -0.028 (0.002) -0.028 (0.002)
County MA penetration squared (%) 0.000 (0.000) 0.000 (0.000) 0.000 (0.000)
Log HRR cost 0.002 (0.063) 0.104 (0.064) 0.091 (0.063)
Male -0.050 (0.021) -0.117 (0.030)
Age-65 0.046 (0.033) 0.038 (0.033)
(Age-65)^2 -0.025 (0.017) -0.023 (0.017)
(Age-65)^3 0.004 (0.002) 0.004 (0.002)
Disabled eligibility
Race group
-0.019 (0.021) -0.015 (0.021)
Asian 0.012 (0.049) 0.031 (0.053)
Black -0.208 (0.024) -0.204 (0.025)
Other
Education group
-0.040 (0.044) -0.050 (0.044)
High school 0.083 (0.024) 0.085 (0.024)
Less than college 0.153 (0.025) 0.156 (0.026)
College or more 0.192 (0.032) 0.190 (0.033)
Log income 0.003 (0.010) 0.001 (0.009)
Married 0.023 (0.019) 0.023 (0.020)
Working 0.011 (0.022) 0.013 (0.022)
Served in armed forces
Self-reported health
-0.020 (0.029) -0.022 (0.029)
Excellent -0.108 (0.049)
Very good -0.141 (0.048)
Good -0.112 (0.045)
Fair -0.092 (0.046)
Diagnosis indicators Yes
Treatment indicators Yes
IADL indicators Yes
Year FE Yes Yes Yes
Log pseudolikelihood/R-squared 16.4% 20.4% 22.8%
N 3,556 3,556 3,556
Source: 2000-2005 MCBS; Weiss Ratings
Notes: Parameter estimates from 2SLS estimates of linear probability model for Medigap choice. The sample is restricted
to individuals with Medigap, HMO or no supplemental coverage who turned 65 in 2000 or later. The omitted race group is
white, education group is less than high school, and health status is poor. Bootstrap standard errors clustered by
individual in parentheses.

CHAPTER 3. MEDIGAP 152
Table 3.12: Third Stage: 3SLS Estimates of Medical Costs on Medigap Indicator
(1)
Dependent Variable: Log Medical Costs
(2) (3)
Log premium 0.495 (0.152) 0.630 (0.151) 0.585 (0.152)
Log HRR cost -0.019 (0.165) 0.348 (0.171) 0.207 (0.171)
Male -0.044 (0.072) -0.057 (0.102)
Age-65 0.806 (0.128) 0.899 (0.114)
(Age-65)^2 -0.232 (0.065) -0.295 (0.059)
(Age-65)^3 0.021 (0.009) 0.029 (0.008)
Disabled eligibility
Race group
0.421 (0.075) 0.325 (0.062)
Asian -0.276 (0.175) -0.183 (0.179)
Black -0.044 (0.110) -0.129 (0.110)
Other
Education group
0.043 (0.182) -0.061 (0.152)
High school 0.100 (0.079) 0.192 (0.076)
Less than college 0.089 (0.080) 0.221 (0.075)
College or more 0.159 (0.095) 0.335 (0.092)
Log income -0.007 (0.034) 0.054 (0.030)
Married -0.081 (0.069) 0.016 (0.065)
Working -0.435 (0.062) -0.172 (0.059)
Served in armed forces
Self-reported health
-0.100 (0.100) -0.114 (0.096)
Excellent -0.596 (0.137)
Very good -0.547 (0.145)
Good -0.387 (0.137)
Fair -0.192 (0.129)
Diagnosis indicators Yes
Treatment indicators Yes
IADL indicators Yes
Year FE Yes Yes Yes
R-squared 3.5% 10.6% 29.1%
N 3,415 3,415 3,415
Source: 2000-2005 MCBS; Weiss Ratings
Notes: Parameter estimates from 3SLS cost equation. Sample of individuals with HMO, Medigap or no supplemental
coverage who turned 65 no sooner than 2000 excluding those with positive VA expenditure. Omitted race group is white,
education group is less than high school, and health status is poor. Bootstrap standard errors clustered by individual in
parentheses.

Chapter 4
Do Expiring Budgets Lead to
Wasteful Year-End Spending?
Evidence from Federal
Procurement
with Jeffrey Liebman
4.1 Introduction
Many organizations have budgets that expire at year’s end. In the U.S., most budget
authority provided to federal government agencies for discretionary spending requires
the agencies to obligate funds by the end of the fiscal year or return them to the Trea-
sury general fund, and state and municipal agencies typically face similar constraints
(???). 1
For these organizations, unspent funding may not only represent a lost opportunity
1 At the end of the federal fiscal year, unobligated balances cease to be available for the purpose
of incurring new obligations. They sit in an expired account for 5 years in case adjustments are
needed in order to accurately account for the cost of obligations incurred during the fiscal year for
which the funds were appropriated. At the end of the 5 years, the funds revert to the Treasury
general fund.
153

CHAPTER 4. YEAR-END SPENDING 154
but can also signal a lack of need to budget-setters, decreasing funding in future
budget cycles (??). When current spending is explicitly used as the baseline in setting
budgets for the following year, this signaling effect is magnified.
This “use it or lose it” feature of time-limited budget authority has the potential to
result in low value spending—since the opportunity cost to organizations of spending
about-to-expire funds is effectively zero. Exacerbating this problem is the incentive to
build up a rainy day fund over the front end of the budget cycle. Most organizations
are de facto liquidity constrained, facing at the very least a high cost to acquiring mid-
cycle budget authority. When future spending demands are uncertain, organizations
have an incentive to hoard. Thus, at the end of the budget cycle, organizations often
have a buffer-stock of funding which they must rush to spend through.
To illustrate the key mechanisms which create the potential for wasteful year-end
spending, we build a simple model of spending with fixed budgets and expiring funds.
Annual budget cycles are divided into two six-month periods. Spending exhibits de-
creasing returns within each period and is scaled by a productivity parameter that
is unknown in advance. In the face of this uncertainty, organizations engage in pre-
cautionary savings in the first period. In the second period, the prospect of expiring
funds leads to a rush to spend. As a result, average spending is higher and average
quality is lower at the end of the year.
There is some suggestive evidence consistent with these predictions for the U.S.
federal government. A Department of Defense employee interviewed by ? describes
“merchants and contractors camped outside contracting offices on September 30th
(the close of the fiscal year) just in case money came through to fund their contracts.”
A 1980 report by the Senate Subcommittee on Oversight of Government Management
on “Hurry-Up Spending” found that the rush to spend led to poorly defined contracts,
limited competition and inflated prices, and the procurement of goods and services
for which there was no current need (?). At a congressional hearing in 2006, agency
representatives admitted to a “use-it-or-lose-it” mentality and a “rush to obligate”
at year’s end (?). The 2007 Federal Acquisition Advisory Panel concluded that “the
large volume of procurement execution being effected late in the year is having a
negative effect on the contracting process and is a significant motivator for many of

CHAPTER 4. YEAR-END SPENDING 155
the issues we have noted with respect to, among another things, lack of competition
and poor management of interagency contracts” (?).
Yet despite these accounts, there is no hard evidence on whether year-end spend-
ing surges are currently occurring in the U.S. federal government or whether year-end
spending is lower-value than spending during the rest of the year. Government Ac-
countability Office (GAO) reports in 1980 and 1985 documented that fourth quarter
spending was somewhat higher than spending during the rest of year using aggregate
agency spending data. In a follow-up report, ? stated that because “substantial
reforms in procurement planning and competition requirements have changed the
environment . . . year-end spending is unlikely to present the same magnitude of
problems and issues as before.” However, this later report was unable to examine
quarterly agency spending patterns for 1997 because agency compliance with quar-
terly reporting requirements was incomplete. Other government professionals cite
similar constraints to empirical analysis. In summarizing the response of Department
of Defense contracting officers to the interview question, “How would you measure
the quality of year-end spending?” ? writes, “Absent flagrant abuse that no one
could miss, there is no practical way of weighing year-end spending.”
This paper address the empirical shortfall. It begins by documenting the within-
year pattern of federal spending on government contracts, the main category of spend-
ing where significant discretion about timing exists. The analysis demonstrates that
there is a large surge in the 52nd week of the year that is concentrated in procure-
ments for construction-related goods and services, furnishings and office equipment,
and I.T. services and equipment. It also shows that the date of completion of annual
appropriations legislation has a noticeable effect on the timing of federal contracting,
consistent with anecdotal claims that part of the difficulties agencies face in effective
management of acquisitions comes from tardy enactment of appropriations legisla-
tion. The estimates show that a delay of ten weeks, roughly the average over this
time period, raises raises the share of spending in the last month by 1 percentage
point, from a base of about 15 percent.
We then analyze the impact of the end-of-year spending surge on spending quality
using a newly available dataset on the status of the federal government’s 686 major

CHAPTER 4. YEAR-END SPENDING 156
information technology projects—a total of $130 billion in spending. Consistent with
the model, spending on these I.T. projects spikes in the last week of the fiscal year,
increasing to 7.2 times the rest-of-year weekly average. Moreover, the spike is not
isolated to a small set of agencies or a subset of years, but rather a persistent feature
both across agencies and over time. In tandem with the spending increase, there
is a sharp drop-off in investment quality. Based on a categorical index of overall
investment performance, which combines assessments from agency information officers
with data on cost and timeliness, we find that projects that originate in the last week
of the fiscal year have 2.2 to 5.6 times higher odds of having a lower quality score.
Ordered logit and OLS regressions show that this effect is robust to agency and year
specific factors as well as to a rich set of project characteristic controls.
Our findings suggest that the various safeguard measures put into place in re-
sponse to the 1980 Senate Subcommittee report (?) and to broader concerns about
acquisition planning have not been fully successful in eliminating the end-of-year
rush-to-spend inefficiency. An alternative solution is to give agencies the ability to
roll over some of their unused funding for an additional year. Provisions of this nature
have been applied with apparent success in the states of Oklahoma and Washington,
as well as in the UK (??). Within the U.S. federal government, the Department of
Justice (DOJ) has obtained special authority to roll over unused budget authority
into a fund that can be used in the following year.
We extend the model to allow for rollover and show that, in the context of the
model, the efficiency gains from this ability are unequivocally positive. To test this
prediction, we study I.T. contracts at the Department of Justice which has special
rollover authority. We show that there is only a small end-of-year I.T. spending spike
at DOJ and that the one major I.T. contract DOJ issued in the 52nd week of the
year has a quality rating that is well above average.
The rest of the paper proceeds as follows. Section 4.2 presents a model of wasteful
year-end spending and discusses the mechanisms that could potentially lead to end-
of-year spending being of lower quality than spending during the rest of the year.
Section 4.3 examines the surge in year-end spending using a comprehensive dataset
on federal procurement. Section 4.4 tests for a year-end drop-off in quality using data

CHAPTER 4. YEAR-END SPENDING 157
on I.T. investments. Section 4.5 analyzes the Department of Justice experience with
rollover authority. Section 4.6 concludes.
4.2 A Model of Wasteful Year-End Spending
In this section, we present a simple model to illustrate how expiring budgets can give
rise to wasteful year-end spending. The model has three key features. First, there
are decreasing returns to spending within each sub-year period. Decreasing returns
could be motivated by a priority-based budgeting rule, where during a given period
organizations allocate resources to projects according to the surplus they provide.
Alternatively, decreasing returns could be motivated by short-run rigidities in the
production function. For example, federal agencies with a fixed staff of contracting
specialists might have less time to devote to each contract in a period with abnormally
high spending.
Second, there is uncertainty over the value of future budget resources. One can
think of uncertainty arising from either demand or supply factors. Shifts in mili-
tary strategy or an influenza outbreak, for example, could generate an unanticipated
change in demand for budget resources. On the supply side, uncertainty could be
driven by variation in the price or quality of desired goods and services. 2
Third, resources that are not spent by the end of the year cannot be rolled over
to produce value in subsequent periods. At the end of this section, we also analyze
what happens when this constraint is relaxed.
4.2.1 The Baseline Model
Consider an annual model of budgeting where an organization chooses how to appor-
tion budget authority, B, normalized to 1, over two six-month periods, denoted by
t = {1, 2}, to maximize a Cobb-Douglas objective. Denote spending in each period by
xt, and normalize its price to 1. To model decreasing returns and uncertainty, assume
that the Cobb-Douglas elasticity parameters αt are stochastic i.i.d. draws from the
2 As an example of supply side uncertainty, during the recent recession many agencies have experienced
construction costs for Recovery Act projects that were below projections.

CHAPTER 4. YEAR-END SPENDING 159
To summarize, decreasing returns and uncertainty create an incentive for organi-
zations to build up a rainy day fund in the first period, spending less than half of their
budget on average. At the end of the year, spending increases and average quality is
lower than in the earlier part of the year.
It is worth emphasizing that this model illustrates two different channels through
which end-of-year spending can be of lower quality. The first channel comes from
the uncertainty about future needs and will lead organizations to engage in only
high value projects early in the year and wait to undertake lower value projects once
the full-year demands on resources are clearer. The second channel comes from the
production-function rigidities motivated by the anecdotal evidence that contracting
officers become over-extended in the end-of-year rush to get money out the door.
Either channel by itself would be enough to produce an empirical relationship in
which end-of-year spending is of lower value.
4.2.2 Rollover Budget Authority
The inefficiency from wasteful year-end spending raises the question of whether any-
thing can be done to reduce it. Reducing uncertainty would be helpful, but is infea-
sible in practice for many organizations due to the inherently unpredictable nature
of some types of shocks. Another way to potentially increase efficiency would be to
allow organizations to roll over budget authority across fiscal years. Under such a
system, budgeting would still occur on an annual basis, but rather than expiring at
year’s end, unused funds would be added to the newly granted budget authority in
the next year.
The idea that budget authority should last for longer than one year is not new. As
? has pointed out, granting Congress the power to collect taxes, Article 1, Section 8
of the U.S. Constitution gave Congress the power of taxation, “To raise and support
armies, but no appropriation of money to that use shall be for longer term than two
years.” Not only does this suggest that the Founding Fathers though that two-year
limits were reasonable in some instances, but by failing to attach this clause to other
forms of federal expenditure, they implied by omission that periods longer than two

CHAPTER 4. YEAR-END SPENDING 162
For many types of government spending, there is little potential for a spike in
spending at the end of the year. The 65 percent of spending that is made up of
mandatory programs and interest on the debt is not subject to the timing limitations
associated with annual appropriations. The 13 percent of spending that pays for
compensation for federal employees is unlikely to exhibit an end-of-year surge since
new hires bring ongoing costs. This leaves procurement of goods and services from
the private sector as the main category of government spending where an end-of-year
spending surge could potentially occur. We therefore focus our empirical work on the
procurement of goods and services that accounted for $538 billion or 15.3 percent of
government spending in 2009 (up from $165 billion dollars and 9.2 percent in 2000).
It is worth noting that even within procurement spending, there are some cat-
egories of spending for which it would be unlikely to observe an end-of-year spike.
Some types of appropriated spending, such as military construction, come with longer
spending horizons to provide greater flexibility to agencies. Moreover, there are limits
to what kinds of purchases can be made at year end.
In particular, Federal law provides that appropriations are available only to “meet
the bona fide needs of the fiscal year for which they are appropriated.” Balances
remaining at the end of the year cannot generally be used to prepay for the next
year’s needs. A classic example of an improper obligation was an order for gasoline
placed 3 days before the end of the fiscal year to be delivered in monthly installments
throughout the following fiscal year (?). That said, when there is an ongoing need
and it is impossible to separate the purchase into components performed in different
fiscal years, it can be appropriate to enter into a contract in one fiscal year even
though a significant portion of the performance is in the subsequent fiscal year. In
contrast, contracts that are readily severable generally may not cross fiscal years
(unless specifically authorized by statute). 6
6 Over the past two decades, Congress has significantly expanded multi-year contracting authorities.
For example, the General Services Administration can enter into leases for periods of up to 20
years, and agencies can contract for services from utilities for periods of up to 10 years.

CHAPTER 4. YEAR-END SPENDING 163
4.3.1 The Federal Procurement Data System
Falling technology costs and the government transparency movement have combined
to produce an extraordinary increase in the amount of government data available on
the web (??). As of October 2010, Data.gov had 2,936 U.S. federal executive branch
datasets available. The Federal Funding Accountability and Transparency Act of
2006, sponsored by Senators Coburn, Obama, Carper, and McCain, required OMB
to create a public website, showing every federal award, including the name of the
entity receiving the award and the amount of the award, among other information.
USAspending.gov was launched in December 2007 and now contains data on federal
contracts, grants, direct payments, and loans.
The data currently available on USAspending.gov include the full Federal Procure-
ment Data System (FPDS) from 2000 through 2009. FPDS is the data system that
tracks all federal contracts. Every new contract awarded as well as every follow-on
contracting action such as a contract renewal or modification results in an observation
in FPDS. Up to 176 pieces of information are available for each contract including
dollar value, a four digit code describing the product or service being purchased, the
component of the agency making the purchase, the identity of the provider, the type
of contract being used (fixed price, cost-type, time and materials, etc.), and the type
of bidding mechanism used. While FPDS was originally created in 1978, agency re-
porting was incomplete for many years, and we are told that it would be difficult to
assemble comprehensive data for years before 2000. Moreover, while FPDS is thought
to contain all government contracts from 2000 on, data quality for many fields was
uneven before the 2003 FPDS modernization. Therefore, for most of the FPDS-based
analyses in this paper, we limit ourselves to data from fiscal years 2004 through 2009. 7
Table 4.1 shows selected characteristics of the FPDS 2004 to 2009 sample. There
were 14.6 million contracts during this period or an average of 2.4 million a year.
The distribution of contract size is highly skewed. Ninety-five percent of contracts
were for dollar amounts below $100,000, while 78 percent of contract spending is
7 FPDS excludes classified contracts. Data are made available in FPDS soon after an award,
except during wartime the Department of Defense is permitted a 90 day delay to minimize the
potential for disclosure of mission critical information.

CHAPTER 4. YEAR-END SPENDING 164
accounted for by contracts of more than $1 million. Seventy percent of contract
spending is by the Department of Defense. The Department of Energy and NASA,
which rely on contractors to run large labs and production facilities, and the General
Services Administration, which enters into government-wide contracts and contracts
on behalf of other agencies, are the next largest agencies in terms of spending on
contracts. Twenty-nine percent of contracts were non-competitive, 20 percent were
competitive but received only a single bid, and 51 percent received more than 1 bid.
Sixty-five percent were fixed price, 30 percent were cost-reimbursement, and 6 percent
were on a time and materials or labor hours basis.
4.3.2 The Within-Year Pattern of Government Procurement
Spending
Figure 4.1 shows contract spending by week, pooling data from 2004 through 2009.
There is a clear spike in spending at the end of the year with 16.5 percent of all
spending occurring in the last month of the year and 8.7 percent occurring in the
last week. The bottom panel shows that when measured by the number of contracts
rather than the dollar value, there is also clear evidence of an end-of-the-year spike,
with 12.0 percent occurring in the last month and 5.6 percent occurring in the last
week.
Table 4.1 shows that the end of the year spending surge occurs in all major
government agencies. If spending were distributed uniformly throughout the year,
we would expect to see 1.9 percent in the final week of the year. The lowest agency
percentage is 3.6 percent.
Table 4.3 shows the percent of spending on different types of goods and services
that occurs at the end of the year. The table shows some of the largest spend-
ing categories along with some selected smaller categories that are very similar to
the large categories. Construction-related goods and services, furnishings and office
equipment, and I.T. services and equipment all have end-of-year spending rates that
are significantly higher than the average. These categories of spending often represent
areas where there is significant flexibility about timing for performing maintenance

CHAPTER 4. YEAR-END SPENDING 165
or upgrading facilities and equipment, and which, because they represent on-going
needs, have a reasonable chance of satisfying the bona fide needs requirement even if
spending happens at the end of the year.
The categories of spending under the “Services” heading have end-of-year spending
rates that are near the average. For these kinds of services it will often be difficult to
meet the bona fide needs requirements unless the services are inseparable from larger
purchases, the services are necessary to provide continuity into the beginning of the
next fiscal year, or the services are covered by special multiyear contracting authori-
ties. Thus it is not surprising that their rate of end of year spending is lower than that
for construction, for example. There are two categories of spending where there is
very little year-end surge. The first is ongoing expenses such as fuels where attempts
to spend at the end of the year would represent a blatant violation of prohibitions
against paying for the following year’s expenses with current year appropriations.
The second is military weapons systems where because of long planning horizons and
the flexibility provided by special appropriations authorities, one would not expect
to see a concentration of spending at the end of the year.
Figure 4.1 also shows a spike in spending in the first week of the year, along with
smaller spikes at the beginning of each quarter. The spending patterns for these
beginning of period contracts are very different from those at the end of the year.
Appendix Table 4.9 shows that leases and service contracts are responsible for most
of the beginning-of-period spikes.
4.3.3 The Impact of Appropriations Timing on the Within-
Year Pattern of Government Procurement Spending
It is the exception rather than the rule for Congress to pass annual appropriations
bills before the beginning of the fiscal year. Over the 10 years from 2000 to 2009, the
full annual appropriations process was never completed on time. Although defense
appropriations bills were enacted before the start of the fiscal year 4 times, in 8 of
the ten years, appropriations for all or nearly all of the civilian agencies were enacted
in a single consolidated appropriations act well after the start of the fiscal year.

CHAPTER 4. YEAR-END SPENDING 166
Analysts have attributed some of the challenges facing federal acquisition to the
tardiness of the appropriations process, since the delays introduce uncertainty and
compress the time available to plan and implement a successful acquisition strategy
(?). In this subsection we analyze the relationship between the timing of the annual
appropriations acts and the within-year pattern of government contract spending.
For this analysis, we use the full 2000 to 2009 FPDS data, even though the data prior
to 2004 are of lower quality. In particular, in these earlier years it appears that for
some agencies, contracts are all assigned dates in the middle of the month and the
within-month weekly pattern is therefore not fully available.
Figure 4.2 shows results from regressing measures of end-of-year spending on the
timing of annual appropriations. This analysis has two data points for each year,
one representing defense spending and the other representing non-defense spending.
For each observation we measure the share of annual contract spending occurring in
the last quarter, month, and week of the year and the “weeks late” of the enactment
of annual appropriations legislation (enactment is defined by the date the President
signs the legislation). “Weeks late” measures time relative to October 1 and takes on
negative values when appropriations were enacted prior to the start of the fiscal year.
For defense spending, “weeks late” measures the date that the defense appropriations
bill was enacted. For non-defense spending the date is assigned from the date of
the consolidated appropriations act, or, in the case of the two years in which there
was not a consolidated act, a date that is the midpoint of the individual non-defense
appropriations acts. 8
There is a clear pattern in the data in which later appropriation dates result
in a greater fraction of end-of-year spending. In the plots, we show the separate
slopes of the defense and non-defense observations. Defense spending tends to be
appropriated earlier and to have less end of year spending, but the slopes for the two
types of spending are similar. The labels show the regression coefficients, including
the coefficients from a pooled regression in which defense and non-defense spending
8 We aggregate all non-defense spending because it facilitates communication of the pattern of
results while capturing nearly all of the available variation. We have also run analysis in which we
assign each non-defense agency the date of its individual appropriations act and obtain very similar
results.

CHAPTER 4. YEAR-END SPENDING 167
have different intercepts but are constrained to have the same slope. The estimates
show that a delay of ten weeks, roughly the average over this time period, raises the
share of spending in the last quarter by 2 percentage points from a base of about
27 percent. A ten-week delay raises the share of spending in the last month by 1
percentage point, from a base of about 15 percent. Both coefficients are statistically
significant at the 1 percent level. As we mentioned above, we do not have reliable
within-month data on timing for the years before 2004, so we exclude years before
2004 for the analysis of spending during the last week of the year. The estimates
indicate that a 10 week delay raises the share of spending by 1 percentage point on
a base of 9 percent. Due to the smaller sample, the estimate is less precise, with a
p-value of .07.
Overall, the analysis in this section shows that the end-of-year spending surge
is alive and well, thirty years after Congress and GAO focused significant attention
on the problem and despite reforms designed to limit it. Moreover, claims that late
appropriations increase the end-of-year volume of contracting activity are accurate,
suggesting that late appropriations may be exacerbating the already adverse effects
of having an acquisition workforce operating beyond capacity at the end of the year.
A surge in end-of-year spending does not necessarily imply bad outcomes. Agency
acquisition staff can plan ahead for the possibility that extra funds will be available.
Indeed, for large contracts weeks and even months of lead time are generally necessary.
The next section of the paper therefore analyzes the relative quality of end-of-year
contract spending to explore whether there are any adverse effects of the end-of-year
spending surge.
4.4 Is End of Year Spending of Lower Quality?
Our model predicts that end-of-year spending will be of lower quality both because
agencies will spend money at the end of the year on low value projects and because
the increased volume of contracting at the end of the year will lead to less effective
management of those acquisitions. As the discussion in the introduction indicated,
it has been challenging historically to study contract quality because of the limited

CHAPTER 4. YEAR-END SPENDING 168
availability of data. In this section of the paper, we use a new dataset that includes
quality information on 686 of the most important federal I.T. procurements to study
whether end-of-the year procurements are of lower quality.
4.4.1 I.T. Dashboard
Our data come from the federal I.T. Dashboard (www.itdashboard.gov) which tracks
the performance of the most important federal I.T. projects. The I.T. Dashboard
came online in beta form in June, 2009 and provides the public with measures of
the overall performance of major I.T. projects. Like the USAspending.gov data dis-
cussed earlier, the I.T. Dashboard is part of the trend toward “open government”
and part of a shift in federal management philosophy toward monitoring performance
trends rather than taking static snapshots of performance and of making the trends
public both for the sake of transparency and to motivate agencies to achieve high
performance (?). 9
Along with providing credible performance data for a portion of contract spending,
studying federal I.T. projects has two other advantages. The first is the ubiquity of
I.T. spending. Major information technology projects are carried out by nearly all
components of the U.S. federal government. Compared to an analysis of, say, the
purchase of military or medical equipment, an analysis of I.T. spending shines a much
broader light on the workings of government, allowing us to test our hypotheses across
agencies with a wide range of missions and organizational cultures.
The second advantage is that federal I.T. spending is an important and growing
federal activity. Federal I.T. expenditure was $81.9 billion in 2010, and has been grow-
ing at an inflation-adjusted rate of 3.8 percent over the past 5 years. 10 . Moreover,
these expenditure levels do not account for the social surplus from these projects.
9 The legislative foundation for the I.T. Dashboard was laid by the Clinger-Cohen Act of 1996,
which established Chief Information Officers at 27 major federal agencies and called on them to
“monitor the performance of the information technology programs of the agency, [and] evaluate
the performance of those programs on the basis of applicable performance measurements.” The
E-Government Act of 2002 built upon this by requiring the public display of these data.
10 Analytical Perspectives: Budget of the U.S. Government, 2010

CHAPTER 4. YEAR-END SPENDING 169
It is reasonable to think that information systems used to monitor terrorist activi-
ties, administer Social Security payments, and coordinate the health care of military
veterans could have welfare impacts that far exceed their dollar costs.
Finally, it should be noted that while we are duly cautious about external validly,
the widespread nature of I.T. investment across all types of organizations, including
private sector ones, makes a study of I.T. purchases more broadly relevant than would
be certain other categories of spending where the federal government may be the only
purchaser. Not only do non-federal organizations buy similar products under similar
budget structures, but they often purchase these products from the same firms that
sell to U.S. federal agencies. These firms know the end-of-year budgeting game, and if
they play it at the U.S. federal level, there may be reason to believe that they operate
similarly elsewhere. 11
4.4.2 Data and Summary Statistics
The I.T. Dashboard displays information on major, ongoing projects made by 27 of
the largest agencies of the federal government. The information is gleaned from the
Exhibit 53 and Exhibit 300 forms that agencies are required to submit to the Office
of Management and Budget and is constantly updated on the Dashboard website,
allowing users to view and conduct simple analysis of the data. The data we use was
downloaded in March, 2010 at which time there were 761 projects being tracked.
For the analysis, we drop the 73 observations that are missing the quality mea-
sures, date of award, or cost variables. We also drop two enormous projects because
their size would cause them to dominate all of the weighted regression results and
because they are too high-profile to be indicative of normal budgeting practices. 12
This leaves us with a baseline sample of 686 projects and $130 billion in planned
total spending.
11 See ? for a discussion of the incentives facing government contractors.
12 These projects are a $45.5 billion project at the Department of Defense and a $19.5 billion
project at the Department of Homeland Security; the next largest project is $3.9 billion and the
average of the remaining observations is $219 million. Because the dropped observations have above
average overall ratings and are not from the last week of the year, omitting the observations works
against us finding the effect predicted my our model.

CHAPTER 4. YEAR-END SPENDING 170
Table 4.4 shows the year of origination of these projects and the agencies at which
they occurred. Almost two-thirds of these projects (64.6 percent) and half of the
spending (50.3 percent) originated in 2005 or later, although there are some ongoing
projects that originated more than 20 years ago. 13 The projects are distributed
fairly broadly across agencies. Although the Department of Defense, Department of
Transportation, and Department of Veteran’s Affairs have higher levels of spending,
the vast majority of the agencies have at least 10 projects (21 of 27) and at least $1
billion in aggregate spending (20 of 27).
The main performance measure tracked on the I.T. dashboard is the overall rating
for the project. The canonical approach to tracking acquisitions is to measure cost,
schedule, and performance. The overall rating therefore combines three subindexes.
The cost rating subindex is based on the absolute percent deviation between the
planned and actual cost of the project. Projects that are on average within 5 percent
of the scheduled cost receive a score of 10, projects that are within 5 percent to 10
percent on average receive a score of 9, and so on down to zero. Because the symmetric
treatment of under and over-cost projects is somewhat unnatural, in our analysis we
also construct an alternative index, “cost overrun” which gives under-cost projects
the highest scores and over-cost projects the lowest. In this index, projects that are
at least 45 percent under-cost receive a score of 10, projects that are 35 percent to 45
percent under-cost receive a score of 9, and so on.
The schedule rating subindex is based on the average tardiness of the project
across milestones. Projects that are no more than 30 days overdue on average receive
a score of 10, projects that are no more than 90 days overdue on average receive a
score of 5, and projects that are more than 90 days overdue on average receive a score
of 0.
The third subindex is a subjective CIO evaluation score, designed to incorporate
an assessment of contract performance. The rating is intended to reflect the CIO’s
“assessment of the risk of the investment’s ability to accomplish its goals,” with
the CIO instructed to “consult appropriate stakeholders in making their evaluation,
13 We address sample selection issues in the sensitivity section below.

CHAPTER 4. YEAR-END SPENDING 172
To classify year-end projects, we use the date the first contract of the projects was
signed, creating an indicator variable for projects that originated in the last seven
days of September, the end of the fiscal year. Most I.T. projects are comprised of a
series of contracts that are renewed and altered as milestones are met and the nature
of the project evolves. We think that using the date the first contract was signed
to classify the start date of the project is the best approach as the key structure of
the project is most likely determined at its onset. While future contract awards may
affect the quality of the project, we only observe outcomes at the project level. We
view any potential measurement error from our approach as introducing downward
bias in our coefficient of interest as contracts initially awarded before the last week
of the year may be contaminated by modifications made in the last week of a later
year and contracts initially awarded at the rush of year’s end may be rectified at a
later point.
Figure 4.3 shows the weekly pattern of spending in the I.T. Dashboard sample.
As in the broader FPDS sample, there is a spike in spending in the 52nd week of the
year. Spending and the number of projects in the last week increase to 7.2 and 8.3
times their rest-of-year weekly averages respectively. Alternatively put, while only
accounting for 1.9 percent of the days of the year, the last week accounts for 12.3
percent of spending and 14.0 percent of the number of projects. Activity is tilted
even more strongly towards the last week if the sample of projects is restricted to
the 65.1 percent of contracts that are for less than $100 million. Given the longer
planning horizon for larger acquisitions, it is not surprising that we see more of a
year-end spike for the smaller contracts. 17
4.4.3 The Relative Quality of Year-End I.T. Contracts
Figure 4.4 shows the distributions of the overall rating index for last-week-of-the-year
projects and projects from the rest of the year. In these histograms, the ratings on
the 0 to 10 scale are binned into 5 categories with the lowest category representing
overall ratings less than 2, the second lowest representing overall ratings between 2
17 As in the broader FPDS sample, the end of year spike in the I.T. data is a broad phenomenon,
not limited to a few agencies.

CHAPTER 4. YEAR-END SPENDING 174
Recall that odds ratios capture the proportional change in the odds of a higher cat-
egorical value associated with a unit increase in the dependent variable, so that an
odds ratio of 1/2 indicates that the odds of a higher categorical value are 50 percent
lower, or reciprocally that the odds of a lower categorical variable are 2 times as great.
The results in this table are weighted by inflation-adjusted spending.
The first column of the table shows the impact of a last week contract on the
rating in a regression with no covariates. Columns 2 through 4 sequentially add in
fixed effects for year, agency, and project characteristics. In all of the specifications,
the odds ratios are well below one—ranging from 0.18 to 0.46—implying that last week
spending is of significantly lower quality than spending in the rest of the year (the
p-values are less than 0.01 in all specifications). The estimates imply that spending
that originates in the last week of the fiscal year has 2.2 to 5.6 times higher odds of
having a lower quality score.
4.4.4 Sensitivity Analysis
This subsection explores the robustness of the basic estimates. It shows how the
results vary with different treatment of large contracts, with different functional form
assumptions, and when selection into the sample is modeled.
Figure 4.4 showed that the finding that year end projects are of lower quality was
more pronounced in the dollar weighted analysis than in the unweighted analysis,
suggesting that a few large poor performing contracts may be heavily affecting the
results. The first four columns of Table 4.7 analyze this issue. The first two columns
split the sample at the median contract size of $62 million. In the sample of smaller
contracts, the coefficient of 0.60 is substantially below one but is less precisely es-
timated (p-value of 0.17). The point estimate in column (3) from an unweighted
regression is quite similar to the estimate in column (1) for the smaller contracts,
but with added precision from doubling the sample size by including the full sam-
ple (p-value of .02). Results in which we Winsorize the weights, assigning a weight
of $1 billion to the 4 percent of projects that are larger than $1 billion, are about
half way between the full sample weighted and unweighted results (p-value less than

CHAPTER 4. YEAR-END SPENDING 175
0.01). Overall, it is clear that the pattern of lower rating for end of year contracts is
a broad phenomenon. It is also clear that the sample contains several very large low
rated projects that were originated in the last week of the year—possibly providing
evidence that it is particularly risky to rush very large contracts out the door as the
fiscal year deadline approaches.
Column (5) of Table 4.4 shows results from an ordinary least squares (OLS) model
in which the raw overall rating is regressed on an indicator for the contract originating
in the last week of the year and on controls. The regression coefficient of -1.00 shows
that I.T. spending contracted in the last week of the year receives ratings that are
on average a full point lower on the 0 to 10 rating scale. This estimate also confirms
that the finding of lower quality year end spending is not limited to the ordered logit
functional form.
An important feature of our sample is that it reflects only active I.T. projects.
Projects that have already been completed or projects that were terminated without
reaching completion are not in our sample. Unfortunately, because the I.T. dashboard
and the CIO ratings are brand new, it is not possible to acquire rating information
on the major I.T. projects that are no longer ongoing.
Ideally, one would want a sample of all major I.T. projects that originated in a
particular period in time. The bias introduced by the way in which our sample was
constructed most likely leads us to underestimate the end-of-year-effect. In particular,
very bad contracts begun in the last week of the year are likely to be canceled and
would not appear in our data set. Similarly, very well executed contracts from earlier
in the year are likely to be completed ahead of schedule and also not appear in our
data set. Thus, our estimates likely understate the gap in quality that we would find
if we could compare all contracts from the last week of the year with all contracts
from the rest of the year.
To explore the extent of bias that a selection mechanism like the one just described
might introduce into our estimates, we assembled a data set of all 3,859 major I.T.
projects that originated between 2002 and 2010. We were able to assemble this data
set using the annual Exhibit 53 reports that allow the Office of Management and
Budget to track I.T. projects across the major federal agencies. These data show that

CHAPTER 4. YEAR-END SPENDING 176
more recently originated projects are significantly more likely to be in our sample.
Our sample contains 85 percent of the total spending on projects that originated in
2007 or later and only 28 percent of the spending on projects that originated before
this date.
A simple way to assess whether there is selection is to estimate the model on
samples split into earlier and later years. A difference in the coefficient of interest
across samples, given the assumption that there is no time trend in the effect, would
be indicative of selection bias. Given this assumption, however, we can estimate the
parameter of interest exactly by using the date of project origination to identify a
selection correction term. Column (6) implements this strategy, showing estimates
from a standard Heckman selection model where the year or origination is excluded
from the second stage. The results show a larger effect than the corresponding OLS
estimate, but the lack of precession means that we cannot rule out that the effects
are the same. 19 The negative coefficient on the selection term, although statistically
indistinguishable from zero, suggests that lower quality projects are on net more likely
to remain in the sample over time.
4.4.5 Why Are Year End Contracts of Lower Quality?
The results from the I.T. dashboard show that, consistent with the predictions of our
model, year end spending is of lower quality than spending obligated earlier in the
year. 20 Our model posited two channels: agencies may save low priority projects for
the end of the year and undertake them only if they have no better uses for the funds,
and the high volume of contracting activity at the end of the year might allow for
less management attention per project.
19 Consistent with this finding, OLS estimates on a sample split in 2007 show a larger point estimate
in the later years, but we cannot reject the hypothesis that the coefficients are the same.
20 In addition to the last week of the year results described above, we have also examined last
month of the year spending and find that spending in the balance of the last month of the year is of
moderately lower quality than that in the first 11 months of the year. We have also examined the
quality of first week of the year spending (which also spikes). The point estimate for the first week
of the year suggests somewhat higher spending quality, but the odds ratio difference from 1.0 is not
statistically significant.

CHAPTER 4. YEAR-END SPENDING 177
There is also a third possible mechanism. 21 Some program managers or contracting
officers may be inclined toward procrastination and these same employees may be ones
who do a worse job of planning for, writing, or managing contracts. Thus, we may
see a surge of low quality contracts at the end of the year because that is when the
least effective acquisition professionals get their contracts out the door. This third
mechanism could have different policy implications from the first two because allowing
agencies to roll over funds would not necessarily improve outcomes—the least effective
acquisition professionals would simply issue their contracts at a different time of year.
One way to evaluate this procrastination hypothesis would be to examine the within-
year calendar distribution of other contracts issued by the contracting officers who
issued last week contracts in our I.T. Dashboard sample to see if they appear to be
procrastinators who consistently complete their contracts late in the year. While we
can identify the contracting officers responsible for a contract in both the dashboard
sample and the FPDS sample, we cannot currently link the two samples, though we
believe we ultimately will be able to do so.
Another way to explore the possible mechanisms behind the poor outcomes for
end-of-year contracts is to examine the subcomponents of the overall rating to see
which subindices are responsible for the result. Appendix Table 4.10 repeats our main
ordered logit analysis with each subindex as the dependent variable. The results show
clearly that it is the evaluation by the agency CIO that is responsible for the main
finding. Neither the cost rating nor the schedule rating has an odds ratio that is
significantly different from 1. The CIO evaluation shows that the odds of having a
higher rating are one-sixth as high for last-week-of-the-year contracts. The coefficient
in the CIO regression is insensitive to adding the cost rating and scheduling rating
into the regression, suggesting that it is information in the CIO rating that is not
incorporated in the other rating that is responsible for the result. This finding is not
all that surprising. As we mentioned above, the I.T. dashboard explicitly places more
faith in the CIO’s assessment than in the other components by allowing the CIO
assessment to override the other components if it is lower than the other components.
Moreover, the ability to reset milestone targets makes it difficult to assess the cost
21 We thank Steve Kelman for suggesting this third mechanism.

CHAPTER 4. YEAR-END SPENDING 178
and schedule ratings. But while not surprising, the fact that it is the CIO evaluation
that is driving the result means that we cannot learn much about the mechanism
from the subindices, since the CIO evaluation in a comprehensive measure of the I.T.
project’s performance.
Another way to explore possible mechanisms is to examine whether other ob-
servable features of end-of-year contracts are different from those earlier in the year.
Specifically, we examine whether features that policymakers often define as high risk—
such as lack of competitive bidding or use of cost-reimbursement rather than fixed
cost pricing—are more prevalent in end of year contracts. For this analysis we return
to the FPDS sample of all contracts from 2004 to 2009. To facilitate the analysis,
we aggregate the 14.6 million observations up to the level of the covariates. We then
estimate linear probability models with indicators for contract characteristics (e.g.,
a non-competitively sourced indicator) as the dependent variable on an indicator
for last week of the fiscal year and controls. The regressions are weighted by total
spending in each cell.
The first three columns of Appendix Table 4.11 examine shifts in the degree of
competitive sourcing at the end of the year. The use of non-competitive contracts
shows little change. However, contracts that are competitively sourced are signifi-
cantly more likely to receive only one bid perhaps because the end of year rush leaves
less time to allow bidding to take place. The estimates indicate that there is almost
a 10 percent increase in the percent of contracts receiving only a single bid—a 1.7
percentage point increase one a base of 20 percent. On net, then, there is a modest
increase in “risky” non-competitive and one bid contracts at the end of the year.
Column (3) shows that spending on contracts that are either non-competitive or one
bid contract increases by 1 percentage points on a base of 49 percent in the last week
of the year.
The second three columns investigate the type of contract used. Contracts that
provide for cost reimbursement rather than specifying a fixed price are often seen as
high risk because they have significant potential for cost overruns. Time and material
or labor hours (T&M/LH) contracts raise similar concerns because they involve open
ended commitments to pay for whatever quantity of labor and materials are used

CHAPTER 4. YEAR-END SPENDING 179
to accomplish the task specified in the contract. Column (4) shows that end-of-year
contracts are less likely to include these sorts of high risk contract terms possibly
because agencies are attempting to use up defined sums of excess funds and therefore
limit contracts to the available amounts. The use of T&M/LH contracts increases
by 0.4 percentage points, which is substantial compared to a base of 5.5 percent.
Because T&M/LH contracts are infrequent, column (6) shows a net decrease in the
combined use of risky cost-reimbursement and T&M/LH contract spending of about
3 percentage points on a base of 36 percent.
Overall, the analysis in this section does not offer any clear insights into what
might be causing lower performance among end-of-year contracts. There is no reason
to expect it to be a single mechanism. All three mechanisms could be reinforcing
each other in contributing to the phenomenon.
4.5 Do Rollover Provisions Raise Spending Qual-
ity?
The third prediction of the model is that allowing for the rollover of unused funding
unambiguously improves quality, both overall and at year’s end. Intuitively, organi-
zations are less likely to engage in wasteful year-end spending when the funding could
be used for higher value projects in the next budget period.
As we noted in the introduction, since 1992 the Department of Justice (DOJ)
has had special authority to roll over unused funds to pay for future I.T. needs. In
this section of the paper, we examine whether the quality of DOJ’s end-of-year I.T.
spending is higher than that of federal agencies who lack rollover authority.
4.5.1 The DOJ’s Rollover Authority
The DOJ authority provides that “unobligated balances of appropriations available
to the Department of Justice during such fiscal year may be transferred into the
capital account of the Working Capital Fund to be available for the department-wide
acquisition of capital equipment, development and implementation of law enforcement

CHAPTER 4. YEAR-END SPENDING 180
or litigation related automated data processing systems, and for the improvement
and implementation of the Department’s financial management and payroll/personnel
systems.” 22 While other agencies have ongoing working capital funds, appropriated
funds contributed to those funds retain their fiscal year restrictions.
Between 1992 and 2006, approximately $1.8 billion in annual appropriations were
transferred to the DOJ working capital fund from unused appropriations balances (?).
Nonetheless, Table 4.2 shows that DOJ has an end-of-year spending surge comparable
to that of other agencies when all spending is taken into account, with 9.4 percent of
its spending occurring in the last week of the year.
Even with rollover authority, there remain incentives for agencies to use up their
full allocation of funding. Large balances carried over from one period to another
are likely to be interpreted by OMB and Congressional appropriators as a signal that
budget resources are excessive and lead to reduced budgets in subsequent periods.
For example, Senator Coburn issued a report in 2008 entitled “Justice Denied: Waste
and Management at the Department of Justice” in which he stated: “Every year
Congress appropriates more than $20 billion for the Department of Justice to carry
out its mission, and every year the Department ends the year with billions of unspent
dollars. But instead of returning this unneeded and unspent money to the taxpayers,
DOJ rolls it over year to year, essentially maintaining a billion dollar bank account
that it can dip into for projects for which the money was not originally intended”
(?). 23
It is not just external pressure that may lead an organization to spend all of its
resources even in the presence of rollover authority. Components of an agency may
not be willing to return resources to the center if they are not ensured of being able
to spend those resources after they are rolled over. Indeed, in 2006 testimony before
22 Public Law 102-104: 28 USC 527 note.
23 Coburn goes on to add “. . . perhaps without the pressure to rush to spend funds before
they are canceled, DOJ may, in fact, make more prudent spending decisions with unobligated funds.
This has not been studied and warrants examination for potential cost savings across the federal
government. As long as DOJ is banking billions of dollars from year to year that the Department
has some discretion to spend on its priorities as they arise, however, Congress should more carefully
review how much it is appropriating for DOJ programs. If a particular initiative or office does not
need or spend as much as Congress has appropriated, then Congress should consider appropriating
less for that particular office and the Department overall.”

CHAPTER 4. YEAR-END SPENDING 181
Congress, Deputy Assistant Attorney General Lee Lofthus may have been suggesting
a connection between those components of DOJ that contribute expiring funds to the
working capital fund and those that benefit from it when he noted that the FBI was
both the biggest contributor to the fund and the biggest beneficiary (?).
Given that rollover funds at DOJ are used for I.T. purposes, one might expect
to see little or no end-of-year spike in I.T. spending, because I.T. components of
the agency will know that they will directly benefit from funds contributed to the
working capital fund. This is indeed the case. In the comprehensive FPDS data,
only 3.4 percent of DOJ’s I.T. spending occurs in the last week—the 19th lowest of
the 21 major agencies. In the I.T. dashboard data, DOJ has 16 ongoing major I.T.
investments, with planned total spending of just over $5.1 billion. Only 1 investment
occurred in the last week of year. The $99 million cost of this investment implies
that 1.9 percent of DOJ spending occurred in the last week compared to an average
of 11.0 percent across the other agencies.
Although the sample size of investments is very small, the quality of DOJ’s year-
end spending is high. The one last week investment has the highest quality score of all
major I.T. investments at the agency. Difference-in-differences estimates presented in
Table 4.8, which allow us to control for investment characteristics, suggest that the
DOJ pattern is sufficiently unusual to be statistically significant—even though it is
identified off of a single end-of-year observation. Specifically, difference-in-differences
point estimates indicate that year-end quality increases by 2.0 to 3.5 categorical levels
at DOJ relative to other agencies. The p-values on the interaction variable are less
than 0.01 in all specifications.
Despite the statistical and economic significance of the estimates, we are hesitant
to draw strong conclusions from the estimated effect. In addition to the small number
of projects, the fact that the effect is identified off of a single agency raises the potential
for bias from unobserved organizational characteristics. Nevertheless, the evidence
that exists appears consistent with the prediction that rollover increases year-end
quality.

CHAPTER 4. YEAR-END SPENDING 182
4.6 Conclusion
Our model of an organization facing a fixed period in which it must spend its budget
resources made three predictions. We have confirmed all three of them using data on
U.S. federal contracting. First, there is a surge of spending at the end of the year.
Second, end of year spending is of lower quality. Third, permitting the rollover of
spending into subsequent periods leads to higher quality.
Because we cannot identify the exact mechanism producing the decline in spending
quality at the end of the year, it is difficult to draw firm policy conclusions from the
result. If the low spending quality comes from agencies squandering end of year
resources on low priority projects, possibly compounded by insufficient management
attention during the end of year spending rush, then allowing for rollover of unused
balances or switching to two-year budgeting might improve spending quality. But as
long as future agency budgets are based in part on whether agencies exhausted their
resources in the current period, there will still be an incentive for year-end spending
surges. And unless the rollover balances stay with the same part of the organization
that managed to save them, agency subcomponents will still have an incentive to
use up their allocations. An alternative approach would be to apply greater scrutiny
to end-of-the-year spending with the presumption that any spending above levels
occurring earlier in the year was unwarranted. This latter approach could also be the
proper management prescription if low quality end-of-year spending results from a
correlation between acquisition officer skill and a tendency to procrastinate—agencies
would want to give greater attention to end of year contracts because the acquisition
officials responsible for them need more oversight.
In evaluating possible policy reforms one should not lose sight of the potential
benefits of one-year budget periods. The annual appropriations cycle may provide
benefits from greater Congressional control over executive branch operations. More-
over, the use-it-or-lose it feature of appropriated funds may push projects out the
door that would otherwise languish due to bureaucratic delays.

CHAPTER 4. YEAR-END SPENDING 183
Figure 4.1: Federal Contracting by Week, Pooled 2004 to 2009 FPDS
Spending (billions)
Number of contracts (thousands)
$250
$200
$150
$100
$50
$-
800
600
400
200
1 2 3 4 5 6 7 8 910111213141516171819202122232425262728293031323334353637383940414243444546474849505152
Week
(a) Spending
1 2 3 4 5 6 7 8 910111213141516171819202122232425262728293031323334353637383940414243444546474849505152
Week
(b) Number of Contracts
Source: Federal Procurement Data System, accessed October, 2010 via
www.usaspending.gov.
Note: Total spending and number of contracts by week of the fiscal year. Spending values
inflation-adjusted to 2009 dollars using the CPI-U.

CHAPTER 4. YEAR-END SPENDING 184
20 25 30 35
Last quarter spending (percent)
Non-defense
2005
2001
2007
2004
Figure 4.2: Year-End Spending by Appropriations Date
2003
2006
2008
2002
2000
2000
2005
2008
2001
2006
2002
2004
2007
2003
2009
Defense
-10 0 10 20 30
Weeks late
2009
Non-defense slope = .193 (.094)
Defense slope = .180 (.053)
Pooled slope = .183 (.044)
(a) Last Quarter Spending
6 8 10 12 14 16
Last week spending (percent)
Non-defense slope = .112 (.146)
Defense slope = .081 (.045)
Pooled slope = .089 (.047)
2005
2007
Defense
2004
2006
2008
2005
10 15 20
Last month spending (percent)
2008
2006
2001
2005
Non-defense
2006
2003
2007
2004
2008
2002
2008
2000
20052001
2006
2000
2002
2004
2003
2007
2009
Defense
-10 0 10 20 30
Weeks late
2004
2007
2009
Non-defense slope = .130 (.078)
Defense slope = .089 (.030)
Pooled slope = .100 (.029)
(b) Last Month Spending
Non-defense
-10 0 10 20 30
Weeks late
(c) Last Week Spending
Source: Federal Procurement Data System, accessed October, 2010 via
www.usaspending.gov and Library of Congress.
Note: Vertical axes show the percent of annual spending occuring in the last quarter,
month, and week of the fiscal year. Horizontal axes shows the passage dates for the
non-defense and defense appropriation bills, relative to the first day of the fiscal year in
weeks. For defense spending, weeks late measures the date that the defense appropriations
bill was enacted. For non-defense spending the date is assigned from the date of the
consolidated appropriations act, or, in the case of the two years in which there was not a
consolidated act, a date that is the midpoint of the individual non-defense appropriations
acts. Plots show fitted lines and slope coefficients from bivariate regressions on defense
and non-defense spending. Pooled coefficients from a regression where defense and
non-defense spending have different intercepts but are constrained to have the same slope.
Robust standard errors in parentheses.
2009
2009

CHAPTER 4. YEAR-END SPENDING 185
Spending (millions)
Number of projects
$16,000
$12,000
100
$8,000
$4,000
80
60
40
20
0
$0
Figure 4.3: I.T. Contracting by Week
1 2 3 4 5 6 7 8 910111213141516171819202122232425262728293031323334353637383940414243444546474849505152
1 2 3 4 5 6 7
8 9101112131415
16 171819202122
Weeks
(a) Spending
23 24252627282930
Weeks
31 323334353637
(b) Number of Projects
38 39404142434445
46 474849505152
Source: I.T. Dashboard data, accessed March, 2010 via http://it.usaspending.gov.
Note: Total spending and number of I.T projects by week of the fiscal year. Spending
values inflation-adjusted to 2009 dollars using the CPI-U.

CHAPTER 4. YEAR-END SPENDING 186
Figure 4.4: Year-End and Rest-of-Year Overall Ratings
40.0%
30.0%
20.0%
10.0%
0.0%
45.0%
30.0%
15.0%
0.0%
25.8%
8.5%
Last week
Rest of year
3.0%
22.9%
5.5%
11.0%
32.4%
15.5%
36.5%
24.8%
1 2 3 4 5
Last week
Rest of year
2.9%
2.1%
2.8%
Overall Rating
(a) Spending
25.5%
24.4%
37.2%
40.6%
26.6%
1 2 3 4 5
Overall Rating
(b) Number of projects
Source: I.T. Dashboard data, accessed March, 2010 via http://it.usaspending.gov.
Note: Overall rating histograms for I.T. projects originating in the last week and rest of
the year. To construct this figure, ratings are binned into 5 categories with the lowest
category representing overall ratings less than 2, the second lowest representing overall
ratings between 2 and 4, and so on. See text for details on the overall rating index. Panel
A weights projects by inflation-adjusted spending. Panel B shows unweighted values.
22.6%
29.3%

CHAPTER 4. YEAR-END SPENDING 187
Table 4.1: Summary Statistics: Federal Contracting, Pooled 2004 to 2009 FPDS
Spending
Contracts
Billions Percent Count Percent
Totals $2,597 100.0% 14,600,000 100.0%
Year
2004 $304 11.7% 1,413,316 9.7%
2005 $355 13.7% 1,857,960 12.7%
2006 $405 15.6% 2,719,479 18.6%
2007 $452 17.4% 2,977,426 20.4%
2008 $542 20.9% 3,292,063 22.5%
2009 $538 20.7% 2,307,904 15.8%
Contract size
Less than $100K $166 6.4% 13,800,000 94.5%
$100K to $1M $398 15.3% 626,134 4.3%
At least $1M $2,033 78.3% 98,001 0.7%
Agency
Agriculture $25 1.0% 241,626 1.7%
Commerce $13 0.5% 112,756 0.8%
Defense $1,824 70.2% 3,536,530 24.2%
Education $8 0.3% 12,806 0.1%
Energy $142 5.5% 37,756 0.3%
Environmental Protection Agency $8 0.3% 62,713 0.4%
General Services Administration $82 3.2% 4,830,748 33.1%
Health and Human Services $76 2.9% 249,907 1.7%
Homeland Security $74 2.8% 255,461 1.7%
Housing and Urban Development $6 0.2% 15,666 0.1%
Interior $25 1.0% 377,743 2.6%
Justice $33 1.3% 420,379 2.9%
Labor $13 0.5% 41,229 0.3%
National Aeronautics and Space Administration $83 3.2% 81,211 0.6%
National Science Foundation $2 0.1% 4,201 0.0%
Other $37 1.4% 179,283 1.2%
Small Business Administration $0 0.0% 3,361 0.0%
State $34 1.3% 239,019 1.6%
Transportation $21 0.8% 57,235 0.4%
Treasury $25 1.0% 177,662 1.2%
Veterans Affairs $67 2.6% 3,630,856 24.9%
Competition type
Non-competitive $745 28.7% 3,553,453 24.3%
Competitive with one bid $521 20.0% 3,883,273 26.6%
Competitive with more than one bid $1,332 51.3% 7,131,422 48.8%
Contract type
Fixed price $1,675 64.5% 14,200,000 97.3%
Cost-reimbursement $780 30.0% 151,362 1.0%
Time and materials/labor hours $142 5.5% 249,705 1.7%
Source: Federal Procurement Data System, accessed October, 2010 via www.usaspending.gov
Note: Contract spending inflation adjusted to 2009 dollars using the CPI-U.

CHAPTER 4. YEAR-END SPENDING 188
Table 4.2: Year-End Contract Spending by Agency, Pooled 2004 to 2009 FPDS
Spending
Percent of spending
(billions) Last month Last week
Agriculture $24.8 17.0% 6.2%
Commerce $13.4 21.4% 5.6%
Defense $1,820.0 16.0% 8.6%
Education $8.2 18.6% 11.2%
Energy $142.0 6.6% 4.0%
Environmental Protection Agency $8.1 22.3% 10.4%
General Services Administration $82.0 12.9% 7.0%
Health and Human Services $76.4 25.5% 12.2%
Homeland Security $73.6 22.7% 9.4%
Housing and Urban Development $5.7 18.5% 11.7%
Interior $25.3 23.2% 7.6%
Justice $32.6 17.9% 9.4%
Labor $12.7 12.9% 5.9%
National Aeronautics and Space Administration $82.7 16.9% 11.0%
National Science Foundation $2.0 27.7% 11.5%
Small Business Administration $0.4 31.9% 16.3%
State $33.5 34.9% 20.4%
Transportation $20.5 17.6% 3.6%
Treasury $24.9 15.3% 9.6%
Veterans Affairs $66.9 18.2% 9.5%
Other $37.4 28.6% 18.9%
Total $2,600.0 16.5% 8.7%
Source: Federal Procurement Data System, accessed October, 2010 via www.usaspending.gov.
Note: Contract spending inflation adjusted to 2009 dollars using the CPI-U.

CHAPTER 4. YEAR-END SPENDING 189
Table 4.3: Year-End Contract Spending By Selected Product or Service Code, Pooled
2004 to 2009 FPDS
Spending
Percent of spending
(billions) Last month Last week
Construction-related
Construction of structures and facilities $136.0 40.9% 28.6%
Maintenance, repair, or alteration of real property $72.5 34.8% 20.1%
Architect and engineering services $32.8 26.1% 13.8%
Installation of equipment $4.0 33.9% 20.4%
Prefabricated structures and scaffolding $3.7 34.9% 18.4%
Furnishings and office equipment
Furniture $8.0 37.3% 18.4%
Office supplies and devices $4.0 24.9% 16.6%
Household and commercial furnishings and appliances $1.2 37.8% 20.7%
Office machines, text processing systems and equipment $1.1 33.5% 17.0%
I.T. services and equipment
Automatic data processing and telecom. services $145.0 21.0% 12.3%
Automatic data processing equipment $53.7 29.2% 14.9%
Services
Professional, admin, and management support services $336.0 19.1% 9.9%
Research and development $309.0 11.3% 5.3%
Utilities and housekeeping services $73.7 15.6% 9.1%
Ongoing
Fuels, lubricants, oils and waxes $72.7 13.2% 0.7%
Medical services $68.8 4.9% 1.7%
Chemicals and chemical products $6.2 3.3% 1.3%
Tires and tubes $1.0 8.7% 2.7%
Toiletries $0.3 12.2% 3.0%
Military weapons systems
Aircraft and airframe structural components $141.0 5.7% 2.9%
Ships, small craft, pontoons, and floating docks $48.5 7.5% 2.1%
Guided missiles $38.0 8.1% 3.5%
Other $1,111.6 13.6% 6.8%
Total $2,600.0 16.5% 8.7%
Source: Federal Procurement Data System, accessed October, 2010 via www.usaspending.gov.
Note: Contract spending in the last month and week of the fiscal year by selected 2-digit product or service code, inflation
adjusted to 2009 dollars using the CPI-U. Categories jointly account for 57.2 percent of total spending.

CHAPTER 4. YEAR-END SPENDING 190
Table 4.4: Summary Statistics: Major I.T. Projects as of March, 2010
IT spending
IT projects
Millions Percent Count Percent
Total $129,729 100.0% 686 100.0%
Agency
Agency for International Development $265 0.2% 3 0.4%
Agriculture $1,864 1.4% 33 4.8%
Commerce $11,042 8.5% 46 6.7%
Corps of Engineers $4,012 3.1% 11 1.6%
Defense $14,889 11.5% 46 6.7%
Education $1,407 1.1% 25 3.6%
Energy $4,914 3.8% 26 3.8%
Environmental Protection Agency $3,166 2.4% 20 2.9%
General Services Administration $2,162 1.7% 25 3.6%
Health and Human Services $8,990 6.9% 64 9.3%
Homeland Security $13,068 10.1% 70 10.2%
Housing and Urban Development $1,605 1.2% 10 1.5%
Interior $4,557 3.5% 39 5.7%
Justice $4,376 3.4% 15 2.2%
Labor $2,434 1.9% 34 5.0%
National Aeronautics and Space Administration $9,722 7.5% 22 3.2%
National Archives and Records Administration $649 0.5% 8 1.2%
National Science Foundation $374 0.3% 6 0.9%
Nuclear Regulatory Commission $515 0.4% 16 2.3%
Office of Personnel Management $497 0.4% 7 1.0%
Small Business Administration $269 0.2% 9 1.3%
Smithsonian Institution $58 0.0% 9 1.3%
Social Security Administration $1,236 1.0% 13 1.9%
State $3,705 2.9% 13 1.9%
Transportation $12,514 9.6% 42 6.1%
Treasury $4,921 3.8% 41 6.0%
Veterans Affairs $16,521 12.7% 33 4.8%
Year of origination
1981 $2,706 2.1% 1 0.1%
1991 $61 0.0% 1 0.1%
1992 $322 0.2% 1 0.1%
1993 $409 0.3% 2 0.3%
1994 $155 0.1% 2 0.3%
1996 $3,050 2.4% 7 1.0%
1997 $1,430 1.1% 3 0.4%
1998 $2,891 2.2% 5 0.7%
1999 $2,814 2.2% 10 1.5%
2000 $2,855 2.2% 15 2.2%
2001 $8,463 6.5% 17 2.5%
2002 $12,577 9.7% 32 4.7%
2003 $13,860 10.7% 60 8.7%
2004 $12,818 9.9% 87 12.7%
2005 $13,529 10.4% 95 13.8%
2006 $16,169 12.5% 126 18.4%
2007 $17,935 13.8% 121 17.6%
2008 $14,176 10.9% 75 10.9%
2009 $3,508 2.7% 26 3.8%
Source: I.T. Dashboard data, accessed March, 2010 via http://it.usaspending.gov
Notes: Major I.T. investments by federal agency and year of origination, inflation adjusted to 2009 dollars using the CPI-U.

CHAPTER 4. YEAR-END SPENDING 191
Table 4.5: Summary Statistics: Quality Indexes and Project Characteristics for Major
I.T. Projects
Mean Std. Dev. Min Max
Planned cost (millions) 189.11 447.06 0.10 4770.89
Overall rating 7.07 2.30 0.00 10.00
Rating subindexes
CIO evaluation 3.95 0.94 1.00 5.00
Cost rating 8.72 2.52 0.00 10.00
Cost overrun 5.25 1.49 0.00 10.00
Schedule rating 8.43 3.09 0.00 10.00
Count Percent
Investment phase
Full-Acquisition 59 8.6%
Mixed Life Cycle 304 44.3%
Multi-Agency Collaboration 29 4.2%
Operations and Maintenance 278 40.5%
Planning 16 2.3%
Service group
Management of Government Resources 124 18.1%
Missing 2 0.3%
Service Types and Components 125 18.2%
Services for Citizens 344 50.2%
Support Delivery of Services to Citizen 91 13.3%
Line of business
Administrative Management 15 2.2%
Controls and Oversight 12 1.8%
Defense and National Security 30 4.4%
Disaster Management 20 2.9%
Economic Development 9 1.3%
Education 16 2.3%
Energy 5 0.7%
Environmental Management 32 4.7%
Financial Management 81 11.8%
General Government [CA] 45 6.6%
General Science and Innovation 22 3.2%
Health 55 8.0%
Homeland Security 40 5.8%
Human Resource Management 24 3.5%
Income Security 17 2.5%
Information and Technology Management 85 12.4%
International Affairs and Commerce 7 1.0%
Law Enforcement 12 1.8%
Natural Resources 16 2.3%
Planning and Budgeting 8 1.2%
Public Affairs 13 1.9%
Revenue Collection 8 1.2%
Supply Chain Management 25 3.6%
Transportation 45 6.6%
Workforce Management 5 0.7%
Other 39 5.7%
Total 686 100.0%
Source: I.T. Dashboard data, accessed March, 2010 via http://it.usaspending.gov
Notes: Planned total cost is inflation-adjusted to 2009 dollars using the CPI-U. Overall rating is a quality index that combines that
CIO evaluation, cost rating, and scheduling rating subindexes (see text for details). It takes values from 0 to 10, with 10 being the
best score. The CIO evaluation is the agency CIO's assessment of project quality. It takes values from 1 to 5, with 5 being the best.
The cost rating is based on the absolute percent deviation between the planned and actual cost of the project. The cost overrun is a
non-absolute measure that assigns over-cost projects the lowest scores. The schedule rating is based on the average tardiness of
the project. The cost and schedule indexes takes values from 0 to 10, with 10 being the best. The line of business other category
combines all categories with 4 or fewer projects.

CHAPTER 4. YEAR-END SPENDING 192
Table 4.6: Ordered Logit Regressions of Overall Rating on Last Week and Controls
Odds ratio of higher overall rating
(1) (2) (3) (4)
Last week 0.26 0.46 0.30 0.18
(0.07) (0.14) (0.10) (0.07)
Year FE X X X
Agency FE X X
Project characteristics FE X
N 671 671 671 671
Source: I.T. Dashboard data, accessed March, 2010 via http://it.usaspending.gov
Notes: Odds ratios from ordered logit regressions. Overall rating is a quality index that
combines that CIO evaluation, cost rating, and scheduling rating subindexes (see text for
details). It takes values from 0 to 10, with 10 being the best score. Project characteristics are
fixed effects for investment phase, service group, and line of business (see Table 5).
Observations weighted by inflation-adjusted spending. Standard errors in parentheses.

CHAPTER 4. YEAR-END SPENDING 193
Table 4.7: Alternative Overall Ratings Specifications
Odds ratio of higher overall rating from ordered logit Coefficients from linear model
Winsorized
Heckman
Contracts < $62M Contracts ≥ $62M Unweighted weights OLS selection model
(1) (2) (3) (4) (5) (6)
Last week 0.60 0.18 0.56 0.37 -1.00 -1.57
(0.23) (0.11) (0.14) (0.12) (0.39) (0.64)
Year FE X X X X X
Agency FE X X X X X X
Project characteristics X X X X X X
Weighted by spending X X X X X
! -0.87
(0.85)
R-squared 0.69
N 335 336 671 671 671 3,803
Source: I.T. Dashboard data, accessed March, 2010 via http://it.usaspending.gov and 2003 to 2010 Exhibit 53 reports, available
at http://www.whitehouse.gov/omb/e-gov/docs/.
Notes: Columns (1) to (4) show odds ratios from ordered logit regressions. Columns (1) and (2) split the sample at the median
value. Column (3) shows odds ratios from an unweighted regression. Column (4) Winsorizes the spending weight at $1 billion
(96th percentile).Columns (5) and (6) show regression coefficients from linear regressions. Column (5) reports coefficients from
a simple OLS regression. Column (6) reports coefficient from a Heckman selection model with a linear second stage. In this
regression the sample is all major I.T. projects recorded in the Exhibit 53 reports. The excluded variable in this selection model
is the year of project origination. The coefficient ! is the implied coefficient on the inverse Mill's ratio selection term. Overall
rating is a quality index that combines that CIO evaluation, cost rating, and scheduling rating subindexes (see text for details).
It takes values from 0 to 10, with 10 being the best score. Project characteristics are fixed effects for investment phase, service
group, and line of business (see Table 5). Observations weighted by inflation-adjusted spending unless otherwise mentioned.
Robust standard errors in parentheses.

CHAPTER 4. YEAR-END SPENDING 194
Table 4.8: Difference-in-Differences Estimates of Overall Rating on Justice and Last
Week
OLS Estimates
(1) (2) (3) (4) (5) (6)
Justice X last week 3.54 2.29 2.85 2.36 2.251 2.49
(1.19) (1.16) (0.75) (0.65) 0.593 0.898
Last week -1.91 -1.06 -0.93 -0.99 -0.814 -0.468
(1.10) (0.82) (0.48) (0.39) 0.391 0.238
Justice 0.06 -0.59 -3.33 -3.88 -4.022 -2.028
(0.51) (0.49) (0.47) (0.59) 0.578 1.01
Year FE X X X X X
Agency FE X X X X
Project characteristics X X X
Weighted by spending X X X X Winsorized*
R-squared 0.06 0.22 0.58 0.68 0.60 0.48
N 686 686 686 686 686 686
Source: I.T. Dashboard data, accessed March, 2010 via http://it.usaspending.gov
Notes: Coefficients from OLS regressions of overall rating on fully interacted Justice and last week
indicators and controls. Overall rating is a quality index that combines that CIO evaluation, cost rating,
and scheduling rating subindexes (see text for details). It takes values from 0 to 10, with 10 being the
best score. Project characteristics are fixed effects for investment phase, service group, and line of
business (see Table 5). Robust standard errors in parentheses.
*Spending weight Winsorized at $1 billion (96th percentile).

CHAPTER 4. YEAR-END SPENDING 195
Table 4.9: First Week Contract Spending for Selected Product or Service Codes,
Pooled 2004 to 2009 FPDS
Spending First week
(billions) (percent)
Leases
Lease or rental of facilities $29.2 26.2%
Lease or rental of equipment $5.4 13.1%
Service contracts
Utilities and housekeeping services $73.7 11.1%
Medical services $68.8 11.3%
Transportation, travel and relocation services $39.3 15.5%
Social services $5.5 9.3%
Other $2,378.1 3.1%
Total $2,600.0 4.0%
Source: Federal Procurement Data System, accessed October, 2010 via www.usaspending.gov
Note: Percent of contract spending in the last month and week of the fiscal year by selected 2digit
product or service code, inflation adjusted to 2009 dollars using the CPI-U. Categories
account for about 8.5% of spending but 29.7% of first week spending.

CHAPTER 4. YEAR-END SPENDING 196
Table 4.10: Ordered Logit Regressions of Subindexes on Last Week and Controls
Odds ratio from ordered logit
Evalutation by Agency Cost Cost Schedule
CIO
rating overrun rating
(1) (2) (3) (4) (5)
Last week of September 0.14 0.16 0.80 0.74 1.15
(0.06) (0.07) (0.36) (0.30) (0.66)
Cost and schedule rating X
Agency FE X X X X X
Year FE X X X X X
Project characteristics X X X X X
Weighted by spending X X X X X
N 671 671 671 671 671
Source: I.T. Dashboard data, accessed March, 2010 via http://it.usaspending.gov
Notes: Odds ratios from ordered logit regressions. The CIO evaluation is the agency CIO's
assessment of project quality. It takes values from 1 to 5, with 5 being the best. The cost
rating is based on the absolute percent deviation between the planned and actual cost of
the project. The cost overrun is a non-absolute measure that assigns over-cost projects the
lowest scores. The schedule rating is based on the average tardiness of the project. The
cost and schedule indexes takes values from 0 to 10, with 10 being the best. Project
characteristics are fixed effects for investment phase, service group, and line of business
(see Table 5). Standard errors in parentheses.

CHAPTER 4. YEAR-END SPENDING 197
Table 4.11: Year-End Contract Characteristics Regressions
Dependent Variable:
Cost-
Non-competitive Costreimbursement
or
Non-competitive One bid
or one bid reimbursement T&M/LH
T&M/LH
(1) (2) (3) (4) (5) (6)
Last week -0.002 0.017 0.010 -0.032 0.004 -0.028
(0.002) (0.002) (0.002) (0.002) (0.001) (0.002)
Year FE X X X X X X
Agency FE X X X X X X
Product or service code FE X X X X X X
R-squared 0.41 0.31 0.30 0.52 0.21 0.51
N 402,400 402,400 402,400 402,400 402,400 402,400
Mean of dependent variable 28.7% 20.0% 48.7% 30.0% 5.5% 35.5%
Source: Federal Procurement Data System, accessed October, 2010 via www.usaspending.gov
Note: Table shows coefficients from linear probability model regressions of contract type and competition type indicators on last week
and controls. To facility the analysis, the data is aggregated to the level of the covariates and the regressions are weighted by
inflation-adjusted spending in each cell.

CHAPTER 4. YEAR-END SPENDING 198
Table 4.12: Percent of Projects in I.T. Dashboard Data
Spending Projects
All I.T Dashboard
% in I.T.
Dashboard All I.T Dashboard
% in I.T.
Dashboard
Year of origin
≤ 2001 $68,460 $14,538 21.2% 813 48 5.9%
2002 $114,668 $12,848 11.2% 1,018 61 6.0%
2003 $115,286 $51,004 44.2% 653 113 17.3%
2004 $53,151 $10,309 19.4% 467 71 15.2%
2005 $35,027 $16,456 47.0% 250 56 22.4%
2006 $13,023 $5,172 39.7% 191 77 40.3%
2007 $61,953 $55,665 89.8% 248 183 73.8%
2008 $19,864 $19,752 99.4% 135 127 94.1%
2009 $498 $491 98.7% 16 13 81.3%
2010 $285 $273 95.5% 13 10 76.9%
Total $482,215 $186,509 38.7% 3,803 759 20.0%
Source: I.T. Dashboard data, accessed March, 2010 via http://it.usaspending.gov and 2003 to 2010
Exhibit 53 reports, available at http://www.whitehouse.gov/omb/e-gov/docs/.
Notes: All spending and projects are totals from agency Exhibit 53 reports. I.T. Dashboard spending and
projects are totals in the I.T. Dashboard dataset (including projects dropped from the baseline sample due
to missing values). Spending values inflation-adjusted using the CPI-U.