27.08.2015 Views

Using Instrumental Variable Analyses to Inform ... - The Lewin Group

Using Instrumental Variable Analyses to Inform ... - The Lewin Group

Using Instrumental Variable Analyses to Inform ... - The Lewin Group

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

Anirban Basu, PhD<br />

<strong>The</strong> University of Chicago &<br />

<strong>The</strong> National Bureau of Economic Research<br />

<strong>Lewin</strong> <strong>Group</strong>, Center for Comparative Effectiveness Research<br />

CER Symposium: Responding <strong>to</strong> the National CER Agenda:<br />

Evolving Data Sources and Analytics<br />

June 15, 2010


• Review basic principles of IV analysis<br />

• Present a graphical representation of counterfactual<br />

outcomes and treatment effects<br />

• Study how IV identifies casual treatment effects –<br />

conceptual discussions rather than statistics<br />

• Complications in the presence of heterogeneity of<br />

treatment effects<br />

• Local <strong>Instrumental</strong> <strong>Variable</strong>s (LIV) approach<br />

• An empirical example on evaluating breast cancer<br />

treatments


• Identify instrumental variables that:<br />

• Are correlated with treatment receipt but are uncorrelated<br />

with outcomes.<br />

• Can effectively randomize subjects across treatment arms<br />

achieve equal distribution of both observed and<br />

unobserved confounders across treatment groups.<br />

• Address both overt and hidden biases in estimating the<br />

average treatment effect.<br />

• Split the variation in treatment variable in<strong>to</strong> an exogenous<br />

part and an endogenous part ‐ only the exogenous part is<br />

used <strong>to</strong> estimate treatment effect


Outcome<br />

0 1 2 3 4<br />

Potential Outcomes<br />

With Control Treatment<br />

Levels of Unobserved confounder


Outcome<br />

0 1 2 3 4<br />

Potential Outcomes<br />

With New Treatment<br />

Potential Outcomes<br />

With Control Treatment<br />

Levels of Unobserved confounder


Outcome<br />

0 1 2 3 4<br />

AVERAGE TREATMENT<br />

EFFECT<br />

Potential Outcomes<br />

With New Treatment<br />

Potential Outcomes<br />

With Control Treatment<br />

Levels of Unobserved confounder


Outcome<br />

0 1 2 3 4<br />

Levels of Unobserved confounder


Outcome<br />

0 1 2 3 4<br />

Patients Select <strong>to</strong> Receive<br />

New Treatment<br />

Levels of Unobserved confounder<br />

Patients Select <strong>to</strong> Receive<br />

Control Treatment


Outcome<br />

0 1 2 3 4<br />

Levels of Unobserved confounder<br />

An IV effect<br />

= ATE


Outcome<br />

0 1 2 3 4<br />

Another IV effect<br />

= ATE<br />

Levels of Unobserved confounder


Outcome<br />

0 1 2 3 4<br />

Levels of Unobserved confounder


Outcome<br />

0 1 2 3 4<br />

Levels of Unobserved confounder


Outcome<br />

0 1 2 3 4<br />

Effect of IV2<br />

Effect of IV3<br />

Effect of IV1<br />

Levels of Unobserved confounder<br />

Angrist, Imbens, Rubin, 1996; Brooks, Chrischilles, Scott, Chen‐Hardee, 2003


• LATE ‐ the average treatment effect for<br />

individuals who would change their treatment<br />

choice when instrument level moves<br />

• But who are these “marginal” patients?<br />

• At what margins will my policy induce changes in<br />

treatment choices?<br />

• IV effects (combinations of LATEs) are<br />

misleading as they can be arbitrarily weighted.<br />

Heckman, 1997; McClellan, Newhouse, 1998; Harris, Remler, 1998; Heckman, Vytlacil, 1999


• Mirrors control function and selection models in<br />

economics<br />

• Requires an explicit choice model<br />

• LIV is a method <strong>to</strong> estimate the Marginal Treatment<br />

Effects:<br />

• treatment effect among patients who are at the margin of<br />

choice – the combination of IVs & observed confounders<br />

balances the effects of unobserved confounders on choice<br />

• a small perturbation in IV will change the treatment choice<br />

for these patients<br />

Heckman and Vytlacil, 1999, 2000, 2001; Heckman, Urzua, Vytlacil, 2006


Outcome<br />

0 1 2 3 4<br />

TE for combination<br />

of margins<br />

TE for a margin<br />

Levels of Unobserved confounder<br />

TE for a margin<br />

Heckman and Vytlacil, 1999, 2000, 2001


Outcome<br />

0 1 2 3 4<br />

Levels of Unobserved confounder


• Identify the profile of MTE over the entire support of<br />

the [1‐Pr(Trt Choice)] ~ scaled version of<br />

unobserved confounders at the margins<br />

• Compute ATE, TT, TUT and other policy effects by<br />

simple weighting of the MTEs.<br />

• Can identify situations where current data wont<br />

allow <strong>to</strong> estimate an ATE.<br />

• Can identify situations where complications due <strong>to</strong><br />

heterogeneity can be ignored.<br />

Heckman and Vytlacil, 1999, 2000, 2001


MTE(UD) ( in $)<br />

-100000 0 100000 200000 300000<br />

MTE(UD) Profile and 95% CI with IV = NORTH, FEEDIF<br />

0 .2 .4 .6 .8 1<br />

UD<br />

Propensity <strong>to</strong> choose BCSRT due <strong>to</strong> unobserved confounders <br />

Linear IV approach:<br />

IV(North): $45K<br />

IV(Feedif): $13K<br />

Local IV approach:<br />

ATE(emp): $41K<br />

ATE(extrp): $88K<br />

TT: $52K<br />

Basu, Heckman, Navarro, Urzua, 2007


• In IV analyses, the focus has been more on trying <strong>to</strong><br />

find good instrument and less on interpretation of<br />

results.<br />

• In health, unique opportunity <strong>to</strong> exploit providerlevel<br />

and system level variations that are not linked<br />

<strong>to</strong> outcomes.<br />

• Careful attention is necessary <strong>to</strong> interpret IV results<br />

in the face heterogeneity of treatment effects<br />

• Local instrumental variable approaches can help in<br />

identifying the proper treatment effects relevant for<br />

policy making.


V <br />

( Z , X ) U E ( U )=0<br />

D 1( V<br />

<br />

0),<br />

V V V<br />

V ( Z, X) U E( U D 1( V 0),<br />

V V V )=0<br />

Population level:<br />

P( z) Pr( D 1| Z z )<br />

Pr( U ( z ))<br />

V<br />

V<br />

1 F ( ( z ))<br />

<br />

U<br />

V<br />

F ( ( z )) = 1 - P( z )<br />

U<br />

V<br />

V<br />

V<br />

Individual level:<br />

D 1( V 0)<br />

1( U ( z ))<br />

V<br />

V<br />

1( F ( U ) F ( ( z )))<br />

U V U V<br />

V<br />

1( F ( U ) 1 P( z ))<br />

U<br />

V<br />

V<br />

<br />

1( U 1 P( z))<br />

D<br />

V<br />

where F ( U ) U ~ Uniform[0,1]<br />

U V D<br />

V


D 1( V 0)<br />

=<br />

=<br />

Let<br />

D 1( V 0) 1( U 1 P( z))<br />

D<br />

S( z) 1 P( z)<br />

UD < S(z)<br />

=> D=0<br />

UD > S(z)<br />

=> D=1<br />

Treatment Effects<br />

S(z)<br />

0 1<br />

UD


D 1( V 0)<br />

=<br />

=<br />

Let<br />

D 1( V 0) 1( U 1 P( z))<br />

D<br />

S( z) 1 P( z)<br />

UD < S(z)<br />

=> D=0<br />

UD > S(z)<br />

=> D=1<br />

UD S(z’)<br />

Treatment Effects<br />

=> D=0 => D=1<br />

S(z’)<br />

S(z)<br />

0 1<br />

UD


D 1( V 0)<br />

=<br />

=<br />

Let<br />

D 1( V 0) 1( U 1 P( z))<br />

D<br />

S( z) 1 P( z)<br />

--> S(z')

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!