USING PROPENSITY SCORES IN QUASI-EXPERIMENTAL DESIGNS

William M. Holmes

STATA COMMANDS FOR PROPENSITY USE

Shenyang Guo and Mark W. Fraser (2010) have written an excellent book on propensity score analysis that provides many Stata commands. While their treatment is more technical than this book, the Stata commands are quite useful. The support site for their book is: http://ssw.unc.edu/psa/.  The Stata Journal article by Becker and Ichino (2002) provides a good overview of propensity score use and extensive commands with Stata. It can be accessed at http://www.stata-journal.com/sjpdf.html?articlenum=st0026.

IMBALANCE ASSESSMENT

            Imbalance may be assessed with the ANOVA command or ttest.

            . anova confounder1 treatment

            .ttest confounder 1 treatment

PROPENSITY ESTIMATION

            Propensities may be estimated with logistic, probit, or linear regression or with discriminant function analysis commands when combined with the predict command.

. logistic confounder1 confounder2 confounder3 i.treatment

. predict propen, xb

. probit treatment confounder1 confounder2 confounder3

. predict propen

. regress treatment confounder1 confounder2 confounder3

. predict propen

. candisc confounder1 confounder2 confounder3, group(treatment)

. predict propen

MATCHING

            The psmatch2 program provides a means for propensity score matching within Stata.  Alternative matching programs may be accessed with the R interface for Stata.

            Exact and Coarsened Exact

            See psmatch2 within Stata

            Nearest Neighbor and Caliper

            See psmatch2 within Stata

            1-Many

            See psmatch2 within Stata

            Optimized, Full, or Genetic

            To do optimized, full, or genetic matching within Stata, the R interface must be used to access the corresponding R programs.

STRATIFYING

            Creating propensity strata can be done with the recode command.

recode propen 0/.19999=1 .2/.39999=2 .4/.59999=3 .6/.79999=4 .8/.99999=5, gen(propenstrata)

REGRESSION AND THE GENERAL LINEAR MODEL

            Regression and General Linear Model estimation may be done with the regress and glm commands.

             . regress outcome1  i.treatment

            . glm outcome1 i.treatment, family(gaussian) link(identity)

TWO-STAGE LEAST SQUARES

            2SLS regression can be done with the ivregress command.

. ivregress 2sls outcome1(treatment = instrument1) , vce(robust)

SAMPLE WEIGHTING

Sample weighting is done within the analytical commands using   pweight. For example with a logit analysis, the command would be:

logit treatment confounder1 confounder2 [pweight=ipa]

WEIGHTED LEAST SQUARES

Weighted Least Squares may be done within analytical program (see sample weighting above) or with the wls0 command.

.wls0 outcome treatment, wvar(ipa) type(abse) noconst graph

GENERALIZED LINEAR MODEL

            . glm income educ jobexp i.black, family(gaussian) link(identity)

MISSING DATA ANALYSIS

            Analysis of missing data can combine mdesc with mvpatterns or with the misstable command.

            mdesc confounder1 confounder2 confounder3        

            mvpatterns confounder1 confounder2 confounder3

            . misstable summarize

IMPUTATION OF MISSING DATA

Multiple imputation can be done with the mi impute command. Variables to be imputed must first be registered.  The example below imputes 5 values with a random seed of 123. Estimation of imputed data in Stata may be done with the mi estimate command.

            . mi register imputed confounder1 confounder2

            . mi impute regress confounder1  confounder2 confounder3, add(5) rseed(123)

            . mi estimate: logit treatment confounder1 confounder2 confounder3

return to Home

Revised 1/31/2013