Practical Issues for Estimating Causal Effects with Propensity-Score Matching: Testing Covariate Balance

• In Event: 343 - Section on Medical Sociology Roundtables (one-hour)
  In Refereed Roundtable: Table 13. Chronic Health Conditions

Abstract

Propensity-score matching is becoming a popular method for estimating causal effects in the sociological literature, yet relatively few scholars appropriately test for adequate covariate balance after the matching algorithm has been run. This oversight may, and often does, lead to biased estimates of causal effects. I use my research on the effects of parental obesity on offspring income as a working example to provide conceptual intuition and practical advice, including STATA 13® code that may be adapted to test covariate balance in any propensity-score analysis. While there is little agreement about which balance tests are necessary or sufficient, I explore common tests including comparing means and prevalences of covariates using standardized differences (“standardized bias”), comparison of higher order moments and interactions, quantile-quantile plots, side-by-side boxplots, five-number summaries, and variance ratios. The first section of the paper introduces the substantive topic and addresses several conceptual issues as a prerequisite to any causal analysis. The remainder of the paper focuses on the practical issue of estimating a propensity score model and testing covariate balance to diagnose model inadequacies.

Author

• Ian F Wall, University of Wisconsin-Madison