On Confidence Interval Coverage of Causal Effect Estimators After Variable Selection

• In Event: Propensity Scores and Causal Inferences

Abstract

The ability of a conditioning strategy to consistently estimate an average treatment effect in an observational study depends on the ignorability assumption, which is more likely to be satisfied if the conditioning set is large. Nevertheless, there are compelling reasons not to simply condition on all available variables. Variable selection adds stochastic variability to statistical estimators that is not accounted for by classical theory. The use of sample data for both variable selection and inference worsens confidence interval coverage. We investigate the effect of variable selection on CI size and coverage for causal effect estimators. We describe and evaluate several analytical and resampling-based standard error estimators through Monte Carlo simulation, and conclude with discussion and recommendations.

Author

• Bryan Sean Keller, Teachers College, Columbia University