Comparing Heckit and Propensity Methods Under Conditions When There Is Evidence of Selection Into Treatment

• In Event: Poster Session 10
  In Poster Session: Division D Section 2 Poster Session 2

Abstract

This study compares different approaches to controlling for selection bias in observational studies by utilizing the probability derived from a selection model. The two general approaches we compared were Heckit treatment effect models and propensity balancing methods. The Heckit models assume that selection is related to outcomes, whereas propensity methods assume conditional independence of treatment selection from the outcome. We compared differences in the approaches on an empirical selection model of school choice. We then explored three conditions of selection (i.e., 1 of 4, 2 of 4, and 3 of 4 significant predictors) with a Monte Carlo simulation. Results indicate that Heckit models become preferable to propensity score models when there is significant prediction in the selection model.

Authors

• Paul Wesley Scott, University of Pittsburgh

• Clement A. Stone, University of Pittsburgh

• Heather J. Bachman, University of Pittsburgh

• Leanne Elliott, University of Pittsburgh