Assessing the Robustness to an Unobserved Confounder of the Average Treatment Effect on the Treated Estimated by Propensity Score Matching

Abstract

One of the serious drawbacks of observational studies is the selection bias caused by the selection process to the treatment group. Propensity Score Matching (PSM), which allows for the reduction of the selection bias when estimating the average treatment effect on the treated (ATT), is a method recommended for the evaluation of projects and programmes co-financed by the European Union. PSM relies on a strong assumption known as the Conditional Independence Assumption (CIA) which implies that selection into the treatment group is based on observable variables, and all variables influencing both the selection process and outcome are observed by the researcher. If this does not hold, the estimated effect may be not so much the result of the treatment as of the lack of balance of an unobserved confounder, which affects both the selection process and the outcome. Rosenbaum’s sensitivity analysis allows researchers to determine how strong the impact of such a potential unobserved confounder on selection into treatment and the outcome must be to undermine conclusions about ATT estimated by PSM. Rosenbaum’s primal and simultaneous approaches are applied in the paper to assess robustness to an unobserved confounder of the net effect of internships for unemployed young people with a maximum age of thirty-five (estimated with PSM) organized by one of the biggest district employment offices in Malopolska.