Estimation and Inference on Treatment Effects Under Treatment-Based Sampling

TL;DR

This paper develops methods for estimating and making valid inferences about treatment effects in stratified sampling designs, accounting for unknown population shares and external validity issues.

Contribution

It introduces confidence set construction for treatment effects under unknown population shares and proposes an optimal sampling design to improve efficiency.

Findings

Confidence sets valid for a range of population shares

Method to discover external validity scope with error control

Optimal sampling design minimizes efficiency bound

Abstract

Causal inference in a program evaluation setting faces the problem of external validity when the treatment effect in the target population is different from the treatment effect identified from the population of which the sample is representative. This paper focuses on a situation where such discrepancy arises by a stratified sampling design based on the individual treatment status and other characteristics. In such settings, the design probability is known from the sampling design but the target population depends on the underlying population share vector which is often unknown, and except for special cases, the treatment effect parameters are not identified. In this paper, we propose a method of constructing confidence sets that are valid for a given range of population shares. When a benchmark population share vector and a corresponding estimator of a treatment effect parameter are given, we develop a method to discover the scope of external validity with familywise error rate control. Finally, we derive an optimal sampling design which minimizes the semiparametric efficiency bound given a population share associated with a target population. We provide Monte Carlo simulation results and an empirical application to demonstrate the usefulness of our proposals.