Improving Causal Inference with TSCS Data Using Counterfactual Estimators

• In Event: Causal Inference for Time-Series Data

Abstract

We introduce a group of treatment effects estimators in time-series cross-sectional (TSCS) settings that directly estimating counterfactuals for observations under the treatment condition. They include (1) the fixed-effect counterfactual estimator, (2) the interactive fixed-effect estimator and (3) the matrix completion estimator, which differ in the underlying models of predicting treated counterfactuals. These estimators provide more reliable estimates for treatment effects in a TSCS setting than the conventional two-way fixed effects approach when the constant treatment effect and strict exogeneity assumptions are not satisfied. Moreover, we develop two diagnostic tests—a placebo test and an equivalence test—accompanied by visualization tools to help researchers gauge the validity of the identifying assumptions of these estimators. We provide two empirical examples from political economy to illustrate their performance in applied settings. These methods and tests are implemented with the fect package in R.

Authors

• Licheng Liu, Tsinghua University

• Ye Wang, New York University

• Yiqing Xu, Stanford University