Constructing generalizable geographic natural experiments

Abstract

A natural experiment is a real-world situation that generates as-if random or haphazard assignment to treatment. Geographic or administrative boundaries can be exploited as natural experiments to construct treated and control groups. Previous research has demonstrated that matching can help enhance these designs by reducing imbalances on observed covariates. An important limitation of this empirical approach, however, is that the results are inherently local. While the treated and control groups may be quite similar to each other, they could be substantially different from the target population of interest (e.g., a country). We propose a simple design inspired by the idea of template matching to construct generalizable geographic natural experiments. By matching our treated and control groups to a template (i.e., the target population), we obtain groups that are similar to the target population of interest and to each other, which can increase both the internal and external validity of the study.

Immigration and Nationalistic Attitudes: Panel Evidence from Chile

“…22 Unfortunately, the unmatched sample is too small and the groups too different to include more constraints such as constructing a representative matched sample (Bennett et al, 2019;Kuffuor et al, 2022). However, cardinality matching is particularly good at addressing problems of limited overlap in small samples (Visconti and Zubizarreta, 2018).…”