Robust estimation and the impact of outliers on causal Inference


In recent decades, the estimation of causal effects has attracted increasing professional interest. However, the success of this effort depends on the ability to generate precise estimates that can aid in the implementation of new laws and regulations. Unfortunately, extreme values or outliers, which are often found in microdata, can decrease accuracy if not taken into account properly. In this paper, we conduct a replication exercise to investigate the impact of outliers on estimated causal effects. Our analysis is based on about 92 articles that were published in 2017 in nine top economic journals and are reproducible. After replicating the results using robust estimators (MDPD and MM), we found that outliers change the magnitude of coefficients and, in some cases, even the sign. This was the case for a substantial number of the articles we examined. Taking into account the causality assertions made in the original articles, this finding appears interesting and opens the discussion for a more accurate analysis of outliers.

Work in progress (first draft available soon)