Social scientists routinely evaluate the effect of economic policies. For instance, what is the effect of raising the minimum wage on employment? Angrist and Pischke (2010) argue that quantitative social sciences, and in particular applied economics, have recently experienced a “credibility revolution'”: by switching to transparent methods to evaluate policies, social scientists have increased the credibility of their findings, and the impact of their work. However, the credibility revolution is not complete. Two-way fixed effects regression (TWFE) is a policy-evaluation method used in as many as 26% of the most-highly cited papers recently published in the American Economic Review. Social scientists have long believed that this method is similar to a difference-in-differences (DID) method. DID yields an unbiased estimator of the policy’s effect under a parallel trends assumption: without the treatment, treated and control units would have experienced the same outcome evolution. Parallel trends is an appealing assumption, because it is partly testable: one can check whether prior to treatment, treated and control units were experiencing parallel outcome evolutions. However, recent papers have shown that TWFE regressions also rely on the assumption that the treatment effect is constant, both across units and over time. This constant-effect assumption is implausible in most applications. For instance, the effect of the minimum wage on employment could be more negative in locations with a less qualified workforce, and that effect could also change over time. This constant-effect assumption is also much harder to test than parallel trends. Therefore, the project proposes a series of so-called “generalized DID estimators” (GDID), that only rely on parallel trends assumptions and do not rely on the constant effect assumption, and that can be used in virtually all the applications where two-way fixed effects regressions have been used. The project also proposes software implementation of those new estimators, in Stata and R, the two statistical software most commonly used by quantitative social scientists. Thus, it hopes to provide easily accessible tools that quantitative social scientists can use to produce more credible policy evaluations.
