IZA DP No. 9491: New Evidence on Linear Regression and Treatment Effect Heterogeneity
superseded by IZA Discussion Paper No. 11866
It is standard practice in applied work to rely on linear least squares regression to estimate the effect of a binary variable ("treatment") on some outcome of interest. In this paper I study the interpretation of the regression estimand when treatment effects are in fact heterogeneous. I show that the coefficient on treatment is identical to the outcome of the following three-step procedure: first, calculate the linear projection of treatment on the vector of other covariates ("propensity score"); second, calculate average partial effects for both groups of interest ("treated" and "controls") from a regression of outcome on treatment, the propensity score, and their interaction; third, calculate a weighted average of these two effects, with weights being inversely related to the unconditional probability that a unit belongs to a given group. Each of these steps is potentially problematic, but this last property – the reliance on implicit weights which are inversely related to the proportion of each group – can have particularly severe consequences for applied work. To illustrate the importance of this result, I perform Monte Carlo simulations as well as replicate two applied papers: Berger, Easterly, Nunn and Satyanath (2013) on the effects of successful CIA interventions during the Cold War on imports from the US; and Martinez-Bravo (2014) on the effects of appointed officials on village-level electoral results in Indonesia. In both cases some of the conclusions change dramatically after allowing for heterogeneity in effects.