Mostly Harmless Econometrics

Joshua Angrist and Jorn-Steffen Pischke wrote this as a practical guide to causal inference for applied economists, and it changed how empirical economics is taught. The organising idea is that the credible identification of causal effects, not the sophistication of the estimator, determines whether a study means anything.

The framework is the potential outcomes model: for each unit there exists an outcome under treatment and an outcome without it, only one is observed, and the entire problem is constructing a credible comparison. The authors work through the research designs that achieve this. Randomised experiments supply the benchmark. Regression is treated as a way of approximating the conditional expectation function and as valid for causal purposes only under selection on observables. Instrumental variables are presented through the local average treatment effect framework, which specifies that the estimate applies to the compliers and not to the whole population. Differences-in-differences requires the parallel trends assumption, which the authors insist must be argued rather than asserted. Regression discontinuity exploits arbitrary thresholds in assignment rules.

The book is opinionated, funny in places, and dismissive of methods it considers unidentified, particularly structural modelling and much of time series macroeconometrics. It is aimed at graduate students and assumes regression at a solid level. The subsequent literature has qualified several of its recommendations, most notably on differences-in-differences with staggered treatment timing, where the two-way fixed effects estimator the book presents has been shown to be biased. The design-based approach it advocates is now standard, and Angrist shared a Nobel prize in 2021 for the work behind it.