Regression's weaknesses and strengths: A reply to Gerring

Assumptions are the rule, not the exception, in both descriptive and causal inference in the social sciences. This fact has long been used as a defense of the specific families of assumptions used to make causal inferences on the basis of regression-type models (Freedman 2004: 195). Yet the defense is weak.  Inferences differ in terms of the strength, complexity, plausibility, and testability of the assumptions they require. On all of these fronts, regression-type analysis of observational data often performs so poorly that it is difficult to give the results a persuasive causal interpretation. In what follows, I will make this argument by showing how hard it can be to assign causal interpretations to regression models that show either unstable or stable results across the range of models that the discipline considers plausible, as well as the challenges involved with drawing causal conclusions from either the unconditional or the conditional analysis of quantitative observational data. For these reasons, I disagree with Gerring’s argument that the regression analysis of messy data is a good default option for social scientists; instead, it is a weak default, and one far more suited to describing initial facts, discovering puzzles, and characterizing patterns than to causal inference. I then argue that the importance of difficult-to-research questions is a weak defense of the status quo, and I conclude by briefly sketching the valuable but carefully delimited role that regression should play in our research.