Change scores and baseline adjustment: splitting the difference (in differences)

Abstract A lively debate exists on how best to incorporate repeated measurements of an outcome in a statistical analysis. Two popular approaches are change-score analyses (the difference in baseline and follow-up outcome measurements are regressed on the exposure) and baseline adjustment (in which baseline outcomes are included as a covariate in a regression model). In this article, we compare both through the lens of the causal inference literature. We draw in particular on the connections between change-score analyses and ‘difference-in-differences’—a popular approach for inferring causal effects in economics and social sciences, which has received less attention in epidemiology. We make practical recommendations for which approach to use, based on one’s knowledge or belief about the confounding mechanism.