When the Baseline Is Not Exogenous in Random-Intercept Models

Researchers routinely adjust for a baseline score in random-intercept models, usually for efficiency or confounding control. Because the baseline is a noisy measure of the same trait as the random intercept, the two are correlated, yet standard multilevel software treats the baseline as exogenous. We examine the consequences. The baseline coefficient does not estimate the structural effect of initial level; it estimates a reliability-weighted blend, the structural effect plus a regression-to-the-mean projection. A structural equation formulation makes explicit when the structural effect itself, rather than this blend, can be identified. The focal predictor’s effect is unaffected when that predictor is uncorrelated with the trait, and otherwise depends on whether its confounding runs through the observed baseline or the latent trait.