More Precise Estimation of Lower-Level Interaction Effects in Multilevel Models

In hierarchical data, the effect of a lower-level predictor on a lower-level outcome may often be confounded by an (un)measured upper-level factor. When such confounding is left unaddressed, the effect of the lower-level predictor is estimated with bias. Separating this effect into a within- and between-component removes such bias in a linear random intercept model under a specific set of assumptions for the confounder. When the effect of the lower-level predictor is additionally moderated by another lower-level predictor, an interaction between both lower-level predictors is included into the model. To address unmeasured upper-level confounding, this interaction term ought to be decomposed into a within- and between-component as well. This can be achieved by first multiplying both predictors and centering that product term next, or vice versa. We show that while both approaches, on average, yield the same estimates of the interaction effect in linear models, the former decomposition is much more precise and robust against misspecification of the effects of cross-level and upper-level terms, compared to the latter.