To address the resulting model complexity, we trim non-significant interactions between predictor variables and study membership at each stage of the model building process. This practice of model trimming is maximally conservative and designed to maintain parsimony, to provide ease of interpretation, and to support greater stability in model estimation. The main effect codes for study membership, however, we do not trim. In our pooled analyses, even our initial models include study membership as a main effect. This may be a particularly important point for testing growth models and using other longitudinal approaches in which the initial model is typically an unconditional model estimated with the goal of identifying the functional form of change over time in a given construct (i.e., in the absence of covariates). However, in the context of IDA, failure to include study membership as a covariate (and predictor of change over time) may occlude the functional form of these trajectories, with differences in the studies contributing observations to the pooled data set over time or age leading to seeming changes in trajectories that are driven not by time trends in the underlying factor but instead by the pooled data design.
