Abstract:
|
This paper attempts to synthesize "latent variable" (LV) and "empirical" approaches to regression with outcomes that must be assessed by multiple measurements. Empirical approaches would directly analyze the measured responses, perhaps aggregated. LV approaches treat measured responses as surrogates of an idealized, unobservable outcome and infer predictor associations with that outcome through a model. I review an approach for identifying how each of a LV model's statistical assumptions may be contradicted in data being analyzed, acknowledging ramifications for estimates in interpretation. Building on this, I propose an analytic method that regresses observed variables directly on predictors without incorporating LVs but retains LV analysis advantages. I argue that the method minimizes biases in estimating predictor effects as specified by an LV model. Inferential strategies are evaluated. In application to health data, the method is used to contrast distinct LV approaches to analysis. The paper seeks to provide an analytic tool that approaches the advantages of LV analysis but ameliorates concerns that model assumptions may overpower data in deriving findings.
|