A PRIISM Seminar by University of Delaware's Sanford Student

Join PRIISM and Dr. Sanford Student to learn how to apply moderated nonlinear factor analysis to causal inference in randomized controlled trials and distinguish genuine intervention effects from measurement bias.

Abstract

In principle, item-level data allow researchers to estimate causal effects on latent variables, sidestepping issues of measurement error inherent in the use of observed scores as outcomes. In doing so, the researcher introduces a set of assumptions around how the items relate to the underlying latent variable of interest in the treatment versus control group, and these assumptions are easily violated. For example, a subset of items on the outcome measure might be overly aligned to the intervention, such that effects on these items are larger than effects on the rest of the items. Ignoring this can lead to misleading estimates of treatment effects. I will discuss the idea of causal parameter moderation, which applies moderated nonlinear factor analysis to causal inference in randomized controlled trials. MNLFA combines item response theory and structural equation modeling to model relations between an arbitrary number of covariates and the parameters of a latent variable model, including both the mean and variance of the latent variable (known in the educational measurement literature as impact) as well as the parameters that relate it to the items (known as differential item functioning). This allows the researcher to evaluate effects of the treatment, and possible heterogeneity with respect to person-side covariates, on all aspects of the latent variable model. In doing so, researchers with item-level data can gather richer information about just what it is that an intervention does, and learn about potential threats to the generalizability of their findings.