I am an Associate Professor and also the Director of the Quantitative Methods (QM) program within the Department of Psychology and Human Development at Vanderbilt University.
As a quantitative psychologist, my research involves evaluating, improving, and disseminating best-practices for modeling data from psychology and other social sciences. Below I provide an overview of some of my research. If you are interested in applying to our QM Ph.D. program to work with me, feel free to reach out with questions!
Psychologists are often interested in understanding individual heterogeneity in the longitudinal course of problem behaviors, such as externalizing or internalizing symptoms. In this context, psychologists commonly employ mixture models to describe and predict individual differences in change over time. This approach involves positing that individual heterogeneity in change can be explained by the existence of a mixture of latent classes (unobserved subgroups) of persons whose patterns of change differ across classes. Some of my research involves developing assessment tools for evaluating mixture models, relating mixture models to more traditional longitudinal/multilevel methods, developing methods to handle alternative kinds of missing data when fitting mixture models, studying how mixtures can be used to recover interactive relationships, and disseminating pedagogical information about mixtures to clarify how they can be interpreted in practice.
Psychology researchers frequently have other kinds of clustered (or nested) data structures as well, such as students nested within school or children nested within family. Because observations within a cluster tend to be similar to each other, clustered data violate independence assumptions of most conventional statistical models and thus necessitate the use of specialized methods, such as multilevel models. Some of my research involves developing R-squared effect size indices (e.g., here, here, here, and here) to convey practical (rather than simply statistical) significance of terms in multilevel models and adapting traditional multilevel models to handle complex nesting structures (e.g., partial nesting, wherein patients are nested within therapy groups in a study’s treatment arm, but remain non-nested in the study’s wait-list-only control arm).
Psychologists fitting structural equation models (SEMs) often have many items that are intended to measure an underlying latent construct (e.g., depression). Reviews indicate that between one fourth and two thirds of SEM applications currently use parceling, often to aid model estimability with small sample sizes. Parceling has been in use for over 50 years and entails averaging or summing subsets of items and using the resultant “parcel” scores, in lieu of the item scores, as indicators of the latent construct when fitting a SEM. Some of my research studies unintended negative consequences of this practice and provides tools (e.g., here, here, and here) to help quantify and account for these consequences.
© 2010-2021, Sonya K. Sterba