Dr. Wes Bonifay
College of Education & Human Development
University of Missouri
Title: The Hidden Depths of Complexity in Statistical Modeling
Abstract: It is no secret that statistical model complexity affects goodness-of-fit to the observed data (i.e., accommodation) and generalizability to future data (i.e., prediction). Less obvious is that typical methods of addressing complexity fail to tell the whole story: Familiar statistics such as BIC account for parametric complexity (due to the presence of many model parameters) while configural complexity (due to the particular configuration of model variables) remains undetected. To obscure matters further, a model may be simultaneously contaminated by both sources of complexity, and the knowledge that can be gained from application of such a model will be severely limited. In this talk, I will present several methods that have been recently developed to detect and counteract the effects of configural complexity, and to thereby improve inference, generalizability, and replication in model-based research.
Wes Bonifay’s research interests are in the area of psychological measurement, with particular focus in item response theory and model evaluation. He has published a number of quantitative research articles on psychometric topics such as dimensionality assessment, subscale analysis, and model complexity. He has also collaborated with substantive educational and psychological researchers, applying item response theory and structural equation models to better understand certain issues in school psychology, psychiatric treatment, and cognitive behavioral therapy.