Diagnostics for Bayesian Hierarchical Models of Response Time Data

• Activity Number: 327
• Type: Topic Contributed
• Date/Time: Tuesday, August 3, 2010 : 10:30 AM to 12:20 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Peter F. Craigmile and Mario Peruggia*+ and Trisha Van Zandt
• Companies: The Ohio State University and The Ohio State University and The Ohio State University
• Keywords: autoregressive errors ; generalized residuals ; goodness-of-fit ; predictive distribution ; prequential analysis ; time series

Abstract

Human response time data are used in experimental psychology to evaluate theories of mental processing. The sequential nature of the experiments yielding the data produces modeling challenges due to trends and serial dependencies. Extreme observations must also be considered. We present goodness-of-fit diagnostics that assess how well a model can capture certain essential features of the data. We describe a method based on the introduction of artificial autoregressive error structures at various levels of the model hierarchy and two methods that focus on the predictive performance of the model. The first predictive method compares the marginal and higher order serial dependence properties of posterior predictive samples with those of the observed data. The second predictive method is a modified prequential approach which examines the empirical distribution of the generalized residuals.