Detecting Overdispersion in Discrete Data

• In Event: AERA Roundtable Session Saturday 11:40 am Four Seasons Ballroom 4
  In Roundtable Session: Complexity in Data: Simulations and Applications in Complex, Multivariate, Multilevel, and Nonnormal Models (Table 2)

Abstract

Overdispersion, where observed variance exceeds model expectations, can compromise statistical inference in discrete data analysis. This study develops a unified framework for defining and detecting overdispersion in count, binary, and ordinal data. We highlight the unique characteristics and benchmark models for each data type, extending existing detection methods for count data to binary and ordinal data. Using publicly available data from the 2017-18 School Survey on Crime and Safety, we demonstrate various tests and visualization techniques to identify overdispersion. Our results emphasize the necessity of addressing overdispersion to ensure accurate and robust statistical models. This study provides practical recommendations for researchers, enhancing the methodological rigor in categorical data analysis.

Authors

• Nivedita Bhaktha, Indian Institute of Technology Kanpur

• Ann A. O'Connell, Rutgers University

• Latif Kadir, The Ohio State University