Multilevel Zero-Inflation and Hurdle Models for Count Data: Making Sense of Sampling and Structural Zeros

• In Event: Fresh Perspectives on Estimating Complex Multilevel Models: Counts, Mixtures, Interactions, and Machine Learning

Abstract

Many applications involving counts as outcomes include excess zeros, relative to the expected counts given an assumed underlying distribution even after adjusting for violations in dispersion assumptions. Excess zeros can pose a challenge to analysis and interpretation, particularly when the data are clustered or multilevel. Zeros are often not all the same, and different approaches are used to model counts depending on understanding of how the zeros arise. Using data from a student self-report social environmental assessment (the School Success Profile), we demonstrate two multilevel alternatives to the treatment of excess zeros: Hurdle (H) and Zero-Inflation (ZI) models. Our comparison clarifies use and interpretation of how these models approach structural (H) or combinations of structural and sampling zeros (ZIP).

Authors

• Ann A. O'Connell, Rutgers University

• Nivedita Bhaktha, Indian Institute of Technology Kanpur

• Latif Kadir, The Ohio State University

• Natasha K. Bowen, The Ohio State University