First Advisor

Merchant, William

First Committee Member

Yu, Han

Second Committee Member

Wamboldt, Frederick

Third Committee Member

Brown, Corina

Degree Name

Doctor of Philosophy

Document Type


Date Created



College of Education and Behavioral Sciences, Applied Statistics and Research Methods, ASRM Student Work

Embargo Date



Commonly in applied research settings, investigators wish to study counts of some event of interest over time, such as number of infections, or absences from school. However, in many cases, investigators may only have access to data sets in which measurements with individuals are not available at each measurement timepoint. A foundational statistical method for the study of panel data is generalized linear mixed effect modeling, which typically uses the maximum likelihood estimation (MLE) to estimate parameters. However, Bayesian estimation (BE) offers an alternative approach. In this simulation study, data were generated with varying sample sizes, proportions of missing data, and treatment effect sizes, as well as a time-dependent nuisance variable. Missing data were treated by way of multiple imputation, and parameters were estimated using BE and MLE. To date, there has been no study comparing BE and MLE in this manner for the case of count outcome data with unbalanced panels. A demonstration of simulation study methods using a real data set is provided. By virtue of the simulation, the true parameters were known and the performance of the two methods across all simulation replications was compared using bias, mean squared error, coverage, and Monte Carlo error of these estimates. BE outperformed MLE in 50 of 96 total comparisons and performed equally well in 18 of 96 comparisons. In many cases differences were very small. BE was found to be much more computationally demanding, requiring 40 to 179 times the computation time of MLE, depending on the conditions. On the basis of these findings, applied researchers and statisticians studying the case of unbalanced panels with count outcomes are advised to consider the computational and time requirements of fitting many models using multiple imputation and BE, and in many cases may consider MLE the preferable estimation method because of its speed and comparable accuracy.

Abstract Format



111 pages

Local Identifiers


Rights Statement

Copyright is held by the author.

Available for download on Wednesday, May 13, 2026