Count outcomes are common in clinical studies of multiple sclerosis (MS), but the methods for the analysis of count outcomes are not commonly discussed in introductory statistics courses.
To introduce commonly used statistical approaches for clinical studies of MS with a focus on the similarities and differences between the approaches.
For this session, data was simulated to mimic two recent MS clinical trials. The first dataset was simulated to mimic a phase III clinical trial with the number of relapses as the outcome. The second dataset was simulated to mimic a phase II clinical trial with the number of new gadolinium enhancing lesions on a brain MRI as the outcome.
When the goal is to analyze the number of new relapses, a commonly used approach is Poisson regression. This approach estimates the rate ratio and easily accommodates different follow-up intervals for each subject by including an offset in the model. This approach is related to a comparison of incidence rates and the analysis of recurrent events using survival analysis. When the goal is to analyze the number of new lesions, Poisson regression is usually not appropriate because the mean and variance of the number of new lesions are often quite different due to a small number of subjects having a large number of new lesions. This leads to highly skewed data. To analyze this outcome, Poisson regression with robust standard errors or negative binomial regression are preferred because these approaches accommodate overdispersion. Results from the statistical software package Stata will be shown to demonstrate the commands described in the session.
Count outcomes are common in MS clinical studies, and specific analysis approaches are used for these outcomes. Many of the analysis approaches used for count outcomes are related, but each approach makes different assumptions so the best approach will depend on the outcome of the study.