POOLED TIME-SERIES REGRESSION-ANALYSIS IN LONGITUDINAL-STUDIES

Many longitudinal studies attempt to examine changes in outcome measures over time in groups of patients. Applying conventional analytic techniques, such as a single classical linear regression model, to these data will often not result in minimum variance estimates, and may affect the results of tests of significance. Pooled time series regression analyses comprise a set of techniques that may be used in these instances to model changes in outcome measures over time. Pooling of time series data from many individuals may be done using two types of models: fixed effect models, which specify differences among individuals in separate intercept terms, and random effects models, which allow for differences among individuals by including an additional error component in the model. The choice between these alternative model specifications is guided by both theoretical and statistical considerations. This paper describes the use of pooled time series analysis, contrasts these methods with two classical linear regression approaches, and demonstrates these differences using two examples: a hypothetical study of serum glucose measurements in patients with diabetic ketoacidosis, and a longitudinal study of the development of functional disability in a cohort of patients with rheumatoid arthritis. These methods may be applicable to the study of outcomes in many chronic illnesses.