Longitudinal Research Design

In epidemiology, a longitudinal study refers to the collection of data from the same unit (e.g., the same person) at two or more different points in time. The great advantage of longitudinal studies is that each subject serves as his or her own control in the study of change across time. This reduces between-subject variability and requires a smaller number of subjects compared with independent subject designs, and it allows the researcher to eliminate a number of competing explanations for effects observed—most important, the cohort effect. The main disadvantages of longitudinal designs are that they are expensive and time-consuming relative to cross-sectional designs, and that they are subject to difficulties with retention, that is, subjects may drop out of the study. In addition, special statistical techniques are needed to account for the fact that repeated measurements taken on the same person or unit will be more similar than the same number of measurements taken on different people.

Designing a longitudinal study is a complex task that involves a number of decisions. The primary decision is whether the data should be recorded prospectively (from the starting point of the study into the future) or retrospectively (collecting data on events that have already occurred). The investigator must also determine how to select a sample of subjects that will represent the target population and how large a sample is needed to have adequate power. Finally, the investigator must choose the variables that will represent the phenomenon under investigation and the frequency at which these variables should be measured.

Methods for the analysis of data in longitudinal studies depend primarily on whether time is considered as a covariate or as an outcome. When the time is viewed as a covariate, regression techniques that account for within-subject association in the data can be used to study the change across time. For time-toevent data, survival analysis that takes into account the potential censoring of data, that is, the unavailability of end points, is required. This entry is concerned with studies in which time is considered to be a covariate. Survival analysis and related methods, which consider time-to-event as the outcome, are treated in a separate entry.