ABSTRACT
Longitudinal data arise when an outcome variable such as pulmonary function, exercise testing, or quality of life is measured repeatedly over time. Standard approaches such as ordinary least squares (OLS) assume that errors are independent and identically distributed (i.i.d.). For longitudinal (or repeated measures) data, the i.i.d. assumption breaks down because individual observations are naturally associated with larger groups or clusters-two data points taken from one cluster will likely be more similar to one another than two data points taken from different subjects. Usually the clusters comprise repeated observations from individual subjects, although there can be other types of clusters such as clinical trial sites, schools, and geographical regions. In this chapter we will discuss a number of ways to model this correlation and account for it in assessing treatment differences. While the techniques discussed in this chapter can be easily adapted to other types of clustering, we will assume throughout that clusters correspond to subjects.
