Introduction to GEEs

Regression models come with many assumptions that need to be relaxed for many types of real-world problems. For instance, linear regression assumes a normally distributed outcome. In many problems, we may wish to work with binary outcome data (yes/no, survived/died, recurred/did not recur…), count data (number of events in a period of time, such as this chapter’s outcome), or failure rates for a manufactured product (likelihood of failure outcome).

One of the most common outcome distributions comes from the binomial distribution, in which binary data is collected across a population. For example, we may have a sample of patients in a glioblastoma study, where we compare 6-month survival rates of patients across different treatment groups. Each patient will either survive or die (binary outcome); aggregated by group, these outcomes form binomial distributions, which can be compared statistically to determine if an optimal treatment exists for glioblastoma...