Poisson Regression in Epidemiology

Various authors [3, 9, 11, 12] have noted that Poisson regression can be used to analyze cohort survival data (see Cohort Study). This formulation also leads to a unification of risk estimation based on internal comparison of rates among members of a cohort with various exposure levels and classical epidemiologic methods based on external rates that yield standardized mortality ratios or standardized incidence ratios [2, 5] (see Standardization Methods). Poisson regression is an important alternative to partial-likelihood-based analysis of the proportional hazards model (see Cox Regression Model) and to parametric analyses of such models (see Survival Analysis, Overview) for two main reasons. First, it provides an efficient and intuitive method for dealing with cumulative exposures and other time-dependent covariates and for allowing risk to depend on multiple time scales (e.g. attained age, time since exposure, or calendar time). Secondly, it facilitates the consideration of a broad range of risk models including those that allow for the direct parametric description of baseline rates, absolute excess rates, and relative risks. Breslow & Day [4] offer a general discussion of the use of Poisson regression in the analysis of cohort survival data. Some of the most extensive applications of these methods have involved studies of radiation effects on mortality and cancer incidence in the atomic bomb survivors [14].