More on Poisson Regressions

Abstract

In the Chap. 4, binary Poisson regressions were assessed of parallel groups with a binary outcome. In the Chap.5, Poisson regressions were used for data with polytomous outcomes. This chapter will address additional models applying Poisson distributions. For rate analyses Poisson regression is very sensitive, and, generally, better so than standard linear regression. Linear regression measures events per population (or person), but does not explicitly include time as a covariate, although, implicitly, it is often assumed, albeit not plainly expressed. Poisson regression cannot only be used for counted events per person per period of time, but also for numbers of yes/no events per population per period of time. It is, then, similar to logistic regression, but different from it, in that it uses a log instead of logit (log odds) transformed dependent variable. It is more adequate and often provides better statistics than logistic regression does, because, again, time is explicitly included. In observational research event rates are often very much age and sex dependent and a model routinely adjusting these confounders are welcome. Examples and the analysis in SPSS statistical software is given, including intercept only Poisson regressions and loglinear Poisson models for incident rates with varying incident risks.