Bayesian Inference for Generalized Linear Models

Alan Agresti, Maria Kateri, Ranjini Grove et al. · Chapman & Hall/CRC eBooks · 2026

Generalized linear models enable us to assume distributions other than the normal for the response variable Y and to relate a link function of E(Y) to the explanatory variables. GLMs apply both to discrete and to continuous Y. Bayesian fitting typically uses normal prior distributions for effect parameters. Positively-valued responses often have variance increasing as the mean increases, which motivate assuming a gamma distribution for Y using the log link function. The log link is also relevant for exponential relationships. A Cauchy distribution for Y or for a prior distribution has thick tails that permit extreme outcomes or unusual effects. For binary Y with outcomes 0 and 1, the logistic regression model describes the log of the odds, P(Y=1)/P(Y=0), called the logit. When Y is a count, GLMs called loglinear models use the log link function. GLMs that assume a Poisson distribution for Y force its variance to equal the mean, and the negative binomial is more appropriate when overdispersion occurs. Multilevel models are a hierarchical model that account for a nesting of observations at more than one level. The generalized linear mixed model handles matched sets of observations that are naturally correlated, such as in a longitudinal study.

Cited by 0 works. doi:10.1201/9781003715924-5

Authors