Bayesian Inference for Linear Models
Alan Agresti, Maria Kateri, Ranjini Grove et al. · Chapman & Hall/CRC eBooks · 2026
This chapter presents Bayesian linear modeling methods for analyzing how the outcome on a quantitative response variable is associated with or can be predicted by a set of explanatory variables. Inferential methods assume that the response variable, given the explanatory variables, has a normal distribution. The standard normal linear model is the multiple regression model. The model can include a categorical explanatory variable by adding an indicator variable for each category. A Bayesian approach assumes normal prior distributions for the effect parameters. With an inverse gamma prior distribution for the error variance, the marginal posterior distribution of each effect parameter is a t distribution. From the bias/variance tradeoff, simpler models that have higher bias but smaller variance may have smaller mean squared error for estimating effects. This is taken into account by indices such as BIC, which increasingly penalizes the log-likelihood as the number of model parameters increases. The lasso penalized likelihood method relates to a Bayesian fit using double-exponential prior distributions for the effect parameters, which place much weight close to 0.
Cited by 0 works. doi:10.1201/9781003715924-4