A weakly informative default prior distribution for logistic and other regression models

Abstract

We propose a new prior distribution for classical (nonhierarchical) logistic regression models, constructed by first scaling all nonbinary variables to have mean 0 and standard deviation 0.5, and then placing independent Student-t prior distributions on the coefficients. As a default choice, we recommend the Cauchy distribution with center 0 and scale 2.5, which in the simplest setting is a longer-tailed version of the distribution attained by assuming one-half additional success and one-half additional failure in a logistic regression. Cross-validation on a corpus of datasets shows the Cauchy class of prior distributions to outperform existing implementations of Gaussian and Laplace priors. We recommend this prior distribution as a default choice for routine applied use. It has the advantage of always giving answers, even when there is complete separation in logistic regression (a common problem, even when the sample size is large and the number of predictors is small), and also automatically applying more shrinkage to higher-order interactions. This can be useful in routine data analysis as well as in automated procedures such as chained equations for missing-data imputation. We implement a procedure to fit generalized linear models in R with the Student-t prior distribution by incorporating an approximate EM algorithm into the usual iteratively weighted least squares. We illustrate with several applications, including a series of logistic regressions predicting voting preferences, a small bioassay experiment, and an imputation model for a public health data set. © Institute of Mathematical Statistics

Differences in the risk of cervical cancer and human papillomavirus infection by education level

“…Owing to small numbers, the last two groups were merged in the case-control studies. Regularised logistic regression (Gelman et al, 2008) was used to calculate odds ratios (ORs) and 95% confidence intervals (CIs) by education level for cervical cancer in the case-control studies and for HPV infection among control women only in case-control studies and among the general female population in prevalence surveys. The reference category for education was set to the most common category (i.e., 1-5 years in case-control studies and 6-10 years in prevalence surveys).…”

Identifying Vulnerable Populations through an Examination of the Association Between Multipollutant Profiles and Poverty

“…Since it is possible for a particular θ c to be associated with an empty cluster, these parameters must be assigned a proper prior. Therefore, we assign to each θ c a proper t density function with 7 degrees of freedom and scale 2.5 as a prior, as discussed by Gelman et al, 29 which corresponds to the baseline case of onehalf of a success and one-half of a failure for a single binomial trial with probability p = logit À1 (θ c ). Our response model, which links the clusters with poverty counts, y i for CT, i, is simply y i ≈ Bin (n i ,p i ) with…”

Problems modelling behavioural variation across Western North American Indian societies

“…Model comparison is a powerful tool for evaluating alternative models given the data [12,17]. However, given so many predictors, the best model is often one of a large set of models with very similar AIC values (e.g.…”