Summary
The episode works through four questions that students often find difficult in generalized linear models. It shows why one odds ratio from an ordinal model changes probabilities by different amounts at different splits, why a relative risk ratio from a multinomial model can exceed one while the share in that category falls, how leaving out an offset can reverse the direction of a rate ratio, and how unmeasured differences between people produce overdispersion and extra zeros. Context comes from Statistics Canada data on self-rated health and on access to a regular health care provider, from a British Columbia study of frequent emergency department users, and from the origin of generalized linear models in 1972. The hosts close with a discussion of whether a five-point rating can be analysed as a number, and they agree on a practical rule for reporting ordinal outcomes.