The log-binomial model (the generalized linear model with binomial errors and log link) makes it possible to directly estimate the relative risk from cohort follow-up data, or the prevalence ratio from cross-sectional data, with adjustment for confounders. One of the problems with the use of...