Collapsibility of some association measures and survival models (Q1680804)

There are many variations of collapsibility discussed in the literature, and this paper focuses on one of them, average collapsibility. The association measures under consideration are expectation dependence and mixed derivative of interaction, applicable to linear and logistic regression models, together with a new suggestion by the author, log-expectation dependence with application to Poisson and negative binomial regression models. The relative risk in contingency tables is discussed in an example. The connection to Simpson's paradox and and reverse association is pointed out. Sufficient conditions for average collapsibility are given for these association measures. Finally, collapsibility results for survival models are given in terms of the linear transformation model, which includes proportional hazards and proportional odds regression. As the author points out: ``\dots, there is not much work devoted to the testing aspects of collapsibility in general, \dots''. We can only hope that this situation changes in the future.