This question is a follow up that extends this issue to GLMs.
I have come accross this blogpost that presents several ways of “standardizing” the coefs from a logistic regression. They seem to suggest that the Agresti method is the most simple and intuitive.
Since the the method presented by bgoodri and jgabry “works” with GLMs (i.e., does not crash), and given my rather beginer level mathemathical skills, I would like to know if it was ok to use it for GLMs to obtain standardized log odds ratios.
Thank you
I would say it does not matter that much which way of standardizing the coefficients you use because the results are not very useful anyway.
The purported goal of standardizing the coefficients is to be able to say that one variable has more of an effect than another variable. But putting all of the variables in fake units does not accomplish that. Also, too, measuring in terms of log-odds is unintuitive. What does accomplish that goal in an intuitive way is manipulating the posterior predictions (or the posterior probabilities in the case of a logit model) by passing differently constructed data.frames to the newdata argument in rstanarm and seeing what changes your model implies.
I agree, but “standardized” coefficients can be required by reviewers as indices of “effect size” (especially in psychology), and there have been some attempts to develop “rules of thumb” for odds ratios. I just wanted to know if standardizing the log odds using the method mentioned above, then transforming them to odds ratios made sense to answer reviews. Thanks a lot anyway!
I would try explaining to the editor that the reviewers do not understand Bayesian estimation of generative models well enough to review a paper that uses Bayesian estimation of a generative model.
You’re experiences dealing with reviewers must be better than mine have been!
2 Likes