Brms marginal_effects in probit/logit models

Thanks very much @paul.buerkner and @matti for the tutorial on ordinal regression models using brms!

I have two questions regarding the interpretation/plotting of marginal effects, which is suggested as a good way to understand the model’s implications. If you or another member could help, it would be much appreciated!

Firstly, I have one model in which there are unequal variances, as assessed by estimating ‘disc’. Are these unequal variances automatically incorporated into the probability estimates that are displayed when you plot/get summaries for the marginal_effects, or do I have to stipulate this explicitly? At the moment, I have simply used the conditions argument to specify that I want estimates for an interaction, and hope that the resulting marginal effects incorporate the unequal variances.

Secondly, how are the probabilities exactly calculated from the posterior estimates, and how are unequal variances included here? I can receive the estimates used for plotting the marginal_effects graphs by saving the marginal effects as an object and then seeing what it contains with $, rather than simply running it and getting the plots returned, but I also like to calculate them myself to check I understand. Do you simply add together the appropriate components from the posterior samples for each level of the outcome variable, and then use something like inv_logit_scaled to get the probabilities for logit models, and perhaps a z-score-to-percentile conversion to get the implied probabilities for the probit models?

Thank you very much!

  • Operating System: Mac 10.12.6
  • brms Version: 2.9.0

(1) Yes. They are automatically incorporated.

(2) marginal_effects uses the fitted method behind the scenes to compute the probabilities,
which roughly works as you suggested.