Understanding brms functions for summarizing a logistic model

Hi everyone,

Sorry if this is a beginner question: I am at the point of creating some output tables to report the results of my experiment; specifically the population-level fixed effects. It is a logistic (bernoulli, link = logit) model in brms, and I would like to report it in the most intuitive way possible (i.e. in probability space instead of log-odds space).

I am unsure of the best way to go about this, as I don’t feel I’ve seen one consistent method in my searches. There are the following functions in the brms package that I think could be useful to me, and I’d like to understand the differences between them, and which ones incorporate a transformation into probability space:

  • summary (seen here)
  • as_draws_arrayposterior::summarize_draws (seen here)
  • coef, fixef, ranefinv_logit_scaled (see here)
  • posterior_predict : includes an option transform, where I can enter inv_logit_scaled but this is deprecated? (seen here)
  • posterior_epredstr (seen here)
  • fitted (seen here)
  • posterior_table, with input from one of the above functions (see here)

For reasons specified here, I cannot plot conditional effects due to a bug in the current version of brms. [Edit: this is a bug with scale() and poly() in general, not brms-specific]

Also note that I have a fairly complex model, and to my understanding I would need to include all the variables and interactions each time I want to use hypothesis() with inv_logit_scaled() or exp(). But perhaps there is something I am not understanding here.

Maybe it is not possible, or not meaningful, to have such a table? Should I report my results in log-odds space instead?

As far as reporting your model parameters, I’d recommend a table with all major parameters on the log-odds scale. That table would basically include the information returned by the summary() function.

However, I agree that it’s often easier to understand the model predictions on the probability space. For that, I’d try to pin down which parameters or predictions from your model most directly address your primary research questions. Then I’d find a way to convert those to the probability metric (or difference in probability metric) and report them in either prose or plots.

1 Like