Question about Est.Error column in predict() with zero-one inflated beta family

Interfaces brms
1

Hi forum,

I’m somewhat new to bayesian statistics, but am loving brms - it’s brilliantly solving challenging variable selection and distributional regression problem I have. I have a question about the interpretation of the Est.Error column when using extended families - in my case I’m using a zero-one inflated beta distribution. What I’m curious about is what factors effect the choice of summary statistic when interested in the “width” and flatness of the posterior, in particular with the zero-one inflated beta distribution. My use case involves predicting hundreds of thousands of out-of-sample values, so inspecting the posteriors of individual predictions isn’t feasible

In addition, I’m interested in comparing the accuracy of different summary statistics of central tendency of the posterior against point-estimates obtained from some frequentist models, but I’m stalled on how to pull the appropriate parameters out of the fitted brms model object.

Can anyone recommend a primer on summary stats for zero-one inflated beta?

Thanks!

2

Hi,

By default Est.Error is the standard deviation of the posterior. You can easily look at the mean, median, and mode with quantile intervals or highest density intervals using functions from the tidybayes package: https://mjskay.github.io/tidybayes/articles/tidy-brms.html#point-summaries-and-intervals.

A reasonable thing to do might be to look at e.g. means+qis and medians+hdis and examine the differences.

Some people argue that the hdi is more sensible, but the default quantile intervals should be reasonable in most cases. I don’t see why the zoib model would be any different to other models in regard to the choice of how to summarize the posterior.

3

Perfect, thank you for the link.