blog/2022/11/29/conditional-marginal-marginaleffects/index

[giscus[bot]]

blog/2022/11/29/conditional-marginal-marginaleffects/index

Use the {marginaleffects} package to calculate tricky and nuanced marginal and conditional effects in generalized linear mixed models

https://www.andrewheiss.com/blog/2022/11/29/conditional-marginal-marginaleffects/index.html

[sstew45]

The detail used here in teasing out the different estimates, and understanding their interpretation, is really nice to see. What's even better is how you summarized each part of the analysis in a beautiful and digestable narrative, complete with journal-quality tables and plots and well-organized code. I will be bookmarking this for multiple use-cases. Great post!

[epkanol]

Sorry if I'm dumb, but I'm trying to get my head around Bayesian inference, and your excellent posts help me a lot in that regard, Andrew.

When you say "There are a few problems with these estimates though: (1) they’re on the log odds scale, "
shouldn't it really be "...on the log scale..."

(No less hard to interpret, it's just that I reserved "log odds" for the "logit" link in my brain, and here we are using a log link, i.e. the lognormal distribution)

[nahorp]

Thanks once again Andrew! What would the emmeans equivalent of the marginaleffects code look like? Even if only in the summary table, it'd be great to see a comparison!

[gugushvili]

Thanks for a nice blog post. Just wanted to point out this article (page 327), where the terms conditional and marginal are defined in the opposite way to what you wrote: there marginal predictions are obtained by setting random offsets to zero. They are not the only ones to do it.

[DanSimonet]

Hey Andrew awesome blog posts - your marginalia overview is one of my favorites! Question for your code on Population-level ATE (marginalized) . Should it also include the unique clusters in data grid like your example for Average population-level outcomes (marginalized)? So like this

comparisons(
  fit, 
  variables = "TX", 
  newdata = datagrid(TX = c(0, 1), 
                     cluster = unique), 
  re_formula = NULL
) %>% 
  tidy()

I receive very different results for some test cases if the cluster=unique is omitted. Thanks!