Extract estimates for subject-level correlation in multivariate model

Hi folks,

I’m modeling data from two tasks, call them A and B, and they’re both influenced by the same underlying trait. I have data from 30 subjects doing both tasks, and I’ve fit a multivariate model of the form below.

bf_a <- (DV_A ~ (1|a|Subject)+(1|AItem))+bernoulli()

bf_b <- (DV_B ~ TaskSpecificFixedEffects + (1|a|Subject)+(1|BItem))+bernoulli()

The output of the model contains a group-level correlation between the intercepts for DV_A and DV_B, but I can’t seem to find a subject-level correlation parameter.

I see a bunch of parameters like “r_Subject_DV_A[XYZ, Intercept]” and “r_Subject_DV_B[XYZ, Intercept]”, where XYZ is the individual subject’s unique identifier. Would it obtain the same information to just extract estimates for the DV_A and DV_B r_ parameters for each individual subject and then correlate them within draw?

Thanks very much,

  • Operating System: Windows 10
  • brms Version: 2.14.4

Ignoring the TaskSpecificFixedEffects term in your bf_b–which makes the rest of the second model conditional on that term–, my interpretation of your set-up is that the “group-level correlation between the intercepts for DV_A and DV_B” is your “subject-level correlation parameter.” What other correlation are you looking for?

Hi Solomon,

I think so; what I was wondering was whether there was an individual-level correlation as well as a group-level correlation in a hierarchical relationship or not, but I guess what I’m hearing is that as it’s done right now, there’s not?


If you were using the Gaussian likelihood, you’d also get a residual correlation. But there’s no corresponding parameter within a bivariate Bernoulli framework. Perhaps that’s what you were thinking of.

I think so, thanks very much.

1 Like