this is my first post.
I am trying to fit a (simple?) multivariate model in
brms with two responses coming from different distributions. Both are intercept-only models with the same grouping structure (‘random effects’, see reprex below), so the model I have estimates two intercepts and two group-level variances.
bf(resp1 | trials(resp1_trials) ~ 1 + (1|g)) + binomial() + bf(resp2 ~ 1 + (1|g)) + poisson()
Since I know (in the context of simulated data) that the underlying intercept and group level variance is the same for both responses, I would like to constrain the model such that it returns a single (‘shared’) intercept and a single (‘shared’) variance (instead of two each). Is it possible in
brms to write the model formula so that it adheres to that constraint (rather than coding the model directly in
I know that I could add a fixed (via a constant prior) or estimated correlation term for the group-level intercepts (e.g. here), which gets me closer to what I have in mind, but still estimates two group-level terms.
Any pointer would be highly appreciated.
library(brms) set.seed(123) g_val <- rnorm(50, mean = 0.3, sd = 1.2) resp1 <- rbinom(n = 50, size = 100, prob = inv_logit_scaled(g_val)) resp2 <- rpois(n = 50, lambda = exp(g_val)) dat <- data.frame(resp1, resp1_trials = 100, resp2, g = seq_along(g_val)) form1 <- bf(resp1 | trials(resp1_trials) ~ 1 + (1|c|g)) + binomial() form2 <- bf(resp2 ~ 1 + (1|c|g)) + poisson() fit <- brm(form1 + form2 + set_rescor(FALSE), data = dat) fit par(mfrow = c(1, 2)) p <- ranef(fit)$g[, , "resp1_Intercept"][, "Estimate"] plot(g_val, p, xlab = "truth", ylab = "estimated for response 1") p <- ranef(fit)$g[, , "resp2_Intercept"][, "Estimate"] plot(g_val, p, xlab = "truth", ylab = "estimated for response 2")
- Operating System: macOS 12.5
- brms Version: 2.17.0