search

stan-dev / projpred

Projection predictive variable selection
https://mc-stan.org/projpred/
Other
110 stars 25 forks source link

Allow dispersion parameter to be observation-specific #481

Open fweber144 opened 10 months ago

fweber144 commented 10 months ago

This is a feature request to allow the dispersion parameter to be observation-specific, as, for example, in models like the following:

library(brms)
set.seed(4734)
data_het <- data.frame(
  y = c(rnorm(50), rnorm(50, 1, 2)),
  x = factor(rep(c("a", "b"), each = 50)),
  se_y = rgamma(100, shape = 1)
)
fit6 <- brm(y | se(se_y, sigma = TRUE) ~ x, data = data_het)
stancode(fit6)
str(standata(fit6))

This came up in https://discourse.mc-stan.org/t/accounting-for-measurement-error-during-variable-selection-with-projpred-possibly-with-rstanarm-or-brms/11789.

fweber144 commented 10 months ago

I guess we will need special submodel fitters for this (or at least adapt the currently available ones), so that the observation-specific dispersion parameter values are taken into account when performing the projection.