Modeling independent variance between groups

Modeling
,
1

Hello,

I’m trying to create a simple linear model with a random intercept and slope, *edit in the brms package in R. This works just fine, but I noticed that my group slope/intercept parameters have almost equivalent estimated error, which as I understand it is by design. Reading the documentation, if I follow, using the gr() by parameter can allow for the modeling of separate variance.

Here is my original formula:

y ~ 1 + x  + (1 + x|group)

Which I tried to alter to:

y ~ 1 + x + (1 + x|gr(group, by=x)

This throws an error

Error: Some levels of 'x' correspond to multiple levels of 'group'.

I can try to re-create this with some toy data and post it here, but my sense is that I’ve simply misunderstood the syntax. Any insight would be appreciated.

2

From: Set up basic grouping terms in brms — gr • brms

by

An optional factor variable, specifying sub-populations of the groups. For each level of the by variable, a separate variance-covariance matrix will be fitted. Levels of the grouping factor must be nested in levels of the by variable.

So I think the problem with x|gr(group, by=x) is the nesting. You need each level of x to correspond to a single level of group.

3

Just to spell it out, I think you simply want:

y ~ 1 + x + (1 | gr(group, by = x)

which estimates separate variances among groups for each level of x.

4

Thanks for the response. This adjustment throws the same error. Maybe it was unclear, but in this model x is a continuous float predictor and group is a categorical. Is using gr() to estimate separate variances for x amongst each level of group for both slope and intercept parameters the correct approach in that situation?