Varying gradient by group variable

[mattmoo]

I know that it is possible in mcp to calculate varying changepoints by a grouping variable, but is it possible to do the same with gradients?

[lindeloev]

If you mean continuous variables, then yes, it will be in the future (and this would be unique to mcp AFAIK). They still have to be observed within a group, but the syntax would be something like:

model = list(
    y ~ x,
    1 + (0 + age | participant) ~ 0 + x
)

So each participant has one age and you would get an estimate on how the change point location is affected by age. If that's what you had in mind? The grouping (participant) of the continuous predictor (age) is necessary for model identifiability, I think.

[mattmoo]

I think so, the specific situation that I have right now is to compare trajectories in patient outcomes between social groups. I guess that it would need different intercepts per group also (and zero changepoints!). It is easy enough to work around for now by fitting separate models to each group.

A future situation that I have in mind is a stepped-wedge trial, wherein the implementation is applied to each site at a different timepoint (already supported), but sites may also differ in intervention-agnostic factors, so some further clustering in the model would be necessary.