Suggestion: Change-in-estimate projection

[StaffanBetner]

In epidemiology we are concerned with the selection of confounders. One way to select confounders for an exposure is to fit a full model (reference model) with all potential confounders and then remove potential confounders in such a way that the effect estimate for the exposure(s) is as similar to the estimate for the full model as possible, while keeping the model as small as possible. Typically only the point estimate is considered which throws away information.

I wonder whether the projection could be adjusted to work towards this goal, instead of selecting the minimal submodel that keeps the same predictive performance. Maybe something to look into?

[avehtari]

Thanks for the suggestion, this makes sense. This is feasible as it would not change the projection part, just the criterion for whether the model is good enough.

[avehtari]

Checking if you would like to collaborate on this?

[StaffanBetner]

In which sense?

[avehtari]

In sense of developing theory, running experiments, or implementing code? Have you already done experiments on this?