cov_adj's in lmitt.f()

[benthestatistician]

Presently lmitt.formula() accepts a prior covariance adjustment model via offset=cov_adj(cmod), just as in the lm() |> lmitt() path. Should we it also accept such models via a new argument cov_adj()? I.e.

lmitt(y~s, design=des, cov_adj=cmod)

?

[benthestatistician]

Not at all urgent, just creating a place for discussion.

The clear advantage of this would be to simplify calls to lmitt(). A less-obvious disadvantage is that it would frustrate or complicate adding additional arguments to cov_adj(), as envisioned in #4.

[jwasserman2]

My vote would be to stick with offset. Even though we're being less rigorous about how much the functionality vialmitt or lm aligns, I think this change would bring about unnecessary confusion. I don't have hesitation about the code implementation--under the hood the changes would be small--I just think it's a somewhat extraneous addition.

[benthestatistician]

Excellent point. Related: on its face lmitt(y~s, design=des, cov_adj=cmod) is ambiguous as to whether the predictions from the covariance adjustment model function as an offset or as a regressor. lmitt(y~s, design=des, offset=cov_adj(cmod)) and lmitt(y~s + offset(cov_adj(cmod)), design=des) both resolve that ambiguity.

[benthestatistician]

A number of downsides having emerged, with no overriding upsides, we elected to table this indefinitely.