Own kind of analysis - just to make sure

[adrianolszewski]

Hello Dear Authors!

I'd like to confirm if I got the basics right.

Let's assume I'm interested in a 4 kinds of analysis:

1. ANCOVA but fit via GEE (due to violated normality of the Cholesky-decorrelated (normalised) residuals). I mean - GEE with Gaussian conditional response + identity link + independence covariance + small-sample bias-adjusted covariance estimator (Mancl-deRouen, as implemented in the glmtoolbox::geeglm). I often use this as the sensitivity under mildly violated normality remembering the requests from local regulators. So, it's basically ANCOVA only fit via GEE rather than OLS.

As far as I understand, for such thing I just provide my own function returning the list with: estimate, the SE and the DFs, right? (typically I use those from Kenward-Roger, sometimes residual one for Mancl-DeRouen).

2. Now let's assume that per SAP I want the inference at multiple selected timepoints (say, 3 selected ones). Will analyse() accept only a function returning a 2-sample ("per-timepoint") result? I noticed the "subset" in the examples.

The manual says: "This analysis function fits an ANCOVA model to the outcomes from each visit separately, and returns the “treatment effect” estimate as well as the corresponding least square means for each group"

In the naive approach, with MICE, I would:
a) iterate over the imputed datasets,
b) fit separate MMRM via GLS or via GEE (to retain power; also for GEE - to bypass the MCAR problem),
c) calculate the EM-mean for each timepoint,
d) pool the emmeans
e) perform the contrast testing using the MVT (multivariate t distribution) or some sequential adjustment.

If impossible with analyse(), is it possible to "bypass" it just by:
a) extracting the imputed datasets,
b) apply the analysis one-by-one (just like via lapply over mira-class casted to a regular list),
c) and then pool on my own? (Rubin/Barnard, von Hippel, etc)?

In other words, - is it possible to use only the imputation part of the rbmi, then leave this friendly ecosystem "equipped" with the imputed datasets?

2a) MMRM but used for the cLDA (constrained longitudinal data analysis) - where baseline is part of the response vector (no adjustment for it), and the baseline per-arm means are forced to be equal via appropriate model specification. 2 of ours sponsors love cLDA and don't want anything else, so I have to ask :)

3. Now I'd like to use the logistic regression (normal or Firth) for the clinical success. Most of the time, I report the result on the response (probability) scale and then it's always the AME (average marginal effect over the individual per-Patient covariates) if there are covariates (rather than MEM (marginal effect at the mean(s)), so unlike emmeans does). For it I employ the marginaleffects package. So, I understand that using analyse() will make it all transparent: just provide the est = AME, the SE, the DFs and that's all?

PS: just one last question - and please forgive me if it's an idiotic one, as I'm still recognising the broad topic of MI types....
For most of our sponsors, so far, we haven't dealt with the estimand framework (ICE handling) yet, following rather the old naive approach: ITT + MICE. The sponsor provided us with the definitions of imputations chains, i.e. which variables will be used to impute each variable of interest: the response, the covariates, and sometimes the baseline values).

For the univariate "imputer", my sponsors practically always (90% or so) asked for the PMM method (either imputed separately at each timepoint via MidasTouch extension, or by using the 2-level PMM "longitudinally at once"). It was asked practically for all data types (numerical, binary, ordinal, counts), rarely also the ordinal logistic regression univariate "imputer" was used for Likert-type data. The decision was justified as follows: the PMM preserves the nature and distribution of the data (doesn't assume any specific theoretical distribution; even including extreme data if they occur), which was especially important for 3 of the most common statistical reviewers we corresponded with (even despite the well-known limitation, that PMM cannot extrapolate outside the observed domain).

Is PMM "combinable" with RBMI and this package?

[wolbersm]

Hi @adrianolszewski

I'll give it a quick shot but others, please chime in:

1. Yes (assuming you are using method_bayes() or method_approxbayes())
2. Yes, you can extract the imputed datasets from impute() using extract_imputed_dfs() and then do things from then on "by hand".
   2a) It's not officially implemented. I haven't tried but you might be able to trick draws() into doing this by defining the covariates smartly.
3. Yes, I think so. You can write your own analysis function which provides your preferred estimates (with corresponding SE and df).
4. Not sure but my guess is that you are likely better of doing this with mice or similar packages. Depending on your use case, retrieved drop-out models might serve a similar purpose to PMM and these are available in rbmi, see our vignette.

[adrianolszewski]

Thank you wholeheartedly for the responses! It's very helpful.

PS: If you agree, I would like to keep this discussion open, as maybe someone will want to share some additional warnings or ideas, and people will be able to find it.

[wolbersm]

Thanks for the response, glad my comments were helpful. Yes, fine with me to keep the discussion open for now.