How to display and tabulate interaction effects in a longitudinal model

rms
interaction

#1

New poster here. I did not see this topic in the list of topics going back several months. If there is an old topic covering this, point this out to me.

A stat colleague of mine and I are working with clinical researchers in Type 1 diabetes and writing a manuscript showing what baseline factors are important in the rate of loss of c-peptide from diagnosis. The analysis is a straightforward multiple regression longitudinal model with several main effects and their interaction with time in the model. Our struggle has been with how to best show this in tabular form for a clinical journal. Our clinical colleagues are struggling with the usual tables that are produced in SAS or R. We have some nice plots of the estimated linear degradation rates over levels of the factors – everyone is fine with those figures but the table of the estimated effects of main effects and interactions is causing a lot of confusion. Has anyone come up with some nice tabular examples for clinical journals? Additional graphics suggestions are also welcome.


#2

I don’t know of a good way to tabulate estimates from the model if either time or the c-peptide is continuous. (Note: rate of loss of c-peptide assumes that the most recent level of c-peptide is not all-important. This assumption should be questioned.) I think that a graphical display is preferred and we have found such displays to be well accepted in biomedical journals.

If the predictor were to be binary, this is simple. Compute the difference in two groups as a function of time along with compatibility (aka confidence) intervals and plot time on the x-axis. Simultaneous confidence bands may be called for. If the predictor is continuous, you might choose 5 values for it and produce 5 estimated curves with confidence bands. If one of the values is “special” then you might show 4 differences in curves over time, with compatibility bands.

:new: If the predictor is continuous and time is discrete, then show the predictor on the x-axis and a separate curve for each time.