BBR Session 16: Transformations and Change

This is a topic for questions, answers, and discussions about session 16 of the Biostatistics for Biomedical Research course for 2020-04-24. Session topics and links are listed here . The session covers transforming variables, measuring change, problems with change scores, and regression to the mean.

I have a question about the change of exposure. Is it the same strategy as the change of the outcome? If I want to evaluate the change of ordinal exposure, the right way to model it should be:

Assume Y is the outcome and X is the exposure (without considering covariates), then
library (rms)
ols (Y~pol(Xbaseline, 2)+pol(Xfollow-up, 2))

I am wondering whether it is the right way to think about this issue.

It is simple to think about this if you have two baseline measures (say one historical one and one measurement just as follow-up starts). But if your Xfollow-up is measured at the same time as Y this gets harder to interpret. It’s then a cross-correlation analysis, and you might find some good thinking about that in the econometrics literature.

1 Like

Thank you so much, Frank!

You really got my point which I initially had a rough idea.

It is the actual situation I encountered. Xfollow-up was measured at the same time as Y. But the good thing is I have only one Xhistorical, not several. It seems I could handle it through a simple way rather than a cross-correlation analysis.

Not sure. The fact that you will be analyzing Xfollow-up means that cross-correlation analysis is what you are doing.

1 Like

Hi @f2harrell ,

I watched the YouTube video on this session, and I am interested in the topic discussed towards the end (40:37min BBR16: Biostatistics for Biomedical Research Session 16 - YouTube).

If I understood correctly, you suggest fitting a model with multiple time points as:

y1 + y2 + … + yk ~ y0 + treatment

You mentioned in the video that further explanations could be found in the RMS’ Course Notes. Yet, I wasn’t able to find notes on this topic in specific.

Could you provide further references on this type of modeling? I would love to delve into it.

Thank you!

Please see the Modeling Longitudinal Responses chapter in the RMS course notes. At the end of that chapter you’ll see new material about Markov longitudinal models, which are quite flexible.

Thank you for the quick reply. I will definitely take a look.