Including both the time varying and baseline value of a covariate

data analysis
1

I am looking for a deeper theoretical explanation on why one would need to include both the baseline value prior to treatment along a variable that varies with time for a covariate like Blood Pressure. There would be correlation between the two predictors. In PMx we use an NLME hammer and the explanation I have been told is that both are needed to accurately detect a covariate effect with the BL value accounting for the BSV and the time variation capturing the treatment impact.

Does that explanation hold water? I would love to hear the thoughts on this issue by the community here.

2

You would need to have special justification to not include the baseline value such as strong evidence that were it to be included its regression coefficients would all be zero. Otherwise you’ll get an artificially high estimate of the time-varying effect. Think of the baseline value as a centering constant if nothing else.

1 Like
3

Thanks for the reply Professor.

One clarification and one follow up…

If the baseline is centering should the time varying component be captured as change from baseline? And when does it make sense to express the difference as percentage? (Like if I had a huge range of values like platelet count or something).

Thanks in advance.

4

To use change from baseline assumes the baseline is unimportant as an absolute effect, the effect of the variables at all times is linear, and the slope of baseline on on times is 1.0. No way those are satisfied.

3 Likes