I have data for a single arm pre-post study for which one of the outcomes of interest only has measures at the follow-up visit (this is part of the study design). Several potential predictors have measures at baseline and follow-up.
I am wanting to know in the simple case of a univariate regression with one of these predictors, what are the potential implications of using both (baseline and follow up) measures vs the change score vs follow-up measures only? (remember this is for the independent, not dependent variable case).
I have done several analyses already and included both measures of each predictor in these regression models. I largely based this decision on comments in this stack exchange thread (for which Frank commented also):
It’s now been suggested to me now that this is not an optimal approach and I should consider using just the follow-up measures.
I’m trying to work out if one approach is better. Interested in people’s thoughts - thank you.
There are already quite a few good points in that topic you refer to including:
Functional form of using a change score can be more restrictive (as you will have just 1 coefficient for the change score, rather than 2 separate ones for the baselines and follow-up measurement)
Measurements closer to the end-point on average are more closely associated with the outcomes
I think the main question is what you want to achieve with your analyses?
For example, if you include just the baseline measures, you are effectively trying to see if some baseline measures predict the outcome/response (which might be useful if you are trying to predict in advance which individuals might benefit from treatment).
Including both measures will give you an idea of how well the combination of baseline and follow-up measurements predict the outcome. I think you can do this for several reasons, but I’m not competely sure. It could just be something you find interesting to know from a descriptive point of view (i.e. so you know that the measurements of baseline and follow-up predictors have a association with the outcome). Alternatively, I can imagine if you find the baseline and follow-up measurements predict the outcome quite well and you can’t measure the outcome in all cases (i.e. it’s difficult or expensive to measure) then the predictions from your baseline and follow-up measurements could be used as a surrogate for the actual outcome.
It really depends on what one wants to even do, which was not clear fro. the original post.
For seeing whether an intervention works or for whom it works best, this is usually the wrong study design to get any meaningful answers (unless the natural history is well known and you have a massive treatment effective relative to what changes are plausible a-priori). The suggestion to see who benefits from a treatment is off the mark: How would you know who benefits when you don’t have a control group?
Of course, predicting outcomes could be a more plausible question. Or perhaps variables x and y are correlated. Or some patients will experience xyz (without necessarily saying that this had anything to do with any intervention). Or patients receiving this intervention had the following characteristics: … All of those might be the questions the original poster had in mind.
Fair point, I was assuming because they do a single-arm study with pre-post measurements to determine the intervention effect that no effect is expected without treatment (i.e. the natural history is known as you said).
But agree, some more information on the goal of the analysis would be welcome!