RCTs: uneven temporal spacing and changes in direction and magnitude?

data analysis
1

RCTs: How should we model uneven temporal spacing and changes in direction and magnitude?

Hey everyone,

I was peer-reviewing a paper and I got this doubt. Which model would you pre-register for this study?

Assessments were conducted at baseline (V1), approximately one month later but before surgery (V2), and six weeks after surgery (V3).

The expected trajectory was an improvement from V1 to V2, followed by a 50% loss of that improvement from V2 to V3.

8b1e904ed595567e015e853bec6ce888
8b1e904ed595567e015e853bec6ce888558×376 125 KB

The challenge is that the time intervals are unequal, and the expected changes differ in both direction and magnitude.

Which approach would you choose?

  1. Splines: With only three assessment points, I suspect that modelling the trajectory using splines would be unnecessarily complex and potentially unstable.

  2. Conventional change scores: I believe these would underestimate at V2 and overestimate at V3.

  3. Two separate ANCOVAs: One ANCOVA comparing V1 and V2, and another comparing V1 and V3. This would allow us to estimate the effect at each follow-up separately, but would not directly model the full trajectory.

Would you recommend any other approach?

Bonus question: Is there a meaningful reason to test the contrast between V2 and V3? I often see authors report this comparison, but I am not always sure what additional information it provides beyond comparing each follow-up with baseline.

Thank you!

2

The graph is for a single-arm study. Doesn’t the RCT have two arms? Say more about the question of interest. Depending on the question you might have two baselines and only one follow-up.

Avoid change scores. Use days from the index date treat time as continuous.

1 Like
3

I forgot about the other line.:blush: Yes, it is an RCT. I was assuming that the control group remains flat.

It is about EXERCISE PREHABILITATION IN CANCER PATIENTS.

Meanwhile I was playing with LLMs and to my surprise, using time as categorical, all models behave nicely.

But whenever I treated time as continuous, only the ancova multi-level model behaves nicely

m1 ← lmer(y ~ time*group + (1|ID)) # y0+y1+y2, group main effect kept

m2 ← lmer(y ~ time*group + y0 + (1|ID)) # y1+y2 multilevel ANCOVA

4

Not a safe assumption.

5

You are right.

To which model would you gravitate if you assume that control remains flat?

To which model would you gravitate if you assume that control will change in the same direction as intervention group but let’s say with half magnitude?

Actually I imagined that you would endorse the same model for both cases.