I have data from a randomised control trial of approximately 60 people (30 people in the intervention group, 30 in the control group). I would like to assess whether change in one variable (M) is associated with change in another (Y), which are continuous variables. Ideally I'd run a mediation analysis to see if M mediates the effect of the intervention on the outcome Y, however I am low in power.
Therefore instead I am planning to assess the association between change in M and change in Y within the intervention group only (i.e., ignoring data from the control group). To do this, I am thinking of:
- Creating "change scores" for M and Y by using the post-intervention score for each variable and regressing out the respective baseline score and covariates (i.e., residuals from post-intervention score ~ pre-intervention score + age + sex + other baseline covariates). I will use the same set of covariate for both.
- Then I plan to use a regression to assess whether change in M predicts change in Y. That is, "change in Y" ~ "change in M".
- I have quite a lot of missing data at follow-up so plan to address missing data I am thinking of using multiple imputation, though I am unsure whether to apply this or when, probably at stage 1.
Questions: Does this sound like a reasonable approach that would have greater power than a mediation analysis? Is there a way to complete a similar anlysis using linear mixed models so that I can use MLE to address missing data?
I should probably mention that in the above plan, I use data from two-timepoints. However, for my variable M I have three timepoints (i.e., an additional measurement mid-intervention at 6 months). I also plan to run analyses for several outcomes (Ys) some of which have 3 measures (baseline, 6 months, 12 months), others only have two (baseline, 12 months).
