Analyzing 'change scores' in an RCT: Alternative to ANCOVA — assumptions violated

I am working on an RCT measuring change in scores from baseline; Total N = 60 with 3 arms (1:1:1, no dropouts). I was going for ANCOVA comparing the follow up values between the 3 groups while adjusting for baseline ones.

However, the variances are quite heterogeneous, and outliers exist. I am not sure if the model would be robust in this scenario. I thought of rank transformation, but is there a better alternative? Perhaps a “non-parametric” ANCOVA or another suitable linear model?

I appreciate your input!

I think the proportional odds model implemented in rms package will be appropriate.

1 Like

Yes semiparametric ordinal models such as the PO model are good choices. But since it is an RCT isn’t the statistical analysis plan already in place?


Thank you so much for your input! Re SAP: The initially-assigned statistician, who later had to leave, powered the trial to analyze the follow up scores while not adjusting for baseline. I did this bit as planned, and the results are great, but I thought adjustment would be more informative.

Adjustment is more effective because it doesn’t make any of these assumptions

  • baseline is relevant (estimated slope can be zero)
  • baseline is linearly related to Y
  • slope is 1.0 for baseline if it is linearly related to Y

But for power/sample size calculations we frequently approximate the final analysis by using the standard deviation of change scores.