Hi all,
We received a question about the Vickers and Altman article about change scores.
From Dr Hilton:
"Repeated assessments of an outcome within a respondent are correlated. For example, if we weigh ourselves every day, we will see little variation among the measurements because weights are highly correlated from day to day. Accounting for intraperson correlation in outcomes reduced the overall variance of the outcome, which also accounts for interperson variation".
Vickers and Altman proposes three types of analysis for when your RCT has baseline and follow-up measurements of your outcome: (1) analyze follow-up outcome data only, ignoring baseline differences; (2) analyze change scores (pre minus post); or (3) adjust for the baseline outcome measurement as a covariate in your model.
For (1) and (2), ANOVA can be used to test differences in means between groups of a single continuous variable (i.e. change scores or follow-up outcome data), while for (3) ANCOVA can be used to test differences in the follow-up outcome between groups, while statistically controlling for other covariates (in this case, baseline values of your outcome).
Though it seems that the authors advocate for using ANCOVA, the use of either one of these three methods in your randomized trial will dependent on various assumptions, including the correlation between pre and post values -- which they point out in the paper.
P.S. This paper is focused on randomized studies and recommendations may not apply for non-randomized studies.
Michelle