Re: Analysing baseline and follow-up measurements

Richard Ulrich wrote: 
On 20 Oct 2005 00:03:26 -0700, andrea.meyer@xxxxxxxxx wrote:


Hello everybody
In a recent statistical note, Vickers & Altman (British Medical
Journal, January 2005) have shown that in randomised trials where for
example two treatment outcomes are compared, the ANCOVA aproach using
the baseline score of the outcome as covariate is superior to analysing
either change scores (i.e. outcome minus baseline score) or simply
outcomes ignoring baseline scores. If, however, the focus is not only
on comparing two treatments at follow-up but at the same time on
testing for changes between baseline and follow-up (and also whether
thes changes vary across treatments, i.e. time x treatment interaction)
the ANCOVA approach obviously does not address the right question it
does not analyse changes.


This final statement is implicitly wrong. Under assumptions of the model, and with randomization
(so the PRE's are the same), the "regressed change" analyzed
by the ANCOVA is "change" without a certain amount of noise. Thus, it *is* an analysis of change, and it is the preferred one.

To reinforce what Rich said, in a parallel group design, the primary contrast is the between-grouop comparison. Keep it simple. Get the estimate of the parallel group treatment effect, adjusted for baseline. Changes within group are irrelevant - they have secular trend effects, natural history of disease effects, etc. I'm assuming there is only one follow-up measurement.

Frank Harrell


When the groups differ at PRE, nothing is as simple.


                  My question is whether the same ANCOVA
approach would work if instead of taking follow-up scores as dependent
variable we simply chose change scores. The so obtained model has
change scores (outcome minus baseline score) as dependent and treatment
and baseline score both as independent variables in the model. The
baseline score is only used to correct for baseline differences between
treatments. This model leads to exactly the same results for the
treatment and the treatment x time interaction as the ANCOVA model


- Perhaps I am misunderstanding something. If PRE is the
covariate, with POST as the outcome, I see TREATMENT as measuring what was tested as TR x Time in the repeated measures; I do not see any separate interaction in the ANCOVA.



suggested by Vikers and Altman but has the advantage of focussing on
temporal changes rather than singe points in time.
I wonder what others think of this approach and whether e.g. the fact
that the baseline scores are necessarily correlated with the outcomes
(changes scores) would lead to problems (i.e. inflated correlation)?
Any suggestion s are welcome.



Bruce responded to this.

.

Relevant Pages

  • Re: Analysing baseline and follow-up measurements
    ... In a recent statistical note, Vickers & Altman have shown that in randomised trials where for example two treatment outcomes are compared, the ANCOVA aproach using the baseline score of the outcome as covariate is superior to analysing either change scores or simply outcomes ignoring baseline scores. ... My question is whether the same ANCOVA approach would work if instead of taking follow-up scores as dependent variable we simply chose change scores. ...
    (sci.stat.consult)
  • Analysing baseline and follow-up measurements
    ... the baseline score of the outcome as covariate is superior to analysing ... outcomes ignoring baseline scores. ... thes changes vary across treatments, ... variable we simply chose change scores. ...
    (sci.stat.consult)
  • Re: Analysing baseline and follow-up measurements
    ... January 2005) have shown that in randomised trials where for ... > example two treatment outcomes are compared, the ANCOVA aproach using ... > the baseline score of the outcome as covariate is superior to analysing ... > thes changes vary across treatments, ...
    (sci.stat.consult)
  • Why ANCOVA?
    ... Data was collected from different individuals in each treatment group. ... ANCOVA analyses for each continuous test score (36 separate ... used to control for continuous "covariates", ...
    (sci.stat.edu)