Hi Ioanna,
Regarding question 1, I wouldn't say you necessarily have to covary for
baseline scores but there are times when it is meaningful to do so. For
instance, we studied an intervention to reduce cognitive decline in the
elderly over 3-5 years, measured by using individual slopes of change.
We tested the equivalence of the randomised groups, as you have done,
and that reassured us that the randomisation worked as we'd hoped and
the groups were balanced. However, we still included baseline cognitive
scores as a covariate because the literature suggests that where
patients start from has an influence on how they decline. In other
studies we routinely used change from baseline scores as the outcome,
and here we included baseline as a covariate because the change scores
obscured where the subjects had started from. You do lose some degrees
of freedom by including more variables, so I would suggest only
including if you have an apriori rationale for doing so (other than
'just in case'), or if you're attempting to control for an imbalance in
your groups that was unforeseen.
Hope that helps,
Brian
On 24/04/2010 03:47, Ioanna Vrouva wrote:
> Dear All,
> I would really appreciate your advice regarding the following. I have
> a set of data (pre-post continuous scale as within subjects factor),
> and 2 groups (experim. vs control)-as within subjects factor.
>
>
>
> QUESTIONS:
> 1) I have read in some places that if you have pre-test scores, an
> ANCOVA should be done, using pre-test as the covariate. I have done
> some t-test on pre-test data between the 2 groups and they are not
> different, so should the pre test scores be still used as a covariate
> or not?
>
> 2) I went through my positively variables to transform the ones that
> were skewed, however this did not make them any better and Kilgromov
> stat was still violated normality, as was skewness z-scores. As there
> is no non-parametric test equivalent to a mixed design (as far as i
> know) I am tempted to go ahead with the repeared measures anova ahead
> leaving the data as it is! Is this a problem and how could I correct it?
>
> 3) I have therefore stuck with a 2x2 mixed design anova (on
> untransformed data), with group (exp vs cont) as the between subjects
> factor and time (pre vs post) as the within subjects factor.
> In the output the Mauchly's test wasn't reported (see below)-is it
> because there are only 2 levels to my factors? Is it a problem?
>
> Mauchly's W: 1.000
> Approx. Chi-Square .00
> df 0
> Sig. .
>
> Greenhouse-Geisser 1.
> Huynh-Feldt 1.
> Lower-bound 1.
>
>
>
> many thanks
> Ioanna
|