Reader Query
Question:
I notice that you have reported Within-Subjects Contrasts, here,
rather than Within-Subjects Effects. Is there any reason for this? In my analyses, most
times the Contrasts and Effects tables come out the same, but in some of my analyses,
these differ. Which is the correct table to use?
Answer:
For a two-level factor, these should be equivalent.
For a factor with three or more levels, the main effect considers whether
the factor as a whole exerts a reliable influence on the dependent variable.
The contrasts analysis compares pairs of levels from the multiple levels,
its choice of levels being specified in the options of SPSS. It may for
instance give the reliability of the biggest difference between any two levels.
You should therefore see that the degrees of freedom differ between the main
effects and simple contrasts when the factor has more than two levels.
Where you have a reliable difference in simple contrasts
but not in main effect, it tells you that a pair of the levels considered alone
would show a difference (which you should probably be able to tell from
your trajectory plots), but the multiple levels as a whole do not, either due
to greater noise in other levels or carrying out the comparison with greater
degrees of freedom.
Reader Query
Question:
I've been working through your worksheet to analyse some data.
I notice that you always talk about the main effects of repeated measure factors
using the within subjects contrasts table. This is presumably because your
repeated measures factor has two levels, and so there is no difference between
the within subjects effect table and the within subjects contrast table.
However, if the repeated measures factor has more than two levels (mine has 3),
I only use the within subjects contrast table if the effect is best described
as linear, and if not I use the within subjects effects table and explain the
effect using pairwise comparisons.
Yes, a previous reader noticed that the main effects and
simple contrasts tables differ when the repeated measures factor has more than
two levels :0)
So, in light of this, is it fair to say that when looking at
the ANCOVA, then if your main effect of the repeated measures factor was not
linear, then any interactions between this variable and group or the covariate
should be taken from the within subjects effect table, and but that if my main
effect of the repeated measure factor was significantly linear, then I look at
these interactions using the within subjects contrast table? Or is more
complicated than that, and I should always use the within subjects contrast
table (as in your worksheet) for a reason that I'm not aware of?
Answer:
My intuition is that you should look at the trajectories
to decide. A non-linear repeated measures factor would imply that the difference
in intercepts between the 3 (or more) levels are not evenly spaced (collapsed
over any other factors). Interactions with other factors would indicate how
this spacing was altered at different levels of other factors. The main effects
table is giving you summary of all levels, the simple constrasts a (pre-determined)
choice of two out of these levels. Quite often, when interpreting factors with
more than 3 levels, you end up doing posthoc pairwise comparisons anyhow to see
where the reliable difference is coming from.
The main guideline, I think, is examine the data, and see
which comparisons give the honest picture of what you think is going on.
Reader Response:
Thanks, yes, that does all make sense. I'm glad to hear
that there wasn't some reason for using the contrasts table that I wasn't
aware of. I think I'll extend the rule that I use for ANOVAs then. So, if
my main effect of the repeated measure is best explained as linear, then
when looking at the ANCOVA, I'll look at interaction with that factor using
the contrasts table, as this then explains how the linear main effect
interacts with group or the covariate. For the same reason, when the main
effect of the repeated measures factor is not linear, then it makes more sense
to look at the interactions using the main effect table.
Reader Response:
Having thought a bit more, I think your advice to look
at the pattern of the data is the better approach, as it's perfectly
possible to have a linear effect in one group, but not the other, and
for this to cause an interaction which is therefore not linear, but
related to a group difference at a specific level of the repeated
measures factor only.
© Michael Thomas 2010
Last edited by MT 21/4/10
|