Dear Stephen,
Thanks for pointing us out again to the tutorial - and to the topic we
ourselves started when we noted what Radek noted: suspiciously strong
between-group effects. Since then we got convinced of what we stated in
our previous email, that you should not model the subject factor when making
between-subject comparisons, i.e. modeling main effects of group. I remain
convinced of that after your replies, though IANASE (I am not a statistician
either) I will try to explain why.
The trouble is not, as you state in a previous email:
"You're absolutely right that specifying the effect of subject will remove the
effect of group if you're not careful."
The problem is, it removes variance from the error term that should not be
removed. Your explanation shows this:
>
>The model is
>
> y_ijk = s_i + g_j + c_k + gc_jk + e_ijk
>
>where
> s = subject term
> g = group term (and j = j(i) depends on i)
> c = condition term
> gc = interaction of group and condition term
> e = residual
>
>After estimating and fitting, we have (I'm omitting hats and carats, and not
>taking into account prewhitening/nonsphericity correction)
>
> y_ijk = s_i + g_j + c_k + gc_jk
>
So, the subject means are modeled, are therefore not in the residual, which is
removed after estimating.
Now, I fully agree the above model provides a better model for the y-term.
However, the whole point of ANOVA's is testing any difference in means
against the appropriate error term. When testing for within-subjects effects,
the appropriate term is within-subject variance - it is then appropriate to
remove the between-subject variance from the residual by modeling it.
However, when modeling between-subject effects, as you state:
"In terms of breaking down variance, the intersubject variance is split between
group differences and differences between subjects _within groups_."
The first part is the variance we are trying to explain with a main effect of
group, the second part is the error variance that gives an indication of how
reliable the group effect is. By inserting the subject factor, this variance is
removed from the residual - which leaves only within-subject variance. Jan
made the same claim in his reply to us last summer:
"It is these subject constants that absorb much of the inter-subject
variability present in most imaging data, which in turns leads to more
sensitivity for the experimental effects (including group differences)."
In conclusion, though no doubt including the subject means in the model
provides a better fit (and therefore stronger group effects), I think this is not
appropriate _when looking at between-subject effects_.
Laura
|