Dear Guilherme,
The sequence of regressors in the design matrix does not change the
beta-values, their contrasts or the T values.
The value of each beta reflects the contribution of its regressor that
cannot be explained by the remaining regressors in the design matrix
(c.f. a partial correlation)
The fact that you get significant results from the first regressor
(all events) means that there must have been some events that were
not modelled by the seven semantic-specific regressors. It may be that
the semantic attributes of the events omitted from the seven have an
activating effect.
This also means that correlations or collinearity among the seven
regressors will reduce the apparent significance of your results using
a T contrast. Perhaps if you grouped certain semantic categories with an
SPM{F} you would get a better picture of what is going on. For example if
regressors 1, 2 and 3 all modelled words with some attribute (e.g.
referred to animate objects), you could use an F contrast
0 1 0 0 0 0 0 0
0 0 1 0 0 0 0 0
0 0 0 1 0 0 0 0
I hope this helps - Karl
Estimated Prof. Friston,
I have a short question about the interpretation of my parametric design. I
tried firstly the SPM discussion group, but I didn't obtain a response, then
I thought I should try to contact You directly.
My design is rapid event-related. I have created a design matrix with a
single vector of onsets cond_1 which comprises all events in the experiment
and I modeled cond_1 using 7 regressors reg_1 to reg_7, which represent
diverse semantic properties of stimuli which are of interest. Thereafter I
created t contrasts in order to look at the effect of each regressor as well
as at the effect of their combinations.
My question concerns the contrasts. Does the sequence of entering the
regressors in the design matrix change the beta-values of the contrasts and
consequently their t-values like in conventional multiple regression models?
That means, when I look the contrast for reg_2, am I looking at the
equivalent to the null-order correlation (activation vs. reg_2) or am I
looking at the partial correlation (activation vs. reg_2 -> reg_1)?
I found very much activation in the t contrast taking only cond_1 and much
weaker activation in the contrasts taking only single regressors. That makes
me confused because the semantic effects I am pursuing are very strong in
the RT and also in a prior EEG study. I would be very pleased if You have
any comment about my data.
Best regards,
Guilherme Wood
|