t-test degrees of freedom

Submitted by Marieke on

Hi you all,

I am using Matlab Version 2019a, and GraphVar 2.02. I am calculating group comparisons with 4 subgroups that arise when combining 2 factors (factor A and B) so that I am calculating e.g., A0 + B0 vs. A1 + B0. Hence, I have 6 comparisons in total. My total n is 211, but my subgroups have an n of 60, 47, 24, and 80, respectively.

Upon analyzing the output, I encountered two difficulties:


(1) When I it understand correctly, the t value should indicate whether one group or the other has e.g., nodes with greater strength in a specific brain region. If we have positive values, the first group (e.g., A0 +B0) has greater values, if the t-value is negative ,the second group (e.g., A1 +B0) does. Why does the sign of the t-value differ for one brain region depending on the threshold?


(2) Why are the df 207 and not, e.g., (60+47) -2 = 105 for the t-test? As we are comparing separate groups and not the whole sample, I do not understand why the df is caculated 211 (total n) - 4 (number of groups)? If I am not mistaken, this is not the way df for t-tests are calculated?


Thank you so much for making this toolbox available! I am looking forward to your answers! :)





Johann Kruschwitz

Tue, 10/29/2019 - 09:31

Hi Marieke,
as we have talked already on the phone, this is just to not leave this question blanc.
(1) we have figured out that you did not specify the GLM correctly and made clear that GraphVar runs ONE GLM with the specified variables. If you enter several regressors this equals a multiple regression. Thus, GraphVar computes the weights for the unique influence of each regressor (and partials out the shared influence). When you enter highly correlated regressors (as you did by entering different groupings for the same subjects) then the weights are not easily interpretable as one factor may 'steal' the weight of the others (i.e., multicollinearity), which could lead to strange effects.
(2) the Df were different as you not only included a factor with 2 groups but entered several other variables in the GLM that led to a reduction of the total Df.