help
help > RE: Contrast specification for Group comparison in a multi-center RCT (controlling for scanner)
Jan 26, 2021 12:01 AM | Alfonso Nieto-Castanon - Boston University
RE: Contrast specification for Group comparison in a multi-center RCT (controlling for scanner)
Hi Till and Marcel,
Regarding over-parametrized models, that is perfectly fine. The general linear model handles rank-deficient design matrices without any problem. In this particular case, that also means that you will get exactly the same results to your [-1 1 0 0... 0] contrast evaluating between-group differences whether you enter the 8-sites as eight 0/1 control covariates or as seven mean-centered control covariates (and the same for sex, you may enter that as a single 1/-1 variable, or as two 0/1 variables, or a single 0/1 variable, and the results of the between-group difference analysis will not change at all)
If you want to double-check that the design is correct I would suggest to click on the 'n=..." text on the second-level results tab, which will bring up the design matrix and stat details. You should see something like the image attached (this example that is just for a two-group comparison with three control sites and a [-1 1 0 0 0] contrast). In the bottom-right corner, the "5 predictors (4 independent)" text tells you that this design had 5 regressors, but it is over-parameterized as it has only 4 independent factors (because in this example the three sites were entered using three 0/1 control covariates). In Till's case, with an analysis including [patients controls age sex center1 center2 center3 center4 center5 center6 center7 center8] and a [-1 1 0 0 ... 0] contrast, you should see "12 regressors (11 independent)", and if you decide to enter sex as two different covariates, e.g. [patients controls age male female center1 center2 center3 center4 center5 center6 center7 center8] you would then see "13 regressors (11 independent)", as that additional regressor does not add any information to the design.
Hope this helps
Alfonso
Originally posted by Marcel Daamen:
Regarding over-parametrized models, that is perfectly fine. The general linear model handles rank-deficient design matrices without any problem. In this particular case, that also means that you will get exactly the same results to your [-1 1 0 0... 0] contrast evaluating between-group differences whether you enter the 8-sites as eight 0/1 control covariates or as seven mean-centered control covariates (and the same for sex, you may enter that as a single 1/-1 variable, or as two 0/1 variables, or a single 0/1 variable, and the results of the between-group difference analysis will not change at all)
If you want to double-check that the design is correct I would suggest to click on the 'n=..." text on the second-level results tab, which will bring up the design matrix and stat details. You should see something like the image attached (this example that is just for a two-group comparison with three control sites and a [-1 1 0 0 0] contrast). In the bottom-right corner, the "5 predictors (4 independent)" text tells you that this design had 5 regressors, but it is over-parameterized as it has only 4 independent factors (because in this example the three sites were entered using three 0/1 control covariates). In Till's case, with an analysis including [patients controls age sex center1 center2 center3 center4 center5 center6 center7 center8] and a [-1 1 0 0 ... 0] contrast, you should see "12 regressors (11 independent)", and if you decide to enter sex as two different covariates, e.g. [patients controls age male female center1 center2 center3 center4 center5 center6 center7 center8] you would then see "13 regressors (11 independent)", as that additional regressor does not add any information to the design.
Hope this helps
Alfonso
Originally posted by Marcel Daamen:
Dear Till,
I wonder whether the model is overparameterized because you should only need 7 variables for dummy-coding the 8 centers/ scanners (i.e. scanner 8 implicitly specified by a "0" for all seven regressors)?
Best wishes,
Marcel
Originally posted by Till Langhammer:
I wonder whether the model is overparameterized because you should only need 7 variables for dummy-coding the 8 centers/ scanners (i.e. scanner 8 implicitly specified by a "0" for all seven regressors)?
Best wishes,
Marcel
Originally posted by Till Langhammer:
Dear Alfonso and Forum members,
I would like to compare different groups. Groups are diagnosis groups in an RCT for different anxiety disorders.
We have 8 different sites and therefore different scanners.
My Model
[1 -1 0 0 0 0 0 0 0 0 0 0]
[patients controls age sex center1 center2 center3 center4 center5 center6 center7 center8]
I am not sure if this works. The results explorer and folder is named
"patients(1).controls(-1).age(0).sex(0).Dr7573199821223284"
I would have expected:
"patients(1).controls(-1).age(0).sex(0).center1(0).center2(0).center3(0).center4(0).center5(0).center6(0).center7(0).center8(0)"
I chose user defined between subject contrast and inserted [1 -1 0 0 0 0 0 0 0 0 0 0], as it is not possible to choose this contrast in the drop down menu.
How can I check if CONN calcualted the correct model?
Thanks
Till
I would like to compare different groups. Groups are diagnosis groups in an RCT for different anxiety disorders.
We have 8 different sites and therefore different scanners.
My Model
[1 -1 0 0 0 0 0 0 0 0 0 0]
[patients controls age sex center1 center2 center3 center4 center5 center6 center7 center8]
I am not sure if this works. The results explorer and folder is named
"patients(1).controls(-1).age(0).sex(0).Dr7573199821223284"
I would have expected:
"patients(1).controls(-1).age(0).sex(0).center1(0).center2(0).center3(0).center4(0).center5(0).center6(0).center7(0).center8(0)"
I chose user defined between subject contrast and inserted [1 -1 0 0 0 0 0 0 0 0 0 0], as it is not possible to choose this contrast in the drop down menu.
How can I check if CONN calcualted the correct model?
Thanks
Till
Threaded View
| Title | Author | Date |
|---|---|---|
| Till Langhammer | Jan 12, 2021 | |
| Marcel Daamen | Jan 13, 2021 | |
| Alfonso Nieto-Castanon | Jan 26, 2021 | |
| Till Langhammer | Jan 26, 2021 | |
| Till Langhammer | Jan 14, 2021 | |
| Marcel Daamen | Jan 14, 2021 | |