Hi Matthieu,

Yes, that's right: if the interaction is significant, we would demean then, even though these are group comparisons. I posted the link to Jeanette so as not to have to write the same over and over. I should have been more specific, sorry if this caused confusion.

All the best,

Anderson

On 2 February 2017 at 17:40, Matthieu Vanhoutte <[log in to unmask]> wrote:

Dear Anderson,

I have followed the discussion below concerning design of a three-groups and two continuous variables. I have been blocked concerning demeaning question according to both your answer and Jeanette Mumford's page.

Could you please take a look at my question below inline ?

Best regards,

Matthieu

Le 21 août 2016 à 22:29, Anderson M. Winkler <[log in to unmask]> a écrit :

Hi Elena,

Please, see below:

On 21 August 2016 at 16:37, elena mazza <[log in to unmask]> wrote:

Good evening Anderson,
doing the different designs, some doubts come up

2. Regarding the intercept. i think I didn't understand exactly when I have to use it in the designs.

    a) Why in the design above you suggested to inclued it, whereas in the following design is not necessary?

EV1: BP (coded as 0/1)
EV2: SCZ (coded as 0/1)
EV3: CON (coded as 0/1)
EV4: age BP (put zeros for the non-BP)
EV5: age SCZ (put zeros for the non-SCZ)
EV6: age CON (put zeros for the non-CON)
EV7: sleep efficacy BP (put zeros for the non-BP)
EV8: sleep efficacy SCZ (put zeros for the non-SCZ)
EV9: sleep efficacy CON (put zeros for the non-CON)
EV10: EV4*EV7 (that is, interaction age by sleep efficacy in BP)
EV11: EV5*EV8 (that is, interaction age by sleep efficacy in SCZ)
EV12: EV6*EV9 (that is, interaction age by sleep efficacy in CON)

Note that there is no explicit EV for intercept in this case.

In this design, with 3 groups, the intercept is in fact there, but split across the three groups in EV1, EV2 and EV3. In the other design, with two groups, the intercept is shown explicitly. However, it is possible to reorganise both designs so that the opposite happens, i.e., it is possible to reorganise the design for 2 groups so that the intercept is split across, and it is also possible to reorganise the design with 3 groups so that there is a column full of ones for the intercept. These reorganised designs would lead to identical results. Choosing one or another is just a matter of readability.

 

  b) I also read that demeaning can be skipped if an intercept is entered in the design. So are they (demean-intercept) mutually exclusive?

For these designs that will use contrasts comparing groups there is no need for demeaning.

Please have a look at Jeanette Mumford's page on this topic: http://mumford.fmripower.org/mean_centering/

When we comparing groups with these contrasts:

Differences between groups:
C1: 1 -1 0 0 0 0 0 0 0 0 0 0 F1
C2: 1 0 -1 0 0 0 0 0 0 0 0 0 F1
C3: 0 1 -1 0 0 0 0 0 0 0 0 0

We are in the third case, first contrast of the Jeanette Mumford's table below, aren’t we ? This position seems to indicate to demean data but in your answer you said «  For these designs that will use contrasts comparing groups there is no need for demeaning ». Could you explain me what I didn’t understand ?

 

3. Last question concern the randomise script: when do I have to add -D option?
I read in the forum that if I demean by hand the values, in the design, i don't have to add the -D option. If i don't demean by hand i have to add it.
Is that correct?

For these contrasts there is no need for -D as the intercept is in one way or another in the model and also because there are only group comparisons. Jeanette's page should clarify it.

 

So, demeaning the values by hand AND adding in the script the  -D option is an error?!

Not an error. Randomise would demean again. Since the mean would already be zero, it would subtract zero, to no effect.

All the best,

Anderson

 

Best,
Elena

________________________________
Da: FSL - FMRIB's Software Library <[log in to unmask]> per conto di Anderson M. Winkler <[log in to unmask]>
Inviato: martedì 14 giugno 2016 09.19.27
A: [log in to unmask]
Oggetto: Re: [FSL] interaction design with two continuous variables in two groups

Hi Elena,

With three groups, perhaps the most intuitive way is this:

EV1: BP (coded as 0/1)
EV2: SCZ (coded as 0/1)
EV3: CON (coded as 0/1)
EV4: age BP (put zeros for the non-BP)
EV5: age SCZ (put zeros for the non-SCZ)
EV6: age CON (put zeros for the non-CON)
EV7: sleep efficacy BP (put zeros for the non-BP)
EV8: sleep efficacy SCZ (put zeros for the non-SCZ)
EV9: sleep efficacy CON (put zeros for the non-CON)
EV10: EV4*EV7 (that is, interaction age by sleep efficacy in BP)
EV11: EV5*EV8 (that is, interaction age by sleep efficacy in SCZ)
EV12: EV6*EV9 (that is, interaction age by sleep efficacy in CON)

Note that there is no explicit EV for intercept in this case.

The contrasts are then:

Differences between groups:
C1: 1 -1 0 0 0 0 0 0 0 0 0 0 F1
C2: 1 0 -1 0 0 0 0 0 0 0 0 0 F1
C3: 0 1 -1 0 0 0 0 0 0 0 0 0

Interaction age by group:
C4: 0 0 0 1 -1 0 0 0 0 0 0 0 F2
C5: 0 0 0 1 0 -1 0 0 0 0 0 0 F2
C6: 0 0 0 0 1 -1 0 0 0 0 0 0

Interaction sleep efficacy by group:
C7: 0 0 0 0 0 0 1 -1 0 0 0 0 F3
C8: 0 0 0 0 0 0 1 0 -1 0 0 0 F3
C9: 0 0 0 0 0 0 0 1 -1 0 0 0

Interaction age by sleep efficacy by group:
C10: 0 0 0 0 0 0 0 0 0 1 -1 0 F4
C11: 0 0 0 0 0 0 0 0 0 1 0 -1 F4
C12: 0 0 0 0 0 0 0 0 0 0 1 -1

The t-contrasts (C1-C12) can all be repeated with the signs flipped, for a total of 24 tests. The F-tests, however, remain the same for already being bidirectional (two-tailed), i.e., no need to duplicate them up to F8.

All the best,

Anderson

On 13 June 2016 at 14:30, elena mazza <[log in to unmask]<mailto:[log in to unmask]>> wrote:
Hello Anderson,
Thank you very much for your reply!
I have another question: what if I have three groups (such as bipolar-schizophrenia-controls)?

Thank you in advance.

All the best,
Elena
________________________________
Da: FSL - FMRIB's Software Library <[log in to unmask]<mailto:FSL@JISCMAIL.AC.UK>> per conto di Anderson M. Winkler <[log in to unmask]<mailto:[log in to unmask]>>
Inviato: sabato 11 giugno 2016 08.56.03
A: [log in to unmask]<mailto:FSL@JISCMAIL.AC.UK>
Oggetto: Re: [FSL] interaction design with two continuous variables in two groups

Hi Elena,

This would be a three-way interaction. The design is almost fine, only need to replace the intercept (last EV) for two new EVs one for each group. Instead, how about coding as this:

EV1: intercept
EV2: +1/-1 for group
EV3: age
EV4: sleep efficacy
EV5: EV2*EV3 (interaction group by age)
EV6: EV2*EV4 (interaction group by sleep efficacy)
EV7: EV3*EV4 (interaction age by sleep efficacy)
EV8: EV2*EV3*EV4 (interaction group by age by sleep efficacy)

The contrasts are then:
C1: [0 0 0 0 0 0 0 1]
C2: [0 0 0 0 0 0 0 -1]

All the best,

Anderson

On 10 June 2016 at 15:22, Elena Mazza <[log in to unmask]<mailto:[log in to unmask]>> wrote:
Dear FSL users,
I had a question about how to set up a design.
I want to see if there is an interaction effect between age and sleep efficacy on white matter integrity, and if this interaction differs in the two groups that I considered (Healthy subjects VS Patients).

I know that considering just ONE sample the design will be as follows:

EV1: age
EV2: sleep efficacy
EV3: product EV1*EV2.
EV4: intercept

where the interaction is tested with:

C1: [0 0 1  0]

C2: [0 0 -1 0]

- Considering TWO groups (Healthy control->first group Vs Patients -> second group), do I have to split each variables in two columns as follows?

EV1: age of Healthy controls
EV2: age of Patients
EV3: sleep efficacy of Healthy controls
EV4: sleep efficacy of Patients
EV5: product EV1*EV3
EV6: product EV2*EV4
EV7: intercept

where the difference between groups in the interaction is tested with:

C1: [0 0 0 0  1  -1 0]

C2: [0 0 0 0 -1   1  0]

Is it correct?
how about the intercept?


Thank you in advance!

Best,
Elena Mazza