I'm unable to replicate my visual data with %signal change values in a between group analysis. If anyone has time to look at this and offer help, I'd be very appreciative. Here are the steps I'm taking:

1. I'm running an event-related analysis with 5 conditions. I'm interested in the difference between conditions 3 and 4 (specifically, condition 3 minus condition 4)

2. I run the analysis above on all my subjects, then do a between group analysis where I observe a significant difference in activation in group A minus group B in vmPFC

3. I create and save a label in the vmPFC in the between group image, and use label2label to apply that label to each individual

4. I then use func2roi and roisummary to extract activation values for each subject

5. In the roisummary spreadsheet, I use the offset value to calculate %signal change for each condition, I split my individual subjects that I used in my between group analysis (step 2) into two groups to compare %signal change values and am getting no significant differences.

How is this possible if there is very robust activation when I run the analysis with selxavg, isxconcat, and mriglmfit?? Where might I be making a mistake?

thanks for your quick answer. I think that the assumption of linearity can often be made in genetics and in large GWAS studies the coding is done as suggested by Nick. However, Nick's solution does not suit this particular case since one of the homozygous groups has less than 10 subjects which, in my opinion, might skew our results. Therefore we would like to lump homozygotes and the carriers together and have 2 groups (instead of 3).

Unfortunately I and Stefan Brauns do not understand your reply, Doug. We were suggesting to include the binary variable for genotype as a "variable" ( covariate) instead of a "class" (factor) in the FSGD. This would enable us to get around the problem of "small cell sizes". Why do you say this is not possible with an FSGD?

We would just specify

and SNP would be 0 and 1 instead of 0, 1 and 2 (as suggested by Nick)

This is how the header would look like:

GroupDescriptorFile 1
Title    rs8216888_status
MeasurementName thickness
Class SCZMALEMGH
Class SCZFEMALEMGH
Class HCMALEMGH
Class HCFEMALEMGH
Class SCZMALEIA
Class SCZFEMALEIA
Class HCMALEIA
Class HCFEMALEIA
Class SCZMALEUMN
Class SCZFEMALEUMN
Class HCMALEUMN
Class HCFEMALEUMN
Class SCZMALEUNM
Class SCZFEMALEUNM
Class HCMALEUNM
Class HCFEMALEUNM
Variables age SNP

In contrast, if we would include it as a class - we would have 32 instead of 16 classes and then "Variables age"

The question is if it violates some assumptions if we specify a "variables" (covariate) which is binary.

Message: 6
Date: Wed, 02 Dec 2009 13:18:20 -0500
From: Douglas N Greve <greve@nmr.mgh.harvard.edu>
Subject: Re: [Freesurfer] "dummy variable" in mri-glmfit
To: Stefan Brauns <stefan.brauns@googlemail.com>
Cc: freesurfer <freesurfer@nmr.mgh.harvard.edu>
Message-ID: <4B16AF6C.9080103@nmr.mgh.harvard.edu>
Content-Type: text/plain; charset=UTF-8; format=flowed

Do you mean having just another column in your design matrix with 0s and
1s? You can do this, but not with an FSGD. You'll have to supply your
own matrix. An easy way to do this would be to run mri_glmfit with and
FSGD without the genotype. This will create a matrix Xg.dat in the
output dir, then just modify that matrix and pass it to a new call to
mri_glmfit

doug

Stefan Brauns wrote:
> Hi there,
>
> we would like to test the effect of a binary variable (genotype
> = carrier vs. homozygous) on cortical thicknes in mri-glmfit. Since we
> are also controlling for gender and aquisition site (4 sites) we
> already have 16 groups. In order to control for age as a covariate we
> need at least 2 subjects per group to be able to estimate an age slope.
>
> If we include the aforementioned binary variable (genotype) as a
> factor (two different "groups"), we would have 32 groups and
> unfortunately not enough subjects per group.
>
> Is it possible to include binary variables ("dummy variable" coded as
> 0 and 1) such as genotype or gender as covariates (slope), in order to
> reduce the number of groups and examine the effect on thickness? In
> simple regression this would not affect the results - what would we
> expect here?
>
> Many thanks,
>
> Stefan
>