Very insightful. Thank you so much Mark.
A follow up question: How should we divide the data if we want to
divide them into three groups? By keeping the approx. percentage? or
by keeping approx. number of group members?
On Sat, May 10, 2014 at 2:24 AM, Mark Jenkinson
<[log in to unmask]> wrote:
> Hi,
>
> The question of normality doesn't really apply to the covariates - the distribution really relates to the acquired data, and the noise within that, which will be largely Gaussian for fMRI. Instead, the relevant question is really whether linear correlation, which is what you would get with the GLM, is the appropriate thing to look at here. Transforming the covariate data prior to putting it in the GLM might be a very good option, but the choice of transformation is something that you should consider in light of what you expect to see for the relationship with the MRI data.
>
> Another option, as suggest in point 3, would be to divide the results into a small number of categories and enter each one as a separate EV, as you can then look for any changes between the categories with a 1-way ANOVA style of analysis. It would also allow you to also look at the changes in signal across the categories in a more non-parametric way.
>
> All the best,
> Mark
>
>
>
>
> On 10 May 2014, at 01:03, "Xue, Feng" <[log in to unmask]> wrote:
>
>> Dear folks,
>>
>> We would like to correlate a questionnaire data set (shown in the
>> attached snapshot, first column - these are count data for the
>> frequency of an activity) with our group GLM analysis. However, as you
>> can see, we have a very non-normal and long tailed distribution. We
>> had a few questions:
>>
>> 1. Will automatic outlier de-weighting be sufficient to account for
>> the non-normalcy of the data? What's the mechanism of automatic
>> de-weighting?
>> 2. If it isn't sufficient, would it be recommended to use another
>> transformation (such as square root or log transform) to make the data
>> more normal before running the GLM?
>> 3. Would it be better to create bins/groups from the continuous
>> measure and then to analyze differences among those groups with the
>> GLM?
>>
>> Thanks in advance.
>>
>> --
>> Best Regards
>>
>> Xue, Feng Ph.D.
>>
>> Psychology Department
>> University of Southern California
>> Los Angeles CA, 90089
>> Mobile: +1 213-249-1040
>> ==============================================
>> National Key Laboratory of Cognitive Neuroscience and Learning
>> Beijing Normal University
>> Beijing, China. 100875
>> Mobile: +86 13810154455
>> web: http://psychbrain.bnu.edu.cn
>> <Quant.gif>
--
Best Regards
Xue, Feng Ph.D.
Psychology Department
University of Southern California
Los Angeles CA, 90089
Mobile: +1 213-249-1040
==============================================
National Key Laboratory of Cognitive Neuroscience and Learning
Beijing Normal University
Beijing, China. 100875
Mobile: +86 13810154455
web: http://psychbrain.bnu.edu.cn
|