Hi Ilana,

Normality isn't a requirement for FDR to work. However, if the distribution of your statistic isn't normal *under the null hypothesis*, neither follows t, F or any other well known parametrizable distribution, then you have to obtain p-values in a different way, not from the cdf/pdf of these distributions. Most likely you'd have to use permutations.

But do you have any evidence that your cortical thickness measurements are not normal after regressing out your model? The histogram that appears in qdec window tells you the frequency of the statistics across space for all the vertices. This is not the same as the distribution of the residuals, nor of the statistic for each vertex if the same experiment were to be repeated many, many times. You could apply a normality test to each vertex using the residuals of the GLM. I wouldn't bother much with that, though. I see no reason why thickness measurements wouldn't be normal.

All the best!

Anderson


On 05/16/2011 03:18 PM, Ilana Hairston wrote: 
thanks!  very helpful.
not to be argumentative, but if my Z distribution is not normal, is there meaning to FDR?


On May 16, 2011, at 2:08 PM, Anderson Winkler wrote:


Hi Ilana,

I'll let the first question to the FS experts to answer. For the second:


1. Is it legitimate toconclude from this that the regressor's overall effect is that it's associated with a thinner cortex (e.g., older subjects have a thinner cortex), eventhough most of the cortex does not survive FDR or multiple comparison correction?

Even if the effect exists, your experiment wasn't able to detect it as 
significant, since nothing survived the multiple testing correction you 
used. Most scientific journals wouldn't accept conclusions based on 
non-significant results, despite how compelling the hypothesis might be.


2.  If yes, is there a way of exporting the distribution, or getting some numerical representation of the negative and positive values across the cortex, or within aparc annotations.

The p-values are in the files sig.mgh, in subdirectories of the qdec 
directory where the results of the analyses were stored. There are 
multiple ways to open and a suggestion is to convert to ASCII using the 
-c option of mris_convert, then open in Octave/Matlab with dlmread.


3. Is there a way to set FDR separately for positive and negative values?

The plain answer is no, because for the FDR procedure to work, it's 
necessary that the p-values under the null are uniformly distributed 
between 0 and 1 (from the definition of p-values). If you drop part of 
your distribution, then the p-values are no longer distributed like 
that. It turns out, however, that in some particular cases, part of the 
distribution does not actually exist (see Self & Liang, J Am. Stat 
Assoc, 1987 for some cases, particularly the Cases 5 and 7). A 
workaround is to divide the q by a constant that depends on the fraction 
of the distribution that known to be missing (e.g., for Case 5, instead 
of q=0.05, one could use q=0.025 to obtain the same 5% of false 
discoveries). However, these particular cases I believe don't apply to 
your scenario, so please, refrain from applying FDR separately only on 
positives or negatives.

There is also a second reason for not applying FDR separately: for each 
set (positive and negative), the number of tests would be reduced, on 
average, by half, alleviating the multiple testing problem and making, 
on average, twice as easy for results to survive the threshold, 
inflating the amount of false discoveries, something undesired.

Hope this helps!

All the best,

Anderson


On 05/16/2011 09:53 AM, Ilana Hairston wrote:

Hi there,
First the simple question - is there a way to run mris_anatomical_stats on all my subjects at once generating a single output file?

Second -  actually comprised from several questions:  when looking a the unthresholded group analysis on cortical thickness (in the averaged space), the distribution of negative and positive z values is not necessarily normal.  i.e., the range of negative values maybe be 0 to -5, and the range of positive values 0 to +3.
1. Is it legitimate toconclude from this that the regressor's overall effect is that it's associated with a thinner cortex (e.g., older subjects have a thinner cortex), eventhough most of the cortex does not survive FDR or multiple comparison correction?
2.  If yes, is there a way of exporting the distribution, or getting some numerical representation of the negative and positive values across the cortex, or within aparc annotations.
3. Is there a way to set FDR separately for positive and negative values?

thanks
ilana






Ilana Hairston
hairstonster@gmail.com
*************************
Our ignorance is not so vast as our failure to use what we know.
—M. King Hubbert, peak oil prophet




_______________________________________________
Freesurfer mailing list
Freesurfer@nmr.mgh.harvard.edu
https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer


The information in this e-mail is intended only for the person to whom it is
addressed. If you believe this e-mail was sent to you in error and the e-mail
contains patient information, please contact the Partners Compliance HelpLine at
http://www.partners.org/complianceline . If the e-mail was sent to you in error
but does not contain patient information, please contact the sender and properly
dispose of the e-mail.




_______________________________________________
Freesurfer mailing list
Freesurfer@nmr.mgh.harvard.edu
https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer

Ilana Hairston
hairstonster@gmail.com
*************************
Our ignorance is not so vast as our failure to use what we know.
—M. King Hubbert, peak oil prophet