 |
 |
Hi Alva, The input for FDR (-i) has to be the image with the uncorrected p-values. This goes for either for conventional voxelwise statistics (produced in randomise with -x) or for TFCE (produced with -T or --T2). The corrected p-values shouldn't be used with FDR, neither the statistics themselves. To get the uncorrected p-values, use the --uncorrp option in randomise then (just add it to the dual_regression script). Consider using the option "-a" in FDR instead of the options -q and --othresh, so that an image with all the adjusted p-values is produced. This image you can then supply to the command cluster, with a threshold of 0.99 (and indeed, these are all 1-p). All the best, Anderson On 24 May 2015 at 02:55, Alva Tang <[log in to unmask]> wrote: > Thanks for your explanations Anderson, > > Following your second point to run FDR on the contrasts, I tried the > following: > > fdr -i dr_stage3_ic0014_tstat1.nii.gz --oneminusp -q 0.01 -m > avg_mask.nii.gz --othresh=fdr_thresh_ic0014_tstat1_withmask -v > > > cluster -i fdr_thresh_ic0014_tstat1_withmask -t 0.99 > > > I followed this thread, > https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind1502&L=FSL&D=0&1=FSL&9=A&I=-3&J=on&d=No+Match%3BMatch%3BMatches&z=4&P=211203 , > but am not sure about the first command. Stage 3 of the dual regression > by default derives 2 images for each contrast: ic#_tfce_corrp and > ic#_tstat. Is the input file the ic#_tstat or do I need to run randomise > again with the --uncorrp option? Also is the ic#_tstat a 1-p image, or > would I have to transform it? I am not sure, I read on the discussion > board that randomise gives 1-p images? > > > > Sorry if this is repetitive and thank you for your help, > > > Alva > > > > On Fri, May 22, 2015 at 4:34 AM, Anderson M. Winkler < > [log in to unmask]> wrote: > >> oops, there was a "Paste" issue and a sentence fell off place at the top >> of the message, but it should be clear as you read down. >> >> On 22 May 2015 at 09:19, Anderson M. Winkler <[log in to unmask]> >> wrote: >> >>> Significant interaction without individual group effectsHi Alva, >>> >>> Please, see below: >>> >>> >>> On 22 May 2015 at 01:09, Alva Tang <[log in to unmask]> wrote: >>> >>>> Dear FSL experts, >>>> >>>> I am looking at differences in resting state networks using dual >>>> regression to see whether connectivity differs across 3 groups as a >>>> function of anxiety. After setting up the design matrix with the >>>> mean-centered values of anxiety, I set up the contrasts below. >>>> >>>> >>>> normal SGA AGA >>>> Anxiety_norm Anxiety_SGA Anxiety_AGA >>>> Slope normal > SGA 0 0 0 1 -1 >>>> 0 >>>> Slope SGA > Slope norm 0 0 0 -1 1 0 >>>> Slope normal > slope AGA 0 0 0 1 0 -1 >>>> Slope AGA > slope normal 0 0 0 -1 0 1 >>>> Slope SGA > Slope AGA 0 0 0 0 1 -1 >>>> Slope AGA > Slope SGA 0 0 0 0 -1 1 >>>> >>>> >>>> 1. If a region is statistically significant for the first contrast, >>>> does it mean that region has increased connectivity for the normal > the >>>> SGA group in relation to anxiety, such that, anxiety has different effects >>>> on this region between groups? >>> >>> >>> >>> Yes. >>> >>> >>>> If this interpretation is correct, how do I visualize the direction of >>>> this effect in the program (to see whether increases or decreases in >>>> anxiety is related to the increased connectivity of this region); Is there >>>> a way to plot the anxiety scores and the activity of that region and across >>>> groups or would I extract these values then plot in another program? >>>> >>> >>> There isn't a direct way to visualise. You'll have to run a few more >>> contrasts to check the direction of each interaction EV (as opposed to the >>> differences), and see if the signs of the statistics in the regions where >>> you found significance are positive or negative. There is a lengthy thread >>> in the mailing list discussing how it can be done, search for "Significant >>> interaction without individual group effects" to see all. The first in the >>> thread should be this one >>> <https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=FSL;3b4f778f.1311>. >>> >>> >>>> >>>> >>>> 2. To correct for multiple comparisons using false discovery rate, >>>> there's a paper (Veer et al., 2010) that inputted the tfce_corrp difference >>>> images and then spatially masked with the binary representation of the >>>> pooled group main effects images. This was to decrease susceptibility to >>>> type 1 errors. I am not understanding how the masking here contributes to >>>> a more stringent threshold or which masks to select? >>>> >>> >>> I just found the paper. They seem to have used TFCE-corrected for the >>> main group effects, and FDR for the between group-effects, and these were >>> masked after doing FDR. I'm not reading the full paper, and perhaps there >>> is justification for this somewhere. The way I would do is not test the >>> main group effects at all (these are known to be different than zero), and >>> test just the between-group differences, using then perhaps FDR. If >>> masking, I'd do it before FDR. Maybe the main group effects were tested in >>> order to produce a mask, which otherwise perhaps would haven't been >>> available, I'm not sure. >>> >>> All the best, >>> >>> Anderson >>> >>> >>> >>> >>>> >>>> If you could please help me, that would be much appreciated. Thank you. >>>> >>>> Alva >>>> >>> >>> >> > > > >