Thank you very much Anderson,
I am not able to digest the two type of voxels in the region(s) where  interaction is significant.
Just forget our interaction [-1 1], coming back to our interaction [1 -1]
In first type of voxels, I have regions which shows increased FA values as a function of Age for my Group 1, while decreased FA values as a function of age for my Group2.
In second type of voxels since positive sign is not significant we assume that FA values for both Group 1 and Group 2 are decreasing as a function of age.

But these second type of voxels is looking like union of two regions mentioned and some other regions in first type of voxels.

I can not understand both Group 1 and Group2 decrease with age keeping with findings of my first type of voxels.
How is this possible?

Thank you for your kind assistance.



On Thu, Jun 12, 2014 at 2:02 PM, Anderson M. Winkler <[log in to unmask]> wrote:
Hi Himanshu,

Please, see below:


I am afraid when you suggest for C1[1 -1 0 0] and C2[-1 1 0 0] in this design


Hmm, actually I wasn't suggesting... I thought that these were your initial contrasts because normally we don't know beforehand whether the interaction is significant. The existence of C1 and C2 was only useful in my earlier email so that the *_tstatN.nii.gz files could be identified properly.

 
Does this C1 and C2 make sense with the current interaction design (Two groups with continious covariate interaction)

Given that the interaction is significant, no, these two contrasts aren't all that useful.

 
Contrast C1 and C2 makes sense when the two groups are coded as 1's and 2's (i.e for Two sample unpaired T-Test)
while in this it is suggested and advised to use only 1's in group field refer---http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/GLM#Two%20Groups%20with%20continious%20covariate%20interaction

Since C1 and C2 are not used anywhere and as per your suggestions,I got these results

For the interaction [1 -1]
1) Positive signs group 1 (using the contrast [1 0])---Significant
2) Negative signs group 1 (using the contrast [1 0])---Significant
3) Positive signs group 2 (using the contrast [0 1])---NOT Significant
4) Negative signs group 2 (using the contrast [0 1])---Significant


For the interaction [-1 1]
5) Positive signs group 1 (using the contrast [1 0])---Not Significant
6) Negative signs group 1 (using the contrast [1 0])---Not Significant
7) Positive signs group 2 (using the contrast [0 1])----Not Significant
8) Negative signs group 2 (using the contrast [0 1])---Not significant


Ignore the significance images (p-values) for the contrasts [1 0] and [0 1]. They aren't useful and you don't even need to open them. It's just the signs that matter.

However, if by "significant" above you actually mean that there are some voxels marked as 1 in an otherwise full of zeros image, and as "not significant" an all-zero map, then ok. If that is the case, and only if that is the case, then what you are showing tells you that the region(s) in which the interaction was significant there are two types of voxels:
- In some, the slope is positive for group 1 and negative for group 2, that is, FA increases with age for group 1, but decreases with age for group 2.
- In others, the slope is negative for group 1, and also negative for group 2, meaning that FA decreases with age for both groups, although faster for group 2 (we know it's faster for group 2 and not for group 1 because the significant interaction was the [1 -1], not [-1 1]).

This also tells that nowhere in the region with significant interaction the slope for group 2 is positive, i.e., it's always negative, meaning that for group 2, FA decreases with age in every voxel where the interaction was significant.

 
Thus it can be concluded that the group difference for mean FA varies with age, but can we predict anything  else (lets say direction)  from these results.if yes then How?
Kindly rectify me, if I may be wrong, and help me in constructing the conclusion in a more representable form.

Hope the above helps. Please feel free to ask if not.


All the best,

Anderson




On Tue, Jun 10, 2014 at 11:10 PM, Anderson M. Winkler <[log in to unmask]> wrote:
Hi Himanshu,

You need to add the two more contrasts to check the signs for the slopes of each group. I thought your contrasts were:

C1: 1 -1 0 0
C2: -1 1 0 0
C3: 0 0 1 -1
C4: 0 0 -1 1

And I was suggesting that you added two more, to check the signs:

C5: 0 0 1 0
C6: 0 0 0 1

Please, try again to follow the email that I sent yesterday, now with this in mind.

All the best,

Anderson



On 10 June 2014 14:05, Himanshu Joshi <[log in to unmask]> wrote:
I think there is a confusion, I am unable to find error so I am sending my design and the procedure.
Kindly rectify .
During my interaction analysis  I had two contrasts to study interaction
1 1.000000e+00 0.000000e+00 3.000000e+00 0.000000e+00
1 1.000000e+00 0.000000e+00 -4.000000e+00 0.000000e+00
1 1.000000e+00 0.000000e+00 1.000000e+00 0.000000e+00
1 1.000000e+00 0.000000e+00 -2.000000e+00 0.000000e+00
1 1.000000e+00 0.000000e+00 3.000000e+00 0.000000e+00
1 1.000000e+00 0.000000e+00 -4.000000e+00 0.000000e+00
1 1.000000e+00 0.000000e+00 -5.000000e+00 0.000000e+00
1 1.000000e+00 0.000000e+00 2.000000e+00 0.000000e+00
1 1.000000e+00 0.000000e+00 -4.000000e+00 0.000000e+00
1 1.000000e+00 0.000000e+00 4.000000e+00 0.000000e+00
1 1.000000e+00 0.000000e+00 2.000000e+00 0.000000e+00
1 1.000000e+00 0.000000e+00 7.000000e+00 0.000000e+00
1 1.000000e+00 0.000000e+00 1.000000e+00 0.000000e+00
1 1.000000e+00 0.000000e+00 -3.000000e+00 0.000000e+00
1 1.000000e+00 0.000000e+00 1.000000e+00 0.000000e+00
1 0.000000e+00 1.000000e+00 0.000000e+00 -2.000000e+00
1 0.000000e+00 1.000000e+00 0.000000e+00 7.000000e+00
1 0.000000e+00 1.000000e+00 0.000000e+00 2.000000e+00
1 0.000000e+00 1.000000e+00 0.000000e+00 2.000000e+00
1 0.000000e+00 1.000000e+00 0.000000e+00 3.000000e+00
1 0.000000e+00 1.000000e+00 0.000000e+00 -5.000000e+00
1 0.000000e+00 1.000000e+00 0.000000e+00 2.000000e+00
1 0.000000e+00 1.000000e+00 0.000000e+00 -1.000000e+00
1 0.000000e+00 1.000000e+00 0.000000e+00 1.000000e+00
1 0.000000e+00 1.000000e+00 0.000000e+00 -2.000000e+00
1 0.000000e+00 1.000000e+00 0.000000e+00 -4.000000e+00
1 0.000000e+00 1.000000e+00 0.000000e+00 -4.000000e+00
1 0.000000e+00 1.000000e+00 0.000000e+00 -4.000000e+00
1 0.000000e+00 1.000000e+00 0.000000e+00 2.000000e+00
1 0.000000e+00 1.000000e+00 0.000000e+00 -4.000000e+00





Age1>Age2 0.00E+000 0.00E+000 1.00E+000 -1
Age2>Age1 0.00E+000 0 -1 1.00E+000

for contrast 1 (Age1>Age2) I get one tstat file (i.e. xx_tstat1)
for contrast 2(Age2>Age1)  I get one tstat file (i.e. xx_tstat2)

Then
For the interaction [1 -1]
fslmaths xx_tstat1-thr 0 -bin maskpos_tstat1  followed by fslmaths xx_tfce_corrp_tstat1 -thr 0.95 -mul maskpos_tstat1 inter1_grp1_pos
fslmaths xx_tstat1 -mul -1 -thr 0 -bin maskneg_tstat1 followed by fslmaths xx_tfce_corrp_tstat2 -thr 0.95 -mul maskneg_tstat1 inter1_grp1_neg

For interaction [-1 1]
fslmaths xx_tstat2 -thr 0 -bin maskpos_tstat2 followed by  fslmaths xx_tfce_corrp_tstat1 -thr 0.95 -mul maskpos_tstat2 inter2_grp1_pos
fslmaths xx_tstat2 -mul -1 -thr 0 -bin maskneg_tstat2 followed by  fslmaths xx_tfce_corrp_tstat1 -thr 0.95 -mul maskneg_tstat2 inter2_grp1_neg

How  and where and when to use contrast [1 0] and contrast [0 1] to get inter1_grp2_pos ,inter1_grp2_neg, inter2_grp2_pos and inter2_grp2_neg files ??

Kindly rectify If something is wrong

Kindly advise the needful

Regards
Himanshu





































































































































































On Tue, Jun 10, 2014 at 2:47 PM, Anderson M. Winkler <[log in to unmask]> wrote:
Hi Himanshu,

There is information missing. You need to have 8 files overall:

For the interaction [1 -1]
1) Positive signs group 1 (using the contrast [1 0])
2) Negative signs group 1 (using the contrast [1 0])
3) Positive signs group 2 (using the contrast [0 1])
4) Negative signs group 2 (using the contrast [0 1])

For the interaction [-1 1]
5) Positive signs group 1 (using the contrast [1 0])
6) Negative signs group 1 (using the contrast [1 0])
7) Positive signs group 2 (using the contrast [0 1])
8) Negative signs group 2 (using the contrast [0 1])

Each of these is calculated using fslmaths as above. Make the necessary changes to pick the right tstats for thresholding at 0 and then masking the appropriate corrected p-value images. I'm sorry that my example suggested filenames that would not prevent name collisions (e.g. inter_grp1_pos). Please, use instead something as inter1_grp1_pos.

All the best,

Anderson



On 10 June 2014 10:55, Himanshu Joshi <[log in to unmask]> wrote:
Thanks Anderson,

Now I got the signs too..
This shows only inter_grp1_pos is significant and rest all inter_grp1_neg, inter_grp2_pos and inter_grp2_neg are not significant.

Thus it can be concluded that the group difference for mean FA varies with age in a positive direction for the voxel in which the interaction was significant.
Kindly rectify me if I am wrong, and help me in constructing the conclusion in a more representable form.

Regards
Himanshu




On Mon, Jun 9, 2014 at 8:08 PM, Anderson M. Winkler <[log in to unmask]> wrote:
Hi Himanshu,

You can ignore the p-values for these 2 additional contrasts. It doesn't really matter that neither are significant, because you already know that the difference between them, i.e., the interaction, is so. So don't even bother opening the images with p-values. Just go straight to the image with the t-statistic, and look at the signs (positive or negative) for these two contrasts for the voxels in which the interaction was significant.

To look at these voxels, use fslmaths to make a mask of the positive tstats, another with the negative t-stats. Then use these to further mask the image of the significant interaction thresholded at 0.95. Something like this (make the changes as needed):

fslmaths xx_tstat5 -thr 0 -bin maskpos_tstat5
fslmaths xx_tstat5 -mul -1 -thr 0 -bin maskneg_tstat5
(repeat for the xx_tstat6)

Then:
fslmaths xx_tfce_corrp_tstat3 -thr 0.95 -mul maskpos_tstat5 inter_grp1_pos
fslmaths xx_tfce_corrp_tstat3 -thr 0.95 -mul maskneg_tstat5 inter_grp1_neg
(repeat for the xx_tstat6, and also everything for the negative interaction, that I think is your _tstat4).

All the best,

Anderson





On 9 June 2014 13:53, Himanshu Joshi <[log in to unmask]> wrote:
Hi Anderson,
Thanks a lot, your reply was helpful.

Regarding checking of signs to see what's really going on -- You have advised [0 0 1 0] and [0 0 0 1] .For these recommended contrasts, I didn't find any significant  result in _tfce_corp images at p<0.05.
Can you please elaborate on
How to look at just the signs of the statistics for the same voxels where the interaction is significant.

Thanks in advance


On Mon, Jun 9, 2014 at 2:58 PM, Anderson M. Winkler <[log in to unmask]> wrote:
Hi Himanshu,

On 9 June 2014 09:01, Himanshu Joshi <[log in to unmask]> wrote:
Hello Quin and FSL experts,
I am also getting similar results  in certain brain regions for a contrast of 0 0 1 -1 ;the order for EV is Group1 Group2 Continious_variable_group1 Continious_variable_group2.
I agree with mark's comment that the difference of slope is significant. As you have cross checked it by stating your positive correlation results between group 1 mean FA and group1 Age, where as negative correlation between group2 mean FA and group2 Age.
If my group 1 is control group and group 2 is diseased group can I infer that as the age increases mean FA for controls also increases within particular region but mean FA for diseased decreases as a function of age.
And thus it can be concluded group difference for mean FA varies with age.

A significant interation means that the slopes are different. This can mean that one is positive and the other negative, or both positive (one being steeper than the other) or both negative (one plummeting faster than the other). You can check the directions by looking at the signs of the statistics of the contrasts [0 0 1 0]' and [0 0 0 1]', even if you don't use these contrasts for inference (i.e., the p-values don't need to be used, just the signs of the statistics for the same voxels where the interaction is significant).

How to report these findings in a more general way (say for instance to a layman) ?

I think the way you explained is quite clear, only need to check the signs to see what's really going on.

 All the best,

Anderson

Thanking you in anticipation

Regards
Himanshu


On Tue, Nov 26, 2013 at 3:22 PM, Qin Xu <[log in to unmask]> wrote:
Dear Mark,
Thank you very much for your answer!
At first, I thought my question must be too simple to be answered. Sorry, I need some time to improve myself. Several days ago, I created a mask image with the tstats which showed the significant interaction effect, and then used the fslmeants to calculate the subjects' mean FA values within the mask. There was a positive correlaton between the mean FA values and subjects' age in the group1, while negative correlation in the group2. This result was consistent with what you told me.
Thanks for your help.

Sincerely,
Qin



--
Himanshu Joshi
M.Tech. Cognitive & Neuroscience.
Senior Research Fellow
Ph.D Scholar,
Department of Psychiatry
NIMHANS, Bangalore

http://mbial.weebly.com/himanshu-joshi.html




--
Himanshu Joshi
M.Tech. Cognitive & Neuroscience.
Senior Research Fellow
Ph.D Scholar,
Department of Psychiatry
NIMHANS, Bangalore

http://mbial.weebly.com/himanshu-joshi.html














--
Himanshu Joshi





--
Himanshu Joshi
M.Tech. Cognitive & Neuroscience.
Senior Research Fellow
Ph.D Scholar,
Department of Psychiatry
NIMHANS, Bangalore

http://mbial.weebly.com/himanshu-joshi.html