Print

Print
Hi Hanneke,

I finally looked into the images. For some reason none of the contrasts files that came included the interaction, and I wonder if this was on purpose or by accident (?). In any case, using the modified (but equivalent) model proposed in the earlier email, with and without the interaction term, the results are essentially the same. I named them models A and B:

Model A:
EV1: intercept
EV2: group (+/-1), mean-centered
EV3: risk (+/-1), mean centered
EV4: interaction: product of EV2 and EV3.
EV5-7: age, sex, scanner.

In this model, the interaction isn't significant, and among the main effects, only the contrast C4: [0 -1 0 0 0 0 0] shows significant results. 

Model B:
Same as Model A, except that EV4 was removed. In this, the only contrast with significant results is C3: [0 -1 0 0 0 0].

Both contrasts, i.e., C4 in Model A and C3 in Model B, correspond to a negative main effect of group (pat < con). Both gave very similar maps.

I wonder if the issue wasn't that the contrast files weren't entered correctly in the initial model, i.e. the interaction wasn't actually done, but instead just the 4 positive and negative comparisons of effects of group and genotype. Could this have been the case?

All the best,

Anderson

PS: I'm sending the relevant maps in a separate email; they are small, but not enough for the list server.
 


On 26 August 2015 at 07:49, Anderson M. Winkler <[log in to unmask]> wrote:
Hi Hanneke,

You should have just received an email with instructions to upload. Please, be patient as it will take several days until I can look into it (at least two weeks I think). If other users in the list could offer help, maybe you could get feedback faster.

All the best,

Anderson



--
Anderson M. Winkler
FMRIB / Analysis Group
[ Blog | Twitter ]


On 25 August 2015 at 14:25, Ewijk, H. van <[log in to unmask]> wrote:
Hi,

Regarding the correlations:
Scansite is correlated with group (i.e. group is not equally distributed over both scansites). Same for gender. Age isn't correlated with either main EV, but is included anyway because we have a very broad age range. But none of those nuisanvce EV's are correlated with genotype, or with the group*genotype interaction term.

Good to know about randomise; I'm currently using FSL 5.0.7.

For clarity, I've named the model you proposed Model 3.
When I look at the uncorrected tstat images (no tfce) for the main group effect at p>.95 I get many voxels, very widespread, for all models, but Model 1 shows less voxels (~26000) than Models 2/3 (~38000 each). This difference becomes much larger when looking at the tfce_p images (~9000 for Model 1 vs. ~17000 voxels for Model 2/3) and tfce_corrp (0 voxels for Model 1 and ~17000 for Model 2/3).

I'm still confused why Model 1 seems to give a different output than Model 2/3. I must have made a mistake somewhere but can't seem to figure out where. It would be very helpful if you could have a look at the images/files. Where should I upload the files?

Kind regards,
Hanneke


On 8/20/15, 10:05 AM, "FSL - FMRIB's Software Library on behalf of Anderson M. Winkler" <[log in to unmask] on behalf of> wrote:[log in to unmask]

Hi,

Please see below,

On 19 August 2015 at 13:54, Ewijk, H. van <[log in to unmask]> wrote:
Hi Anderson,

Thanks for your suggestions! Unfortunately, I'm still confused.. 

I checked the correlations using your approach, and none of the nuisance EVs are significantly correlated to the interaction term. All r<.10.

I assume this means not correlated either with the main effects' EVs (?).
 

I also ran TBSS using the model you proposed. Results were similar to the Model 2 I'd previously run. So main effect of group on FA, but no main effect for genotype (corrected p=.060). Somehow both your model and my Model 2 give similar output, but different to Model 1, in which I used the Cell Means approach. 

Something doesn't sound right. The model I suggested should give the same as the Model 1. Did you change the contrasts accordingly?
 

When I extract the mean FA value of the voxels with unexpected behaviour and run a GLM in SPSS, it shows a significant effect for group, but not for genotype or group*genotype. So similar to Model 2 and your model. The same (but the other way around) goes for mean MD and genotype; the GLM in SPSS shows the same main effect for genotype as did Model 1 in TBSS.

I haven't found anything strange when looking at the plotted data either (but will keep looking)..

When looking at the uncorrected tstat images, the pattern of results for the group contrast is similar for both models, except that Model 1 is less significant. I need to lower the threshold to p<.10 for Model 1 to get similar results as Model 2 at p<.05. Again, this is the same (but the other way around) for genotype and MD: when I set Model 2 at p<.10 and Model 1 at p<.05 I get identical results. 
(For this, I used the _tfce_p_tstatX.nii.gz images (instead of the tfce_corrp images), because I think these are the ones you're talking about? Or should I really re-run randomise with the -uncorrp option?)

Sounds you're using an earlier version of randomise. Currently, to have uncorrected it's necessary to add the option -uncorrp. Since you already have these images, no need to run again. But better for diagnostic purposes is to look at the raw t-statistic (vox_p_tstat*), instead of TFCE or cluster, because the last take contiguity of signals into account, and may make it more complicated to see what might be going on.

Try looking into these. If nothing comes out, I can offer to look into the images but you'd have to wait for a few days until I could do it. Perhaps others in the list could bring other suggestions.

All the best,

Anderson



 

So it apears that Model 1 has more power to detect a group effect, and Model 2 has more power to detect a genotype effect..? How come? Which one would most correct to report?

Thanks again for your help!


 






On 8/19/15, 9:08 AM, "FSL - FMRIB's Software Library on behalf of Anderson M. Winkler" <[log in to unmask] on behalf of [log in to unmask]> wrote:

Hi,

Is there any chance that one or more of the nuisance EVs is correlated with the interaction term? To check for this, consider an equivalent model that has:

EV1: intercept.
EV2: -1/+1 for group, mean centered
EV3: -1/+1 for genotype, mean centered.
EV4: product of EV2 and EV3 after these two have been mean centered.
EV5 and others: nuisance EVs, mean centered.

Then compute the correlations between these EVs (in Matlab, corr(X), where X is the design, will return all pairwise correlations at once).

Also, consider making scatter plots of the voxels with unexpected behaviours to see if any elucidative pattern arises.

Another thing is, when investigating, use uncorrected p-values (option --uncorrp in randomise). Although these aren't appropriate in general, the relationship between significance and effect size is better preserved and is unrelated to what goes on in the other voxels. It's probably clearer (to investigate) than the corrected.

I'm not sure this helps, but curious to see the result of your investigation.

All the best,

Anderson



On 18 August 2015 at 11:49, Ewijk, H. van <[log in to unmask]> wrote:

Dear FSL experts,


I have a question about my GLM design. 

I'm running TBSS on 2 groups (patient/control) with 2 different genotypes, while controlling for scansite, age and gender.

 

Because I was interested in a possible group*genotype interaction, I first set up the following matrix and contrasts (Model 1):

patient_genotype1   patient_genotype2   control_genotype1   control_genotype2   scansite   age   gender

1   0   0   0   x   x   x 

0   1   0   0   x   x   x 

0   0   1   0   x   x   x 

0   0   0   1   x   x   x

. . .

Where the last 3 columns are all demeaned over all subjects. 

With (some of the) contrasts:

1   1   -1   -1   0   0   0   Main effect group: patients > controls

1   -1   1   -1   0   0   0   Main effect genotype: 1 > 2

1   -1   -1   1   0   0   0   Interaction group*genotype

. . .

 

This design did not reveal any interactions or main effects on FA (all p>.20).

 

Hence, I set up a new model to simply test the main effect of group on FA without the (nonsignificant) group*genotype interaction, with a different design (Model 2):

patients   controls   scansite   age   gender   genotype

1   0   x   x   x   x

0   1   x   x   x   x

. . .

Where genotype was demeaned over all subjects, using contrasts for the main group effect:

1   -1   0   0   0   0

-1   1   0   0   0   0


This design DID reveal a (very) significant main effect for group (p=.008).

 

I'm a little puzzled by this; I thought that both designs would test the main group effect while controlling for genotype, so why did Model 1 not reveal the main effect for group while Model 2 did? In theory, both designs account for genotype-related variance, though coded in different ways (in Model 1 I used the Cell Means approach for a 2x2 ANOVA according to the FSL GLM wiki. In Model 2 I added genotype as a covariate similarly to gender and scansite). Which one is correct? I'm assuming Model 2, since it probably doesn't make much sense to use the design used in Model 1 when there's no significant interaction.

 

To make it more confusing though, a similar thing happened for the main effect for genotype on MD, but the other way around: Model 1 DID reveal a main effect for genotype (p=.024), while Model 2 did NOT (p=.066) (I made a Model 2 for genotype in a similar way as the Model 2 for group that I described above -- so genotype coded in 2 columns with 1s and 0s, and patient-group in 1 column, demeaned).

 

Could you please explain why this happens, and which model would be best to run for the main effects of group and genotype in the absence of significant interactions?

Many thanks!