Re: [SPM] SnPM group comparison

[Thomas Nichols]

Dear Mandy,

Sorry for the delay; please be sure to cc snpm-s...@googlegroups.com so I see your message faster.

i have a question regarding an analysis between 3 groups. My groups have different sample sizes (group1:29 subjects, group2: 14 subjects, group3: 14 subjects), because of this we decided to do a non-parametric approach.

Just to clarify, there is no assumption about standard methods that the group sizes must be equal.  But it *is* true that standard t-test/ANOVA methods are *less* robust to violation of Gaussian assumptions when the group sizes are unequal.

 

I was able to do and interprete a two-sample t-test in SnPM with one scan per subject for the comparison of 2 groups (like post-hoc-tests).

But we are also interested in compare the 3 groups directly with a non-parametric test under one special condition. So i did a between-group ANOVA and selected the 3 groups with one scan (contrast) each. I got the results of the F-Test but its not clear for me what has been calculated or how to interpret the results.

The F-test is for the null hypothesis that all groups have equal means, and is sensitive to any departures of that null hypothesis (i.e. of any mean difference between the groups).   It is *not* an test of the null hypothesis "All groups have zero mean" (which would probably only be of interest in fMRI).

 

Do you know how to interpret the F-test results of the 3 groups under a special condition and do you know if there is another non-parametric possibility to test for comparisons between 3 groups?

What precisely are you looking for?  A correction over all the possible contrasts you might want to examine among the 3 groups?

-Tom

 

Best regards,
Mandy

--

__________________________________________________________
Thomas Nichols, PhD
Professor, Head of Neuroimaging Statistics
Department of Statistics & Warwick Manufacturing Group
University of Warwick, Coventry  CV4 7AL, United Kingdom

Web: http://warwick.ac.uk/tenichols

Email: t.e.n...@warwick.ac.uk
Phone, Stats: +44 24761 51086, WMG: +44 24761 50752
Fax:  +44 24 7652 4532

[Skunde, Mandy]

Dear Tom,

 

thanks for your answer.  So do u think we could also report the results of the parametric test? This would be much more easier for us to understand and interpret the results.

 

In case we follow the non-parametric approach: We want to compare the three groups (one control group, 2 patient groups) regarding their brain activation during a special contrast (activation when seeing a non-target on which they have to inhibit their reaction vs activation when seeing a target on which they have to press a button) it means we want to see how the activation is in this groups under this contrast. I choosed the between-group ANOVA in SnPM13.

 

I recognized that I made the mistake to make the assumption that “groups are all zero” for the null hypothesis. I changed it to “all groups have equal means“. Now the resulting contrast looks clearer for me because it includes all three groups. But the problem is now that the whole brain is activated. I get a huge cluster with 16.000 voxel. I choosed “cluster inference: Yes, set cluster forming threshold now (fast): 0.001” and in the inference settings “Cluster-Forming Threshold:NaN” as well as “Significance Level: Uncorrected k : 10 voxel”. When I change it to FWE correction, I don’t get anything. Now I don’t know how to deal with this result.

 

When I do the same group comparison in SPM (One-way ANOVA) I get clear results by the FWE-correction with the largest cluster = 1946 voxel. (F-contrast 1 0 0, 0 1 0, 0 0 1).

 

Greetings

Mandy

[Thomas Nichols]

Dear Mandy,

thanks for your answer.  So do u think we could also report the results of the parametric test? This would be much more easier for us to understand and interpret the results.

If the parametric test is giving reasonable answers, then that's fine.  I'm only adamant about using nonparametric when (a) parametric and nonparametric give drastically different answers and/or (b) the parametric assumptions are clearly invalid.  I don't think this is the case here, so you should be fine.

 

In case we follow the non-parametric approach: We want to compare the three groups (one control group, 2 patient groups) regarding their brain activation during a special contrast (activation when seeing a non-target on which they have to inhibit their reaction vs activation when seeing a target on which they have to press a button) it means we want to see how the activation is in this groups under this contrast. I choosed the between-group ANOVA in SnPM13.

 

I recognized that I made the mistake to make the assumption that “groups are all zero” for the null hypothesis. I changed it to “all groups have equal means“. Now the resulting contrast looks clearer for me because it includes all three groups. But the problem is now that the whole brain is activated. I get a huge cluster with 16.000 voxel. I choosed “cluster inference: Yes, set cluster forming threshold now (fast): 0.001” and in the inference settings “Cluster-Forming Threshold:NaN” as well as “Significance Level: Uncorrected k : 10 voxel”. When I change it to FWE correction, I don’t get anything. Now I don’t know how to deal with this result.

I guess I would disagree with "whole brain is activated", since with FWE correction you get nothing.

But I'm also surprised... as I would have thought that "Groups are all zero" would (mistakenly) give a result of a whole active brain, while "Groups are all equal" would give a more sensible result.

How does a voxel-wise (FWE or FDR) result look?

When I do the same group comparison in SPM (One-way ANOVA) I get clear results by the FWE-correction with the largest cluster = 1946 voxel. (F-contrast 1 0 0, 0 1 0, 0 0 1).

Well, *that* F-contrast is "groups are all zero".  If you want "all groups have equal means, you need something like

1 -1 0

1 0 -1

Are you sure you're comparing like with like?

-Tom

 

--
You received this message because you are subscribed to the Google Groups "Statistical Nonparametric Mapping" group.
To unsubscribe from this group and stop receiving emails from it, send an email to snpm-support...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.