Hi Martin and Beatriz,


I'm using SPHARM for interaction test in a group, but I'm a little confused with the test. First, please tell me if I'm doing it correctly. Here is one line of my glm matrix file :

1 1 /NAS/.../pat31_auto_lha_hippo_bin_close_to_templateSPHARM_ellalign.meta 66 1 11

I have 3 independent variables, age(66), gender(1) and my clinical score(11). I want to perform a correlation with the clinical score after a correction with age and gender, then I use : 

shapeAnalysisMANCOVA glm.txt --numIndependent 3 --interactionTest --testColumn 2

Is it correct ? I have no way to check if the correct variable is tested. 
Second, I'm confused with the results. I have correctly the Pearson and Spearman correlation statistics BUT, I obtain a map of FDR corrected p-value which correspond to the MANCOVA permutation. How is it possible given that I erform any group test ? What is the meaning of this p-value map of MANCOVA ?


Thank you for your help.

Clément B

Hi Clement,

I've been wondering about the same issues. I've come to some conclusions, but would certainly appreciate conformation from Martin or Beatriz.

Initially, I believe your command line is missing an argument specifying your input columns, something like: –columnIndependent 2,3,4 (with 2 being your clinical score and 3 and 4 being your covariates).

1. While it looks like this should work, the correlation test (Pearson & Spearman) does NOT consider the covariates, just runs against the --testColumn. It even says so much in the output, that the result is uncorrected, or something similar. You can confirm this by running the analysis with and without the covariates and the Pearson/Spearman output will be identical. Additionally, in the source code there is consideration for correlation with covariates but it not implemented.

2. From my understanding, the MANOVA test is valid and DOES consider covariates, and so can provide p-value/FDR p-value maps. However, the MANOVA test cannot provide a single coefficient to provide the direction of the change (like the Pearson R). Both these points can be demonstrated in SPSS by running a MANOVA with some covariates – a group test/factor is not necessary. Obtaining a direction is possible using correlation testing though, because only one dimension is used (inflation/deflation along the difference vector or the normal) and so can provide a single description of inflation/deflation. This also explains why the MANOVA significance maps and correlation significance maps are slightly different – the first is using a MANOVA on 3D coordinates, whilst the second is using vectors collapsed from the 3D coordinates.

In summary, MANOVA shows covariate-corrected significance tests, but doesn't tell you the direction, and Pearson/Spearman can tell you direction and significance, but not using covariates.

This is my working assumption – very happy to be corrected by the SPHARM team though!

Hope this helps,
Dave.

Hi D J, 

thank you for your answer.

1. Ok, I see. But the fact is that I used an example from Beatriz and Martin in their paper in Insight Journal about shape analysis MANCOVA where they showed a similar matrix with, age, IQ and clinical score at the end of the matrix. I quote : " Hypothesis : Is there a significant clinical score influence AFTER CORRECTION for group, gender, age and IQ ? Example command :
infile --numGroupTypes 2 --numIndependent 3 --interactionTest --testColumn 2"

I assume that is something similar to my arguments. But I will try with the arguments "columnIndependent". I have another question about that, how can you know the number of the column for the covariates. In my case, I thought that my independent variables were 0,1,2 columns. 

2. Oh, I thought MANOVA permutation test were designed for group comparison.. Then do you know how to combine correlation and covariates correction ? I a little bit lost. 



Thank you.

Clément B

Hi Clement,

I think the interaction test provides MANOVA output which does correct for covariates, but the Pearson and Spearman maps are still uncorrected.

The example command line is correct, but you still need to specify the columnsIndependent and columnInputfile. The defaults may work nevertheless.

The column numbers are given by the csv file you use. Be aware that the numbering starts from zero. In your example above, assuming the first 1 and 1 before the .meta file are separate columns, your test column should be 5 and columnIndependnt should be 3,4,5 (which means that 3 and 4 are the covariates).

Unfortunately I don't believe correlation with covariates (i.e. Partial correlation) is supported at this time. There is some mention of an update, so perhaps in the future.

Regards,
David.

Hi Marco and DJ,



I am catching up with your messages and in my opinion there is a problem of what is what here.

The correlation test can be performed in only one or both groups, and it can include covariates. It has an associated p-value to it. 
An interaction test without specifying with an argument that it is a correlation test, will perform morphological group testing checking for the influence of a certain variable (that you do not correct for, you have to indicate the variables that you correct for, I do believe that in that case scenario command line syntax the columnIndependent is missing). This sort of experiment uses both groups.



I hope that makes sense. I will give it a little bit of thought myself, but please feel free to ask any follow up questions if this is confusing.
Thank you,



Beatriz

Hi Beatriz, 


Thank you. Then, could you tell me if I'm right ? If I want to check the correlation with one variable, with this glm file : 

1 1 /NAS/.../pat31_auto_lha_hippo_bin_close_to_templateSPHARM_ellalign.meta 66 1 11

Are the arguments correct :

--numIndependent 3 --interactionTest --testColumn 5 --columnIndependent 3,4,5

where 66 and 1 are my covariates to correct with and 11 the variable I want to test for the correlation.

Thank you for your help, 

Clément