I am also finding that some off-diagonal elements in this matrix are also
greater than 1 indicating this matrix is not a correlation matrix.
In [5]: img
Out[5]:
memmap([[ 8.66434002e+00, 1.96847185e-01, 1.66294336e-01, ...,
1.01449557e-01, 7.45474100e-02, 1.15624115e-01],
[ 1.96847185e-01, inf, 3.36383432e-01, ...,
-5.70017472e-03, -5.49946353e-02, 3.72834280e-02],
[ 1.66294336e-01, 3.36383432e-01, inf, ...,
-4.45242636e-02, -6.07097335e-02, -1.51601573e-02],
...,
[ 1.01449557e-01, -5.70017472e-03, -4.45242636e-02, ...,
inf, 1.91883039e+00, 9.20160294e-01],
[ 7.45474100e-02, -5.49946353e-02, -6.07097335e-02, ...,
* 1.91883111e+00*, 8.31776619e+00, 8.82132888e-01],
[ 1.15624115e-01, 3.72833721e-02, -1.51601573e-02, ...,
9.20160294e-01, 8.82132888e-01, 8.66434002e+00]],
dtype=float32)
Any insight would be appreciated
Thanks,
Sang
On Mon, May 15, 2017 at 1:13 PM Sang-Yun Oh <[email protected]> wrote:
> Thank you for the response.
>
> I am, too, confused by some being non-zero finite values, and others being
> infinities.
>
> Before computing a correlation matrix, if standardized by subtracting the
> mean and scaling by variance, all diagonal elements should be exactly 1.
>
> What I am concerned about is how the whole matrix was computed, since a
> fundamental characteristic of correlation matrix is not satisfied
>
> Best,
> Sang
>
> On Mon, May 15, 2017 at 11:33 AM Timothy Coalson <[email protected]> wrote:
>
>> Per the name "zcorr", the correlation values have been z-transformed
>> (fisher's small z transform). I am somewhat confused as to why some
>> elements in the diagonal are not infinite. The "true" value of applying
>> this transform would be infinite on the entire diagonal, as arctanh(1) is
>> infinite. I am guessing this result was generated in matlab, as wb_command
>> actually prevents infinities when using the z transform, putting a cap on
>> the correlation (when not using z-transform, it shows correlations of 1 as
>> expected).
>>
>> Whatever analysis you do with correlation matrices like this should
>> ignore the diagonal anyway, since it is correlation to itself.
>>
>> Tim
>>
>>
>> On Mon, May 15, 2017 at 3:57 AM, Sang-Yun Oh <[email protected]> wrote:
>>
>>> I downloaded group average functional correlation
>>> file: HCP_S900_820_rfMRI_MSMAll_groupPCA_d4500ROW_zcorr.dconn.nii
>>>
>>> Some diagonal elements of the square matrix (91282x91282) are infinite
>>> (Please see below).
>>>
>>> I want to use this matrix in ananalysis; however, I am not sure how to
>>> understand or deal with infinite diagonal values.
>>>
>>> I appreciate any insight
>>>
>>> Thanks,
>>> Sang
>>>
>>> ======================
>>>
>>> In [1]: import nibabel
>>>
>>> In [2]: asdf =
>>> nibabel.load('HCP_S900_820_rfMRI_MSMAll_groupPCA_d4500ROW_zcorr.dconn.nii')
>>>
>>> In [3]: img = asdf.get_data()
>>>
>>> In [4]: img.shape
>>> Out[4]: (1, 1, 1, 1, 91282, 91282)
>>>
>>> In [5]: S = img[0,0,0,0,:,:]
>>>
>>> In [6]: S
>>> Out[6]:
>>> memmap([[ 8.66434002e+00, 1.96847185e-01, 1.66294336e-01, ...,
>>> 1.01449557e-01, 7.45474100e-02, 1.15624115e-01],
>>> [ 1.96847185e-01, inf, 3.36383432e-01, ...,
>>> -5.70017472e-03, -5.49946353e-02, 3.72834280e-02],
>>> [ 1.66294336e-01, 3.36383432e-01, inf, ...,
>>> -4.45242636e-02, -6.07097335e-02, -1.51601573e-02],
>>> ...,
>>> [ 1.01449557e-01, -5.70017472e-03, -4.45242636e-02, ...,
>>> inf, 1.91883039e+00, 9.20160294e-01],
>>> [ 7.45474100e-02, -5.49946353e-02, -6.07097335e-02, ...,
>>> 1.91883111e+00, 8.31776619e+00, 8.82132888e-01],
>>> [ 1.15624115e-01, 3.72833721e-02, -1.51601573e-02, ...,
>>> 9.20160294e-01, 8.82132888e-01, 8.66434002e+00]],
>>> dtype=float32)
>>>
>>> In [7]: S.diagonal()
>>> Out[7]:
>>> memmap([ 8.66434002, inf, inf, ..., inf,
>>> 8.31776619, 8.66434002], dtype=float32)
>>>
>>>
>>> _______________________________________________
>>> HCP-Users mailing list
>>> [email protected]
>>> http://lists.humanconnectome.org/mailman/listinfo/hcp-users
>>>
>>
>>
_______________________________________________
HCP-Users mailing list
[email protected]
http://lists.humanconnectome.org/mailman/listinfo/hcp-users
