**correlation analysis with language scores**

Hello！

I hope this message finds you well.

I've been using the NBS software to explore the correlation between fluency scores from the Aphasia Battery of Chinese and the connectivity across brain regions (AAL atlas). The goal is to identify networks related to fluency function, while controlling for age, sex, and education level as covariates.

Despite reviewing the manual and forum discussions, I still have some uncertainties regarding the design matrix and the choice of statistical tests, and I would greatly appreciate your insights on the following queries:

Q1: Does the attached design matrix appropriately represent the variables of interest for my analysis?

my design matrix is:

1 2 31 12
9

1 1 34 16 9

1 1 20 14
10

1 2 38 14 8

1 2 36 14 9

1 2 34 16
10

1 2 41 12 9

1 2 39 12
10

1 2 33 9 9

1 2 58 9 8

1 2 24 9 10

1 2 37 14 6

1 2 43 8 8

1 1 34 21

1 2 43 20 8

1 1 59 16 9

1 1 66 0 8

1 2 30 16 9

1 1 58 9 9

1 1 50 9 9

1 1 25 9 9

1 1 44 16 6

1 2 36 12 9

1 1 43 16 8

1 1 59 12
10

1 1 40 14
10

1 1 53 12
10

1 1 17 10
10

1 1 36 12
10

1 2 44 9 10

1 2 52 14 9

1 1 38 14
10

1 1 30 12
10

1 2 67 3 9

1 1 42 12 9

1 2 37 14
10

1 1 50 9 9

1 1 65 12
10

1 2 36 16
10

1 2 33 14
10

1 1 43 14 9

1 2 19 14
10

1 1 40 3 10

1 1 36 9 10

1 1 59 12

1 2 34 19
10

1 1 54 12 6

1 1 57 9 8

1 2 37 16
10

1 1 34 16
10

1 1 39 19 9

1 1 46 6 8

1 1 55 6 4

1 1 34 16
10

1 1 62 9 8

with the col 1: global mean, cols 2-4(covariates): sex, age, education-level, col 5 fluency scores.

Q2: Given my interest in both positive and negative correlations, should I employ both [0 0 0 0 1] and [0 0 0 0 -1] in separate t-tests, or would a single F-test using [0 0 0 0 1] be sufficient? How do the interpretations of the results compare?

Q3: After I set the constrast as [0 0 0 0 1] with t test, I computed the Pearson correlation coefficients for each edge I got from the significant network, but I found some of them are negative, I am not sure where the issue occurred, And then I tried to convert all scores in the col 5 to negative values, expecting results would have shown an inverse correlation of positive and negative values, however, there was no significant result. Could you provide some insight into why negative correlations emerged initially, and why inverting the fluency scores didn't result in significant inverse correlations?

Q4: Is the 'threshold' value in NBS intended to be the t-statistic
for t-tests and the square of the t-statistic (t^2) for F-tests?
Does that mean when I use the F test, I should input the value of
t^2 as the threshold?

Q5: When I tried 12.25 as the threshold for the F test, NBS gave a significant network, but when I tried 6.25 and intended to obtain a denser network, there was no significant result. In my understanding, a lower threshold implies a more intricate network. Could there be an error in my approach?

I have attached the full connectivity matrices for your review. Your guidance on these matters would be invaluable to my research.

Thank you for your time and the invaluable resource you have provided to the community.

Best regards,

Hua Song

Hi, please see below for responses:

Originally *posted by Hua Song:*

Hello！

I hope this message finds you well.

I've been using the NBS software to explore the correlation between fluency scores from the Aphasia Battery of Chinese and the connectivity across brain regions (AAL atlas). The goal is to identify networks related to fluency function, while controlling for age, sex, and education level as covariates.

Despite reviewing the manual and forum discussions, I still have some uncertainties regarding the design matrix and the choice of statistical tests, and I would greatly appreciate your insights on the following queries:

Q1: Does the attached design matrix appropriately represent the variables of interest for my analysis?

my design matrix is:

1 2 31 12 9

1 1 34 16 9

1 1 20 14 10

1 2 38 14 8

1 2 36 14 9

1 2 34 16 10

1 2 41 12 9

1 2 39 12 10

1 2 33 9 9

1 2 58 9 8

1 2 24 9 10

1 2 37 14 6

1 2 43 8 8

1 1 34 21

1 2 43 20 8

1 1 59 16 9

1 1 66 0 8

1 2 30 16 9

1 1 58 9 9

1 1 50 9 9

1 1 25 9 9

1 1 44 16 6

1 2 36 12 9

1 1 43 16 8

1 1 59 12 10

1 1 40 14 10

1 1 53 12 10

1 1 17 10 10

1 1 36 12 10

1 2 44 9 10

1 2 52 14 9

1 1 38 14 10

1 1 30 12 10

1 2 67 3 9

1 1 42 12 9

1 2 37 14 10

1 1 50 9 9

1 1 65 12 10

1 2 36 16 10

1 2 33 14 10

1 1 43 14 9

1 2 19 14 10

1 1 40 3 10

1 1 36 9 10

1 1 59 12

1 2 34 19 10

1 1 54 12 6

1 1 57 9 8

1 2 37 16 10

1 1 34 16 10

1 1 39 19 9

1 1 46 6 8

1 1 55 6 4

1 1 34 16 10

1 1 62 9 8

with the col 1: global mean, cols 2-4(covariates): sex, age, education-level, col 5 fluency scores.

The design matrix looks fine.

Q2: Given my interest in both positive and negative correlations, should I employ both [0 0 0 0 1] and [0 0 0 0 -1] in separate t-tests, or would a single F-test using [0 0 0 0 1] be sufficient? How do the interpretations of the results compare?

**
Both contrasts are acceptable. The F-test is a two-sided test and
will be sensitive to positive and negative correlations. The t-test
is one-sided and is either sensitive to positive or negative
correlations, depending on whether the polarity of the 1 is
negative or positive. **

Q3: After I set the constrast as [0 0 0 0 1] with t test, I computed the Pearson correlation coefficients for each edge I got from the significant network, but I found some of them are negative, I am not sure where the issue occurred, And then I tried to convert all scores in the col 5 to negative values, expecting results would have shown an inverse correlation of positive and negative values, however, there was no significant result. Could you provide some insight into why negative correlations emerged initially, and why inverting the fluency scores didn't result in significant inverse correlations?

If the fluency scores are multiplied by -1, the sign of the correlation will flip and thus a signifciant positive correlation will no longer be signifciant (because it will become negative). For the [0 0 0 0 1] contrast, it may be that the other variables are having a signifciant confounding effects and that is why you find some of the edges associated with negative correlations.

Q4: Is the 'threshold' value in NBS intended to be the t-statistic for t-tests and the square of the t-statistic (t^2) for F-tests? Does that mean when I use the F test, I should input the value of t^2 as the threshold?

Yes.

Q5: When I tried 12.25 as the threshold for the F test, NBS gave a significant network, but when I tried 6.25 and intended to obtain a denser network, there was no significant result. In my understanding, a lower threshold implies a more intricate network. Could there be an error in my approach?

A lower threshold does not necessarily a imply a denser network. Lowering the threshold could yield no signifciant results.

I have attached the full connectivity matrices for your review. Your guidance on these matters would be invaluable to my research.

Thank you for your time and the invaluable resource you have provided to the community.

Best regards,

Hua Song

I sincerely appreciate your invaluable assistance, Andrew, your guidance has been instrumental in helping me make significant progress. Thanks to your insights, now I can keep forward.

However, I still have two lingering questions:

Q1:I conducted both t-tests with a threshold of 2.5 and F-tests with a threshold of 6.25. Surprisingly, the F-test, which I anticipated would reveal more links, did not yield any significant results. On the other hand, the t-test produced a network. Could you shed some light on why this might be the case?

Q2:I conducted the fluency scores that multiplied by -1, with the t test and contrast [0 0 0 0 -1], the result shown no significance while the original version of design matrix with the contrast [0 0 0 0 1] provided a useful network. What caused such results?

Thank you once again for your unwavering support.

Best regards,

Hua Song

Hi Hua,

Q1. It is certainly possible that the t-test is more sensitive than the F-test, given that the t-test is one-sided and F-test is two-sided. A one-sided test will be be more sensitive to effects in a common direction.

Q2. Using -1 instead of 1 in the contrast will test for a negative correlation with the fluency score. It is not unusual to find effects with 1 but not with -1.

Andrew

Originally *posted by Hua Song:*

I sincerely appreciate your invaluable assistance, Andrew, your guidance has been instrumental in helping me make significant progress. Thanks to your insights, now I can keep forward.

However, I still have two lingering questions:

Q1:I conducted both t-tests with a threshold of 2.5 and F-tests with a threshold of 6.25. Surprisingly, the F-test, which I anticipated would reveal more links, did not yield any significant results. On the other hand, the t-test produced a network. Could you shed some light on why this might be the case?

Q2:I conducted the fluency scores that multiplied by -1, with the t test and contrast [0 0 0 0 -1], the result shown no significance while the original version of design matrix with the contrast [0 0 0 0 1] provided a useful network. What caused such results?

Thank you once again for your unwavering support.

Best regards,

Hua Song

Thank you very much for your guidance. I can now proceed to work effectively using the NBS！