Hi,
I'm trying correlation analysis using NBS, and I'm not sure if i'm using correct design matrix and contrast.
first column consists of ones (for global mean), and second column age, third IQ, fourth age*IQ, and fifth is behavioral score.
1   8.9 113 1005.7  41
1   8.5 101 858.5 117
1   7.9 99 782.1 66
1   7.3 112 817.6 46
1   7.9 112 884.8 48
1   9.7 88 853.6 83
1   8.8 98 862.4 60
1   5.9 129 761.1 63
1   10.6 101 1070.6 108
1   9 120 1080 54
1   10 105 1050 44
1   9.2 97 892.4 60
1 8.7 105 913.5 48
1   5.5 72 396 59
1 7.6 122 927.2 39
1   5.4 88 475.2 59

And I applied [0,0,0,0,1] for contrast to investigate positive correlation. Is this correct?
Any comment would be appreciated.

Thanks,


YM Huh.

Hi Youngmin,

Yes - the contrast of [0 0 0 0 1] will test for a positive association with behavior if you select the t-test. It will test for any association (positive or negative) if you select F-test.

Please be careful with the age*IQ regressor, as this interaction will need to be interpreted carefully. You might want to demean the variables before computing the interaction, or it might be best not to include age*IQ. 

Andrew


Originally posted by Youngmin Huh:

Hi, Andrew
Thank you for your reply. It helped me a lot.
I should try NBS without interaction term (age*IQ).

Thanks again!


YM Huh

Dear Professor ;
Will appreciate your comments related to a problem with a design similar to this example, using 5 predictors , F test [0 0 0 0 1] were i get a lot of connections differences for every node (p 0.033) but when i try a T test to define if the correlation is positive [0 0 0 0 1] or negative [0 0 0 0 -1] NBS is unable to detect any difference at all (not significant results).

Using the F test, NBS also detects correlations for the rest of the predictors , for the contrasts [0 1 0 0 0], [0 0 1 0 0] and for the  [0 0 0 1 0], but ultimately a T test is only able to detect differences for the contrast [0 -1 0 0 0] and not for any of the others regressors.

Thanks for your time and precious help.
Juan P

Hi Juan,

If the F-test is replaced with a t-test, it might be useful to lower the statistic threshold. Remember the F=t^2. Therefore, if you use an F-stat threshold of 9, you should probably use a t-stat threshold of sqrt(9)=3 to make the two tests comparable. 

Note however that F is two-sided whereas t is one-sided. So even if you sent the statistic threshold to match, there still may be differences in the connections that are found significant.This is expected.

Overall, I think you have just forgot to lower the statistic threshold when using a t-test compared to a F-test.

Andrew


Originally posted by Juan Pablo Princich: