Thanks for your valuable time in advance. I would like to ask you guys some questions about how to use connectomestats.
Suppose that I have a group of healthy controls (10 subjects), this group is scanned by two different scanners to investigate if the scanner difference would have some adverse impact on connectome reconstruction accuracy. My aim is to use connectomestat to show that there is no difference between two scanners and I designed the following steps (For simplicity sake, I only use 5 subjects as examples). Could you please help me to check?
Create design matrix
1 0
1 0
1 0
1 0
1 0
0 1
0 1
0 1
0 1
0 1
Create contrast matrix
1 -1
run âconnectomestats connectome.csv tfnbs designmatrix.txt contrast.txt tfnbs_outputâ and got several output files: abs_effect.csv, beta_0.csv, beta_1 csv, enhanced.csv, fwe_pvalue.csv, null_dist.txt, std_dev.csv, std_effect.csv, tvalue.csv, uncorrected_pvalue.csv.
Analysed the results.
From my understanding, fwe_pvalue.csv, enhanced.csv, and tvalue.csv are three main metrics to measure the differences between two scanners. Regarding enhanced.csv and tvalue.csv, the lower value, the better agreement between two scanners. The difference between enhanced.csv and tvalue.csv is that enhanced.csv used enhanced statistic to generate enhanced t value. I am curious about which of them is more accurate to describe the difference. With respect to few_pvalue.csv, I think I should only check pixels of which values are less than 0.95 (because in connectomestats, fwe_pvalue = 1 - p). The pixels possessing values higher than 0.95 mean that two scanners are not consistent with each other regarding these edges. Could you please help me check my understanding. Please correct me if I am wrong.
Given the nature of the experiment that you are performing, I am not entirely confident that utilising the TFNBS method is appropriate. The premise of that method is that if there is a genuine biological effect in some pathway, then there is an increased likelihood that other pathways involving the same grey matter regions are more likely to also be affected than are completely unrelated pathways. However you very explicitly have a cohort where you expect there to be no genuine biological effect. So the nature of that âstatistical enhancementâ changes quite drastically, becoming something more like âif the change in scanner induces a difference in one pathway, then it is more probable that another pathway involving one of those nodes will also have an erroneously induced change than it is for some unrelated pathway to have suchâ. Which is probably true mind you; but itâs unclear whether interpreting any statistically significant result as âthe network influenced by scanner changeâ makes any sense.
Beyond that, youâre right about the t-value and enhanced statistic outputs, as well as the p-value. But I would also question whether or not rigorous familywise error correction is appropriate in your case. You are probably more interested in discovering any potential differences if they are there, than determining with robust correction for multiple comparisons whether or not it is possible to measure such a difference in the presence of all sources of data variance. Perhaps an FDR correction across edges might be more appropriate in your case?
Thank you so much for your help. I checked the FDR and FWE and find that they are very easily to get them mixed up despite they differ from each other fundamentally. I also find that connectomestats seems do not provide a way to compare two groups using FDR. Could you please tell me what the most convenient way is to compare two groups using the FDR mentioned by you. Thanks
You are correct in that the statistical inference commands in MRtrix3 do not currently support FDR correction. You might find that some other software that is more specific to connectome analysis may provide such a functionality, though Iâm not myself familiar with such.
Alternatively, if you are not requiring advanced GLM functionalities and will not be using statistical enhancement, then you may well find that implementing something yourself may not be too difficult. Such an analysis would essentially ignore the network structure of the data: the e.g. upper triangular portion of the matrix just becomes an unstructured set of numbers, for which you perform an independent t-test in each and apply FDR correction. Permutation would not be too difficult to implement. In e.g. MatLab / Octave, itâs probably not a very large number of lines of code.