help
help > RE: Design matrix and contrast for 2x2 design
Aug 12, 2017 01:08 AM | Andrew Zalesky
RE: Design matrix and contrast for 2x2 design
Hi David,
1. The choice between full, partial and sparse correlation is difficult. Pros and cons associated with each method. I recommend full correlation in the first instance, but some people in the field would disagree.
2. If you are simply interested in the effect of drug versus no drug, you should delete the music column and the interaction column and rerun the model. You will still need a column for each subject to model each subject's mean. I don't think you should throw away the music scans for this analysis.
3. No. You MUST run at least 50000 - 100000 permutations with FDR. The results will be highly unreliable otherwise. Refer to this as FDR with p-values computed using permutation testing.
4. Generally, the name of the function provides clues about whether the measure is suitable for signed networks. For example, "bu" at the end of the function name means "binary undirected". See BCT webpage for more details.
5. Your design is repeated in both factors. Therefore a 2 x 2 with only one repeated measure is not appropriate. You could import your data into SPSS and perform testing there.
Andrew
Originally posted by David de Wide:
1. The choice between full, partial and sparse correlation is difficult. Pros and cons associated with each method. I recommend full correlation in the first instance, but some people in the field would disagree.
2. If you are simply interested in the effect of drug versus no drug, you should delete the music column and the interaction column and rerun the model. You will still need a column for each subject to model each subject's mean. I don't think you should throw away the music scans for this analysis.
3. No. You MUST run at least 50000 - 100000 permutations with FDR. The results will be highly unreliable otherwise. Refer to this as FDR with p-values computed using permutation testing.
4. Generally, the name of the function provides clues about whether the measure is suitable for signed networks. For example, "bu" at the end of the function name means "binary undirected". See BCT webpage for more details.
5. Your design is repeated in both factors. Therefore a 2 x 2 with only one repeated measure is not appropriate. You could import your data into SPSS and perform testing there.
Andrew
Originally posted by David de Wide:
Hey Andrew,
Thank you again for all your help. The NBS is a great tool and has been very useful so far. I've been working hard at applying the NBS and the BCT over the past week and have made quite a bit of progress. However, I do have some remaining quesstion that I haven't been able to find an answer to.
1) After reading several of your articles using the NBS and BCT, I decided to apply partial correlation to my data (similar to the ADHD study) instead of full correlation (similar to the schizoprenia study). However, this changed many of the implicated ROIs dramatically. I am considering presenting both, but am unsure of the merit of partial correlation in this case, as the results are now less intuitive. Which form of correlation do you think is most appropriate? (I also looked into Sparse Partial Correlation, but could not find a matlab plugin).
2) Using the NBS with all 48 scans, and looking at the contrast of the drug effect, I am effectively using both the music and non-music scan together (which themselves are different). Would it be "cleaner" to load a seperate design and dataset of 24 scans and look at the effect of drug seperately for both music and no-music? Or is there an additional contrast I can put in to control for the effect of music on drug (and vice versa)?
3) Before the NBS, I was only aware of regular methods of FDR correction. Using the NBS I initially found significant edges using FDR, despite not finding any when using traditional FDR methods on my 90x90 ANOVA table. I later read about the high number of permutations required, and although this removed all FDR finding, 50.000 is taking a very long time. Is there a lower number that would still be considered "safe"? How do I refer to this method of FDR in my article (i.e. permutation based FDR)?
4) Is there an overview somewhere of which BCT metrics require values to be postive/unsigned?
5) Using Graphvar/GTG, I've been able to do some statistical testing on the network metrics. However, these only allow for a single factor repeated measures, so I still can't look at the contrast of a 2x2 design. Would it be a good to compare each pair seperately and plot all 4 in the same graph?
Thank you for your willingness to answer questions. It's been a massive help, as neither of my supervisors (nor anyone on my floor) has any experience using graph theory based metrics.
Sincerely,
David
Thank you again for all your help. The NBS is a great tool and has been very useful so far. I've been working hard at applying the NBS and the BCT over the past week and have made quite a bit of progress. However, I do have some remaining quesstion that I haven't been able to find an answer to.
1) After reading several of your articles using the NBS and BCT, I decided to apply partial correlation to my data (similar to the ADHD study) instead of full correlation (similar to the schizoprenia study). However, this changed many of the implicated ROIs dramatically. I am considering presenting both, but am unsure of the merit of partial correlation in this case, as the results are now less intuitive. Which form of correlation do you think is most appropriate? (I also looked into Sparse Partial Correlation, but could not find a matlab plugin).
2) Using the NBS with all 48 scans, and looking at the contrast of the drug effect, I am effectively using both the music and non-music scan together (which themselves are different). Would it be "cleaner" to load a seperate design and dataset of 24 scans and look at the effect of drug seperately for both music and no-music? Or is there an additional contrast I can put in to control for the effect of music on drug (and vice versa)?
3) Before the NBS, I was only aware of regular methods of FDR correction. Using the NBS I initially found significant edges using FDR, despite not finding any when using traditional FDR methods on my 90x90 ANOVA table. I later read about the high number of permutations required, and although this removed all FDR finding, 50.000 is taking a very long time. Is there a lower number that would still be considered "safe"? How do I refer to this method of FDR in my article (i.e. permutation based FDR)?
4) Is there an overview somewhere of which BCT metrics require values to be postive/unsigned?
5) Using Graphvar/GTG, I've been able to do some statistical testing on the network metrics. However, these only allow for a single factor repeated measures, so I still can't look at the contrast of a 2x2 design. Would it be a good to compare each pair seperately and plot all 4 in the same graph?
Thank you for your willingness to answer questions. It's been a massive help, as neither of my supervisors (nor anyone on my floor) has any experience using graph theory based metrics.
Sincerely,
David
Threaded View
| Title | Author | Date |
|---|---|---|
| David de Wide | Aug 3, 2017 | |
| yue zhang | Jan 11, 2020 | |
| Andrew Zalesky | Aug 5, 2017 | |
| Liam Nestor | Sep 27, 2019 | |
| Andrew Zalesky | Sep 28, 2019 | |
| Liam Nestor | Sep 28, 2019 | |
| Andrew Zalesky | Sep 28, 2019 | |
| Liam Nestor | Sep 30, 2019 | |
| Andrew Zalesky | Oct 1, 2019 | |
| Liam Nestor | Oct 1, 2019 | |
| David de Wide | Aug 6, 2017 | |
| Andrew Zalesky | Aug 7, 2017 | |
| David de Wide | Aug 7, 2017 | |
| Andrew Zalesky | Aug 7, 2017 | |
| David de Wide | Aug 11, 2017 | |
| Andrew Zalesky | Aug 12, 2017 | |
| David de Wide | Aug 18, 2017 | |
| Andrew Zalesky | Aug 19, 2017 | |