[color=#000000]Hi Jeff,[/color]
[color=#000000]The p-FDR analysis-level correction is typically used when you want to make inferences about individual connections. It corrects the individual connection-level statistics for the total number of individual connections in your entire analysis (e.g. the size of the ROI-to-ROI matrix for the selected ROIs). In this case (for connection-level inferences) you typically just want to use this connection-level threshold (and disregard/uncheck any seed-level or network-level additional thresholding options). A typical example would be if you have say 10 ROIs of interest and would like to know whether among all the 45 connections between those ROIs any of them show significant differences between two subject groups. To be able to do this you perform a two-sample t-test to evaluate between-group differences in those ROI-to-ROI connections, but then you still need to correct the individual statistics (e.g. one T-value for each connection) by the total number of connetions tested (in this exaple 45), so one way to do this is by selecting a p-FDR analysis-level <.05 threshold which will apply an FDR correction to those 45 individual p-values. If any connection survives this threshold then you can confidently say that [i]those individual connections [/i]show different strengths between the two subject groups. [/color]
[color=#000000]The "intensity"-based thresholding options are part of the Network Based Statistics (NBS) analyses, and these are typically used instead when you want to make inferences either about individual ROIs or about individual networks of ROIs (instead of inferences about individual connections). Often times (when looking at a relatively large number of ROIs and connections) connection-level inferences require a very strong correction and the analysis sensitivity/power may simpy be too low to reach any sort of significance at this level (e.g. for 100 ROIs you now have 4500 individual connections to test, so you may simply not have the power to identify individual connections that survive such a strong correction). Seed- and network- level inferences offer higher sensitivity at the cost of lower specificity. The way they work is by combining a (typically uncorrected) connection-level threshold with a properly corrected seed- or network-level threshold. For example, for the same between-group comparison in the example above, you may now use a connection-level threshold of p<.01 uncorrected (to threshold the individual connection results at this somewhat arbitrary level), and then, for each seed-ROI you may want to simply count the number of significant connections emanating from this seed-ROI (this is what the "NBS (by size)" statistics compute), or alternatively compute the weighted sum of those significan connections emanating from this seed-ROI weighted by the strength of those individual connection effects (this is what the "NBS (by intensity)" statistics compute), and then determine whether those counts are themselves significant (this is performed in NBS using permutation/randomization analyses). If you have more than a single seed-ROI of interest then you would also need to apply a multiple-comparison correction of those seed-level statistics for the number of seeds tested (and this is what the associated seed-level p-FDR threshold does). So, summarizing, in this example you would simply activate/check both a connection-level threshold (and enter there p-uncorrected p<.01) and a seed-level threshold (select "seed-ROI (NBS by size)", and enter a p-FDR < .05 threshold there; note that you would need to click on the 'enable permutation analyses' button first to enable this thresholding option). If any ROI survives this threshold then you can confidently say that [i]those individual ROIs[/i] show different patterns of connectivity between the two subject groups. [/color]
[color=#000000]Let me know if this clarifies. I realize that the sheer number of potential thresholding combinations in ROI-to-ROI analyses might be a bit excessive/confusing and we are thinking of ways to simplify this interface and/or make it a bit more intuitive so any thoughts/suggestions are most welcome.[/color]
[color=#000000]Best[/color]
[color=#000000]Alfonso[/color]
[i]Originally posted by Jeff Browndyke:[/i][quote]What do the p-FDR intensity correction and p-FDR analysis-level correction options actually denote or measure?
Which is more appropriate? To not p-FDR at the intensity level (whatever that corrects for or denotes), but correct at the F-test ROI multiple comparison correction level? Or, p-FDR at the intensity level and not correct for subsequent F-test ROI comparisons?
Thanks,
Jeff[/quote]
[color=#000000]The p-FDR analysis-level correction is typically used when you want to make inferences about individual connections. It corrects the individual connection-level statistics for the total number of individual connections in your entire analysis (e.g. the size of the ROI-to-ROI matrix for the selected ROIs). In this case (for connection-level inferences) you typically just want to use this connection-level threshold (and disregard/uncheck any seed-level or network-level additional thresholding options). A typical example would be if you have say 10 ROIs of interest and would like to know whether among all the 45 connections between those ROIs any of them show significant differences between two subject groups. To be able to do this you perform a two-sample t-test to evaluate between-group differences in those ROI-to-ROI connections, but then you still need to correct the individual statistics (e.g. one T-value for each connection) by the total number of connetions tested (in this exaple 45), so one way to do this is by selecting a p-FDR analysis-level <.05 threshold which will apply an FDR correction to those 45 individual p-values. If any connection survives this threshold then you can confidently say that [i]those individual connections [/i]show different strengths between the two subject groups. [/color]
[color=#000000]The "intensity"-based thresholding options are part of the Network Based Statistics (NBS) analyses, and these are typically used instead when you want to make inferences either about individual ROIs or about individual networks of ROIs (instead of inferences about individual connections). Often times (when looking at a relatively large number of ROIs and connections) connection-level inferences require a very strong correction and the analysis sensitivity/power may simpy be too low to reach any sort of significance at this level (e.g. for 100 ROIs you now have 4500 individual connections to test, so you may simply not have the power to identify individual connections that survive such a strong correction). Seed- and network- level inferences offer higher sensitivity at the cost of lower specificity. The way they work is by combining a (typically uncorrected) connection-level threshold with a properly corrected seed- or network-level threshold. For example, for the same between-group comparison in the example above, you may now use a connection-level threshold of p<.01 uncorrected (to threshold the individual connection results at this somewhat arbitrary level), and then, for each seed-ROI you may want to simply count the number of significant connections emanating from this seed-ROI (this is what the "NBS (by size)" statistics compute), or alternatively compute the weighted sum of those significan connections emanating from this seed-ROI weighted by the strength of those individual connection effects (this is what the "NBS (by intensity)" statistics compute), and then determine whether those counts are themselves significant (this is performed in NBS using permutation/randomization analyses). If you have more than a single seed-ROI of interest then you would also need to apply a multiple-comparison correction of those seed-level statistics for the number of seeds tested (and this is what the associated seed-level p-FDR threshold does). So, summarizing, in this example you would simply activate/check both a connection-level threshold (and enter there p-uncorrected p<.01) and a seed-level threshold (select "seed-ROI (NBS by size)", and enter a p-FDR < .05 threshold there; note that you would need to click on the 'enable permutation analyses' button first to enable this thresholding option). If any ROI survives this threshold then you can confidently say that [i]those individual ROIs[/i] show different patterns of connectivity between the two subject groups. [/color]
[color=#000000]Let me know if this clarifies. I realize that the sheer number of potential thresholding combinations in ROI-to-ROI analyses might be a bit excessive/confusing and we are thinking of ways to simplify this interface and/or make it a bit more intuitive so any thoughts/suggestions are most welcome.[/color]
[color=#000000]Best[/color]
[color=#000000]Alfonso[/color]
[i]Originally posted by Jeff Browndyke:[/i][quote]What do the p-FDR intensity correction and p-FDR analysis-level correction options actually denote or measure?
Which is more appropriate? To not p-FDR at the intensity level (whatever that corrects for or denotes), but correct at the F-test ROI multiple comparison correction level? Or, p-FDR at the intensity level and not correct for subsequent F-test ROI comparisons?
Thanks,
Jeff[/quote]