Dear Renzo,

The low (<80%) medium (80-95%) high (>95%) ranges for these variables are only rough heuristics. In general the "low" category is meant to signify the presence of big problems in quality or sensitivity which should be clearly visible from the raw data and which should be addressed before any analyses, and the "high" category is meant to mark a reasonable good quality threshold where any remaining limitations in quality or sensitivity are not expected to affect dramatically the analyses. For Data Validity the 95% "high" threshold is set at a value where the observed biases in FC distributions signaled by the peak/mode are smaller than 5% of the FC standard deviation (signifying a relatively small departure from a centered distribution), for Data Quality the 95% threshold is set to a value where the null-hypothesis distribution differences in QC-FC correlations from randomization/permutation analyses are below 5% (overlap coefficient, again signifying a relatively small departure from the null-hypothesis expectation of no QC-FC associations), and for Data Sensitivity the 95% threshold, while it departs from the 80% power/sensitivity threshold commonly used in power analyses from the same measure, it is simply meant to encourage higher-powered studies and also highlight the difference between group-level and subject-level sensitivity (most current studies have >95% sensitivity at the group-level to identify "significant" connections in the group/population, but they suffer from considerably lower sensitivity at the subject level to precisely quantify connectivity in individual subjects). In all cases, at least from my own experience across many different datasets from varied acquisition characteristics (e.g. 1000FCP, HCP), ~95% values may require some work but are in most/all cases achievable through some combination of denoising techniques and subject exclusion criteria, but I am of course looking forward to hearing from your own experience and happy to adapt my recommendations/thresholds as needed from the feedback! 

Hope this helps

Alfonso

Originally posted by Renzo Torrecuso:

Dear Alfonso,

Thanks a lot for your reply.

Could you please clarify for us, hard CONN users, one more step in this direction?

I have analyzed the same dataset using the liberal, intermediate and conservative preprocessing options.

On the (excelent) new version of conn we get in the denoising step 3 metrics about the data quality in the Quality Control display. 

Here´s my question, after reading the clarifying explanation provided in CONN´s documentation (https://web.conn-toolbox.org/fmri-method...) :

Could you please provide any reference or explanation on the thresholds you mention "low (<80%), intermediate (80–95%), and high (>95%) values."?

Thanks a lot

All best

Renzo