Hi Harriet,

This change was somewhat overdue but I have now added the possibility of defining subject-specific TR values (instead of assuming the same TR for all subjects) to the latest CONN release (CONN14h). Please let me know if you run into any issues.

A few things to keep in mind when including subjects with different acquisition TRs.

1) make sure that your band-pass filter defined in the Preprocessing step is below half the sampling frequency of *all* subjects. By default, if you use a higher 'low-pass' threshold (e.g. set the band-pass filter to [.01 inf]), CONN will automatically convert that 'inf' to the maximum frequency of your data (1/TR/2). This works perfectly fine if all of your subjects have the same TR, but if they don't you will effectively end up having different frequency filters applied to different subjects, which can introduce potential biases in your connectivity estimates. By making sure your 'low-pass' threshold is below 1/TR/2 for all subjects (i.e. below 1/max(TR)/2) you will end up looking at exactly the same frequency band across all subjects.

2) depending on whether differences in TR are accompanied by equivalent differences in scanning length the most likely scenario is that you will end up with different number of acquisitions across subjects as well. So even if the resulting connectivity measures will not have biases they still will have different degrees of freedom and different associated measurement errors across subjects. The impact of this is typically reduced by the fact that the between-subject variance is often considerably larger than the within-subject variance / measurement errors, and standard second-level analyses behave robustly to this 'heteroscedasticity' condition, but this is just something to keep in mind, and if the differences are very large perhaps even consider using alternative methods such as permutation tests for your second-level analyses. 

3) last but not least, despite our best efforts to 'equalize' the effect of different TR's there might still be other unaccounted-for between-group differences introduced by the different scanning parameters across your different sites, so consider adding additional second-level regressors that account for these potential differences (e.g. 'site' dummy-coded regressors), as well as keeping these potential confounding effects in mind when interpreting other potentially-related between-subject effects.

Hope this helps
Alfonso


Originally posted by Harriet Johnston: