Wednesday, September 15, 2021

"The Dual Mechanisms of Cognitive Control Project": post-publication analyses

The recently-published "The Dual Mechanisms of Cognitive Control Project" paper (preprint publisher) describes the motivations and components of the project as a whole, but also contains several analyses of its task fMRI data. The supplemental information for the manuscript has the R (knitr) code and input files needed to generate the results and figures in the manuscript. 

The supplemental information now has more results than those included in the paper, however: versions of the analyses using different sets of participants, parcels, and estimating the HDR curves (GLMs), which I will briefly describe here.

For example, Figure 5 (below) shows the estimated HDR curves for each task (high - low control demand) in the Baseline (blue, BAS) and Reactive (green, REA) sessions. The grey shading marks the target; when we expect the control demand difference to be greatest. Of interest is that the estimates are greater in the target window for all tasks and sessions, with the Baseline session estimates larger than those from the Reactive session (see manuscript for more explanation and framing).

The top version (supplemental file) includes 80 participants in the estimates (some related), averages the estimates from a set of 35 parcels (from the Schaefer 400x7 parcellation) found to be particularly sensitive to DMCC tasks, and uses GLMs estimating one knot for every 2 TRs.

The second version (supplemental file) shares the theoretically interesting aspects: curves mostly peak in the target (grey) area, blue curves mostly above green. There are many differences, though: the second graph is from a post-publication analysis using the DMCC55B participants (55 unrelated people; 48 of whom are in both 55B and the 80-participant set), the set of 32 Schaefer 400x7 parcels approximating the Core Multiple Demand network (12 parcels are in both sets), and GLMs estimating one knot for every TR.

It is reassuring to see that the analysis results are generally consistent despite these fairly substantial changes to its inputs. Sometimes results can look great, but are due to a statistical fluke or overfitting; in these cases small changes to the analysis that shouldn't matter (e.g., removing or replacing several participants) often make large changes in the results. The opposite occurred here: fairly substantial changes to the parcels, participants, and (to a lesser extent) GLMs led to generally matching results.

The paper's osf site now contains results files for all the different ways to set up the analyses, within the "postPublication_coreMD" and "postPublication_1TRpK" subdirectories. The variations:

  • 80 or 55 participants. Files for analyses using the DMCC55B participants have a "DMCC55B" suffix; files for the original set of 80 participants has either no suffix or "DMCC80".
  • 35 or 32 parcels. The set of 35 parcels identified via DMCC data are referred to as the 35-megaparcel or "parcels35"; the 32 parcels approximating the core MD are referred to as "core32".
  • GLMs with 1 or 2 knots per TR. The original analyses all used GLMs with 2 TRs per knot ("2TRpK"). The 1 TR per knot GLMs are abbreviated "1TRpK", including in the file names.

No comments:

Post a Comment