5. Modeling

The ExtraLong data comes from many different projects, including, as of April 2021: PNC_CS, PNC_LG, AGGY, ONM, CONTE, GRMPY, FNDM, NEFF, NODRA, DAY2, SYRP and MOTIVE. Additional sessions and subjects will be added over time as MOTIVE and the Evolution of Psychosis projects continue to collect more data. The imaging data was frozen from September 2019, so there will be over a hundred, if not hundreds, of sessions to add down the line. At the time that the data was compiled, any subject on Flywheel that had more than one session across the initial projects listed was included in the dataset. Some subjects were never part of the PNC, and as such would not be appropriate to include in many analyses.

No matter the subset of ExtraLong that you plan the use, you will invariably have different acquisition parameters and scanners in your dataset. Therefore, it is crucial that in the modeling stage, you utilize longitudinal ComBat. Longitudinal ComBat adjusts data to handle batch effects, such as scanner/parameter groups. Note that it is important that your variables of interest are not confounded with these groups. The closer they are to being orthogonal to one another, the better. Considering protocols changed over time, age and parameter groups will be somewhat associated. If you would like to use GAMMs, please contact Dr. Joanne Beer, as she is currently working on implementing longitudinal ComBat with GAMMs. For more details on the sample, processing and analysis considerations, see the slides from the Longitudinal PNC meeting on April 16, 2021.