I have a collaborator that is interested in statistical testing of pathway activation between groups of cells (from different tissues) within specific cell type clusters using data from this Pan-cancer T cell atlas. But I have concerns/questions about applying statistical tests b/c the atlas is comprised of 27 distinct data sets, which they integrated with Seurat to anchor cells and generate clusters, but otherwise only the raw gene counts are available. They labeled cells by data set but did not provide individual sample labels for data sets with multiple samples (applies to most data sets).
I am concerned about the effects of data set-level and sample-level batch effects confounding the statistical testing. I did come across the stratified wilcoxon test (aka Van Elteren test) which may be more robust to batch effect (and has been applied to scRNA-seq for differential expression testing).
I also came across SCPA but the developer explicitly states in GitHub issue 33 that for SCPA "there's no inbuilt method to model for batch effects directly in SCPA -- it just relies on you having the cleanest possible upstream processing." which is rather hand wavy IMO.
I would greatly appreciate any thoughts or ideas people may have, I'm concerned I may be overthinking the issue.
