testNhoods with multiple contrasts

[gianfilippo]

testNhoods with multiple contrasts

[MikeDMorgan]

Hi @wmacnair
This comes from the behaviour of glmQLFTest and topTags in edgeR, which computes (I assume) an omnibus p-value with multiple contrasts, which is what is fed into the spatialFDR correction, but returns a log fold change for each contrast. Note, the exact behaviour of multiple contrasts in glmQLFTest is also undocumented, but was clarified by Gordon K Smyth here: https://support.bioconductor.org/p/117466/

I agree this should be documented better, and I will add it to the to-do list for the next Bioc release.

[MikeDMorgan]

Hi @gianfilippo I'd recommend runnning each contrast separately for now, as the formal testing will only return the results for one contrast, not an omnibus test. If individual nhoods are either entering or leaving the "significant" set the it is probably because they are close to what ever threshold you are using.

[gianfilippo]

Ok, i asked because the the help function mentions a "string vector" and I assumed it would take a vector of contrasts.
Thanks for explaining

[wmacnair]

Hi, just following up on the recommendation to run contrasts separately.

I think if the recommendation is not to run multiple contrasts at the moment, this should be made clear in the package messages and in the documentation. The safest option would be to not allow it until it is ready (although it should be pretty straightforward to add a loop, a contrast column and a call to rbind, I would assume?), but otherwise it should at least raise a warning.

Out of curiosity, if I run multiple contrasts, what test does the resulting PValue correspond to? Something like ANOVA?

Thanks!
Will

[wmacnair]

Hi @MikeDMorgan, thanks for your explanation :)

I'm wondering what the most common use case for specifying multiple contrasts would be. My expectation is that most users doing that would be interested in having separate p and FDR values for each contrast, rather than an omnibus p-value. If that's the case, then it could be worth adding an internal loop to get values for a list of contrasts.

Cheers
Will