Firth p-values in PLINK vs logistf

[Joelle Mbatchou]

Hi Chris,

One of the REGENIE user recently pointed out discrepancies between running exact Firth in REGENIE (without using whole genome regression) vs running it in PLINK which got us to do some testing and led us to compare with logistf (which I believe PLINK also followed for the Firth implementation). 

We tested on a small dataset which I included here (in data.tgz) using the script in "check_exact_firth.sh" (requires 'plink1.9' and 'plink2' binaries, and packages 'RServe' and 'logistf' in R -- I commented out the code which runs REGENIE). The plot of the resulting -log10 p-values is attached in 'ex_pval_plots.png'. 

As I am not familiar with the Firth implementation in PLINK, I just wanted to bring this to your attention in case this was not something to be expected.

Cheers,

Joelle

[Joelle Mbatchou]

I probably should mention the binary versions (should also be included in the log files in 'output.tgz')

PLINK 1.9: PLINK v1.90b6.25 64-bit

PLINK 2: PLINK v2.00a3 AVX2

This was run on OSX platform.

[Christopher Chang]

Hi Joelle,

Thanks for the heads-up.  PLINK 1.9 and 2.0 have been using mostly single-precision arithmetic for logistic and Firth regression since the work done in https://academic.oup.com/gigascience/article/6/5/gix009/3057432 , and 2.0's Firth implementation is using slightly looser convergence parameters than logistf's defaults.  So some small differences from logistf are expected.

I am open to moving everything back to double-precision if there is a clear scientific problem with the current implementation.

[Chris Chang]

fyi, PLINK 2.0's default Firth implementation should now match the latest logistf package (except with Wald-test p-values).

--
You received this message because you are subscribed to the Google Groups "plink2-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to plink2-users...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/plink2-users/3664a31e-9d46-4b75-8651-4c85b044ad7fn%40googlegroups.com.