Only Firth regression in latest plink2 version.

[Tore Eriksson]

Hi,

We are re-running some GWAS that we previously ran using PLINK v2.00a1LM, using alpha 3 this time. We are doing logistic regression with Firth regression as a fallback. Using alpha 1 we had one Firth regression per 1000 SNPs. However, using alpha 3 all SNPs use Firth regression accoring to the FIRTH? column. We get the same result whether we use the 'firth-residualize' modifier or not.

Did the convergence criteria for the logistic regression change between alpha 2 and 3? If so, can these criteria be adjusted back to the previous setting or is this not recommended?

Regards,

Tore

[Christopher Chang]

Using the older version is definitely NOT recommended.

[Tore Eriksson]

Understood. Alpha 2.3 final still runs without Firth regression though. We are analyzing UK Biobank data and the case/control ratio is quite unbalanced (10x or more). Is this the reason that plink2 only uses Firth regression? I was also wondering if there was any possibility that logistic regression was used but that there's a bug when writing the output, but I guess such a bug would have been caught by now.

[Chris Chang]

If you provide e.g. two variants and a phenotype/covariate set where the behavior is different, I can take a look and explain what’s going on.

--
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/3a9716be-04a8-4094-a761-dc0c1a3c0d66n%40googlegroups.com.

[Christopher Chang]

Ok, there was an actual separation-detection bug here which only manifested when a categorical covariate was present.  This is fixed in today's build.

[Tore Eriksson]

The latest build works as expected. Many thanks for fixing the problem this quickly!