Inflation of beta-values from logistic analysis - it is version specific #684
Description
We have very discrepant results between versions 2 and 4 of REGENIE. Namely, in V4, the beta-values are substantially higher than what they are in V2, yet standard errors are comparable between runs. This is apparently resulting in massive inflation of test statistics in our V4 analyses. Example output headers from versions for the leading variants:
Version 4
CHROM GENPOS ID ALLELE0 ALLELE1 A1FREQ N TEST BETA SE CHISQ LOG10P
1 149368746 chr1:149368746:AT:A AT A 0.58238 173192 ADD 0.260464 0.0130714 398.193 87.8658
1 99834082 chr1:99834082:G:GAA G GAA 0.509135 172078 ADD 0.166545 0.00901437 345.49 76.3906
1 214357735 chr1:214357735:GA:G GA G 0.507556 173309 ADD 0.163555 0.00893237 339.313 75.0454
1 234589389 chr1:234589389:C:T C T 0.50116 173313 ADD 0.164097 0.00892871 341.687 75.5625
1 151715118 chr1:151715118:G:C G C 0.521015 173304 ADD 0.164103 0.00899881 336.699 74.4762
1 99823004 chr1:99823004:G:A G A 0.506091 173300 ADD 0.163032 0.00895211 335.578 74.2322
1 99831645 chr1:99831645:A:G A G 0.50638 173187 ADD 0.163076 0.00895848 335.291 74.1695
1 151731469 chr1:151731469:C:CA C CA 0.515142 171377 ADD 0.164302 0.00906022 332.918 73.6528
1 99816937 chr1:99816937:G:A G A 0.505958 173308 ADD 0.162349 0.00895072 332.862 73.6405
1 99813581 chr1:99813581:G:A G A 0.506033 173285 ADD 0.162363 0.00895208 332.819 73.6311
Version 2
CHROM GENPOS ID ALLELE0 ALLELE1 A1FREQ N TEST BETA SE CHISQ LOG10P
1 149368746 chr1:149368746:AT:A AT A 0.58238 173192 ADD 0.0191721 0.0150508 1.62263 0.693093
1 99834082 chr1:99834082:G:GAA G GAA 0.509135 172078 ADD 0.0340108 0.00967343 12.3623 3.35845
1 214357735 chr1:214357735:GA:G GA G 0.507556 173309 ADD 0.0379514 0.00957412 15.7139 4.13264
1 234589389 chr1:234589389:C:T C T 0.50116 173313 ADD 0.0331702 0.00957834 11.9922 3.27226
1 151715118 chr1:151715118:G:C G C 0.521015 173304 ADD 0.0368848 0.00959743 14.7742 3.91655
1 99823004 chr1:99823004:G:A G A 0.506091 173300 ADD 0.0311296 0.00960826 10.497 2.92239
1 99831645 chr1:99831645:A:G A G 0.50638 173187 ADD 0.0311383 0.00961564 10.4869 2.92
1 151731469 chr1:151731469:C:CA C CA 0.515142 171377 ADD 0.034597 0.00970546 12.7093 3.43908
1 99816937 chr1:99816937:G:A G A 0.505958 173308 ADD 0.0304739 0.00960859 10.0588 2.81924
1 99813581 chr1:99813581:G:A G A 0.506033 173285 ADD 0.0305154 0.00960987 10.0836 2.82508
For each version, I have bolded the beta and SE for the leading variant to further illustrate the problem.
We have also now seen this problem from another research group who use regenie 4.1. I'm attaching their log file and the header of the results file to help diagnosis.
runRegenie.20260225.10h47m.log
step2_BROAD_PTSD+pval.regenie.top100.log
Also, per this other person's post #631, note how there beta-values look scaled oddly high yet the SEs appear to be on a different measure scale.
So my conclusion is that there is a recently introduced problem in how output beta values are being scaled, which is subsequently results in the miscalculation of p-values.
Do you have any insights?Thanks for your help!
BW
Adam