different betas and standard errors with different versions of rvtest

[lubapardo]

To whom may it concern, We are using rvtest for genetic associations and noted a problem when doing logistic regression GWAS.

Commands where: rvtest --inVcf /home/chr${chr}.dose.vcf.gz --out trialCov_chr${chr} --covar COV.txt --covar-name var1, sex, var2 --pheno pheno.PHE.txt --dosage DS --single score --freqLower 0.010 --hide-covar FOR EXAMPLE: for Program version: 20170613 [INFO] Git Version: 444e2f1ebacf96ce1a299c17557150804ed7b42d I get a SNP with effects of:

CHROM POS REF ALT N_INFORMATIVE AF U V STAT DIRECTION EFFECT SE PVALUE 1 152285861 G A 657 0.0163387 6.21877 1.81368 21.323 + 25.858 5.59978 3.88042e-06

With different versions of the rvtest you get different effect sizes for the same input files;

  | For another version (Program version: 20190205; git unknown) I got    

1       152285861       G       A       657     0.0163387       6.21877 1.81368 21.323  +       3.42882 0.74254 3.88042e-06   I g

  | For another version (Program version: 20190205; git unknown) I got     1       152285861       G       A       657     0.0163387       6.21877 1.81368 21.323  +       3.42882 0.74254 3.88042e-06  
I got the same p-values but the effects are different and this is making the results unreliable. Could you explain why this happens?  
It also happens with other snps. We have used the same input files and commands. I hope you can clarify this.

Regards, Luba Pardo

ot the same p-values but the effects are different and this is making the results unreliable. Could you explain why this happens?   It also happens with other snps. We have used the same input files and commands. I hope you can clarify this.       Regards, Luba Pardo     Met vriendelijke groeten, -- | --

[zhanxw]

Sorry about this. The two versions in the comparison are 20170613 and 20190205. It is probably due to some bugs discovered in the earlier versions. Please use the newer version as there are fewer bugs. Thanks.

[epxlp]

We have observed a similar issue in some single variant meta-analysis that we are running with some rvtests results that were produced with earlier versions of rvtests. Interestingly we concluded that studies using the 2015 version look mostly ok. But those studies using 2016 or 2017 versions the beta and SE reported cannot be used to recapitulate the rvtest p-vals. Interestingly in the example you give, in the 2019 version the beta and SE seems to match the p-val in terms of what you would expect when converting a Z-statistic to a p-val, so it looks like something strange was introduced around 2016 and rectified by 2019.

I don't know if this means we need to 'throw away' any results that used the 2016/2017 versions or if we can correct them somehow......?