search

saigegit / SAIGE

Development for SAIGE and SAIGE-GENE(+)
GNU General Public License v3.0
66 stars 29 forks source link

Can --invNormalize be turned off for any value of variance? #55

Closed mariosaare closed 1 year ago

mariosaare commented 2 years ago

Dear developer,

I am trying to compare SAIGE output to REGENIE output. I am using a continuous variable with values close to zero (range -0.03 - 0.66). REGENIE runs without the option --apply-rint (apply rank inverse normal transformation), but I cannot run SAIGE without normalization. I get the following error message:

Error in glmmkin.ai_PCG_Rcpp_Quantitative(bedFile, bimFile, famFile, Xorig, : WARNING: variance of the phenotype is much smaller than 1. Please consider invNormalize=T Calls: fitNULLGLMM -> system.time -> glmmkin.ai_PCG_Rcpp_Quantitative

Is there any way to avoid normalization in SAIGE? If not, are the normalization methods equivalent, so that I could compare the outputs after normalization in REGENIE as well?

Best regards, Mario Saare

saigegit commented 1 year ago

Hi @mariosaare,

Very sorry for our late reply. SAIGE tries to fit a linear mixed model. When the variance of the phenotype is too small or too large, the algorithm can't converge well, so the inverse normalization is required.

Thanks, Wei