Closed mpx353 closed 2 years ago
joellembatchou
commented
2 years ago Hi Stefan,
We are working on improving the algorithm for the level 1 logistic regression to reduce convergence issues like yours. It seems you have already tried with no covariates and are already using LOOCV & also tried increasing the number of iterations. I don't have an obvious suggestion but will be looking into solutions to reduce convergence issues.
Cheers, Joelle
joellembatchou
commented
2 years ago Hi Stefan,
We have improved the ridge logistic regression algorithm in Step 1 to reduce convergence issues. You will need to pull from most recent commit 52648a49d38f518e6ea2b440f103ba159567850a and re-compile the binary. Let me know if this addresses your issue.
Cheers, Joelle
mpx353
commented
2 years ago Fabulous, I am re-compiling now and will let you know if the issue has been resolved.
Best, Stefan
mpx353
commented
2 years ago Just to let you know that the convergence issues have disappeared now. Many thanks for your help. I will proceed to Step 2 now and will be back in contact in case I encounter related issues.
Best, Stefan
Hi,
I just wondered whether there are any guidelines re analysis of very imbalanced BTs? I am currently trying to complete Step 1. with a case:control ratio of ~2500:50k (no covariates considered) in --loocv mode. When attempting Step 1 regression, I am receiving a non-converging warning.
Following other posts here and the FAQ I tried to play with --niter, increasing it to 1000 without luck.
I just wondered whether there are some further things I could try in order to be able to run step 2? At the moment I am using ~150k HQ SNPs in the model, should I try to add more (lower quality) SNPs? Happy to hear any thoughts.
Many thanks, Stefan