scale of MCMC sampling of beta

privefl / bigsnpr

R package for the analysis of massive SNP arrays.

https://privefl.github.io/bigsnpr/

183 stars 43 forks source link

scale of MCMC sampling of beta #499

Open yidingdd opened 1 month ago

yidingdd commented 1 month ago

Hi Florian,

Could you help to confirm the scale of betas from the MCMC sampling in your newest bigsnpr release?

It looks like the beta sample returned by snp_ldpred2_grid is on allelic scale, but the beta sample returned by snp_ldpred2_auto is on standardized scale, is it correct?

https://github.com/privefl/bigsnpr/blob/ee9f2ed5a069cf63401b9b922bae0b9492c46afc/R/LDpred2.R#L139 https://github.com/privefl/bigsnpr/blob/ee9f2ed5a069cf63401b9b922bae0b9492c46afc/R/LDpred2.R#L182

Best, Yi

privefl commented 1 month ago

Hi Yi,

For LDpred2-auto,

$beta_est is also on allelic scale
but $sample_beta is indeed on the standardized scale. You can probably use Matrix::Diagonal(x = scale) %*% auto$sample_beta to transform it to the allelic scale, where scale <- with(df_beta, sqrt(n_eff * beta_se^2 + beta^2)).

yidingdd commented 1 month ago

Hi Florian,

Thank you for confirming LDpred2-auto, for ldpred2-grid, do we need to use the same rescaling?

Best, Yi

privefl commented 1 month ago

No, the output is already on the allelic scale for LDpred2-grid.