susie_rss lbfs vs susie_ss lbfs

[zhana-optima]

Dear all,

Thank you for this fantastic package and all the hard work in the great statistics results.

I have a more general question regarding the log Bayes Factors produce by the different functions : susie_rss and susie_ss; single_effect_regression_rss and single_effect_regression_ss. At first exploration, it appears that rss uses approximate Bayes factors (Proposition 1, SuSie RSS manuscript) while ss uses results for irrelevance of null SNPs (Proposition 3, Susie RSS manuscript). Is this correct? In addition, (under the assumption that z, R, n are all well known summary data) in the current master branch it appears that even susie_rss defaults to susie_suff_stats -> defaults to using single_effect_regression_ss. Could you please provide some code infrastructure to potentially highlight the difference between susie_rss and susie_ss resulting BFs, CS, etc., and when the BFs are computed using the rss and the ss estimators? Which is the latest branch version you would recommend using? Which is the current cran version that is used in R?

Thank you very much in advance!

[pcarbo]

@zhana-optima We recommend using the latest version available on GitHub in the master branch. The version currently available on CRAN is 0.12.35.

In Proposition 1, the BFs are exact; there is no approximation.

We haven't done thorough empirical comparisons of the CS Bayes factors returned by susie_suff_stat and susie_rss, so if you find inconsistencies, please let us know.

[zhana-optima]

Thank you for the response. I have a specific question - is at any point the single_effect_regression_rss.R function used? Or is only the single_effect_regression_ss.R function used? In the current master branch it appears that all calculations default to susie_suff_stat and for that matter single_effect_regression_ss.R. Is this correct?

I am looking at the Susie RSS supplementary information (https://doi.org/10.1371/journal.pgen.1010299.s002) - the BFs from Prop1 appear to be Wakefield's asymptotic bayes factors.

[pcarbo]

I am looking at the Susie RSS supplementary information (https://doi.org/10.1371/journal.pgen.1010299.s002) - the BFs from Prop1 appear to be Wakefield's asymptotic bayes factors.

Perhaps there is a connection to the Wakefield asymptotic BF, but the Bayes factors in Proposition 1 are exact for the SER model; no approximation is being introduced here.

[stephens999]

Peter is correct - these BFs are exact for the linear regression model (conditional on sigma^2). Wakefield's asymptotic BF has the same form though.

[zhana-optima]

Thank you both for clarifying this - I now see how this easily flows to Prop3 for SER model (again in the supplementary information). Highly appreciate your speedy responses!

As a final clarification, is it safe to assume that in the current master branch susie_rss defaults to susie_suff_stat and for that matter to single_effect_regression_ss.R ?

[pcarbo]

Is at any point the single_effect_regression_rss.R function used?

@zhana-optima This was used in an earlier version of the package, but is now no longer used.

[zhana-optima]

Thank you @pcarbo - this is incredibly helpful!