Model selection for rank/beta penalty criterion of the SLR

[erfanshekarriz]

Dear Zachary, I hope all is well.

I wanted to inquire about your pre-print and the beta/rank penalty parameter tuning for model selection.

Specifically in your manuscript, you've written :

• "Rather than varying the β parameter, we directly consider the rank r of the latent variable component L. For each r ∈ [2, rmax] along the “rank”-path with rmax ≪ p, we solve the SE-slr model and use the StARS routine along the λ-path. Across all equi-stable sparse networks, we select r that minimizes the extended Bayesian Information Criterion (BIC) of the full inverse covariance model."

I understand what you mean, but can't seem to find the "rank"-path in the SE object. A reviewer has asked me to at minimum undergo the same model selection using the extended Bayesian Information Criterion (eBIC) for my final revision, and your help would be most appreciated.

Is there a piece of code you could provide to select the optimal rank?

Best,

Erfan

[erfanshekarriz]

Nevermind!

I seem to have found it.

ranks <- round(exp(seq(log(2), log(32), len=6)))
se.slr <- spiec.easi(phy_taxfilt, method='slr', nlambda=50,
           lambda.min.ratio=1e-2, r=ranks, lambda.log=TRUE,
           pulsar.params=list(ncores=32, rep.num=30))
se.slr$ebic <- sapply(se.slr, function(x)
                       ebic(x$refit$stars, x$est$data,
                            x$est$loglik[x$select$stars$opt.index]))

Please correct me if I'm wrong, but I assume the model with the lowest eBIC is the best to go with.

Thanks for the awesome package!

Erfan

[zdk123]

That's correct. I'll add docs for this.