bayesfactor_models() recomputes brms model with rstan. Can we avoid that?

[DominiqueMakowski]

For the (I think many) people struggling with having a functioning rstan (typically on servers or with restricted laptops), cmdstanr has been great. However, bayesfactor_models recomputes the models using rstan, rendering it unusable with cmdstanr. Is there a solution?

m1 <- brms::brm(mpg ~ wt, data=mtcars, backend="cmdstanr", refresh=0)
m2 <- brms::brm(mpg ~ wt + hp, data=mtcars, backend="cmdstanr", refresh=0)


bayestestR::bayesfactor_models(m1, m2)
#> Warning: Bayes factors might not be precise.
#>   For precise Bayes factors, sampling at least 40,000 posterior samples is
#>   recommended.
#> Computation of Marginal Likelihood: estimating marginal likelihood,
#>   please wait...
#> Recompiling the model with 'rstan'
#> Error in compileCode(f, code, language = language, verbose = verbose): /Users/domma/AppData/Local/R/win-

^{Created on 2025-04-02 with reprex v2.1.1}

[strengejacke]

@mattansb ?

[mattansb]

Nope. CmdStan does not support bridge sampling, which is required for estimating the marginal likelihood and in turn the Bayes factor.

[DominiqueMakowski]

That's good to know :)

For future ref, some attempts were made: to make it compatible: quentingronau/bridgesampling#38

[mattansb]

Yeah, I've seen this. Once this is implemented, it should automatically work here as well (the re-fitting with rstan is done by brms, not us).