"Error in if (varx == 0) { : missing value where TRUE/FALSE needed" with fit$loo() but not loo::loo(fit$draws("log_lik"))

Hi,

Unfortunately I don't have a reproducible example because this is just popping up with a few versions of a big model. When I fit the model with cmdstanr, the following works:

> loo::loo(fit_mvn$draws("log_lik"))

Computed from 400 by 590 log-likelihood matrix.

         Estimate     SE
elpd_loo -31300.8 3489.9
p_loo       323.5   14.9
looic     62601.5 6979.8
------
MCSE of elpd_loo is NA.
MCSE and ESS estimates assume independent draws (r_eff=1).

Pareto k diagnostic values:
                          Count Pct.    Min. ESS
(-Inf, 0.62]   (good)     367   62.2%   37      
   (0.62, 1]   (bad)      177   30.0%   <NA>    
    (1, Inf)   (very bad)  46    7.8%   <NA>    
See help('pareto-k-diagnostic') for details.

But this doesn't:

> fit_mvn$loo()
Error in if (varx == 0) { : missing value where TRUE/FALSE needed

I checked and there's no NAs anywhere:

> fit_mvn$draws("log_lik") |> anyNA()
[1] FALSE

Does anyone have any idea? Sorry I can't be more helpful with reproducible code. I'd be happy to share the Stan program with simulation file.

Thanks,

Matt

Start by comparing the fit$loo() entry point with loo::loo(fit$draws("log_lik")), using the reported error and the 400-by-590 log-likelihood matrix as the reference case. Reproduce the discrepancy with the Stan program and simulation file if they are shared, then verify that both entry points handle this input consistently without the missing-value error.