Different rhat for Rstan and CmdStanR

[Alex-Cremers]

Hello,

I have a model fitted with cmdstanR which did not converge properly. The Rhat for one of the parameters is 1.33 when I look at model_fit$summary(), however if turn the model into a stanfit object, the summary now indicates a Rhat of 1.48.

I first thought that the two summary functions call different rhat functions, but the following lines do show the same discrepancy:

bayesplot::rhat(model_fit,pars="alpha")
rstan_fit = rstan::read_stan_csv(model_fit$output_files())
bayesplot::rhat(rstan_fit,pars="alpha")

Is there a reason why this could happen, or it it a bug? If needed I can upload the draws for this particular parameter. Using posterior::rhat on the matrix of draws returns 1.33, as with the summary method for the CmdStanFit object.

Best regards,

[jgabry]

Hi @Alex-Cremers, sorry nobody responded until now!

Internally bayesplot::rhat uses the values computed by CmdStanR and RStan (bayesplot doesn't actually compute anything itself), so the issue is that CmdStanR's summary method is using posterior's Rhat whereas RStan's summary method is still using the older version of Rhat. After you convert model_fit to rstan_fit the Rhat values that are used will come from RStan's summary method and will be different.

Sorry again for the (very) slow reply!