convergence diagnostics by n_eff?

[sophie]

Hi guys,

For my model fitted in Stan, I implement repeated sampling, which fits the same model on different datasets. I notice that among those unsatisfactory results, most values of n_eff are quite large, close to the number of post warm-up draws, for example, ranging from 500 to 750, where 750 is the number of post warm-up draws. Among the analysis which is reasonable, the n_eff values are smaller, ranging from 200 to 400 for 750 post warm-up draws. The trace-plots are almost non-informative for the convergence check. I am wondering whether you guys have some sense for this issue.  Thanks a lot!

Best,

Yajuan

[Daniel Lee]

Hi Sophie,

If I understand correctly, the models that are incorrect have high numbers of effective samples? When you look at the traceplots for the parameters here, are they actually moving? As in, if you did something like this in R, you'd find a high number:
length(unique(unlist(extract(fit, "<param name>"))))


Daniel

> --
> You received this message because you are subscribed to the Google Groups "stan users mailing list" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to stan-users+...@googlegroups.com.
> For more options, visit https://groups.google.com/groups/opt_out.
>
>

[Michael Betancourt]

If you run with --diagnostics=diagnostics.csv and send the file I can take a look at the possible lack of convergence.