Error on bridgesampling

General
1

I’m fitting Negative Binomial in Bayesian frame work. I want to check the sensitivity of priors in the dispersion parameters. I used two different priors and compared it using the Bayes Factor, and I chose bridgesampling. Unfortunately, I got the following error for many times. I read here about the related error, and they recommend increasing the number of iteration. I tried the number of iteration with 30,000, but I got the same.
FYI – My number of observations is around 6,000.

Error in jj[2,]: Subscript out of bounds
In addition: Warning messages:
1: Infinite value in iterative scheme, returning NA.
    Try rerunning with more samples. 
2: logml could not be estimated within maxiter, rerunning with adjusted starting value. 
     Estimate might be more variable than usual.

Therefore:
1- Would you please recommend something to solve it? Or
2 - Do we have other packages or techniques use to compare the model with different priors?

Regards

1 Like
2

Can you post

traceback()

after that happens?

3

image
image1056×648 18.6 KB

4

Can’t tell if this is a rstan and / or a bridgesampling issue. @Henrik_Singmann ?

5

Hmm, failing with an error should not happen. Are you using the latest version of our package (1.0-0)?

In any case, I suspect the true underlying issue is, as indicated, a too low number of samples. If you have a rather high number of parameters, than you just need a lot of samples. I would suggest something like 100,000 as a starting point. However, if possible, you can still post a reproducible example to https://github.com/quentingronau/bridgesampling/issues
In case it fails, it should do so with NA and not with an error.

6

I’ve actually run into this with models before, I’ve got an example with data that I’ll put up on the issue tracker

7

Thanks so much for your reproducible example. Was indeed a bug. The development version now correctly returns NA in this case.

2 Likes