I keep getting Rhat below 1 (for usually about half of my parameters of interest at the level of 9.99e-01). A recent run (see below) reported Rhat < 1 for even smaller values (eg, 9.94e-01, 9.96e-01, 9.97e-01). Anyone knows why Rhat < 1 is reported? ( @bbbales2 kindly told me that Rhat in theory sho…

It’s not a problem, it’s not a rounding error, it’s just that as the B and W are estimated from finite number of draws this can be slightly less than 1 just by chance. It is more likely to observe this if you run very short chains as you are now doing with just 150 iterations. It is more likely to o…

[image] tlyim: Anyone knows why Rhat < 1 is reported? ( @bbbales2 kindly told me that Rhat in theory should be at least 1.) I think that you must have misunderstood him. Rhat should be close to 1, but it doesn’t matter if it’s slightly more or slightly less than 1.

Theoretically, it should be bigger than 1, According to https://arxiv.org/pdf/1903.08008.pdf the formula is \hat{R}^2 = \frac{N-1}{N} + \frac{1}{N}\frac{B}{W}, where B and W are both non-negative. So if you get something less than one it indicates some numerical instability in the computation. But …

[image] avehtari: We could threshold it so that values below 1 are reported as 1, but then we would miss warning examples like this where very shirt chains are used :) Thanks for your clear explanation. I think better to leave it as it is. I hope this post would be useful to others who also h…

[image] tlyim: Though not showing in the cmdstan report is that I ran warmup=800, which is about 5 times of the sampling (150). It’d be worth doing a save_warmup=1 run and plotting traceplots. Plot a few parameters and see what’s happening. If it doesn’t look like (from a far off view) that …

Rhat < 1 (as low as 9.94e-01). Why?

General

tlyim June 19, 2019, 4:28pm 7

Thanks for your suggestions. I am giving a reply in my another post.

Topic		Replies	Views
Divergent Transitions with R_hat < 1.1 Modeling fitting-issues , performance , shinystan	14	2519	June 13, 2019
Understanding Rhat with respect to the results from Bayesian Hierarchical modeling General	2	1512	March 19, 2020
Rhat of one estimate always close to 1 Modeling fitting-issues	2	216	February 29, 2024
False warning about max Rhat? - rstan General	3	3047	August 11, 2019
Traceplot vs Rhat for convergence General	11	2572	April 3, 2024

Rhat < 1 (as low as 9.94e-01). Why?

Related topics