1: \begin{abstract}

2:   Markov chain Monte Carlo is a key computational tool in Bayesian 

3:   statistics, but it can be challenging to monitor the convergence of an iterative stochastic algorithm.

4: In this paper we show that the convergence diagnostic \Rhat\ 

5: of \citet{Gelman+Rubin:1992} has serious flaws. Traditional \Rhat\ will fail to correctly diagnose

6: convergence failures when the chain has a heavy tail or when the variance varies across 

7: the chains. In this paper we propose an alternative rank-based diagnostic that fixes these 

8: problems. We also introduce

9:   a collection of quantile-based local efficiency

10:   measures, along with a practical approach for computing Monte Carlo error

11:   estimates for quantiles. We suggest that common trace plots should

12:   be replaced with rank plots from multiple chains. Finally, we give

13:   recommendations for how these methods should be used

14:   in practice.

15: \end{abstract}

16: