← All research spotlights

Research spotlight · Unit 12

Aki Vehtari

Aalto University

Research field
Bayesian statistics, Markov chain Monte Carlo diagnostics, probabilistic programming

Aki Vehtari is Academy Professor in the Department of Computer Science at Aalto University in Finland and one of the developers of the probabilistic programming language Stan. With Andrew Gelman, Daniel Simpson, Bob Carpenter, and Paul-Christian Bürkner he showed that the classic R-hat diagnostic of 1992, which compares the variation within several Markov chains with the variation between them, misses convergence failures when a posterior has heavy tails or when chains agree in their means but differ in their spread. Their improved version works on the ranks of the draws and on the draws folded about their median, and it comes with an effective sample size for each quantile a report quotes. The unit presents this work as the modern answer to the question of whether a sampler can be trusted.

In the masterclass

Section 7, after the Metropolis algorithm, the trace plot, and the R-hat statistic of 1992, presents the improved diagnostic.

The research is presented in Unit 12: Bayesian Inference, in a research spotlight box that explains the result at the level of the unit and links it to what you have just learned.

The paper

Vehtari, A., Gelman, A., Simpson, D., Carpenter, B., and Bürkner, P.-C. (2021). Rank-Normalization, Folding, and Localization: An Improved R-hat for Assessing Convergence of MCMC (with Discussion). Bayesian Analysis, 16(2), 667‑718. https://doi.org/10.1214/20-BA1221

The scientists named here are not contributors to the Statistics Masterclass, are not affiliated with it, and have not endorsed it. The program presents and explains their published research, with references to the original papers.