Estimating Monte Carlo variance from multiple Markov chains
share
The ever-increasing power of the personal computer has led to easy parallel implementations of Markov chain Monte Carlo (MCMC). However, almost all work in estimating the variance of Monte Carlo averages, including the efficient batch means (BM) estimator, focuses on a single-chain MCMC run. We demonstrate that simply averaging covariance matrix estimators from multiple chains (average BM) can yield critical underestimates in small sample sizes, especially for slow mixing Markov chains. We propose a multivariate replicated batch means (RBM) estimator that utilizes information from parallel chains, thereby correcting for the underestimation. Under weak conditions on the mixing rate of the process, the RBM and ABM estimator are both strongly consistent and exhibit similar large-sample bias and variance. However, in small runs the RBM estimator can be dramatically superior. This is demonstrated through a variety of examples, including a two-variable Gibbs sampler for a bivariate Gaussian target distribution. Here, we obtain a closed-form expression for the asymptotic covariance matrix of the Monte Carlo estimator, a useful result for benchmarking in the future.
READ FULL TEXT