Gelman and Rubin's diagnostic

Usage

gelman.diag(data, confidence = 0.95, transform=FALSE)

Arguments

data An mcmc.list object with more than one chain, and with starting values that are overdispersed with respect to the posterior distribution.
confidence the coverage probability of confidence interval for the shrink factor
transform a logical vector indicating whether variables in data should be transformed to improve the normality of the distribution. If set to TRUE, a log transform or logit transform, as appropriate, will be applied. If transform is of length 1 it is replicated to be equal to the number of variables in data.

Description

The "shrink factor" is calculated for each variable in data, together with upper and lower confidence limits. Approximate convergence is diagnosed when the upper limit is close to 1.

The confidence limits are based on the assumption that the stationary distribution of the variable under examination is normal. Hence the `transform' parameter may be used to the skewness of the distribution.

theory

Gelman and Rubin (1992) propose a general approach to monitoring convergence of MCMC output in which two or more parallel chains are run, with starting values that are overdispersed relative to the posterior distribution. Convergence is diagnosed when the chains have "forgotten" their initial values, and the output from all chains is indistinguishable. The gelman.diag diagnostic is applied to a single variable from the chain. It is based a comparison of within-chain and between-chain variances, and is similar to a classical analysis of variance.

There are two ways to estimate the variance of the stationary distribution. item{the mean of the empirical variance within each chain, W, and} item{the empirical variance from all chains combined, which can be expressed as

sigma.hat^2 = (frac{(n-1) B }{n}) + frac{W}{n}

where B is the empirical between-chain variance.} If the chains have converged, then both estimates are unbiased. Otherwise the first method will underestimate the variance, since the individual chains have not had time to range all over the stationary distribution, and the second method will overestimate the variance, since the starting points were chosen to be overdispersed.

The convergence diagnostic is based on the assumption that the target distribution is normal. A Bayesian credible interval can be constructed using a t-distribution with mean

mu.hat = Sample mean of all chains combined

and variance

V.hat = sigma.hat^2 + frac{B}{mn}

where m is the number of chains, and degrees of freedom estimated by the method of moments

d = frac{2*V.hat}{Var(V.hat)}

Use of the t-distribution accounts for the fact that the mean and variance of the posterior distribution are estimated.

The convergence diagnostic itself is

R = sqrt{frac{(d+3) V.hat}{(d+1)W}}

Values substantially above 1 indicate lack of convergence. If the chains have not converged, Bayesian credible intervals based on the t-distribution are too wide, and have the potential to shrink by this factor if the MCMC run is continued.

See Also

gelman.plot


[Package Contents]