Gelman and Rubin (1992) proposed a general approach to monitoring
convergence of MCMC output in which \(m > 1\) parallel chains are
updated with initial values that are overdispersed relative to each
target distribution, which must be normally distributed. Convergence
is diagnosed when the chains have `forgotten' their initial values,
and the output from all chains is indistinguishable. The
Gelman.Diagnostic function makes a comparison of within-chain
and between-chain variances, and is similar to a classical analysis of
variance. A large deviation between these two variances indicates
non-convergence.
This diagnostic is popular as a stopping rule, though it requires
parallel chains. The LaplacesDemon.hpc function is an
extension of LaplacesDemon to enable parallel chains.
As an alternative, the popular single-chain stopping rule is based on
MCSE.
Gelman.Diagnostic(x, confidence=0.95, transform=FALSE)This required argument accepts an object of class
demonoid.hpc, or a list of multiple objects of class
demonoid, where the number of components in the list
is the number of chains.
This is the coverage probability of the confidence interval for the potential scale reduction factor (PSRF).
Logical. If TRUE, then marginal posterior
distributions in x may be transformed to improve the
normality of the distribution, which is assumed. A log-transform is
applied to marginal posterior distributions in the interval \((0,
\infty]\), or a logit-transform is applied to marginal
posterior distributions in the interval \((0,1)\).
A list is returned with the following components:
This is a list containing the point-estimates of the
potential scale reduction factor (labelled Point Est.) and
the associated upper confidence limits (labelled Upper C.I.).
This is the point-estimate of the multivariate potential scale reduction factor.
To use the Gelman.Diagnostic function, the user must first have
multiple MCMC chains for the same model, and three chains is usually
sufficient. The easiest way to obtain multiple chains is with the
LaplacesDemon.hpc function.
Although the LaplacesDemon function does not
simultaneously update multiple MCMC chains, it is easy enough to
obtain multiple chains, and if the computer has multiple processors
(which is common), then multiple chains may be obtained simultaneously
as follows. The model file may be opened in separate, concurrent R
sessions, and it is recommended that a maximum number of sessions is
equal to the number of processors, minus one. Each session constitutes
its own chain, and the code is identical, except the initial values
should be randomized with the GIV function so the chains
begin in different places. The resulting object of class
demonoid for each chain is saved, all objects are read into one
session, put into a list, and passed to the Gelman.Diagnostic
function.
Initial values must be overdispersed with respect to each target
distribution, though these distributions are unknown in the
beginning. Since the Gelman.Diagnostic function relies heavily
on overdispersion with respect to the target distribution, the user
should consider using MCMC twice, first to estimate the target
distributions, and secondly to overdisperse initial values with
respect to them. This may help identify multimodal target
distributions. If multiple modes are found, it remain possible that
more modes exist. When multiple modes are found, and if chains are
combined with the Combine function, each mode is
probably not represented in a proportion correct to the distribution.
The `potential scale reduction factor' (PSRF) is an estimated factor
by which the scale of the current distribution for the target
distribution might be reduced if the simulations were continued for
an infinite number of iterations. Each PSRF declines to 1 as the
number of iterations approaches infinity. PSRF is also often
represented as R-hat. PSRF is calculated for each marginal posterior
distribution in x, together with upper and lower confidence
limits. Approximate convergence is diagnosed when the upper limit is
close to 1. The recommended proximity of each PSRF to 1 varies with
each problem, but a general goal is to achieve PSRF < 1.1. PSRF is an
estimate of how much narrower the posterior might become with an
infinite number of iterations. When PSRF = 1.1, for example, it may be
interpreted as a potential reduction of 10% in posterior interval
width, given infinite iterations. The multivariate form bounds above
the potential scale reduction factor for any linear combination of the
(possibly transformed) variables.
The confidence limits are based on the assumption that the
target distribution is stationary and normally distributed. The
transform argument may be used to improve the normal
approximation.
A large PSRF indicates that the between-chain variance is
substantially greater than the within-chain variance, so that longer
simulation is needed. If a PSRF is close to 1, then the associated
chains are likely to have converged to one target distribution. A
large PSRF (perhaps generally when a PSRF > 1.2) indicates convergence
failure, and can indicate the presence of a multimodal marginal
posterior distribution in which different chains may have converged
to different local modes (see is.multimodal), or the
need to update the associated chains longer, because burn-in (see
burnin) has yet to be completed.
The Gelman.Diagnostic is essentially the same as the
gelman.diag function in the coda package, but here it is
programmed to work with objects of class demonoid.
There are two ways to estimate the variance of the stationary distribution: the mean of the empirical variance within each chain, \(W\), and the empirical variance from all chains combined, which can be expressed as
$$ \widehat{\sigma}^2 = \frac{(n-1) W}{n} + \frac{B}{n}$$
where \(n\) is the number of iterations and \(B/n\) 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 initial values were chosen to be overdispersed (and this assumes the target distribution is known, see above).
This convergence diagnostic is based on the assumption that each
target distribution is normal. A Bayesian probability interval (see
p.interval) can be constructed using a t-distribution
with mean
$$\widehat{\mu}=\mbox{Sample mean of all chains combined,}$$
variance
$$\widehat{V} = \widehat{\sigma}^2 + \frac{B}{mn},$$
and degrees of freedom estimated by the method of moments
$$d = \frac{2\widehat{V}^2}{\mbox{Var}(\widehat{V})}$$
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
$$R=\sqrt{\frac{(d+3) \widehat{V}}{(d+1)W}}$$
Values substantially above 1 indicate lack of convergence. If the chains have not converged, then Bayesian probability intervals based on the t-distribution are too wide, and have the potential to shrink by this factor if the MCMC run is continued.
The multivariate version of Gelman and Rubin's diagnostic was proposed by Brooks and Gelman (1998). Unlike the univariate proportional scale reduction factor, the multivariate version does not include an adjustment for the estimated number of degrees of freedom.
Brooks, S.P. and Gelman, A. (1998). "General Methods for Monitoring Convergence of Iterative Simulations". Journal of Computational and Graphical Statistics, 7, p. 434--455.
Gelman, A. and Rubin, D.B. (1992). "Inference from Iterative Simulation using Multiple Sequences". Statistical Science, 7, p. 457--511.
Combine,
GIV,
is.multimodal,
LaplacesDemon,
LaplacesDemon.hpc,
MCSE, and
p.interval.
# NOT RUN {
#library(LaplacesDemon)
###After updating multiple chains with LaplacesDemon.hpc, do:
#Gelman.Diagnostic(Fit)
# }
Run the code above in your browser using DataLab