Hangartner et al. (2011) proposed a convergence diagnostic for discrete Markov chains. A simple Pearson's Chi-squared test for two or more non-overlapping periods of a discrete Markov chain is a reliable diagnostic of convergence. It does not rely upon the estimation of spectral density, on suspect normality assumptions, or determining overdispersion within a small number of outcomes, all of which can be problematic with discrete measures. A discrete Markov chain is split into two or more non-overlapping windows. Two windows are recommended, and results may be sensitive to the number of selected windows, as well as sample size. As such, a user may try several window configurations before concluding there is no evidence of non-convergence.
As the number of discrete events in the sample space increases, this diagnostic becomes less appropriate and standard diagnostics become more appropriate.
Hangartner.Diagnostic(x, J=2)This required argument is a vector of marginal posterior
samples of a discrete Markov chain, such as selected from the output
of LaplacesDemon.
This argument specifies the number \(J\) of windows to be used, and defaults to \(J=2\).
The Hangartner.Diagnostic returns an object of class
hangartner, including the output from a Pearson's Chi-squared
test. A frequentist p-value less than or equal to 0.05 is usually
considered to be indicative of non-convergence.
Hangartner, D., Gill, J., and Cranmer, S., (2011). "An MCMC Diagnostic for Purely Discrete Parameters". Paper presented at the annual meeting of the Southern Political Science Association, Hotel InterContinental, New Orleans, Louisiana Online.
# NOT RUN {
library(LaplacesDemon)
N <- 1000
K <- 3
x <- rcat(N, rep(1/K,K))
hd <- Hangartner.Diagnostic(x, J=2)
hd
# }
Run the code above in your browser using DataLab