mcmcOutput (version 0.1.3)

Get diagnostics: Get diagnostic statistics

Description

These functions calculated diagnostic statistics for objects of class mcmcOutput. Optionally, only values that are worse than a predefined threshold will be returned, and values can be sorted so that the worst are at the start of the output vector.

Usage

getMCE(x, pc=TRUE, bad=5, sort=TRUE)
getNeff(x, bad=10000, sort=TRUE)
getRhat(x, bad=1.1, sort=TRUE)

Value

A named vector with the values of the diagnostic. Values of NA will be excluded unless bad = NA. It may have length 0 if no values are bad.

Arguments

x

an object of any class with MCMC output that can be coerced to class mcmcOutput.

pc

if TRUE, the value of the MC error as a percentage of the posterior SD will be returned.

bad

threshold for "bad" values: only values above this (for getMCE or getRhat) or below this (for getNeff) will be returned. If bad = NA, all values will be returned, including NAs.

sort

if TRUE, the values will be sorted, with the worst at the top.

Author

Mike Meredith

Details

getRhat returns the Brooks-Gelman-Rubin (BGR) convergence diagnostic (Brooks & Gelman 1998), a non-parametric 'interval' estimator of the 'potential scale reduction factor' for MCMC output. Similar to the function coda::gelman.diag, but faster when thousands of parameters are involved and will not cause R to crash.

getMCE returns the Monte Carlo standard error calculated using the batch method of Lunn et al (2013, p77); see also Roberts (1996).

getNeff returns the effective number of draws taking account of autocorrelation within each chain. It is a wrapper for coda::effectiveSize.

References

Brooks, S.P. & Gelman, A. (1998) General methods for monitoring convergence of iterative simulations. Journal of Computational and Graphical Statistics, 7, 434-455.

Lunn, D., Jackson, C., Best, N., Thomas, A., & Spiegelhalter, D. (2013) The BUGS book: a practical introduction to Bayesian analysis, Chapman and Hall.

Roberts, G.O. (1996). Markov chain concepts related to sampling algorithms. In Markov Chain Monte Carlo in practice (eds W.R. Gilks, D.J. Spiegelhalter & S. Richardson). Chapman & Hall, London.

Examples

data(mcmcListExample)
mco <- mcmcOutput(mcmcListExample)

getMCE(mco, bad=2)
getMCE(mco, bad=0)  # returns all except NAs
getMCE(mco, bad=NA)  # returns all including NAs
getMCE(mco, bad=NA, sort=FALSE)  # returns all, in original order

getNeff(mco, bad=2800)
getNeff(mco, bad=Inf)  # returns all except NAs
getNeff(mco, bad=NA)   # returns all including NAs

getRhat(mco)
getRhat(mco, bad=0)
getRhat(mco, bad=NA, sort=FALSE)

# Extract the values with 'bad' MCE and do plots:
( badNodes <- names(getMCE(mco, bad=2)) )
( badMco <- mco[badNodes] )
plot(badMco)