Description Usage Arguments Details Author(s) References See Also Examples
Monte Carlo Standard Error (MCSE) is an estimate of the inaccuracy of
Monte Carlo samples, usually regarding the expectation of posterior
samples, E(theta), from Monte Carlo or
Markov chain Monte Carlo (MCMC) algorithms, such as with the
LaplacesDemon
or LaplacesDemon.hpc
functions. MCSE approaches zero as the number of independent posterior
samples approaches infinity. MCSE is essentially a standard deviation
around the posterior mean of the samples,
E(theta), due to uncertainty associated with
using an MCMC algorithm, or Monte Carlo methods in general.
The acceptable size of the MCSE depends on the acceptable uncertainty
associated around the marginal posterior mean,
E(theta), and the goal of inference. It has
been argued that MCSE is generally unimportant when the goal of
inference is theta rather than
E(theta) (Gelman et al., 2004, p. 277), and
that a sufficient ESS
is more important. Others perceive
MCSE to be a vital part of reporting any Bayesian model, and as a
stopping rule (Flegal et al., 2008).
In LaplacesDemon
, MCSE is part of the posterior
summaries because it is easy to estimate, and Laplace's Demon prefers
to continue updating until each MCSE is less than 6.27% of its
associated marginal posterior standard deviation (for more information
on this stopping rule, see the Consort
function), since
MCSE has been demonstrated to be an excellent stopping rule.
Acceptable error may be specified, if known, in the MCSS
(Monte Carlo Sample Size) function to estimate the required number of
posterior samples.
MCSE
is a univariate function that is often applied to each
marginal posterior distribution. A multivariate form is not
included. By chance alone due to multiple independent tests, 5% of
the parameters should indicate unacceptable MSCEs, even when
acceptable. Assessing convergence is difficult.
1 2 
x 
This is a vector of posterior samples for which MCSE or MCSS will be estimated. 
a 
This is a scalar argument of acceptable error for the mean of

method 
This is an optional argument for the method of MCSE
estimation, and defaults to Geyer's 
batch.size 
This is an optional argument that corresponds only
with 
warn 
Logical. If 
The default method for estimating MCSE is Geyer's Initial Monotone Positive Sequence (IMPS) estimator (Geyer, 1992), which takes the asymptotic variance into account and is timeseries based. This method goes by other names, such as Initial Positive Sequence (IPS).
The simplest method for estimating MCSE is to modify the formula for
standard error, sigma(x) /
sqrt(N), to account for nonindependence in the sequence
x of posterior samples. Nonindependence is
estimated with the ESS
function for Effective Sample Size (see
the ESS
function for more details), where m = ESS(x), and MCSE is
sigma(x) / sqrt(M). Although this
is the fastest and easiest method of estimation, it does not
incorporate an estimate of the asymptotic variance of
x.
The batch means method (Jones et al., 2006; Flegal et al., 2008) separates elements of x into batches and estimates MCSE as a function of multiple batches. This method is excellent, but is not recommended when the number of posterior samples is less than 1,000. These journal articles also assert that MCSE is a better stopping rule than MCMC convergence diagnostics.
The MCSS
function estimates the required number of posterior
samples, given the userspecified acceptable error, posterior samples
x
, and the observed variance (rather than asymptotic
variance). Due to the observed variance, this is a rough estimate.
Statisticat, LLC. software@bayesianinference.com
Flegal, J.M., Haran, M., and Jones, G.L. (2008). "Markov chain Monte Carlo: Can We Trust the Third Significant Figure?". Statistical Science, 23, p. 250–260.
Gelman, A., Carlin, J., Stern, H., and Rubin, D. (2004). "Bayesian Data Analysis, Texts in Statistical Science, 2nd ed.". Chapman and Hall, London.
Geyer, C.J. (1992). "Practical Markov Chain Monte Carlo". Statistical Science, 7, 4, p. 473–483.
Jones, G.L., Haran, M., Caffo, B.S., and Neath, R. (2006). "FixedWidth Output Analysis for Markov chain Monte Carlo". Journal of the American Statistical Association, 101(1), p. 1537–1547.
Consort
,
ESS
,
LaplacesDemon
, and
LaplacesDemon.hpc
.
1 2 3 4 5 6 
[1] 0.03110484
$est
[1] 0.04336594
$se
[1] 0.02947241
[1] 0.03064745
[1] 9393
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.