Rev Language Reference

srGelmanRubin(RealPos R, String filename, IntegerPos frequency, String burninMethod, Probability burnin)

R :	RealPos (pass by value)
	The maximum allowed potential scale reduction factor.
filename :	String (pass by value)
	The name of the file containing the samples.
frequency :	IntegerPos (pass by value)
	The frequency how often to check for convergence.
	Default : 10000
burninMethod :	String (pass by value)
	Which type of burnin method to use.
	Default : ESS
	Options : ESS|SEM|fixed
burnin :	Probability (pass by value)
	The fraction of samples to discard as burnin (only used when burninMethod is "fixed").
	Default : 0.25

The Gelman-Rubin statistic, also referred to as the potential scale reduction factor (PSRF), compares the variance of the sample pooled from multiple runs to the sum of variances calculated from individual runs. Accordingly, it can only be calculated when two or more independent runs are performed, and its value tends to unity (1) as the runs converge. The number of samples to be removed as burnin before calculating the test statistic is determined using the `burninMethod`. If the `"ESS"` (default) or `"SEM"` options are chosen, different burnin lengths are tested, increasing from 0 to 50% (for `"ESS"`) or 100% (for `"SEM"`) of the length of the trace in increments of 10 samples. The `"ESS"` option calculates effective sample sizes (ESS) for all monitored parameters after removing the number of samples corresponding to each candidate burnin length. The best burnin length for a given parameter is the one that maximizes its ESS value. The `"SEM"` option instead calculates the standard error of the mean (SEM), and the best burnin length for a given parameter is the one that minimizes its SEM value. In both cases, the final burnin length is set to the maximum of the parameter-specific burnin lengths. Alternatively, the user may set `burninMethod` to `"fixed"`, which discards a constant fraction of the samples collected up to that point. This fraction can be specified using the `burnin` argument, and is set to 0.25 by default. The `burnin` argument has no effect if the `"ESS"` or `"SEM"` options are chosen, and a corresponding warning is displayed if the user explicitly sets the argument without specifying `burninMethod="fixed"`. The `"fixed"` option is appropriate for analyses with very long parameter traces and large numbers of monitored variables, for which the automatic burnin determination may be too computationally demanding. See also the tutorial on [convergence assessment](https://revbayes.github.io/tutorials/convergence/).

# Binomial example: estimate success probability given 7 successes out of 20 trials
r ~ dnExp(10)
p := Probability(ifelse(r < 1, r, 1))
n <- 20
k ~ dnBinomial(n, p)
k.clamp(7)
mymodel = model(k)

moves = VectorMoves()
moves.append( mvSlide(r, delta=0.1, weight=1) )

paramFile = "parameters.log"

monitors = VectorMonitors()
monitors.append( mnModel(filename=paramFile, printgen=100, p) )

# Stop when the potential scale reduction factor falls below 1.01
stopping_rules[1] = srGelmanRubin(1.01, file = paramFile, freq = 1000)

# Create the MCMC object.
# Set nruns = 2 to ensure the Gelman-Rubin statistic is applicable
mymcmc = mcmc(mymodel, monitors, moves, nruns = 2)

# Begin the MCMC run
mymcmc.run(rules = stopping_rules)

• mcmc
• srGeweke
• srMinESS
• srStationarity