tutorial - Johannes Karreth

Using JAGS via R
Johannes Karreth
University at Albany, SUNY
jkarreth@albany.edu
ICPSR Summer Program 2015
All code used in this tutorial can also be found on my course website at http://www.jkarreth.
net/bayes-icpsr.html. Updated code will be posted there.
If you find any errors in this document or have suggestions for improvement, please don’t hesitate
to email me.
Contents
1
Using R as frontend
2
2
What are JAGS, R2jags, ...?
2
3
Installing JAGS and R2jags
2
4
Fitting Bayesian models using R2jags
4.1 Preparing the data and model . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 Fitting the model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3 Diagnostics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3
4
6
6
5
Using runjags
5.1 Output and diagnostics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15
17
6
Using JAGS via the command line
18
7
Using the coda package
20
8
MCMCpack
8.1 Output and diagnostics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
22
23
9
Following this course using JAGS
23
10 Other software solutions for Bayesian estimation
1
23
This tutorial focuses on using JAGS for fitting Bayesian models via R. There are other options
for fitting Bayesian models that we will briefly discuss during the workshop. You can find more
information about them at the end of this tutorial and on my workshop website.
1
Using R as frontend
A convenient way to fit Bayesian models using JAGS (or WinBUGS or OpenBUGS) is to use R
packages that function as frontends for JAGS. These packages make it easy to process the output of
Bayesian models and present it in publication-ready form. In this tutorial, I focus on the R2jags and
runjags packages. rjags is another package for the same purpose.
2
What are JAGS, R2jags, ...?
JAGS (Plummer, 2011) is Just Another Gibbs Sampler that was mainly written by Martyn Plummer
in order to provide a BUGS engine for Unix. More information can be found in the excellent JAGS
manual at http://sourceforge.net/projects/mcmc-jags/.
R2jags (Su and Yajima, 2012) is an R package that allows fitting JAGS models from within R. Almost all examples in Gelman and Hill’s Data Analysis Using Regression and Multilevel/Hierarchical
Models (2007) can be worked through equivalently in JAGS, using R2jags.
rjags (Plummer, 2013) is another R package that allows fitting JAGS models from within R. R2jags
depends on it. Simon Jackman’s Bayesian Analysis for the Social Sciences (2009) provides many
examples using rjags, and so does John Kruschke’s Doing Bayesian Data Analysis (2011).
runjags (Denwood, Under review) allows some additional functionalities, including parallel computing.
In this tutorial, I focus on the use of R2jags and runjags, as well as using JAGS directly from the
Terminal.
3
Installing JAGS and R2jags
1. Install the most recent R version from the CRAN website: http://cran.r-project.org/
bin/macosx or http://cran.r-project.org/bin/windows.
2. Mac users: Install the GNU Fortran (gfortran-4.2.3.dmg) library from the CRAN tools
directory: http://cran.r-project.org/bin/macosx/tools.
3. Install JAGS version 3.4.0 from Martyn Plummer’s repository: http://sourceforge.net/
projects/mcmc-jags/files/JAGS/3.x/.
4. Start the Terminal (Mac) or Console (Windows) and type
jags
If JAGS is installed, you will receive the following message:
Welcome
JAGS is
Loading
Loading
.
to JAGS 3.4.0 on Tue Dec 23 20:11:23 2014
free software and comes with ABSOLUTELY NO WARRANTY
module: basemod: ok
module: bugs: ok
You can quit the Terminal/Console again.
5. Install the packages R2jags, coda, R2WinBUGS, lattice, and (lastly) rjags from within R or
RStudio, via the Package Installer, or by using
> install.packages("R2jags", dependencies = TRUE, repos = "http://cran.us.r-project.org")
The dependencies = TRUE option will install the aforementioned packages automatically.
2
6. Install the package runjags from within R or RStudio, via the Package Installer, or by using
> install.packages("runjags", dependencies = TRUE, repos = "http://cran.us.r-project.org")
7. Install the package MCMCpack from within R or RStudio, via the Package Installer, or by
using
> install.packages("MCMCpack", dependencies = TRUE, repos = "http://cran.us.r-project.org")
8. Download a scientific text editor for writing R and JAGS code. I recommend TextWrangler for
Mac (http://www.barebones.com/products/textwrangler/), Sublime Text (http://
www.sublimetext.com) or Notepad++ for Windows (http://notepad-plus-plus.org).
RStudio is a very neat integrated environment for using R, but a separate text editor will be
useful from time to time to inspect JAGS model files and other files.
9. Note for users of Mac OS X 10.5 (Leopard): Due to a particular behavior of the JAGS installer
on Leopard, the JAGS files that rjags requires to run are not located where rjags is looking for
them. See http://martynplummer.wordpress.com/2011/11/04/rjags-3-for-mac-os-x/
#comments If you would like to use R2jags or rjags on Mac OS X 10.5, you need to manually
relocate these files from /usr to /usr/local. Ask me if you would like help with this.
10. You should now be able to run the following code in R, taken directly from the help file for
the jags function:
>
>
>
>
>
>
>
>
>
>
>
>
>
+
+
>
>
>
+
library(R2jags)
# An example model file is given in:
model.file <- system.file(package = "R2jags", "model", "schools.txt")
# data
J <- 8.0
y <- c(28.4,7.9,-2.8,6.8,-0.6,0.6,18.0,12.2)
sd <- c(14.9,10.2,16.3,11.0,9.4,11.4,10.4,17.6)
jags.data <- list("y","sd","J")
jags.params <- c("mu","sigma","theta")
jags.inits <- function(){
list("mu"=rnorm(1),"sigma"=runif(1),"theta"=rnorm(J))
}
# Fit the model
jagsfit <- jags(data=list("y","sd","J"), inits = jags.inits,
jags.params, n.iter = 10, model.file = model.file)
## Compiling model graph
##
Resolving undeclared variables
##
Allocating nodes
##
Graph Size: 41
##
## Initializing model
This should take less than a second and you should see the output above in your R console.
4
Fitting Bayesian models using R2jags
Just like R2WinBUGS,1 the purpose of R2jags is to allow fitting JAGS models from within R, and
to analyze convergence and perform other diagnostics right within R. A typical sequence of using
R2jags could look like this:
1 See Appendix C in Gelman and Hill (2007) or their online appendix http://www.stat.columbia.edu/ gelman/
~
bugsR/runningbugs.html for more info on how to use R2WinBUGS and R.
3
4.1
Preparing the data and model
For an example dataset, we simulate our own data in R. For this tutorial, we aim to fit a linear
model, so we create a continuous outcome variable y as a function of two predictors x1 and x2 and a
disturbance term e. We simulate a dataset with 100 observations.
• First, we create the predictors:
>
>
>
>
n.sim <- 100; set.seed(123)
x1 <- rnorm(n.sim, mean = 5, sd = 2)
x2 <- rbinom(n.sim, size = 1, prob = 0.3)
e <- rnorm(n.sim, mean = 0, sd = 1)
• Next, we create the outcome y based on coefficients b1 and b2 for the respective predictors
and an intercept a:
>
>
>
>
b1 <- 1.2
b2 <- -3.1
a <- 1.5
y <- a + b1 * x1 + b2 * x2 + e
• Now, we combine the variables into one dataframe for processing later:
> sim.dat <- data.frame(y, x1, x2)
• And we create and summarize a (frequentist) linear model fit on these data:
>
>
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
freq.mod <- lm(y ~ x1 + x2, data = sim.dat)
summary(freq.mod)
Call:
lm(formula = y ~ x1 + x2, data = sim.dat)
Residuals:
Min
1Q Median
-1.3432 -0.6797 -0.1112
3Q
0.5367
Max
3.2304
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 1.84949
0.28810
6.42 5.04e-09 ***
x1
1.13511
0.05158
22.00 < 2e-16 ***
x2
-3.09361
0.20650 -14.98 < 2e-16 ***
--Signif. codes:
0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Residual standard error: 0.9367 on 97 degrees of freedom
Multiple R-squared: 0.8772,Adjusted R-squared: 0.8747
F-statistic: 346.5 on 2 and 97 DF, p-value: < 2.2e-16
Now, we write a model for JAGS and save it as the text file "bayes.mod" in your working directory.
(Do not paste this model straight into R yet.) You can set your working directory via:
> setwd("~/Documents/Uni/9 - ICPSR/2015/Applied Bayes/Tutorials/2 - JAGS and R")
The model looks just like the JAGS models shown in throughout this course:
model {
for(i in 1:N){
y[i] ~ dnorm(mu[i], tau)
mu[i] <- alpha + beta1 * x1[i] + beta2 * x2[i]
}
alpha
beta1
beta2
tau ~
}
~ dnorm(0, .01)
~ dunif(-100, 100)
~ dunif(-100, 100)
dgamma(.01, .01)
4
Instead of saving the model in your WD, you can also enter it in your R script:
> bayes.mod <- function()
{
+
+
+
+
for(i in 1:N){
y[i] ~ dnorm(mu[i], tau)
mu[i] <- alpha + beta1 * x1[i] + beta2 * x2[i]
}
+
+
+
+
alpha
beta1
beta2
tau ~
~ dnorm(0, .01)
~ dunif(-100, 100)
~ dunif(-100, 100)
dgamma(.01, .01)
+ }
Now define the vectors of the data matrix for JAGS:
>
>
>
>
y <- sim.dat$y
x1 <- sim.dat$x1
x2 <- sim.dat$x2
N <- nrow(sim.dat)
Read in the data frame for JAGS
> sim.dat.jags
<- list("y", "x1", "x2", "N")
(You could also do this more conveniently using the as.list command on your data frame:)
> sim.dat.jags
<- as.list(sim.dat)
Note, though, that you will also need to specify any other variables not in the data, like in this case
N. So here, you would need to add:
> sim.dat.jags$N
<- nrow(sim.dat)
Define the parameters whose posterior distributions you are interested in summarizing later:
> bayes.mod.params <- c("alpha", "beta1", "beta2")
Now, we need to define the starting values for JAGS. Per Gelman and Hill (2007, 370), you can
use a function to do this. This function creates a list that contains one element for each parameter.
Each parameter then gets assigned a random draw from a normal distributio as a starting value. This
random draw is created using the rnorm function. The first argument of this function is the number
of draws. If your parameters are not indexed in the model code, this argument will be 1. If your
jags command below then specifies more than one chain, each chain will start at a different random
value for each parameter.
> bayes.mod.inits <- function(){
+
list("alpha" = rnorm(1), "beta1" = rnorm(1), "beta2" = rnorm(1))
+
}
Alternatively, if you want to have control over which starting values are chosen, you can provide
specific separate starting values for each chain:
>
>
>
>
inits1 <- list("alpha" = 0, "beta1" = 0, "beta2" = 0)
inits2 <- list("alpha" = 1, "beta1" = 1, "beta2" = 1)
inits3 <- list("alpha" = -1, "beta1" = -1, "beta2" = -1)
bayes.mod.inits <- list(inits1, inits2, inits3)
Before using R2jags the first time, you need to load the package, and you might need to set a random
number seed. To do this, type
> library(R2jags)
> set.seed(123)
directly in R or in your R script. You can choose any not too big number here. Setting a random seed
before fitting a model is also good practice for making your estimates replicable. We will discuss
replication in more detail in Weeks 3–4.
5
4.2
Fitting the model
Fit the model in JAGS:
> bayes.mod.fit <- jags(data = sim.dat.jags, inits = bayes.mod.inits,
+
parameters.to.save = bayes.mod.params, n.chains = 3, n.iter = 9000,
+
n.burnin = 1000,
+
model.file = "~/Documents/Uni/9 - ICPSR/2015/Applied Bayes/Tutorials/2 - JAGS and R/bayes.mod")
## Compiling model graph
##
Resolving undeclared variables
##
Allocating nodes
##
Graph Size: 511
##
## Initializing model
Note: If you use as your model file the function you gave directly to R above, then remove the
quotation marks:
> bayes.mod.fit <- jags(data = sim.dat.jags, inits = bayes.mod.inits,
+
parameters.to.save = bayes.mod.params, n.chains = 3, n.iter = 9000,
+
n.burnin = 1000, model.file = bayes.mod)
## Compiling model graph
##
Resolving undeclared variables
##
Allocating nodes
##
Graph Size: 511
##
## Initializing model
Update your model if necessary - e.g. if there is no/little convergence:
> bayes.mod.fit.upd <- update(bayes.mod.fit, n.iter=1000)
> bayes.mod.fit.upd <- autojags(bayes.mod.fit)
This function will auto-update until convergence.
4.3
Diagnostics
One of the neat things about running JAGS or BUGS from R is that this offers seamless, and therefore
quick, access to convergence diagnostics after fitting a model. See for yourself:
> print(bayes.mod.fit)
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
Inference for Bugs model at "/var/folders/8j/p9nqyxk5295bnlk0n50mm5ch0000gq/T//RtmpgkE7yZ/model470543d32d1.txt", fit using jags,
3 chains, each with 9000 iterations (first 1000 discarded), n.thin = 8
n.sims = 3000 iterations saved
mu.vect sd.vect
2.5%
25%
50%
75%
alpha
1.854
0.289
1.281
1.660
1.863
2.051
beta1
1.134
0.052
1.036
1.097
1.134
1.168
beta2
-3.092
0.210 -3.496 -3.239 -3.087 -2.943
deviance 271.766
2.871 268.240 269.647 271.079 273.257
97.5% Rhat n.eff
alpha
2.396 1.003
820
beta1
1.237 1.003
810
beta2
-2.692 1.001 3000
deviance 278.737 1.001 3000
For each parameter, n.eff is a crude measure of effective sample size,
and Rhat is the potential scale reduction factor (at convergence, Rhat=1).
DIC info (using the rule, pD = var(deviance)/2)
pD = 4.1 and DIC = 275.9
DIC is an estimate of expected predictive error (lower deviance is better).
> plot(bayes.mod.fit)
> traceplot(bayes.mod.fit)
6
Bugs model at "/var/folders/8j/p9nqyxk5295bnlk0n50mm5ch0000gq/T//RtmpgkE7yZ/model470543d32d1.txt", fit using jags, 3 chains, each with 9000 iterations (first 1000 discarded)
R−hat
80% interval for each chain
−4 −2 0
2
4
1 1.5 2+
medians and 80% intervals
alpha
beta1
beta2
2.5
−4
−2
0
2
4
1
1.5 2+
2
alpha
1.5
1
1.25
1.2
beta1
1.15
1.1
1.05
−2.8
−3
beta2
−3.2
−3.4
280
275
deviance
270
265
2.0
1.0
1.5
alpha
2.5
3.0
alpha
100
200
300
400
500
600
700
800
900 1000
600
700
800
900 1000
iteration
1.1
1.0
beta1
1.2
1.3
beta1
100
200
300
400
500
iteration
7
−3.0
−3.5
beta2
−2.5
beta2
100
200
300
400
500
600
700
800
900 1000
700
800
900 1000
iteration
280
270
275
deviance
285
deviance
100
200
300
400
500
600
iteration
If you want to print and save the plot, you can use the following set of commands:
• > pdf("~/Documents/Uni/9 - ICPSR/2015/Applied Bayes/Tutorials/2 - JAGS and R/bayes_trace.pdf")
... defines that the plot will be saved as a PDF file with the name "bayes trace.pdf" in
your working directory. 2
• > traceplot(bayes.mod.fit)
creates the plot in the background (you will not see it).
• > dev.off()
finishes the printing process and creates the PDF file of the plot. If successful, R will display
the message "null device 1".
More diagnostics are available when you convert your model output into an MCMC object. You can
generate an MCMC object for analysis with this command:
> bayes.mod.fit.mcmc <- as.mcmc(bayes.mod.fit)
> summary(bayes.mod.fit.mcmc)
2 LAT X cannot process file names with periods, so if you use LAT X and try to include the graphics file angell.trace,
E
E
LATEX will not compile your document.
8
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
Iterations = 1001:8993
Thinning interval = 8
Number of chains = 3
Sample size per chain = 1000
1. Empirical mean and standard deviation for each variable,
plus standard error of the mean:
Mean
alpha
1.854
beta1
1.134
beta2
-3.092
deviance 271.766
SD
0.28863
0.05185
0.20974
2.87084
Naive SE Time-series SE
0.0052697
0.008836
0.0009466
0.001616
0.0038294
0.003890
0.0524142
0.057127
2. Quantiles for each variable:
2.5%
25%
50%
75%
97.5%
alpha
1.281
1.660
1.863
2.051
2.396
beta1
1.036
1.097
1.134
1.168
1.237
beta2
-3.496 -3.239 -3.087 -2.943 -2.692
deviance 268.240 269.647 271.079 273.257 278.737
With an MCMC object, you can use a variety of commands for diagnostics and presentation using
the CODA package:
• Plot:
> library(lattice)
> xyplot(bayes.mod.fit.mcmc)
• You can customize the plot layout (you can use other Lattice options here as well):
> xyplot(bayes.mod.fit.mcmc, layout=c(2,2), aspect="fill")
• Density plot:
> densityplot(bayes.mod.fit.mcmc)
> densityplot(bayes.mod.fit.mcmc, layout=c(2,2), aspect="fill")
deviance
Density
0.0
0.00
0.5
0.05
1.0
0.10
1.5
0.15
beta2
−4.0
−3.5
−3.0
−2.5
265
270
275
280
285
290
beta1
0
0.0
2
0.5
4
1.0
6
8
alpha
1.0
1.5
2.0
2.5
3.0
0.9
1.0
1.1
1.2
1.3
• Trace- and density in one plot, print directly to your working directory:
> pdf("~/Documents/Uni/9 - ICPSR/2015/Applied Bayes/Tutorials/2 - JAGS and R/bayes_fit_mcmc_plot.pdf")
> plot(bayes.mod.fit.mcmc)
> dev.off()
• Autocorrelation plot, print directly to your working directory:
9
> pdf("~/Documents/Uni/9 - ICPSR/2015/Applied Bayes/Tutorials/2 - JAGS and R/bayes_fit_mcmc_autocorr.pdf")
> autocorr.plot(bayes.mod.fit.mcmc)
> dev.off()
• Other diagnostics using CODA:
> gelman.plot(bayes.mod.fit.mcmc)
> geweke.diag(bayes.mod.fit.mcmc)
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
[[1]]
Fraction in 1st window = 0.1
Fraction in 2nd window = 0.5
alpha
-1.65474
beta1
beta2 deviance
1.71754 -0.03899 -0.30234
[[2]]
Fraction in 1st window = 0.1
Fraction in 2nd window = 0.5
alpha
-0.3987
beta1
0.1764
beta2 deviance
0.8337 -1.0954
[[3]]
Fraction in 1st window = 0.1
Fraction in 2nd window = 0.5
alpha
1.477
beta1
-1.651
beta2 deviance
1.003
-2.182
> geweke.plot(bayes.mod.fit.mcmc)
> raftery.diag(bayes.mod.fit.mcmc)
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
[[1]]
Quantile (q) = 0.025
Accuracy (r) = +/- 0.005
Probability (s) = 0.95
You need a sample size of at least 3746 with these values of q, r and s
[[2]]
Quantile (q) = 0.025
Accuracy (r) = +/- 0.005
Probability (s) = 0.95
You need a sample size of at least 3746 with these values of q, r and s
[[3]]
Quantile (q) = 0.025
Accuracy (r) = +/- 0.005
Probability (s) = 0.95
You need a sample size of at least 3746 with these values of q, r and s
> heidel.diag(bayes.mod.fit.mcmc)
## [[1]]
##
##
##
## alpha
## beta1
Stationarity
test
passed
passed
start
p-value
iteration
1
0.185
1
0.139
10
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
beta2
passed
deviance passed
1
1
Halfwidth Mean
test
alpha
passed
1.87
beta1
passed
1.13
beta2
passed
-3.09
deviance passed
271.74
0.589
0.827
Halfwidth
0.02971
0.00527
0.01308
0.18831
[[2]]
Stationarity
test
alpha
passed
beta1
passed
beta2
passed
deviance passed
start
iteration
1
1
1
1
Halfwidth Mean
test
alpha
passed
1.86
beta1
passed
1.13
beta2
passed
-3.09
deviance passed
271.80
p-value
0.157
0.393
0.982
0.604
Halfwidth
0.0296
0.0055
0.0129
0.2042
[[3]]
Stationarity start
test
iteration
alpha
passed
1
beta1
passed
1
beta2
passed
1
deviance passed
101
Halfwidth Mean
test
alpha
passed
1.83
beta1
passed
1.14
beta2
passed
-3.10
deviance passed
271.81
p-value
0.4468
0.3556
0.0964
0.0839
Halfwidth
0.03066
0.00568
0.01360
0.20054
Quick convergence diagnostics: superdiag
A very convenient function to analyze numerical representations of diagnostics in one sweep is the
superdiag package (Tsai, Gill, and Rapkin, 2012).
• First, install the package:
> install.packages("superdiag", dependencies = TRUE, repos = "http://cran.us.r-project.org")
• Then, load it:
> library(superdiag)
• Next, all you need to do is:
> superdiag(bayes.mod.fit.mcmc, burnin = 100)
##
##
##
##
##
##
##
##
##
Number of chains = 3
Number of iterations = 1000 per chain before discarding the burn-in period
The burn-in period = 100 per chain
Sample size in total = 2700
********** The Geweke diagnostic: **********
Z-scores:
chain1
chain 2
chain 3
alpha
-0.32702322 -0.8681148 -1.804617
11
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
beta1
0.47904271 0.5804246 1.541857
beta2
-1.89980599 -0.2904349 1.577197
deviance
-0.05770752 0.4806812 -1.169611
Window From Start 0.10000000 0.4542000 0.719760
Window From Stop
0.50000000 0.0827600 0.226630
********** The Gelman-Rubin diagnostic: **********
Potential scale reduction factors:
alpha
beta1
beta2
deviance
Point est. Upper C.I.
1.01
1.03
1.01
1.04
1.00
1.00
1.00
1.00
Multivariate psrf
1.01
********** The Heidelberger-Welch diagnostic: **********
Chain 1, epsilon=0.1,
Stationarity
test
alpha
passed
beta1
passed
beta2
passed
deviance passed
alpha=0.05
start
p-value
iteration
1
0.558
1
0.441
1
0.491
1
0.477
Halfwidth Mean
test
alpha
passed
1.87
beta1
passed
1.13
beta2
passed
-3.09
deviance passed
271.76
Halfwidth
0.03087
0.00553
0.01380
0.18102
Chain 2, epsilon=0.062, alpha=0.1
Stationarity start
p-value
test
iteration
alpha
passed
1
0.178
beta1
passed
1
0.300
beta2
passed
1
0.907
deviance passed
1
0.752
Halfwidth Mean
test
alpha
passed
1.86
beta1
passed
1.13
beta2
passed
-3.09
deviance passed
271.84
Halfwidth
0.03183
0.00595
0.01363
0.21074
Chain 3, epsilon=0.111, alpha=0.05
Stationarity start
p-value
test
iteration
alpha
passed
1
0.219
beta1
passed
1
0.377
beta2
passed
1
0.122
deviance passed
1
0.145
Halfwidth Mean
test
alpha
passed
1.83
beta1
passed
1.14
beta2
passed
-3.10
deviance passed
271.81
Halfwidth
0.0323
0.0060
0.0137
0.2005
********** The Raftery-Lewis diagnostic: **********
Chain 1, converge.eps = 0.001
Quantile (q) = 0.025
Accuracy (r) = +/- 0.005
Probability (s) = 0.95
12
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
You need a sample size of at least 3746 with these values of q, r and s
Chain 2, converge.eps = 5e-04
Quantile (q) = 0.001
Accuracy (r) = +/- 0.005
Probability (s) = 0.95
Burn-in
(M)
alpha
2
beta1
2
beta2
2
deviance 2
Total
(N)
173
173
173
173
Lower bound
(Nmin)
154
154
154
154
Dependence
factor (I)
1.12
1.12
1.12
1.12
Chain 3, converge.eps = 0.0025
Quantile (q) = 0.25
Accuracy (r) = +/- 0.001
Probability (s) = 0.99
You need a sample size of at least 1244044 with these values of q, r and s
Note that the R2jags object by default retains 1000 iterations (through thinning), hence the
burn-in period you provide for superdiag must be less than 1000.
Quick diagnostic plots: mcmcplots
A convenient way to obtain graphical diagnostics and results is using the mcmcplots package (Curtis,
2012):
• First, install the package:
> install.packages("mcmcplots", dependencies = TRUE, repos = "http://cran.us.r-project.org")
• Then, load it:
> library(mcmcplots)
• Some commands for plots:
> denplot(bayes.mod.fit.mcmc)
> denplot(bayes.mod.fit.mcmc, parms = c("alpha", "beta1", "beta2"))
> traplot(bayes.mod.fit.mcmc, parms = c("alpha", "beta1", "beta2"))
0.0
0.4
0.8
1.2
alpha
1.0
1.5
2.0
2.5
3.0
0
2
4
6
8
beta1
0.9
1.0
1.1
1.2
1.3
0.0
0.5
1.0
1.5
beta2
−4.0
−3.5
−3.0
13
−2.5
1.0
1.5
2.0
2.5
3.0
alpha
2000
4000
6000
8000
6000
8000
6000
8000
1.0
1.1
1.2
1.3
beta1
2000
4000
−3.5
−3.0
−2.5
beta2
2000
4000
As always, check the help files for options to customize these plots.
• Or, for quick diagnostics, you can produce html files with trace, density, and autocorrelation
plots all on one page. The files will be displayed in your default internet browser.
> mcmcplot(bayes.mod.fit.mcmc)
• If you want to produce a coefficient dot plot with credible intervals, use caterplot:
> caterplot(bayes.mod.fit.mcmc)
> caterplot(bayes.mod.fit.mcmc, parms = c("alpha", "beta1", "beta2"),
+
labels = c("alpha", "beta1", "beta2"))
alpha
beta1
beta2
−3
−2
−1
0
1
2
More plots: ggmcmc
Yet another option for plotting output is the ggmcmc package (Fernández i Marı́n, 2013):
• First, install the package:
> install.packages("ggmcmc", dependencies = TRUE, repos = "http://cran.us.r-project.org")
• Then, load it:
> library(ggmcmc)
• To use this package, you need to first convert your MCMC object into a ggs object, using the
ggs function. This object can then be passed on to the various ggmcmc plotting commands:
14
> bayes.mod.fit.gg <- ggs(bayes.mod.fit.mcmc)
> ggs_density(bayes.mod.fit.gg)
alpha
1.0
0.5
0.0
1.0
1.5
2.0
beta1
8
2.5
3.0
6
4
density
2
Chain
0
1.0
1.1
beta2
1.2
1.3
1
2
3
1.5
1.0
0.5
0.0
−3.6
−3.2
deviance
−2.8
−2.4
0.15
0.10
0.05
0.00
270
275
280
285
value
• Or, you can use the ggmcmc command to create a PDF file containing a variety of diagnostic
plots:
> ggmcmc(bayes.mod.fit.gg,
+
file = "~/Documents/Uni/9 - ICPSR/2015/Applied Bayes/Tutorials/2 - JAGS and R/bayes_fit_ggmcmc.pdf")
##
##
##
##
##
##
##
##
##
##
##
##
5
Plotting
Plotting
Plotting
Plotting
Plotting
Plotting
Plotting
Plotting
Plotting
Plotting
pdf
2
histograms
density plots
traceplots
running means
comparison of partial and full chain
autocorrelation plots
crosscorrelation plot
Potential Scale Reduction Factors
Geweke Diagnostic
caterpillar plot
Using runjags
Instead of R2jags, you can also use the runjags package to command JAGS via R. Here is an abbreviated version of the workflow for runjags. The workflow mimics what you saw for R2jags above:
> library(runjags)
For an example dataset, we simulate our own data in R. For this tutorial, we aim to fit a linear
model, so we create a continuous outcome variable y as a function of two predictors x1 and x2 and a
disturbance term e. We simulate 100 observations.
• First, we create the predictors:
>
>
>
>
n.sim <- 100; set.seed(123)
x1 <- rnorm(n.sim, mean = 5, sd = 2)
x2 <- rbinom(n.sim, size = 1, prob = 0.3)
e <- rnorm(n.sim, mean = 0, sd = 1)
• Next, we create the outcome y based on coefficients b1 and b2 for the respective predictors
and an intercept a:
>
>
>
>
b1 <- 1.2
b2 <- -3.1
a <- 1.5
y <- a + b1 * x1 + b2 * x2 + e
15
• Now, we combine the variables into one dataframe for processing later:
> sim.dat <- data.frame(y, x1, x2)
• And we summarize a (frequentist) linear model fit on these data:
>
>
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
freq.mod <- lm(y ~ x1 + x2, data = sim.dat)
summary(freq.mod)
Call:
lm(formula = y ~ x1 + x2, data = sim.dat)
Residuals:
Min
1Q Median
-1.3432 -0.6797 -0.1112
3Q
0.5367
Max
3.2304
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 1.84949
0.28810
6.42 5.04e-09 ***
x1
1.13511
0.05158
22.00 < 2e-16 ***
x2
-3.09361
0.20650 -14.98 < 2e-16 ***
--Signif. codes:
0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Residual standard error: 0.9367 on 97 degrees of freedom
Multiple R-squared: 0.8772,Adjusted R-squared: 0.8747
F-statistic: 346.5 on 2 and 97 DF, p-value: < 2.2e-16
Create the model object:
> bayes.mod <- "model{
for(i in 1:N){
y[i] ~ dnorm(mu[i], tau)
mu[i] <- alpha + beta1 * x1[i] + beta2 * x2[i]
}
alpha
beta1
beta2
tau ~
~ dnorm(0, .01)
~ dunif(-100, 100)
~ dunif(-100, 100)
dgamma(.01, .01)
}"
Convert the data frame to a list:
> sim.list
<- as.list(sim.dat)
and add the number of observations:
> sim.list$N
<- nrow(sim.dat)
Convert the data for runjags:
> sim.dat.runjags <- dump.format(sim.list)
Specify initial values:
> inits1 <- list(alpha = 1, beta1 = 1, beta2 = 1)
> inits2 <- list(alpha = 0, beta1 = 0, beta2 = 0)
> inits3 <- list(alpha = -1, beta1 = -1, beta2 = -1)
Fit the model in using run.jags:
> bayes.mod.fit2 <- run.jags(model = bayes.mod, monitor = c("alpha", "beta1", "beta2"),
+
data = sim.dat.runjags, n.chains = 3, inits = list(inits1, inits2, inits3),
+
burnin = 1000, sample = 5000, keep.jags.files = TRUE)
16
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
Calling the simulation...
Welcome to JAGS 3.4.0 on Wed May 27 19:07:26 2015
JAGS is free software and comes with ABSOLUTELY NO WARRANTY
Loading module: basemod: ok
Loading module: bugs: ok
. . Reading data file data.txt
. Compiling model graph
Resolving undeclared variables
Allocating nodes
Graph Size: 511
. Reading parameter file inits1.txt
. Reading parameter file inits2.txt
. Reading parameter file inits3.txt
. Initializing model
. Adapting 1000
-------------------------------------------------| 1000
++++++++++++++++++++++++++++++++++++++++++++++++++ 100%
Adaptation successful
. Updating 1000
-------------------------------------------------| 1000
************************************************** 100%
. . . . Updating 5000
-------------------------------------------------| 5000
************************************************** 100%
. . . . . . Updating 0
. Deleting model
.
Simulation complete. Reading coda files...
Coda files loaded successfully
Calculating summary statistics...
Calculating the Gelman-Rubin statistic for 3
variables....
Finished running the simulation
JAGS files were saved to the 'runjagsfiles_2' folder
in your current working directory
Note a few things:
• The model is read into R within quotation marks.
• By setting keep.jags.files = TRUE, we get run.jags to create a folder “runjagsfiles” in
your WD. This folder will contain the same files you obtain from JAGS in the next step below.
This can be useful for further processing.
• The option method allows for parallel chains on separate cores of your computer and other
higher-end computing options. See the help file for run.jags for more information.
5.1
Output and diagnostics
runjags also allows you to use the same set of commands for diagnostics and accessing the fitted
objects as you explored above.
• Summarize the model object
> print(bayes.mod.fit)
##
##
##
##
##
##
##
##
##
##
##
##
##
##
Inference for Bugs model at "/var/folders/8j/p9nqyxk5295bnlk0n50mm5ch0000gq/T//RtmpgkE7yZ/model470543d32d1.txt", fit using
3 chains, each with 9000 iterations (first 1000 discarded), n.thin = 8
n.sims = 3000 iterations saved
mu.vect sd.vect
2.5%
25%
50%
75%
alpha
1.854
0.289
1.281
1.660
1.863
2.051
beta1
1.134
0.052
1.036
1.097
1.134
1.168
beta2
-3.092
0.210 -3.496 -3.239 -3.087 -2.943
deviance 271.766
2.871 268.240 269.647 271.079 273.257
97.5% Rhat n.eff
alpha
2.396 1.003
820
beta1
1.237 1.003
810
beta2
-2.692 1.001 3000
deviance 278.737 1.001 3000
17
##
##
##
##
##
##
For each parameter, n.eff is a crude measure of effective sample size,
and Rhat is the potential scale reduction factor (at convergence, Rhat=1).
DIC info (using the rule, pD = var(deviance)/2)
pD = 4.1 and DIC = 275.9
DIC is an estimate of expected predictive error (lower deviance is better).
• Convert to an MCMC object. Here, we use as.mcmc.list to keep the 3 chains separate in
the resulting MCMC object:
> bayes.mod.fit2.mcmc <- as.mcmc.list(bayes.mod.fit2)
> summary(bayes.mod.fit2.mcmc)
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
Iterations = 2001:7000
Thinning interval = 1
Number of chains = 3
Sample size per chain = 5000
1. Empirical mean and standard deviation for each variable,
plus standard error of the mean:
Mean
SD Naive SE Time-series SE
alpha 1.829 0.2994 0.0024447
0.012116
beta1 1.138 0.0536 0.0004377
0.002189
beta2 -3.092 0.2075 0.0016942
0.002772
2. Quantiles for each variable:
2.5%
25%
50%
75% 97.5%
alpha 1.242 1.635 1.830 2.022 2.420
beta1 1.032 1.104 1.139 1.173 1.244
beta2 -3.496 -3.233 -3.090 -2.950 -2.687
• Use mcmcplot and superdiag for diagnostics:
> mcmcplot(bayes.mod.fit2.mcmc)
> superdiag(bayes.mod.fit2.mcmc, burnin = 1000)
6
Using JAGS via the command line
JAGS can also be operated straight from the command line—on Windows and Unix systems alike.
This can be useful if you don’t want to have R busy fitting models that take a longer time. The most
feasible way to do this is to write a script file with the following parts, and save it, for instance as
bayes.jags. Before running this script, you will need to create some files (see below).
model clear
data clear
load dic
model in "bayes.mod"
data in "sim.dat"
compile, nchains(3)
inits in "bayes.mod.inits1", chain(1)
inits in "bayes.mod.inits2", chain(2)
inits in "bayes.mod.inits2", chain(3)
initialize
update 2500, by(100)
monitor alpha, thin(2)
monitor beta1, thin(2)
monitor beta2, thin(2)
monitor deviance, thin(2)
update 2500, by(100)
coda *
The following instructions are for the Terminal on the Mac, but you can reproduce the same steps on
Windows or Linux. You run this script file by opening a Terminal window, changing to the working
directory in which all the above files are located 18
cd "~/Documents/Uni/9 - ICPSR/2015/Applied Bayes/Tutorials/2 - JAGS and R"
and then simply telling JAGS to run the script:
jags bayes.jags
In your Terminal, you will see something like this:
In more detail, this is what each line of the script does:
• model clear
data clear
load dic
Remove previous data and models (if applicable), load the dic module so you can monitor
the model deviance later.
• model in "bayes.mod"
Use the model bayes.mod, which is saved in your WD, and looks like a regular JAGS model.
Make sure you use the exact and full name of the model file as it is in your working directory,
otherwise JAGS will not find it. Look out for hidden file name extensions. 3 You wrote this
model earlier and saved it in your WD as bayes.mod:
model {
for(i in 1:N){
y[i] ~ dnorm(mu[i], tau)
mu[i] <- alpha + beta1 * x1[i] + beta2 * x2[i]
}
alpha
beta1
beta2
tau ~
}
~ dnorm(0, .01)
~ dunif(-100, 100)
~ dunif(-100, 100)
dgamma(.01, .01)
• data in "sim.dat"
Use the data sim.dat. These data can be saved as vectors in one file that you name sim.dat.
This file will look something like this:
"y" <- c(6.94259729574987, 4.61661626624104, ...
"x1" <- c(3.87904870689558, 4.53964502103344, ...
"x2" <- c(0, 1, ...
"N" <- 100
You can create this data file by hand (inconvenient), or you can work with some R commands
to do this automatically. If sim.dat is a data frame object in R, you can do the following:
3 On the Mac, you can set the Finder to display all file extensions by going to Finder > Preferences > check “Show all
filename extensions.”
19
>
>
>
>
sim.dat.list <- as.list(sim.dat)
sim.dat.list$N <- nrow(sim.dat)
dump("sim.dat.list", file = "sim.dat.dump")
bugs2jags("sim.dat.dump", "sim.dat")
The first command converts the data frame into a list; the second command writes out the
data in BUGS format as a file to your working directory; the third command (part of the coda
package) translates it into JAGS format. Both angell.dump and angell.dat are written to your
working directory, so be sure that you have specified that in the beginning. Also be sure to
visually inspect whether your data file was created correctly.
• compile, nchains(2)
Compile the models and run two Markov chains.
• inits in "bayes.mod.inits1", chain(1)
inits in "bayes.mod.inits2", chain(2)
inits in "bayes.mod.inits3", chain(3)
Use the starting values you provide in bayes.mod.inits1, bayes.mod.inits2, and bayes.mod.inits3,
each of which could look like this:
"alpha" <- c(0)
"beta1" <- c(0)
"beta2" <- c(0)
This way, you can specify different starting values for each chain.
• initialize
Initialize and run the model.
• update 2500, by(100)
Specify 2500 updates before you start monitoring your parameters of interest.
• monitor alpha, thin(2)
monitor beta1, thin(2)
monitor beta2, thin(2)
monitor deviance, thin(2)
Monitor the values for these parameters, in this case the three regression coefficients and the
model deviance. thin(2) specifies that only each second draw from the chains will be used
for your model output.
• update 2500, by(100)
• coda *, stem(bayes_out)
Tell JAGS to produce coda files that all begin with the stem bayes out and are put in your
WD. These can then be analyzed with the tools described in this tutorial.
7
Using the coda package
You can use the coda package in R to analyze your JAGS output. These are the key steps to get your
data into R to use CODA:
• Fit your model in JAGS, and identify where your software saved the chains and index files –
most likely in the working directory where the other components of your model (data, model,
inits) are.
• In R, load the coda package:
> library(coda)
• Read in your JAGS output. This requires that the chains and index files (see above) are in your
working directory.
20
> setwd("~/Documents/Uni/9 - ICPSR/2015/Applied Bayes/Tutorials/2 - JAGS and R/")
> chain1 <- read.coda(output.file = "CODAchain1.txt",
+
index.file = "CODAindex.txt")
##
##
##
##
Abstracting
Abstracting
Abstracting
Abstracting
alpha ... 1250 valid values
beta1 ... 1250 valid values
beta2 ... 1250 valid values
deviance ... 1250 valid values
> chain2 <- read.coda(output.file = "CODAchain2.txt",
+
index.file = "CODAindex.txt")
##
##
##
##
Abstracting
Abstracting
Abstracting
Abstracting
alpha ... 1250 valid values
beta1 ... 1250 valid values
beta2 ... 1250 valid values
deviance ... 1250 valid values
> chain3 <- read.coda(output.file = "CODAchain3.txt",
+
index.file = "CODAindex.txt")
##
##
##
##
Abstracting
Abstracting
Abstracting
Abstracting
alpha ... 1250 valid values
beta1 ... 1250 valid values
beta2 ... 1250 valid values
deviance ... 1250 valid values
> bayes.chains <- as.mcmc.list(list(chain1, chain2, chain3))
• Now you can analyze using some of the commands listed in
> help(package = "coda")
for instance:
> summary(bayes.chains)
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
Iterations = 2501:4999
Thinning interval = 2
Number of chains = 3
Sample size per chain = 1250
1. Empirical mean and standard deviation for each variable,
plus standard error of the mean:
Mean
alpha
1.803
beta1
1.143
beta2
-3.088
deviance 271.768
SD
0.28359
0.05142
0.20954
2.93772
Naive SE Time-series SE
0.0046310
0.014504
0.0008397
0.002786
0.0034218
0.004091
0.0479728
0.071076
2. Quantiles for each variable:
2.5%
25%
50%
75%
97.5%
alpha
1.238
1.614
1.808
1.992
2.355
beta1
1.041
1.109
1.143
1.179
1.244
beta2
-3.496 -3.228 -3.089 -2.949 -2.678
deviance 268.138 269.671 271.082 273.063 279.708
> traceplot(bayes.chains)
You should be able to play around with graphical parameters and different printing devices
this way.
• Convenient: the superdiag, mcmcplots, and ggmcmc packages also work like described above
once you have declared your JAGS output an MCMC object.
21
8
MCMCpack
MCMCpack (Martin, Quinn, and Park, 2011) is an R package that we will also use in this course. It is
as easy to use as the lm() command in R and produces the same MCMC output you analyzed above.
One potential downside of MCMCpack is that it does not offer as much options for customization
as writing your full model in JAGS. For more information on MCMCpack, see http://mcmcpack.
wustl.edu/.
Because MCMCpack uses an R-like formula, it is straightforward to fit a Bayesian linear model.
> library(MCMCpack)
For an example dataset, we again simulate our own data in R. We create a continuous outcome
variable y as a function of two predictors x1 and x2 and a disturbance term e. We simulate 100
observations.
• First, we create the predictors:
>
>
>
>
n.sim <- 100; set.seed(123)
x1 <- rnorm(n.sim, mean = 5, sd = 2)
x2 <- rbinom(n.sim, size = 1, prob = 0.3)
e <- rnorm(n.sim, mean = 0, sd = 1)
• Next, we create the outcome y based on coefficients b1 and b2 for the respective predictors
and an intercept a:
>
>
>
>
b1 <- 1.2
b2 <- -3.1
a <- 1.5
y <- a + b1 * x1 + b2 * x2 + e
• Now, we combine the variables into one dataframe for processing later:
> sim.dat <- data.frame(y, x1, x2)
• And we summarize a (frequentist) linear model fit on these data:
>
>
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
##
freq.mod <- lm(y ~ x1 + x2, data = sim.dat)
summary(freq.mod)
Call:
lm(formula = y ~ x1 + x2, data = sim.dat)
Residuals:
Min
1Q Median
-1.3432 -0.6797 -0.1112
3Q
0.5367
Max
3.2304
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 1.84949
0.28810
6.42 5.04e-09 ***
x1
1.13511
0.05158
22.00 < 2e-16 ***
x2
-3.09361
0.20650 -14.98 < 2e-16 ***
--Signif. codes:
0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
Residual standard error: 0.9367 on 97 degrees of freedom
Multiple R-squared: 0.8772,Adjusted R-squared: 0.8747
F-statistic: 346.5 on 2 and 97 DF, p-value: < 2.2e-16
MCMCpack only requires us to specify one single model object. For the linear model, we use
MCMCregress. For other models available in MCMCpack, see help(package = "MCMCpack").
22
> bayes.mod.fit3 <- MCMCregress(y ~ x1 + x2, data = sim.dat, burnin = 1000,
+
mcmc = 5000, seed = 123, beta.start = c(0, 0, 0),
+
b0 = c(0, 0, 0), B0 = c(0.1, 0.1, 0.1))
> summary(bayes.mod.fit)
##
##
##
##
##
##
##
Length
model
8
BUGSoutput
24
parameters.to.save 4
model.file
1
n.iter
1
DIC
1
Class
jags
bugs
-none-none-none-none-
Mode
list
list
character
character
numeric
logical
The resulting object is already an MCMC object.
8.1
Output and diagnostics
MCMCpack also allows you to use the same set of commands for diagnostics and accessing the
fitted objects as you explored above. For example, you may use mcmcplot and superdiag for
diagnostics:
> mcmcplot(bayes.mod.fit3)
> superdiag(bayes.mod.fit3, burnin = 1000)
9
Following this course using JAGS
You can find modified (if necessary) JAGS code for all models presented in this course on my website at http://www.jkarreth.net/bayes-icpsr.html.
The most recent version of this file is posted at http://www.jkarreth.net/files/Lab2_JAGS.
pdf.
Many JAGS-related questions are answerd at http://stackoverflow.com/questions/tagged/
jags and the discussion board at http://sourceforge.net/projects/mcmc-jags/forums/
forum/610037.
10
Other software solutions for Bayesian estimation
Other options to fit Bayesian models include:
• WinBUGS/OpenBUGS. These programs have a GUI, but can also be accessed from R. Gelman and Hill (2007) provide extensive information on how to integrate BUGS and R. OpenBUGS (http://www.openbugs.net/w/FrontPage) is the open-source continuation of WinBUGS (http://www.mrc-bsu.cam.ac.uk/software/bugs/the-bugs-project-winbugs/)
with similar functionality. A convenient way to fit Bayesian models using WinBUGS or
OpenBUGS is to use R packages that function as frontends for WinBUGS or OpenBUGS,
such as R2WinBUGS (Sturtz, Ligges, and Gelman, 2005) or BRugs (Thomas et al., 2006).
R2WinBUGS works very similar to R2jags, so you can adapt the code in this tutorial easily.
Mac users can find setup instructions on my course website.
• Stan (http://mc-stan.org) is a new and fast program that can be accessed via R (and other
statistical packages). The documentation on the Stan website is very easy to follow, and offers
tutorials on fitting some example models in Stan. We will show you how to use Stan in Week
4 of this course.
23
References
Curtis, S. McKay. 2012. mcmcplots: Create Plots from MCMC Output. R package version 0.4.1.
URL: http://CRAN.R-project.org/package=mcmcplots
Denwood, Matthew J. Under review. “runjags: An R package providing interface utilities, parallel computing methods and additional distributions for MCMC models in JAGS.” Manuscript,
University of Copenhagen .
Fernández i Marı́n, Xavier. 2013. ggmcmc: Graphical tools for analyzing Markov Chain Monte
Carlo simulations from Bayesian inference. R package version 0.5.1.
URL: http://xavier-fim.net/packages/ggmcmc
Gelman, Andrew, and Jennifer Hill. 2007.
Data Analysis Using Regression and Multilevel/Hierarchical Models. New York, NY: Cambridge University Press.
Jackman, Simon. 2009. Bayesian Analysis for the Social Sciences. Chichester: Wiley.
Kruschke, John. 2011. Doing Bayesian Data Analysis: A Tutorial Introduction with R. Oxford:
Academic Press / Elsevier.
Martin, Andrew D., Kevin M. Quinn, and Jong Hee Park. 2011. “MCMCpack: Markov Chain
Monte Carlo in R.” Journal of Statistical Software 42 (9): 22.
Plummer, Martyn. 2011. “JAGS Version 3.1.0 User Manual.”.
Plummer, Martyn. 2013. rjags: Bayesian graphical models using MCMC. R package version 3-10.
URL: http://CRAN.R-project.org/package=rjags
Sturtz, Sibylle, Uwe Ligges, and Andrew Gelman. 2005. “R2WinBUGS: A Package for Running
WinBUGS from R.” Journal of Statistical Software 12 (3): 1–16.
Su, Yu-Sung, and Masanao Yajima. 2012. R2jags: A Package for Running jags from R. R package
version 0.03-08.
URL: http://CRAN.R-project.org/package=R2jags
Thomas, Andrew, Bob O’Hara, Uwe Ligges, and Sibylle Sturtz. 2006. “Making BUGS Open.” R
News 6 (1): 12–17.
Tsai, Tsung-han, Jeff Gill, and Jonathan Rapkin. 2012. superdiag: R Code for Testing Markov Chain
Nonconvergence. R package version 1.1.
URL: http://CRAN.R-project.org/package=superdiag
24