ggdmc is an R package for modelling cognitive processes. Although its focus is on the challenging hierarchical Bayesian models, fitting them with Bayesian MCMC. It can also fit cognitive models with conventional methods, such as maximum likelihood estimation and least squares. The package uses the sampling method of population-based Markov chain Monte Carlo (pMCMC).
This example demonstrates the Wiener diffusion model. For other models, see my tutorials site. The naming of R functions in ggdmc attempts to inform the user what the functions are for. For example, BuildModel is to build a model object.
As the user is often reminded in using Bayesian tools, it is always a good practice to check the result of a model fit. Note that the sequence of parameters in a parameter vector (i.e., p.vector) must follow the sequence in the p.vector reported by BuildModel. Some built-in checks will try to safeguard this, but they are far from bulletproof.
## Set up model ----
## Fixing sv & sz to 0 to set up a Wiener diffusion model
require(ggdmc)
model <- BuildModel(
p.map = list(a = "1", v="1", z="1", d="1", sz="1", sv="1", t0="1",
st0="1"),
match.map = list(M = list(s1 = "r1", s2 = "r2")),
factors = list(S = c("s1", "s2")),
responses = c("r1","r2"),
constants = c(st0 = 0, d = 0, sv = 0, sz = 0),
type = "rd")
npar <- model@npar ## Note this works for version > 0.2.7.5;
## npar <- length(GetPNames(model)) ## Use GetPNames instead in 0.2.6.0
p.vector <- c(a=1, v=1.5, z=0.5, t0=.15)
dat <- simulate(model, nsim = 50, ps = p.vector)
dmi <- BuildDMI(dat, model)
p.prior <- BuildPrior(
dists = rep("tnorm", npar),
p1=c(a=1, v=0, z=1, t0=1),
p2=c(a=1, v=2, z=1, t0=1),
lower = c(0, -5, rep(0, 2)),
upper = rep(NA, npar))
## Fit model -------------
fit0 <- StartNewsamples(dmi, p.prior)
fit <- run(fit0)
## Check model -----------
plot(fit)
plot(fit, den = TRUE)
plot(fit, pll = FALSE)
plot(fit, pll = FALSE, den = TRUE)
isconv <- gelman(fit)
est <- summary(fit, recovery = TRUE, ps = p.vector, verbose = TRUE)
require(ggdmc);
model <- BuildModel(
p.map = list(a = "1", v ="1", z ="1", d ="1", sz ="1", sv ="1", t0 ="1",
st0 ="1"),
match.map = list(M = list(s1 = "r1", s2 = "r2")),
factors = list(S = c("s1", "s2")),
responses = c("r1","r2"),
constants = c(st0 = 0, d = 0, sv = 0, sz = 0),
type = "rd")
npar <- model@npar
pop.mean <- c(a = 2, v = 4, z = 0.5, t0 = 0.3)
pop.scale <- c(a = 0.5, v = .5, z = 0.1, t0 = 0.05)
pop.prior <- BuildPrior(
dists = rep("tnorm", npar),
p1 = pop.mean,
p2 = pop.scale,
lower = c(0,-5, 0, 0),
upper = c(5, 7, 1, 1))
## Simulate some data
dat <- simulate(model, nsub = 50, nsim = 30, prior = pop.prior)
dmi <- BuildDMI(dat, model)
ps <- attr(dat, "parameters")
p.prior <- BuildPrior(
dists = rep("tnorm", npar),
p1 = pop.mean,
p2 = pop.scale*5,
lower = c(0,-5, 0, 0),
upper = c(5, 7, 1, 1))
plot(p.prior, ps = ps) ## Check if all true values are in the range
## Sampling separately
fit0 <- StartNewsamples(dmi, p.prior, ncore = 4)
fit <- run(fit0, 5e2, ncore = 4)
fit <- run(fit, 1e2, add = TRUE, ncore = 4) ## add additional 100 samples
## Check model -----
isconv <- gelman(fit, verbose = TRUE)
plot(fit)
est0 <- summary(fit, recovery = TRUE, ps = ps, verbose = TRUE)
## Sampling hierarchically
mu.prior <- BuildPrior(
dists = rep("tnorm", npar),
p1 = pop.mean,
p2 = pop.scale*5,
lower = c(0,-5, 0, 0),
upper = c(5, 7, 1, 1))
sigma.prior <- BuildPrior(
dists = rep("beta", npar),
p1 = c(a=1, v=1, z=1, t0=1),
p2 = rep(1, npar),
upper = rep(1, npar))
## !!!The names are important!!!
priors <- list(pprior = p.prior, location = mu.prior, scale = sigma.prior)
names(priors)
## [1] "pprior" "location" "scale"
## Fit hierarchical model ----
fit0 <- StartNewsamples(dmi, priors)
fit <- run(fit0, 5e2)
p0 <- plot(fit, hyper = TRUE)
p0 <- plot(fit, hyper = TRUE, den = TRUE, pll=FALSE)
## Check model -----------
## hgelman function is deprecated
res <- gelman(fit, verbose = TRUE)
est0 <- summary(fit, recovery = TRUE, ps = ps, verbose = TRUE)
est1 <- summary(fit, hyper = TRUE, recovery = TRUE, ps = pop.mean, type = 1, verbose = TRUE)
est2 <- summary(fit, hyper = TRUE, recovery = TRUE, ps = pop.scale, type = 2, verbose = TRUE)
- The LBA model, type = "norm",
- The DDM, type = "rd",
- The Wiener diffusion, type = "rd" and set sv=0 and sz=0
- The Piecewise LBA model 0; CPU-based PDA likelihoods; type = "plba0",
- The Piecewise LBA model 1; CPU-based PDA likelihoods; type = "plba1",
- The Piecewise LBA model 0; GPU-based PDA likelihoods; type = "plba0_gpu",
- The Piecewise LBA model 1; GPU-based PDA likelihoods; type = "plba1_gpu",
- The LBA model; GPU-based PDA likelihoods;, type = "norm_pda_gpu",
- The leaky, competing accumulator model (Experimental!).
4 to 8 are separated from the latest version of the package. For these PDA-based models, see my BRM paper and associated packages there (osf project). 9 is in a separate module, which has yet to be incorporated. See the LCA tutorial for its testing result using MLE.
For the details regarding PLBA types, please see Holmes, Trueblood, and Heathcote (2016)
- 2-D/circular drift-diffusion model, type = "cddm"
- Prospective memory model, type = "norm" (see tutorial for more details)
- Time-varying changes in other free parameters
One aim in designing ggdmc is to read objects from DMC, which share similarities. They have, however, some differences. For example, in the latest version of ggdmc, the dimension of theta and phi arrays are 'npar x nchain x nmc'. DMC uses 'nchain x npar x nmc'. To reduce the computation time for manipulating the matrices and arrays, we change this to accommodate the Armadillo convention change to accommodate the convention in Armadillo. Similarly, the dimension of the 'log_likelihoods' and 'summed_log_prior' matrices are 'nchain x nmc'. DMC uses 'nmc x nchain'. Remember to transpose them if you want to operate objects back and forth. Currently, we use the R functions, 'aperm' and 't', to transpose matrices and arrays when operating have to operate between DMC and ggdmc. The following two convenient functions are designed for doing this operation.
DMC2ggdmc <- function(x) {
## x is an object of posterior samples from individual subject fit
x$theta <- aperm(x$theta, c(2, 1, 3))
x$summed_log_prior <- t(x$summed_log_prior)
x$log_likelihoods <- t(x$log_likelihoods)
class(x) <- c('list', 'model')
return(x)
}
ggdmc2DMC <- function(x) {
## Should change $ to @ whenoperatinge output from ggdmc's run function,
## because ggdmc now uses S4 class
x$theta <- aperm(x$theta, c(2, 1, 3))
x$summed_log_prior <- t(x$summed_log_prior)
x$log_likelihoods <- t(x$log_likelihoods)
return(x)
}
Note Dstats.dmc in DMC is also affected by the issue of the different array and
matrix dimensions because Dstats.dmc calculates the means of the theta/phi
array across the column,
apply(samples$theta,2,mean)
ggdmc provides DIC function, which uses a back-end function, deviance_model to attain the same operation.
The tutorial in 3-accumulator LBA model illustrates an example of doing the back-and-forth operation.
Note that we start to use S4 class after version 0.2.7.5, so switch to use "@" operator to extract object components (i.e., slot).
- R (>= 3.3.0)
- R packages: Rcpp (>= 0.12.10), RcppArmadillo (>= 0.7.700.3.0), ggplot2 (>= 2.1.0), coda (>= 0.16-1), matrixStats, data.table
- Windows users need Rtools (>= 3.3.0.1959)
Mac OS users need to make Clang understand the OpenMP flag.Linux/Unix users may need to install the Open MPI library if it has not been installed.Armadillo may need a recent g++ compiler > 4.6
We now use S4 class after version 0.2.7.5. The new design enables a more user-friendly interface in help pages.
From CRAN (0.2.6.0):
install.packages("ggdmc")
From source:
install.packages("ggdmc_0.2.8.0.tar.gz", repos = NULL, type="source")
From GitHub (need devtools) (0.2.8.0):
devtools::install_github("yxlin/ggdmc")
For Microsoft R users:
As of 06-01-2020, because Microsoft R uses R version 3.5.3, the user who wishes to deploy ggdmc on Microsoft R may encounter two challenges. First is RcppArmadillo on MRAN is behind the one on R CRAN. The RcppArmadillo on MRAN has yet to introduce recent Armadillo functions, for instance, randperm in C++. This can be resolved by installing RcppArmadillo directly from its source tarball, downloaded from CRAN. Secondly, the default installation process on Windows is to look for the package binary matching the R version on a Windows machine. This may result in Microsoft R looking for a version of ggdmc matching R 3.5.3, and thereby, it cannot find one. This can be resolved similarly by installing from the source tarball.
For Mac Users:
1. Install gfortran.
As of 27, Aug, 2018, the gfortran version has to be 6.1, even if you are using a
macOS High Sierra Version 10.13.4. gfortran 6.3 may not work.
2. Install clang4-r.
James Balamuta
has created a convenient tool, clang4-r.
Once you install clang4-r, your clang will understand the OpenMP flag
in ggdmc. The aim is to allow macOS to understand the OpenMP flag, so you may use
other methods if you do not want to install clang4-r. The
clang4-r is the most straightforward we have found so far.
However, we have yet to look into the source code of clang4-r. You can use it at your
own risk.
The configure script now disables OpenMP, so macOS users should be able to install without encountering the OpenMP problem.
FAQ:
- When the error message arises, "/usr/bin/ld: cannot find -lgsl" and/or "/usr/bin/ld: cannot find -lgslcblas", installing libgsl-dev may resolve this problem (Ubuntu).
Lin, Y.-S and Strickland, L. (2020). Evidence accumulation models with R: A a practical guide to hierarchical Bayesian methods. The Quantitative Methods for Psychology.
The R documentation, tutorials, C++ codes, parallel computations, new genetic algorithm, R helper functions and R packaging are developed by Yi-Shin Lin. A substantial part of R codes for handling experimental designs are adapted from the DMC, developed by Andrew Heathcote (Heathcote et al., 2018). You could find different and more interesting cognitive models in DMC.
Please report bugs to me or start an issue here.
The help page for the function, likelihood in Density.cpp states that it returns log-likelihood (v2.8.0). An inspection of the source code found that it returns likelihood, not log likelihood. (13-06-2022; v2.8.1). Thanks for Nachshon Meiran pointing it out.
- The PDF, CDF and random number generation of DDM were derived from Voss & Voss's fast-dm 30.2 and rtdists 0.9-0.
- Truncated normal functions were originally based on Jonathan Olmsted's RcppTN 0.1-8 at https://github.com/olmjo/RcppTN, Christopher Jackson's R codes in msm package, and Robert's paper (1995, Statistics & Computing).
- Thanks to Matthew Gretton's consultation regarding the rtdists package.
- Thanks to Andrew Heathcote for lending me his MacBook Air. ggdmc works on OS X (macOS High Sierra Version 10.13.4)
- The PDF and random number generation of the 2-D diffusion/circular diffusion model is based on Smith (2016).
- Heathcote, A., Lin, Y.-S., Reynolds, A., Strickland. L. Gretton, M., & Matzke, D. (2018). Dynamic models of choice, Behavior Research Methods. https://doi.org/10.3758/s13428-018-1067-y
- Lin, Y.-S. and Strickland, L. (2020). Evidence accumulation models with R: A practical guide to hierarchical Bayesian methods. The Quantitative Methods for Psychology.
- Smith, P. (2016). Diffusion Theory of Decision Making in Continuous Report, Psychological Review, 123(4), 425-451. http://dx.doi.org/10.1037/rev0000023