fddm provides the function dfddm(), which
evaluates the density function (or probability density function, PDF)
for the Ratcliff diffusion decision model (DDM) using different methods
for approximating the full PDF, which contains an infinite sum.
fddm also provides the family of functions
d*_dfddm(), which evaluate the first-order partial
derivatives of the DDM density function with respect to the parameter
indicated by the * in the function name; the available
parameters are listed below. Similarly, fddm provides the
family of functions d*2_dfddm(), which evaluate the
second-order partial derivatives of the DDM density function for the
same parameters. Based on the density function and its partial
derivatives, fddm provides the function ddm(),
which fits the DDM to provided data. fddm also provides the
function pfddm(), which evaluates the distribution function
(or cumulative distribution function, CDF) for the DDM using two
different methods for approximating the CDF.
Our implementation of the DDM has the following parameters: a ϵ (0, ∞) (threshold separation), v ϵ (-∞, ∞) (drift rate), t0 ϵ [0, ∞) (non-decision time/response time constant), w ϵ (0, 1) (relative starting point), sv ϵ (0, ∞) (inter-trial-variability of drift), and sigma ϵ (0, ∞) (diffusion coefficient of the underlying Wiener Process).
You can install the released version of fddm from CRAN with:
install.packages("fddm")And the development version from GitHub with:
# install.packages("devtools")
devtools::install_github("rtdists/fddm")As a preliminary example, we will fit the DDM to the data from one
participant in the med_dec data that comes with
fddm. This dataset contains the accuracy condition reported
in Trueblood et al. (2018), which investigates medical decision making
among medical professionals (pathologists) and novices (i.e.,
undergraduate students). The task of participants was to judge whether
pictures of blood cells show cancerous cells (i.e., blast cells) or
non-cancerous cells (i.e., non-blast cells). The dataset contains 200
decisions per participant, based on pictures of 100 true cancerous cells
and pictures of 100 true non-cancerous cells. Here we use the data
collected from the trials of one experienced medical professional
(pathologist). First, we load the fddm package, remove any
invalid responses from the data, and select the individual whose data we
will use for fitting.
library("fddm")
data(med_dec, package = "fddm")
med_dec <- med_dec[which(med_dec[["rt"]] >= 0), ]
onep <- med_dec[ med_dec[["id"]] == "2" & med_dec[["group"]] == "experienced", ]
str(onep)
#> 'data.frame': 200 obs. of 9 variables:
#> $ id : Factor w/ 37 levels "1","2","3","4",..: 2 2 2 2 2 2 2 2 2 2 ...
#> $ group : Factor w/ 3 levels "experienced",..: 1 1 1 1 1 1 1 1 1 1 ...
#> $ block : int 3 3 3 3 3 3 3 3 3 3 ...
#> $ trial : int 1 2 3 4 5 6 7 8 9 10 ...
#> $ classification: Factor w/ 2 levels "blast","non-blast": 1 2 2 2 1 1 1 1 2 1 ...
#> $ difficulty : Factor w/ 2 levels "easy","hard": 1 1 2 2 1 1 2 2 1 2 ...
#> $ response : Factor w/ 2 levels "blast","non-blast": 1 2 1 2 1 1 1 1 2 1 ...
#> $ rt : num 0.853 0.575 1.136 0.875 0.748 ...
#> $ stimulus : Factor w/ 312 levels "blastEasy/AuerRod.jpg",..: 7 167 246 273 46 31 132 98 217 85 ...ddm()The ddm() function fits the 5-parameter DDM to the
user-supplied data via maximum likelihood estimation. Each DDM parameter
can be modeled using R’s formula interface; the model parameters can
either be fixed or estimated, except for the drift rate which is always
estimated.
We will demonstrate a simple example of how to fit the DDM to the
onep dataset from the above code chunks.
Because we use an ANOVA approach, we set orthogonal sum-to-zero contrasts.
op <- options(contrasts = c('contr.sum', 'contr.poly'))Now we can use the ddm() function to fit the DDM to the
data. Note that we are using formula notation to indicate the
interaction between variables for the drift rate. The first argument of
the ddm() function is the formula indicating how the drift
rate should be modeled. By default, the boundary separation and
non-decision time are estimated, and the initial bias and inter-trial
variability in the drift rate are held fixed.
fit0 <- ddm(rt + response ~ classification*difficulty, data = onep)
summary(fit0)
#>
#> Call:
#> ddm(drift = rt + response ~ classification * difficulty, data = onep)
#>
#> DDM fit with 3 estimated and 2 fixed distributional parameters.
#> Fixed: bias = 0.5, sv = 0
#>
#> drift coefficients (identity link):
#> Estimate Std. Error z value Pr(>|z|)
#> (Intercept) -0.5924 0.1168 -5.073 3.91e-07 ***
#> classification1 -2.6447 0.1168 -22.647 < 2e-16 ***
#> difficulty1 0.2890 0.1168 2.475 0.0133 *
#> classification1:difficulty1 -1.4987 0.1168 -12.834 < 2e-16 ***
#> ---
#> Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
#>
#> boundary coefficients (identity link):
#> Estimate Std. Error
#> (Intercept) 2.064 0.058
#>
#> ndt coefficients (identity link):
#> Estimate Std. Error
#> (Intercept) 0.3938 0.007This output first shows the input to the function call and which
parameters are held fixed; this information is useful to verify that the
formula inputs to the ddm() function call were correct. For
the model of the drift rate that we input, we can see the estimates and
summary statistics for each coefficient. Below this, we can see the
simple estimates for the boundary separation and non-decision time
(default behavior).
We can reset the contrasts after fitting.
options(op) # reset contrastsnlminb()Although we strongly recommend using the ddm() function
for fitting the DDM to data because it is faster and more convenient, we
will also show how to use the probability density function in a manual
optimization setup. We further prepare the data by defining upper and
lower responses and the correct response bounds.
onep[["resp"]] <- ifelse(onep[["response"]] == "blast", "upper", "lower")
onep[["truth"]] <- ifelse(onep[["classification"]] == "blast", "upper", "lower")For fitting, we need a simple likelihood function; here we will use a
straightforward log of sum of densities of the study responses and
associated response times. This log-likelihood function will fit the
standard parameters in the DDM, but it will fit two versions of the
drift rate v: one for when the correct response is
"blast" (vu), and another for when the
correct response is "non-blast" (vl). A
detailed explanation of the log-likelihood function is provided in the
Example Vignette (vignette("example", package = "fddm")).
Note that this likelihood function returns the negative log-likelihood
as we can simply minimize this function to get the maximum likelihood
estimate.
ll_fun <- function(pars, rt, resp, truth) {
v <- numeric(length(rt))
# the truth is "upper" so use vu
v[truth == "upper"] <- pars[[1]]
# the truth is "lower" so use vl
v[truth == "lower"] <- pars[[2]]
dens <- dfddm(rt = rt, response = resp, a = pars[[3]], v = v,
t0 = pars[[4]], w = pars[[5]], sv = pars[[6]], log = TRUE)
return( ifelse(any(!is.finite(dens)), 1e6, -sum(dens)) )
}We then pass the data and log-likelihood function to an optimization
function with the necessary additional arguments. As we are using the
optimization function nlminb for this example, the first
argument must be the initial values of our DDM parameters that we want
optimized. These are input in the order: vu,
vl, a, t0, w,
and sv; we also need to define upper and lower bounds for each
of the parameters. Fitting the DDM to this dataset is basically
instantaneous using this setup.
fit <- nlminb(c(0, 0, 1, 0, 0.5, 0), objective = ll_fun,
control = list(iter.max = 300, eval.max = 300),
rt = onep[["rt"]], resp = onep[["resp"]], truth = onep[["truth"]],
# limits: vu, vl, a, t0, w, sv
lower = c(-Inf, -Inf, .01, 0, 0, 0),
upper = c( Inf, Inf, Inf, min(onep[["rt"]]), 1, Inf))
fit
#> $par
#> [1] 5.6813063 -2.1886615 2.7909130 0.3764465 0.4010115 2.2813001
#>
#> $objective
#> [1] 42.47181
#>
#> $convergence
#> [1] 0
#>
#> $iterations
#> [1] 231
#>
#> $evaluations
#> function gradient
#> 251 1656
#>
#> $message
#> [1] "relative convergence (4)"