Build Regression models over Drift Diffusion Model parameters using MCMC!
You can install latest version of RegDDM
from CRAN:
install.packages("RegDDM")
For RStudio users, you may need the following:
install.packages("rstudioapi")
First, load the package and the example dataset.
library(RegDDM)
data(regddm_data)
data1
is the subject-level dataset:
head(regddm_data$data1)
#> id iq age gender race education
#> 1 1 112.80 22 F White 14
#> 2 2 114.32 22 F White 16
#> 3 3 116.96 22 F Black 13
#> 4 4 111.68 31 F Black 16
#> 5 5 121.36 21 M Asian 16
#> 6 6 124.24 29 F White 18
data2
is the subject-level dataset:
head(regddm_data$data2)
#> id memload response rt
#> 1 1 1 1 1.1772806
#> 2 1 1 1 2.2207544
#> 3 1 6 1 4.4166550
#> 4 1 6 1 0.8540982
#> 5 1 3 1 1.3794191
#> 6 1 3 0 0.8278006
Specify the model using a list. In this example, the drift rate v
is
influenced by memload
, which is the memory load of the trial. The
subject’s iq
is predicted by baseline drift rate v_0
(drift rate
when memload
is 0), the influence of memload
on drift rate
v_memload
and covariates age
and education
:
model = list(
v ~ memload,
iq ~ v_memload + v_0 + age + education
)
Use the main function of RegDDM
to automatically generate the RStan
model and summary the results. This could take ~20 minutes to run. The
rows starting with ‘beta_’ are the posterior distributions of
regression parameters:
fit = regddm(
regddm_data$data1,
regddm_data$data2,
model
)
print(fit)
#> RegDDM Model Summary
#> Number of subjects: 49
#> Number of trials: 6032
#> Model:
#> v ~ memload
#> iq ~ v_memload + v_0 + age + education
#> Family: gaussian
#> Sampling: 4 chains, 500 warmups and 1000 iterations were used. Longest elapsed time is 3218 s.
#>
#> Regression coefficients:
#> variable mean sd 2.5% 97.5% n_eff Rhat
#> 1 beta_0 112.8863 12.308 87.834 135.485 1621 0.999
#> 2 beta_v_memload -54.3336 27.418 -110.982 -3.216 958 1.001
#> 3 beta_v_0 -3.6667 2.003 -7.883 0.156 2602 0.999
#> 4 beta_age 0.1293 0.329 -0.498 0.809 2541 0.998
#> 5 beta_education -0.0422 0.601 -1.216 1.083 2652 0.999
#> 6 sigma 6.8007 0.826 5.350 8.599 2004 0.999
#> Maximum R-hat: 1.005
In this example, iq
is negatively correlated with v_memload
. The
higher the influence of memload
on drift rate, the lower the iq
of
the subject.
If you want to fit the model on your own data, you need to specify
data1
, data2
and model
.
data1
is subject-level data table. It should contain the following: *
id
: unique indexing column for each subject. * other subject-level
variables that we want to include in the regression. Missing values are
supported
data2
is trial-level data table. It should contain the following: *
id
: the subject of each trial using the same index in data1
. *
rt
: response time of the trial in seconds. * `response``: response
the trial. must be either 0 or 1. * trial-level variables. These are
the variables that differ by trial, such as difficulty of the task or
different numbers on the screen. We assume that subjects’ behavior
changes according to these variables. These variables cannot contain
missing values.
model
is the proposed dependency between these parameters. Default is
an empty list. It must be a list of 0 - 5 formulas. The outcome of these
formulas can be either: * one of the four DDM parameters a
, t
, z
,
v
, modeling the relationship between DDM parameters and trial-level
variables. * one formula for GLM regression, modeling the relationship
between estimated DDM parameters and other subject-level variables.
family
is the family of distribution of GLM. It can be either
"gaussian"
, "bernoulli"
or "poisson"
. Default is "gaussian"
.
init
is how we initialize the MCMC algorithm. The "default"
initialization should work in most conditions
prior
determines whether to use the default prior for DDM parameters
or not. Default is TRUE
stan_filename
is the file location for the automatically generated
stan model. If an empty string ’’ is provided, a temporary file will be
created and deleted after the model is fit. Default is
"stan_model.stan"
gen_model
determines whether to generate the model or not. Default is
TRUE
.
fit_model
determines whether to fit the model or not. Default is
TRUE
.
...
: additional parameters used by rstan
, including
warmup
,iter
,chains
,cores
etc.
to be added