Given a data file (text file, `data.frame`

) with the
correct variable names (see section 1.3) and a model file (see section
1.1 and 2) the steps needed to fit the RT-MPT model (Hartmann et al.,
2020; Hartmann & Klauer, 2020; Klauer & Kellen, 2018) to the
data is described in Section 1:

1.1 Convert the equation/model file to a `rtmpt_model`

via
the `to_rtmpt_model()`

function 1.2 If necessary, specify
some restrictions 1.3 Convert the data file to a `rtmpt_data`

format via the `to_rtmpt_data()`

function 1.4 Fit the
specified model to the data via the `fit_rtmpt()`

function
1.5 Check convergence of the model parameters and getting a summary

In Section 2, 3, and 4 more information about the equation/model files, the data and additional functions is provided.

The equation/model file can be provided by a text file or directly
like the following. One might use a syntax that is comparable to the
`EQN`

syntax by Heck, Arnold, and Arnold (2018) or Hu (1999)
or a syntax that is comparable to the `MDL`

syntax by
Singmann and Kellen (2013). In the `EQN`

syntax you can
separate trees, categories and paths either by using semicolons (like in
the example below) or commas, but not mixed:

```
= "
eqn # CORE MODEL
## EQN SYNTAX
## tree ; cat ; mpt
0 ; 0 ; Do
0 ; 0 ; (1-Do)*g
0 ; 1 ; (1-Do)*(1-g)
1 ; 3 ; Dn
1 ; 2 ; (1-Dn)*g
1 ; 3 ; (1-Dn)*(1-g)
"
```

In this `eqn`

object we specify a Two-High Threshold (2HT)
model. Comments can be done with a `#`

symbol at the
beginning of a line.

The same model can be specified with the `MDL`

syntax:

```
= "
mdl # CORE MODEL
## MDL SYNTAX
### targets
Do+(1-Do)*g
(1-Do)*(1-g)
### lure
(1-Dn)*g
Dn+(1-Dn)*(1-g)
"
```

All paths that lead to the same response category will be added
together with a `+`

symbol.

More information about these two syntaxes are written in Section 2.

The conversion to an `rtmpt_model`

list can then be done
by the `to_rtmpt_model()`

function:

```
library(rtmpt)
# using the MDL syntax:
<- to_rtmpt_model(mdl_file = mdl)
TwoHTM # using the EQN syntax:
<- to_rtmpt_model(eqn_file = eqn)
TwoHTM TwoHTM
```

```
##
## MODEL OVERVIEW
##
##
## MDL syntax for the MPT part:
## MDL_lines
## 1 # 0 [ 0 , 1 ]
## 2 Do+(1-Do)*g
## 3 (1-Do)*(1-g)
## 4
## 5 # 1 [ 2 , 3 ]
## 6 (1-Dn)*g
## 7 Dn+(1-Dn)*(1-g)
##
## * NOTE 1: Each line in the MDL syntax represents one response category.
## ----------------------------
##
## Process probability parameters (thetas):
## Dn Do g
## 1 NA NA NA
##
## * NOTE 1: NA means the parameter will be estimated.
## * NOTE 2: A value larger than 0 and smaller than 1 means it will be held constant.
## * INFO: 0 of the process probabilities will be held constant.
## -------------------------------
##
## Process completion time parameters (taus):
## Dn Do g
## minus NA NA NA
## plus NA NA NA
##
## * NOTE 1: "minus" refers to the negative outcome (1-P) and "plus" to the positive.
## * NOTE 2: NA means the parameter will be estimated. 0 means it will be suppressed.
## * INFO: 0 of the process completion times will be suppressed.
## -----------------------------------
##
## Mapping of response categories and encoding plus motor execution times (deltas):
## TREE CAT MAP
## 1 0 0 0
## 3 0 1 0
## 5 1 2 0
## 4 1 3 0
##
## * NOTE 1: Unique representation of trees and categories.
## * NOTE 2: Each mapping number corresponds to a distinct encoding plus motor execution time.
## * INFO: 1 distinct delta(s) assumed, namely with mapping [ 0 ].
## -----------------------------------
```

`# `

There are five functions with which an `rtmpt_model`

object can be changed. Two functions can be used to hold a process
probability parameter (theta) constant or to make some process
probability parameters equal. Two functions can be used to suppress
process completion times (taus), which means the process times are set
to zero, or to make some process completion times equal. The last
function handles how many different motor and execution times (deltas)
are used in the model.

To hold a process probability constant you can use
`theta2const()`

like this:

`theta2const(model = TwoHTM, names = "g", constants = 0.5)`

```
##
## MODEL OVERVIEW
##
##
## MDL syntax for the MPT part:
## MDL_lines
## 1 # 0 [ 0 , 1 ]
## 2 Do+(1-Do)*g
## 3 (1-Do)*(1-g)
## 4
## 5 # 1 [ 2 , 3 ]
## 6 (1-Dn)*g
## 7 Dn+(1-Dn)*(1-g)
##
## * NOTE 1: Each line in the MDL syntax represents one response category.
## ----------------------------
##
## Process probability parameters (thetas):
## Dn Do g
## 1 NA NA 0.5
##
## * NOTE 1: NA means the parameter will be estimated.
## * NOTE 2: A value larger than 0 and smaller than 1 means it will be held constant.
## * INFO: 1 of the process probabilities will be held constant.
## -------------------------------
##
## Process completion time parameters (taus):
## Dn Do g
## minus NA NA NA
## plus NA NA NA
##
## * NOTE 1: "minus" refers to the negative outcome (1-P) and "plus" to the positive.
## * NOTE 2: NA means the parameter will be estimated. 0 means it will be suppressed.
## * INFO: 0 of the process completion times will be suppressed.
## -----------------------------------
##
## Mapping of response categories and encoding plus motor execution times (deltas):
## TREE CAT MAP
## 1 0 0 0
## 3 0 1 0
## 5 1 2 0
## 4 1 3 0
##
## * NOTE 1: Unique representation of trees and categories.
## * NOTE 2: Each mapping number corresponds to a distinct encoding plus motor execution time.
## * INFO: 1 distinct delta(s) assumed, namely with mapping [ 0 ].
## -----------------------------------
```

Here, `"g"`

will be set to be equal to 0.5 and will not be
estimated therefore.,You might also use this function with a vector of
`names`

(and a vector of `constants`

). Note that
the new model is not saved because we did not assign it to an R
object.

To set two or more process probabilities equal you can use
`theta2theta()`

like this:

`theta2theta(model = TwoHTM, names = c("Do", "Dn"), keep_consts = FALSE)`

```
##
## MODEL OVERVIEW
##
##
## MDL syntax for the MPT part:
## MDL_lines
## 1 # 0 [ 0 , 1 ]
## 2 Do+(1-Do)*g
## 3 (1-Do)*(1-g)
## 4
## 5 # 1 [ 2 , 3 ]
## 6 (1-Dn)*g
## 7 Dn+(1-Dn)*(1-g)
##
## * NOTE 1: Each line in the MDL syntax represents one response category.
## ----------------------------
##
## Process probability parameters (thetas):
## Dn Do g
## 1 NA Dn NA
##
## * NOTE 1: NA means the parameter will be estimated.
## * NOTE 2: A value larger than 0 and smaller than 1 means it will be held constant.
## * INFO: 1 of the process probabilities will be held constant.
## -------------------------------
##
## Process completion time parameters (taus):
## Dn Do g
## minus NA NA NA
## plus NA NA NA
##
## * NOTE 1: "minus" refers to the negative outcome (1-P) and "plus" to the positive.
## * NOTE 2: NA means the parameter will be estimated. 0 means it will be suppressed.
## * INFO: 0 of the process completion times will be suppressed.
## -----------------------------------
##
## Mapping of response categories and encoding plus motor execution times (deltas):
## TREE CAT MAP
## 1 0 0 0
## 3 0 1 0
## 5 1 2 0
## 4 1 3 0
##
## * NOTE 1: Unique representation of trees and categories.
## * NOTE 2: Each mapping number corresponds to a distinct encoding plus motor execution time.
## * INFO: 1 distinct delta(s) assumed, namely with mapping [ 0 ].
## -----------------------------------
```

The process probabilities for all processes specified in
`names`

are set equal. The first of the process
`names`

in alphabetical order (here `"dn"`

)
denotes the reference probability that will be estimated. The other(s)
will not be estimated but will be set to be equal to the reference
process probability. The `keep_consts`

argument can be used,
if you want to keep constants that you already specified and all other
process probabilities should have the same constant but we suggest to
use `theta2const()`

for that.

To suppress a process completion time (i.e., set it to zero) you can
use `tau2zero()`

like this:

`tau2zero(model = TwoHTM, names = "g", outcomes = "minus", values = 0)`

```
##
## MODEL OVERVIEW
##
##
## MDL syntax for the MPT part:
## MDL_lines
## 1 # 0 [ 0 , 1 ]
## 2 Do+(1-Do)*g
## 3 (1-Do)*(1-g)
## 4
## 5 # 1 [ 2 , 3 ]
## 6 (1-Dn)*g
## 7 Dn+(1-Dn)*(1-g)
##
## * NOTE 1: Each line in the MDL syntax represents one response category.
## ----------------------------
##
## Process probability parameters (thetas):
## Dn Do g
## 1 NA NA NA
##
## * NOTE 1: NA means the parameter will be estimated.
## * NOTE 2: A value larger than 0 and smaller than 1 means it will be held constant.
## * INFO: 0 of the process probabilities will be held constant.
## -------------------------------
##
## Process completion time parameters (taus):
## Dn Do g
## minus NA NA 0
## plus NA NA NA
##
## * NOTE 1: "minus" refers to the negative outcome (1-P) and "plus" to the positive.
## * NOTE 2: NA means the parameter will be estimated. 0 means it will be suppressed.
## * INFO: 1 of the process completion times will be suppressed.
## -----------------------------------
##
## Mapping of response categories and encoding plus motor execution times (deltas):
## TREE CAT MAP
## 1 0 0 0
## 3 0 1 0
## 5 1 2 0
## 4 1 3 0
##
## * NOTE 1: Unique representation of trees and categories.
## * NOTE 2: Each mapping number corresponds to a distinct encoding plus motor execution time.
## * INFO: 1 distinct delta(s) assumed, namely with mapping [ 0 ].
## -----------------------------------
```

Here, the process completion time of `"g"`

with outcome
`"minus"`

(meaning the time for 1-g) is set to zero and will
not be estimated therefore (which typically does not make sense to do).
Also this function can be used with vectors.

To set two or more process completion times with the same outcome
equal you can use `theta2theta()`

like this:

`tau2tau(model = TwoHTM, names = c("Do", "Dn"), keep_zeros = FALSE)`

```
##
## MODEL OVERVIEW
##
##
## MDL syntax for the MPT part:
## MDL_lines
## 1 # 0 [ 0 , 1 ]
## 2 Do+(1-Do)*g
## 3 (1-Do)*(1-g)
## 4
## 5 # 1 [ 2 , 3 ]
## 6 (1-Dn)*g
## 7 Dn+(1-Dn)*(1-g)
##
## * NOTE 1: Each line in the MDL syntax represents one response category.
## ----------------------------
##
## Process probability parameters (thetas):
## Dn Do g
## 1 NA NA NA
##
## * NOTE 1: NA means the parameter will be estimated.
## * NOTE 2: A value larger than 0 and smaller than 1 means it will be held constant.
## * INFO: 0 of the process probabilities will be held constant.
## -------------------------------
##
## Process completion time parameters (taus):
## Dn Do g
## minus NA Dn NA
## plus NA Dn NA
##
## * NOTE 1: "minus" refers to the negative outcome (1-P) and "plus" to the positive.
## * NOTE 2: NA means the parameter will be estimated. 0 means it will be suppressed.
## * INFO: 0 of the process completion times will be suppressed.
## -----------------------------------
##
## Mapping of response categories and encoding plus motor execution times (deltas):
## TREE CAT MAP
## 1 0 0 0
## 3 0 1 0
## 5 1 2 0
## 4 1 3 0
##
## * NOTE 1: Unique representation of trees and categories.
## * NOTE 2: Each mapping number corresponds to a distinct encoding plus motor execution time.
## * INFO: 1 distinct delta(s) assumed, namely with mapping [ 0 ].
## -----------------------------------
```

The process completion times for all processes specified in
`names`

are set equal (i.e., all process completion times in
minus-direction will be equal and all process completion times in
plus-direction will be equal). The first of the process
`names`

in alphabetical order (here `"dn"`

)
denotes the reference process times (in minus and plus direction) that
will be estimated. The other(s) will not be estimated but will be set to
be equal to the reference process times. The `keep_zeros`

argument can be used, if you want to keep the zeros that you already
specified and all other process times with the same outcome should be
zero as well. Also here we suggest using the `tau2zero()`

function.

You can choose which categories of the (RT-)MPT model should have the
same encoding plus motor execution time. Unlike the processes the motor
plus execution times have no names which we can use to set them equal.
Therefore a mapping is used in the function `delta2delta()`

where the response categories (`categories`

) can be mapped to
encoding plus motor execution times by using a mapping argument
(`mappings`

). Because the response categories are not always
unique the tree (`trees`

) has to be provided as well.

`delta2delta(model = TwoHTM, trees = c(0,1), categories = c(1,3), mappings = c(1,1))`

```
##
## MODEL OVERVIEW
##
##
## MDL syntax for the MPT part:
## MDL_lines
## 1 # 0 [ 0 , 1 ]
## 2 Do+(1-Do)*g
## 3 (1-Do)*(1-g)
## 4
## 5 # 1 [ 2 , 3 ]
## 6 (1-Dn)*g
## 7 Dn+(1-Dn)*(1-g)
##
## * NOTE 1: Each line in the MDL syntax represents one response category.
## ----------------------------
##
## Process probability parameters (thetas):
## Dn Do g
## 1 NA NA NA
##
## * NOTE 1: NA means the parameter will be estimated.
## * NOTE 2: A value larger than 0 and smaller than 1 means it will be held constant.
## * INFO: 0 of the process probabilities will be held constant.
## -------------------------------
##
## Process completion time parameters (taus):
## Dn Do g
## minus NA NA NA
## plus NA NA NA
##
## * NOTE 1: "minus" refers to the negative outcome (1-P) and "plus" to the positive.
## * NOTE 2: NA means the parameter will be estimated. 0 means it will be suppressed.
## * INFO: 0 of the process completion times will be suppressed.
## -----------------------------------
##
## Mapping of response categories and encoding plus motor execution times (deltas):
## TREE CAT MAP
## 1 0 0 0
## 3 0 1 1
## 5 1 2 0
## 4 1 3 1
##
## * NOTE 1: Unique representation of trees and categories.
## * NOTE 2: Each mapping number corresponds to a distinct encoding plus motor execution time.
## * INFO: 2 distinct delta(s) assumed, namely with mapping [ 0, 1 ].
## -----------------------------------
```

The numbers for the mappings have to start at zero and no whole
number can be skipped. For example in the above code the
`mappings`

cannot be `c(2,2)`

since the number 1
would be missing then in the mappings.

The data should be provided with the following variables in this
order: `subj`

, `group`

, `tree`

,
`cat`

, and `rt`

. All of these variables except
`rt`

should have values from zero to the number of unique
values minus one. This will get clear with an example; if we have 60
subjects, two groups, two trees and four categories we would have the
following values in the variables:

`subj = {0,1,2,...,58,59}`

,`group = {0,1}`

,`tree = {0,1}`

, and`cat = {0,1,2,3}`

.

`rt`

should be provided in milliseconds and as integers.
If it has decimal places it will be rounded to a whole number.

Given the right order of the variables and the right values one can
use a `data.frame`

or a path to a text file containing the
variable names in the first line. If not, the
`to_rtmpt_data()`

function can be used to reorder the
variables and transform the values of those variables. Next to the new
`data.frame`

the transformation that is used will be saved in
an `rtmpt_data`

list. Note that **at least** the
variable names should be correctly provided in order to make
`to_rtmpt_data()`

work. The `to_rtmpt_data()`

function is also used automatically in the `fit_rtmpt()`

function, but should be used at least once, when the model is specified
the first time, to check whether the data is transformed according to
the equation/model file.

Here is an example of a data set with unordered variables and wrong values (not starting from zero) and a transformation of this data set:

```
set.seed(2021)
<- data.frame(tree = rep(1:2, 8), rt = round(1000*(runif(16)+.3)),
raw_data group = rep(1:2, each=8), subj = rep(1:4, each = 4), cat = rep(1:4, 4))
raw_data
```

```
## tree rt group subj cat
## 1 1 751 1 1 1
## 2 2 1084 1 1 2
## 3 1 1010 1 1 3
## 4 2 682 1 1 4
## 5 1 936 1 2 1
## 6 2 1001 1 2 2
## 7 1 940 1 2 3
## 8 2 567 1 2 4
## 9 1 1115 2 3 1
## 10 2 1283 2 3 2
## 11 1 327 2 3 3
## 12 2 1137 2 3 4
## 13 1 903 2 4 1
## 14 2 867 2 4 2
## 15 1 1120 2 4 3
## 16 2 552 2 4 4
```

```
<- to_rtmpt_data(raw_data = raw_data, model = TwoHTM)
data data
```

```
##
## DATA TRANSFORMATION OVERVIEW
##
##
## Reordered variables:
## subj, group, tree, cat, rt
## * NOTE1: Additional variables are attached next to these five.
## * NOTE2: To see your data frame use <object name>$data.
## --------------------
##
## Transformed variable(s):
## "subj"
## old new
## 1 1 0
## 2 2 1
## 3 3 2
## 4 4 3
##
## "group"
## old new
## 1 1 0
## 2 2 1
##
## "tree"
## old new
## 1 1 0
## 2 2 1
##
## "cat"
## old new
## 1 1 0
## 2 3 1
## 3 2 2
## 4 4 3
##
## * NOTE: "old" refers to the used labels and "new" to the ones that will be used.
## ------------------------
```

Note that it is never a good idea to specify the tree and category labels/numbers differently from the data file, or vice versa. It is bad practice to define a model with tree and category values starting from zero and using a data set with tree and category values starting from one; Make sure the labels or numbers match for these two match.

Finally, with the function `fit_rtmpt()`

one can fit the
model to the data. This code should not be run. There are way too few
data points.

```
## do not run
<- fit_rtmpt(model = TwoHTM, data = data)
rtmpt_out ## end not run
```

In this function call we used the default values for all the
parameters that are not listed in the call. `model`

and
`data`

do not have default values. The full call of the
function would look like this:

```
## do not run
<- fit_rtmpt(model = restr_2HTM,
rtmpt_out data = data,
n.chains = 4,
n.iter = 5000,
n.burnin = 200,
n.thin = 1,
Rhat_max = 1.05,
Irep = 1000,
prior_params = NULL,
indices = FALSE,
save_log_lik = FALSE)
## end not run
```

`n.chains`

is the number of chains with a maximum of 16,
`n.iter`

is the number of samples to draw from the posterior
distribution, `n.burnin`

is analogue to the burn-in phase for
R package `rjags`

, `Rhat_max`

is a threshold for
the potential scale reduction factor and forces the sampling to start
only when the maximal potential scale reduction factor of the parameters
is lower or equal to this value, `n.thin`

is the thinning
parameter, `Irep`

is the number of samples until the next
summary is printed, `prior_params`

is a list of some prior
parameter specifications, `indices`

is a logical parameter
with which one can allow to compute the WAIC and LOO, and
`save_log_lik`

is a logical parameter that allows one to save
the log-likelihood for each posterior sample and data point.

Convergence of the model can be checked by using the fitted object
(here `rtmpt_out`

). Just get the diagnostics and then R-hat.
In order to get a traceplot one can use the `traceplot()`

function from `coda`

. You can also use the
`summary()`

function to get a summary of the estimates.

```
$diags$R_hat
rtmpt_out::traceplot(rtmpt_out$samples[, 1:9])
codasummary(rtmpt_out)
```

In previous versions it was possible to set some restrictions directly in the equation/model file, but due to the number of restrictions that are possible we decided to use now functions instead (see Subsection 1.2).

As we have seen above one can specify an equation file
(`eqn_file`

) like this:

```
= "
eqn # CORE MODEL
## tree ; cat ; path
0 ; 0 ; Do
0 ; 0 ; (1-Do)*g
0 ; 1 ; (1-Do)*(1-g)
1 ; 3 ; Dn
1 ; 2 ; (1-Dn)*g
1 ; 3 ; (1-Dn)*(1-g)
"
```

The trees, categories, and paths need to be separated by a semi-colon
or a comma, but should not be mixed. As mentioned above the
`eqn_file`

can be provided by a text file through a path to
the file or written directly in R/Rstudio like above.

Another way of specifying the same model is by providing a model file
(`mdl_file`

). This is also not yet the model itself as needed
for the `fit_rtmpt()`

call, but much closer. It is based on
the syntax developed by Singmann and Kellen (2013).

```
= "
mdl # CORE MODEL
## targets:
Do+(1-Do)*g
(1-Do)*(1-g)
## lures
(1-Dn)*g
Dn+(1-Dn)*(1-g)
"
```

Between two trees (here targets and lures) there must be an empty
line. All paths which lead to the same response category are added
together on one line by a `+`

symbol.

Internally an `eqn_file`

will be converted to an
`mdl_file`

. From that a text file will be written locally
onto your computer, used by the `fit_rtmpt()`

call, and
removed afterwards.

When labels are used in the data set one has to be cautious when
specifying the equation file. The labels of the trees and categories in
the equation file must match those of the data. For example, if the tree
labels “target” and “lure” are used in the data set, one has to use
those also in the equation file (`eqn_file`

). If you specify
a model file (`mdl_file`

) instead of the equation file the
labels in the data set need to be changed to numerics. Otherwise no
mapping between those can be done.

When numbers are used in the data set for the trees and categories it is best to also check that they match with the ones used in the equation / model file, but the program checks for the order. So for example, if you are using the numbers {1, 2, 3, 4} for your four trees in the data set and {0, 1, 2, 3} in the equation file (which is the standard for model files anyway) they will be mapped through the order. The new data, that can be used, will have a zero instead of a one, a one instead of a two, and so on.

One could use labels in the equation file even though you are using
numbers in the data file, but it is **strongly advised
against** it. It will be assumed, that the order in which you
specify the equations matches the order of the numbers in the data file
starting from the lowest to the highest. So if you are using again {1,
2, 3, 4} for the four trees, then the first tree you need to define in
the equation files (no matter what label you use) must correspond to
tree 1 in the data and so on.

With the function `sim_rtmpt_data()`

it is possible to
generate data from an RT-MPT model. Required is a model object generated
with the function `to_rtmpt_model()`

, a random seed number,
the number of subjects, the number of trials per tree in the model, and
a named list of parameters. Here it is also possible to set the mean of
`mean_of_mu_alpha`

; A value of zero refers to the probability
of `0.5`

. If a given probability is desired, say
`0.4`

, then one might specify
`params <- list(mean_of_mu_alpha = qnorm(.4))`

. If a mean
process time of `200`

ms is required, one might specify
`params <- list(mean_of_exp_mu_beta = 1000/200)`

. The mean
parameter of the motor times is easier to transform, since they are
already on the second scale; Just write `500/1000`

if your
mean motor time should be `500`

ms. The code below shows a
more or less good balance between too large variances (leading to too
many extreme values like probabilities of zero and one and RTs of zero)
and too small variances (unrealistic data).

Important: If no variances are specified in the `params`

argument, then the corresponding mean values are fixed, else they will
be randomly drawn.

```
# randomly drawn group-level mean values
<- "
mdl_2HTM # targets
do+(1-do)*g ; 0
(1-do)*(1-g) ; 1
# lures
(1-dn)*g ; 0
dn+(1-dn)*(1-g) ; 1
# do: detect old; dn: detect new; g: guess
"
<- to_rtmpt_model(mdl_file = mdl_2HTM)
model
# random group-level parameters
<- list(mean_of_mu_alpha = 0,
params var_of_mu_alpha = 1, # delete this line to fix mean_of_mu_alpha to 0
mean_of_exp_mu_beta = 10,
var_of_exp_mu_beta = 10, # delete this line to fix mean_of_exp_mu_beta to 10
mean_of_mu_gamma = 0.5,
var_of_mu_gamma = 0.0025, # delete this line to fix mean_of_mu_gamma to 0.5
mean_of_omega_sqr = 0.005,
var_of_omega_sqr = 0.000025, # delete this line to fix mean_of_omega_sqr to 0.005
df_of_sigma_sqr = 10,
sf_of_scale_matrix_SIGMA = 0.1,
sf_of_scale_matrix_GAMMA = 0.01,
prec_epsilon = 10,
add_df_to_invWish = 5)
<- sim_rtmpt_data(model, seed = 123, n.subj = 40, n.trials = 30, params = params)
sim_dat head(sim_dat$data_frame)
```

```
## subj group tree cat rt
## 1 0 0 0 0 560
## 2 0 0 0 0 456
## 3 0 0 0 0 538
## 4 0 0 0 0 427
## 5 0 0 0 1 420
## 6 0 0 0 0 574
```

With the function `fit_rtmpt_SBC()`

it is possible to
generate data and fit the model to that data using the same priors /
ground truths. The main output of this function will be the rank
statistic for each parameter in the model, which is needed for the
simulation-based calibration (SBC) method by Talts, Betancourt, Simpson,
Vehtari, and Gelman (2018). For example, if this function is repeated,
say `2000`

times, with different `seeds`

one might
add all the rank statistics to a matrix (`2000`

rows and
number of parameters as columns) and then calculate the pearsons’
chi-squared statistic. This allows one to test for how many of the
parameters the rank statistics are not uniformly distributed. A value of
about `0.05`

is typical for a test with an alpha level of
`0.05`

.

Here is an example of one replication of the SBC procedure:

```
mdl_2HTM <- "
# targets
d+(1-d)*g ; 0
(1-d)*(1-g) ; 1
# lures
(1-d)*g ; 0
d+(1-d)*(1-g) ; 1
# d: detect; g: guess
"
model <- to_rtmpt_model(mdl_file = mdl_2HTM)
params <- list(mean_of_exp_mu_beta = 10,
var_of_exp_mu_beta = 10,
mean_of_mu_gamma = 0.5,
var_of_mu_gamma = 0.0025,
mean_of_omega_sqr = 0.005,
var_of_omega_sqr = 0.000025,
df_of_sigma_sqr = 10,
sf_of_scale_matrix_SIGMA = 0.1,
sf_of_scale_matrix_GAMMA = 0.01,
prec_epsilon = 10,
add_df_to_invWish = 5)
SBC_out <- fit_rtmpt_SBC(model, seed = 123, prior_params = params)
SBC_out$ranks
```

With the following code it would essentially be possible to calculate the rank statistics of each parameter and then test it with the pearsons’ chi-squared test, but much is not considered here (e.g. the convergence and the effective sample size).

```
# For 2000 replications
## This takes too long to run and in addition, Rhat should always be
## checked as well as the effective sample size.
R = 2000
rank_mat <- data.frame()
for (r in 1:R) {
SBC_out <- fit_rtmpt_SBC(model, seed = r*123, prior_params = params, n.eff_samples = 99)
rank_mat <- rbind(rank_mat, SBC_out$ranks)
}
## pearsons' chi-square for testing uniformity
x <- apply(rank_mat[1:R,], 2, table)
expect <- R/100 # 100 = number of bins/cells (0:99)
pearson <- apply(X = x, MARGIN = 2, FUN = function(x) {sum((x-expect)^2/expect)})
z95 <- qchisq(0.95, 99) # 99 = degrees of freedom
sum(pearson>z95) / length(pearson)
```

With the following code a rank matrix is generated from a uniform distribution and then tested for uniformity:

```
<- 2000
R <- matrix(data = sample(0:99, R*393, replace=TRUE), nrow = R, ncol = 393)
rand_rankmat
## pearsons' chi-square for testing uniformity
<- apply(rand_rankmat[1:R,], 2, table)
x <- R/100 # 100 = number of bins/cells (0:99)
expect <- apply(X = x, MARGIN = 2, FUN = function(x) {sum((x-expect)^2/expect)})
pearson <- qchisq(0.95, 99) # 99 = degrees of freedom
z95 sum(pearson>z95) / length(pearson)
```

`## [1] 0.05343511`

- Hartmann, R., Johannsen, L., & Klauer, K. C. (2020). rtmpt: An R
package for fitting response-time extended multinomial processing tree
models.
*Behavior Research Methods, 52*(3), 1313-1338. doi:10.3758/s13428-019-01318-x - Hartmann, R., & Klauer, K. C. (2020). Extending RT-MPTs to
enable equal process times.
*Journal of Mathematical Psychology, 96*, 102340. doi:10.1016/j.jmp.2020.102340 - Heck, D. W., Arnold, N. R., & Arnold, D. (2018). TreeBUGS: An R
package for hierarchical multinomial-processing-tree modeling.
*Behavior Research Methods, 50*(1), 264-284. doi:10.3758/s13428-017-0869-7 - Hu, X. (1999). Multinomial processing tree models: An
implementation.
*Behavior Research Methods, Instruments, & Computers, 31*(4), 689-695. doi:10.3758/BF03207714 - Klauer, K. C., & Kellen, D. (2018). RT-MPTs: Process models for
response-time distributions based on multinomial processing trees with
applications to recognition memory.
*Journal of Mathematical Psychology, 82*, 111-130. doi:10.1016/j.jmp.2017.12.003 - Singmann, H., & Kellen, D. (2013). MPTinR: Analysis of
multinomial processing tree models in R.
*Behavior Research Methods, 45*(2), 560-575. doi:10.3758/s13428-012-0259-0 - Talts, S., Betancourt, M., Simpson, D., Vehtari, A., & Gelman,
A. (2018). Validating Bayesian inference algorithms with
simulation-based calibration.
*arXiv preprint arXiv:1804.06788*.