Overview

The edstan package for R provides convenience functions and predefined Stan models related to item response theory (IRT). Its purpose is to make fitting common IRT models using Stan easy. edstan relies on the rstan package, which should be installed first. See here for instructions on installing rstan.

The following table lists the models packaged with edstan. Each of these may optionally included a latent regression of ability. The table includes also links to case studies for the models, though it should be noted that the Stan code and prior distributions have changed somewhat this they were written.

Model	Stan file
Rasch	rasch_latent_reg.stan
Partial credit	pcm_latent_reg.stan
Rating scale	rsm_latent_reg.stan
Two-parameter logistic	2pl_latent_reg.stan
Generalized partial credit	gpcm_latent_reg.stan
Generalized rating scale	grsm_latent_reg.stan

The next table lists the functions packaged with edstan.

Function	Purpose
`irt_data()`	Prepares data for fitting
`irt_stan()`	Wrapper for Running MCMC
`print_irt_stan()`	Show table of output
`stan_columns_plot()`	Create plot of convergence statistics
`labelled_integer()`	Create vector of consecutive integers
`rescale_binary()`	Appropriately scale binary person covariates
`rescale_continuous()`	Appropriately scale continuous person covariates
`edstan_model_code()`	Print the Stan file for a given model

Brief tutorial

Dichotomous IRT with spelling data

The R code below first loads edstan, which implicitly loads rstan. Then the two commands that follow set options related to rstan, which are generally recommended. The first causes compiled Stan models to be saved to disc, which allows models to run more quickly after the first time. The second causes Stan to run multiple MCMC chains in parallel.

# Load packages and set options
library(edstan)
options(mc.cores = parallel::detectCores())

Next we preview the spelling data set.

# Preview the spelling data
spelling[sample(1:nrow(spelling), 6), ]

The data set is a response matrix indicating whether the 658 respondents spelled four words correctly. It also includes a dummy variable for whether the respondent is male. Data is fed into Stan in list form, and the next block of code demonstrates how a suitable data list may be constructed using the irt_data() function. The response matrix (that is, all columns but the first) is provided as the response_matrix option.

# Make a data list
simple_list <- irt_data(response_matrix = spelling[, -1])
str(simple_list)

This data list may be fed to the irt_stan() function to fit a Rasch or 2PL model, but for this example we will add a person covariate to the model using the optional arguments covariates and formula. The covariates option takes a data frame of person-related covariates and the formula option takes a formula to be applied to that data frame. The left side of the formula should be left empty, though implicitly it is latent ability. The choice of what to include in the latent regression (if anything) is made when calling irt_data() to assemble a data list.

In the next block a data list is made that includes male as a covariate. We use the rescale_binary() function to rescale this covariate for compatibility with the prior distributions specified in edstan models.

# Make a data list with person covariates
latent_reg_list <- irt_data(
  response_matrix = spelling[, -1],
  covariates = spelling,
  formula = ~ rescale_binary(male)
)

str(latent_reg_list)

The only difference between the two lists are K and W. The first list has W with \(K=1\) columns, which is a constant for the model intercept. The second list has W with \(K=2\) columns, which correspond to the constant for the model intercept and the male indicator variable. In either list, W has one row per person.

Next we fit the model using the irt_stan() function, which is a wrapper for rstan::stan(). latent_reg_list is provided as the data, and the name of one of the .stan files to use is provided as the model argument. We also choose the number of chains and number of iterations per chain. By default the first half of each chain is discarded as warm up.

# Fit the Rasch model
fit_rasch <- irt_stan(latent_reg_list, model = "rasch_latent_reg.stan",
                      iter = 2000, chains = 4)

The stan_columns_plot() function shows convergence statistics for the fitted model. As an aside, it may also be used to show other statistics such as posterior means or numbers of effective samples.

# View convergence statistics
stan_columns_plot(fit_rasch)

print_irt_stan() provides a summary of parameter posteriors. It is a wrapper for the print() method for fitted Stan models, selecting and organizing the most interesting parameters. Labels for the items are automatically taken from the column names of the response matrix. Regarding the parameters in the output: beta refers to item difficulties, lambda refers to latent regression coefficients, and sigma is the standard deviation of the ability distribution.

# View a summary of parameter posteriors                      
print_irt_stan(fit_rasch, latent_reg_list)

A convenient way to view the Stan code for the model is the edstan_model_code() function.

edstan_model_code("rasch_latent_reg.stan")

If we prefer to fit the 2PL model with latent regression, we could call irt_stan() and select that .stan file.

# Fit the Rasch model
fit_rasch <- irt_stan(latent_reg_list, model = "2pl_latent_reg.stan",
                      iter = 2000, chains = 4)

Polytomous IRT with verbal aggression data

The second example will use the verbal aggression data to fit a polytomous item response theory model.

# Preview the data
head(aggression)

The data consists of 316 persons responding to 24 items with responses scored as 0, 1, or 2. The variable person is a person ID, item is an item ID, and poly is the scored response. In this way, the data may be said to be in “long form” as there is one row per person-item combination. The data includes two person covariates: male is a dummy variable for whether the respondent is male, and anger is the person’s trait anger raw score, a separate measure.

Next we make the data list using irt_data(). However, we’ll use a different set of arguments because the data are in long form: y takes the responses in vector form, ii takes a vector of item IDs, and jj takes a vector of person IDs. The use of covariates and formula are the same as before. For the latent regression part, we use the rescale_binary() and rescale_continuous() functions to rescale the covariates for compatibility with the prior distributions specified in edstan models.

# Make the data list
agg_list <- irt_data(
  y = aggression$poly,
  ii = aggression$description,
  jj = aggression$person,
  covariates = aggression,
  formula = ~ rescale_binary(male) * rescale_continuous(anger)
)

str(agg_list)

The generalized partial credit model is fit to the these data.

# Fit the generalized partial credit model
fit_gpcm <- irt_stan(agg_list, model = "gpcm_latent_reg.stan",
                     iter = 2000, chains = 4)

As before, a plot of convergence statistics is made.

# View convergence statistics
stan_columns_plot(fit_gpcm)

And lastly we obtain a table of posterior summaries. Here, alpha refers to discrimination parameters, beta refers to step difficulty parameters, and lambda refers to latent regression coefficients.

# View a summary of parameter posteriors    
print_irt_stan(fit_gpcm, agg_list)

Technical notes

Users will be able to fit the edstan models without full knowledge of the technical details, though these are provided in this section. All that is really needed for interpreting results is to know the meanings assigned to the Greek letters.

Notation

Variables and parameters are similar across edstan models. The variables used are:

\(i = 1 \ldots I\) indexes items.
\(j = 1 \ldots J\) indexes persons.
\(m_i\) is simultaneously the maximum score and the number of step difficulty parameters for item \(i\) for partial credit models. Alternatively, \(m\) is the same across all items for rating scale models.
\(s = 1 \ldots m_i\) or \(s = 1 \ldots m\) indexes steps within items.
\(y_{ij}\) is the scored response of person \(j\) to item \(i\). The lowest score for items must be zero (except for rating scale models).
\(w_{j}\) is the vector of covariates for person \(j\). \(w_{j}\) may be assembled into a \(J\)-by-\(K\) covariate matrix \(W\), where \(K\) is number of elements in \(w_j\).

The parameters used are:

For the Rasch and 2PL models, \(\beta_i\) is the difficulty for item \(i\). For the rating scale models, \(\beta_i\) is the mean difficulty for item \(i\). For partial credit models, \(\beta_{is}\) is the difficulty for step \(s\) of item \(i\).
\(\kappa_s\) is a step difficulty for the (generalized) rating scale model.
\(\alpha_i\) is the discrimination parameter for item \(i\) (when applicable).
\(\theta_j\) is the ability for person \(j\).
\(\lambda\) is the vector of latent regression parameters of length \(K\) (when applicable).
\(\sigma\) is standard deviation for the ability distribution (when applicable).

The .stan files and the notation for the models below closely adhere to these conventions.

Rasch family models

Rasch model

rasch_latent_reg.stan

\[ \mathrm{logit} [ \Pr(y_{ij} = 1 | \theta_j, \beta_i) ] = \theta_j - \beta_i \]

Partial credit model

pcm_latent_reg.stan

\[ \Pr(Y_{ij} = y,~y > 0 | \theta_j, \beta_i) = \frac{\exp \sum_{s=1}^y (\theta_j - \beta_{is})} {1 + \sum_{k=1}^{m_i} \exp \sum_{s=1}^k (\theta_j - \beta_{is})} \]

\[ \Pr(Y_{ij} = y,~y = 0 | \theta_j, \beta_i) = \frac{1} {1 + \sum_{k=1}^{m_i} \exp \sum_{s=1}^k (\theta_j - \beta_{is})} \]

Rating scale model

rsm_latent_reg.stan

\[ \Pr(Y_{ij} = y,~y > 0 | \theta_j, \beta_i, \kappa_s) = \frac{\exp \sum_{s=1}^y (\theta_j - \beta_i - \kappa_s)} {1 + \sum_{k=1}^{m} \exp \sum_{s=1}^k (\theta_j - \beta_i - \kappa_s)} \]

\[ \Pr(Y_{ij} = y,~y = 0 | \theta_j, \beta_i, \kappa_s) = \frac{1} {1 + \sum_{k=1}^{m} \exp \sum_{s=1}^k (\theta_j - \beta_i - \kappa_s)} \]

Models featuring discrimination parameters

Two-parameter logistic model

2pl_latent_reg.stan

\[ \mathrm{logit} [ \Pr(y_{ij} = 1 | \alpha_i, \beta_i, \theta_j) ] = \alpha_i \theta_j - \beta_i \]

Generalized partial credit model

gpcm_latent_reg.stan

\[ \Pr(Y_{ij} = y,~y > 0 | \theta_j, \alpha_i, \beta_i) = \frac{\exp \sum_{s=1}^y (\alpha_i \theta_j - \beta_{is})} {1 + \sum_{k=1}^{m_i} \exp \sum_{s=1}^k (\alpha_i \theta_j - \beta_{is})} \]

\[ \Pr(Y_{ij} = y,~y = 0 | \theta_j, \alpha_i, \beta_i) = \frac{1} {1 + \sum_{k=1}^{m_i} \exp \sum_{s=1}^k (\alpha_i \theta_j + w_{j}' \lambda - \beta_{is})} \]

Generalized rating scale model

grsm_latent_reg.stan

\[ \Pr(Y_{ij} = y,~y > 0 | \theta_j, \lambda, \alpha_i, \beta_i, \kappa_s) = \frac{\exp \sum_{s=1}^y (\alpha_i \theta_j - \beta_i - \kappa_s)} {1 + \sum_{k=1}^{m} \exp \sum_{s=1}^k (\alpha_i \theta_j - \beta_i - \kappa_s)} \]

\[ \Pr(Y_{ij} = y,~y = 0 | \theta_j, \lambda, \alpha_i, \beta_i, \kappa_s) = \frac{1} {1 + \sum_{k=1}^{m} \exp \sum_{s=1}^k (\alpha_i \theta_j - \beta_i - \kappa_s)} \]

Priors

For Rasch family models, the prior distributions for the person-related parameters are

\(\theta_j \sim \mathrm{N}(w_j' \lambda, \sigma^2)\)
\(\lambda \sim t_7(0, 2.5)\)
\(\sigma \sim \mathrm{gamma}(2, 1)\)

For models with discrimination parameters, the priors are

\(\theta_j \sim \mathrm{N}(w_j' \lambda, 1)\)
\(\lambda \sim t_7(0, 2.5)\)

In either case, the prior for \(\lambda\) is applied to centered and scaled versions of the person covariates. Specifically: (1) continuous covariates are mean-centered and then divided by two times their standard deviations, (2) binary covariates are mean-centered and divided their maximum minus minimum values, and (3) no change is made to the constant, set to one, for the model intercept. \(t_7\) is the Student’s \(t\) distribution with seven degrees of freedom.

The priors for the item parameters are

\(\alpha \sim \mathrm{lognormal}(.5, 1)\)
\(\beta \sim \mathrm{N}(0, 9)\)
\(\kappa \sim \mathrm{N}(0, 9)\)

Writing your own Stan models

It is expected that edstan users will eventually want to write their own Stan models. In this case, the edstan functions may still be useful. If the user-written models use the same conventions for naming variables and parameters as the edstan models, then the irt_data() and print_irt_stan() functions will work just fine for the user-written models. The function stan_columns_plot() works with any Stan model, and labelled_integer() may help in data preparation for Stan models in general.

A brief manual

Daniel C. Furr