Computerized Adaptive Testing with Cognitive Diagnostic Models

cdCAT is an R package that provides a session-based engine for Cognitive Diagnostic Computerized Adaptive Testing (CD-CAT). Unlike simulation-focused packages, cdCAT is designed for real-time, item-by-item adaptive applications – making it suitable for integration with APIs, Shiny apps, and online assessment platforms.

Theoretical Framework

cdCAT implements adaptive testing under Cognitive Diagnosis Models (CDMs).

The adaptive cycle follows the standard CD-CAT framework:

1. Posterior estimation of latent attribute profiles
2. Item selection based on information criteria
3. Response update
4. Stopping rule evaluation

Implemented criteria are grounded in the CD-CAT literature:

• Xu, Chang & Douglas (2003) – KL
• Cheng (2009) – PWKL / MPWKL
• Shannon (1948) – Entropy
• Tatsuoka (2002) – Attribute-level classification
• Hsu, Wang & Chen (2013) – Dual posterior threshold
• Kingsbury & Zara (1991) – Content balancing
• Sympson & Hetter (1985) – Exposure control
• van der Linden (2005) – Shadow CAT

This ensures theoretical consistency with established CD-CAT methodology.

Features

• Models: DINA, DINO, and GDINA
• Item selection criteria: KL, PWKL, MPWKL, SHE, sequential, random
• Estimation methods: MAP, MLE, EAP
• Custom prior: informative prior over skill profiles via cdcat_prior()
• Fixed or adaptive start item: control the first item via start_item
• Content balancing: blueprint-driven domain proportions (Kingsbury & Zara, 1991)
• Exposure control: Sympson-Hetter and Randomesque methods
• Shadow CAT: user-defined constraint functions for arbitrary test assembly
• Stopping rules: maximum items, posterior threshold classification
• Built-in simulated datasets: cdcat_sim — 16 datasets across K, model, and item quality
• Session-based architecture: stateful R6 object manages the full CAT lifecycle
• Lightweight: depends only on R6

Architecture

cdCAT is built around a stateful CdcatSession R6 object that:

• Stores the full posterior distribution
• Tracks administered items
• Updates responses sequentially
• Evaluates stopping criteria at each step

The item selection pipeline at each step is:

Content balancing filter  ->  Criterion scores  ->  Exposure control  ->  Selected item
                                    |
                          (Shadow CAT replaces the full pipeline
                           when constr_fun is supplied)

This design makes the package suitable for:

• REST APIs
• Shiny applications
• Online assessment engines
• Real-time adaptive testing systems

Installation

# Development version from GitHub
remotes::install_github("thiagofmiranda/cdCAT")

Quick Start

library(cdCAT)
# 1. Define the item bank
Q <- matrix(c(
  1, 0,
  0, 1,
  1, 0,
  0, 1,
  1, 1
), nrow = 5, ncol = 2, byrow = TRUE)
items <- cdcat_items(
  q_matrix = Q,
  model    = "DINA",
  slip     = c(0.10, 0.10, 0.15, 0.10, 0.10),
  guess    = c(0.20, 0.20, 0.15, 0.20, 0.15)
)
# 2. Start a session
session <- CdcatSession$new(
  items     = items,
  criterion = "PWKL",
  max_items = 5L,
  threshold = 0.8
)
# 3. Administer items
simulated_responses <- c(1, 1, 0, 1, 0)
repeat {
  item <- session$next_item()
  if (item == 0) break
  session$update(item, simulated_responses[item])
}
# 4. Results
res <- session$result()
cat("Estimated profile:", res$alpha_hat, "\n")
#> Estimated profile: 1 1
cat("Items used       :", res$n_items, "\n")
#> Items used       : 2
cat("Stop reason      :", res$stop_reason, "\n")
#> Stop reason      : single threshold reached

Item Selection Criteria

All adaptive criteria operate on the conditional response distribution P(X_j = 1 | alpha) and use one of the following strategies:

Supported Models

Advanced Features

Custom Prior

Specify an informative prior over skill profiles using cdcat_prior(). Pattern names follow expand.grid order (first attribute varies fastest).

prior <- cdcat_prior(
  "00" = 0.4,
  "10" = 0.3,
  "01" = 0.2,
  "11" = 0.1
)
session <- CdcatSession$new(items, prior = prior)

Content Balancing

Enforce a test blueprint by restricting each selection step to the most under-represented content domain (Kingsbury & Zara, 1991).

content      <- c("Algebra", "Algebra", "Geometry", "Geometry", "Mixed")
content_prop <- c(Algebra = 0.4, Geometry = 0.4, Mixed = 0.2)
session <- CdcatSession$new(
  items        = items,
  content      = content,
  content_prop = content_prop
)

Exposure Control

Limit overuse of specific items with two detection-free methods:

# Sympson-Hetter: per-item acceptance probabilities in [0, 1]
exposure_sh <- c(0.9, 0.9, 0.9, 0.5, 0.9)   # item 4 accepted only 50% of the time
# Randomesque: top-n pool size per test position (all values >= 1)
exposure_rq <- c(3L, 2L, 2L, 1L, 1L)         # position 1 draws from top-3
session <- CdcatSession$new(items, exposure = exposure_sh)

Fixed or Adaptive Start Item

Control which item is administered first without touching the main adaptive pipeline. Accepts a fixed index, "random", "seq", or any adaptive criterion name.

# Always start with item 3
session <- CdcatSession$new(items, start_item = 3L)
# Draw the first item randomly
session <- CdcatSession$new(items, start_item = "random")
# Use KL for the first item, then switch to PWKL
session <- CdcatSession$new(items, criterion = "PWKL", start_item = "KL")

Shadow CAT

Enforce arbitrary test assembly constraints via a user-supplied function. cdCAT is solver-agnostic: use any optimisation library inside constr_fun.

# constr_fun receives (scores, items, administered) and returns one item index
greedy_shadow <- function(scores, items, administered) {
  scores[administered] <- -Inf
  which.max(scores)
}
session <- CdcatSession$new(items, constr_fun = greedy_shadow)

See vignette("advanced-item-selection") for a full LP-based example using lpSolve.

Built-in Simulated Data

cdcat_sim provides 16 ready-to-use datasets covering K ∈ {3, 5} attributes, DINA and GDINA models, and four levels of item quality (HD-LV, HD-HV, LD-LV, LD-HV). Each dataset contains the Q-matrix, true attribute profiles, item parameters, and simulated responses.

# List all available datasets
names(cdcat_sim)
# Use the K=3 GDINA high-discrimination dataset
d     <- cdcat_sim[["k3_GDINA_HDLV"]]
items <- cdcat_items(d$Q, "GDINA", gdina_params = d$parameters)

See vignette("simulation-study") for a worked simulation experiment comparing all item selection criteria and stopping rules.

How cdCAT Differs from Existing Packages

Unlike simulation-oriented packages such as mirtCAT or GDINA, cdCAT focuses on:

• Real-time session management
• Modular adaptive components (prior, content, exposure, shadow independently composable)
• Lightweight dependency structure (only R6)
• Explicit posterior tracking at every step
• Solver-agnostic shadow CAT (no lpSolve dependency in the package core)

Roadmap

Planned extensions include:

• Polytomous extensions
• pkgdown documentation site

Authors

• Thiago Miranda – thiagofm.pa@gmail.com
• Bruno Silvestre – bruno.ap.silvestre@gmail.com

License

MIT