Skip to content

PACER 3.2.0

Latest
Latest

Choose a tag to compare

View all tags
@hdoran hdoran released this 13 Oct 13:16
V.3.2.0
6c6b6c3

PACER 3.2.0 — Nexus

Six cognitive diagnostic models on one engine. This release adds the three classic additive CDMs — A-CDM, LLM and R-RUM — alongside G-DINA, DINA and DINO, in calibration, in the Skill Profiler, and from Claude Desktop. Estimates were verified against the GDINA R package.

What's new

A-CDM, LLM and R-RUM

  • The additive models of de la Torre (2011): G-DINA with the interaction terms removed, so each required skill adds a fixed main effect on the probability scale (A-CDM, identity link), the log-odds scale (LLM, the linear logistic model, logit link) or the log scale (R-RUM, log link). An item requiring k skills has k + 1 parameters instead of 2^k.
  • Fitted by the same multi-start EM engine as G-DINA with a Newton (Fisher-scoring) M-step. One-skill items use the closed form.
  • Every item still reports pattern probabilities, δ, guess, slip, IDI, the monotonicity check, and the fit indices (deviance, parameters, AIC, BIC, CAIC, AICc, SABIC), so all six models can be compared on the same data.

A model chooser that explains the link

  • The Model option is now a chooser in three families — Saturated (G-DINA), Conjunctive / disjunctive (DINA, DINO), Additive (A-CDM, LLM, R-RUM) — with a one-line description of the selected model.
  • For G-DINA, DINA and DINO the Link Function stays a reporting scale for δ. For the additive models it is part of the model: the selector locks and says fixed by the model.
  • The item table shows the model and its link; the item plot charts P(correct) by attribute pattern for every model.

Enforce monotonicity

  • New Estimation Options checkbox (off by default; the R package's mono.constraint) that constrains every main effect of an additive model to be non-negative, so P(correct) never falls when a skill is mastered. Other models keep the monotonicity check.

Skill Profiler

  • Profiles examinees under all six models. The pattern-level parameter file gains a model column (G-DINA when absent) and the profiler locks its model to the loaded calibration or to the file.
  • Unknown model names are reported instead of silently becoming DINA.

Claude Desktop (MCP) and the AI Advisor

  • calibrate_cdm accepts ACDM, LLM and RRUM plus mono_constraint, rejects a link that conflicts with an additive model, and writes the model column; score_cdm reads it back and locks to that model. MCP server 0.4.0.
  • The AI Advisor's knowledge base covers the additive models, the two roles of the link function, the Newton M-step and the monotonicity constraint.

Learn it

  • CDM Calibration and Skill Profiler guided tours describe the six models in all eight languages.
  • PACER University's CDM module adds Additive Models: When the Link Is the Model with a six-model fit comparison.
  • The technical reference documents the additive models, their design matrix and the constrained Newton M-step.

Fixes

  • /api/cdm/calibrate and /api/cdm/score now return a clear 400 error for an unknown method or link name instead of silently using DINA and the identity link.

Verification

On a 10-item, 3-skill, 1,000-examinee dataset, PACER and GDINA 2.13.1 agree on deviance to within 0.01 for A-CDM, LLM and R-RUM (and to 0.02 with the monotonicity constraint); pattern probabilities agree to three decimals. G-DINA, DINA and DINO results are unchanged from 3.1.0.

Upgrading

  • Desktop users: after installing, open Settings → Connect to Claude Desktop again (the config path includes the version number), then restart Claude Desktop.
  • Pattern-level parameter files exported by 3.1.0 (no model column) still import and score as G-DINA.

Downloads: PACERSetup.exe (Windows) · PACER.dmg (macOS, Apple Silicon and Intel) · web app at paceronline.com

Assets 4
Loading