Changelog • DPMirt

DPMirt 0.2.0

DPMirt 0.2.0 is the overhaul release. It expands the package from the initial development release into a release-candidate implementation with a clearer model contract, richer posterior-draw metadata, broader diagnostics, updated vignettes, and a rebuilt pkgdown site.

Models

Rasch, 2PL, and 3PL item response theory models.
Parametric (Normal) and semiparametric (Dirichlet Process Mixture) latent trait priors.
IRT and slope-intercept (SI) parameterizations for 2PL and 3PL.
Three identification strategies: constrained_item (Rasch default), constrained_ability, and unconstrained with post-hoc rescaling (2PL/3PL default).
3PL models include item guessing parameters (delta), with posterior summaries, rescaled draw extraction, and plots where applicable.
Some identification combinations are intentionally unavailable: constrained_item is rejected for 3PL models, and constrained_ability is rejected for DPM models.

Workflow

dpmirt() — all-in-one model fitting (spec → compile → sample → rescale).
Modular pipeline: dpmirt_spec() → dpmirt_compile() → dpmirt_sample() → dpmirt_rescale().
dpmirt_resume() — continue sampling without recompilation while the compiled NIMBLE object is live in the current R session. Saved RDS objects cannot restore compiled pointers.
sampler_config accepts an advanced function hook for NIMBLE sampler customization. List-based sampler schemas are reserved.
item_priors is reserved; use NULL or list() for DPMirt’s fixed item priors.

Posterior Summaries

Three estimators via dpmirt_estimates():
- PM (Posterior Mean) — optimal individual-level MSE.
- CB (Constrained Bayes; Ghosh, 1992) — moment-matched to the posterior predictive distribution.
- GR (Triple-Goal; Shen & Louis, 1998) — simultaneous estimation, ranking, and distributional recovery.
dpmirt_draws() — extract posterior samples in matrix or long format. Current extraction is for rescaled draws only; raw draw extraction is reserved.
dpmirt_loss() — evaluate MSEL, MSELR, KS, and custom loss metrics.

Diagnostics

dpmirt_fit objects carry schema_version, chain_info, draw_index, and run_history metadata.
dpmirt_diagnostics() — ESS, optional chain-aware R-hat, trace summaries, WAIC provenance, timing, and DPM cluster diagnostics.
dpmirt_compare() — WAIC-based model comparison across fits, including waic_aggregation provenance.
dpmirt_dp_density() — posterior density estimation from the DP mixture. This can be computationally expensive because it reconstructs posterior DP measure samples through NIMBLE’s public modelValues and getSamplesDPmeasure() workflow.
save_draws = FALSE and save_path are reserved; DPMirt currently returns in-memory draw-retaining fit objects.

Simulation

dpmirt_simulate() — generate IRT data with Rasch, 2PL, and 3PL response models. 3PL fallback simulation includes Beta-distributed guessing parameters (delta) and does not use target_rho.
IRTsimrel integration for Rasch/2PL reliability-targeted simulation via EQC calibration by default, with SAC calibration for MSEM targets; 3PL uses DPMirt’s internal fallback simulator.

Prior Elicitation

dpmirt_alpha_prior() — principled concentration-parameter selection via DPprior, with graceful fallback to Gamma(1, 3).

Visualization

S3 plot() methods for dpmirt_fit, dpmirt_estimates, and dpmirt_sim objects.
12 standalone ggplot2 functions: dpmirt_plot_trace(), dpmirt_plot_density(), dpmirt_plot_caterpillar(), dpmirt_plot_items(), dpmirt_plot_icc(), dpmirt_plot_info(), dpmirt_plot_dp_density(), dpmirt_plot_clusters(), dpmirt_plot_wright_map(), dpmirt_plot_parameter_trace(), dpmirt_plot_density_compare(), dpmirt_plot_pp_check().

Documentation

Eight vignettes covering quick start, models and workflow, posterior summaries, prior elicitation, simulation studies, mathematical foundations, and NIMBLE internals.
Precomputed vignette fixtures follow an 11 .rds file compact fixture contract in inst/extdata, plus an inst/extdata/README.md note. Fit fixtures now ship as xz-compressed, evenly thinned 250-draw dpmirt_fit objects with single-chain chain-aware-v1 metadata and plot-ready DP-density summaries; session-bound compiled/raw NIMBLE storage is not included.
31 manual pages generated with roxygen2 and including the current reserved interfaces and diagnostic contracts.

Compatibility Notes

New and upgraded dpmirt_fit objects use the chain-aware-v1 draw-storage schema. Fits created by older development versions may not contain schema_version, chain_info, draw_index, or run_history metadata.
R-hat is reported only when at least two labeled chain/run streams are available. Single-chain fits should be assessed with ESS, trace summaries, posterior predictive checks, and substantive model diagnostics.
dpmirt_resume() requires a live compiled NIMBLE object in the current R session. Saved RDS files cannot restore compiled external pointers.
dpmirt_draws() currently returns retained rescaled posterior draws. Raw/unrescaled draw extraction and disk-backed draw storage are reserved for a future release.
DP density recomputation no longer uses private NIMBLE namespace helpers. It loads retained DP draws through public modelValues accessors before calling getSamplesDPmeasure().
Transformed-scale DP density output for 2PL/3PL IRT and SI fits should be interpreted as a diagnostic summary. Full scale/Jacobian-adjusted density reconstruction remains a documented future enhancement.

DPMirt 0.1.0

Initial development release.