bayesqm 0.2.0
CRAN release: 2026-08-20
The model changed. 0.1.0 scored the grid positions and fit a Student-t factor model through Stan. 0.2.0 models the forced Q sort as an ordered partition of the statements, computes the probability of the observed sorting event exactly, and fits it with a Gibbs sampler written in R. Stan is no longer used. Scripts written against 0.1.0 need the changes below.
Changes from 0.1.0
-
fit_bayesian()fits the new model. The Stan-era arguments give migration errors that name their replacements. - Functions keeping their names now return the new model’s quantities.
compute_loadings()gives bounded loadings with credible intervals,compute_zscores()scores with intervals from the aligned draws,compute_factor_array()a quota-exact array built from posterior column probabilities, andmatchalign()takes and returns a whole fit. - Removed with the score-scale model, each erroring with a pointer to its successor:
run_bayes(),select_k_peak(),select_k_sivula(),compute_dominant_prob(),compute_dominant_sign(),compute_threshold_prob(),classify_membership(),compute_divergence(),critical_delta(),compute_posterior_scalars(),demo_run(), andplot_hyper().suggest_delta()is replaced bydelta_grid(), andresiduals()on a fit errors because a rank likelihood has no residual scale. - The fit object was rebuilt. The 0.1.0 slots are gone, and every table now comes from its own function.
New
- One rule behind every claim.
claims()selects the reported flags, distinguishing statements, consensus statements, and pairwise stars at a posterior false-discovery level and gives the expected number of false claims. - A two-part choice of the number of factors.
fit_ladder()fits the candidate models andselect_k()checks each for adequacy and for per-factor support, refusing to select when no K passes both.loo_ladder()adds PSIS-LOO as directional corroboration. - Model checks.
check_fit()runs the posterior-predictive checks andcheck_persons()separates sorts the model spans from shared viewpoints it does not. Every fit passes a convergence check before it returns, andextend()continues a chain draw for draw. - The tables.
compute_flags()reports flag probabilities with an explicit unclassified state,compute_qdc()judges statements against a posterior critical difference and the one-grid-column consensus region with two-level stars, andfactor_characteristics()gives the per-factor summary block. - Alignment ends in a stabilized orientation, so the delivered rotation does not inherit one draw’s sampling noise.
- Two real panels ship with the package.
obesity_sortsis the childhood obesity panel of Akhtar-Danesh (2023), 33 participants sorting 42 statements on a nine-column grid, andgrizzly_sortsthe grizzly bear reintroduction panel of Easter et al. (2025), 67 participants sorting 41 statements on an eleven-column grid. - Plots rebuilt in one style, from
plot_sorts(), which previews every participant’s completed sort before any model runs, toplot_choice_k(), which draws the whole choice-of-K decision with its verdict written on the rows.plot_zscores(),plot_statement(),plot_factor_array(),plot_contrasts(),plot_flags(),plot_loading_posterior(),plot_person_check(),plot_convergence(), andplot_ppc()cover the remaining views, withggplot2::autoplot()methods for loadings, flags, contrasts, and the array. -
rotate_factors()applies judgmental rotation to every draw,flip_factor()reverses a pole, andrename_factors()relabels every table at once. - Data import is unchanged, and grids with any distinct printed labels are accepted as they are.
- A second vignette, the output codebook, defines every column of every reported table, and the website gains three articles, the validation checks, the decision rules, and a start-to-finish walkthrough of the obesity panel.