Changelog

Source: NEWS.md

rasch 1.13.3

Sixteenth review round: two identification edge cases beneath the 1.13.2 checks, and one policy refinement. Development branch only; the CRAN submission of 1.11.7 is untouched.

• Connectivity now counts only INFORMATIVE response pairs. A pair whose every observed total is 0 or the maximum has a single feasible conditional allocation and carries no information, so one respondent scoring (0, 0) across two otherwise disjoint item blocks no longer “connects” them (every intermediate total has at least two allocations, so any response off the extreme-total corners is a real link). Applied to the item-pair graph of rasch()/pcml() and to the MFRM co-observation components.
• As the review prescribed, the graph test is now backed by a rank/conditioning check of the projected information at the solution in the shared solver. This also catches what no graph can see: a block bridged by a SINGLE informative response is perfect separation in the conditional pair logit – the pair estimate runs to the boundary and the information vanishes there – and previously returned converged fits with zero standard errors.
• The Ford (1957) check now respects anchors. Strong connectivity is the existence condition for the FREE model only; an anchored fit requires each free object to be tied to an anchor in both win directions (reachability to and from the anchor set over the points digraph – the constrained recession-direction condition). Two anchored components joined by a one-directional edge between their anchors is now correctly accepted, while removing one anchor still errors.
• EFRM: the practical weak-identification cutoff (SE of log phi above
  5. is now a loud warning plus a note with the estimates RETAINED, supporting reproducible sensitivity analyses; the error is reserved for structural non-identification (a flat direction of the joint information), where no estimate exists to retain.

rasch 1.13.2

Fifteenth review round: the 1.13.1 identification guards were the right idea executed too weakly (or, in one case, too strongly), and this release replaces heuristics with information-based checks throughout. Development branch only; the CRAN submission of 1.11.7 is untouched.

• EFRM: the threshold-spread heuristic is gone. Group units are now checked on the joint information itself – a flat direction loading on a unit, or a unit whose analytic SE exceeds 5 log-units (uncertain beyond a factor of ~150), is refused with an error naming the group, since every common-unit quantity would silently depend on it. Weakly identified units with real spread are KEPT with their honest large SEs; the previous heuristic wrongly replaced some of them with NA.
• MFRM: full column rank of the structural design is necessary but not sufficient. The fit now also checks that no between-block shift of person-disjoint response blocks is expressible by the facet map (a column-space intersection test on the co-observation components): designs such as two person groups answering disjoint item pairs – or one rater per person – previously returned converged fits whose between-block contrasts were ridge artefacts, and now error with the blocks named. A test fixture with exactly this nested-rater flaw was corrected accordingly.
• btl_efrm stage 1: the per-set (locations, panel-ratio) information is now checked for rank at the solution; panels observing disjoint object pairs leave the ratio underdetermined, and such sets are screened out of the panel-unit reconciliation (refit at the reconciled units) or, when no ratio information remains anywhere, the fit stops.
• btl_efrm stage 2: rank failures are now CLASSIFIED by where the information fails. A flat direction confined to a set’s log-unit – its within-set locations are indistinguishable, so the unit has nothing to scale – refits with that unit fixed at the conventional 1 and reports alpha as NA with a note; the set’s origin kappa and its objects’ placements remain identified and keep valid SEs (this is the GermanParties2009 near-even-set pattern, which a blanket refusal would have broken). A flat direction on an origin means the set cannot be placed at all and is an error. Bootstrap replicates treat an expected NA (the declared unit) as normal and any unexpected rank failure as a failed replicate.
• BTL: a Ford (1957) violation is now an error before estimation, not a post-fit warning – a separated cluster has no finite maximum likelihood locations, and the previous warning still presented the optimiser’s boundary values with converged = TRUE.
• EFRM input: data-frame factors columns are removed from the item matrix (a numeric factor column could previously become an implicit (rest) item).
• MFRM input: rows dropped for missing person/item/facet identifiers now drop from factors too (both the column-name and data-frame forms), instead of failing with an internal length error.
• Test honesty: the wide-MFRM factor-equivalence test compared a nonexistent $table component (NULL == NULL, vacuously true); it now compares the $summary F and p columns, on a properly linked design – the equivalence itself was verified to hold.

rasch 1.13.1

Fourteenth review round: three defects in the 1.13.0 additions and the review’s nine-item identification backlog. Development branch only; the CRAN submission of 1.11.7 is untouched.

Defects in the new code:

• EFRM crossed-frame labels: the solver orders cells as sorted strings while the frame construction used factor-crossing order, so with factor labels that sort differently the per-cell units were assigned to the wrong cells (a pure-noise 2x2 gave a spurious region effect at p = 2e-5). Cells are now matched by label throughout.
• MFRM dif_size() no longer lets facet severity leak into a between-group magnitude: the joint split now pools over the COMMON observed cells with common precision weights, so severity differences cancel by construction. A design that previously manufactured a -1.75-logit phantom effect (z = -6.5) now reports -0.11 (ns) while a planted 0.7 effect is still recovered.
• phi_factorial was a diagonally-weighted fit with coefficient-wise z tests; it is now a generalised least-squares decomposition on the JOINT covariance of the cell log-units (bootstrap draws when available, otherwise the solver’s centred covariance, inverted spectrally along its identified directions, intercept excluded as inestimable under the centring), with multi-degree-of-freedom Wald tests per term in phi_factorial_tests. A planted 1.5x region unit ratio is recovered at 1.52 with p = 2e-6 and null factors stay null.

Identification and input-handling backlog:

• btl() now checks Ford’s (1957) existence condition: when the directed win graph is not strongly connected the maximum-likelihood locations diverge, and the fit warns and notes the separated objects instead of presenting the optimiser’s boundary values as measures.
• MFRM: wide input now carries factors= through the internal melt (character columns are replicated like facets, data-frame factors row-wise); duplicate person-by-item-by-facet responses are an error (keeping the first made results depend on row order); and a structural rank check on the facet design errors when facet levels are confounded with the items instead of returning a valid-looking but unidentified decomposition.
• EFRM: a group whose frames show no threshold spread beyond estimation noise has nothing for its unit to scale – previously the optimiser returned phi near 1 with a healthy-looking SE regardless of the true unit; such units are now reported NA with a note. The guard is signal-to-noise based (spread under 1.5x the pooled threshold SE in every set the group answers) and leaves even modest real spreads untouched.
• EFRM: under se_method = "bootstrap" the frame-level se_log_rho now comes from the joint replicate draws of log(alpha) + log(phi), capturing cross-stage dependence; the hybrid fallback still combines the stagewise errors as if uncorrelated, as documented.
• btl_efrm(): stage 2 checks the conditioning of its observed information; when the cross-set comparisons barely identify the set units (reciprocal condition number below 1e-10), the conditional SEs for alpha, kappa and the linked values are withheld with a note rather than reported from a near-singular inverse.
• btl_efrm() frame table: rho is now phi / alpha as documented (the within-set logit is phi (beta_a - beta_b) with beta = (v - kappa) / alpha), not phi * alpha.
• rasch_efrm() validates item_sets (named list, no overlapping or unknown items); rasch() removes data-frame factor columns from the response matrix even when items is not given, so a numeric factor can no longer be silently treated as an item.

rasch 1.13.0

The three capability extensions from the review wishlist. Development branch only; the CRAN submission of 1.11.7 is untouched.

• MFRM DIF pools to the underlying items by default: dif_anova() on an MFRM fit now answers “does item A show DIF” by pooling residuals over each item’s facet cells (pool_facets = FALSE keeps the per-virtual-cell tests). With few items, the compensating artificial DIF the plant creates on the remaining items is visible – as it should be – with the planted item carrying the dominant F.
• dif_size() works on underlying MFRM items: every facet cell of the item is split by the groups in one joint unstructured refit of the virtual matrix (the facet decomposition is not reimposed, which the note states), and per-level locations are precision-weighted means over the cells with the full covariance carried. A planted 0.7-logit sex effect is recovered at 0.82 (SE 0.17), and dif_anova(mf, sizes = TRUE) now yields magnitude rows.
• rasch_efrm() accepts several frame-defining factors: the frames are their crossed cells, per-cell units appear in phi_table, and phi_factorial reports a factorial decomposition of the cell units (sum-coded mains, plus the interaction when every cell is observed), weighted by the cells’ unit precisions with the independence approximation stated. All frame-defining factors (cell and components) are excluded from DIF testing; misspelled group and factor column names now error instead of silently falling through.

rasch 1.12.3

Thirteenth review round: the BTL, EFRM, and MFRM analogues of the DIF and identification work. Development branch only; the CRAN submission of 1.11.7 is untouched.

• btl_dif() runs on the same order-invariant machinery as the person DIF ANOVA: between-judge terms by Type II sums of squares on band-centred judge margins, band-crossing terms through orthonormal contrasts with the Greenhouse-Geisser correction. Verified exactly order-invariant across correlated judge factors; the tuned planted detections are unchanged.
• btl_dif() rejects factors that vary within a judge (the judge-level analysis would silently take whichever row came first).
• btl_efrm() validates within-set object-graph connectivity (relative locations inside a disconnected set are unidentified, exactly as in pcml()) and alpha identification (cross-set comparisons must touch at least two objects of every non-reference set; one object identifies only the origin).
• rasch_efrm() group linkage requires at least two SHARED ITEMS of a common set between groups: sharing a set label with disjoint item subsets left the unit ratio unidentified but was accepted.
• dif_anova() works on EFRM fits: the frame group is excluded (it is the frame structure – every frame has its own virtual items – so it has no within-item contrast) with a note, and other person factors are tested per virtual item; nominating only the frame group is an informative error.
• rasch_mfrm() gains factors= (person factors, constant within person, validated – a facet passed as a person factor errors with a pointer to interaction=), so dif_anova() works on MFRM fits directly; unavailable DIF magnitudes are reported in notes instead of silently returning nothing.
• The frames table’s se_log_rho independence approximation (log alpha and log phi estimated in different stages; cross-stage covariance unavailable) is stated in the documentation.

rasch 1.12.2

Twelfth review round: the incomplete-panel edges of the DIF engine. Development branch only; the CRAN submission of 1.11.7 is untouched.

• Between-person margins are centred within each within-cell BY CLASS INTERVAL (falling back to the cell mean where a combination is empty): cell-only centring removed common occasion effects but not occasion-by-trait structure, which with differential missingness masqueraded as non-uniform group DIF at F = 214.7; it now reads F = 0.01 with the occasion terms intact.
• A between level that loses every complete within panel (for example a group observed only at baseline) no longer crashes lm() with a one-level factor: its within-stratum interactions are reported NA with an explanatory note, and every other test proceeds.
• Tukey candidates are filtered to between-only terms, so a significant multilevel within term can no longer reach TukeyHSD and crash on its empty result.
• Transparency: the count of person panels dropped from the within-person tests is returned in notes, alongside the non-estimable-term note.

rasch 1.12.1

Eleventh review round on the rebuilt DIF engine: five findings, all verified and closed. Development branch only; the CRAN submission of 1.11.7 is untouched.

• Multi-within contrast alignment. interaction() orders cells with the FIRST factor fastest while the Kronecker contrast matrices assume the LAST fastest; with two within factors a pure w1 effect therefore loaded onto w1:w2 (F 7.4 on the interaction, 3.2 on w1). The within cells are now built in Kronecker order; the same effect reads F = 16.9 on w1 and 0.35 on the interaction.
• Comparable between-person margins. Person means for the between tests are computed on within-cell-centred residuals: with a common occasion effect and one group missing 80% of an occasion, raw means reported group DIF at F = 37.6 (p = 2e-9); centred margins give F = 0.29 while the occasion effect stays on the occasion term.
• Tukey follow-ups in mixed designs run on the person-level between-factor aov (one row per person); ordinary Tukey on repeated within cells is no longer offered.
• An NA in a between factor no longer flips it to within-subject (detection counts distinct non-missing values).
• The Greenhouse-Geisser correction is fully reported: terms carries df_denom and gg_epsilon, the p-value is exactly pf(F, epsdf, epsdf_denom), and the stale “sums of squares are sequential” sentence is replaced by the Type II description.

rasch 1.12.0

The multi-factor DIF ANOVA engine rebuilt for statistical validity (tenth review round; three P1 findings, every one reproduced before fixing). This release is on the development branch only – it does NOT supersede the CRAN submission of 1.11.7.

• Order-invariant tests. Between-person terms now use Type II sums of squares on person-level residual means: every term is adjusted for every term not containing it, with the class interval always among them. The old sequential (Type I) tests with factors entered first let entry order decide which of two correlated factors flagged – in the reproduction, reversing the order LOST a planted 0.8-logit DIF entirely – and let a group factor absorb pure trait variance (reported p = 2e-25 where the adjusted test gives p = 0.83). Verified exactly order-invariant.
• Persons are the units whenever ids repeat. Residuals aggregate to one mean per person (per within cell) before any test, and the automatic class-interval rule counts persons, not rows: exactly duplicating every person previously changed F = 4.06 (unflagged) to F = 8.21 (flagged); it now changes nothing (verified to 2e-13).
• Within-person terms are tested on person-by-cell means through orthonormal contrasts with the Greenhouse-Geisser epsilon (Maxwell & Delaney 2004), replacing raw multi-stratum aov: a nonspherical four-level null rejected ~9% at nominal 5% before (and could crash on tied person means – also fixed, in the interval allocator); it now runs at 2.8%, with the spherical null at 6.1% and planted within-occasion DIF still detected. Persons missing a within cell are dropped from the within tests explicitly rather than projected murkily. The two-level case still agrees with its paired-t gold standard.
• Input honesty: unknown within names error; declaring a factor within-subject with no repeated ids errors; and a factor that VARIES within persons can no longer be forced between-subjects (the old row-level treatment pseudo-replicated).
• The BH family choice (per term, across items) is now documented explicitly.

rasch 1.11.7

Ninth review round: one finding.

• print() distinguishes the three saved-fit dependence schemas by their df field: current fits label the statistic t; 1.11.4 transitional fits (z name, df present, t-based p) print as t; older fits (z, no df, normal-reference p) KEEP the z label – relabelling them t misrepresented how their p-values were computed (a legacy z = 3.17 with normal p = .002 is not the same claim as t = 3.17 with p = .016 at five clusters). The regression test now reconstructs both legacy schemas faithfully.

rasch 1.11.6

Eighth review round: label and migration polish.

• print() reads the dependence statistic from either the current t column or the pre-1.11.5 z column, so fits saved before the rename print completely; the printed label and the rank-deficiency note both say t, matching the t(G - 1) reference. No duplicate z alias column is added: the package has never had a released version with the old name, the rename is documented here, and old saved fits are read transparently.

rasch 1.11.5

Seventh review round: three completions of round-six fixes, one label.

• margin= gets the same explicit-ordering rule as response=: an ORDERED factor (smallest to largest margin) or a numeric magnitude; plain factors AND character columns are refused, since alphabetical order could silently reverse which margin counts as the big win.
• Invalid graded numeric responses (“abc”, Inf) error through the shared guard instead of becoming missing and being dropped – the fit no longer succeeds on silently reduced data.
• The simulator’s item resolver rejects fractional indices (4.9 no longer truncates to item 4), an empty second_dim$items, and a vector-valued rho.
• The clustered dependence/position statistic is labelled t, matching its t(G - 1) reference (at infinite df it is the familiar z). The reviewer’s independent 400-run null simulation of the cluster-t correction found 5.25% rejection at nominal 5%.

rasch 1.11.4

Sixth review round: four findings, all verified and closed.

• The integer-score guard reads factors through their LABELS: as.numeric(factor) returns level codes, so a factor score of “1.9” slipped past as the integer 3 and was then truncated. Non-numeric and non-finite scores now also error at every entry point instead of silently becoming missing.
• Graded btl() requires an ORDERED response factor: a plain factor’s alphabetical level order silently defined – and could reverse – the response scale (worse < same < better read as better < same < worse). Integer scores 0..m remain the alternative.
• Clustered dependence and position tests use a t reference with G - 1 degrees of freedom (the standard few-cluster correction) instead of normal theory; the table carries the df. Five judges now give honestly wide p-values.
• simulate_rasch() validates the second-dimension specification: rho outside [-1, 1] and unknown item names/indices are errors (rho = 1.2 used to produce an all-missing second dimension that looked like a successful simulation), and the item-set resolver now range-checks every nominated item list centrally.

rasch 1.11.3

Fifth review round: seven findings, all verified and closed.

• Every public estimator now rejects fractional scores before integer coercion could truncate them: pcml(), pcml_pc(), and rasch_mfrm() truncated 1.9 to 1 silently, and graded btl() rounded it; only rasch() rejected. One shared guard serves all entry points.
• item_moments() is overflow-stable (log-sum-exp): a 31-category item, or wide categories with a large discrimination, overflowed the uncentred exponent and turned every person_wle() location NA.
• When the judge-clustered covariance is rank-deficient (clusters <= parameters) the OSI is withheld as NA – understated SEs would overstate the separation reliability – and the note now spells out how to read the marginal SEs and dependence tests. The SEs themselves stay reported: they are consistent estimates whose understatement is disclosed, which is standard few-cluster practice.
• btl_efrm() rejects a judge assigned to more than one panel (a panel is a judge attribute; the judge bootstrap would otherwise silently reclassify rows).
• Undefined diagnostics are NA, not numbers: Cronbach’s alpha under zero total-score variance (was -Inf), and the omnibus item-trait test when no item is testable (was chi-square 0 on 0 df with p = 1).
• simulate_rasch()’s secondary trait keeps the requested mean and sd (combining two mean-mu components had shifted the mean by the factor rho + sqrt(1 - rho^2)) and is returned in the truth metadata.
• The MLE score-table root drops the common discrimination, which cancels from the ML equation exactly as it does from Warm’s (ordinary unit-discrimination fits were unaffected).

rasch 1.11.2

Edge cases from the fourth review round, verified and closed:

• The BTL dimensionality reference now simulates at the PAIR level: count weights are summed per unordered pair and rounded, so half-tie rows (weight 0.5 in each direction) recombine into whole comparisons. Row-level simulation broke on fractional weights (as.integer(0.5) is a zero binomial size – every reference draw degenerated to 0). The per-replicate leading strengths are exposed as $reference$draws.
• btl_efrm(se_method = "judge_bootstrap") refuses a panel with a single judge (resampling one judge returns the same data every time, so its SEs would be a spurious zero) and notes panels with fewer than five judges as rough.
• btl() notes when there are no more judge clusters than parameters: the clustered covariance is then rank-deficient by construction and no scalar (CR1) correction repairs it – the SEs likely understate.

rasch 1.11.1

The follow-up review’s residual findings, all verified and closed:

• The BTL dimensionality reference now simulates binomial (multinomial) sums for count-weighted rows instead of one weighted Bernoulli, whose w^2 p(1-p) variance overdispersed the reference and made the test conservative on aggregated data; aggregated and expanded data now give the same reference (verified to < 1%).
• The pairwise chi-square degrees of freedom subtract every estimated parameter (position and dependence covariates were omitted), and a design with nothing testable reports NA instead of df = 1.
• Judge-clustered covariance carries the standard CR1 small-sample factor G/(G-1), on top of the single-judge error and few-cluster note.
• btl_efrm(se_method = "judge_bootstrap"): judges resampled with replacement within panels and the whole pipeline refitted, so the errors carry within-judge dependence as well as stage-one uncertainty – closing the documented gap in the parametric bootstrap. On model-true data it agrees with the parametric bootstrap; estimates are unchanged by construction.
• The next-pair priority documentation is internally consistent (ranking heuristic; the Sherman-Morrison update is exact for information matrices, a scoring device on the sandwich).

rasch 1.11.0

Statistical validity, from a second external review. Every finding was re-derived or reproduced by simulation before fixing; the fixes are verified by new calibration tests (test-statistical-validity.R).

• Recentred covariance for mixed max scores. pcml() recentres thresholds so the mean item location is zero, but the covariance stayed under the sum-of-thresholds constraint – identical only when every item has the same maximum score. With mixed 1/2/3-threshold items, per-item empSD/SE ran 0.90-1.21 by item type; the linear transform (I - 1a') cov (I - a1') restores 0.96-1.09. Affects threshold and item SEs, separation, and equating weights on mixed designs.
• Warm WLE discrimination cancellation. The common-discrimination weighted score is disc[(R-E) + mu3/(2V)] – the discrimination cancels; a stray disc factor in the correction biased WLEs by up to 0.26 logits at disc = 0.5 (exact at disc = 1, so ordinary analyses were unaffected; the vector-disc EFRM path was already correct).
• Honest chi-square degrees of freedom. An item whose responders occupy fewer than two class intervals has no estimable item-trait interaction: chi-square, df, and p are now NA (df was manufactured as 1 with p = 1), and the total test sums over testable items only. The adjust_N proportional-scaling semantics are now documented explicitly.
• Covariance-correct equating drift tests, both families. The shift c0 is estimated from the same common items the tests examine, and locations within a calibration are correlated through the identification constraint: drift denominators now use [(I - 1u') Sigma (I - u1')]_ii with Sigma the sum of the two calibrations’ location covariances (a bank contributes diag(se^2)). Null calibration: 4.6% rejection at nominal 5%, t-SD 0.97.
• Judge-clustered SEs refuse a single judge (they collapse to ~1e-16) and note fewer than 10 clusters as likely understating.
• btl_dif() now tests at the judge level: residuals aggregate to one weighted mean per judge (per opponent band) and enter a split-plot ANOVA with the judge as the error unit. The comparison-level ANOVA pseudo-replicated – 6 of 10 null datasets with judge heterogeneity and arbitrary groups falsely flagged uniform DIF; the judge-level design flags 0 of 10 while still detecting a planted 1-logit effect. Band breaks moved to count-weighted quantiles, making aggregated and expanded data exactly equivalent. Power now scales with judges per group – the honest unit.
• Non-convergence now warns loudly at fit time in rasch() and btl() (it was only visible in print); btl()’s convergence flag is scale-free (gradient per comparison).
• Documentation: the test-of-fit suite’s calibration status is stated plainly (approximate, convention-faithful diagnostics; simulation shows near-but-not-nominal rejection), with parametric-bootstrap references via sim_replicate() recommended where exact calibration matters; the BTL pair chi-squares are marked descriptive under judge clustering; the next-pair priority is documented as a ranking heuristic; the BTL dimensionality reference’s no-refit design is stated.

rasch 1.10.3

Input honesty, from an external code review (five findings, all confirmed and fixed):

• Misspelled id, factors, or items column names in rasch() are now errors naming the missing columns; they were silently ignored (one misspelled factor silently dropped ALL factors), producing valid-looking analyses of the wrong data. Numeric item indices still work.
• Fractional scores (e.g. 1.9) are now an error naming the columns; as.integer() used to truncate them silently, altering response data. Integer-valued doubles (“2.0”) still pass.
• MFRM rows with a missing person, item, or facet identifier are dropped with a note; paste() used to coerce NA to the literal string “NA”, silently pooling unrelated rows under a phantom level.
• equate_tests() excludes common items whose location or SE is unavailable (e.g. weakly determined items with honest NA SEs) from the precision-weighted shift and drift tests, notes them, and keeps their rows with NA test columns – one such item used to turn every statistic NA. plot_equate() handles the excluded rows.
• report_html() escapes all data-derived text (title, notes, item and facet names) – crafted column names could inject markup into reports; table cells were already escaped.

rasch 1.10.2

• App: hover identification is now consistent across the app, through one shared mechanism. Beyond the person-fit plot and item fit map, hovering identifies the point (or cell) on: the Guttman scalogram (person, item, score), both residual-correlation heatmaps (item pair and its Q3/Q3*), the common-item equating plot and the tailored-analysis comparison (item, reference and current locations), and – for paired comparisons – the unexpected-judgements map (matchup, surprise, location), the group characteristic curves’ observed means (object, panel, mean), and the equating plot (object, both calibrations). Every tooltip is resolved server-side against exactly the rows its plot draws. Plots whose points are already labelled (kidmap, Wright map, caterpillars) and curve or distribution displays are deliberately left without hover.

rasch 1.10.1

• App: hovering a point on the person-fit plot or the item fit map now shows who it is – a floating tooltip with the person ID (or item name), location, and fit residual, resolved server-side against exactly the subset each plot draws. The navbar wordmark is left-justified.

rasch 1.10.0

• Model comparison, calibrated for composite likelihood. compare_fits() gains cl_aic and cl_bic (Varin & Vidoni 2005; Gao & Song 2010): -2 cl penalised by the effective parameter count tr(H^-1 J) from the Godambe matrices – the same quantity whose eigenvalues calibrate lr_test() – which absorbs the pairwise over-counting a nominal AIC/BIC would ignore (empirically 3.5-5x the nominal count). Verified by plant-and-detect: the criteria select RSM on rating-structured data, PCM under varying threshold spreads, and a planted BTL position effect – and reject the richer model when the effect is absent. n for CL-BIC counts independent units: persons contributing an informative pair, or judges (count-weighted comparisons when unclustered). MFRM/EFRM fits report NA with the reason noted.
• compare_fits() now accepts btl fits (all-BTL comparisons: free vs principal-component thresholds, with/without a position effect or within-judge dependence), reporting judges/objects/comparisons and OSI. btl() fits carry their composite-likelihood ingredients in $cl; anchored pcml() and pcml_pc() now return their Godambe matrices. The app’s Compare page shows the criteria and accepts paired-comparison fits.
• Cross-validation against independent implementations, in the test suite whenever the packages are installed (new in Suggests: eRm, sirt, psychotools): dichotomous, PCM, and RSM parameters against eRm’s full CML (agreement to sampling precision; pairwise SEs at or just above CML’s, the documented efficiency price), the same pairwise family against sirt::rasch.pairwise (near-identity), and btl() against psychotools::btmodel to machine precision (same likelihood).

rasch 1.9.3

• Missing-data identification, verified and enforced. A booklet-design probe (zero complete cases, linked forms, MCAR on top) confirmed the pairwise conditional estimator recovers planted parameters with nominal coverage – and exposed two silent failure modes at the design’s edges, both now honest:
• Disconnected designs error instead of fitting. When no person answers items in more than one block, the pairwise likelihood is flat in the between-block shift; the fit used to return arbitrary cross-block spacing (within-block orders fine, origins meaningless) with no warning. pcml(), pcml_pc(), and everything built on them now stop with an error naming the disconnected blocks. Anchors rescue a block (disjoint-form equating): with an anchored item in every block the fit proceeds; with any block unanchored it stops and names it.
• Thresholds beside near-empty categories report NA, not a lie. A category with 1-2 responses can send its threshold toward a boundary (observed: a single top-category response put an item location at +13.9 logits) while the ridged covariance printed SE = 0. Such thresholds are now flagged (thr$weak), their SEs and the item-location SE reported as NA, and a note names the item, the category, and its count, pointing to pc_components or category collapsing. Estimates are still returned; categories with zero responses are rescored away as before.

rasch 1.9.2

• Case study: inst/casestudies/party_blocs_crisis.R applies btl_efrm() to the Tuebingen 2009 party-preference data (via psychotools), with ideological blocs as object sets and economic-crisis concern as judge panels. Running the model against real data surfaced three defects, all fixed below.
• A set whose within-set contests are all near-even (or all one-sided) carries no stable information about the panel-unit ratios: its stage-one ratio estimate diverged and poisoned the phi reconciliation with a spurious covariance (on the party data, the right bloc’s single CDU/CSU-FDP pair at 55:45 drove phi to 3e4 and a fatal error). Such sets are now screened out of the reconciliation, refit with the panel units held at the reconciled phi – which the frame model says apply to them regardless – and named in a note. Stage one also gains the same trust-region step cap as btl().
• Bootstrap replicates in which a set unit reaches its boundary (log alpha driven to -Inf when a resampled within-set order flips against the cross-set evidence, the signature of a two-object set with a near-even internal pair) made the standard error silently NaN, printed as blank. The SE is now reported as NA with a note naming the parameter and the boundary count: no number is manufactured for a sampling distribution that is not normal. A replicate whose convergence flag was NA could also crash the bootstrap loop; it now counts as a failure.
• The stage convergence checks compared the gradient to an absolute threshold, which is scale-dependent: on large designs a converged fit could be flagged unconverged and – with the new screen – silently rerouted, changing the estimates. Both stages now use a per-comparison criterion, verified invariant to 50-fold duplication of the data.
• Documentation: the bootstrap’s two honesty notes (model-based resampling versus judge-clustered conditional errors, and boundary-unstable parameters) are spelled out in ?btl_efrm.

rasch 1.9.1

• Shiny app: the Frames page gains its paired-comparison variant – nominate a judge-panel column and an object-set map (CSV, or inferred from object-name prefixes), estimate the frame units on demand (bootstrap standard errors by default), and read the panel units, set units and origins, the unit caterpillar, and the per-frame fit. On the app’s own paired-comparison demo the two judge panels differ significantly in discriminating power – the panel containing the erratic judge carries the smaller unit – which is precisely the phenomenon the model makes visible.

rasch 1.9.0

• The extended frame of reference model for paired comparisons: btl_efrm() estimates object locations when the unit of the scale differs across judge-panel by object-set frames – the package’s extension of Humphry’s persons-by-items EFRM to the Bradley-Terry-Luce family. Panel units (phi, a panel’s discriminating power) are identified conditionally wherever two panels judge a common set; set units and origins (alpha, kappa) are identified from cross-set comparisons, which are themselves conditional outcomes – so, unlike the persons-by-items case, the linking makes no distributional assumption about the objects. Everything is estimated by the staged conditional estimator: the within-frame calibrations are invariant to the linking data (the frame-of-reference property), a deliberate trade of efficiency for invariance. The cross-frame discrimination convention (cross-set contests judged at exactly phi) is stated plainly in the documentation as an assumption, with the per-frame fit residuals as its check.
• Inference for btl_efrm() defaults to a parametric bootstrap of the whole two-stage pipeline: simulation showed the fast conditional errors cover at a third of nominal on chain-linked designs (they omit the stage-one uncertainty that flows into the linking), while the bootstrap restores 12/12 coverage and nominal false-flag rates at both small and large designs. The conditional errors remain available as se_method = "conditional" for quick inspection, labelled as understating. plot_btl_units() draws the unit caterpillar; simulate_btl_efrm() generates frame-structured data with planted units for the plant-to-detect loop.

rasch 1.8.0

Three further paired-comparison analogues of the Rasch tool set, and CRAN housekeeping.

• Anchored estimation for paired comparisons: btl(anchors = c(Object = value, ...)) holds the named objects at fixed locations and estimates the rest on that scale (no sum-zero constraint; anchored objects report se 0) – the equating workflow’s second half, mirroring rasch(anchors=). An anchored object at a response boundary is an error rather than a silent removal.
• First-position advantage: btl(position = TRUE) estimates the pure positional (first/left-presented) advantage of the Davidson & Beaver
  1977. order-effect device as a constant covariate alongside any exposure and carry-over effects; it is identified by triangle closure even under fixed presentation orders (more weakly, with honest SEs). The dimensionality reference simulation carries the fitted coefficient.
• Common-object equating: btl_equate() compares two paired-comparison calibrations (or a calibration against a bank) through their common objects – precision-weighted origin shift, per-object drift t-tests with multiplicity adjustment, equated locations – with plot_btl_equate(); the standards-maintenance workflow of comparative judgement across panels or years.
• Targeting and adaptive pairing: btl_information() (per-comparison Fisher information and per-object design information, with the naive 1/sqrt(information) SE beside the judge-clustered sandwich), plot_btl_targeting(), and btl_next_pairs() – the greedy most-informative-next-pair step of adaptive comparative judgement (Pollitt 2012), with the Bramley (2015) reliability-inflation caution documented.
• CRAN housekeeping: pure-ASCII sources, quoted ‘shiny’ in DESCRIPTION, heavy Monte-Carlo recovery tests wrapped in skip_on_cran() (they run locally and on CI), and cran-comments.md drafted from a full --as-cran check (about 3 minutes; slowest example 1.4s).
• An adversarial verification round over the new features also caught and fixed a PRE-EXISTING inference bug: btl(count=) inflated every sandwich standard error by about the square root of the count (the meat weighted aggregated rows twice; point estimates were always correct). Count- weighted fits now reproduce the expanded-data standard errors exactly; OSI and all z/p values follow. Additionally from that round: btl_equate()’s shift standard error now uses the stored covariance of each calibration (it treated the common-object differences as independent, overstating the SE several-fold; 95 per cent coverage verified) and documents the majority-drift limitation of single-shift equating; btl_next_pairs() ranks by the exact one-step reduction in total location variance (a Sherman-Morrison update on the fit’s covariance), verified to beat low-priority pairs empirically; and btl_information()’s se_naive is documented honestly as a single-parameter lower bound rather than an independence benchmark.
• Shiny app: the Targeting and Equating tabs gain paired-comparison variants (design information + targeting plot + adaptive next-pair recommendations; reference-calibration upload with drift table and plot), and the Data roles gain a first-position-advantage switch and an object-anchors upload.

rasch 1.7.1

A statistical audit of the new simulation and paired-comparison independence code (23 findings across two adversarial rounds, all fixed), plus a vignette and a logo.

• Misfit layers now compose correctly in simulate_rasch(): the response-dependence and response-style regeneration steps previously rebuilt their target items from the base structure, silently erasing any DIF or second-dimension loading planted on those items (and leaking a DIF-like signal from a DIF-carrying source item into its dependent partner). Every regeneration now honours the item’s own generating structure – trait, group-shifted thresholds and slope, guessing.

• btl_dimensionality() now simulates its noise reference sequentially WITH the fitted within-judge dependence effects when the fit carries them: real order effects previously inflated the false-positive rate to ~36 per cent on one-dimensional data (now ~0-5 per cent). The documented price is conservatism, since carry-over and a judge-camp second attribute are partially confounded.

• Doc examples pass id = "id" so they no longer emit spurious dropped-column notes.

• simulate_rasch(): model = "PCM" now draws each item’s threshold spacings and span afresh, as the partial credit model allows (previously PCM and RSM generated identical rating-scale-structured data); dif without n_groups >= 2 is an error instead of a silent no-op; polytomous guessing warns and is ignored rather than falsely recorded as planted; the response-dependence term now carries the guessing asymptote of the pair’s first item (no more spurious easiness on the dependent item); disordered thresholds work with 3 categories and no longer shift the item’s location; careless responders are no longer double-counted as response-style persons in the recorded truth.

• simulate_mfrm(): halo raters are capped at the eligible pool (no NA raters when erratic raters shrink it); the example runs.

• sim_recovery(): person abilities are mean-centred as documented (an asymmetric difficulty range no longer masquerades as person bias); the many-facet branch reads the item margins (item_effects), restoring the documented item-difficulty recovery for MFRM fits.

• btl_dimensionality(): null calibration verified for graded fits (mildly conservative, never anticonservative) and documented; docs also clarify the leading-bimension reference columns and the no-ties precondition of Kendall’s zeta.

• Shiny app (Simulate page): the parameter-recovery card only renders when the current fit was estimated on the currently loaded simulation (a stale fit can no longer be compared against new truth); the reproducible call now carries every argument the generator received (MFRM person/item SDs, EFRM person SD, graded categories); non-integer seeds are rounded safely.

• New vignette: “Planting misfit and watching the diagnostics fire” — the plant-to-detect loop end to end, closing with a small Monte Carlo power estimate via sim_replicate().

• A hex logo (an item characteristic curve at its location), reproducible from tools/logo.R.

rasch 1.7.0

• Simulation gains the standard controls of a proper simulation tool.
  • Population parameters: the person distribution (mean, SD, shape) and the item-difficulty range are now set directly (in the package and the app).
  • More misfit: response styles (extreme / midpoint categories) and speededness (a not-reached tail) for , and a halo effect for .
  • generates a batch of datasets for Monte Carlo power, Type-I, or recovery studies.
  • and compare the parameters a fit recovers with the ones that were planted (correlation, RMSE, bias, and a true-vs-estimated scatter) for every layout.
• Shiny app: the Simulate page renames the wide layout to “Rasch”, exposes the population parameters, shows a parameter-recovery report and the true-vs-estimated scatter after Run, and prints the reproducible call.

rasch 1.6.1

• Shiny app: a Simulate page (under More) exposes all four simulators with a data-type selector and per-layout misfit controls. Pressing Simulate generates the data, loads it as the current dataset with its roles set, and lists what was planted; pressing Run then analyses it so the matching diagnostic can be watched firing. Verified end to end for every layout (e.g. a planted DIF item is flagged by dif_anova, a many-facet fit recovers its rater severities).

rasch 1.6.0

• Data simulation across all four model families, each with dial-in departures from the model – so a known pathology can be planted and the matching diagnostic watched as it fires. Every result carries the true parameters () and prints what was planted.
  • : dichotomous/PCM/RSM with discrimination misfit, guessing, a second dimension, local dependence, DIF (uniform and non-uniform), careless responders, disordered thresholds, and missing data.
  • : paired comparisons with erratic judges, a second object attribute (two judge camps), and within-judge exposure/carry-over dependence; dichotomous or graded.
  • : rated data with rater-severity spread, erratic raters, and a rater-by-item interaction.
  • : frames with differing set and group units. Verified end to end: each planted departure is recovered or flagged by the corresponding fit / diagnostic (an app panel follows).

rasch 1.5.1

• The judge unexpected-judgements map () is now drawn at the matchup level: each pair the judge met is a segment spanning its two objects on the consensus location axis, placed by how surprising the verdict was, with the judge’s backed object filled. returns the per-matchup residuals behind it. The object-level is retained for per-object residuals.

rasch 1.5.0

• judge_surprise() and plot_btl_judge_map(): the paired-comparison counterpart of the kidmap. A judge has no ability to condition on, so the reference is the consensus object scale; each object a judge met gets a standardised residual saying how much more or less that judge favoured it than its location predicts. A strong object the judge under-rated, or a weak object over-rated, is an unexpected judgement. It catches systematic bias – a distinct failure mode from the noise that judge fit and judge consistency flag. In the Shiny app the Persons > Judge fit panel is now a master-detail: the judges table drives the selected judge’s map, mirroring person -> kidmap.

rasch 1.4.0

• Paired comparisons gain the pair-structure analogues of the Rasch independence diagnostics – the tools that a persons-by-items residual matrix would give but paired-comparison data cannot.
  • btl_transitivity() counts circular triads (preference loops: A beats B, B beats C, C beats A) against the chance rate, reports Kendall’s coefficient of consistency for complete designs, and gives each judge’s own consistency (Kendall & Babington Smith 1940) – the single-dimension check, and a judge-fit analogue.
  • btl_dimensionality() decomposes the skew-symmetric object-by-object residual preference matrix into Gower (1977) bimensions – the paired-comparison counterpart of residual PCA – and judges the leading bimension against a model-simulated noise reference, so a coherent residual “swirl” (a second attribute steering some contests) is distinguished from noise.
  • plot_btl_scree(), plot_btl_dim_map(), and plot_btl_transitivity() display them; the Shiny app’s Independence > Trait tab now carries these for a BTL fit (it previously hid for paired comparisons).

rasch 1.3.1

• The package is now called rasch (it was developed under the working title rmt). The name rmt belongs to an unrelated package on CRAN, and rasch says plainly what this is – and mirrors mirt::mirt() with rasch::rasch(). No functions, arguments, or results change; only the package and library name. The auxiliary result classes were unified under the rasch_ prefix (rasch_btl, rasch_dif, …) to match the fit classes.

rasch 1.3.0

A full statistical audit of the independence, invariance, and paired- comparison procedures, with fixes throughout.

• Paired comparisons: btl() now interrogates within-judge dependence – dependence_data holds every comparison’s exposure and carry-over covariates (count-weighted), plot_btl_dependence() displays an effect as a partial-residual curve, and a separated effect (one-sided informative comparisons) is set aside with a note instead of reported as a runaway estimate. The graded engine’s threshold-by-dependence Hessian block is corrected (free-threshold fits with five or more categories and an order column now converge), rows with a missing judge are dropped with a note, category checks re-run after boundary objects are removed, and the dichotomous path now routes through the one (vectorised) estimator.
• btl_dif() holds fitted dependence effects fixed in its residual moments and resolution refits (dependence is no longer absorbed as spurious judge-group DIF), weights count-aggregated comparisons exactly as expanded rows, restricts supersession to group terms, and notes every significant term it does not resolve.
• dif_anova() keeps each ANOVA term in its own error stratum, so a mixed (within-subject) design with missing responses no longer un-flags real DIF; factors named like internals (ci, f1) are safe; a term’s own class-interval crossing no longer supersedes it (so resolve_dif() again splits items with both uniform and non-uniform DIF); mixed designs say plainly that Tukey comparisons are unavailable; and dif_size() notes when a within-person factor makes its standard errors conservative.
• plot_scree()’s parallel-analysis reference is now model-simulated (responses drawn from the calibrated model, persons re-estimated), so model-true data sits at the reference instead of always “showing structure” (Raiche 2005; Chou & Wang 2010).
• Shiny app: judge-group DIF results freeze their run’s factors and alpha (later sidebar edits cannot silently regroup the overlay or break the reproducible snippets), superseded terms are shown, the within-judge dependence panel explains dropped effects, and fit flags are consistent on every table (fit residual |2.5|; outfit 0.7-1.3; infit 0.8-1.2).

rasch 1.2.0

• plot_pca_biplot() draws the item loadings on the first two residual principal components – the pair that usually carries any interpretable contrast – on equal axes, coloured by the sign of the PC1 loading.

• residual_correlations() now also returns the adjusted-Q3 star_matrix (each Q3 less the average off-diagonal), and plot_resid_cor() gains a stat argument to colour either the raw Q3 or the adjusted Q3*, drawing the lower triangle only.

• Shiny app: the trait- and local-dependence pages pair each table with its plot (loadings with a PC1-PC2 biplot; Q3 and Q3* matrices with heatmaps), the raw and adjusted correlations are each flagged by their own rule (|Q3| and Q3*), and every data-restructuring action (subtest, item split, automatic DIF resolution) offers an in-place reset to the original data.

• dif_anova() is now the single DIF analysis-of-variance function. One factor is analysed one-way; several factors are modelled jointly – the statistically correct treatment – with main effects by default and factor-by-factor interactions optional (effects = "factorial"). It handles within-subject factors (repeated-measures / mixed ANOVA) and returns a classed object with a summary, the full terms table, and Tukey comparisons. The separate dif_anova_factorial() is removed.

• resolve_dif() resolves DIF iteratively by item splitting, largest effect first, to clear artificial DIF.

rasch 1.0.0

First stable release. The package delivers a complete Rasch Measurement Theory workflow built entirely from published measurement theory, with a pairwise conditional estimation core and a Shiny interface that exposes every analysis with reproducing R code attached to every output.

Models

• rasch(): dichotomous and polytomous (partial credit and rating scale) Rasch models by pairwise conditional maximum likelihood (Andrich & Luo 2003; Zwinderman 1995), with Godambe sandwich standard errors, sum-zero identification, Warm (1989) weighted likelihood person estimation, and the principal-component threshold parameterisation as an estimation option.
• rasch_mfrm(): many-facet models with additive facet severities or item-by-facet interactions, accepting wide (one column per item) or long data.
• rasch_efrm(): the extended frame-of-reference model (Humphry) with frame units estimated by within-frame pairwise conditional likelihood.
• btl(): paired comparisons as the conditional form of the dichotomous Rasch model (Bradley & Terry 1952; Andrich 1978), including the adjacent-categories graded extension (Tutz 1986; Agresti 1992) with symmetric thresholds – the Davidson (1970) ties model is its three-category case – winner-plus-margin data entry, judge-clustered errors, and within-judge exposure and carry-over effects estimated jointly with the locations.

Diagnostics

• The test-of-fit suite follows Andrich & Marais (2019): log-of-mean-square fit residuals with apportioned degrees of freedom, the item-trait chi-square on automatically sized, tie-preserving class intervals with per-interval detail, the ANOVA item fit, and person fit.
• Unidimensionality: residual principal components with parallel analysis, the Smith (2002) subset t-test, and magnitude estimation.
• Local dependence: Yen’s (1984) Q3 with the Christensen, Makransky & Horton (2017) flagging convention, response-dependence magnitude (Andrich & Kreiner), subtest formation, and the spread test.
• Guessing: tailored analysis with origin-equated comparison calibrations.
• DIF: two-way residual ANOVA per factor, the joint factorial analysis with a compact uniform/non-uniform summary, resolved DIF magnitudes in logits with familywise control and a practical-significance criterion, planned contrasts derived from the factor structure (with repeated-measures support via person-level residual scores), and class intervals sized to the cells each analysis actually uses. For paired comparisons, btl_dif() tests object-by-judge-group interaction by the same two routes.
• Equating and invariance: common-item comparison with drift flags, and calibration anchoring.

Display and reporting

• A complete base-graphics plot suite: expected value curves with observed class-interval means, category and threshold probability curves, the person-item threshold distribution, Wright maps, kidmaps (Wright, Mead & Ludlow 1980), object characteristic curves for paired comparisons, threshold and item maps, test characteristic and information curves, residual displays, and Guttman scalograms – all with adjustable class intervals and scale ranges, and batch export to multi-page PDF or ZIP at publication resolution.
• fit_summary_table() and targeting_table() return the headline statistics as tidy tables; save_outputs() writes every table and plot to disk; report_html() produces a single-file HTML report.
• run_app() launches the Shiny interface: model-adaptive navigation, reproducing R code beneath every output, and one-click export.