Changelog

Source: NEWS.md

meddecide 1.0.4 (2026-08-07)

This release is the result of end-to-end release reviews of fourteen analyses: decisioncompare, decision, agreement, lassologistic, decisioncombine, decisioncalculator, nogoldstandard, cotest, sequentialtests, enhancedROC, psychopdaROC, and all three kappaSize sample-size analyses — kappaSizePower, kappaSizeCI and kappaSizeFixedN, each reviewed in its own right and then cross-checked against the other two. Each was traced from its user interface through to its printed output, every statistic was checked against an independent R package (epiR, DescTools, psych, binom, poLCA, pROC, cutpointr, kappaSize, stats), and the confirmed defects were fixed. Most of what follows is a correction to a number a clinician could already have copied into a report.

Breaking changes

• decisioncompare() has four new required arguments: goldNegative, test1Negative, test2Negative and test3Negative. They name the level of each variable that represents a genuine negative result, so that a third, equivocal level can be told apart from a negative one. jamovi does not permit a default on a Level option, so all four compile to bare parameters of the R wrapper: existing scripts that call decisioncompare() will fail with argument "goldNegative" is missing, with no default until they are updated, passing NULL for any test that is not in use. Nothing changes for users driving the analysis from the jamovi GUI. The reason for the change is the excludeIndeterminate defect described below — without a declared negative level the analysis has no way to know which rows are equivocal.
• decisioncombine()’s filterPattern lost the levels serial, parallel and majority. That option selects rows of the observed pattern table, but those three names describe decision rules, which is what the strategy table reports — “serial” as a filter selected the +/+/+ pattern while the Serial (AND) rule is a different row of a different table. The remaining levels are all, allPositive, allNegative and mixed. A script passing one of the removed values now fails with Argument 'filterPattern' must be one of ...; use allPositive for the old serial and all for the old parallel. The jamovi GUI shows only the current levels.
• nogoldstandard()’s default method changed from all_positive to latent_class. The old default defines the reference standard as “every test positive”, which fixes sensitivity and NPV at 1 for every test on every dataset — it cannot estimate accuracy, which is the entire point of the analysis. latent_class is the only method here that estimates sensitivity and specificity without building the reference standard out of the tests themselves. A new analysis will therefore produce different numbers than it did in 1.0.3, and because latent_class needs three or more tests, a two-test analysis that previously returned (meaningless) results now refuses and explains why. Saved .omv files keep the method they were saved with; pass method = "all_positive" explicitly to restore the old behaviour.

Fixed

decisioncompare (Compare Medical Decision Tests)

• Every results table grew on each re-run. The six calls meant to clear old rows used clearRows(), which is not a jmvcore Table method — only deleteRows() exists — and each call was wrapped in try(silent = TRUE), so the error was swallowed and nothing was ever cleared. In jamovi, toggling any option re-runs the analysis on the same object, so the comparison table went 4 → 8 → 12 rows across three runs and each test appeared twice, then three times, then four. Anyone who changed a checkbox mid-session was reading a duplicated table.
• excludeIndeterminate was a complete no-op. It filtered on c(positiveLevel, setdiff(levels, positiveLevel)) — that is, on every level — so equivocal results were still collapsed into Negative and still inflated specificity and NPV, the exact harm the checkbox promises to prevent. On a 60-case fixture with 20 Equivocal results, specificity read 0.950 with the option both off and on; excluding the equivocal cases gives 0.900. With the new negative-level options supplied, equivocal rows are now genuinely dropped; when they are not supplied the analysis says so explicitly rather than pretending to have acted.
• The manuscript-ready report named a winner the data could not support. On a fixture where Cochran’s Q gives p = 0.076 and no pairwise comparison survives Holm correction, the report read “t1 demonstrated OPTIMAL diagnostic performance” one sentence before “did not reveal a statistically significant difference”. It now says “had the highest observed accuracy” with an explicit caveat, and the clinical recommendation panel carries a matching caution block.
• Four result panels were permanently visible. The visible: expressions began with ! ((!is.null(test3) && test3 != "")), and jmvcore routes an expression to the R evaluator only when it matches ^\([\$A-Za-z].*\)$; a leading ! fails that test, so the raw — and therefore truthy — string was returned instead of being evaluated. An empty “Test 3 — Recoded Data” table sat under every two-test analysis, and the stratified table showed with no stratifier selected.
• The manuscript text printed a literal placeholder, 95% CI: [see confidence interval table]. It now prints real Clopper-Pearson intervals.
• A tie for best-performing test was broken by whichever test came first, silently. Ties are now disclosed.
• The difference table’s “95% Confidence Interval” header did not say which method produced it. It is now footnoted as the paired Wald interval for correlated proportions, and distinguished from the separate “CI Method for Agreement” option that governs the Overall Percent Agreement table.
• The three tests in test-decisioncompare-critical-fixes.R never called the module at all — they re-implemented the logic inline and asserted on their own arithmetic. They now exercise the real analysis.
• Verified unchanged and correct: all per-test metrics against epiR::epi.tests (to 1e-8); Cochran’s Q = 5.150442, p = 0.076137 against DescTools::CochranQTest (six decimal places); McNemar with Holm correction against stats::mcnemar.test + p.adjust (1e-10); Wilson intervals against binom::binom.confint; exact intervals against Clopper-Pearson; the paired Wald standard error by hand. McNemar correctly compares diagnostic correctness against the gold standard rather than raw positivity — proven with a fixture on which the wrong design gives p = 1.000 and the right one p = 0.0015.

decision (Medical Decision)

• With a user-supplied population prevalence, the results row was arithmetically impossible. The Prevalence cell was overwritten with the user’s prior while PPV and NPV stayed at the raw 2×2 values computed at the study prevalence, so a row could read “Prevalence 5.0% … PPV 88.9%” — for a test with sensitivity 0.80 and specificity 0.90, Bayes gives 29.6%. A pathologist reading PPV off a screening-prevalence run would have overstated positive predictive value roughly threefold, and the footnote made it worse by asserting the predictive values had already been prior-adjusted. PPV and NPV now report the post-test probabilities that correspond to the stated prevalence, both branches carry a footnote naming which prevalence was used, and the epiR interval pane — whose exact binomial values cannot be moved to a different prevalence — gains a note saying it is at the observed sample prevalence.
• The epiR “number” table’s footnotes described the wrong statistics. They were attached by hard-coded row number while the rows render LR+, LR−, DOR, Youden and NNDx, so LR+ was described as the diagnostic odds ratio, LR− as the number needed to diagnose, and DOR as Youden’s index. Footnotes are now keyed to the epiR statistic name carried alongside the data.
• The Fagan nomogram silently failed to render whenever a 2×2 cell was zero — precisely the sparse tables that most need it. A zero cell makes sensitivity or specificity exactly 1, which nomogrammer rejects. The likelihood ratios passed to it were already Haldane-Anscombe corrected, so the proportions now come from the same corrected table, keeping the plot self-consistent instead of clamping to an arbitrary epsilon. Plr/Nlr are no longer passed: nomogrammer ignores them when sensitivity and specificity are supplied, and warned on every render.
• With a zero cell, the interval pane disagreed with the estimates beside it. LR+, LR− and DOR in the main tables are computed from Haldane-Anscombe corrected counts (LR+ 145), while epi.tests() was run on the raw table and returned Inf with a NaN lower bound — two numbers for one statistic on one screen. Those three rows now come from a second epi.tests() call on the corrected table. Sensitivity, specificity, PPV and NPV deliberately stay on raw counts.
• The “Getting Started” welcome panel was permanently visible, sitting above every completed analysis, for the same leading-! reason described under decisioncompare. Rewritten as (length(gold) == 0 || length(newtest) == 0 || length(goldPositive) == 0 || length(testPositive) == 0); a bare option name could not be used because a Variable’s value is a list, which R’s && rejects.
• Rows dropped because their level was neither the positive nor the negative one were reported as “cases with missing values removed” — 40 rows with nothing missing. Rows are dropped twice: once by jmvcore::naOmit() for genuine missingness, and again when an explicit negative level is set and other levels are recoded to NA. The summary now counts the two separately, says plainly that level-excluded cases are not missing values, and reports “N of M cases analysed”.
• A likelihood ratio of exactly 1.0 was called flawed, and in the wrong direction. It fell through the lr_pos > 1 band into “Decreases probability of disease (test may be flawed)”; it now reads “Uninformative”.
• The clinical summary called the user’s population prior the sample prevalence. The copy-ready report quoted PPV and NPV with neither the prevalence they were computed at nor a confidence interval, and closed by concluding “the test may be clinically useful” regardless of the results. The summary and report now quote predictive values at their prevalence and the report includes Clopper-Pearson sensitivity and specificity intervals.
• Verified: all statistics match epiR::epi.tests exactly (sensitivity 0.8, specificity 0.9, PPV 0.888889, NPV 0.818182, LR+ 8, LR− 0.222222), with the 2×2 not transposed.

agreement (Interrater Reliability)

• Weighted kappa and Gwet’s AC2 laid their weight matrices over an alphabetically sorted category order instead of the factor’s declared level order, scrambling any ordinal scale whose labels do not happen to sort into clinical order — which is most of them. Low/Moderate/High sorts to High/Low/Moderate; Negative/Weak/Strong to Negative/Strong/Weak; Absent/Focal/Diffuse to Absent/Diffuse/Focal. On the test set recorded in the code, weighted kappa read 0.751 where the correct value is 0.597 — across the Landis & Koch moderate/substantial boundary. Three separate places were at fault: .pairKappaWithCI() used sort(unique(c(a, b))), irr::kappa2 was handed a data frame (which irr coerces with as.matrix() to character and then re-derives categories alphabetically) and is now fed integer factor codes, and irrCAC::gwet.ac1.raw now receives an explicit categ.labels. Unweighted kappa and Gwet’s AC1 are order-invariant and were unaffected; only the weighted options bite.
• Weighted kappa was applied to nominal variables. Ordinality was inferred from category count alone (length(all_levels) >= 3), so quadratic weights — which assume a meaningful distance between categories — were applied to any nominal variable with three or more levels, such as tumour type or mutation class. It now also requires is.ordered() on at least one of the two rating columns.
• The headline Cohen’s/Fleiss’ kappa table reported kappa, z and p but no confidence interval, while four secondary tables (Krippendorff, all-pairs, item-modal, hierarchical) all carried one. For two raters the table now shows an interval from the non-null asymptotic standard error via vcd::Kappa, which matches psych::cohen.kappa exactly, with a footnote explaining that z and p test H0: kappa = 0 off the null SE and so z is not kappa divided by the interval’s SE. For three or more raters the interval is deliberately left blank with a note, because irr supplies only the null-SE test for Fleiss’/Conger’s kappa.
• A non-finite kappa crashed the entire analysis, not just the panel that produced it. The “Clinical Meaning” block ran if (kappa_val >= 0.60) without the is.na() guard the Landis & Koch chain above it had, so if (NA >= 0.60) threw “missing value where TRUE/FALSE needed”. Reachable with weighted kappa on nominal data, exact kappa with two raters, a single rating category, or Fleiss returning a non-finite value. Separately, irr::kappam.fleiss returns -Inf when three or more raters all use one category — 100% agreement — and that value was written straight into the table and graded “poor agreement (worse than chance)”, the exact opposite of the data. Kappa is now blanked with a note explaining that there is no chance-agreement baseline when there is no variation.
• Subgroup forest-plot intervals used the null-hypothesis standard error and ignored confLevel. The subgroup path derived se <- abs(kappa / z) from irr’s z — the H0 SE, the same defect corrected elsewhere in the file — and hard-coded 1.96. Two-rater subgroups now use the non-null ASE, falling back to the null SE only when it is unavailable, and the multiplier is qnorm() at the user’s confidence level. The plot subtitle, previously hard-coded to “95% confidence intervals”, now prints the level actually used.
• The same kappa was given two different words in one output. The subgroup table used its own unattributed cut-points (0.40/0.60/0.75/0.90 → Poor/Fair/Good/Excellent/Outstanding) while the plain-language summary used Landis & Koch 1977 (0.20/0.40/0.60/0.80), so 0.61 read “substantial” in the summary and “Good” in the table, and 0.56 read “moderate” there and “Fair” here. One named scale (Landis & Koch) is now used throughout.
• Wald intervals for kappa were not clamped to the parameter space, and an unclamped interval reported an upper limit of 1.18 for a statistic bounded on [−1, 1]. Both branches now clamp, matching the sibling inter/intra path that already did.
• Bootstrap confidence intervals computed the wrong statistics, in four separate ways. (a) The categorical-versus-continuous decision was made per replicate with length(unique(na.omit(x))) <= 20; a bootstrap resample retains only about 0.63n distinct values, so at n ≈ 30–40 some replicates fell under the threshold, were treated as categorical, produced no ICC, and silently dropped the ICC row — at exactly the study sizes most common in pathology. It is now decided once, on the original data. (b) Each rater column was coded independently with as.numeric(factor(x)), so the same clinical category received different codes in different columns whenever raters used different subsets of the scale, corrupting Krippendorff’s alpha; a single shared category set is now used. (c) Krippendorff’s level of measurement was hard-coded to nominal and now follows the data. (d) The bootstrap ICC was hard-coded to ICC(2,1) two-way absolute agreement while the row was labelled with whatever iccType the user had selected; the user’s choice is now honoured in both paths.
• The Bland-Altman plot did not refresh when the proportional-bias option was toggled — proportionalBias was missing from clearWith although the plot adds a geom_smooth line under it. Separately, the pairwise-agreement figure silently truncated at max_pairs: with five raters, 4 of 10 pairs were dropped with no note, so the figure read as the complete set. The omission count is now stamped on the image itself, so it travels with an exported figure.

lassologistic

• Coefficients and odds ratios were reported on the standardized scale, not the original measurement scale. The design matrix is standardized in-module and glmnet is then called with standardize = FALSE, so nothing was back-transformed and every coefficient was per 1 SD. For a 0/1 dummy from a balanced marker (sd ≈ 0.5) the printed per-SD odds ratio is roughly the square root of the model’s actual present-versus-absent odds ratio: 1.81 printed where the model implies 3.25. The column standard deviations are now retained and coefficients divided through, reproducing exactly what glmnet(standardize = TRUE) would return. The Importance column keeps the per-SD magnitude, which is the quantity that is comparable across predictors, and the table note was rewritten accordingly.
• The integer scoring system was internally inconsistent, and its rule was never published. Points were derived from the raw per-SD coefficients while scores were awarded on a median split, so binary and continuous predictors were weighted on two different contrasts — a 0/1 dummy’s per-SD coefficient is about half its real effect, while a continuous median split spans roughly 1.6 SD — mis-ranking them against each other and making the Scoring System table’s odds ratio disagree with the Selected Variables table for the same predictor (2.11 against 4.46 for p53). Points are now derived from the log-odds contribution of meeting each criterion, and the cut is resolved once so that the printed rule, the derivation contrast and the applied cut cannot drift apart. A new “Award points when” column publishes the criterion on the original measurement scale; previously the median cut never reached the output at all, so a clinician had no way to apply the score to a new patient.
• The Brier score was graded against fixed cut-offs and mislabelled as calibration. It is an overall accuracy score whose scale is driven by outcome prevalence: a no-information model that always predicts the base rate scores p(1−p), already 0.09 at 10% prevalence, which the old code graded “Excellent calibration”. It is now graded by the Brier skill score against that null model (“Good (x% better than predicting the base rate)” / “Marginal” / “No better than predicting the base rate”).
• With zero selected variables, a model that calls everyone positive was presented as perfectly sensitive. Every predicted probability is identical, the ROC is degenerate, and pROC returns -Inf, which was printed as “Optimal threshold: −Inf” beside Sensitivity 1.000 and Specificity 0.000. Non-finite thresholds now fall back to 0.5.
• Selecting exactly one predictor produced a completely blank result — no panel, no notice, no error. The welcome/To Do panel shows only when explanatory is empty, and the analysis returned early with one predictor, so both guards stayed silent. A warning now explains that LASSO performs variable selection and needs candidates to choose among.
• Listwise deletion was invisible. “Total observations” in the model summary actually held the complete-case count, so it read as the full cohort while rows had silently been removed, and the suitability assessment then green-lit the reduced N. The summary now reports “Complete cases analysed” and “Excluded (incomplete data)” separately, and a warning fires whenever any case is excluded, pointing out that a single sparsely-measured predictor can remove a large share of the cohort.
• The bootstrap optimism correction did not disclose how many replicates it rested on. Failed replicates are caught by tryCatch, left as NA and dropped by na.rm = TRUE, so a correction computed from 50 survivors of 200 looked identical to one computed from all 200. The table now footnotes the completed and failed counts, and warns separately when fewer than 20 replicates succeeded.

decisioncombine (Combine Medical Decision Tests)

• Five of the seven optional outputs were completely non-functional. asDF is an R6 active binding on jmvcore::Table, so tbl$asDF already returns the data frame; the code wrote tbl$asDF(), invoking that data frame as a function — “attempt to apply non-function”. The recommendation and all four plots were affected.
• The tables grew on every re-run and then broke. Nothing called deleteRows(), and jamovi re-runs .run() on the same object whenever an option changes, so rows went 5 → 10 → 15; the duplicated row keys then made $asDF fail with duplicate 'row.names' are not allowed, taking down the second run entirely.
• The analysis explained nothing when it stopped. .renderNotices() sat after three early returns in .run(), so the notice saying why it had stopped was collected and then discarded, leaving a blank analysis with no message.
• The “optimal” rule was an uncorrected argmax. It ranks 5 candidate rules with two tests, or 10 with three, using no interval and no test, so on pure noise it still names a winner — and called it optimal. It now discloses how many rules were compared and that the winner is not an established finding. Serial (AND) and the all-positive pattern are the same 2×2 under two labels; counting both manufactured a tie and inflated that count.
• Proportions and odds ratios shared one estimate column — sensitivity appeared as a percentage in one table and as 0.813 in another. Now split. Serial (AND) also gained its own named row: it was numerically identical to the “+/+/+” pattern and had been omitted, leaving a reader to know that the pattern was the serial rule.
• The “mixed” pattern filter dropped genuinely mixed patterns. It excluded anything starting with +/+ or -/-, removing +/+/- and -/-/+; and when no pattern matched it fell back to the whole unfiltered table rather than to nothing. A multi-level variable is now flagged rather than silently dichotomised, and a gold standard with one outcome yields NA with a notice instead of a silent number.
• Pattern and strategy statistics were confirmed against epiR::epi.tests, the Wilson intervals against binom::binom.confint, and the OR/AND identities and the eight-pattern partition were checked as mathematical facts rather than estimates.

decisioncalculator (Medical Decision Calculator)

• The tables grew on every re-run — verified at the jmvcore level: addRow() with an existing rowKey duplicates rather than replaces (rowCount 2 → 4 → 6 over three passes, after which $asDF fails). Fixed on the three tables that populate via addRow.
• The Fagan nomogram silently failed on any table with a zero cell — precisely the sparse tables that most need one — because nomogrammer rejects a sensitivity or specificity of exactly 0 or 1. The proportions now come from the same Haldane-Anscombe corrected table the likelihood ratios already used, and a test whose positive result argues against disease declines with an explanation instead of crashing.
• The most clinically useful part of the nomogram was invisible. nomogrammer prints its reading — prevalence, likelihood ratios, post-test probabilities — to stdout under Verbose = TRUE, and jamovi never shows stdout. It is now rendered beside the figure, at the tables’ precision rather than nomogrammer’s whole percents, and tracks the supplied prior.
• The cut-off comparison named the wrong alternative. The verdict was an if/else-if chain, so when both alternatives beat the current cut-off only the first was ever named, however much better the second was. It now picks the best of the three, declines to call a trivial advantage better performance, flags cut-offs whose totals differ (moving a threshold cannot change how many patients there are, so differing totals mean separate studies), and reports whether the Wilson intervals on the accuracies overlap — a formal paired test is not possible from four marginal counts per scenario, and the table now says so.
• Point estimates and every interval bound were confirmed against epiR::epi.tests; the 2×2 is not transposed; a supplied prior moves PPV and NPV while leaving sensitivity and specificity alone, and the table states which prevalence was used — the defect found in the sibling decision is not present here.

nogoldstandard (Analysis Without Gold Standard)

• Latent-class analysis reported sensitivity and specificity swapped. The diseased class was identified with probs[[i]][class, outcome] but read back as probs[[i]][2, disease_class] — the transpose. [2,2] and [1,1] coincide, so the error surfaced only when poLCA happened to label the diseased group as class 1, which is about half of all runs. Verified against poLCA on identical data and against the known truth of simulated data.
• The default method could not estimate accuracy at all, and now does not hide it. Under all_positive the reference standard is “every test positive”, so a diseased case can never be test-negative: FN is identically 0 and sensitivity and NPV are 1 for every test on every dataset. It printed “100% (95% CI 100–100%)”. any_positive is the mirror image (FP ≡ 0, specificity and PPV fixed at 1), and composite with two tests is any_positive, because a 1-of-2 tie passes a rowMeans >= 0.5 majority. Those quantities are now blanked with an explanation. See Breaking changes for the change of default.
• The confidence intervals were too narrow — by about 1.8× for sensitivity at 30% prevalence. The standard error used the total n for both metrics; the denominators are n × prevalence for sensitivity and n × (1 − prevalence) for specificity.
• The analysis claimed its latent-class model handles the assumption it actually makes. The always-visible method guide advertised “Handles conditional dependence” and “No identifiability issues”, while poLCA(nclass = 2, ~ 1) assumes the tests err independently given true status. The guide now says so; a new Conditional Independence Check table reports bivariate residuals per test pair (above 3.84 is evidence of a shared error source, which inflates estimated accuracy); and it explains that four or more tests are needed for the check to be informative — with three the model is just-identified and reproduces every table exactly. Two tests cannot identify a two-class model at all (5 parameters against 3 degrees of freedom) and are now refused rather than answered with numbers determined by the starting values.
• The Bayesian method disclosed neither its priors nor its nature. Beta(2, 1) on both sensitivity and specificity has mean 2/3 and increases toward 1, so it pulls estimates upward; nothing mentioned any prior, or that the results are not draws from a posterior. Both are now stated.
• One undefined cell could kill the analysis. if (ppv_denominator > 0) with an NA sensitivity threw “missing value where TRUE/FALSE needed”, ending the run instead of blanking a number. An invalid positive level produced Level 'test1_result' not found in variable 'negative, positive'. Available levels: {} — jmvcore::reject(formats, code = NULL, ...) takes code as its second positional argument, so the substitution values were swallowed and shifted.
• Bootstrapping is now a single seeded pass with warm-started latent-class fits, and an Analysis Diagnostics panel (under Verbose output) reports sample size, method, convergence, random starts and bootstrap failures. The clinical summary no longer announces a sensitivity “Range from 100.0% to 100.0%”, and no longer presents the fraction of cases on which all tests agree as an estimate of disease prevalence.

cotest (Co-Testing Analysis)

• A 0% post-test probability was reported for a test combination the model had made impossible. When two tests are modelled as conditionally dependent, the joint probability P(Test1+, Test2+) cannot exceed the smaller of the two marginals — a Fréchet bound. Beyond it the requested dependence is unattainable and cotest truncated it to the bound. That truncation is correct, but it forces one of the four test combinations to have probability exactly zero in one disease group, and the likelihood-ratio helper turned that into a printed post-test probability of 0.000000 — which reads as “this combination rules out disease”. With the standard cervical co-testing pair (HPV 95%/85%, cytology 55%/97%) the bound binds from a dependence of 0.25 upward, so the analysis reported that an HPV-negative, cytology-positive woman has no chance of disease. When both groups hit their bounds the ratio was 0/0 and the same 0.000000 appeared for a combination that cannot occur at all. The three degenerate cases are now distinguished: 0/0 is undefined and the row is blank; a combination impossible only in non-diseased subjects gives exactly 1 (previously a fake-precision 0.999991, an artefact of an internal likelihood-ratio cap of 1e6); and one impossible only in diseased subjects still gives 0. All three, and the truncation behind them, now raise a warning stating that the value follows from the assumed model rather than from data. The truncation was previously reported at “info” severity and never said what it implied.
• The Fagan nomogram emitted a raw R error into the results pane. nomogrammer() refuses a positive likelihood ratio below 1, which the permitted specificity range (down to 0.01) can produce, and the call was unguarded. The nomogram is now suppressed with an explanation naming the offending ratio.
• The “Understanding Test Dependence” panel was displayed even under conditional independence, where it describes a model that is not being fitted — the visible: (!indep) case of the leading-! defect described under decisioncompare. Rewritten as (indep == FALSE).
• Post-test probabilities under conditional independence were confirmed against Bayes’ theorem to nine decimal places, and the dependent model was confirmed to reduce to them exactly at zero dependence. The shipped tests had asserted four different values, implying joint likelihood ratios of 120/2.105/3.333/0.0585 where the correct ones are 112/2.526/3.111/0.0702 — the module was right and the expected values were fabricated. Eight further assertions passed an error message containing parentheses as a regular expression, so they never matched, and because testthat re-raises a non-matching error they masked one another.
• The joint-probability validation could never fail. It compared the sum of the four cells against 1, but the caller defines the fourth cell as one minus the other three, so the sum was 1 by construction and the check passed on any input — including a set whose cells no longer matched the sensitivities they were built from. It now also verifies that every cell is a probability and that P(both) + P(one only) still reproduces the marginal it came from, which is the invariant the clamping could actually break.
• Negative conditional dependence is now accepted. The parameter was bounded at 0, so tests that partly compensate for each other’s errors could not be expressed. The permitted range is now −1 to 1 across the option definition, the UI clamp and the backend validation. The existing Fréchet clamping already handled negative values; the resulting joint distributions were verified to be valid and to reproduce their marginals across the full range. For tests with high specificity the feasible negative range is narrow, and values beyond it are truncated with the warning described above.
• Post-test probabilities carry no uncertainty, and now say so where they are read. Sensitivity, specificity and prevalence are typed in by the user and treated as exact, so no reported probability has a confidence interval. That was stated only in the collapsible “Getting Started” panel and is now a note on the results table itself.

sequentialtests (Sequential Testing Analysis)

• Clinical presets were ignored whenever the analysis was called from R. The eight presets are applied by jamovi/js/sequentialtests.events.js, which writes the values into the interface controls; nothing runs that JavaScript outside jamovi, so asking for the HIV scenario silently analysed the panel defaults - 0.95/0.70 with 0.80/0.98 at 10% prevalence - instead of ELISA and Western Blot at 2%. preset was the only declared option the backend never read. Presets are now applied in the backend as well, and a regression test parses the JavaScript configuration and compares it field by field against the R table, so the two cannot drift apart.
• Illustrative teaching values were presented as evidence. The preset control was documented as loading “evidence-based test parameters and optimal strategies from medical literature”. They are rounded approximations with no citation, no confidence interval and no population behind them, chosen to make each strategy’s behaviour visible. Selecting a preset now raises a warning that the figures are for demonstration only and must not be used to design a protocol or advise on a patient; the same caveat appears beside the preset in the interface and in the option documentation. The eleven bundled sequentialtests_* example datasets had no documentation at all; they are now documented, leading with the same warning and spelling out that a single sensitivity figure hides variation by assay, operator and disease stage, that the prevalences are illustrative settings rather than yours, and that a label such as "RT-PCR" names the scenario rather than asserting anything about that test as actually performed.
• A Fagan nomogram was advertised that does not exist. The word appeared in the analysis description and nowhere else in the module - no option, no results item, no code - while every sibling analysis in the same menu does provide one.
• Two of the three strategies are the same rule, and nothing said so. Serial testing of negatives and parallel testing both call a subject positive if either test is positive, so they are algebraically identical in sensitivity, specificity, PPV and NPV; they differ only in how many second tests are performed, which only the cost table reveals. Verified identical to 1e-12. A note now explains it.
• The conditional-independence caveat went stale. It was attached only to parallel testing, and because the summary table declares no clearWith it survived a change of strategy - leaving a note about parallel testing under a row labelled “Serial Testing”. The assumption applies to all three strategies and is now stated for all three, alongside a note that sensitivity, specificity and prevalence are treated as exact so the combined figures carry no confidence interval. Negative conditional dependence, previously forbidden by a lower bound of 0, is now permitted across the -1 to 1 range in the option, the interface clamp and the backend validation; the existing Frechet clamping already handled it correctly. All three strategies were confirmed against a four-million-subject simulation of their actual decision rules, and the conditional-independence figures against Bayes’ theorem to nine decimal places.

enhancedROC (Clinical ROC Analysis)

• Every results table doubled on each re-run, and then broke. Twenty-one addRow() calls and not one deleteRows(). jamovi re-runs an analysis on the same object whenever an option changes, so the AUC summary went 2 to 4 to 6 rows and the cut-off table 32 to 64 to 96; from the second run onward the tables could not be read at all, failing with duplicate 'row.names' are not allowed. Anyone who changed a checkbox mid-session was reading each predictor two or three times over. Fixed on all 18 tables populated by addRow; the two driven by setRow cannot duplicate and were left alone.
• Nothing in 4,500 lines set a random seed, while the analysis resamples in four places: bootstrap AUC confidence intervals, bootstrap ROC comparisons, internal-validation resamples and cross-validation fold assignment. Two identical runs returned 95% intervals of 0.775-0.862 and 0.771-0.862, so a figure copied into a manuscript could not be reproduced. A new Random Seed option (default 0) seeds all four; the caller’s own random number stream is restored afterwards.
• Automatic direction detection biases the AUC upward, and now says by how much. pROC‘s auto-detection reads the direction from the data by comparing the two groups’ medians, and it is the default. Because the direction is then fitted from the same data that supply the AUC, the AUC is inflated: a marker carrying no information at all gives a mean reported AUC of 0.593 at n = 20 and 0.565 at n = 40, against 0.502 when the direction is fixed in advance, and exceeds 0.60 in 43% of runs at n = 20 - enough to turn a null pilot study into an apparently promising biomarker. The notice named the direction chosen but not this consequence, and was filed as informational. It is now a warning quoting the expected AUC for an uninformative marker at the study’s own sample size, and recommending that Direction be set explicitly.
• Three comparison options did nothing, without explanation. Pairwise comparisons, metric differences and statistical comparison all require Analysis Type to be “Comparative ROC Analysis”, but the interface offers them as plain checkboxes with no such dependency; ticking one under the default Analysis Type produced no output and no message. They now say what to change.
• Nineteen options were documented as working features but do nothing. They reach the public R wrapper and ?enhancedROC described them in the present tense - “Calculate Harrell’s concordance index for time-to-event outcomes” - while the backend only lists them in a “planned features” notice at run time. None has an interface control, so this affected R callers and the help page rather than the jamovi menus. All twenty such options, including splineKnots which configures one of them, are now prefixed “NOT YET IMPLEMENTED - selecting this produces no output”.
• test-enhancedroc-pr-clinical.R never exercised the analysis. It built a hand-rolled jmvcore (MockTable, MockHtml, MockImage), assigned a fake base class into the global environment and source()d the backend; the mock implemented only addRow, setRow and asDF, drifted from the real backend, and the whole file died with “attempt to apply non-function”. Rewritten against the real function, keeping its assertions and adding a check that PPV at a user-supplied prevalence equals Bayes’ theorem applied to the table’s own likelihood ratio. AUC itself was confirmed against pROC to nine decimal places under every direction setting.

psychopdaROC (Advanced ROC Analysis)

• Fourteen of the sixteen results tables doubled on every re-run. Twenty-one addRow() calls against two deleteRows(). jamovi re-runs an analysis on the same object whenever any option changes, so rows went 1 to 2 to 3 per predictor and the decision curve 40 to 80 to 120. Every table is now cleared once at the top of the run - after the manual-run gate, so that manual mode keeps its results when an option is edited.
• Leaving the positive class unset ran the whole analysis backwards. The fallback took the FIRST factor level, which for every ordinary coding - Healthy/Disease, Negative/Positive, Control/Case, 0/1 - is the NEGATIVE group. On the bundled example data it reported AUC 0.1001 where naming the positive class gives 0.8999, with no error and no warning. Worse, the positive class was worked out in four places that could disagree, one of which fell back to unique(classVar)[1] - the first value in DATA ORDER, so the answer depended on how the rows happened to be sorted. All four now go through one resolver, which takes the last level and says so on the results tables. With three or more levels it refuses instead of guessing, matching enhancedROC: on a three-level recode of the bundled data, assuming “Disease” gives AUC 0.8012 and assuming “Severe” gives 0.8263, and there is no basis for preferring either.
• The reported “optimal cutpoint” was usually not the optimum. The metric tolerance defaulted to 0.05 while the underlying cutpointr package uses 1e-06. Every cutpoint scoring within 0.05 Youden of the best was treated as equivalent and averaged: on the bundled data 39 of 200 candidate thresholds, spanning 52.1 to 67.7, giving a cutpoint with 84.5 percent sensitivity in place of the 94.4 percent available at the true optimum. The default now matches cutpointr, and any tolerance still in force is disclosed beside the cutpoint.
• The Hanley-McNeil fallback footnote pointed the wrong way. It warned that the approximation “may produce narrower confidence intervals than appropriate”. On the shipped data its standard error is 0.025863 against DeLong’s 0.021092 - 22.6 percent WIDER, i.e. conservative. The formula itself is the standard Hanley-McNeil variance and is correct; only the footnote was wrong.
• Six warning() calls remain invisible to jamovi users (R sends them to stderr, which the GUI does not display). Two of them used to change the analysis rather than merely inform; those are gone, and the analysis has no general notices panel - see Remaining work.
• DeLong’s test could report an AUC of 0.794 for a marker the AUC table reported as 0.206. DeLong’s test needs every marker read in the same direction, so when a marker’s AUC came out below 0.5 the fallback implementation flipped its scores and then displayed 1 - AUC. The primary implementation does not flip - it passes the user’s Classification Direction straight to pROC and reports the honest figure - so the two tables disagreed for exactly the markers where the direction is in question, and the only signal was an R warning(), which a jamovi user never sees. Measured on a constructed marker: 0.206 in the AUC table, 0.794 in the DeLong output. The flip still happens, because the test requires it, but the affected markers are now named both in the DeLong output and in a note on the comparison table, with the instruction to change Classification Direction and re-run if the reversed direction is the clinically correct one.
• DeLong’s test printed confident p-values on samples too small to support them. Its variance is a large-sample approximation, so with few cases in either class the p-values and confidence intervals are too narrow. A note now appears when either class holds fewer than ten cases - the conventional floor for a stable AUC variance - and says that any difference should be treated as provisional. The check sits in the rendering path, so it covers both implementations.
• The test suite was substantially broken and is repaired: test-psychopdaroc.R was 29 failures and 6 errors, and source()d ../../R/psychopdaROC.h.R - capitalised paths that do not exist, so the file only loaded on a case-insensitive filesystem. Its helper in five other files filled every argument lacking a default with "", including dependentVars and classVar, so a test that omitted a required variable was silently handed an empty string. All nine documented examples were uncopyable because none passed refVar, a type: Level option that the jamovi compiler forbids a default on and which is therefore a required argument. AUC, the DeLong test and the optimal cutpoint were confirmed against pROC and cutpointr.

Cross-checking the two ROC analyses

enhancedROC and psychopdaROC sit in the same menu, so the same data can be put through either. They were compared directly on eight datasets. The estimators agree to machine precision: AUC and DeLong confidence intervals were bit-identical across all of them - maximum absolute AUC difference 3.2e-15, maximum CI difference exactly zero - and both matched pROC and a hand-computed Mann-Whitney AUC, including on data rounded to only twenty distinct values. What differs is defaults and scope, and one of those matters:

• Their direction defaults disagree, and both analyses now explain this on screen. For a marker where LOWER values indicate disease - a falling haemoglobin, a falling ejection fraction

  • the same column of numbers came back as AUC 0.8999 from enhancedROC and 0.1001 from psychopdaROC, with nothing on screen saying why. The cause is a default, not a calculation: enhancedROC works the direction out from the data, while psychopdaROC assumes higher values mean disease unless told otherwise. Both answers are correct for the question each is asking - 0.1001 is what you get if you insist higher means disease, and 0.8999 is the same marker read the right way round - and told the same direction the two agree exactly.

  Each analysis now prints the same plain sentence beside its AUC, worded identically so the two outputs can be read side by side:

  Reading of the test values: HIGHER values of biomarker were taken to indicate Disease.

  It goes on to say where that reading came from (specified by you, or worked out from the data) and that every sensitivity, specificity, cutpoint and AUC is reversed if it is the wrong way round. psychopdaROC’s “AUC below 0.5” warning now also names the current Classification Direction setting and states that switching it gives an AUC of 1 minus the value shown, with sensitivity and specificity swapped - so the remedy is explicit rather than left to the reader. Set Direction explicitly whenever the marker’s orientation is known clinically.

• enhancedROC honours a confidence-level option; psychopdaROC has none and always reports 95 percent, so at 99 percent the two give [0.8456, 0.9542] and [0.8585, 0.9412].

• enhancedROC refuses any dataset below n = 20 and returns nothing; psychopdaROC computes and agrees with pROC (n = 16: 0.6406 [0.3236, 0.9576]).

• Their optimal cutpoints differ in the last decimal by construction: pROC places the threshold at the MIDPOINT between two adjacent observed values (54.1979) while cutpointr returns an observed value (54.2904). Sensitivity and specificity are identical either way.

• Given a positive class that does not exist, enhancedROC returns empty tables with an explanatory panel while psychopdaROC raises an R error.

The kappaSize sample-size family (kappaSizePower, kappaSizeCI, kappaSizeFixedN)

• Certain option combinations froze the analysis permanently. kappaSize’s root finder does not converge when the significance level is at or above the target power: a direct PowerBinary(alpha = 0.90, power = 0.20) call was still running when killed after 60 seconds and could not be interrupted, while the same call at alpha 0.05 and power 0.80 returns instantly. The option bounds permitted alpha up to 0.99 and power down to 0.01, so the combination was reachable from the interface and left jamovi with no way to recover. It is now refused before the engine is entered, in 0.07 seconds, with an explanation that a study whose power does not exceed its type I error rate provides no evidence. The same guard removes two related absurdities: power 0.01 used to report “A minimum of 1 subjects”, and alpha just below power produced a sample size of zero rendered as one subject.
• kappa0 was documented backwards in two of the three analyses. kappaSize’s own documentation defines PowerBinary’s kappa0 as “the null hypothesis for the kappa hypothesis test” and CIBinary’s and FixedNBinary’s as “the preliminary value of kappa” - two different quantities that happen to share a name. kappaSizePower described its null as the “Expected value of kappa”, and kappaSizeCI described its anticipated value as “the null hypothesis value of kappa”: each carried the other’s meaning. Someone following the help text would enter the wrong quantity and get a different answer without any indication. All three descriptions now match the package, each says explicitly what it is NOT, and kappaSizeCI’s printed output no longer labels its anticipated kappa “Null hypothesis kappa”.
• The three analyses now agree on their shared options. The significance level was bounded 0.01-0.99 in Power and CI but 0.01-0.20 in FixedN; all three are now 0.001-0.20. The old floor made a Bonferroni-adjusted alpha inexpressible, and the old ceiling admitted values that are not significance levels at all. Proportions parse identically everywhere (commas, semicolons or spaces) - the string 0.30 0.70 was accepted by kappaSizeFixedN and rejected by kappaSizePower. A binary outcome may be entered as one prevalence or as two proportions in all three; kappaSizeFixedN alone demanded two, although kappaSize::FixedNBinary accepts either. A European decimal comma (0,30 0,70) used to report “Each proportion must be strictly between 0 and 1”, which is true and useless; it now names the decimal separator as the problem.
• The power and confidence-interval approaches answer different questions, and each now says so. For the same study they can differ by nearly threefold in required sample size, with nothing in either output explaining why. Each now opens by stating whether it is sizing the study to reject a null value or to achieve a target interval width, and names the other analysis.
• An alternative kappa below the null was answered silently. It is a legitimate question - how many subjects to show agreement is WORSE than the null - but it is more often a transposition, and it returns a different sample size from the mirrored input (149 against 191 on the default scenario). The output now states which way round it read the two values.
• The study explanation never reported the sample size it was explaining, and read “the prevalence of the categories are” for three or more categories. The result panel rendered the raw kappaSize object, duplicating the summary panel and repeating the same expected-cell-count warning ten times in the five-category case; it now shows the headline sentence.
• Sample sizes were confirmed against kappaSize across 540 combinations - four cardinalities, five rater counts, three kappa pairs, three significance levels and three power targets - with zero divergences, plus a 95-case rounding check. The module is a faithful wrapper; everything fixed here sits around the calculation rather than in it.

kappaSizeCI

• A narrow confidence interval froze the analysis with no way out. kappaSize::CIBinary finds its answer by brute force - n <- 10; while (...) n <- n + 1 in interpreted R, with no cap - and the required n grows as roughly one over the square of the distance from kappa0 to the nearer limit. Measured on the binary engine: distances of 0.20, 0.05, 0.01 and 0.005 give 118, 1,625, 38,203 and 151,533 subjects in 0.00 to 1.38 seconds, while 0.0005 had not returned after eight seconds. jamovi cannot abort a running analysis, so the user was simply stuck with no error and no progress. The search now runs under a bounded wall-clock budget (setTimeLimit, which does interrupt an interpreted loop) and, if it expires, refuses with a message naming the distance to the nearer limit and telling the user to widen the interval. Demanding but finite designs are untouched: 0.55-0.65, 0.58-0.62 and 0.59-0.61 still return 1,625, 9,707 and 38,203.
• The explanation reported the wrong quantity as the driver of the sample size. It showed the “Precision width” of the interval, but kappaSize sizes on whichever limit lies nearer kappa0: with kappaL = 0.55 the answer is 1,625 for every kappaU from 0.65 to 0.99. In one-sided mode kappaU is not used at all. The output now states the distance to the nearer limit and says explicitly that this, not the full width, is what determines the number of subjects.
• Proportions were parsed with a character class that did not mean what it looked like. The separator set was written "[,;|\\t]+", which in R is the set of characters , ; | \ and the letter t - it matched a literal backslash and the letter t, but neither an actual tab nor a space. So 0.2, 0.3; 0.5 was accepted and 0.2, 0.3 0.5 was rejected. It is now [,;|[:space:]]+, which handles tabs and spaces properly. A European decimal comma is diagnosed as such rather than reported as an out-of-range proportion.
• kappaSize’s own sparse-cell caveat reached only the Summary pane. When the computed sample size leaves a category with fewer than five expected subjects, the package says so in its summary text - a real warning about the large-sample approximation it relies on, in the one place a reader looking for caveats would not think to look. It is now raised in the Notes panel as well.
• Sample sizes were confirmed against kappaSize across 2,560 comparisons, including the one-sided path, with zero divergences.

kappaSizeFixedN

• The analysis could display a lower bound of -23.78 for Cohen’s kappa. Kappa is bounded below by -1. kappaSize searches by decrementing rho from kappa0 in steps of 0.001 with no floor, so a combination that is inside every declared option bound - kappa0 = 0.01, n = 11, a 2% prevalence, alpha = 0.001 - walks straight past -1, and the result was printed as an ordinary answer. It is now refused, with an explanation that the large-sample approximation has broken down at this combination and what to change. A bound that is negative but still valid (for example -0.293) is reported as before, now with a sentence saying that a bound at or below zero means this many subjects cannot rule out agreement no better than chance.
• A sample size of infinity hung the analysis forever. The guard read is.na(n) || n < 2 || n != round(n), and Inf passes all three clauses - is.na(Inf) is FALSE, Inf < 2 is FALSE, and Inf != round(Inf) is FALSE. Inside the engine the test statistic becomes NaN and the search never terminates. n is now bounded 11 to 1,000,000 at the option level and checked for finiteness and integrality in the backend.
• Sample sizes below 11 produced the package’s own message, which is off by one. Every kappaSize FixedN engine contains if (n <= 10) stop("Sorry, your study should enroll at least 10 subjects."), while the option permitted n down to 2 - so 2 through 10 reached the engine only to bounce back as a raw vendor string that names the wrong threshold. The option floor is now 11 and the message explains that the method is a large-sample approximation.
• The significance level was checked against the wrong bounds in the backend. The backend accepted anything in (0, 1) while the option compiles to 0.001-0.20, so an R caller passing alpha = 0.5 reached the engine and got “missing value where TRUE/FALSE needed”. The two now agree. The proportions-sum tolerance likewise used all.equal, which accepts a sum within 0.001 inclusive, while the engine rejects at 0.001 exclusive; the module now uses the engine’s own predicate so no input can slip past the clear message into the vendor one.
• A rejected re-run left the previous run’s numbers on screen. jamovi calls .run() on the same object for every option change, and validation rejects before anything new is written, so an invalid edit produced a red error above a stale, still-plausible-looking result. All panels are cleared first, as the two sibling analyses already did.
• The study explanation invited a misreading. It opened “Researchers would like to determine the expected lower bound for kappa0=0.6”, which reads as though the bound belonged to kappa0. It does not: kappa0 here is the agreement the researchers anticipate observing, and the bound is the worst case still compatible with it at this sample size. The wording now says so, states the confidence level, reports the answer it is explaining, and says “the prevalence of the trait is” rather than “the proportions of the outcome categories are” when a single binary prevalence was entered.
• A Notes panel was added. It states the method and, in particular, that kappa0 here is the anticipated agreement and not a null hypothesis value as it is in kappaSizePower - the two analyses sit in the same menu and take an identically named argument with different meanings. It warns when a category is expected to contain fewer than five subjects, and warns in red when the achievable bound is at or below zero, the case where the planned study cannot demonstrate agreement at all regardless of what it observes.
• Lower bounds were confirmed against kappaSize across 2,000 combinations with a maximum difference of 6e-16 (print rounding alone). Both monotonicity properties - more subjects raise the bound, a higher significance level raises it - held in 400 of 400 checks each, and the returned bound was below kappa0 in all 2,000.

Changed

• lassologistic score performance figures are now labelled as apparent. “Score AUC” became “Score AUC (apparent)” and “Optimal score cutoff” became “Optimal score cutoff (chosen on this data)”, with a note that these figures are optimistic twice over — the points were derived from a model fitted to these data, and the cutoff was Youden-optimised on the same rows. The model’s own table already carried an optimism caveat; the score’s did not.

Added

• lassologistic gained a configurable cut point for continuous predictors in the scoring system. scoreCutMethod offers median, mean, upper tertile, upper quartile and manual (default median, i.e. the previous behaviour), and scoreCutPoints is a free-text field taking variable=value pairs on the original measurement scale (for example ki67=20, age=65), so an established clinical threshold can be used instead of a data-derived split. Unparseable entries are dropped rather than guessed; any continuous predictor without a manual cut falls back to the sample median, and that fallback is named in a table note. The Scoring System note states that a dataset-derived cut is not an externally established threshold and will differ in another cohort.

Documentation

All 92 files under vignettes/ were audited against the shipped analyses and their current option definitions. These articles are published to https://www.serdarbalci.com/meddecide/articles/; they are excluded from the package tarball by .Rbuildignore, so none of what follows affects R CMD check.

• Twenty-eight articles document an analysis that meddecide does not ship, and now say so. decisionpanel, decisiongraph, decisioncurve, modelbuilder, screeningcalculator, bayesdca, icccoeff, latentbiomarker, advancedtree, decision2 and ppv are all still on development or test menu routes, in this module and in the umbrella ClinicoPath module alike, so they reach no user today. Each affected article carries a note at the top - or, where the article is mainly about a shipped analysis and only borrows one of these, at the point of use - saying that the analysis is documented ahead of a future release, that its options and output may change, and that it will not appear in the jamovi menu yet. No article was deleted.
• 07-kappasizeci-comprehensive.Rmd described kappa0 as a null hypothesis value. For kappaSizeCI it is the anticipated kappa; only kappaSizePower treats it as a null. The article now states which it is and names the other analysis, since the two sit in the same menu and take an identically named argument meaning different things. The same article claimed that the width of the confidence interval drives the required sample size; the calculation actually sizes on whichever limit lies nearer kappa0, so with kappaL = 0.55 the answer is 1,625 subjects for every kappaU from 0.65 to 0.99, and in one-sided mode kappaU is ignored entirely. It also listed the rater options as 2 to 5 where the analysis accepts 2 to 6.
• nogoldstandard-documentation.md still gave the old default. It listed method as defaulting to all_positive in two places; the 1.0.4 default is latent_class, and the reason for the change is now stated alongside it.
• Four articles illustrate agreement() with an argument list that no longer exists. Twenty- eight code chunks use names such as rater1_var, agreement_type, pathologyContext and diagnosticStyleAnalysis, none of which have existed for some time - copying them yields unused argument. Every chunk in these articles is eval = FALSE, so nothing broke and the narrative and statistics remain sound, but the code was uncopyable. Each now carries a notice to that effect together with a verified working call, checked against 1.0.4 on three raters and 60 cases (Fleiss’ kappa 0.734, Krippendorff’s alpha 0.735). A full rewrite of those examples against the current API is still outstanding.
• Argument names in six further call sites were corrected outright: psychopdaROC(class =, value =) is psychopdaROC(dependentVars =, classVar =, positiveClass =), and kappaSizeCI/kappaSizeFixedN take alpha rather than a conf_level, with kappaL/kappaU in place of a single width.
• Every call to a shipped analysis across all 92 files - 345 of them - was checked against the generated wrapper signatures. Apart from the 49 described above, all use argument names the functions accept.

Removed

• Pruned NAMESPACE exports for analyses that this module does not ship, together with unused importFrom declarations. Seven exports were removed (clinicalscore, decisioncurve, latentbiomarker, leaveonecenterout, mageeequation, misclassificationbias, timedependentdca) along with about 25 unused imports (dplyr::arrange/filter/summarise, several ggplot2 and graphics functions, stats::complete.cases, and others). No shipped analysis was affected. .Rbuildignore gained temp and backups patterns.
• boot and survival were dropped from Imports. Neither is used anywhere in the package: no :: call, no NAMESPACE import, and no bare call to any function they export. jamovi installs every package in Imports the first time a module is used, so two unused dependencies and their own dependency trees were being downloaded by every user for nothing.

meddecide 1.0.3 (2026-08-04)

No user-visible changes. Version alignment across the ClinicoPath submodules: the package version and every analysis version string were bumped, and no R code, option definition or results definition was touched.

meddecide 1.0.2 (2026-08-03)

Fixed

• Three analyses named the wrong confidence-interval method. epiR::epi.tests() — the source of the sensitivity, specificity and predictive-value intervals in decision, decisioncalculator and decisioncompare — defaults to method = "exact", which is the Clopper-Pearson interval, not Wilson. On a 2×2 of 67/56/38/79 the two differ in the third decimal (sensitivity 0.5385–0.7296 exact against 0.5428–0.7236 Wilson). The footnotes in decision and decisioncompare said “Wilson score method” and now name Clopper-Pearson. No interval changed — only the label was wrong. decisioncombine was checked and left alone: its own .calcWilsonCI() genuinely is Wilson and genuinely populates the CI table it describes, so its labelling was already correct.

• decisioncompare advertised likelihood-ratio confidence intervals it never computes. The per-test table is filtered to a row set that excludes lr.pos/lr.neg, and the LRP/LRN columns carry no lower/upper bounds — so no likelihood-ratio interval is rendered anywhere in that analysis. The Assumptions panel now states that likelihood ratios are point estimates only, and describes the intervals that are shown: Clopper-Pearson for the proportions, the user’s chosen method (Wilson by default) for Overall Percent Agreement, and Wald for paired differences.

• decisioncalculator overstated the reach of its continuity correction. The notice said the Haldane-Anscombe 0.5 correction was applied to “likelihood ratios, diagnostic odds ratio, and confidence intervals”. It is applied to the point estimates only — epi.tests() is fed the raw table — so the intervals are uncorrected and may be undefined for those statistics when a cell is zero. The notice now says which is which.

• decision crashed whenever “Disease Absent Level” or “Test Negative Level” was left unset. dplyr::case_when() evaluates every branch regardless of which one matches, so the unreachable test == <negative level> comparison still ran with the level NULL, produced a zero-length condition, and aborted the analysis with Can't recycle ... size 0. Both levels now use NA_character_ as the “unset” sentinel, which compares to a full-length, never-matching condition. Behaviour with the levels set is unchanged: an explicitly chosen negative level still restricts the analysis to those two levels and drops rows with any other level.

• Test and gold-standard variables were required arguments of the R function. An option with no default in its jamovi definition compiles to a bare parameter, so calling the analysis from R without it failed with argument "X" is missing, with no default before the analysis’s own validation could produce a usable message. Now defaulting to NULL: agreement (vars), decision (gold, newtest), decisioncombine (gold, test1), decisioncompare (gold, test1, test2) and nogoldstandard (test2). Behaviour in the jamovi GUI is unchanged; no statistical method was altered.

Added

• Automated GitHub release (.github/workflows/release.yaml). A push to the default branch touching DESCRIPTION or jamovi/0000.yaml cross-checks the two version strings, refuses to proceed if they disagree, and — if the tag does not already exist — tags v<version> and publishes a release whose notes are the matching section of this file.

meddecide 0.0.47 (2026-07-05)

Bug Fixes

• Restored correct inter-rater agreement statistics on jamovi installs. vcd and lme4 are used by agreement() but were missing from the package Imports. Because jamovi installs only a package’s Imports, on a clean install they were unavailable: pairwise kappa confidence intervals silently fell back to the narrower irr::kappa2 null-SE method, and the entire ICC / Lin’s CCC / continuous-agreement suite (which relies on lme4) was non-functional. Both are now declared, so agreement CIs use the intended vcd::Kappa asymptotic-SE method and the continuous-agreement measures work.
• Clarified the all-pairs kappa fallback note so it distinguishes a missing vcd package (with install guidance) from a genuinely near-degenerate table.
• Declared DescTools and lmerTest, previously used via :: but undeclared.

meddecide 0.0.46 (2026-07-04)

This release consolidates every change since 0.0.32.69 (unreleased versions 0.0.33 through 0.0.46 roll into this entry). The headline is a large expansion of agreement() into a comprehensive interrater/intrarater reliability suite (20+ new agreement statistics, tests, and visualizations), robustness and input-validation hardening of the ROC modules, one-sided confidence-interval support in the kappa sample-size tools, and package infrastructure updates (minimum jamovi app raised to 2.7.27, new imports, dataset cleanup).

Major Changes

agreement() — Comprehensive Reliability Suite Expansion

The agreement module was expanded from Cohen’s/Fleiss’ Kappa into a full interrater/intrarater reliability toolkit. Each new statistic ships with its own results table, an “About …” HTML explanation, and a “When to use …” guide notice.

• New chance-corrected / categorical measures:
  • Gwet’s AC1/AC2 (gwet, with gwetWeights: unweighted/linear/quadratic) → gwetTable
  • PABAK with prevalence and bias indices (pabak) → pabakTable
  • Light’s Kappa for 3+ raters (lightKappa) → lightKappaTable
  • Krippendorff’s Alpha guidance (showKrippGuide)
• New continuous-data agreement measures:
  • ICC with six models ICC(1,1)–ICC(3,k) (icc, iccType) → iccTable
  • Mean Pearson correlation (meanPearson) → meanPearsonTable
  • Lin’s Concordance Correlation Coefficient (linCCC) → linCCCTable
  • Total Deviation Index (tdi, tdiCoverage, tdiLimit) → tdiTable
  • Finn coefficient with one-way/two-way models (finn, finnLevels, finnModel) → finnTable
  • Iota multivariate coefficient (iota, iotaStandardize) → iotaTable
• New ordinal / rank-based agreement measures:
  • Kendall’s W coefficient of concordance (kendallW) → kendallWTable
  • Robinson’s A ordinal agreement index (robinsonA) → robinsonATable
  • Mean Spearman rho (meanSpearman) → meanSpearmanTable
• New marginal-homogeneity / rater-bias tests:
  • Rater Bias test (raterBias) → raterBiasTable
  • Bhapkar test (bhapkar) → bhapkarTable
  • Stuart-Maxwell test (stuartMaxwell) → stuartMaxwellTable
  • Maxwell’s RE random-error index (maxwellRE) → maxwellRETable
• New multi-rater / structural analyses:
  • Pairwise Kappa against a reference rater with performance ranking (pairwiseKappa, referenceRater, rankRaters) → pairwiseKappaTable
  • Hierarchical/multilevel Kappa (hierarchicalKappa, clusterVariable) with cluster-specific estimates, variance-component decomposition, hierarchical ICC, cluster-homogeneity test, and shrinkage (empirical Bayes) estimates → hierarchicalOverallTable, clusterSpecificTable, varianceDecompositionTable, hierarchicalICCTable, homogeneityTestTable
  • Mixed-effects condition comparison with Bonferroni/BH/Holm correction (mixedEffectsComparison, conditionVariable, multipleTestCorrection) → mixedEffectsTable, mixedEffectsVarianceTable
  • Inter/intra-rater test-retest reliability (interIntraRater, interIntraSeparator) → interIntraRaterIntraTable, interIntraRaterInterTable
• New machine-learning-style metrics:
  • Confusion matrix table with row/column normalization (confusionMatrix, confusionNormalize) → confusionMatrixTable, perClassMetricsTable
  • Multi-annotator concordance / F1 (multiAnnotatorConcordance, predictionColumn) → concordanceF1Table, concordanceF1PerClassTable
  • Specific (category-focused) agreement indices with optional CIs (specificAgreement, specificPositiveCategory, specificAllCategories, specificConfidenceIntervals) → specificAgreementTable
  • Bootstrap confidence intervals (bootstrapCI, nBoot) → bootstrapCITable
• New visualizations:
  • Agreement heatmap / confusion-matrix plot (agreementHeatmap) with color schemes (blue-red, traffic-light, viridis, grayscale) and count/percentage cell annotations (heatmapColorScheme, heatmapShowCounts, heatmapShowPercentages, heatmapAnnotationSize) → agreementHeatmapPlot
  • Bland-Altman method-comparison output with a Shapiro-Wilk normality check (showBlandAltmanGuide) → blandAltmanHeading, blandAltmanExplanation
  • Rater profile plots — box/violin/bar (raterProfiles, raterProfileType, raterProfileShowPoints) → raterProfilePlot
  • Rater clustering and case clustering with dendrograms and heatmaps — hierarchical/k-means, correlation/euclidean/manhattan/agreement distances, average/complete/single/ward linkage (raterClustering, caseClustering) → raterClusterTable, raterDendrogram, raterClusterHeatmap, caseClusterTable, caseDendrogram, caseClusterHeatmap
  • Stratified agreement-by-subgroup with forest plot (agreementBySubgroup, subgroupVariable, subgroupForestPlot, subgroupMinCases) → subgroupAgreementTable, subgroupForestPlotImage
• New workflow tools:
  • Paired agreement comparison between two rater conditions with bootstrap (pairedAgreementTest, conditionBVars, pairedBootN) → pairedAgreementTable
  • Sample-size calculator for agreement studies supporting Cohen’s Kappa / Fleiss’ Kappa / ICC (agreementSampleSize, ssMetric, ssKappaNull, ssKappaAlt, ssNRaters, ssNCategories, ssAlpha, ssPower) → agreementSampleSizeTable
• New computed output variables:
  • Consensus rating variable with majority/supermajority/unanimous rules and tie handling (consensusVar, consensusName, consensusRule, tieBreaker) → consensusTable, consensusVar
  • Case-level Level-of-Agreement categorization — simple/detailed with custom/quartile/tertile thresholds (loaVariable, detailLevel, simpleThreshold, loaThresholds, loaHighThreshold, loaLowThreshold, loaVariableName, showLoaTable) → loaTable, loaDetailTable, loaOutput
• Other agreement additions:
  • Configurable confidence level for CIs (confLevel)
  • Level-ordering information panel (showLevelInfo) → levelInfoTable
  • Plain-language Summary, About, and Clinical Use Cases panels (showSummary, showAbout) → summary, about, clinicalUseCases
  • New client-side events handler jamovi/js/agreement.events.js (bounds/dependency handling for confLevel, Bland-Altman confidence level, cluster counts, and subgroup minimums)

Enhanced Existing Functions

• enhancedROC(): Robustness and UX overhaul

  • Rewritten input validation: guards for single-value outcomes, outcome-level checks, and positive-class validation
  • HTML-escaped error/warning messages (private$.safeHtmlOutput) and a notices framework (.addNotice/.renderNotices) plus a methods-explanation panel and instructions
  • Sensible defaults now ON: youdenOptimization, rocCurve, aucTable, optimalCutoffs, diagnosticMetrics; customCutoffs now defaults to empty
  • Added clearWith to results so outputs invalidate correctly; removed dead commented-out time-dependent AUC/ROC stubs

• psychopdaroc(): Input hardening and cleanup

  • Bootstrap/threshold count options changed from Number to Integer (boot_runs, maxThresholds, bootstrapReps, idiNriBootRuns)
  • Removed dead commented-out option stubs (effectSizeMethod, advancedMetrics); added clearWith to results

• kappasizeci(): One-sided confidence intervals

  • New citype option (two-sided vs one-sided lower-bound-only), wired to kappaSize::CIBinary/CI3Cats/CI4Cats/CI5Cats, with a UI ComboBox that disables the upper-limit input in one-sided mode
  • New plain-language text_summary output (CI type, lower limit, precision width)

• kappasizefixedn() / kappasizepower(): New plain-language text_summary output panel

• nogoldstandard(): New notices panel (“Important Information”) with plain-text notice rendering that resets on each run

• decisioncalculator(): Sensitivity/specificity confidence intervals now use a logit transformation with continuity correction (sens_se = sqrt(1/TP + 1/FN), spec_se = sqrt(1/TN + 1/FP)) when a zero cell is present — consistent with the existing PPV/NPV CI logic — falling back to exact Clopper-Pearson binomial CIs otherwise

Package Infrastructure

• Version bumped 0.0.32.69 → 0.0.46; release date 2026-07-04; minimum jamovi app raised to 2.7.27 (minApp)
• New Imports: ggraph, grDevices, graphics, htmltools, igraph, irrCAC, knitr, stats, tibble, tools (irrCAC/stats/psych back the new chance-corrected, clustering, and ICC agreement measures; ggraph/igraph back the reimplemented decisiongraph tree visualization; htmltools/knitr/tibble/tools/grDevices/graphics support HTML output and plotting)
• Switched documentation config to Config/roxygen2/version: 8.0.0
• New shared helper files: R/diagnostichelpers.R (reusable sensitivity/specificity/PPV/NPV helpers) and R/error_handling.R (clinical error-handling framework: clinicopath_init, clinicopath_error_handler)
• Added .escapeVariableNames and refactored/hardened existing R/utils.R helpers (%notin%/%!in% rewritten as explicit functions, load_required_package default flipped to install_if_missing = FALSE, print.sensSpecTable given an S3-compliant (x, ...) signature)
• R/decisiongraph_utils.R: decision-tree visualization reimplemented as a real igraph/ggraph dendrogram renderer (graph_from_data_frame, ggraph::geom_edge_diagonal/geom_node_point/geom_node_label, horizontal/vertical/radial layouts) replacing the “Not Yet Implemented” placeholder, plus removal of the hardcoded 0.7/0.2/0.1 Markov transition-matrix stub (decisiongraph is a shipped utility, not a registered menu analysis)

Data

• Removed five raw CSV files from data/ (cancer_biomarker_data.csv, cardiac_troponin_data.csv, dca_test_data.csv, sepsis_biomarker_data.csv, thyroid_function_data.csv)
• Added roxygen dataset documentation in R/data.R for the packaged datasets (histopathology, Bayesian DCA, breast cancer, breast/lymphoma diagnostic-styles, thyroid function, and the cancer/cardiac/sepsis/thyroid diagnostic sets plus the combined master collection)

Minor Changes

• Module-wide label cleanup: replaced “%” with “percent” and removed emojis across decision(), decisioncalculator(), decisioncompare(), decisioncombine(), cotest(), and sequentialtests() labels and descriptions (for decisioncalculator() this covers the label/emoji changes only; the CI methodology change is documented under Enhanced Existing Functions above). decisioncombine() also gained allowNone: true on its Test 3 positive-level option and had its example dontrun flag flipped to true
• Version strings synchronized to 0.0.46 across DESCRIPTION and all jamovi analysis definitions

---

meddecide 0.0.32.69 (2026-01-02)

New Features

• bootstrapNRI(): Exported bootstrapNRI function for Net Reclassification Improvement (NRI) bootstrap confidence interval estimation
  • Enables direct access to NRI bootstrap analysis
  • Provides robust confidence intervals for categorical and continuous NRI
  • Supports custom thresholds for risk category definitions
  • Configurable bootstrap iterations and confidence levels

Bug Fixes

• Fixed critical bug in computeNRI() where risk category labels were incorrectly calculated
  • Corrected the labels vector calculation from 1:length(breaks - 1) to 1:(length(breaks) - 1)
  • This fix ensures proper risk categorization in NRI calculations
  • Affects categorical NRI computations in ROC and psychoPDA analyses
• agreement(): Fixed stability issues and hanging during initial run
  • Refactored agreement.b.R to ensure responsiveness
  • Maintained support for numeric variables in agreement analysis

Minor Changes

• Updated package version to 0.0.32.69
• Synchronized version across DESCRIPTION and jamovi module files (jamovi/0000.yaml)

---

meddecide 0.0.31.84 (2025-10-03)

Major Changes

New Analysis Functions

• decisioncombine(): New function for systematic evaluation of diagnostic test combinations
  • Analyzes all possible test result patterns (2-test: 4 patterns, 3-test: 8 patterns)
  • Calculates sensitivity, specificity, PPV, NPV, and accuracy for each pattern combination
  • Identifies optimal testing strategies based on Youden’s J index
  • Includes visualization options: bar charts, heatmaps, forest plots, and decision trees
  • Supports filtering by statistic type and pattern type
  • Can add test pattern column to dataset for further analysis
• cotest(): New function for analyzing combined results of two concurrent diagnostic tests
  • Calculates post-test probabilities for various scenarios (either positive, both positive, both negative)
  • Supports both parallel and serial testing strategies
  • Provides Fagan nomogram visualizations
• sequentialtests(): New function for sequential testing analysis
  • Analyzes how diagnostic accuracy changes when applying two tests in sequence
  • Compares three different testing strategies: serial positive (confirmation), serial negative (exclusion), and parallel testing
  • Provides comprehensive analysis including population flow, cost implications, and Fagan nomograms
• decisioncalculator(): New calculator for diagnostic test evaluation
  • Designed for when you have the four key counts: TP, FP, TN, FN
  • Calculates comprehensive diagnostic performance metrics
  • Supports confidence interval estimation and Fagan nomogram visualization

Enhanced Existing Functions

• decisioncompare(): Major improvements to test comparison functionality
  • Enhanced comparison plots (bar charts and radar plots)
  • Added statistical comparison using McNemar’s test
  • New summary and explanation options for better interpretation
  • Added manuscript-ready report sentence generation
  • Improved handling of custom prevalence settings
  • Better visualization of confidence intervals for metric differences

Removed Features

• decisionpanel(): Function removed for future redesign
  • Users should use decisioncombine() and decisioncompare() instead
  • These new functions provide more focused and comprehensive analysis

Menu Organization

• Reorganized jamovi menu structure for better user experience
  • Decision: Core diagnostic test evaluation functions
  • Decision Calculators: Calculator-based tools for specific scenarios
  • ROC: ROC curve analysis functions
  • Agreement: Interrater reliability functions
  • Power Analysis: Sample size calculation functions

Minor Changes

• Updated agreement() function with improvements to reliability assessment
• Enhanced documentation across all functions
• Improved error handling and validation
• Updated example datasets and usage examples

Bug Fixes

• Fixed various edge cases in diagnostic metric calculations
• Improved handling of missing data
• Enhanced validation of input parameters

---

meddecide 0.0.31 (2025-09-18)

Package Updates

• Version synchronization across DESCRIPTION and jamovi module
• Updated package metadata and author information
• Enhanced package description with comprehensive feature list

Documentation

• Improved function documentation with clearer examples
• Updated pkgdown website structure
• Added more detailed usage examples for main functions

---

meddecide 0.0.3.91

New Features

• Initial implementation of test comparison framework
• Added support for Fleiss’ Kappa with differentiated method names
• Enhanced Kappa calculation methods

Bug Fixes

• Fixed issues with exact Kappa calculations
• Improved handling of multiple rater scenarios

---

meddecide 0.0.3.90

Initial Release Features

• Basic diagnostic test evaluation functions
• ROC analysis capabilities
• Interrater reliability assessment (Cohen’s Kappa, Fleiss’ Kappa)
• Sample size calculations for reliability studies
• Visualization tools including Fagan nomograms
• jamovi module integration