Changelog
Source:NEWS.md
meddecide 1.1.4 (2026-11-08)
Documentation and release review
- Interrater Reliability UI: an available-category hint beside the positive-category box combines the categories from all selected raters and refreshes after selection or level changes. The welcome panel now stacks Quick Start and Available Methods as headings and lists. Entered categories and statistical calculations are unchanged.
- Corrected the fixed-N kappa guide: the result is an expected one-sided lower confidence bound, not achieved power, a minimum detectable kappa, or a guarantee about a future study’s interval.
- Corrected confidence-limit planning: the limit requiring the larger N governs; it is not always the geometrically nearer limit. Distinguished anticipated kappa in CI/fixed-N planning from the null kappa in power planning, and documented the common-margin model, up to six raters and the need for design-specific simulation.
- Added a guide covering all 15 analyses: reference standards, prevalence transport, denominator differences, conditional dependence, cutpoint optimism, calibration, internal versus external validation, decision-curve utilities and latent-class identifiability. Documented the lack of case-control prevalence correction in DCA.
- Updated README installation requirements to 1.1.4, qualified verification claims, and added a maintained README.Rmd source and links to the methods guide.
- Recorded the 2026-11-08 release review with current test/compiler evidence and unresolved submission checks. Local validation is separate from jamovi library acceptance; bundled dependency licensing and platform acceptance remain review items.
This documentation update changes no estimator, statistical default, clinical cutoff, analysis schema or saved-analysis contract. The 1.1.3 entry below records the earlier implementation changes. See docs/audit/RELEASE_READINESS_1.1.4_20261108.md for the scope and limitations of the current review.
meddecide 1.1.3 (2026-11-07)
No estimate or test changes: every sensitivity, specificity, predictive value, likelihood ratio, kappa, AUC, net benefit and sample size is computed as in 1.1.2, and a pre-release review checked them against independent packages (see the end of this entry). Saved analyses open as before, except that the two ranges below return to their defaults (see the R note).
Corrections from the pre-release review
- Medical Decision no longer calls an indistinct test “poor” or “worse than chance”. When the 95% confidence interval for Youden’s index includes 0, the Clinical Summary now says the test shows “no statistically detectable” discrimination and the Clinical Interpretation panel says “No evidence of discrimination in this sample”, matching the warning above them. Before, the warning said not to call the test poor while the panels below it did.
- Interrater Reliability matches the positive category under the default settings. With “All-category estimates” ticked (the default), typing “pos” for a level “Pos” labelled both rows “(Negative)”; extra spaces and different capitalization are now matched in every mode, a category that matches nothing raises a warning instead of being ignored, and a box holding only spaces counts as empty.
- Medical Decision Calculator: F1 score is 0, not blank, when there is no true positive (2TP / (2TP + FP + FN) = 0), and a positive predictive value that cannot be estimated (no positive results) no longer shows an interval of 0 to 1.
- Interrater Reliability: Gwet’s AC with unanimous ratings leaves the standard error, interval and p-value blank and says why, instead of showing an interval of [1, 1].
-
Advanced ROC Analysis:
- the DeLong comparison table warns when a marker has an AUC below 0.5 in the selected direction, since a “significant” difference then reflects reading that marker backwards;
- the confusion matrices show their note explaining which cells are true and false positives and negatives (it was set and then erased before display);
- no matrix is drawn at an infinite cut-point when no usable cut-point exists, and no empty per-variable matrix appears next to the subgroup matrices;
- Optimization metric and Tie breaking method are disabled when the cutpoint method fixes its own criterion (for example kernel or normal Youden), fixed-sensitivity/specificity shortfall note keys no longer collide for variable names that differ only by spaces versus dots (
A BvsA.B), and failed parametric/kernel ROC curve smoothing emits a user-facing warning.
-
Clinical ROC Analysis:
- sensitivity and specificity 95% confidence interval text in the Diagnostic Performance table always retains three decimal places (
(%.3f, %.3f)); - a partial AUC that falls below the chance diagonal suppresses the raw R console warning and retains its explanatory table note even when explanatory footnotes are hidden;
- the Validation Context table marks the default Unit of observation (
slide (default setting)).
- sensitivity and specificity 95% confidence interval text in the Diagnostic Performance table always retains three decimal places (
-
Interrater Reliability, Medical Decision, Compare Medical Decision Tests, and Decision Curve Analysis:
- enabling Hierarchical kappa without a cluster variable (or Mixed-effects comparison without a condition variable) raises a single warning notice instead of repeating it for each sub-table;
- Medical Decision guards Youden direction against 32-bit integer overflow at very large sample sizes and adds a truncation note when discordant comparator cases exceed
maxCasesShow; - Compare Medical Decision Tests notes missing observations omitted from the original-data cross-tabulation and names the selected positive and negative levels on each 2x2 table;
- Decision Curve Analysis notes when a model’s observed range of benefit reaches the boundary of the analyzed threshold window.
- Corrected sentences. Medical Decision now gives the actual reason a likelihood ratio is not estimable, and no longer prints an NPV of 99.975% as “100.0%”. The specific-agreement interval note states the count it uses (2a + b + c, each agreed case counted twice). A specific agreement between 0.75 and 0.80 is no longer called both “Good” and “Critical”. Decision Curve Analysis says when its intervals are pointwise, and that an interval clearing the treat-all line is not a test of the difference. LASSO Logistic says that high collinearity makes the choice of correlated predictors unstable, not the predictions unreliable, and stops suggesting elastic net when elastic net is already fitted. Medical Decision Calculator’s summary no longer points to predictive values “above” that were moved into the tables.
- Smaller wording fixes. Sequential Testing Analysis no longer reports default values as “entered” when a teaching example is chosen from R. Compare Medical Decision Tests names the level boxes as they appear in the panel, and does not advise reading an interval for a class with no cases. Interrater Reliability’s concordance table names the prediction rater instead of its position. Turkish translations cover every new message.
Changes in this release
- Tables instead of HTML. Advanced ROC Analysis shows its confusion matrices as tables (rows: the decision at the cutpoint; columns: the selected positive and negative levels). Compare Medical Decision Tests shows its original data as two tables: reference-standard counts, and each test against the reference levels. The LASSO Logistic Regression suitability check is a table whose status column says “No concern”, “Marginal” or “Major concern” in words. The bordered summary grids in the Medical Decision Calculator and in Interrater Reliability are gone; those numbers are in their tables. All of them now follow your number format and copy into a spreadsheet as cells.
- Results stay in place when an input is invalid. jamovi greys them and shows the error over them; panels no longer vanish or go blank while you type through an invalid value.
- Tables keep their shape while computing. Comparison, agreement, McNemar and difference tables in Compare Medical Decision Tests, the combination tables in Combine Medical Decision Tests and the workload table in the Medical Decision Calculator show their rows from the start. A pair of tests that cannot be compared, and Cochran’s Q when it is undefined, now keep their row and say “Not computed”.
- Variables first. In Interrater Reliability and Decision Curve Analysis every variable box is in the top variable list, greyed until the analysis that uses it is ticked. In Medical Decision and Compare Medical Decision Tests the variables come before the reference-interpretation box.
- Choose, don’t type, the negative reference level in Clinical ROC Analysis’ frozen-model validation: it is now a level list. In Interrater Reliability the positive category is still typed, and an extra space or different capitalization no longer stops it matching.
-
R users:
enhancedROC()now has a required argumentvalidationNegative, likepositiveClass. PassvalidationNegative = NULLunless you use the frozen-model report. - Clinical presets show what they override. In Clinical ROC Analysis, choosing a clinical preset greys Youden index optimization and, except for Research, the minimum sensitivity and specificity boxes, since the preset sets them.
- Panels grouped by the question you are asking. Interrater Reliability’s 38 option boxes are now nine: General Options, Categorical, Ordinal, Continuous, Any Data Type, Tests, Rater and Case Diagnostics, Computed Variables and Study Design, each method under its own heading. The shared settings (exact kappa, multiple-testing correction, number of bootstrap samples) are in General Options. Combine Medical Decision Tests splits its display options into Tables, Rule Ranking, Plots (with the plot filters) and Save to Data. In Clinical ROC Analysis the analysis settings are in their own box, each heading holds the options under it, and the cost ratio sits under the cutpoint method that uses it. Co-Testing Analysis’ display options are a collapsed box. Nothing is computed differently, and saved analyses keep their settings.
- Interrater Reliability results are grouped under their section headings. Each heading and its tables form one group that shows only when something in it is shown, so the empty heading lines are gone from the output and from copies and exports. Results saved in an older file are recomputed when it is opened.
- Name the prediction rater in Interrater Reliability’s multi-annotator concordance, in a variable box, instead of typing its position in the Raters list. It must also be one of the raters. The “Standardize quantitative ratings” box is gone, as it never changed single-variable iota.
-
R users: in
agreement(),predictionColumn = 1is nowpredictionRater = "<column>". Results that sat under a heading are one level down, for exampleresult$krippTableGroup$krippTable. - Advanced ROC Analysis no longer offers “Explain why within-cohort meta-analysis is unavailable”; the option remains for saved analyses and R calls.
- Clinical ROC Analysis plots follow jamovi’s theme. Plot theme has a new first choice, “jamovi (follow global)”, which is now the default, so these plots match the rest of the document. Classic, Modern and Clinical (the previous default) are still there to override it.
- A more precise verification-bias caution in Combine Medical Decision Tests: when the reference standard was obtained only after a positive test, sensitivity, specificity and the predictive values of single tests and combined rules can be biased, but an exact result pattern’s predictive values are not, if verification depended only on that pattern.
- Worked examples name their tests in your language. In Co-Testing Analysis, the test names a worked example puts in the name boxes are shown translated in the tables and summary; a name you type yourself is kept as typed.
- Fagan nomograms follow jamovi’s colors and name their lines. In Medical Decision, the Medical Decision Calculator and Co-Testing Analysis, the two lines now use the palette chosen in jamovi, and a legend says which is the positive result and which the negative. “Fagan line colors” -> “Red / blue (classic)” brings back the old red and blue.
- Predictor names that start with “<” show in Clinical ROC Analysis tables. A variable named, for example, “<LOD flag” appeared as a blank cell in 18 tables; it now shows as “< LOD flag”.
- Sample-size numbers follow your number format. Interrater Reliability’s Sample Size for Agreement Study table now holds its values as numbers rather than text formatted in R; the agreement metric is named in its first row.
- One label style. Variable boxes and result titles are in title case; options and list choices in sentence case. Youden’s index has one name (columns: “Youden’s J”); “95% confidence intervals” reads the same in Medical Decision, the Medical Decision Calculator and Compare Medical Decision Tests; and two options that shared a title now have their own (“Known population prevalence” and “Population disease prevalence”; “Bootstrap interval for Krippendorff’s alpha”). The kappa sample-size labels and their messages name the inputs one way: κ₀, κ₁, κL and κU instead of the R argument names in the CI and power analyses, and “the anticipated kappa” in the fixed-sample analysis, the prevalence box gives its real range (between 0 and 1), Interrater Reliability’s percent boxes show “%”, and its method headings use one style of brackets. Clinical ROC Analysis says “cutoff” throughout. Only labels changed: saved analyses and R calls work as before.
- Validation context boxes greyed when unused. In Medical Decision, cohort role, sampling design, threshold source, model version and unit of observation are greyed until Validation context is ticked, as they already were in Compare Medical Decision Tests.
- Ranges as number boxes. Clinical ROC Analysis’ partial AUC range is two boxes, Partial AUC from and to, and Advanced ROC Analysis’ treatment threshold range is three, from, to and step; each box checks its own value. A saved analysis that used another range returns to the defaults (0.8 to 1, and 0.05 to 0.50 in steps of 0.05). Lists of any length, such as cutoffs and target prevalences, stay as comma-separated text.
-
R users: in
enhancedROC(),partialRange = "0.8,1.0"is nowpartialAucFrom = 0.8, partialAucTo = 1; inpsychopdaROC(),treatmentThreshold = "0.05,0.5,0.05"is nowtreatmentThresholdMin,treatmentThresholdMaxandtreatmentThresholdStep. - Notices say how serious they are in words in Compare Medical Decision Tests, Decision Curve Analysis and the three kappa sample-size analyses, not by colour alone. An invalid input is reported once, as the analysis error, in Advanced ROC Analysis and LASSO Logistic Regression. Interrater Reliability no longer repeats its “kappa not computed for continuous ratings” note in a second panel, and its computed-variables message is translated.
- A sub-analysis that cannot run says so at the top. In Interrater Reliability, an optional analysis that cannot be computed (rater clustering with two raters, Bland-Altman with fewer than three complete pairs, and so on) raises a warning naming it above the results, instead of a footnote under its empty table.
- Help only when it helps. Co-Testing Analysis no longer shows its guidance panel by default (tick Guidance and explanations to see it), and Clinical ROC Analysis shows its instructions only until an outcome and a predictor are chosen.
- Smaller saved files. Interrater Reliability’s rater-profile plot keeps the category counts it draws, not every rating, and Advanced ROC Analysis’ dot plot stores its threshold and labels once; analyses saved earlier still open and draw as before.
- Severity labels in sentence case. Notices start with “Error:”, “Strong warning:”, “Warning:” or “Info:” instead of capitals, and the confusion-matrix headers in ROC analysis read “Decision based on measure” and “Criterion”.
- American spelling throughout (analyzed, modeled, color, behavior, center, favor). The Turkish translations follow.
- Repository: a personal settings file is no longer tracked.
- Documentation: the vignettes cover only the 15 analyses this module ships, use the current R arguments, and state the Landis and Koch kappa bands correctly; pages about analyses that are not part of meddecide (decision trees, Markov models, decision panels and others) and the legacy copies were removed. The README examples run as written.
What the pre-release review verified
Independent checks against other packages and hand formulas, on the code released here: 2 x 2 accuracy measures, Clopper-Pearson and Wilson intervals, prior-adjusted predictive values, the zero-cell likelihood-ratio rule and Fagan post-test probabilities (binom, DescTools); co-testing and sequential-testing formulas; McNemar, Newcombe paired differences and Cochran’s Q (DescTools); net benefit (dcurves); LASSO coefficients and lambda (glmnet); latent class estimates (poLCA); Cohen’s, weighted, Fleiss’, Conger’s and Light’s kappa, Krippendorff’s alpha, Gwet’s AC1/AC2, all six ICC forms and Lin’s CCC (irr, irrCAC, psych, vcd, DescTools); AUC, DeLong tests, cut-points, partial AUC, NRI/IDI, Brier score and Hosmer-Lemeshow (pROC, ResourceSelection); and every kappa sample-size calculator (kappaSize). These checks use simulated data; they show the calculations are right, not that a test is clinically valid.
meddecide 1.1.2 (2026-10-23)
-
Example dataset update: replaced the no-gold-standard Data Library entry for pathologist readings with
nogoldstandard_validation.omv: 190 simulated patients, three imperfect binary diagnostic tests, a GUI-configured latent-class analysis, 1,000 bootstrap refits, 95% confidence intervals, seed 42, diagnostics, a summary and a teaching annotation. Neither reference test is treated as a gold standard. The example is not clinical validation. - Documentation: a worked example for the no-gold-standard analysis, and updated Turkish translations.
meddecide 1.1.1 (2026-10-21)
Independent-review corrections (2026-10-21):
- psychopdaROC: IDI and NRI no longer replace a marker whose risk runs against the Classification Direction with a constant base rate, which could flip their sign (a stronger “lower = disease” reference read as the weaker marker). Both risk models are fitted freely, and a table note names any marker whose association runs against the selected direction.
- Zero cells (decision, decisioncalculator, decisioncompare, decisioncombine): one shared rule. LR+, LR- and the diagnostic odds ratio are 0.5-corrected only when their denominator cell is empty; an empty numerator gives an exact 0 with no log-scale interval; a ratio whose correction would not stay above 1, or 0/0, is left blank. Previously TP = 0 could print LR+ 3.06. A ratio of exactly 0 shows the interval [0, U], U from the 0.5-corrected log-scale interval, so a perfect small study does not read as certain. Fagan nomograms are not drawn for a zero or blank likelihood ratio.
- enhancedROC: calibration, Brier, Hosmer-Lemeshow, decision and clinical impact curves use the fitted risk model when its slope runs against Direction, with a warning, instead of a silent constant-prevalence model. The analysis is rejected when no predictor can be analysed, and the AUC interval method is always named.
- decision: the negative-level boxes start at None, so jamovi no longer pre-selects the first level (with three levels it could take “Equivocal” as negative). A variable with several candidate negatives waits for the choice; an inferred negative is shown.
- decisioncompare: when tests were done on different cases, the best-test sentence ranks them on the cases common to all; the prevalence sentence names its denominator; an empty level box is an information notice, not an error.
- nogoldstandard: the penalized EM method now shows bivariate residuals and the dependence warning; three-test models carry a warning that dependence cannot be checked. The vignette runs and describes complete-case handling and the fixed priors.
- Stale results: notices are cleared when a re-run has nothing to report (cotest, decision, decisioncombine), and kappaSizeCI no longer leaves the previous sample size under an error. Undefined cells in decisioncombine are blank instead of “NaN”.
- Plot text no longer runs off the edge at the default size: the cotest Fagan caption (which hid its “not for clinical use” disclaimer), two decisioncurve captions, and the enhancedROC clinical-utility legend and CROC caption.
- Wording made accurate: agreement 3+ rater subgroup intervals (Gwet SE, t) and minimum-case rule; kappaSize methodology and governing-limit statements; decisioncurve “range of benefit” (observed, in-sample) and case-control notice; lassologistic optimism-corrected AUC; sequentialtests conditional-independence formulas; rare priors print with enough digits; psychopdaROC tie, fixed-sensitivity and unit-cost notes.
- Turkish translations added for the new and reworded messages.
Earlier 1.1.1 corrections:
- Fix full-response serialization when citation pages are numeric in YAML.
- Report 0% categorical unanimity correctly for valid complete disagreement.
- Correct the explanation of selective zero-cell likelihood-ratio corrections; estimates and interval methods are unchanged.
- Add regressions for citation types, zero unanimity and correction explanations.
- Retain epiR and its calculations. pander declares AGPL-3 or the OSL-3.0 in its LICENSE file; library acceptance can be discussed with the jamovi team if needed.
meddecide 1.1.0 (2026-10-20)
- Packages the medical-decision, ROC, agreement and kappa-planning analyses at version 1.1.0. The statistical and interpretation changes since 1.0.81 are documented in the 1.0.9 notes below.
- Adds installed-package decision and ROC-wrapper regression checks with synthetic, reproducible fixtures and testthat edition 3.
The earlier 1.1.0 archive is superseded by the 1.1.1 candidate. Local checks and library submission status are recorded separately in the release preparation report.
meddecide 1.0.9 (2026-10-17)
The analyses were validated against independent calculations and then independently reviewed. Much of what follows changes a number, label or verdict you may have reported, so behaviour changes come first; a saved .omv keeps its old results until you re-run it. This is the first release since 1.0.81, and it also contains the 1.0.83.03 changes below, never released on their own.
Behaviour changes – saved analyses may produce different results
Across the module
- The positive (event) level is no longer filled in for you. jamovi used to pre-select the outcome’s first level, usually the negative one (“No”, “0”, “Benign”), so a new analysis could compute everything for the wrong event. In Clinical ROC Analysis, Advanced ROC Analysis, Compare Medical Decision Tests, Analysis Without Gold Standard, Decision Curve Analysis and LASSO Logistic Regression the level boxes now start empty, and the analysis waits until you choose. A saved analysis keeps the level it was saved with: check it when you reopen the file.
-
Out-of-range entries are reported, not silently replaced. In Co-Testing Analysis and Sequential Testing Analysis, typing
80for an 80% sensitivity was quietly changed to 0.99 and the analysis ran on that value. It now tells you the entry is out of range.
Interrater Reliability (agreement)
- Kappa for three or more raters uses every observed rating. One missing reading used to drop the whole case. Fleiss’ and Conger’s kappa now keep the other readers’ ratings and add a standard error, interval and p-value; values are unchanged on complete data.
- Weighted kappa keeps the grade order when raters used different grades: G1/G2/G4 for one rater and G1/G2/G3 for another built the scale G1, G2, G4, G3 (0.370 instead of 0.213). All pairs, subgroups and institutions now share one scale.
- Stuart-Maxwell and Bhapkar use the full table. Dropping categories both raters used equally often could give chi-squared = 4.00, p = 0.046 where the correct test gives 0.145, p = 0.930.
- Other coefficients: Krippendorff’s alpha is exact for 3+ raters with complete data (0.4355 becomes 0.4374); Lin’s CCC uses every case both raters scored; Gwet’s AC p-value is always two-sided; macro-averaged F1 counts never-matched classes (0.81 becomes 0.54). The multi-centre ICC is estimated within institution; its G-coefficient row and “Shrinkage estimates” are removed.
- Labels and verdicts. Robinson’s A and Kendall’s W are graded on chance-corrected values (46% of pairs of independent raters had read “Moderate agreement”); every table uses one name per coefficient; Light’s kappa loses a p-value that was not valid; Iota is titled “single variable”, since it never was the multivariate index the guide described (its value is unchanged).
- Rater clustering groups readers by agreement. The default distance is now Concordance (1 - Lin’s CCC); the old correlation distance ignores a constant offset, so a reader who read 25 points high was clustered with the accurate readers. Analyses that never set the metric will re-cluster.
Medical Decision (decision)
- Levels other than the two you select are excluded, never counted as negative. With 40 Benign (35 test-negative) and 10 Equivocal (all test-positive) cases, specificity read 70.0% instead of 87.5%. Excluded cases are counted in the notices and the copy-ready paragraph.
- Only the likelihood ratio with the zero cell is continuity-corrected: TP 20, FP 2, FN 0, TN 19 keeps its observed LR+ of 10.5 (“large”) instead of 8.59 (“moderate”).
- Youden’s index has an Agresti-Caffo interval (the old one covered about 99.7% under a 95% label). Likelihood ratios use one band in every panel, and “worse than chance” is an error only when the Youden interval lies wholly below 0.
Compare Medical Decision Tests (decisioncompare)
- Positive and negative levels must be chosen for the reference and every test. Any other observed level stops the analysis unless Exclude other observed levels is ticked, and a new table shows the levels used and excluded. Saved analyses that relied on an implicit negative level must select one to re-run.
- Pairwise p-values come from the McNemar mid-p test (Fagerland, Lydersen & Laake 2013) instead of the exact or continuity-corrected test. P-values are smaller: with 6 and 16 discordant pairs, 0.052 becomes 0.035. A new McNemar Test option offers the asymptotic test without correction.
- The text names the test each significant comparison favours. It used to name the test with the highest balanced accuracy beside “a significant difference” even when that test was significantly less accurate (95%/70% against 60%/97.5% at 20% prevalence). Sensitivity and specificity get their own p-values in the Differences table.
- A 0/0 likelihood ratio is blank, not invented: a test that never called positive got a finite LR+ labelled “Moderate” or “Strong”.
Combine Medical Decision Tests (decisioncombine)
- Each test alone is now a row and a candidate in the ranking. A strong test (0.90/0.90) with a weak one (0.60) used to give combinations tied at Youden’s J 0.50 while Test 1 alone, at 0.80, appeared nowhere. With three tests every pair also gets Parallel and Serial rows (11 rows become 20).
- “Positive Levels May Be Inverted” checks each test and fires only when its Youden’s J is below 0 beyond sampling variation. Heatmap colours for the ratios use a log scale centred on 1.
Decision Curve Analysis (decisioncurve)
- Curve points sit exactly on the thresholds the axis shows. Rounding had put the points labelled 6, 12, 15, 21 and 22% a hair above them, so patients at the threshold counted differently in plot and table. The clinical decision rule now enters both comparison tables, which changes the Holm-adjusted p-values when there are several models.
- Removed: the “Net benefit confidence intervals” checkbox, a duplicate of “Bootstrap confidence intervals” (a saved analysis that ticked only it reopens without ribbons), and the “Conclusion” column of the exploratory pairwise table. The Clinical Interpretation panel is opt-in.
LASSO Logistic Regression (lassologistic)
- The event level is never guessed. jamovi took the outcome’s first level and R the second, so the same data gave inverted odds ratios. A numeric 0/1 outcome still reads 1 as the event.
- Standardization is repeated inside every cross-validation fold and bootstrap fit, so lambda, coefficients and validation estimates can change.
- The default scoring method is Max-scaled. “Schneeweiss” was not Schneeweiss’s method and is renamed “Smallest-contribution units” (on the bundled example it gave a -23 to 120 point scale). Ordinal predictors get one indicator per level, ending rules such as “grade.L > 0”.
Clinical ROC Analysis (enhancedROC) and Advanced ROC Analysis (psychopdaROC)
- AUC-PR is the interpolated precision-recall area in both, not the trapezoid that overstated tied scores (a constant marker at 10% prevalence reported 0.55, not 0.10).
- Advanced ROC Analysis needs both a Positive Class and a Negative Class; saved analyses must select the Negative Class to re-run. Single-AUC intervals use logit DeLong by default, paired DeLong comparisons add Holm-adjusted p-values, and power tests each marker against AUC 0.5 however many markers are selected (a duplicate marker had raised one example from 12.1% to 99.98% power).
- Clinical ROC Analysis separates fixed score cutoffs, target prevalence and utility thresholds in its impact analysis, so re-enter the intended cutoffs in saved impact analyses. “Stratified bootstrap” now also governs the comparison test and confidence bands. Table notes, the Analysis summary and the Report sentences are opt-in.
Analysis Without Gold Standard (nogoldstandard)
- Penalized EM estimates can change. The fit now starts from several points and is refined numerically, instead of one EM pass clipped to 0.1%-99.9%.
- Unreliable latent estimates are blank, not shown with a caveat. A fit that does not converge, is not identified, or cannot tell its two classes apart (“Class orientation is ambiguous”) leaves prevalence, sensitivities and specificities blank with a strong warning. Bootstrap intervals with fewer than 20 usable resamples are blank and can otherwise be wider than before; a rare test’s interval can now reach 0%; with three tests the dependence check reads “Not assessable”.
Kappa sample-size analyses (kappaSizeCI, kappaSizePower, kappaSizeFixedN)
-
Confidence Interval Approach for the Number of Subjects Required (
kappaSizeCI): larger n for most designs with 3+ raters or 3-5 categories. At the published n, the interval the study should report covered kappa only 91-93% of the time. Rotondi & Donner’s own example now needs 190 subjects, not 171 (147, not 110, with 6 raters); a binary outcome with two raters is unchanged. The n is the exact minimum, shown with its arithmetic (3.84146 / 0.032634, rounded up, = 118). -
Power Approach for the Number of Subjects Required (
kappaSizePower): binary designs are checked by simulation. When the recommended test falls more than 0.03 short of the stated power, the result names both figures: 6 raters, a 5% finding, kappa0 0.40 against 0.60 and power 0.80 need 316 subjects by the published formula and about 572 by simulation (316 give about 0.59). The “0 subjects” result for 5-6 raters and a finding at or below 0.1% (181,129 needed) is fixed. -
Lowest Expected Kappa for a Fixed Sample Size (
kappaSizeFixedN, formerly “Lowest Expected Value for a fixed sample size”): the bound is solved exactly and rounded down (0.419 at the defaults, where kappaSize gives 0.418). With 3+ raters or categories it allows for re-estimating the proportions: 3 raters, kappa0 0.60, prevalence 0.30 and 60 subjects give 0.443 (was 0.445). With fewer than one expected subject per rater in the rarest category it no longer claims the study “can show that agreement exceeds chance”. - In all three, the “Sparse categories” check judges the cells the test actually uses (19 of 40 designs had been flagged spuriously in
kappaSizeCI).
Medical Decision Calculator (decisioncalculator)
- A test that was never positive no longer gets a likelihood ratio. TP 0, FP 0, FN 10, TN 100 showed LR+ 9.18 (a function of the group sizes alone) and a nomogram promising a rise from 9.1% to 47.9%; the cell is now blank with a note. “Post-test Reference-negative Probability”, which repeated the NPV, is replaced by the probability of disease after a negative test (1 - NPV).
- The illustrative cut-off comparison ranks on the Youden index intervals (a scenario with accuracy 0.595 could be called clearly better than one with 0.941), and its scenarios must share the same numbers of reference positives and negatives.
New features
-
Interrater Reliability (
agreement): Pathologist Grading Analysis. For grades, keeping cases some pathologists did not read: grade use per pathologist, agreement with colleagues at each grade boundary, each pathologist’s grading tendency with a forest plot, and model-based agreement (latent ICC and kappa, Nelson & Edwards 2015). Model-based numbers appear only in designs where simulation showed them accurate (for example from 50 cases and 3 pathologists); otherwise a note says why. -
agreement, also: multi-centre analysis of categorical ratings with a stratified kappa; a Case ID so the mixed-effects comparison compares each case with itself; a Kappa interval method option (Wald or Fisher z); an exploratory patient-level bootstrap for specific agreement. -
Medical Decision (
decision) and Compare Medical Decision Tests (decisioncompare): say what the reference is. Against an expert consensus or a comparative method the output reports positive and negative percent agreement, with comparator tables and plots, and leaves out disease-risk figures such as PPV. An optional Validation context table records the cohort, sampling design, threshold source, model version and unit of observation. -
Combine Medical Decision Tests (
decisioncombine): a “Combination Rules vs Each Test Alone” table; Youden’s J with an interval, optionally corrected for picking the best rule on the same data; a notice when results are missing because of another result (reflex testing, partial verification), which biases every complete-case row. Notices now appear at the top. -
Medical Decision Calculator (
decisioncalculator): observed and projected workload – flags per 1000 tested and tests per detected positive at target prevalences (default 0.01%, 0.1%, 1%). Prevalences below 0.1% are accepted here and in Compare Medical Decision Tests. -
Clinical ROC Analysis (
enhancedROC): frozen-model validation report with reference mapping, patient-level bootstrap intervals, paired score comparisons and workload at rare prevalences. -
Advanced ROC Analysis (
psychopdaROC): a Patient ID with mean or median summaries for repeated measurements. LASSO Logistic Regression (lassologistic): an optional patient ID. -
Analysis Without Gold Standard (
nogoldstandard): PPV and NPV have confidence intervals; a warning says when categories other than the positive level (equivocal too) count as negative. -
Decision Curve Analysis (
decisioncurve): warnings for values that run opposite to the positive level, fewer than 25 patients without the outcome, and thresholds with fewer than 10 positive classifications. - Kappa sample-size analyses: the Notes name the model and test each number rests on; new warnings flag a rare category, a small sample and (fixed sample) a design that cannot support a reliability claim. The step-by-step explanation and kappaSize’s own summary are optional in the CI and fixed-sample analyses.
Fixes
-
agreement: “Benjamini-Hochberg (FDR)” in the mixed-effects comparison stopped the analysis; notes and numbers no longer carry over into the next run; interval columns follow the confidence level you chose; “ER+” and “ER-” are no longer merged; long runs can be cancelled. -
decision: the nomogram’s negative-result line no longer vanishes at low prevalence; levels such as<LODdisplay.decisioncompare: paired differences no longer fail on a few hundred cases or more; copy-ready text prints “p < 0.001” and names the tests and intervals used. -
decisioncombine: no longer stops on large samples (about 93,000 cases) with a few missing values; a test column named like an internal working column is no longer silently replaced. -
decisioncurve: bootstrap intervals are about 17 times faster with identical results; undefined cells are blank with a footnote instead of “.” or “NaN”; impact bars are colour-blind safe. -
decisioncalculator: studies of more than about 430 cases (for example 180/20/30/270) no longer stop with an error; blank cells no longer show “NaN”. -
Co-Testing Analysis (
cotest): the numbers were right but several sentences were not. The copy-ready summary now leads with the co-testing rule’s own result, a positive on either test (42.5% at the defaults; only both-positive, 89.1%, was given). Magnitude words follow the likelihood ratio (a rise from 92% to 99.9% had read “minimal change”), and an impossible result combination is footnoted instead of showing “NaN”. -
Sequential Testing Analysis (
sequentialtests): the test-order warning is replaced by a note that appears only when swapping the tests would save second tests (order never changes the combined figures); each test’s PPV and NPV are labelled as values for that test alone; cells that round to 100.0% get a footnote with the real value; long test names wrap in the flow diagram. -
nogoldstandard: a test added without its positive level is named instead of silently dropped.psychopdaROC: numeric columns imported as Nominal keep their codes.enhancedROC: a minimum sensitivity or specificity no longer yields an infinite cutoff, and every resampled result shows its seed.
Packaging and translations
- Turkish: 6,253 of the module’s 6,304 texts are translated, up from 3,738 of 4,881 in 1.0.83.03.
- A message containing its own placeholder could freeze jamovi’s analysis engine, and a table emptied on a re-run could stop the analysis; neither can happen now.
- Syntax copied from the syntax pane quotes free-text fields safely, so it runs for names with quotes or spaces and cannot carry other code.
-
ordinalis added to the install list (for the Pathologist Grading Analysis);tibbleis dropped.
meddecide 1.0.83.03 (2026-09-23)
This release is the result of an external review of the module. Most of it is about results that did not survive leaving the screen.
- The meta-analysis forest plot in ROC analysis exported as a blank file, and came back empty when a saved
.omvwas reopened. It now draws from the data saved with the plot, so exporting and reopening give you the plot you saw. If the plot cannot be built, you now get an explanation instead of the plot silently disappearing. - ROC curve overlays (confidence bands, optimal-cutpoint markers, comparison curves) had the same problem and are fixed the same way.
- Tables and plots no longer redraw when you change an option that only affects appearance, so the results panel stops blinking as you adjust display settings.
- Analyses no longer leave a fixed random seed behind them. Previously, running one analysis that used resampling could make the next analysis in the same session draw from a fixed random stream.
- Interrater Reliability (
agreement) is back in the module after a period out for testing, with the five packages it needs restored to the install list. - Five packages that no analysis actually used were dropped, so installing the module pulls in less.
-
caretis now cited by the analyses whose sensitivity, specificity, PPV and NPV figures it computes, so you can attribute those numbers. - The Turkish and English translation catalogues were inherited from a much larger project and carried about 31,000 message IDs, most from analyses that are not in this module. They are now trimmed to this module’s own text (roughly 4,000), which makes the download substantially smaller. Translatable sentences that were previously split into fragments (
"ERROR: ",", correlation adjustment = %.2f") are now whole sentences, which a translator can actually work with. Some existing Turkish translations were attached to the old fragments and will show in English until they are re-translated. - The “Test method” control in Decision Curve Analysis offered a single choice; its label and description now say plainly that only the bootstrap is implemented.
meddecide 1.0.81 (2026-09-09)
These notes were written after the release, which was published without a description. 1.0.81 was the first release after 1.0.6, so it also delivered the 1.0.6.03 changes listed below.
New analysis
-
Decision Curve Analysis (
decisioncurve) joins the meddecide menu. It compares the net benefit of predicted risks, or of a yes/no clinical rule, with treating everyone and treating no one across a range of threshold probabilities. It reports a Range of Benefit (the thresholds where a model beats both default strategies) rather than an “optimal threshold”, which does not exist for net benefit. Model comparisons are bootstrapped with a fixed seed (default 42), so a re-run gives the same p-values, and are Holm-adjusted; intervals can be pointwise or a simultaneous band over the whole curve. The output states that net benefit is apparent, not optimism-corrected, when the risks were developed on the same data.
Behaviour changes – saved analyses may produce different results
-
Interrater Reliability (
agreement): Robinson’s A is now Robinson’s (1957) coefficient. The value shown before was Goodman-Kruskal’s gamma, an association measure: two raters one grade apart on every case scored 1.0. It now comes with a bootstrap interval. The table headed “Maxwell’s RE” did not compute Maxwell’s statistic; it is now the “Rater Variance Decomposition”, a two-way (case and rater) split of disagreement. The bootstrap ICC row follows the chosen ICC model, and paired-comparison intervals follow the confidence level. -
Compare Medical Decision Tests (
decisioncompare): tests with equivocal results on different cases stay aligned. Paired comparisons could match one patient’s result with another’s. Each test’s estimates now use its own determinate cases, and McNemar, Cochran’s Q and the paired differences use the cases determinate on both, with the denominators shown. The “noninferiority margin”, described as a regulatory convention, is now a descriptive “Minimum OPA criterion”. The same test chosen twice, or the reference reused as a test, is rejected. -
Lowest Expected Value for a fixed sample size (
kappaSizeFixedN): the anticipated kappa now defaults to 0.60 (the agreement you expect, as in the confidence-interval approach, not the 0.40 null of the power approach) and is labelled “Anticipated kappa (kappa0)”. A bound below the lowest agreement the model allows (for example -0.841 against a floor of -0.0004) is refused with an explanation instead of being printed. The description no longer states the bound backwards, and the estimand is named as the intraclass (Fleiss-type) kappa, not Cohen’s. -
Power Approach for the Number of Subjects Required (
kappaSizePower): the sparse-cell warning checks the agreement-pattern cells the method uses, not the category proportions: six raters at 5% prevalence gave N = 316 with no warning while three of seven cells had expected counts below 0.5. New warnings flag N below 10, power below 0.5, a kappa gap under 0.05 and N above 2000. -
LASSO Logistic Regression (
lassologistic): the “Sullivan/D’Agostino” scoring method is renamed “Max-scaled”; a saved analysis that used it must select it again. -
Clinical ROC Analysis (
enhancedROC): custom cutoffs under “lower is positive” follow pROC’s rule for ties; with Youden optimisation off the cutoff is the point closest to the top-left corner; internal validation labels its interval honestly; one-vs-rest curves use one direction per marker. -
Advanced ROC Analysis (
psychopdaROC): the power curve uses the table’s Hanley-McNeil power, and the decision-curve plot uses the table’s risk fit and prevalence; effect-size columns are labelled as differences.
New features
-
Interrater Reliability (
agreement): “Case agreement categorization column” now adds the per-case agreement categories to your data (the option existed, but no column was added). -
Co-Testing Analysis (
cotest): name fields for the two tests and a “Guidance and explanations” switch. Combine Medical Decision Tests (decisioncombine): an optional “About this analysis” panel; “Add Test Pattern to Data” is now a “Test pattern column” output, and the “Optimal Pattern Recommendation” is now a “Descriptive Candidate-Rule Ranking”. -
Analysis Without Gold Standard (
nogoldstandard): optional Plain-Language Summary and Method Guide; tables are retitled “Estimated Class or Rule Proportion” and “Estimated Test Metrics or Rule Agreement”.
Fixes
-
psychopdaROC: re-running after ROC smoothing, an unselected positive class and a missing subgroup value no longer crash the analysis.enhancedROC: plots redraw after a saved file is reopened.sequentialtests: the explanation and formula panels show their numbers instead of{placeholders}.agreement: tables no longer duplicate rows on re-run, and weighted kappa on unordered categories stops with a clear message.
meddecide 1.0.6.03 (2026-08-22)
This release removes clinical advice from analysis output and corrects statistical statements that were false as written. The module now describes what was measured and explains how to read it, rather than issuing verdicts on whether a test is fit for patient care. Two defects produced a wrong or misleading number on screen and are listed first.
Behaviour changes – saved analyses may produce different results
decision: the positive likelihood ratio no longer prints as 0.00 when specificity is 100%. When a test produced no false positives, LR+ is infinite – the strongest possible rule-in evidence. The Clinical Summary collapsed every non-finite value to zero and printed “Positive LR: 0.00”, which reads as the exact opposite: a positive result making disease infinitely less likely. Infinite and undefined values are now distinguished and labelled (“not estimable (specificity is 100% in this sample)”). The same collapse applied to LR-, where the non-finite case arises when specificity is 0%, and the message now names that condition correctly.decisioncompare: predictive values are attributed to the prevalence actually used. When the prior-probability option is on, PPV and NPV are computed by Bayes at the prevalence you supply, while overall accuracy stays at the prevalence of the sample. The copy-ready summary sentence asserted a single sample prevalence for all three, so a manuscript-ready sentence could carry the wrong prevalence for the predictive values. The sentence now branches on the option and states which prevalence each figure rests on.decisioncompare: the best-performing test is described by the criterion actually used. The summary said the winning test “had the highest observed accuracy”, but the selection maximises the Youden index plus accuracy, so the winner need not be the most accurate test in the table.
Changed – results no longer give clinical advice
Recommendation columns now report performance instead of issuing verdicts. In
decisioncalculator,psychopdarocandenhancedroc, columns titled Clinical Recommendation returned strings such as “Excellent performance - Recommended” and “Suitable for clinical use”, keyed on point estimates alone. They are retitled Performance Summary and now report the measured quantities, with the optimism made explicit where a cut-point was chosen on the same data that measured it.sequentialtests’ Clinical Guidance panel is retitled Testing Strategy Notes.enhancedroc: the AUC context column is populated for every setting. The column was blank on every row under the defaultgeneralcontext. It now explains the AUC in plain language for all contexts – for example “Ranks a random case above a random non-case 81% of the time” – and says where a conventional reference level exists and where none does.
Fixed – statements that were not true
agreement: non-significant tests no longer establish agreement. The Stuart-Maxwell and Bhapkar tables reported “marginal homogeneity holds” and “no systematic rater bias” whenever p was at or above 0.05, stating the null hypothesis as established fact. Worked examples went further, concluding from p = 0.68 that a digital pathology platform yields “equivalent” distributions and is “suitable for clinical use”, and from p = 0.12 that two PD-L1 clones (22C3 and SP263) are interchangeable. These now report that no difference was detected and state plainly that this does not establish equivalence.agreement: quantitative claims the module never computed have been removed. The output asserted that confidence intervals overlap in a table that has no confidence-interval columns, and that a test had limited power where no power calculation exists. Kendall’s W and Robinson’s A examples claimed to validate treatment timing, credentialing and regulatory submissions from statistics that contain no outcome data at all.decisioncompare,cotest,nogoldstandard: non-significant comparisons no longer describe the tests as performing equally, and the McNemar note now explains that the test uses only the discordant pairs, so few discordant pairs mean little power.psychopdaroc: AUC is described correctly. An AUC of 0.90 was said to mean the test “can reliably distinguish between diseased and healthy patients”. AUC is the probability that a randomly chosen case outranks a randomly chosen non-case – a property of pairs, not a classification accuracy for any individual patient.
Fixed – output hygiene
R package chatter no longer appears in Analysis Notes. jamovi surfaces
message()andwarning()conditions to the user, so third-party notices leaked into results – a single Lasso-Cox run displayed the glmnet tie-handling notice twelve times. Third-party calls are now wrapped so package chatter and deprecation notices are suppressed while substantive warnings, such as non-convergence, still reach you.A malformed reference no longer prevents results from rendering. A citation with an empty publication year caused a serialization failure that produced no output at all, with an error mentioning
serializerather than anything about the analysis. All references now carry a year.
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,test2Negativeandtest3Negative. 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 aLeveloption, so all four compile to bare parameters of the R wrapper: existing scripts that calldecisioncompare()will fail withargument "goldNegative" is missing, with no defaultuntil they are updated, passingNULLfor any test that is not in use. Nothing changes for users driving the analysis from the jamovi GUI. The reason for the change is theexcludeIndeterminatedefect described below – without a declared negative level the analysis has no way to know which rows are equivocal. -
decisioncombine()’sfilterPatternlost the levelsserial,parallelandmajority. 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 areall,allPositive,allNegativeandmixed. A script passing one of the removed values now fails withArgument 'filterPattern' must be one of ...; useallPositivefor the oldserialandallfor the oldparallel. The jamovi GUI shows only the current levels. -
nogoldstandard()’s defaultmethodchanged fromall_positivetolatent_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_classis 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 becauselatent_classneeds three or more tests, a two-test analysis that previously returned (meaningless) results now refuses and explains why. Saved.omvfiles keep the method they were saved with; passmethod = "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 ajmvcoreTable method – onlydeleteRows()exists – and each call was wrapped intry(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. -
excludeIndeterminatewas a complete no-op. It filtered onc(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 != "")), andjmvcoreroutes 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.Rnever 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 againstDescTools::CochranQTest(six decimal places); McNemar with Holm correction againststats::mcnemar.test+p.adjust(1e-10); Wilson intervals againstbinom::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 2x2 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
epiRinterval 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 theepiRstatistic name carried alongside the data. -
The Fagan nomogram silently failed to render whenever a 2x2 cell was zero – precisely the sparse tables that most need it. A zero cell makes sensitivity or specificity exactly 1, which
nomogrammerrejects. 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/Nlrare no longer passed:nomogrammerignores 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 returnedInfwith aNaNlower bound – two numbers for one statistic on one screen. Those three rows now come from a secondepi.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 underdecisioncompare. 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 toNA. 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 > 1band 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.testsexactly (sensitivity 0.8, specificity 0.9, PPV 0.888889, NPV 0.818182, LR+ 8, LR- 0.222222), with the 2x2 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()usedsort(unique(c(a, b))),irr::kappa2was handed a data frame (whichirrcoerces withas.matrix()to character and then re-derives categories alphabetically) and is now fed integer factor codes, andirrCAC::gwet.ac1.rawnow receives an explicitcateg.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 requiresis.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 matchespsych::cohen.kappaexactly, 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, becauseirrsupplies 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 theis.na()guard the Landis & Koch chain above it had, soif (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.fleissreturns-Infwhen 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 derivedse <- abs(kappa / z)fromirr’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 isqnorm()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 withas.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 whatevericcTypethe 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 –
proportionalBiaswas missing fromclearWithalthough the plot adds ageom_smoothline under it. Separately, the pairwise-agreement figure silently truncated atmax_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
glmnetis then called withstandardize = 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 whatglmnet(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
pROCreturns-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
explanatoryis 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 asNAand dropped byna.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.
asDFis an R6 active binding onjmvcore::Table, sotbl$asDFalready returns the data frame; the code wrotetbl$asDF(), invoking that data frame as a function – “attempt to apply non-function”. The recommendation and all four plots were affected. -
The tables grew on every re-run and then broke. Nothing called
deleteRows(), and jamovi re-runs.run()on the same object whenever an option changes, so rows went 5 -> 10 -> 15; the duplicated row keys then made$asDFfail withduplicate '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 2x2 under two labels; counting both manufactured a tie and inflated that count.
-
Proportions and odds ratios shared one
estimatecolumn – 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 yieldsNAwith a notice instead of a silent number. - Pattern and strategy statistics were confirmed against
epiR::epi.tests, the Wilson intervals againstbinom::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 existingrowKeyduplicates rather than replaces (rowCount 2 -> 4 -> 6 over three passes, after which$asDFfails). Fixed on the three tables that populate viaaddRow. -
The Fagan nomogram silently failed on any table with a zero cell – precisely the sparse tables that most need one – because
nomogrammerrejects 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.
nomogrammerprints its reading – prevalence, likelihood ratios, post-test probabilities – to stdout underVerbose = TRUE, and jamovi never shows stdout. It is now rendered beside the figure, at the tables’ precision rather thannomogrammer’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 2x2 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 siblingdecisionis 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 asprobs[[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 againstpoLCAon 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_positivethe 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_positiveis the mirror image (FP == 0, specificity and PPV fixed at 1), andcompositewith two tests isany_positive, because a 1-of-2 tie passes arowMeans >= 0.5majority. Those quantities are now blanked with an explanation. See Breaking changes for the change of default. - The confidence intervals were too narrow – by about 1.8x for sensitivity at 30% prevalence. The standard error used the total n for both metrics; the denominators are n x prevalence for sensitivity and n x (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 anNAsensitivity threw “missing value where TRUE/FALSE needed”, ending the run instead of blanking a number. An invalid positive level producedLevel 'test1_result' not found in variable 'negative, positive'. Available levels: {}–jmvcore::reject(formats, code = NULL, ...)takescodeas 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 Frechet bound. Beyond it the requested dependence is unattainable and
cotesttruncated 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 of0.000000– 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 same0.000000appeared 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-precision0.999991, 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 underdecisioncompare. 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 Frechet 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%.presetwas 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
clearWithit 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 onedeleteRows(). 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 withduplicate '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 byaddRow; the two driven bysetRowcannot 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
?enhancedROCdescribed 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, includingsplineKnotswhich configures one of them, are now prefixed “NOT YET IMPLEMENTED - selecting this produces no output”. -
test-enhancedroc-pr-clinical.Rnever exercised the analysis. It built a hand-rolledjmvcore(MockTable,MockHtml,MockImage), assigned a fake base class into the global environment andsource()d the backend; the mock implemented onlyaddRow,setRowandasDF, 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 againstpROCto 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 twodeleteRows(). 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, matchingenhancedROC: 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
cutpointrpackage 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 matchescutpointr, 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 topROCand 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 Rwarning(), 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.Rwas 29 failures and 6 errors, andsource()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"", includingdependentVarsandclassVar, so a test that omitted a required variable was silently handed an empty string. All nine documented examples were uncopyable because none passedrefVar, atype: Leveloption 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 againstpROCandcutpointr.
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
enhancedROCand 0.1001 frompsychopdaROC, with nothing on screen saying why. The cause is a default, not a calculation:enhancedROCworks the direction out from the data, whilepsychopdaROCassumes 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. - the same column of numbers came back as AUC 0.8999 from
enhancedROChonours a confidence-level option;psychopdaROChas none and always reports 95 percent, so at 99 percent the two give [0.8456, 0.9542] and [0.8585, 0.9412].enhancedROCrefuses any dataset below n = 20 and returns nothing;psychopdaROCcomputes and agrees withpROC(n = 16: 0.6406 [0.3236, 0.9576]).Their optimal cutpoints differ in the last decimal by construction:
pROCplaces the threshold at the MIDPOINT between two adjacent observed values (54.1979) whilecutpointrreturns an observed value (54.2904). Sensitivity and specificity are identical either way.Given a positive class that does not exist,
enhancedROCreturns empty tables with an explanatory panel whilepsychopdaROCraises 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 directPowerBinary(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. -
kappa0was documented backwards in two of the three analyses.kappaSize’s own documentation definesPowerBinary’skappa0as “the null hypothesis for the kappa hypothesis test” andCIBinary’s andFixedNBinary’s as “the preliminary value of kappa” - two different quantities that happen to share a name.kappaSizePowerdescribed its null as the “Expected value of kappa”, andkappaSizeCIdescribed 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, andkappaSizeCI’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 0.70was accepted bykappaSizeFixedNand rejected bykappaSizePower. A binary outcome may be entered as one prevalence or as two proportions in all three;kappaSizeFixedNalone demanded two, althoughkappaSize::FixedNBinaryaccepts 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
kappaSizeobject, 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
kappaSizeacross 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::CIBinaryfinds its answer by brute force -n <- 10; while (...) n <- n + 1in interpreted R, with no cap - and the required n grows as roughly one over the square of the distance fromkappa0to 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
kappaSizesizes on whichever limit lies nearerkappa0: withkappaL = 0.55the answer is 1,625 for everykappaUfrom 0.65 to 0.99. In one-sided modekappaUis 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 lettert- it matched a literal backslash and the letter t, but neither an actual tab nor a space. So0.2, 0.3; 0.5was accepted and0.2, 0.3 0.5was 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
kappaSizeacross 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.
kappaSizesearches by decrementing rho fromkappa0in 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), andInfpasses all three clauses -is.na(Inf)is FALSE,Inf < 2is FALSE, andInf != round(Inf)is FALSE. Inside the engine the test statistic becomes NaN and the search never terminates.nis 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
kappaSizeFixedN engine containsif (n <= 10) stop("Sorry, your study should enroll at least 10 subjects."), while the option permittedndown 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 = 0.5reached the engine and got “missing value where TRUE/FALSE needed”. The two now agree. The proportions-sum tolerance likewise usedall.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:kappa0here 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
kappa0here is the anticipated agreement and not a null hypothesis value as it is inkappaSizePower- 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
kappaSizeacross 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 belowkappa0in all 2,000.
Changed
-
lassologisticscore 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
-
lassologisticgained a configurable cut point for continuous predictors in the scoring system.scoreCutMethodoffers median, mean, upper tertile, upper quartile and manual (default median, i.e. the previous behaviour), andscoreCutPointsis a free-text field takingvariable=valuepairs on the original measurement scale (for exampleki67=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,decision2andppvare 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.Rmddescribedkappa0as a null hypothesis value. ForkappaSizeCIit is the anticipated kappa; onlykappaSizePowertreats 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 nearerkappa0, so withkappaL = 0.55the answer is 1,625 subjects for everykappaUfrom 0.65 to 0.99, and in one-sided modekappaUis ignored entirely. It also listed the rater options as 2 to 5 where the analysis accepts 2 to 6. -
nogoldstandard-documentation.mdstill gave the old default. It listedmethodas defaulting toall_positivein two places; the 1.0.4 default islatent_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 asrater1_var,agreement_type,pathologyContextanddiagnosticStyleAnalysis, none of which have existed for some time - copying them yieldsunused argument. Every chunk in these articles iseval = 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 =)ispsychopdaROC(dependentVars =, classVar =, positiveClass =), andkappaSizeCI/kappaSizeFixedNtakealpharather than aconf_level, withkappaL/kappaUin place of a singlewidth. - 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
importFromdeclarations. Seven exports were removed (clinicalscore,decisioncurve,latentbiomarker,leaveonecenterout,mageeequation,misclassificationbias,timedependentdca) along with about 25 unused imports (dplyr::arrange/filter/summarise, severalggplot2andgraphicsfunctions,stats::complete.cases, and others). No shipped analysis was affected..Rbuildignoregainedtempandbackupspatterns. -
bootandsurvivalwere dropped fromImports. Neither is used anywhere in the package: no::call, noNAMESPACEimport, and no bare call to any function they export. jamovi installs every package inImportsthe 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 indecision,decisioncalculatoranddecisioncompare– defaults tomethod = "exact", which is the Clopper-Pearson interval, not Wilson. On a 2x2 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 indecisionanddecisioncomparesaid “Wilson score method” and now name Clopper-Pearson. No interval changed – only the label was wrong.decisioncombinewas checked and left alone: its own.calcWilsonCI()genuinely is Wilson and genuinely populates the CI table it describes, so its labelling was already correct.decisioncompareadvertised likelihood-ratio confidence intervals it never computes. The per-test table is filtered to a row set that excludeslr.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.decisioncalculatoroverstated 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.decisioncrashed whenever “Disease Absent Level” or “Test Negative Level” was left unset.dplyr::case_when()evaluates every branch regardless of which one matches, so the unreachabletest == <negative level>comparison still ran with the levelNULL, produced a zero-length condition, and aborted the analysis withCan't recycle ... size 0. Both levels now useNA_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 defaultbefore the analysis’s own validation could produce a usable message. Now defaulting toNULL:agreement(vars),decision(gold,newtest),decisioncombine(gold,test1),decisioncompare(gold,test1,test2) andnogoldstandard(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 touchingDESCRIPTIONorjamovi/0000.yamlcross-checks the two version strings, refuses to proceed if they disagree, and – if the tag does not already exist – tagsv<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.
vcdandlme4are used byagreement()but were missing from the packageImports. Because jamovi installs only a package’sImports, on a clean install they were unavailable: pairwise kappa confidence intervals silently fell back to the narrowerirr::kappa2null-SE method, and the entire ICC / Lin’s CCC / continuous-agreement suite (which relies onlme4) was non-functional. Both are now declared, so agreement CIs use the intendedvcd::Kappaasymptotic-SE method and the continuous-agreement measures work. - Clarified the all-pairs kappa fallback note so it distinguishes a missing
vcdpackage (with install guidance) from a genuinely near-degenerate table. - Declared
DescToolsandlmerTest, 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, withgwetWeights: 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)
- Gwet’s AC1/AC2 (
-
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
- ICC with six models ICC(1,1)-ICC(3,k) (
-
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
- Kendall’s W coefficient of concordance (
-
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
- Rater Bias test (
-
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
- Pairwise Kappa against a reference rater with performance ranking (
-
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
- Confusion matrix table with row/column normalization (
-
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
- Agreement heatmap / confusion-matrix plot (
-
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
- Paired agreement comparison between two rater conditions with bootstrap (
-
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
- Consensus rating variable with majority/supermajority/unanimous rules and tie handling (
-
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 forconfLevel, Bland-Altman confidence level, cluster counts, and subgroup minimums)
- Configurable confidence level for CIs (
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;customCutoffsnow defaults to empty - Added
clearWithto 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); addedclearWithto results
- Bootstrap/threshold count options changed from Number to Integer (
-
kappasizeci(): One-sided confidence intervals- New
citypeoption (two-sided vs one-sided lower-bound-only), wired tokappaSize::CIBinary/CI3Cats/CI4Cats/CI5Cats, with a UI ComboBox that disables the upper-limit input in one-sided mode - New plain-language
text_summaryoutput (CI type, lower limit, precision width)
- New
kappasizefixedn()/kappasizepower(): New plain-languagetext_summaryoutput panelnogoldstandard(): Newnoticespanel (“Important Information”) with plain-text notice rendering that resets on each rundecisioncalculator(): 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/psychback the new chance-corrected, clustering, and ICC agreement measures;ggraph/igraphback the reimplementeddecisiongraphtree visualization;htmltools/knitr/tibble/tools/grDevices/graphicssupport 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) andR/error_handling.R(clinical error-handling framework:clinicopath_init,clinicopath_error_handler) - Added
.escapeVariableNamesand refactored/hardened existingR/utils.Rhelpers (%notin%/%!in%rewritten as explicit functions,load_required_packagedefault flipped toinstall_if_missing = FALSE,print.sensSpecTablegiven an S3-compliant(x, ...)signature) -
R/decisiongraph_utils.R: decision-tree visualization reimplemented as a realigraph/ggraphdendrogram 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 (decisiongraphis 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.Rfor 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(), andsequentialtests()labels and descriptions (fordecisioncalculator()this covers the label/emoji changes only; the CI methodology change is documented under Enhanced Existing Functions above).decisioncombine()also gainedallowNone: trueon its Test 3 positive-level option and had its exampledontrunflag flipped totrue - 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)to1:(length(breaks) - 1) - This fix ensures proper risk categorization in NRI calculations
- Affects categorical NRI computations in ROC and psychoPDA analyses
- Corrected the labels vector calculation from
-
agreement(): Fixed stability issues and hanging during initial run- Refactored
agreement.b.Rto ensure responsiveness - Maintained support for numeric variables in agreement analysis
- Refactored
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()anddecisioncompare()instead - These new functions provide more focused and comprehensive analysis
- Users should use
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