fix: Unloading the package preserves unrelated hooks and tolerates dictionary entries that have already been removed.
fix: cmprsk.auc, cmprsk.brier, and cmprsk.ibs now resolve cause weights before scoring and skip zero-weight causes. Event-frequency weights warn for absent causes and return NA if no events are observed.
feat: cmprsk.auc, cmprsk.brier, and cmprsk.ibs now identify the affected causes when warning about non-finite scores.
feat: cmprsk.auc, cmprsk.brier, and cmprsk.ibs can now warn when scoring requires constant CIF extrapolation after a cause’s final prediction anchor.
feat: cmprsk.auc now explicitly rejects scoring times beyond the maximum test-set follow-up, following the behavior of the Brier score.
feat: cmprsk.ibs now defaults to 100 equally spaced integration points from 0 to the 90th percentile of test-set follow-up. New parameters ntimes, t_max, and t_quantile control the automatic grid size and horizon, using test-set outcomes only. Horizons exceeding the maximum test-set follow-up are reduced with a warning before constructing exactly ntimes points.
feat: CIF sum and extrapolation diagnostics are now disabled by default. Enable them package-wide with options(mlr3cmprsk.warn_cif_sum = TRUE, mlr3cmprsk.warn_cif_extrapolation = TRUE); warnings and errors affecting score validity remain enabled.
refine doc: better learner and measure examples
mlr3cmprsk 0.0.6
Added MeasureCompRisksIntegratedBrierScore (msr("cmprsk.ibs")) via RiskRegression. This is now the main measure for prediction error in the package.
Added config files for code formatting and for AI agents.
Raised the minimum required versions of mlr3 to 1.8.0 and mlr3misc to 0.23.0. (#22)
assert_cif_list() is now exported for validating lists of cumulative incidence matrices.
fix: cmprsk.fg parameter cengroup allows now for vector input.
fix: TaskCompRisks now checks that the event column has 1,2,…,K ordered integer-like numbers for the causes.
feat: more tests, refactoring, better assertions and expectations, better documentation across the package.
fix: TaskCompRisks$filter() now errors when row filtering leaves fewer than two competing events and gives a warning if less causes than before remain.
fix: check_prediction_data() was called (indirectly) twice after a learner finished its .predict() call, which was not required.
mlr3cmprsk 0.0.5
Refactored MeasureCompRisksAUC, added cause_weights parameter, renamed time_horizon to time
Added parameter tests for AJ and FG learners
Added more tests => >80% code coverage
Added MeasureCompRisksBrierScore, i.e. msr("cmprsk.brier") for fixed time point prediction error evaluation
Use survdistr from CRAN
mlr3cmprsk 0.0.4
Added LearnerCompRisksFineGray via cmprsk R package
Refactor merge_cifs() to align_cifs()
Added test coverage report + badge
Added example in README
mlr3cmprsk 0.0.3
Added missings, importance, and selected_features properties to cmprsk.aalen.
Added support for native_model() in cmprsk.aalen
CIF predictions for cmprsk.aalen are now at unique event times (rest can be constantly interpolated).
Minimum required mlr3 version is now 1.4.0.
Added a noise feature to autotest tasks for improved testing.
Improved documentation.
mlr3cmprsk 0.0.2
Use new interp_cif() function to interpolate CIF curves from survdistr@0.0.2
mlr3cmprsk 0.0.1
Initial transfer and update of code from mlr3proba.
