NEWS

multipleOutcomes 0.17.1

netbenefit_() now treats NA in any endpoint column as "tied at that level, fall through" rather than erroring deep in the U-statistic loop, and warns once per fit listing the affected columns.

netbenefit_() adapter for the hierarchical net-benefit (win-difference) statistic, with endpoints built by nb_tte(), nb_continuous(), or nb_binary() and priority encoded by list order.

New test-inplace-formula.R verifies that in-place formula transformations (I(z > 4), log(x), scale(x), factor(x), Surv(t / c, 1 - cnsr), multi-variable Boolean expressions, and LHS arithmetic) produce numerically identical PATED output to the pre-computed-column equivalents, across coxph_() primaries and glm_() prognostic specs.

multipleOutcomes() and the legacy *MO() wrappers (coxphMO, glmMO, geeMO, logrankMO, kmMO, quantileMO) are removed. The package's main entry point is now jointCovariance(), and each component model is specified through a constructor: glm_(), coxph_(), logrank_(), gee_(), mmrm_(), km_(), or quantile_(). pated() accepts the same spec constructors via ....
Datasets supplied to jointCovariance() / pated() must now contain a column pid carrying subject identifiers. Records with the same pid across different data frames refer to the same subject.

mmrm_() adapter for mixed models with repeated measures.
km_() adapter for Kaplan-Meier survival probabilities (bootstrap-only, since the empirical S(t) has no closed-form score).
quantile_() adapter for between-arm quantile differences (bootstrap-only).
gee_() adapter exposing a GEE fit's per-cluster score and Hessian so it slots into the joint asymptotic covariance machinery alongside other engines.
conf_type = "arcsin" is fully supported in km_() (previously silently fell back to logit transformation).

data_index defaults to 1 in every spec constructor, and a single data frame is auto-wrapped, so jointCovariance(spec1(...), spec2(...), data = my_df) is the shortest valid call.
Spec constructors validate data_index (must be a positive integer scalar); jointCovariance() adds an upfront bounds check naming the offending spec when data_index exceeds the number of supplied data frames.
pated() emits a warning when the residual variance goes negative — typically a sign that a prognostic variable is collinear with the primary outcome or with another prognostic.

KMAdapter$fit_model() no longer strips the names off self$estimate, so id_map entries for KM models now carry the time_(strata)_(time) labels that pated()'s arm lookup relies on.
parseTreatmentVariableFromCall() now walks the formula AST instead of regex-parsing the deparsed call, so formulas with nested parentheses (e.g., y ~ arm + us(visit | pid) for mmrm_) parse correctly.
fitKMCurve() uses survival::summary(..., extend = TRUE) consistently, preventing NAs from leaking into the bootstrap covariance matrix when a resample's stratum has no observations past a requested time.
pated() no longer relies on is.null(family), which silently bound stats::family (a function) after the family argument was removed.

New testthat suite (52 tests / 245 expectations) covering every engine through both jointCovariance() and pated(). Tests are split into a fast tier (~4 s) and a Monte Carlo tier (~25 s, opt-in via MULTIPLEOUTCOMES_RUN_MC=1) that validates empirical vs. theoretical covariance for single-engine, cross-engine, partial-overlap, and kitchen-sink configurations.
Regression fixture for the README example pinned to 1e-10 (file inst/testdata/readme_pated_reference.rds).
GitHub Actions workflows for code coverage (codecov) and pkgdown.
README example, vignette, and help-page examples migrated to the new spec-constructor API.

Initial support for Kaplan-Meier in the new spec-constructor interface via km_().
pated() extended to handle KM time-stratum parameter vectors, including transformed-S(t) point estimates, pointwise confidence intervals, and KM-vs-PATED curve comparison plots.