mlumr

• Feature: The baseline hazard is estimated per study by default (aux_by = ".study"), each stratum with its own knots over its own observed support, as multinma::nma() does. aux_by = "none" shares one baseline across studies.
• Feature: New priors prior_aux, prior_aux2 (second generalized-gamma auxiliary) and prior_smooth, with default_prior_aux() and default_prior_smooth().
• Feature: predict() on a survival fit returns "survival", "hazard", "cumhaz", "rmst", "median" and "loghr" curves. times picks evaluation times (snapped to the fitted grid, with a message when they move), pred_times sets the grid, "median" carries p_not_reached, and RMST rows carry their horizon. conditional_effects() and conditional_predict() give covariate-conditional versions.
• Feature: marginal_effects() reports survival effects on the natural scale: the hazard ratio (HR) for proportional-hazards fits, the time ratio (TR) for shared-shape AFT fits, the exponentiated contrast (EXP_DELTA_ETA) otherwise, plus the RMST difference (RMSTD) and ratio (RMSTR). Marginal hazard ratios carry an at_time column because they are time-varying and non-collapsible; predict(type = "loghr") gives the whole curve.
• Improvement: An effect a fit cannot supply is an error naming the alternative, not a substitution: effect = "tr" on a proportional-hazards fit, effect = "hr" on an AFT fit, and either on a relaxed or study-specific-shape AFT fit.
• Feature: Survival benchmarks. naive() returns the unadjusted Cox log hazard ratio and refuses a partial likelihood with no maximum or no convergence. stc() returns the RMST difference by parametric G-computation with flexsurv, with a bootstrap interval (n_boot, default 200) and an rmst_horizon argument.
• Feature: survival_unit on calculate_loo(), calculate_waic() and compare_models() chooses what one held-out unit is for the reconstructed comparator: "observation" (default, optimistic), "arm" or "aggregate".
• Improvement: Coefficient rows in summary() are labeled by covariate name (beta[age], beta_index[age], beta_comparator[age]); the stored variable strings are unchanged.
• Fix: Interval- and left-censored likelihoods under delayed entry are evaluated in whichever of the difference, increment or quadrature forms keeps its digits, and the exponential, Weibull, Gompertz and log-logistic use closed forms. See ?mlumr.
• Fix: mlumr() refuses an M-spline basis column that no subject was at risk under, judged over the merged [entry, exit] intervals of each study, and refuses a shared baseline (aux_by = "none") whose studies never overlap on a spline column. See ?mlumr.
• Improvement: predict(type = "rmst") and predict(type = "median") warn when the prediction grid is too coarse to trust and point at a finer n_rmst_grid or pred_times.
• Improvement: ?set_agd_surv states that reconstruction uncertainty is not propagated and which population the covariate moments must describe under delayed entry.

Transport to any target population

• Feature: A newdata argument on marginal_effects() and predict() standardizes effects and absolute predictions to an arbitrary target population by Bayesian g-computation, for every family. Standardizing to the index covariates reproduces population = "index" exactly. Survival RMST effects and absolute predictions transport; the marginal hazard ratio is reported for a target too, with its at_time.
• Improvement: marginal_effects() on a relaxed fit queried for the index population prints a one-line note with the marginal posterior variance change of beta_comparator per covariate, descriptive only. Silence it with options(mlumr.quiet_relaxed_index = TRUE).

Plotting

• Feature: plot() methods for marginal_effects() (forest), predict() (curves with credible bands, or point-intervals for scalar types) and conditional_effects() (effects by profile). Each forest draws the null its measure implies, uses a log axis for ratio measures, and reads its interval level from the result.
• Feature: geom_km() overlays the observed Kaplan-Meier curves on a model survival plot, colored by treatment, honoring delayed entry, one panel per population. It draws right-censored data only.
• Feature: plot_prior_posterior() overlays each named parameter's posterior on the prior the fit recorded for it.
• Feature: mlumr_forest() draws a forest from a plain data frame, for mixed comparisons such as naive(), stc() and both ML-UMR models on one axis; a single very wide interval is clipped with an arrow instead of squeezing the rest.
• Improvement: marginal_effects(), predict() and conditional_effects() return lightweight data.frame subclasses so the methods dispatch; data frame behavior is unchanged.

Identification of the relaxed comparator coefficients

• Feature: check_identification() reports, before fitting, whether the aggregate subgroup rows can identify beta_comparator in model = "relaxed": the row count against K + 1, and the balance, effective dimension and spread of the centered subgroup-mean matrix, with a descriptive-only label for nonlinear links.
• Feature: prior_beta_comparator argument to mlumr() sets a separate prior on beta_comparator, including a different family. Defaults to prior_beta; ignored for model = "spfa"; reported by prior_summary() and swept by prior_sensitivity().
• Fix: The weak-identifiability warning no longer counts a duplicated set_agd() row as evidence; it triggers on the rank of the mean design under the identity link and on the number of distinct integration grids otherwise.

Data preparation and integration

• Fix: distr() honors arguments passed by position. distr(qnorm, 10, 2) previously integrated a standard normal without any message. Positional and abbreviated arguments are now matched against the quantile function's formals at construction, and an argument matching nothing is an error.
• Feature: Moment-parameterized qgamma()/pgamma()/dgamma() and qlogitnorm()/plogitnorm()/dlogitnorm() accept mean and sd, so distr(qgamma, mean = age_mean, sd = age_sd) works as printed in a baseline table. Without those arguments they forward to stats. See ?GammaDist and ?logitNormal.
• Fix: set_agd() no longer rejects a valid binary covariate SD. The bound was the population SD sqrt(p (1 - p)); it is now the finite-sample maximum with a rounding allowance. See ?set_agd.
• Improvement: set_agd() requires outcome_n when normal aggregate data have more than one row, and the multi-row comparator estimand for the normal family is the sample-size-weighted mean of the strata rather than an inverse-variance average. naive() and stc() use the same weights.
• Improvement: set_ipd() warns about a rank-deficient or nearly collinear covariate design and names the near-redundant covariates. Autoscaling of a covariate with no usable empirical scale falls back to the unscaled prior with a warning.
• Fix: add_integration() rejects cor_adjust = "pearson" with non-Gaussian margins, warns about a nonbinary discrete margin (count or ordered category) that the copula correction does not cover, and warns when a supplied distr() grossly contradicts the declared set_agd() moments.
• Improvement: check_integration() compares realized and target correlations on the same method (Pearson or Spearman), reports fidelity of each grid margin to the declared mean and SD, lists which correlation pairs it could and could not measure (correlation_pairs, with "partial" and "review" verdicts), withholds the comparison under cor_adjust = "none", and reports a comparison it could not make as "unavailable" instead of "close".

Priors and sensitivity

• Breaking: for family = "normal" the default priors are on the outcome's scale. Under the identity link the intercepts and coefficients are in the outcome's units, and the residual SD is under either link, so the fixed normal(0, 10) intercept and half-normal(0, 2.5) residual-SD defaults were informative for an outcome such as a 0 to 100 score or a blood pressure in mmHg. They pulled each intercept toward zero, the comparator's (resting on one aggregate mean) the most, which moved the mean difference, and they understated the residual SD, which narrowed the intervals. With sd(y) the IPD outcome SD, the defaults are now normal(0, 10 * sd(y)) for the intercepts and normal(0, 2.5 * sd(y)) for the coefficients (identity link), and a half-normal with scale 2.5 * sd(y) for the residual SD (either link); autoscale = TRUE multiplies a coefficient's scale by sd(y) as well as dividing it by sd(x) under the identity link. Priors written out in full are used as given. prior_summary() prints the scales the model used, and prior_sensitivity() sweeps a default or autoscaled prior_beta in the same outcome-SD units. Refit normal-family analyses that relied on the defaults or on autoscale.
• Fix: prior_normal(autoscale = TRUE) rescales the prior location as well as the scale, so a nonzero prior mean is on the covariate's own scale.
• Improvement: prior_sensitivity() varies the prior and nothing else: every model-defining setting is replayed from the fit and ... may not override it. Relaxed fits sweep the comparator prior alongside the index one (prior_beta_comparator_scales, reported in scale_comparator), survival rows carry effect and at_time, and a refit that cannot replay a sampler argument the original fit used (such as thin or init) warns and names it.
• Breaking: prior_sensitivity() quantile columns are named q2.5, q50, q97.5, matching marginal_effects(), rather than q2 and q98. Duplicate probs are refused.
• Improvement: prior_sensitivity() no longer claims a scale sweep proves the inference is data-driven; its interpretation text points at check_identification().
• Fix: prior_summary() names constrained priors correctly: a truncated normal or t with a nonzero location is not a half-normal or half-t, and an exponential is not truncated at all.

Benchmarks: naive() and stc()

• Breaking: stc() no longer reports an index-population contrast. That estimate assumed a constant link-scale effect, which population adjustment exists to avoid. stc() returns the comparator-population estimand only; use mlumr() when the index population is the target.
• Breaking: A normal-family stc() under a log link reports the log mean ratio in $estimate and the mean difference in $md, $md_se, $md_lower, $md_upper. Under the identity link the two coincide.
• Fix: stc() refuses a binomial outcome model with complete separation and, when detectseparation is installed, tests exactly for quasi-complete separation; the result records a separation status. A Poisson model with no events is refused the same way. See ?stc.
• Fix: Directly observed arm proportions and rates in naive(), and the observed comparator proportion in a binomial stc(), use exact Clopper-Pearson and Garwood intervals instead of Wald intervals, which excluded the truth at zero counts. Contrasts remain Wald.
• Fix: naive() combines several aggregate rows as strata; the comparator standard error is that of the size-weighted mean of the row proportions.
• Fix: naive() and stc() apply a continuity correction only to an observed proportion of 0 or 1 instead of clamping every proportion, which slightly changes the reported effect for arms with no events or no non-events.

Fitting, diagnostics and model comparison

• Breaking: mlumr() gains model-defining arguments ahead of chains, iter and the other sampler controls. Pass sampler settings by name.
• Improvement: mlumr() centers covariates by default (center = TRUE), which removes the intercept-versus-slope collinearity that pushed the sampler into deep trajectories on raw-scale covariates. Reported contrasts are unchanged; prior_intercept then applies at the pooled covariate mean. qr = TRUE offers a QR-rotated design instead.
• Improvement: The binary, continuous and count IPD likelihoods use Stan's fused GLM densities on their canonical links; results agree with 0.1.0 to Monte Carlo error.
• Fix: With the cmdstanr engine, fit$summary$se_mean is the Monte Carlo standard error of the posterior mean from posterior::mcse_mean(). It was the posterior SD over the square root of the bulk ESS, which is computed on rank-normalized draws and misstates the error for skewed quantities such as a ratio.
• Fix: mlumr() refuses a normal fit whose outcome is constant or reproduced exactly by its covariates, where the posterior for the residual SD is improper, and warns when the design is saturated.
• Fix: Marginal probabilities, means, rates and their contrasts are formed on the log scale; the safe_logit() and safe_divide() clamps are gone, so ratios with a near-zero comparator are no longer understated and lor_* is finite where it used to hit the clamp. A ratio beyond double precision is Inf. See ?mlumr-numerical-evaluation.
• Breaking: predict(type = "link") returns the marginal link g(E[g^-1(eta)]) rather than the mean linear predictor E[eta]; the two agree for the identity link only. This differs from multinma on purpose, since every effect mlumr reports is population-standardized.
• Improvement: seed = NULL uses the documented default of 2026 with a warning instead of drawing an unrecorded seed from the session RNG.
• Improvement: verbose = FALSE silences the cmdstanr sampler banner as well.
• Fix: A caller's rstan control list is merged with the backend's instead of failing argument matching.
• Fix: A cmdstanr run that produced no draws reports how many chains wrote nothing and where CmdStan's messages are, instead of failing inside checkmate on a temporary path.
• Fix: The cmdstanr executable cache key now covers every #include, the CmdStan version and its build flags; a changed include no longer reuses a stale executable. Expect one recompile after upgrading.
• Fix: check_diagnostics() and the fit summary no longer drop an infinite Rhat before taking the maximum, report missing or non-numeric diagnostics as unknown rather than as zero, and compute tail ESS.
• Improvement: predict(), marginal_effects(), conditional_effects() and conditional_predict() warn once per call when NA or NaN draws were dropped from a summary.
• Fix: conditional_predict() names quantile columns from the requested probabilities, so non-round probs no longer return NA.
• Fix: calculate_loo() refuses moment_match = TRUE, which loo ignores for a matrix.
• Fix: compare_models() with "loo" or "waic" refuses fits whose stored outcomes differ. loo_compare() pairs the pointwise scores by position, so the same data fitted in another row order gave a wrong se_diff. The calculate_loo() and calculate_waic() results carry the outcomes as loo's yhash, so loo::loo_compare() warns on the same mismatch.
• Improvement: The compare_models() printout no longer presents se_diff > 2 as a decision rule.

E