marginaleffects

• Many comparisons() and slopes() calls are at least three times faster on large datasets.
• Built-in ratios, lifts, slopes, and elasticities now use exact derivatives.
• Linear hypothesis= expressions and supported formula shortcuts compile to contrast matrices, and post-hoc hypotheses() uses the same exact path for linear expressions, including constants. Nonlinear hypotheses compose with upstream Jacobians instead of re-differentiating the full pipeline. Issue #1735. Thanks to @nremenyi for #1750.
• Bayesian posterior draws processing is much more efficient.
• Set options(marginaleffects_analytic_jacobian = FALSE) to disable analytic Jacobians and force the numerical fallback, which is useful for validation and debugging.

Covariance and inference:

• User-supplied covariance matrices must cover all coefficients. Named matrices are reordered and validated; unnamed matrices warn about positional matching. Errors from vcov functions are now propagated.
• vcov = "stata" now maps to HC1, matching Stata, estimatr, and modelsummary; use vcov = "HC2" for the previous behavior.
• multcomp= and simulated Wald inference now respect the hypothesis null and direction. Bonferroni confidence intervals now match Bonferroni p values.
• Exact zero standard errors are reported as NA, along with their associated inferential quantities. This distinguishes structural or constant estimands from effects estimated with very small uncertainty.
• Per-observation df vectors are rejected with aggregation or hypotheses; df = NULL remains valid. Equivalence tests support vector-valued degrees of freedom and reject reversed bounds.
• Weighted average comparisons no longer return NaN when a group's first weight is zero; groups with zero total weight remain undefined.

Hypotheses and contrasts:

• Numeric variables= functions and data frames honor lo/hi and low/high regardless of order and may return data frames, data tables, or tibbles. Logical comparisons honor their requested direction.
• hypothesis formulas group by datagrid() variables when by is omitted and can resolve external variables such as ~ I(drop(t(www) %*% x)) | am. Thanks to Matt Thomas for #1705 and @mattansb for #1707.
• Hypothesis functions may declare newdata and by; empty results now error. Joint tests prefer exact name matches, reject ambiguous or missing matches, and report null values correctly. Matrix hypotheses preserve posterior and bootstrap draws.
• ~meandev and ~meanotherdev now use O(np) time and memory instead of dense n x n matrices, reducing memory use from 573 MB for 10,000 estimates.
• numderiv = list("richardson", side = ...) is now accepted; the validator previously rejected side and accepted the nonexistent size argument.

Data, grids, and grouping:

• Grouped comparisons and slopes support by = "rowid", avoid collisions with data columns named predicted, term, or contrast, preserve posterior and bootstrap draws when groups are incomplete, and retain comparison = "lift" rows with duplicate labels. Thanks to @coughlanja for #1669.
• Formula terms are parsed to recognize factor(x, levels = ...), as.numeric(x), and as.logical(x), preventing invalid datagrid() values and NA predictions.
• datagrid() now preserves counterfactual row identifiers across groups, supports grouped data.table inputs, honors response, validates grouping columns, and treats numeric fixest fixed effects as categorical. Thanks to @arcruz0 for #1685.
• newdata shortcuts now handle logical, integer, binary, and missing values as documented. Existing rowid columns are no longer modified by reference.
• avg_predictions() with factor variables uses all factor levels when the original model data is unavailable. Thanks to @trose64 for report #1703.

Model compatibility:

• Corrected standard errors for survival::survreg() models with estimated scale, named systemfit() equations, mhurdle scale parameters, and zero coefficients in nnet::multinom().
• Repeated standard-error calculations for nls() models are now reproducible.
• betareg models without a precision formula now align the precision parameter names used by get_coef() and get_vcov(), allowing user-supplied covariance matrices.
• lm and glm estimates work when qr$qr was removed from the fitted object; linear predictions fall back to the model matrix when stats::predict() fails. Thanks to @trose64 for report #1748.
• vcov = "HC0" no longer changes response-scale point estimates for glmmTMB models with random effects; robust covariance extraction now restores the internal TMB state. Thanks to @nremenyi for the report.
• Fixed mira models fitted with formula transformations and subset=, and enabled predictions(<mira>, newdata = datagrid(...)) for mice objects. Thanks to @dmongin for #1682.
• flexsurv prediction standard errors now synchronize transformed coefficients and delegate to flexsurv::predict() when vcov = TRUE. Thanks to @colinorourke for report #1670.
• hypotheses() no longer errors on survey::svyglm() objects when no averaging is required. Thanks to @bdwilliamson for Issue #1723.

Bootstrap and conformal inference:

• Bootstrap and refit() reject models fitted with external weight vectors. Bootstrap refits no longer resolve unrelated global data, and fractional weighted bootstrap uses model-stored data. Thanks to @U8NWXD for #1713.
• Bootstrap inference with a user-supplied estimator uses the conf_level from the estimator result.
• Conformal intervals now use the correct finite-sample quantile, warn and return infinite bounds when calibration data are insufficient, respect transformed outcomes, and preserve absolute residuals for squared-residual CV+ scores.

Predictions and comparisons:

• Comparison rows are now dropped when either prediction is missing.
• transform= handles NaN confidence bounds and reports wrong-length function results clearly.

Plots, printing, and diagnostics:

• Plot functions avoid duplicate lines, handle missing response labels and duplicated rounded condition values, and support points= and rug= with cbind() responses.
• Printing handles NA values in df. print(style = "tinytable") preserves Term: and Comparison: notes when their columns are dropped and bases that decision on the full object rather than the displayed rows.
• Fixed diagnostics and validation in expect_margins() and expect_slopes().
• Models with NA coefficients now trigger a once-per-session rank-deficiency warning. Thanks to @andymilne for #1744.