spicy

• cross_tab() counts observations at an explicit NA factor level as a regular row or column. freq() excludes them from n_valid and the valid percent.
• freq() and cross_tab() replace styled with output. styled = FALSE becomes output = "data.frame", and styled now errors with the replacement.
• cross_tab(output = "data.frame") returns a plain data.frame without metadata attributes. Read them from the default object, for example attr(cross_tab(...), "p_value").
• freq() defaults to rescale = FALSE, matching cross_tab(). Use rescale = TRUE for the previous behavior.
• freq(), table_categorical(), table_continuous(), and table_continuous_lm() no longer print when their result is assigned. freq() drops its unused ..., so unknown arguments error.
• options(OutDec) no longer changes spicy's output. Every number follows decimal_mark alone, or the style or language in force.
• table_categorical() defaults to drop_na = FALSE, showing missing values as a "(Missing)" level. Its labels must be a named vector, and p_digits below 1 errors.
• table_categorical(output = "long") names the association column effect_size and adds effect_size_type. Replace out[["Cramer's V"]] with out$effect_size.
• table_categorical(output = "flextable") no longer writes a .docx when word_path is supplied. Use flextable::save_as_docx().
• table_continuous_lm(output = "data.frame") names the effect-size interval bounds es_ci_lower / es_ci_upper, like the "long" output.
• standardized = "smart" scales continuous inputs by 2 SD and leaves binary inputs unscaled. The rule was applied inverted since 0.12.0, so those betas change.
• table_regression(exponentiate = TRUE) errors on links whose exponentiated coefficient is not a ratio (probit, cauchit, inverse, sqrt).
• keep / drop no longer match the intercept row. show_intercept alone controls it.
• align = "auto" is removed. Use "decimal" (the default), "center", or "right".
• The show_fit_stats information criteria are lowercase tokens ("aic", "aicc", "bic"). show_fit_stats = character(0) errors. Use FALSE to suppress the block.
• With several models, show_columns = "all_b" / "all_ame" drop the CI columns. Request atomic tokens to keep them.
• A robust vcov that cannot be computed is now an error, and a cluster containing NA is refused. It used to warn and label the classical variance robust.
• Weighted n and glance()'s weighted_nobs are NA for an unweighted glm(). They used to repeat n.
• as_structured() describes each row in the body itself (body$.row_role, body$.indent, cell_status). The 0.12.0 row-index vectors are removed, version is 3, and a view built by an older spicy is refused.
• tidy() labels AME rows estimate_type = "ame" (was "AME").
• count_n() warns and returns NA when the selection resolves to no usable column. mean_n() and sum_n() with min_valid = 0 return NA for rows with no valid values.
• copy_clipboard() arguments use snake_case. build_ascii_table() is no longer exported. Use spicy_print_table().
• Association measures with detail = TRUE always include an se element. On degenerate tables, gamma_gk(), kendall_tau_b(), kendall_tau_c(), and uncertainty_coef() return NA with a classed warning instead of a spurious value. conf_level is validated everywhere.
• The package ships a single vignette, *Get started*. The walk-throughs live as articles on the package site, at the same URLs, and vignette("<name>") no longer finds them.

New features

• table_regression() supports more than thirty model classes beyond lm / glm: mixed effects (lmer, glmer, glmmTMB, lme, gls), GEE, Bayesian (stan_glm(), stan_glmer(), brm()), survival (coxph, survreg, cph, flexsurvreg), ordinal (polr, clm), multinomial (multinom, mlogit), two-part counts (zeroinfl, hurdle), fixest, estimatr, ivreg, tobit, rq, rlm, glm.nb, nls, gam, betareg, selection, rms, and the design-based svyglm, svyolr, and svycoxph. ?table_regression_models is the registry. A request a class cannot honor is refused with a classed error, never rendered as an empty column.
• Each family renders with its own conventions. Mixed models report their random effects as a block of rows with SE and CI, the ICC, and a boundary-correct test of the random part. Ordinal models report their thresholds. Two-part models report every component. fixest and estimatr fits disclose their absorbed fixed effects as a Fixed effects: block. Bayesian fits report posterior medians, MAD SD, and credible intervals, with no p-values and a sampler-diagnostics guard.
• New table_continuous_svy() and table_categorical_svy() summarize a survey design object, with every statistic computed by survey and the design degrees of freedom throughout. Design-based regressions (svyglm(), svyolr(), svycoxph()) report both counts and name their variance estimator in the note.
• New table_outcome() summarizes one continuous outcome across several categorical variables, one block per grouping, with the group comparison and an Overall row.
• New table_regression_uv() builds univariable screening tables for lm, glm, and coxph outcomes, one fit per predictor merged beside the multivariable model, with a per-predictor N column.
• New inline() cites one table cell in Quarto or R Markdown text. The returned string is exactly the displayed cell, so a quoted number can never drift from the table.
• New style argument on the four table families, and options(spicy.style = ) for a whole document: "jama", "nejm", "lancet", "annals", "apa", and "aer". A theme applies the rules its journal publishes, as defaults, so any argument you pass wins. spicy_style() builds a style by hand or from a theme.
• options(spicy.language = "fr") prints table labels in French and brings French typography with it: a decimal comma and a leading zero on p-values (0,003). Machine outputs, column names, and messages stay in English. options(spicy.labels = ) overrides one label at a time, and spicy_labels() lists them.
• New show_columns families "rmst" and "risk_diff" for coxph and survreg fits: covariate-adjusted differences in restricted mean survival time and in cumulative incidence, by g-computation with bootstrap inference, in single tables and in the univariable screen.
• New show_columns token "n_events" shows event counts as events/N beside the estimates, for binomial outcomes and coxph fits.
• Heteroskedasticity- and cluster-robust vcov across the supported classes, with each class's field-standard backend. "CR1S" reproduces Stata's regress, vce(cluster) exactly. What no backend supports is refused, never approximated.
• ci_method = "profile" gives profile-likelihood CIs for glm, polr, and clm. ci_method = "boot_percentile" reports percentile CIs from the bootstrap replicates.
• nested = TRUE works across the new classes with the correct test for each, and refuses hierarchies that are not comparable.
• AME columns are available for many more classes, per outcome category for ordinal and multinomial models, and honor a robust vcov. broom::tidy() gains outcome_level for those rows.
• table_categorical() and table_continuous() gain smd = TRUE, a standardized-mean-difference column, the balance diagnostic of a Table 1.
• table_continuous() gains weights and rescale, under a documented convention that matches Stata's [aweight] and survey::svyvar(). Group tests are refused under weights. table_continuous_lm() is the tool for that.
• table_continuous() gains show_columns with median tokens ("med", "q1", "q3", "iqr", "med_iqr", "med_ci"), per variable via a named list. A variable shown as a median is tested as one.
• select is optional in table_categorical(), and table_continuous() gains drop_na = FALSE.
• as_structured() reads the descriptive tables too, and carries everything the printed table shows, with a per-row identity (.variable, .level, .row_role) that survives stacking.
• Seven new articles on the package site: mixed-effects, GEE, multinomial, count and two-part, survival, ordinal regression tables, and categorical predictors.

Minor improvements and bug fixes

The first eight fixes change numbers that 0.12.0 reported.

• kendall_tau_b() reported wrong standard errors, confidence intervals, and p-values in every release from 0.6.0 through 0.12.0. Point estimates were correct. assoc_measures() and cross_tab() were affected too.
• Binomial models fitted with a cbind(successes, failures) response were refitted with squared weights by every internal refit. Bootstrap and jackknife inference, the default McFadden and Nagelkerke R², and standardized = "refit" were wrong for them. Fits with a 0/1, factor, or proportion-plus-weights response were never affected.
• Average marginal effects use the fit's prior weights, so AME values change for weighted fits.
• Partial effect sizes are true Type-II tests. In models with interactions, main effects no longer depend on the factor coding.
• ci_method = "profile" with a robust vcov defers to the vcov and warns.
• table_categorical() computes the ordinal association measures in declared level order under drop_na = FALSE.
• table_continuous_lm() reports correct estimates when by is an ordered factor and correct "balanced" adjusted means with an ordered-factor covariate. It pins treatment contrasts, so options(contrasts = ) no longer alters the results.
• cross_tab() computes weighted totals from the unrounded table.
• Under decimal_mark = "," every surface follows the mark. P-values keep their leading zero (0,018), and the star legend, the change statistics, and the association intervals read the comma.
• nested = TRUE no longer reports a negative chi-square with a p-value when the models are passed largest-first.
• cramer_v(), phi(), and contingency_coef() return NA with a classed warning on a zero margin, and somers_d(direction = "symmetric") returns 0 on equal concordant and discordant pairs.
• The tau_c measure is labelled "Stuart's Tau-c" everywhere.
• freq() keeps its label footer when NA-weight rows are dropped, warns when distinct codes merge under one label, and sorts labelled variables by code under sort = "name+".
• cross_tab() reports excluded missing values in the table note, accepts logical weights like freq(), and no longer swallows the warnings of its association measures.
• table_categorical() displays labelled columns as "[code] label" levels in every path, and keeps both the group and the margin when a by level is named "Total". Its machine outputs carry full-precision values and the documented Chi2 and df columns.
• table_categorical(), table_continuous(), and table_continuous_lm() resolve by data-first, like tidyselect. A by with no level to tabulate is refused.
• table_continuous() forms groups from a non-factor by in order of first appearance, and degrades per variable when a test fails on degenerate data.
• table_continuous() and table_continuous_lm() label an interval with its own coverage (97.5% CI, not 98% CI).
• table_continuous_lm() discloses robust and resampling SEs in the note, treats a value-labelled by as categorical, and degrades cleanly on degenerate fits.
• output = "gt" tables keep their note when saved or printed non-interactively. gt and flextable outputs render in Quarto and R Markdown Word, PowerPoint, and PDF documents, where they silently disappeared. New as_flextable() returns the underlying flextable.
• output = "gt" escapes a by level or model name carrying a quote, an angle bracket, or a backslash.
• stars = TRUE marks the coefficients in every output, not just the console.
• A cell whose statistic applies but has no number shows the console's en dash in every rich output, instead of a blank.
• A table_categorical(by = ) table carries its association note to every output. The descriptive gt and tinytable outputs also draw the title and the rule between variable blocks.
• Factor levels are indented once, not twice, in the tinytable, Word, and Excel outputs.
• Tables without a confidence-interval column lose their empty header strip, multi-line notes keep one disclosure per line, and Typst output no longer forces a column gutter under grouped headers.
• Footer lines cite a model by its displayed label, not Model 1.
• output = "excel" writes the significance stars, blank cells instead of #N/A, honors align, and sizes its columns to the text.
• output = "clipboard" quotes cells that contain the delimiter, ships plain text instead of Excel formulas, and errors clearly on a system without a clipboard.
• Console layout survives NA cells, empty cells, and wide characters (CJK, emoji).
• table_regression(m1, m2) without list() errors helpfully, and an NA or colliding model name no longer crashes the table.
• Factor rows follow levels() order instead of alphabetical, and factors fit with non-default contrasts (successive differences, sum-to-zero, Helmert) group under their parent variable.
• A factor level containing : ("Part-time: 50-89%") stays inside its variable block. It used to be mistaken for an interaction term.
• The statistic column header follows each model's reference distribution (z or t).
• Bootstrap, jackknife, and standardized = "refit" refits no longer leak the caller's environment and work on factor() / log() / poly() formulas.
• varlist() and code_book() render POSIXlt columns and difftime units, and show an explicit NA factor level as <NA>.
• count_n() resolves select and exclude through the same tidyselect path as mean_n() / sum_n(), and errors clearly on an unusable count or special.
• The tabulating and summarizing functions reject bit64::integer64 input with a classed error naming the fix.
• Error messages quote values the same way on every platform, and the enum arguments raise classed errors naming the valid values.
• copy_clipboard() re-emits backend messages and warnings as real R conditions.
• An estimatr fit reports R² and adjusted R² by default like lm, discloses its absorbed fixed effects, and labels its model type after its se_type.