NEWS

Output

• Tables stay one block like SPSS wherever that can be displayed: in a knitted HTML document (R Markdown, Quarto, pkgdown) wide describe() tables and correlation/reliability matrices are no longer split into column blocks at the console width of 80, and labels are not shortened. In the console they are still split only when one block would be wider than the console.
• frequency(), crosstab(), and multiple_response() now print their full tables directly, matching describe(): the result of a descriptive table function is a table, so print() no longer shows a compact placeholder with a "Use summary()" hint. print(x) is now identical to print(summary(x)) with default toggles; summary() remains the place for section toggles (frequency_table=, residuals=, ...) and digits=. The compact one-line print() is unchanged for the hypothesis-test and model classes.

Bug fixes (2026-09 field report)

• logistic_regression() with fractional weights no longer leaks the "non-integer #successes in a binomial glm!" warning under non-English locales (e.g. "Nicht-ganzzahlige #Erfolge in einem binomial-GLM"). The muffling filter matched the English text only; it now rebuilds the message through R's own translation catalog.
• linear_regression(): when a predictor is excluded for perfect collinearity, Tolerance and VIF of the retained terms are now computed without the excluded column. Previously the singular correlation matrix yielded either NA (so summary(collinearity = TRUE) printed no table) or rounding artefacts such as VIF = -2.85e13.
• kendall_tau() is dramatically faster. The pair counts came from an element-wise R loop over all n(n-1)/2 pairs, and on labelled (haven_labelled) data every element access went through vctrs dispatch: 400 labelled cases took 6.5 s, the full ALLBUS sample more than 20 minutes per pair. The counts now come from a vectorized kernel (.kendall_pair_counts() in R/kernels-weighted.R, exploiting that the weighted pair weight sqrt(w_i * w_j) is separable), and the correlation engine strips label classes before computing. Full ALLBUS 2023: 6 pairs in 0.1 s. Results are unchanged.
• group_by() %>% t_test() no longer aborts with "arguments imply differing number of rows: 1, 0" when a group lacks one level of group. The per-group fallback itself built a 0-row data frame and swallowed the real reason. The group now gets an all-NA row, a warning naming the group and the reason, and print()/summary() say "not computed for this group" instead of showing NA tables.
• mann_whitney() and kruskal_wallis(): a variable/group that cannot be tested (e.g. a group lacking a level under group_by()) no longer vanishes silently from $results. The error handler assigned its NA row in its own scope, so the row was discarded. The row is now kept, the warning names the group, and the print methods say "not computed for this group".
• Grouped output with a missing group key (NA, including tagged NAs from SPSS files) is printed correctly. All grouped print methods selected a group's rows with ==, which is NA for a missing key: every group got an extra all-NA ghost row, and the NA group showed only NA rows instead of its statistics (group_by() %>% describe() was the reported case; the pattern sat in ~30 print paths). Matching now goes through one NA-safe helper.
• Group headers in grouped output show factor levels and value labels instead of codes: the compact print() of ~20 classes showed [region = 1] instead of [region = East], and labelled group variables showed their numeric codes everywhere (including the summary() of reliability()/efa() and the grouped regression output).
• Statistics of an empty (all-missing) variable or group are NA instead of NaN/-Inf: describe() reported Mean = NaN and Range = -Inf (plus two R warnings from min()/max()), and the unweighted w_mean() returned NaN. The mean/range kernels and the unweighted branch of the w_* factory now short-circuit empty input the way the weighted branch already did.
• to_label() now keeps the original code of every level in a "codes" attribute, and to_numeric() / to_labelled() restore those codes. Previously the round trip renumbered the values sequentially (6 -> 3, 42 -> 4, 90 -> 5, ...), and to_numeric(keep_labels = TRUE) attached labels with the wrong codes. use_labels = FALSE still gives the sequential 1..k numbering. The attribute survives dplyr verbs; after base subsetting (x[i]) or renaming levels, the level-based conversion applies as before.
• The pipe %>% is re-exported, so the documented survey_data %>% describe(age) works after library(mariposa) alone (it was only imported, giving 'could not find function "%>%"'). No masking message when dplyr is attached as well (identical object).
• spearman_rho() no longer presents itself as a weighted correlation. Weights only filter cases (SPSS NONPAR CORR convention), but the output said "[Weighted]" / "Weighted Spearman's Rank Correlation Analysis", and the help page's description and example spoke of weighted correlations. The output now names the weights variable as "case filter only".
• The correlation vignette's method comparison showed Pearson's r three times: it read a non-existent $correlation column for Spearman and Kendall (NULL), and data.frame() recycled the Pearson value. It now reads $rho and $tau.
• codebook() on a full SPSS survey file is ~6x faster (ALLBUS 2023, 5,246 x 579: 20.4 s -> 3.1 s), and frequency(), to_label(drop_na = FALSE), unlabel(), drop_labels() and the Stata/SAS readers speed up on data with many missing values. Six places read tagged-NA letters with vapply(x, haven::na_tag, ...), splitting labelled vectors into one vctrs object per element; they now share one vectorized reader.
• Weights read from SPSS files work everywhere. A weight imported by read_spss() is haven_labelled_spss with an na_range and format attributes; once it contained NA, every comparison on it (w < 0 in the weights check, w > 0 filters, ...) failed inside haven's cast with "missing value where TRUE/FALSE needed", so every weighted function aborted (reported for crosstab(), affected all of them). Weights are now converted to plain numbers at the entry point.
• crosstab() now validates weights through the package-wide policy: negative weights are an error (previously only a warning, unlike every other weighted function since 0.6.4).
• chi_square() excludes cases with a missing weight. NA weights were summed into the table and made chisq.test() abort ("all entries of 'x' must be nonnegative and finite").

Practice-test fixes (2026-09/10)

Found by a seven-agent practice test of the package in everyday analysis use (survey_data and the ALLBUS 2023 SPSS file), each finding reproduced before it was fixed; SPSS-validated numbers only moved where the SPSS reference output showed mariposa was wrong.

Descriptive statistics

• describe() and the w_* functions with na.rm = FALSE: a variable with missing values now gets NA for every statistic, as in base R. The weighted quantile family returned values computed on shifted positions (the weighted median of income was 3800 instead of NA), and the unweighted describe() / w_iqr() aborted with a base-R quantile() error.

• describe(show = "all") locates the quantile columns by their exact names. The print matched them by a regular-expression prefix, so a second variable such as income_Quintile crashed the print and a name with special characters (Einkommen (EUR)) silently lost its quantiles.

• describe() rejects unknown show values with an error listing the valid ones (show = "min" used to print a table with nothing but N and Missing), names quantile columns with at most two decimals (probs = 1/3 gives Q33.33, not Q33.3333333333333), no longer prints Q50 next to the identical Median, and reports undefined statistics of a constant variable as NA instead of NaN.

• describe() and the w_* functions on grouped data no longer crash when the selection contains a grouping variable (explicitly or through a helper such as where(is.numeric)); errors ranged from "'x' is NULL" and "quantile.haven_labelled() not implemented" to a dplyr internal error. As in dplyr::across(), grouping variables are excluded from the analysed variables, with a message.

• Weighted describe() prints N and Missing like SPSS FREQUENCIES with WEIGHT BY: N is the sum of the weights of the valid cases and Missing the weighted missing count (income: N 2201, Missing 315). The table used to show only Kish's effective sample size (2158.9) under the name Effective_N and no Missing column. The effective N stays available in $results (<variable>_Effective_N); <variable>_Missing is now the weighted missing count for weighted analyses.

• describe() output: the table is sized to its content (the borders were always 40 dashes), grouped output no longer stacks the table rule directly under the group underline, every statistic column uses the same number of decimals (no 50 next to 50.550), and a table wider than the console is split into column blocks that each repeat the Variable column (the continuation block used to lose it).

• w_mean() and the other w_* functions (including w_modus()) print every group combination when data are grouped by two or more variables. The print iterated over the first grouping variable only: region x gender showed two blocks labelled "region = East"/"West" that silently contained the Male rows.

• The w_* functions print one uniform table: Variable, the statistic, N and Missing, one table per group. The prints used to differ across the family (raw column names such as weighted_mean/Effective_N under "--- var ---" headers, w_modus() as a raw tibble, w_quantile() repeating the weights name on every row). With weights, N and Missing are sums of weights as in SPSS; Kish's effective N (previously the only N shown) is displayed by the new summary() methods for all eleven w_* classes. $results has the same columns for one or several variables (Variable, the statistic, n or weighted_n + effective_n, missing); single-variable results no longer carry the duplicated raw columns (age, age_n, age_eff_n).

• w_* input checks: w_mean(1:4, weights = c(1, 3)) says that the lengths differ (was a cryptic base-R error), w_mean(c("a", "b")) says that a numeric vector is needed (was "data must be a data frame"), and w_mean(survey_data, gender) names the non-numeric variable (was NA plus a base-R warning).

• w_modus(): when several values share the highest frequency, the smallest value (first factor level) is returned for weighted and unweighted data alike, as in SPSS (the weighted version took the first value in data order), $results$n_modes counts the tied values, and the print flags the result ("Multiple modes exist").

• Weighted frequency() of several labelled variables no longer aborts with "Can't convert ... due to loss of precision" depending on the order of the variables. The weighted branch kept the haven_labelled class in the value column, so combining variables with different label sets failed; values are now bare numbers, as in the unweighted branch.

• frequency() no longer lists empty factor levels (e.g. "Student 0" after filter(employment != "Student")) unless show_unused = TRUE, as SPSS FREQUENCIES lists observed values only; with show_unused = TRUE they now also appear in weighted tables.

• frequency(show_labels = "auto") also shows the Label column when a labelled missing-value code occurs (ALLBUS age: -32 "NICHT GENERIERBAR"); the automatic mode only looked at the labels of valid values, so metric variables printed their missing codes without the explaining label.

• frequency() (console and write_xlsx()) labels system-missing values "System" as SPSS FREQUENCIES does ("Missing System 79"); the row read "NA".

• frequency() tables follow the SPSS FREQUENCIES layout: valid categories, "Total valid", the missing categories, "Total missing" (only with two or more missing categories) and a grand "Total" row (N and 100 %). Cells without a value are empty instead of reading "NA"; plain numeric, logical and character variables no longer end with two rows both called "Total", and tagged missing values no longer end with a "NA(total)" row. Further:

  • show_valid = FALSE also hides the cumulative percentages (they are cumulative valid percentages).
  • every column is sized to its content: large weighted counts were cut to "2798..." by a fixed 8-character N column.
  • factors, character and logical variables show their categories once (Value and Label columns used to repeat each other), and the summary line leaves out statistics that do not exist ("mean=NA sd=NA", "mean=NaN" for an all-missing variable).
  • labels are left-aligned and never cut (they were cut at 40 characters); when the table is wider than the console, long labels wrap onto extra lines.
  • an all-missing variable prints its missing rows and the Total without a "Valid % 100.00" of zero cases or an R warning.
  • print(x, digits = 0) rounds the percentages only; the summary line keeps two decimals.

• Weighted crosstab() reproduces SPSS CROSSTABS' default /COUNT ROUND CELL: every weighted cell count is rounded first, the margins are sums of the rounded cells and percentages, expected counts and adjusted residuals come from the rounded table. Cells used to be rounded only for display while margins and percentages came from the unrounded sums (402 + 447 shown with a margin of 848; ALLBUS DIVERS 15 | 3 with a row percentage of 81.7). Weighted counts, margins and percentages now match the SPSS reference exactly (validated for the weighted ungrouped and grouped scenarios).

• crosstab(): the Total row shows every requested percentage as SPSS CROSSTABS does (row % = column shares, col % = 100 %, total % = column shares). With percentages = "col" it showed the column shares labelled "col %", with "row"/"total" it had no percentage line and with "all" only that mislabelled line.

• crosstab(digits = 2) is honoured by print() and summary(); their own default of 1 decimal overrode the value stored by crosstab().

• crosstab(): na.rm = FALSE now keeps cases with a missing value as their own "NA" row or column (it had no visible effect), summary(x, percentages = FALSE) says "Counts only" instead of "Row percentages", and crosstab(data, gender) without a column variable gives a clear error (was the base error 'argument "x" is missing').

• crosstab() shows observed categories only, as SPSS CROSSTABS does: an empty factor level (e.g. after filter(employment != "Student")) used to print a "0 0 0" row with a row percentage of "100.0%" of zero cases.

• Weighted crosstab() reports missing cases like SPSS's Case Processing Summary: as the sum of their weights (it printed an unweighted count next to the weighted "N (valid)").

• crosstab() output: category labels are shown in full (they were cut at 20 characters, so eight ALLBUS ISCO categories all read "FUEHRUNGSKRAEFTE,..."), each column is as wide as its own content (every column took the width of the widest label: sex x educ was 182 characters wide), long column headings and row labels wrap when the table would be wider than the console, and the title and the column spanner show the variable labels, as SPSS does, instead of the variable names.

• multiple_response() on data grouped by a labelled variable (e.g. ALLBUS eastwest) no longer aborts with "arguments imply differing number of rows: 1, 2".

• multiple_response(by = ) with a labelled by variable heads the crosstab columns with the value labels in code order (it showed the codes, e.g. "1"/"2" for ALLBUS eastwest).

• multiple_response() tables: counts are whole numbers (unweighted counts printed as "583.0"; weighted counts had one decimal in the frequencies table but none in the crosstab), the frequencies table ends with a Total row, and the crosstab has a Total column and a "Total (cases)" row as in SPSS MULT RESPONSE (replacing the unwrapped "Cases per column" footer). Factor or character indicators whose levels do not contain counted (e.g. "no"/"yes" with the default counted = 1) are an error instead of silently counting 0 mentions.

• codebook() console output (print()/summary()) writes counts without thousands separators ("2500 observations"), like every other table in the package and SPSS's default output.

• describe() gains show = c("min", "max"), SPSS DESCRIPTIVES' Minimum and Maximum (also part of show = "all"). Before show was validated, these names were silently ignored.

Parametric tests

• levene_test(center = "median") on a weighted factorial_anova() is an error instead of silently returning the unweighted test (the median recomputation used the raw, unweighted values). The mean-based test, which SPSS UNIANOVA reports for weighted models, is unchanged.
• t_test(), oneway_anova(), factorial_anova(), ancova() and the post-hoc tests on their results order the groups of a numeric or labelled grouping variable by code, as SPSS does, and show value labels instead of codes. t_test() took the group order from the order of appearance in the data, so sorting the data flipped the sign of t and of the mean difference; labelled groups printed as "Groups compared: 1 vs. 2", Tukey/Scheffe rows as "1 - 2", and factorial descriptives and ANCOVA marginal means by code. (PAR-02, PAR-10)
• A dependent variable without variance is no longer "tested". oneway_anova() on a constant variable reported F = 7256 *** (weighted) or F = 0.992 computed from floating-point noise in sums of squares that are exactly 0, factorial_anova() F = 4.011 *, and tukey_test() adjusted p-values. oneway_anova(), t_test(), tukey_test() and scheffe_test() now report such a variable (and one without any non-missing value) as "not computed ()" with a warning naming the variable; the other variables of the call are still tested. factorial_anova() and ancova() stop with a clear error. t_test() no longer aborts a multi-variable call with the raw base-R message (German locale: "Daten sind praktisch konstant", printed twice as "t_test() failed: ... / Caused by: ..."), no longer reports an all-missing variable as a grouping problem ("Found 0 levels"), and a grouping variable with 3 or more groups gives one error that points to oneway_anova(). (PAR-01, PAR-24)
• A group with a single case (weighted: a sum of weights <= 1) no longer aborts oneway_anova() and t_test() with the raw base-R error "not enough observations" from the Welch part. As in SPSS, the classical ANOVA / Student's t-test is computed and Welch's test is marked "not computed" with the reason; t_test(var.equal = FALSE) then reports Student's t with a warning. A group with zero variance no longer prints a Welch row "NaN 3 NaN NA " (weighted) or an infinite Glass' Delta, and the Welch block of summary(oneway_anova()) is titled "Robust Tests of Equality of Means" (SPSS) instead of "Assumption Tests", with df2 shown with decimals (1229.456, not 1229). (PAR-07, PAR-17)
• Grouped oneway_anova(), tukey_test(), scheffe_test() and levene_test(): a group that cannot be tested (e.g. only one level of group present in that group) is reported with a warning that names the group label and the reason. oneway_anova() printed a silent "Results not available", the post-hoc tests dropped the group without a word, and levene_test() warned "in group 1" (the factor code). (PAR-18, EDGE-13)
• One-sample t_test() follows the SPSS One-Sample Test: mean_diff is now the mean minus the test value mu (was: the mean itself, e.g. 3.628 instead of 0.628) and conf_int_lower/conf_int_upper are the confidence interval of that difference. The weighted interval now follows alternative (it was always two-sided). summary() shows the test value, alternative, confidence level and a One-Sample Statistics table (N, Mean, Std. Deviation, Std. Error Mean) and no longer prints an effect-size legend for effect sizes a one-sample test does not have. (PAR-05, PAR-06)
• summary(t_test()) prints SPSS-style tables: a Group Statistics table (N, Mean, Std. Deviation, Std. Error Mean per group; SD and SE were missing) and an Independent Samples Test table with t, df, p, mean difference, its standard error and the confidence interval for both variance assumptions. Significance stars now come from the exact p-value (p = 0.0008 was rounded to 0.001 first and got "**", p = 0.0497 printed as "0.05" without a star); p-values print as "<.001"/".308" instead of a bare "0", df as 2419 / 2384.147, weighted N as whole numbers ("1149", not "1149.0"), and digits applies to every column. Tables with umlaut labels stay aligned. (PAR-15, PAR-21, PAR-26)
• summary(oneway_anova()) prints an SPSS-style Descriptives table (N, Mean, Std. Deviation, Std. Error and the confidence interval of each group mean at conf.level) and formats the ANOVA, Welch and effect-size tables with fixed decimals and digits. With conf.level = 0.99 the header said 99% but no interval was printed anywhere; weighted N printed as "618.0", Mean Square as "289.79" next to "1.077", and digits had no effect. tukey_test() and scheffe_test() keep their own conf.level (default 95%): as in SPSS, the post-hoc interval follows the post-hoc alpha, not the ANOVA's confidence level (reference Tests 6a/6b). (PAR-19, PAR-21)
• group_by() %>% levene_test() takes its variables through ... like the ungrouped method: several variables, tidyselect helpers (starts_with("trust")) and group = "education" as a string work, and group/weights must be named. The old grouped method had the signature (x, variable, group, weights), so a second variable was silently used as weights ("[Weighted] F(3, 1544896) = 40250"). Grouped results now carry the group keys as columns: two group_by() variables no longer print "region = East, gender = East", and a missing key (NA) is its own group instead of a false "constant values" warning with "F(NA, NA) = ,". An invalid center is an error, a variable without variance prints "not computed (no variance ...)", and weights follow the package policy (negative or non-numeric weights are an error; they were accepted). (PAR-03, PAR-12, EDGE-04)
• summary(levene_test()) prints one SPSS-style table per group (Levene Statistic, df1, df2, Sig.) with fixed decimals and digits (p printed as a bare 0, df2 as 474.2032 next to 3), and its recommendation fits the design: Welch's ANOVA for three or more groups, Welch's t-test only for two groups, a caution for factorial designs. It used to recommend "Welch's t-test" after oneway_anova() and factorial_anova(). The compact line no longer ends in "p = 0.125 , variances equal". (PAR-14, PAR-21)
• tukey_test() and scheffe_test() report every comparison in the SPSS "(I) - (J)" orientation (I before J in the category order, difference = mean(I) - mean(J)) with a spaced separator, identically for the unweighted, weighted, Scheffe and factorial_anova() paths, and add the standard error of the difference. Unweighted Tukey rows came from TukeyHSD() as "Intermediate Secondary-Basic Secondary 0.497" (later minus earlier, unspaced) while weighted Tukey and Scheffe printed "Basic Secondary - Intermediate Secondary -0.490". The comparison tables stay aligned with umlaut labels (they were padded by bytes). (PAR-11, PAR-22)
• factorial_anova() and ancova() output: the Tests of Between-Subjects Effects table no longer wraps at 80 columns (Partial Eta Squared and the stars moved into a second block) and shows sums of squares with fixed decimals instead of scientific notation (1.754652e+09); the header reads "Sum of squares: Type III" instead of "Type III Sum of Squares: Type 3"; Levene's test prints "p < 0.001 ***" instead of "p = <.001"; the compact print shows N once in its title instead of on the last effect line only; descriptives, parameter estimates and marginal means honour digits. The ?factorial_anova examples no longer call a non-existent summary(marginal_means = FALSE) toggle. (PAR-20, PAR-26)
• ancova() Parameter Estimates follow the SPSS UNIANOVA coding: one row per category ("[education=Basic Secondary]"), the last category of each factor (and every interaction cell involving it) is the reference and shown as a redundant 0, and the intercept is SPSS's. They are now validated against the SPSS reference output. R's internal contrasts leaked into the table before (education.L/.Q/.C for the ordered education, gender1, a different intercept). Tiny coefficients print in e-notation instead of "0.000 [0.000, 0.000]". With two or more factors, ancova() also reports the main-effect marginal means (SPSS /EMMEANS=TABLES(factor), new element emm_main_effects) besides the cell means. (PAR-16, EDGE-25)
• factorial_anova() and ancova() with an empty design cell (e.g. no women with a university degree after filtering): ancova() no longer crashes with "Tibble columns must have compatible sizes", and effects whose Type III hypothesis has no degrees of freedom are reported as "not computed (not testable: the design has empty cells)" with a warning that lists the empty cells, instead of "F(0, 2208) = NaN, p = NA" (factorial) or F = -Inf (ANCOVA). The Corrected Model df is the rank of the design minus 1, as in SPSS (unchanged for complete designs). (PAR-08, PAR-09)
• factorial_anova() and ancova() honour group_by(): one complete analysis per group, with the group keys as leading columns of the result tables, per-group print()/summary() output, and grouped tukey_test(), scheffe_test() and levene_test() on the result. Both ignored the grouping silently and reported one pooled table. A group that cannot be analysed (e.g. no variance, a factor with one level) is skipped with a warning naming the group. The weighting of each group's fit is that of an ungrouped call. (PAR-04, EDGE-02)
• Clear errors for arguments that do not exist: t_test(paired = TRUE) explains that paired t-tests are not supported and points to wilcoxon_test() (was: "Can't select columns with TRUE"); t_test(x = , y = ) explains that variables come from data (was: "Variable x is not numeric"); normality_test(weights = ) says the tests are unweighted by design and normality_test(group = ) points to group_by() (was: "Variable weights/group is not numeric"). The errors of tukey_test(), scheffe_test() and levene_test() for unsupported objects now mention factorial_anova() results, which work as well. (PAR-23, PAR-24)
• The compact oneway_anova() line no longer leaves a dangling space ("p = 0.396 , eta2 = ...") and prints the total N also when the weights sum to more than 2^31 (was "N = NA"). (PAR-21)
• normality_test() names the variable (and, for grouped data, the group) it cannot test - no variance or fewer than 3 valid values - in a warning and prints "not computed ()" instead of silent "n/a" results; the Tests of Normality table stays aligned with umlaut variable names. (EDGE-13)
• summary(t_test()) with weights prints the Student and one-sample df as whole numbers like SPSS (2435, not 2434.609) and the Welch df with decimals; the unrounded values stay in $results. (PAR-21)

Non-parametric and categorical tests

• chisq_gof(): expected counts and residuals are stored unrounded and printed half up like SPSS (round() gave 121.2 and -0.2 where SPSS prints 121.3 and -.3), and grouped calls return and print the Frequencies table of every group as SPSS SPLIT FILE does (it was NULL).
• Exact tests with expansion (population) weights: the 2x2 Fisher p-value and the exact McNemar/binomial p-values come from distribution tails instead of enumerating every possible table, with the same result as fisher.test()/binom.test(). Weights summing to billions used to take minutes or run out of memory; an r x c fisher_test() above 2^31 weighted cases is now a clear error suggesting to rescale the weights (it failed with "cannot allocate memory block of size 134217728 Tb").
• chisq_gof(expected = ) is applied to every selected variable. With several variables it was silently dropped (equal proportions were tested while the summary header still showed the custom proportions); a variable whose categories do not fit expected is now a clear error. A named expected vector is matched by category name instead of by position (c(Female = .3, Male = .7) gave Male 30%). As with SPSS /EXPECTED=50 30 20, counts or other relative frequencies are accepted and divided by their sum; proportions that sum to about 1 (e.g. 0.995 from rounding) are rescaled with a message instead of shrinking every expected count, and proportions that clearly do not sum to 1 are an error. Categories with an expected count below 5 now trigger a warning, as in chi_square(), and a group that cannot be tested is reported with a warning naming the group instead of an unexplained NA row.
• chi_square(), phi(), cramers_v(), goodman_gamma() and chisq_gof() use the categories that actually occur in the data, as SPSS does. An empty factor level (typically left over after filter()) made chi-square, V and gamma NaN - also the cause of the silent NA row of grouped chi_square() - and gave chisq_gof() a phantom category with an extra degree of freedom (chi2 = 1257.5 instead of 5.0). A constant variable no longer crashes chi_square() and the effect-size helpers ("replacement has length zero"): the result is NA with a warning naming the variable (and group), and the output says "not computed (x has only one observed category)". The summary() tables show full value labels (no longer cut at 20 characters) and fixed decimals.
• goodman_gamma() and the gamma row of chi_square() are dramatically faster and their p-value now matches SPSS. The concordant/discordant counts came from a quadruple R loop over the table (cramers_v(survey_data, age, income) took ~50 s, because chi_square() always computes gamma); they now come from 2-D cumulative sums (0.03 s). The loop also counted only the pairs below each cell, which mis-stated the null-hypothesis standard error (ASE0): the approximate significance of gamma was wrong (education x employment: p = .122 where SPSS prints .027). The gamma value itself is unchanged.
• chi_square(correct = TRUE) computes Phi, Cramer's V and the contingency coefficient from the Pearson chi-square, as SPSS does; they were computed from the Yates-corrected statistic (gender x region: 0.0119 instead of 0.0129). The Pearson statistic is kept in pearson_chi_squared/pearson_p_value, summary() shows both rows ("Pearson Chi-Square" and "Continuity Correction") like SPSS, and the compact print() marks the statistic as "(continuity-corrected)". The compact line now spells out "negligible" (was "neglig.") and ends with the "Use summary()" hint.
• summary() of chi_square() follows the SPSS "Symmetric Measures" table: Phi and Cramer's V are shown for every table (Phi was hidden outside 2x2 although phi() returned it and SPSS prints it), and Goodman's gamma - with its verbal label - only when both variables are ordinal (ordered factor or numeric). For nominal variables such as gender x region its sign depends on the arbitrary category order. goodman_gamma() still computes gamma on request.
• chi_square() and the effect-size helpers warn once about expected counts below 5. Base chisq.test() added its own "Chi-squared approximation may be incorrect" warning (German: "Chi-Quadrat- Approximation kann inkorrekt sein") next to mariposa's message; that specific warning is now muffled in every locale.
• mann_whitney(mu = , alternative = ): U, Z, r and the p-value now refer to the same hypothesis. With mu the statistics were computed for a shift of 0 while the p-value came from wilcox.test(mu = ) ("Z = -0.696, p < 0.001"); group-1 values are now shifted by mu before ranking, as wilcox.test() does. For one-sided tests Z is directional (positive when group 1 tends to be larger) so that its sign matches the p-value (was Z = -0.226 next to p(less) = .589); the two-sided Z keeps the SPSS convention. The weighted (design-based) test supports only mu = 0 and now says so instead of printing "Null hypothesis (mu): 500" for an unshifted test.
• The unused confidence interval of mann_whitney() is no longer computed: wilcox.test(conf.int = TRUE) ran for every variable and was discarded (the source of German "cannot compute confidence interval" warnings for a constant variable), and summary() no longer advertises a "Confidence level: 95.0%" for which no interval exists. conf.level of mann_whitney(), kruskal_wallis(), wilcoxon_test() and friedman_test() is documented as not used.
• pairwise_wilcoxon(): the interpretation legend and help page now match the direction of the test. The legend said that a positive Z meant the first variable was higher, while it was the second (score_T1 vs score_T2: Z = +5.43 while T2 is higher). Z now follows SPSS (see "SPSS parity audit" below) and the new z_based_on column carries the direction.
• pairwise_wilcoxon() reports the number of cases of every pair (new n column, shown in summary()) and explains why it can exceed the Friedman N: each pair uses all cases with both values (pairwise deletion, exactly like the SPSS /WILCOXON tests the results are validated against), while friedman_test() uses complete cases. The comparison table prints p-values in SPSS style (<.001, .123).
• fisher_test() handles larger tables: instead of aborting with the raw "FEXACT error 501 ... hash table key cannot be computed" it now falls back to a Monte Carlo p-value with a warning (SPSS offers the same "Monte Carlo" option next to "Exact"). The new arguments simulate.p.value and B (default 10000 replicates, the SPSS default) choose it directly; before, simulate.p.value = TRUE was silently swallowed by .... A group that cannot be tested under group_by() is reported with a warning instead of a silent NA row.
• The rank tests treat ordered factors consistently as ordinal: they are ranked by their level order. mann_whitney() aborted with "'x' must be numeric" and wilcoxon_test() with "'-' not meaningful for factors" (printing "Z = ,"), while kruskal_wallis() and friedman_test() accepted them. A nominal (unordered) factor or character variable is now a clear error in all four tests instead of running silently (kruskal_wallis(gender, group = education)) or failing with "not computed for this group" outside any group_by().
• group_by() %>% binomial_test() no longer aborts because the variable has only one category in one group: that group gets an NA row, a warning naming the group and the reason, and the output says "not computed (gender has 1 observed category; ...)". An ungrouped single variable still stops with a clear error.
• Degenerate cases in the rank tests print a reason instead of broken text. wilcoxon_test() with identical variables reports Z = 0 and p = 1 as SPSS does (was "Z = ," and NA); a grouped friedman_test() whose group cannot be tested no longer prints "chi2(NA) = , , W = , N = NA" and its warning names the group; a constant variable in kruskal_wallis()/mann_whitney() is reported ("all values of x are identical") instead of "(see warning)" without a warning or German "cannot compute confidence interval" warnings; an all-missing variable says "x has no valid values" instead of blaming the grouping variable ("Found 0 groups ... use a Kruskal-Wallis test"). "not computed for this group" appears only under group_by(); otherwise the output says "not computed (reason)".
• Labelled (SPSS) variables show their value labels instead of codes in kruskal_wallis() ("Groups: 1, 2, ..., 7"), the pairs of dunn_test(), mann_whitney() ("1 vs. 2"), binomial_test() ("Group 1 (1)") and the categories of chisq_gof(), as chi_square() already did. Grouping variables are ordered by code, as in SPSS: a numeric 0/1 group was ordered by first appearance ("1 vs. 0" while the ranks listed 0 first).
• fisher_test() output: the contingency table is labelled with the variable names and value labels (was "r"/"cc" and codes), empty factor levels are dropped, grouped summary() shows each group's table (it was dropped), the compact line uses 3 decimals like the rest of the family (was "p = 0.5435") and reports the odds ratio with its 95% CI for 2x2 tables (SPSS "Risk Estimate": sample odds ratio, Woolf interval).
• mcnemar_test() output: without discordant pairs the output says "chi2 not computed (no discordant pairs), p = 1.000 (exact)" instead of "chi2 = , (asymp)"; tables are labelled with the variable names and value labels (was "v1"/"v2" and codes); grouped summary() shows each group's table and the "(cc)" continuity-correction marker (both were dropped); the compact line reports the degrees of freedom ("chi2(1) = ..."). Two variables with different category sets (e.g. 0/1 against 1/2) are now an error instead of being tabulated as if the categories matched.
• mann_whitney() without group stops with "group is required" (as kruskal_wallis() does) instead of the internal "Can't extract column with g_name".
• Output of the rank and exact tests is formatted consistently. summary() tables print p-values in SPSS style ("<.001", ".026") instead of a bare "p value 0" (Kruskal-Wallis, Wilcoxon, Friedman, binomial, Mann-Whitney), leave empty cells blank instead of "NA" ("Total 2500 NA", "Ties 502 NA NA"), show counts as integers (Mann- Whitney "n = 1149.0") and give every column a fixed number of decimals (0.52 next to 0.537, Mean Rank 19 next to 40.19, expected 538.24 next to 26.239). Compact lines label Kruskal-Wallis epsilon-squared and Kendall's W with an interpretation, mann_whitney() ends with the "Use summary()" hint, and Mann-Whitney and Wilcoxon share one set of r thresholds. The dunn_test() comparison table no longer wraps at 80 characters with long group labels, and a group or variable that dunn_test() cannot compare is reported with a warning naming the group.
• Skipped groups are reported consistently: mann_whitney() says how many groups with valid values a split has ("gender has 1 group ...") instead of suggesting a Kruskal-Wallis test, and pairwise_wilcoxon() names the group in its warnings; a pair whose values are all tied gets Z = 0, p = 1 (as wilcoxon_test() and SPSS) instead of a silent NA.

Correlation and regression

• logistic_regression() converges with expansion weights. glm()'s binomial start value sits on 0/1 when the weights are large, so the fit diverged (intercept -3.7e15, "algorithm did not converge") for weights summing to billions. The IRLS now starts at the (weighted) share of 1s; the estimates are unchanged.
• linear_regression(), logistic_regression() and marginal_effects() print models with long formulas correctly. A formula longer than about 70 characters (a normal model with five or more predictors) was split into two lines: the compact title was printed twice and summary() aborted with "'length = 2' in coercion to 'logical(1)'". The formula is now always shown on one line.
• confint() and profile() work on logistic_regression() results. Both failed with "incorrect number of dimensions" because profiling called summary() and got mariposa's SPSS-style summary instead of the glm one. confint() now returns Wald intervals by default - the SPSS "95% C.I. for EXP(B)" that summary() prints (exp(confint(model)) reproduces its Lower/Upper columns); confint(model, method = "profile") gives glm's profile-likelihood intervals. broom::tidy(model, conf.int = TRUE) uses the same Wald intervals and no longer prints dozens of "non-integer #successes" warnings for weighted models.
• marginal_effects() computes correct AMEs for transformed predictors. It perturbed one column of the model frame, so the transformed column kept its fitted values: for y ~ x + I(x^2) the AME of x was 0.466 instead of 0.173. Variables that enter only transformed (log(income), poly(x, 2), I(x / 10)), character and logical predictors were dropped without a word. The AMEs are now rebuilt from the original data for each variable (all terms using it move together); character and logical predictors get discrete-change rows, and a numeric variable used only as factor(x) is skipped with a warning that says how to get its AMEs.
• marginal_effects() on a grouped model whose grouping variable is haven-labelled no longer aborts with "arguments imply differing number of rows: 1, 2".
• Weighted linear_regression(): vcov(), confint(), nobs(), df.residual(), anova(), predict() (standard errors/intervals), broom::tidy() and broom::glance() now use the SPSS frequency-weight convention of summary() (N = sum(w), residual df = sum(w) - rank). They were inherited from lm(), which treats weights as analytic weights: with weights = w * 3, summary() showed SE = .0527 but tidy() .0916, and nobs() returned 2115 cases instead of N = 6388. Unweighted models are unchanged.
• logistic_regression() accepts every outcome with exactly two values, as SPSS does: 1/2 codings (lower value = 0, higher = 1), factors with an unused level (as to_label() leaves them; unused levels are dropped, the first remaining level = 0), labelled, character and logical outcomes. It used to demand 0/1 or a factor with exactly two levels. The output now says which category is modelled: the compact print shows [P(y = category)], summary() starts with the SPSS "Dependent Variable Encoding" table, and the classification table is labelled with the categories instead of 0/1. A constant outcome (previously "Nagelkerke R2 = -Inf ... Accuracy = 100%") and outcomes with more than two values now stop with a clear message.
• Grouped linear_regression() and logistic_regression() no longer abort when one group cannot be fitted (too few cases, a constant outcome, ...). Like SPSS SPLIT FILE, the group is skipped with a warning naming it and the reason, the other groups are reported, and print()/summary() list the group as "not computed". Before, one small group stopped the whole analysis with "Insufficient observations for the number of predictors" without saying which group. The message itself now names an all-missing variable ("x has no non-missing values") or gives the case count.
• Model specification in linear_regression() / logistic_regression():
  • dependent = / predictors = work with non-syntactic names such as `my var` or `Zufriedenheit (0-10)` (they were pasted into the formula without backticks: parse error).
  • linear_regression(data, log(income) ~ age) fits the transformed outcome instead of failing with "Variable(s) not found in data: log."; logistic_regression() says the outcome must be a single variable.
  • y ~ . uses all other columns except the weights and grouping variables; y ~ 1 and an outcome that is also a predictor stop with a clear message.
  • A predictor selection that also picks the outcome, the weights or a grouping variable (e.g. predictors = where(is.numeric)) drops them with a message instead of regressing the outcome on itself.
  • Character predictors enter as factors (no more NA descriptives and base-R warnings).
  • weights = accepts an expression such as sampling_weight * 2 (it failed with "Can't convert a call to a string").
  • anova() on a weighted logistic model no longer leaks "non-integer #successes" warnings.
• update() and step() work on linear_regression() and logistic_regression() results. The objects stored lm's internal call (data = data_complete, weights = .wt), so both failed with "object 'data_complete' not found". The stored call is now the user's own call in formula form. A model fitted inside a %>% pipe (no data name to re-use) gets a clear error instead. coef(), residuals(), fitted(), confint(), nobs() and vcov() on grouped results (and all but coef() on pairwise results) now stop with an informative message instead of returning NULL or a base-R error; coef() of a pairwise regression returns its coefficients. nobs() of a weighted logistic model is the sum of the weights (SPSS N).
• Weighted linear_regression(use = "pairwise") uses the unrounded smallest pairwise sum of weights in its degrees of freedom, sums of squares and standard errors (Validation Charter §5.1); it was rounded first. The displayed N stays rounded; results move by a fraction of the rounding error.
• Regression output (linear_regression(), logistic_regression(), marginal_effects()) is readable for every scale of variable:
  • summary(digits = ) is honoured in all tables (it was ignored).
  • Estimates that would round to zero are shown in scientific notation instead of 0.000 / -0.000 (e.g. income B = 3.72e-04, as SPSS shows 3.72E-4); huge values such as separation estimates no longer print as 30-digit numbers. Odds ratios get as many decimals as needed to separate their confidence limits (1.0007 [1.0006, 1.0008] instead of 1.001 [1.001, 1.001]). Fit statistics never show -0.000.
  • Term names are printed in full; the column is sized to the longest term (dummy names were cut to 20/25 characters, e.g. "educationIntermediate S...").
  • Sig. columns use the SPSS style (<.001, .466) instead of 0.000.
  • The CI columns name their level ("95% CI Lower").
  • A very large weighted N (sum of weights above 2^31) no longer breaks the output with "invalid format '%d'".
• Labelled predictors from SPSS files enter linear_regression() and logistic_regression() with their numeric codes (as in SPSS; factors = "dummy" applies to factors only). summary() now says so and points to to_label() for dummy coding. The Descriptive Statistics table shows one row per dummy (the share of each category) for factor predictors instead of the meaningless mean of the level index.
• Correlation output (pearson_cor(), spearman_rho(), kendall_tau(), partial_cor()):
  • The compact print of a matrix shows the N range over the pairs (N = 2008-2421); it showed the N of the first pair only (even N = 0). With more than 15 pairs it lists the strongest significant pairs instead of every pair (45 lines for 10 variables).
  • pearson_cor() labels the interval by conf.level ("90% CI"; every interval was called "95% CI", the column CI_95). A one-sided test (alternative = "less"/"greater") is named in print and summary and gets the matching one-sided interval, as cor.test().
  • Matrices follow SPSS: blank p-value diagonal (was 0.0000), p in table style (<.001 instead of 0.0000), significance flags on the coefficients, digits honoured for any number of variables (it dropped to 2 decimals above 6 variables), and columns split into blocks that fit the console instead of widening the width option.
  • A constant variable gives one warning naming it and "not computed (no variance)" instead of "r = NA, p = NA ," and one base-R warning per pair; partial_cor() with a constant control variable no longer crashes with "missing value where TRUE/FALSE needed".
  • Aligned pair labels and a pairwise table that no longer wraps; a very large weighted N (above 2^31) no longer breaks the output.
• logistic_regression() also finds the value labels of an outcome that carries them as a plain labels attribute (e.g. rec() output) for the encoding table and the [P(y = category)] tag, and both regression functions fit labelled (SPSS) variables on their bare numeric codes, so a fit no longer depends on haven's arithmetic methods being loaded ("<haven_labelled> - <haven_labelled> is not permitted").

Scale analysis

• efa() components and factors now carry SPSS's sign convention. The eigenvectors kept the arbitrary sign eigen() returned, so three positively correlated trust items could load -0.597/-0.475/-0.678 on their single component and signs could flip between groups. Every extracted column is now reflected to a positive loading sum (as SPSS FACTOR does); rotated, pattern and structure matrices and factor correlations follow from the reflected solution. This reproduces every signed loading SPSS prints in the reference runs, now asserted in the SPSS validation tests.
• efa(extraction = "ml") no longer reports the PCA share of variance. The compact line read "Variance explained: 61.0%" (the eigenvalue share of three components) although the three ML factors explain about 24%, and it called ML factors "components". The result gains $extraction_variance (SPSS's "Extraction Sums of Squared Loadings"); the compact line reports its cumulative percentage and says "factors" for ML. summary() prints "Total Variance Explained" as one aligned SPSS-style table (initial eigenvalues, extraction sums, rotation sums) instead of free-text lines whose columns shifted from the 10th component on.
• efa() no longer crashes with the base error "infinite or missing values in 'x'" (German: "unendliche oder fehlende Werte in 'x'") when a correlation cannot be computed. A constant item, an item without valid values, two items without cases in common, or too few complete cases now give an error that names the item(s) and the reason. Under group_by(), such a group is skipped with a warning naming the group, and every other group is still analysed (previously the whole grouped result was lost); print()/summary() show "not computed (...)" for it.
• efa() flags singular correlation matrices like SPSS ("not positive definite"). A duplicated item used to yield KMO 0.500 (from a pseudo-inverse), Bartlett's chi-square Inf and "Sig.: 0.000", and fewer cases than variables gave KMO NaN. Now a warning names the perfectly correlated items or the case shortage, KMO and Bartlett's test are reported as not computed, and ML extraction stops with a clear message instead of "Lapack routine dgesv: system is exactly singular". A separate warning appears when there are no more cases than variables. Bartlett's and the goodness-of-fit significance use the SPSS style ("<.001") instead of "0.000".
• efa() shows its sample size. print() adds "N = 2168 (smallest pairwise)" (or "(listwise)"), summary() an N line and SPSS's "Descriptive Statistics" table (mean, SD, analysis N, missing N; new toggle descriptives). With use = "complete", $item_statistics now describes the complete cases the analysis uses; the analysis N was pairwise before.
• efa() input and output details: n_factors = 2.7 is an error instead of being truncated to 2; use = "listwise" (the SPSS term) is accepted as an alias of "complete", and invalid rotation/extraction/use values give an English error instead of a translated match.arg() message. A requested rotation of a single component is no longer dropped silently: output says "Only one component was extracted. The solution cannot be rotated." as SPSS does. Communalities print with fixed decimals ("1.000" instead of "1" next to "0.457"), and the summary says "N of Components" for PCA.
• reliability(na.rm = FALSE) no longer crashes with "missing value where TRUE/FALSE needed" (German: "Fehlender Wert, wo TRUE/FALSE nötig ist") as soon as a value is missing. All statistics are NA, a warning names the items with missing values and points to na.rm = TRUE (listwise deletion, as SPSS), and print()/summary() say "not computed (...)" instead of "Cronbach's Alpha = NA ()".
• reliability() warnings are clearer. An item with zero variance is removed from the scale with a warning naming it, as SPSS RELIABILITY does; it used to stay in (alpha 0.042 instead of 0.047, standardized alpha NA) next to the German base warnings "Standardabweichung ist Null" and "NaNs wurden erzeugt". When omega cannot be computed (e.g. a duplicated item makes the correlation matrix singular), the warning says why in English and names the perfectly correlated items instead of relaying factanal's translated error. The "omega requires at least 3 items" warning appears once per call instead of once per group, and "Insufficient data (n = 0)" now names the group and the items without valid values.
• reliability() no longer reports McDonald's omega from a Heywood solution. For the trust items in the East region the one-factor model put one uniqueness at its lower bound (loading of about 1), and omega 0.349 was printed next to alpha 0.037 without comment. Omega is now NA in such cases, with a warning that names the item and the group.
• reliability() flags a negative Cronbach's alpha like SPSS: "The value is negative due to a negative average covariance among items ... check item codings." A warning and the summary() footnote name the items with a negative corrected item-total correlation (usually items that need reverse-coding; new $negative_items), and the compact print says "negative; check item coding" instead of classifying alpha -0.929 as "Poor".
• reliability() output is easier to read. The Item-Total Statistics table no longer wraps at 80 columns under snake_case headers (scale_mean_deleted, corrected_r, ...); it has SPSS-style two-line headers ("Scale Mean / if Deleted", "Alpha if / Deleted", ...). The inter-item correlation matrix honours digits for more than six items (it was forced to 2 decimals) and uses numbered columns so it stays narrow. Item statistics print with fixed decimals (no more "1.16" next to "2.615"), and a missing omega reads "not computed" with the reason instead of "NA".
• reliability() and efa() show variable labels, as SPSS does. summary() lists every item with its full label, and the per-item tables (item statistics, communalities, loading matrices) add the label next to the name, shortened with "..." so that rows fit the console width; wide tables keep the names only. Labels with umlauts stay aligned. The labels are stored in $variable_labels.

Data import/export, labels and transformation

• read_sas() reads .sas7bdat files again with the default arguments. It passed catalog_encoding = NULL explicitly, which haven 2.5 rejects ("Expected string vector of length 1"), so every call without an explicit catalog encoding failed. The catalog encoding now falls back to the data file's encoding, as documented.
• untag_na(), na_frequencies(), write_spss(), write_xlsx() and write_xpt() are much faster on large imported files. The NA tag of every missing value was read one element at a time, each through vctrs dispatch, and the Excel "Labels" sheet was built from thousands of one-row data frames. Full ALLBUS 2023 (579 variables): untag_na() over all columns 9.8 s -> 0.15 s, write_spss() 11.5 s -> 1 s, write_xlsx() 22 s -> 8 s (the rest is openxlsx2 itself). Output is unchanged.
• write_spss() no longer crashes ("Failed to insert value ...: The file format does not supported character tags for missing values") after rec(), std(), center(), pomps(), to_numeric() or arithmetic on variables imported with read_spss(). These results kept the tagged-NA payloads but lost the code map. Now rec() keeps the missing-value types, their code map and their labels (so na_frequencies(), frequency() and the SPSS export still show "no answer" etc.), while std(), center(), pomps() and to_numeric() return plain NA by design (as an SPSS COMPUTE gives system-missing). write_spss() writes remaining unmapped tags as system missing with one warning naming the variables.
• Value labels created by rec() (inline [label] syntax, val_labels, mirrored labels of "rev"), kept by strip_tags() or by to_numeric(keep_labels = TRUE) now survive write_spss() and write_stata(). They were attached as a bare labels attribute without the haven_labelled class, which haven's writers ignore. The results are haven_labelled now; to_labelled() picks up an existing labels attribute; the exporters also promote such bare attributes themselves.
• rec(as_factor = TRUE) names the levels by the result's value labels (e.g. the mirrored labels of "rev", previously ignored: levels "1".."7") in code order. Values without a label keep their code as level name instead of becoming NA, and duplicate label texts are disambiguated by their code.
• rec(rules = "rev") reverses on the scale range instead of the observed range. It computed max(x) + min(x) - x over the data, so a 1-5 item answered only with 2-5 became 5..2 instead of 4..1, and the value labels were mirrored to codes that do not exist. The range now comes from the value labels of the valid codes (together with the observed values); without labels the observed range is used with a message, and the new syntax rules = "rev(1, 5)" sets the range explicitly (values outside it are reported).
• rec() on haven_labelled_spss vectors (haven::read_sav(user_na = TRUE)) treats the user-missing codes as missing: they were reversed or recoded like valid values and lost their codes and labels. The input is converted to the tagged-NA form of read_spss() first. val_labels is honoured with rules = "rev" (it was silently ignored).
• rec() syntax is more forgiving and never loses values silently: valid values that match no rule still become NA but now with a warning that lists them and suggests else=copy (SPSS's in-place RECODE keeps them) or else=NA; value lists work ("1,2=1; 3,4:5=2", as SPSS's RECODE (1,2=1)); keywords are case-insensitive ("REV", "Dicho(3)"); a reversed range such as "5:1=1" is an error instead of silently matching nothing; inline labels may contain semicolons ("[niedrig; gering]"); "dicho(x)" and a missing rules argument give clear English errors instead of leaked base-R (German) messages; the " (recoded)" label suffix is no longer appended again on every call.
• to_label() and to_character() no longer merge distinct codes that share a label text. ALLBUS labels the scale points 2-6 of 88 items "..", so e.g. pt12 collapsed into three levels ("GAR KEIN VERTRAUEN", "..", "GROSSES VERTRAUEN") and the to_numeric() round trip turned 3-6 into 2. Duplicate texts now get their code appended (".. (2)", ".. (3)"), the same rule the test functions use for grouping variables.
• to_label(data) and to_character(data) without a variable selection no longer turn metric variables into factors. ALLBUS age and isei08 (whose only labels are missing codes) became all-NA factors and the weight a factor with one level per value, silently. Now only variables whose values are all value-labelled are converted; the others are left unchanged with one message listing them. Selected variables are still converted, and whenever values without a label become NA a warning names the variables and points to add_non_labelled = TRUE.
• copy_labels() no longer forces the source's value labels and class onto columns that were converted or summarised: a to_label() factor became int+lbl 1, 2, 3 carrying the labels 1/5/9 of the source codes, and group means were labelled as if they were codes. Value labels, missing-value metadata and the class are now copied only when the target still holds the source's codes; the variable label is always copied.
• set_na(data, -9, -8, tag = FALSE) (the documented example) no longer strips the labels of every numeric column (survey_data: 15 variable labels -> 6). Variable labels, the value labels of the remaining codes, the haven_labelled class and existing missing-value types are kept; integer columns stay integer.
• val_labels() refuses to set value labels on a factor or character column (they were stored but never used) and points to to_labelled(); set_na() warns when a named variable or a vector is a factor instead of silently returning it unchanged.
• to_dummy() fixes: with suffix = "label", values sharing a label text (ALLBUS ".." scale points) all wrote into one column pt12_ and their dummies were lost - every category now gets its own column (value appended to duplicate or empty labels); umlauts are transliterated ("männlich" -> maennlich, was mnnlich); ref works for factors, by level name or number (it was compared with the level names only, so ref = 1 silently returned all dummies), and a ref that matches no category is an error.
• row_count() counts value sets and missing values: count = c(4, 5) was recycled over the cells (wrong counts without a warning), and count = NA always returned 0. It now counts cells equal to any listed value (SPSS COUNT n = v1 TO v5 (4, 5)), NA counts missing values (SPSS MISSING), and SPSS missing codes of imported data (e.g. -9, a tagged NA after read_spss()) are counted when listed - they were always 0.
• row_means(), row_sums() and row_count() inside a grouped mutate() with the . placeholder now stop with an explanation and the pick() form instead of dplyr's bare size-mismatch error; pick() is the documented, recommended form. Non-numeric columns handed over by pick() are ignored with a warning naming them (they were dropped silently), and a fractional min_valid (e.g. 2.5) is rejected.
• pomps() warns when values lie outside scale_min-scale_max (an unrecoded "don't know" = 9 on a 1-5 scale silently scored 200), checks that scale_min/scale_max are single finite numbers (a vector gave the German base error "Bedingung hat Länge > 1", NA a cryptic one), and says clearly when an all-NA input leaves no range to derive.
• std() and center() keep the variable label of imported (SPSS) variables when overwriting them in place: the label was read after the column had already been replaced, so it was lost. Grouped standardization/centering returns a plain numeric column instead of leaving a dbl+lbl vector behind, vector input keeps its label (with " (standardized)"/" (centered)"), and the zero-spread warning names the variable and the group (e.g. "x (g = a): the spread (sd) is zero"). The weighted mean/SD now come from the shared SPSS kernels (results unchanged).
• write_spss() and write_stata() export a factor created by to_label() with its original codes and value labels: haven renumbered it 1..k (ALLBUS dm06 codes 100, 120, ... became 1, 2, ...), although the factor carries its codes. Other factors are still written as 1..k with their levels as labels.
• read_spss() -> write_spss() round trips the SPSS missing-value definitions exactly and quietly. read_spss() kept only the codes that occur, so an unchanged ALLBUS export produced 133 warnings ("4 discrete missing codes exceed SPSS's limit of 3 ... range -42--8") and rewrote LOWEST THRU -1 as -42 THRU -8. The original definition is now remembered (attribute spss_missing) and written back while it fits (ALLBUS 2023: all 579 definitions and values identical, no warning). Variables with more than 3 codes otherwise use SPSS's "range plus one discrete value" form when that avoids valid values (e.g. codes 0, 7, 8, 9 around valid 1-6, which used to be an error), reported in one message with readable ranges ("-11 to -8 and -42").
• read_spss(), read_por(), read_stata(), read_sas(), read_xpt() and read_xlsx() recognise a file of the wrong type from its first bytes and say what it looks like and which reader to use (e.g. "read_spss() cannot read x.dta: it looks like a Stata file (.dta). Use read_stata() instead."), instead of readstat's cryptic errors; a missing file is reported as such.
• write_xpt() no longer truncates variable names silently: the default SAS transport version 5 allows 8 characters, and truncation even created duplicate names. It now warns (listing old -> new names, suggesting version = 8) and refuses to write when truncation would produce duplicates.
• write_xlsx() makes sheet names unique within Excel's 31-character, case-insensitive limit (two list names sharing their first 31 characters, or an element called "Labels", crashed without writing a file); renamed sheets are listed in one message. A missing output directory is reported clearly, as in codebook(file = ).
• write_xlsx() of a grouped frequency() result writes one block per group, headed by the variable and the group ("gender (Gender) - region = East") with its own N line. All groups were written as one block under the first group's header, without group labels, and the Total row summed the groups (Raw % = 200). Weighted N is rounded like the console print ("N=2516", was "N=5245.99999999998").
• find_var() gains fixed = TRUE for literal (case-insensitive) search, e.g. label text with parentheses: "BEFRAGTE(R)" as a regular expression matched "BEFRAGTER" instead. Regular expressions stay the default; a message points to fixed = TRUE when the literal text would match other variables, an invalid regular expression such as "(" is searched as text (the regex engine's warning no longer leaks), and an empty result is returned invisibly instead of printing <0 rows>.
• strip_tags() and untag_na() accept a data frame (all numeric columns, or the ones selected via ...) instead of failing with a German base-R error, and reject non-numeric vectors with a clear message (strip_tags("a") returned NA). untag_na() keeps the value and variable labels: the labels of the missing types are attached to their restored codes (e.g. -9 = "KEINE ANGABE"), so the result is still haven_labelled.
• na_frequencies() output is reorganised: rows are ordered by code like the missing block of frequency() (they were sorted by count), columns are code, label, n, prc (percent of all cases, new) and tag (the technical tag letter moved last), SPSS codes are numeric (were character), the "(System Missing)" row appears only when system-missing values occur, and a variable without missing values gives a message instead of printing <0 rows>. Data frames are accepted (na_frequencies(data, q1, q2), as the data-io vignette shows) and return one table with a variable column.
• New replacement form var_label(x) <- "Label" (and var_label(data) <- list(age = "Age", sex = "Sex"); NULL removes a label). drop_labels() also drops unused factor levels (keeping the variable label and the codes of to_label() factors); its example filtered on a non-existent category and did nothing. The copy_labels() help no longer claims that filter()/select()/mutate() strip labels (they keep them) and names the operations that do.

Across functions

• A group_by() group in which no case has a valid weight has no cases, as SPSS excludes cases with a missing weight: describe() computed that group unweighted (printed under "Weighted Descriptive Statistics" with N = 0 and one warning per statistic), frequency() failed with "Ersetzung hat 1 Zeile, Daten haben 0" and a grouped crosstab() aborted for all groups. The group now shows N = 0 (describe()) or is left out with a warning naming it (frequency(), crosstab()).

• The weights argument of every analysis function documents the forms it accepts since this release: a column name (unquoted or as a string), an expression such as sampling_weight * 2, or a numeric vector with one weight per row.

• Selecting a grouping variable of group_by() data (explicitly or through a helper such as where(is.numeric)) no longer analyses it within its own groups. pearson_cor(), spearman_rho() and kendall_tau() warned "No variance" in every group and returned NA rows, reliability() used the constant grouping variable as an item, t_test() warned "constant value" per group. As dplyr::across() does, every analysis function now leaves grouping variables out of the selection with a message (previously only describe() and the w_* functions did); rec(), to_dummy() and the row operations, which do not compute per group, still accept them.

• A misspelled argument name is now an error that names the intended argument: "Unknown argument weight of w_mean(). Did you mean weights?". In functions whose ... selects variables, weight = (for weights), na_rm =, groups =, conf = ... were taken as a tidyselect rename: w_mean(data, age, weight = w) returned an unweighted mean plus a bogus "weight" row, binomial_test() a garbage row, and describe(), frequency(), chi_square(), kruskal_wallis() failed with unrelated messages. Inside summarise() the w_* functions ignored ... entirely (w_mean(age, weight = w) was silently unweighted). Renaming selections (describe(data, Age = age)) are refused with an explanation. For consistency, functions whose arguments R would partially match (crosstab(), fisher_test(), mcnemar_test(), wilcoxon_test(), factorial_anova(), ancova(), the regressions, tukey_test(), scheffe_test(), marginal_effects()) no longer accept abbreviated names such as weight = or percent =: the same typo now gives the same error everywhere. fisher_test() and mcnemar_test() no longer ignore unknown arguments.

• weights accepts the same forms in every function: a bare column name, a column name as string ("sampling_weight", all_of(w), !!w), an expression evaluated in the data (sampling_weight * 2, survey_data$sampling_weight), or a numeric vector with one value per row. Outside the regressions only the bare column name worked: weights = survey_data$sampling_weight and weights = all_of(w) failed with "Can't convert a call to a string.", weights = 1 with "Can't convert a double vector to a string". An expression is shown by its text ("Weights: sampling_weight * 2"). weights = w with w holding a column name now points to all_of(w); a vector of the wrong length or an unknown variable inside an expression is named in the error. The w_* functions no longer accept a factor as weights (its level codes were used silently), and in summarise() a non-numeric weights is an error instead of a silent unweighted result. std() and center() return the weights column unchanged (it lost its attributes).

• Labelled data (haven_labelled) work when the haven package is not loaded, e.g. data restored with readRDS() in a fresh session. haven is only suggested, and without its namespace the labelled vectors have no methods for comparison and arithmetic: frequency() failed with "Can't convert x <haven_labelled> to ", describe() and w_mean() with "<haven_labelled_spss> * is not permitted", crosstab(), codebook(), unlabel(), drop_labels(), rec() and write_xlsx() likewise. The entry helpers now load haven's namespace when a selected data set or vector is labelled (and say that haven is needed when it is not installed).

• Calling a function without a required argument gives a clear error naming it ("Argument covariate is missing, with no default."), in every exported function. Previously base-R errors surfaced from inside, often localized and naming internal arguments: 'Argument "data" fehlt (ohne Standardwert)', oneway_anova(data, age) "Can't extract column with g_name", ancova()/factorial_anova() 'Argument "between_expr" fehlt', fisher_test()/mcnemar_test() 'Argument "x" fehlt', crosstab()/write_xlsx() "no applicable method ... class NULL". A bare variable name where a value is expected (linear_regression(data, age), find_var(data, age), a path) is named instead of "object 'age' not found"; set_na(data, age) explains that it takes values, not variable names.

• Weights without any positive value (all zero or missing) are refused with one clear error. The analyses failed deep inside with base-R errors ("'n' must be a positive integer", "missing value where TRUE/FALSE needed", "object 'fit' not found"), and describe() silently fell back to an unweighted analysis when all weights were missing.

• group = given as an expression (group = region == "East") explains that the grouping variable must be created first (it failed with "object 'region' not found"); a group selecting no or several columns is an error naming the selection. spearman_rho() and kendall_tau() accept ordered factors (ranked by level order) like mann_whitney() and kruskal_wallis(); they rejected e.g. education as "not numeric".

• Result tables stay aligned with umlauts and other multi-byte characters in variable names, value labels and terms (Tukey/Scheffe, Dunn, pairwise correlations, regression coefficients, Kruskal-Wallis ranks, goodness-of-fit and many more): the shared table printer padded cells with sprintf("%-20s"), which counts bytes, so every umlaut shifted the rest of its row one column to the left. Cells are now padded by display width. Counts and sums of weights of 2^31 or more (expansion weights) print as whole numbers instead of NA with a coercion warning; leading label columns such as "Group 2" are left-aligned like "Group 1"; table lines no longer end in a blank.

• Sums of weights of 2^31 or more (expansion weights) no longer print as N = NA in the compact lines of the rank tests and the goodness-of-fit test, as NA in the N column of the one-sample t-test, and no longer raise integer-coercion warnings in the t-test, one-way ANOVA, factorial ANOVA/ANCOVA, Levene and reliability tables: every count is formatted as a whole number without integer coercion.

• Grouped output uses one group-header style in every verbose table: describe(), the w_*() functions and the summaries of levene_test(), normality_test(), marginal_effects() and the post-hoc tests printed "Group: region = East " with a trailing blank and no underline, all other summaries an underlined header. Section titles without a suffix ("Pearson Correlation", "Chi-Squared Test of Independence", "Levene's Test for Homogeneity of Variance") no longer end in a blank with an underline one dash too long.

• Counts carry no thousands separators anywhere, as in SPSS tables and the console: the HTML codebook() header wrote "2,500 observations", mann_whitney()'s compact line "U = 776,732" (its summary table 776732). The N/discordant-pair lines of fisher_test(), mcnemar_test(), chi_square() and the Friedman note of pairwise_wilcoxon() use the same whole-number formatter (the latter printed "N = 9.36e+09" for large sums of weights).

• Weighted frequency() of a factor together with a numeric variable (frequency(survey_data, education, life_satisfaction, weights = sampling_weight)) no longer turns the numeric variable's categories into NA (listed under "Total missing") with the base-R warning "invalid factor level, NA generated": the weighted branch kept the factor as the value column when the tables were combined.

• Variable pairs are joined by an ASCII "x" in every output: the titles of chi_square(), fisher_test(), mcnemar_test() and crosstab() and the chi-square "Table size" line used the multiplication sign (U+00D7), the correlation, factorial and post-hoc output an "x".

• print() of linear_regression(), logistic_regression() and normality_test() results accepts digits like every other compact print (R-squared, statistics and p-values ignored it). The grouped normality_test() compact print shows the test results under a [group] line for every group instead of only the number of groups.

• Every compact print() of a test or model result now ends with "Use summary() for detailed output." (t_test(), oneway_anova(), factorial_anova(), ancova(), the three correlation functions, both regressions, reliability() and efa() did not show it, the rank tests, chi-square family, Levene and normality tests did).

• Grouped compact prints of linear_regression() and logistic_regression() put each group on its own [region = East] line (they printed " region = East: R2 = ..."), like every other compact print; a skipped group shows "not computed (reason)" under its [group] line.

• Analysis functions called on a data frame without rows stop with one clear message ("data has no rows") instead of base-R errors from deep inside (frequency(), reliability(), efa()) or seven repeated weight warnings (describe()). Transformations such as rec() still pass empty data through.

Result export and R Markdown

• dunn_test() and pairwise_wilcoxon() store their comparison table as $results, like every other result class (it lived only in $comparisons, so x$results was NULL; $comparisons is kept). A weights-invariance regression test that compared these NULLs with each other now checks the real comparisons.
• Every analysis result can be turned into a plain table: as.data.frame() (and tibble::as_tibble()) now work for all result classes instead of failing with "cannot coerce class ... to a data.frame". The table has one row per test/variable/group (per pair for correlations and post-hoc tests, per term for ANOVA and regression tables, per cell for crosstab(), per item for efa() loadings); list columns are flattened (t_test() gets group1, group2, mean1, mean2, sd1, sd2), written out as text (the values and value labels of a codebook()) or dropped (the observed/expected tables of chi_square()), so write.csv() works; grouping variables are the leading columns, labelled ones as their value labels. With broom loaded, broom::tidy() now works for all test and descriptive classes (it covered only the two regressions), using broom's column names (statistic, p.value, parameter/num.df/den.df, estimate, conf.low, conf.high, adj.p.value, method).
• write_xlsx() exports analysis results. crosstab() results are written in the SPSS table layout (Count and the requested "% within" / "% of Total" rows per category, Total row and column; one block per group), every other result (e.g. describe(), t_test(), oneway_anova(), reliability(), linear_regression()) as its result table followed by the secondary tables (group descriptives, mean ranks, item statistics, model summary, ...). A named list may now mix data frames with any result (list(Descriptives = describe(...), Data = df) was rejected), and unsupported objects get a clear error instead of "no applicable method for 'write_xlsx'".
• codebook() works in R Markdown and Quarto. It returned invisibly and (with view = interactive() being FALSE while knitting) a chunk codebook(data) produced nothing. It now returns visibly, and a knit_print() method embeds the HTML codebook in HTML output (styles scoped to the codebook, so the rest of the document keeps its look); PDF/Word output shows the console overview. While knitting, the viewer is no longer opened by default (rmarkdown::render() from an interactive session used to open it). In the console, the compact overview is now printed after the viewer opens.

SPSS parity: ANCOVA Levene and factor rotations

• ancova(): Levene's test of equality of error variances now matches SPSS UNIANOVA. It was computed on the deviations of the dependent variable from its raw cell means, ignoring the covariates (life_satisfaction BY gender WITH age: F = 1.277, p = .258 instead of SPSS 1.306, p = .253). SPSS tests the absolute residuals of the full model (covariates and factors) across the design cells; mariposa now does the same and reproduces all six unweighted Levene tests of the SPSS reference run. The weighted path is unchanged until the pending weighted reference run. New levene_test() method for ancova() results (ancova(...) |> levene_test(), also per group under group_by()).
• efa(rotation = "promax") now matches SPSS FACTOR /ROTATION PROMAX. mariposa used stats::promax(), which builds the promax target from the raw varimax loadings; SPSS first normalizes every row (Kaiser normalization). Pattern loadings differed by up to .03 and the component correlations were .055 / .155 instead of SPSS -.002 / -.012. The new implementation follows the SPSS algorithm and reproduces every pattern, structure and correlation matrix of the SPSS reference runs (unweighted, weighted, per region), now asserted in the validation suite.
• efa() with an oblique rotation (promax, oblimin): the "Rotation Sums of Squared Loadings" are now the sums of squares of the structure matrix, as SPSS reports them (they were taken from the pattern matrix: 1.604 / 1.065 / 1.045 instead of SPSS 1.599 / 1.039 / 1.021).
• efa() varimax and oblimin rotations now use SPSS FACTOR's own algorithms and stopping rules (Kaiser's cyclic pairwise varimax; the Jennrich-Sampson direct oblimin, delta = 0; at most 25 iterations). stats::varimax() stopped earlier and missed SPSS's component transformation matrix by up to .004 (2-factor solution: 26.653 % instead of SPSS 26.655 % for the first rotated component); GPArotation::oblimin() iterated past SPSS's stopping point (weighted solution: pattern loadings off by up to .002). The rotated factors are reflected to a positive sum and ordered by their sums of squares as in SPSS. Every varimax, oblimin and promax reference solution (unweighted, weighted, per region) now matches SPSS, including SPSS's iteration counts, which summary() prints as SPSS does ("Rotation converged in 4 iterations."). Oblimin no longer needs the GPArotation package.
• efa(extraction = "ml"): a Heywood variable is now bounded at a communality of .999, as SPSS FACTOR does (it was .995, the default bound of stats::factanal()), and the factors keep SPSS's order (by the eigenvalues of the rescaled correlation matrix; factanal() re-sorted them by sums of squares, so the unrotated factor matrix and the extraction sums came in a different order). Where SPSS's ML iteration converges (reference runs per region, West), the whole solution now matches SPSS: communalities, factor and rotated factor matrices, extraction and rotation sums, transformation matrix. The remaining reference runs end without convergence in SPSS itself (a model with 0 degrees of freedom and Heywood cases); see Validation.
• levene_test() compact print: a whole-number weighted df2 above 2^31 (sums of weights in the billions) printed as F(1, NA) with an integer overflow warning; it is now printed in full.
• Weighted ancova(): Levene's test uses sqrt(w) * |WLS residual| of the full model, as SPSS UNIANOVA does with /REGWGT. This reproduces all five weighted SPSS references exactly (e.g. 0.902 instead of 0.880) and restores the rule that weights = 1 gives the unweighted result (the residual-based unweighted test made the old cell-mean-based weighted test disagree with it).

SPSS parity audit: reference runs no test asserted

A value-by-value comparison of every SPSS reference output with the current code found values that had been in the reference files all along but were never asserted. These are now fixed and asserted:

• Weighted tukey_test()/scheffe_test(): the error mean square and its df come from the weighted ANOVA table (df = floor(sum of weights) - k, as SPSS ONEWAY and oneway_anova() use). The pooled df of the group statistics moved SE, confidence limits and Sig. off SPSS (e.g. SE 62.521 instead of 62.534, Sig. .089 instead of .090); about 550 printed values in the reference runs.
• tukey_test() confidence limits use the studentized range quantile to full precision. qtukey() is accurate to about 1e-8, which moved 4-decimal limits of income differences (214.4432 instead of SPSS's 214.4433).
• Phi of a 2x2 table (chi_square(), phi()) carries the sign of the association, as SPSS's Symmetric Measures print it: -.082 instead of .082 for a negative association (reference Fisher Test 1c). Larger tables keep the unsigned sqrt(chi-square / N).
• wilcoxon_test() and pairwise_wilcoxon() report Z as SPSS does: from the smaller rank sum, so it is never positive, with a new z_based_on column ("negative ranks"/"positive ranks") that names that sum like SPSS's footnote; summary() prints it. Before, Z was positive whenever increases dominated (trust_science - trust_government: +25.945 instead of SPSS's -25.945); the tests compared |Z| only. An empty rank category has mean rank 0 (SPSS: .00) instead of NA.
• Weighted two-sided mann_whitney() reports Z with SPSS's sign (from the smaller U, never positive) like the unweighted test; it returned the directional statistic (+1.020 where SPSS prints -.989).
• Weighted factorial_anova()/ancova() cell standard deviations match SPSS's /REGWGT Descriptive Statistics (weighted sum of squares over n - 1): they divided by the sum of weights (1.207 instead of 1.238).
• Printed numbers round halves up as SPSS does: a mean rank of 306/16 = 19.125 printed as 19.12 (round half to even) where SPSS shows 19.13, and a value that rounds to zero no longer prints as "-0.000".

Validation

• New Tier-3 exceptions for efa() ML extraction: EXC-001 (±.002; SPSS stops its Newton-Raphson iteration at ECONVERGE(.001), so loadings, sums of squares and percentages of variance of converged solutions agree to about 1e-3) and EXC-002 (±.05; SPSS reference runs 5a/6a that end without convergence, "More than 25 iterations required").
• efa() rotations are now validated in full against SPSS: 8 varimax, 6 oblimin and 6 promax solutions (transformation, pattern, structure and correlation matrices, rotation sums, iteration counts); the ML initial communalities of all 6 ML reference runs; ancova() Levene tests of the 6 unweighted reference runs.
• The Display-tier p-value tolerance of assert_spss() is one unit of the last printed decimal. A relative 1% was added on top, which let a p of .975 drift by ten units; every reference p matches at the tighter bound.
• The SPSS validation tests now compare about 14,600 SPSS reference values (every value in the reference outputs that mariposa computes, except the weighted runs pending a WEIGHT BY reference): a value-by-value audit had found most of them matching but never asserted, and the few mismatches hiding there (see "SPSS parity audit" above).
• vignette("spss-compatibility") is regenerated: it lists EXC-001/EXC-002 with their reason (it still said "No active Tier-3 exceptions") and counts the SPSS values each validation file compares when it runs (it counted assert_spss() call sites, which data-driven tests understate).

mariposa 0.7.3 (2026-09-28)

CRAN resubmission (theme: make the print/cat console contract lexically visible — the 2026-09 second-round remark on R/kendall_tau.R).

CRAN

• The reviewer-flagged cat() in R/kendall_tau.R sat in .kendall_tau_spec$pair_extras, a display callback that only ever ran inside the summary() print layer — functionally exempt, but lexically indistinguishable from computation code. All three correlation pair_extras callbacks (kendall, pearson, spearman) now return formatted lines; the single cat() lives in the engine's .print_cor_verbose(). Output is byte-identical.
• The same lexical rule is now enforced package-wide: format_stat_table() was renamed to print_stat_table() (it prints, it never returned a formatted string), and for_each_group()'s group header line moved into a new print_group_label() helper. After this, every cat()/print() call in R/ lives in a function whose name starts with print/.print.
• New static meta-test test-console-discipline.R locks the rule in: it parses every file in R/ and fails if a cat(), print(), or writeLines() call appears in any top-level object not named print*/.print* (calls inside capture.output() are exempt as silent). Together with the runtime test-silent-computation.R (zero stdout from every analysis entry point), this makes the CRAN console contract regression-proof.

mariposa 0.7.2

CRAN resubmission (theme: address all four points of the 2026-09 manual review).

CRAN

• DESCRIPTION now cites the published method references in CRAN's auto-link form: IBM SPSS Statistics Algorithms, Dallal and Wilkinson (1986, doi 10.1080/00031305.1986.10475419) for the Lilliefors correction, and Haberman (1973, doi 10.2307/2529686) for adjusted standardized residuals.
• All \dontrun{} blocks are gone. The 15 import/export examples are now genuinely executable \donttest{} roundtrips through tempfile() (guarded by requireNamespace() for the Suggests packages); the two formats R cannot produce (.por, .sas7bdat) run behind a file.exists() guard. The unlabel() example runs unconditionally on the bundled data.
• on.exit() now registers the restoration before the option is changed in ancova(), factorial_anova(), and the correlation-matrix print helper, so an interrupt between the two lines cannot leak a changed options() setting.
• New test-silent-computation.R proves the console contract the reviewer asked about: every analysis entry point produces zero stdout (32 assertions); all cat() calls in the package live exclusively in the print()/summary() display layer, and remaining runtime messages go through suppressable message()-based conditions.

Bug fixes

Making the former \dontrun{} examples actually run uncovered two real crashes on integer columns (tagged NAs are NaN payloads in doubles; haven::na_tag() errors on integer input):

• unlabel() crashed on any data frame with integer columns ("x must be a double vector") — including the bundled survey_data's Likert variables. Integer vectors can never carry tagged NAs and now pass through directly.
• write_xpt() crashed on integer columns for the same reason in its tag-uppercasing step. Both fixes carry regression tests.

mariposa 0.7.1

CRAN resubmission (theme: address the incoming-pretest NOTE) plus two small robustness patches.

CRAN

• The kendall_tau() examples now run on a 300-case subset: Kendall's tau is O(n^2) and the full-sample examples exceeded CRAN's 5-second limit on the Debian pretest machine (7.5s). A comment in the example explains the subsetting.

Improvements

• The Imports now declare minimum versions matching the APIs actually used (cli >= 3.0.0, dplyr >= 1.0.0, rlang >= 1.0.0, tidyselect >= 1.1.0, tibble >= 3.0.0, htmltools >= 0.5.0), so installations with stale libraries get an automatic update instead of runtime errors.
• set_na() on a whole data frame without haven installed now aborts with the friendly "Package haven is required for tagged NAs" hint instead of the raw namespace error (the unnamed-values path reached haven::tagged_na() without its guard; the vector and named-pairs paths were already guarded).

mariposa 0.7.0

Assumption checks and model interpretation (feature set: three new functions closing the most common SPSS/Stata gaps for survey researchers).

New features

• New function normality_test() — the SPSS EXAMINE "Tests of Normality" table: Kolmogorov-Smirnov with Lilliefors significance correction (Dallal-Wilkinson 1986 approximation, as in SPSS) and Shapiro-Wilk (computed for 3 <= n <= 5000, matching the SPSS convention). Supports tidyselect and group_by() (SPSS EXAMINE ... BY factor). Deliberately takes no weights argument: neither test has a well-defined fractional-frequency-weight form (documented in the help page).
• New function partial_cor() — SPSS PARTIAL CORR (Stata pcorr): partial correlations of two or more variables controlling for one or more others, with the zero-order correlation alongside for comparison, df = n - 2 - k, two-tailed t-test, listwise deletion, weights (frequency semantics, unrounded sum(w) in df), grouping, and a partial-correlation matrix for three or more variables.
• New function marginal_effects() — average marginal effects for logistic_regression() models (Stata margins, dydx(*)): average probability-scale derivatives for continuous predictors, average discrete changes vs. the reference level for factors, delta-method standard errors with analytic gradients, weights and group_by() support. Calling it on a linear_regression() explains that B already is the marginal effect there.
• New function multiple_response() — SPSS MULT RESPONSE for "check all that apply" questions (dichotomy sets): the frequencies table with both percentage bases (percent of responses, summing to 100%, and percent of cases, summing above it), and via by = the set-by-demographic crosstab with case-based column percentages. SPSS case rule (valid = at least one non-missing indicator), variable labels as option labels, frequency weights with unrounded sums, group_by() support.

All four ship with the full three-layer print/summary output.

Bug fixes

• pkgdown: formulas on the reference pages render again. The w_* help pages embed LaTeX via \eqn{}, but the site configuration loaded no math engine, so pages showed raw LaTeX source (\frac{...}). _pkgdown.yml now sets template: math-rendering: katex.

Validation

• All three functions carry property-based validation (Charter Tier 4 where an SPSS reference run is still pending): normality_test() against stats::ks.test() and the independent nortest::lillie.test() implementation (new Suggests dependency, test-only); partial_cor() against the residual-of-regressions characterization via lm(); multiple_response() against direct hand-computation from the indicator matrix; marginal_effects() against the analytic logit-AME formula and an independent finite-difference delta-method recomputation (SPSS has no AME procedure — permanently Tier 4). Weighted paths have w == 1 invariance blocks; grouped paths are pinned to per-subset recomputation. SPSS v29 reference syntax for all pending runs lives in .claude/spss-syntax-0.7.0-references.sps.

mariposa 0.6.17

Crosstab cell diagnostics (theme: after the chi-square test, show which cells drive the association).

New features

• crosstab() now computes expected cell counts and adjusted standardized residuals (SPSS CROSSTABS /CELLS=EXPECTED ASRESID, Haberman 1973). Display them with summary(result, residuals = TRUE) — a sub-row per cell, 1 decimal as in SPSS, with a footnote explaining the |adj.res.| > 2 rule of thumb. The raw matrices are available as $expected and $adj_residuals. Weighted tables use the unrounded weighted cell counts (Charter §5.1) and reduce exactly to the unweighted result at weights == 1 (invariance suite extended).

Validation

• The residuals are verified against the independent Haberman-formula implementation in chisq.test()$stdres (unweighted) and a hand-derived weighted recomputation; the SPSS v29 /CELLS=ASRESID reference run is pending and the compatibility vignette discloses them as Tier 4 until it lands.

mariposa 0.6.16

CRAN readiness (theme: the package passes CRAN's submission conventions, not just R CMD check). No statistical behavior changes.

CRAN conventions

• DESCRIPTION: software names are single-quoted per CRAN policy ('SPSS', 'Stata', 'SAS', 'Excel', 'tidyverse'); the promotional "Professional" opener is gone.
• Help pages no longer contain Unicode characters that break the CRAN PDF manual build (>=, <=, -> replace their typographic variants) — local and CI checks had always run --no-manual, so this was never exercised.
• \value{} sections added to the five documented S3 method pages that lacked them (predict/anova for both regressions, print.w_quantile).
• The set_na() example is now runnable (haven-guarded) instead of \dontrun; the efa() oblimin example guards its GPArotation dependency so --run-donttest passes without Suggests installed.
• ?mariposa now resolves: the package help topic is generated from DESCRIPTION (previously suppressed by @noRd).

Test infrastructure

• SPSS-validation assertions (assert_spss() and everything built on it) skip on CRAN. They are golden-number comparisons against SPSS v29 references and remain the release gate in CI; skipping them on CRAN removes the false-positive risk from BLAS/platform numeric variation. Unit, property, invariance, and print/summary tests still run on CRAN.

mariposa 0.6.15

Regression correctness (theme: the two regression functions compute what they claim under weights and degenerate inputs, and say what they show).

Bug fixes

• Weighted logistic_regression(): -2LL, omnibus chi-square, and Cox & Snell / Nagelkerke R-squared were wrong under fractional weights. The -2 log-likelihood came from stats::logLik(), whose binomial method rounds prior weights internally (dbinom(round(m*y), round(m), mu)). All likelihood-based statistics now derive from the residual/null deviance, which equals -2LL exactly for a 0/1 response and honors fractional frequency weights unrounded (Charter §5.1). Unweighted and integer-weighted results are unchanged. A weighted validation test with a hand-summed log-likelihood oracle pins the fix.
• linear_regression(use = "pairwise") no longer silently drops interactions and transformed terms. The pairwise path rebuilds the model from the raw correlation matrix, so y ~ a * b was silently fitted as y ~ a + b. Formulas with non-additive terms now abort with a pointer to use = "listwise".
• Rank-deficient fits no longer crash the unweighted linear_regression() path ("Tibble columns must have compatible sizes"): perfectly collinear terms are excluded from the coefficient table (matching SPSS's excluded-variables handling) with a one-line message, and the ANOVA df now count estimated terms (model$rank) instead of raw coefficients — the same rule the weighted path has used since 0.6.4.

New features

• The linear_regression() coefficients table now prints the confidence intervals for B (SPSS /STATISTICS CI) that it always computed; the level is the conf.level argument. A new summary(..., conf_int = FALSE) toggle hides the two columns.

Documentation

• factors = "dummy" docs (both regressions) now state that ordered factors get R's default polynomial contrasts (.L/.Q/.C terms), not treatment dummies, and how to get dummy coding.
• The SPSS-compatibility vignette now discloses logistic_regression() as Tier 4 (textbook-formula oracle; no SPSS v29 reference run yet) instead of implying SPSS-validated status.
• The logistic classification-table header reads the stored cutoff instead of hardcoding "0.50".

mariposa 0.6.14

Codebook robustness (theme: codebook() survives real-world data and says what it shows). A stress test of the codebook stack (metadata extraction, console print/summary, HTML builder, xlsx export) surfaced a batch of crashes, silent data errors, and display leaks; this release fixes all of them and adds a view argument for side-effect control.

Bug fixes

• Variables with value labels but no variable label no longer show a fake label like "1 | 2 | 3": attr(x, "label") partially matched the labels attribute; all label reads now use exact = TRUE. The "Variables with labels" count excludes such variables accordingly.
• codebook() no longer errors on inline data expressions (e.g. codebook(data.frame(...)) spanning multiple deparse lines); long expressions collapse to the generic dataset name "data".
• List columns no longer crash the frequency computation: they are skipped with a warning ("list column ... skipped - not supported in codebooks") and a data frame consisting only of list columns aborts with a clear error.
• The tagged-NA breakdown is now consistent across all three layers:
  • the HTML builder appends the missing-value rows (codes, labels, counts) below range-displayed variables too — previously an all-user-missing or high-cardinality variable silently lost them (the xlsx export already did this correctly);
  • NA frequencies are computed even for high-cardinality/range-displayed variables;
  • print(summary(cb)) gains a "Missing values:" section (codes, labels, counts) — show_na was a no-op on the console layer before.

Improvements

• Central display formatting for numeric values: empirical values and ranges no longer leak 15-digit doubles or scientific notation into the console/HTML/xlsx output. Fractional values show 4 significant digits ("0.3333"), whole numbers keep their integer look ("1"), tiny values are expanded ("0.00000001", never "1e-08"). Frequency matching is unaffected: display strings and raw matching keys are carried separately (empirical_values vs. new empirical_keys).
• The percentage and effective-n columns (prc, valid_prc, cum_prc, n_eff) on write_xlsx(cb, frequencies = TRUE) sheets are rounded to 2 decimals.
• max_values (single integer >= 1) and max_len (single integer >= 4) are validated up front with a clear error.
• Factor levels now respect max_values and truncate with the same "... (N more)" note used for character values.
• Range displays show the cardinality: "18 - 95 (78 distinct)" instead of just "18 - 95".
• file = into a nonexistent directory aborts early, naming the missing directory.
• Polish: "1 variable" / "1 observation" pluralization in the console and HTML subtitle; very long character values are truncated to max_len with "..." (raw values still drive frequency matching); zero-row data frames say "(no observations)" instead of "(all missing)".

New features

• New view argument for codebook() (default: interactive()): controls whether the HTML codebook opens in the RStudio Viewer. view = FALSE suppresses the Viewer side effect entirely; writing via file = is unaffected. The compact print() only advertises the Viewer when it was actually opened (result gains a viewed flag).

mariposa 0.6.13

McDonald's omega (theme: reliability() learns a second reliability coefficient). reliability() now reports McDonald's omega alongside Cronbach's alpha — a new statistic within an existing function, hence a PATCH per the clarified versioning policy.

New features

• reliability() computes McDonald's omega from a one-factor maximum-likelihood model (stats::factanal on the same (weighted) correlation matrix already used for standardized alpha):
  • omega — raw/total omega in the covariance metric (analogous to raw alpha), reported as "McDonald's Omega";
  • omega_std — standardized omega in the correlation metric (analogous to standardized alpha);
  • omega_if_deleted — a new column in item_total, refitting the one-factor model per deleted item (NA when the reduced scale has fewer than 3 items, where the model is unidentified). Scales with fewer than 3 items get NA omega fields plus a warning (alpha is unaffected); non-convergent factor fits degrade to NA with the factanal message. The compact print() shows omega next to alpha, and summary() adds omega rows to the Reliability Statistics block and an omega column to the Item-Total table.

Validation

• McDonald's omega is Tier 4 (Internal, R-only) for now: SPSS v27+ offers omega in RELIABILITY, but IBM's algorithm documentation is not publicly retrievable and no SPSS v29 reference run exists yet. The pending reference run is prepared in .claude/spss-syntax-omega-references.sps (expected values included); until it lands, omega is guarded by a parameter-recovery test on simulated congeneric data, exact cross-checks against a manual factanal computation, cross-checks against psych::omega() and a lavaan/semTools one-factor CFA (tests/testthat/test-reliability-omega.R), and a w == 1 block in the weights-invariance suite. The help page carries the Tier-4 disclosure; the compatibility vignette flags omega as Internal (Tier 4).
• psych, lavaan, and semTools added to Suggests (cross-check tests only; all gated by skip_if_not_installed()).

mariposa 0.6.12

Weighted-rank correctness and accurate claims (theme: the weighted rank family says exactly what it is). Two formula errors in weighted rank statistics are fixed and a package-wide invariance suite now guards every weighted entry point; alongside, the user-facing claim surface (README, DESCRIPTION, help pages, compatibility vignette) is realigned with what the validation suite actually covers.

Bug fixes

• Weighted kendall_tau(): the tau-b denominator omitted double-tied pairs (ties_both) from the two tie-correction factors, deflating |tau| on tied data. The weighted denominator now mirrors the unweighted (n0 - Tx - Txy)(n0 - Ty - Txy) structure.
• Weighted kruskal_wallis(): the grand mean rank was still the hard-coded N/2 of the pre-0.6.4 rank convention instead of (N+1)/2, inflating H. It is now derived from the weighted mid-ranks themselves.
• Both bugs violated the invariant that weights of exactly 1 must reproduce the unweighted result. A new package-wide invariance suite (tests/testthat/test-weights-invariance.R) enforces this w == 1 reduction for every weighted entry point; intentionally approximate reductions (design-based mann_whitney, weighted Kendall z/p) are documented exceptions with bounded assertions.

Accurate claims

• README and DESCRIPTION no longer overclaim: the paired t-test mode (not yet implemented) is no longer advertised, and "every function is validated ... your results will match" is replaced by the Charter-compliant wording — validated against SPSS v29 within documented per-tier tolerances, with vignette("spss-compatibility") for per-function status.
• The weighted variants of the rank-based family — mann_whitney(), kruskal_wallis(), wilcoxon_test(), friedman_test(), binomial_test(), dunn_test(), pairwise_wilcoxon(), and kendall_tau() — are now disclosed as R-only (Tier 4) in a "Weighted variants" note on each help page: SPSS NPAR TESTS / NONPAR CORR ignore WEIGHT BY, so no SPSS reference exists for these weighted paths. mann_whitney()'s note also states that its design-based U/W may differ from SPSS's expanded-data U (Z and p are the validated quantities); oneway_anova() now documents that omega-/epsilon-squared are truncated at 0 (negative raw estimates occur when F < 1).

Validation

• The SPSS-compatibility vignette is regenerated (was frozen at 2026-05-19) and now carries an "Internal (Tier 4)" marker for the weighted rank variants. Generator fixes: the w_* family and scheffe_test are correctly matched to their shared test files (previously shown as "not validated" despite existing tests), zero-match tier counts no longer report as 1, and assert_spss_count() call sites are tallied as Spec.
• test-t-test-spss-validation.R: the header tier table claimed the t-statistic at Spec (±1e-5) while the assertions use Display(3); the header now matches the assertions.
• The last expect_no_error() in a validation file (test-linear-regression-spss-validation.R) is replaced by real assertions on the per-group predictions; the validation-discipline meta-test now passes with MARIPOSA_VALIDATION_STRICT=TRUE.

mariposa 0.6.11

Deprecation cleanup (theme: the due bridges come out). Two batches of deprecations reached their removal release together: the 0.6.9 argument bridges (originally slated for 0.6.10) and the 0.6.10 duplicate result columns. Removing both here keeps the run-up to the 1.0 API freeze tidy.

Breaking changes

• The 0.6.9 dot-case argument bridges are removed as announced. The old names no longer warn-and-work; they now error (falling through to tidyselect or SET-mode validation):

  • codebook(): show.id, show.type, show.labels, show.values, show.freq, show.na, show.unused, max.values, max.len, sort.by.name (use show_id, show_type, show_labels, show_values, show_freq, show_na, show_unused, max_values, max_len, sort_by_name)
  • val_labels(): drop.na (use drop_na)
  • drop_labels(): drop.na (use drop_na) These bridges were originally slated for removal in 0.6.10 and are batched into this release. The frequency()/rec()/to_label() family of removed-argument errors introduced in 0.6.9 remain in place as permanent guidance (their ... consumes tidyselect, so a clear error beats a silent misinterpretation).

• The deprecated duplicate result columns kept for one release in 0.6.10 are removed; only the canonical column remains:

  • chisq_gof(), friedman_test(): chi_sq removed (use chi_squared)
  • mcnemar_test(): statistic removed (use chi_squared)
  • mann_whitney(): effect_size_r removed (use r_effect)
  • oneway_anova(): F_stat removed (use F_statistic) The statistical values are unchanged; only the redundant column names go away.

mariposa 0.6.10

Result-column harmonization (theme: one statistic, one column name). A style audit found the same statistic carrying different result-column names across sibling functions; the drifted names now converge on the canonical spelling, with the old columns kept as duplicates for one release.

Improvements

• The $results columns for shared statistics are harmonized on the canonical names already used elsewhere in the package:
  • Chi-square statistic: chi_squared (as in chi_square()) - now also in chisq_gof(), friedman_test(), and mcnemar_test()
  • Effect size r: r_effect (as in wilcoxon_test()) - now also in mann_whitney()
  • F statistic: F_statistic (as in levene_test()) - now also in oneway_anova() Print and summary methods read the canonical columns; the statistical values are unchanged.

Deprecations

• The old result-column names remain available as duplicated columns (positioned right after their canonical counterpart) for one release and will be removed in 0.6.11:
  • chisq_gof(), friedman_test(): chi_sq (use chi_squared)
  • mcnemar_test(): statistic (use chi_squared)
  • mann_whitney(): effect_size_r (use r_effect)
  • oneway_anova(): F_stat (use F_statistic)

mariposa 0.6.9

API-cleanup completion (theme: the 0.6.8 bridges come out, the last dot-case stragglers get theirs). One step closer to the 1.0 API freeze.

Breaking changes

• The 0.6.8 deprecation bridges are removed as announced. The dot-case argument names now error instead of warning:
  • frequency()/fre(): sort.frq, show.na, show.prc, show.valid, show.sum, show.labels, show.unused
  • rec(): as.factor, var.label, val.labels
  • to_label()/to_character()/to_numeric(): drop.na, drop.unused, add.non.labelled, use.labels, start.at, keep.labels
  • read_spss()/read_por()/read_stata()/read_sas()/read_xpt(): tag.na Before: frequency(data, x, sort.frq = "desc") warned and worked. After: it errors with a pointer to sort_frq. In the functions whose ... selects variables, the old names raise a clear "removed in 0.6.9" error instead of being silently swallowed by tidyselect; in the readers they fail as unused arguments.

Deprecations

• The remaining dot-case arguments are renamed to snake_case with the usual one-release bridge (old names warn once per session; removal planned for 0.6.10):
  • codebook(): show_id, show_type, show_labels, show_values, show_freq, show_na, show_unused, max_values, max_len, sort_by_name
  • val_labels(): drop_na
  • drop_labels(): drop_na The display options stored on codebook results (result$options) use the snake_case keys as well.

mariposa 0.6.8

API-unification release (theme: snake_case arguments). One release-long deprecation bridge per the versioning policy - old names keep working and warn once per session; they will be removed in 0.6.9.

Breaking changes (with bridge)

• Dot-case arguments renamed to snake_case:
  • frequency()/fre(): sort_frq, show_na, show_prc, show_valid, show_sum, show_labels, show_unused
  • rec(): as_factor, var_label, val_labels
  • to_label()/to_character()/to_numeric(): drop_na, drop_unused, add_non_labelled, use_labels, start_at, keep_labels
  • read_spss()/read_por()/read_stata()/read_sas()/read_xpt(): tag_na Base-R-universal names (na.rm, conf.level, var.equal) are kept.

Breaking changes (no bridge)

• t_test() results no longer carry the duplicated CI_lower/CI_upper alias columns; conf_int_lower/conf_int_upper are the contract.

Improvements

• sort_frq is validated ("none"/"asc"/"desc") - typos used to silently produce an unsorted table; show_labels validates its TRUE/FALSE/"auto" values with a clear error.
• Vignettes showcase the new argument names.

mariposa 0.6.7

Output-layer release (theme: uniform three-layer output). Statistical results are unchanged; what changed is how results present themselves.

Uniform three-layer output (visible change)

Every analysis class now follows the documented pattern that t_test and chi_square pioneered: result prints a compact overview (headline statistic, p-value, significance stars, one line per test), and summary(result) carries the full detailed output behind boolean section toggles. Newly migrated: kruskal_wallis, wilcoxon_test, friedman_test, binomial_test, fisher_test, chisq_gof, mcnemar_test, levene_test, tukey_test, scheffe_test, dunn_test, pairwise_wilcoxon, frequency, crosstab (describe was already compact and gained the summary layer for uniformity). Nothing was removed - everything the old print() showed is in summary(), verified line-by-line.

Internal architecture

• One shared engine for the three correlation functions (pearson/spearman/kendall results verified byte-identical across 21 scenarios; ~500 lines removed).
• The w_* factory now supports multi-value and non-numeric statistics; w_quantile() and w_modus() are ordinary plugins instead of pipeline reimplementations (~420 lines removed, results identical across 31 scenarios).

mariposa 0.6.6

Internal-architecture release (theme: shared cores and formatting utilities). No statistical results change; table rendering in the Tukey/Scheffe output is now aligned and uses SPSS-style p display.

• One home for every weighted formula: the new weighted-statistics kernel file backs describe(), frequency(), the w_* functions and the rank tests; the weighted variance formula previously existed in six files.
• New internal output utilities (bordered table renderer, grouped-results iterator, unified number/p formatting) - the building blocks the print style guide documented; adoption started with the post-hoc tests.
• Tukey and Scheffe now share one engine and one print implementation (results verified byte-identical); t_test() and oneway_anova() were restructured from 500-line nested-closure bodies into short orchestrators with file-level helpers (byte-identical results).
• Weights in summarise() context are captured as quosures (enquo/eval_tidy) instead of frame-walking; shared validators report errors at the user-facing call site.
• Documentation internals standardized on @noRd (man/ shrinks by ~120 internal pages).

mariposa 0.6.5

Housekeeping release (theme: package hygiene). No statistical results change.

• Slimmer dependencies: removed the unused tidyr import and pruned unused importFrom entries.
• Error chains: failures inside grouped analyses are re-thrown with cli_abort(parent = ...) so the original condition is preserved; the haven requirement is enforced by one central guard that reports the calling function instead of an internal helper.
• Import internals: the native-missing-value detection shared by read_stata(), read_sas() and read_xpt() now lives in one helper instead of three copies.
• Mechanical polish: remaining sapply() calls in the oldest files converted to type-stable vapply(); pkgdown reference now lists phi(), cramers_v(), goodman_gamma().
• Test suite: removed a legacy tolerance registry that contradicted the Validation Charter and was no longer used by any test.

mariposa 0.6.4

A quality release. Following an in-depth internal review of the entire statistical codebase, this version sharpens the accuracy of several statistics, makes the package behave more consistently across functions, and adds a dedicated regression-test suite (tests/testthat/test-audit-regressions.R) so these guarantees hold in future releases. Some outputs change slightly as a result - in every case toward the standard reference implementations.

More accurate statistics

• Kendall's tau-b significance test now agrees with stats::cor.test() (and the SPSS formula) to machine precision, which is most noticeable for heavily tied data such as binary variables.
• The weighted Wilcoxon signed-rank test (and pairwise_wilcoxon()) now uses frequency-expansion mid-ranks: with integer weights the statistic equals the expanded-data Wilcoxon exactly, and weights = 1 reproduces the unweighted test. Displayed rank means in the weighted Kruskal-Wallis and Dunn tests follow the same convention.
• The Mann-Whitney asymptotic p-value now matches its reported Z (both follow the SPSS convention without continuity correction).
• Regression degrees of freedom are now derived from the fitted model terms, improving results for models with dummy-coded factors or interaction terms (weighted linear regression and the logistic omnibus test).
• The Kruskal-Wallis effect size is now correctly labelled: the returned field is epsilon_squared (previously named eta_squared).

New and refined API

• linear_regression() gains SPSS-style collinearity diagnostics (Tolerance and VIF per model term), including a collinearity toggle in summary().
• phi(), cramers_v(), and goodman_gamma() now return the requested effect size directly as a numeric value - the convenient behavior their names suggest. For the full test output, use chi_square().
• The weighted two-sample t_test() now honors var.equal for its primary result. oneway_anova() always reports both the classical and Welch results (like SPSS ONEWAY), so its var.equal argument is deprecated; ss_type in factorial_anova()/ancova() is likewise deprecated in favor of the SPSS-standard Type III.

More consistent behavior

• One package-wide weights policy: invalid (negative) weights are now rejected with a clear message at every entry point, instead of being handled differently depending on the function.
• The weighted median now always equals the weighted 50th percentile, and unweighted quantiles follow the SPSS convention (Type 6/HAVERAGE) throughout.
• frequency() header statistics use the same formulas as describe() and the w_* functions.
• Significance stars follow a single boundary convention everywhere, matching the printed legend.

More robust in edge cases

• Correlation functions handle constant variables gracefully (NA instead of an error), frequency(show.unused = TRUE) works on variables tagged via set_na()/read_spss(), and frequency(sort.frq =) now sorts by frequency with a monotone cumulative-percent column.
• write_spss() protects valid values when many missing-value codes must be consolidated into a range, and explains what it is doing.
• logistic_regression() surfaces separation and convergence warnings again; post-hoc tests report when a computation could not be carried out instead of skipping it silently.

Housekeeping

• Internal code paths were consolidated (shared helpers for weighted mid-ranks and the Wilcoxon core) and a substantial amount of unused code was removed, making the codebase easier to maintain.
• Grouped single-variable w_* results print their statistics again.

mariposa 0.6.3.2

rec() reliably matches decimal single values

Single-value recode rules now match decimal codes (e.g. "3.6=2") even when the stored value carries floating-point representation error. The single-value comparison was changed from exact numeric equality (x == value) to a string comparison (as.character(x) == as.character(value)), which rounds to 15 significant digits and thereby absorbs the error.

Reason: a value such as 0.1 + 0.2 is stored as 0.30000000000000004, so the previous exact == test silently failed to match a rule "0.3=...". This mirrors the behaviour of sjmisc::rec(), on which rec()'s string syntax is modelled. Range rules were already robust (they use >=/<=) and are unchanged.

mariposa 0.6.3.1

broom tidiers now work natively

Adds explicit tidy(), glance(), and augment() methods for both linear_regression and logistic_regression results, registered via the standard s3_register() pattern (broom in Suggests, no hard dep).

Reason: with class(r) = c("linear_regression", "lm"), broom::tidy.lm() and broom::glance.lm() dispatched as expected, but internally called summary(x) — which (because of our specialised summary.linear_regression() overriding summary.lm) returned the mariposa SPSS-style summary instead of the lm summary broom needs. The visible failures:

• broom::glance(r) raised object 'r.squared' not found because mariposa's summary stores it as R_squared.
• broom::tidy(r, conf.int = TRUE) returned only 4 columns (term, estimate, conf.low, conf.high) instead of the expected 6+ (term, estimate, std.error, statistic, p.value, conf.low, conf.high).

The new methods strip our linear_regression / logistic_regression class before delegating to broom::tidy.lm / tidy.glm etc., so the inner summary() call dispatches to summary.lm / summary.glm and broom receives its expected shape. The user-facing summary(r) still returns mariposa's SPSS-style output (more specific method wins).

Edge cases stay consistent with the rest of the lm-generic surface: broom::tidy() / glance() / augment() on a grouped or pairwise result raise an actionable error pointing at lapply(r$groups, ...) or use = "listwise".

New tests in test-broom-methods.R cover all three tidiers for both regression types, plus the grouped/pairwise error paths.

mariposa 0.6.3

Behavior Change — regression results inherit from lm / glm

linear_regression() and logistic_regression() results now ARE the fitted lm / glm object (with mariposa-specific tables attached as additional slots), instead of wrapping it in $model. All base-R and broom generics dispatch natively:

r <- linear_regression(survey_data, life_satisfaction ~ age + income)
coef(r)                                 # named numeric vector
predict(r, newdata = head(survey_data)) # works directly
anova(r)                                # sequential SS table
vcov(r); confint(r); residuals(r); fitted(r)
broom::tidy(r); broom::glance(r); broom::augment(r)

Class hierarchy is c("linear_regression", "lm") for linear and c("logistic_regression", "glm", "lm") for logistic. summary(r) still returns the SPSS-style mariposa summary (more specific method wins); for the raw lm/glm summary call stats::summary.lm(r) / stats::summary.glm(r).

Slot renames (breaking)

Two slots collided with lm/glm conventions and were renamed:

| Before | After | |:------------- |:-------------- | | $coefficients (tibble) | $coef_table (tibble) | | $anova (tibble) | $anova_table (tibble) | | $model (lm/glm) | the object IS the model — use r directly |

Migration:

• r$coefficients → r$coef_table (SPSS-style tibble) or coef(r) (named numeric vector).
• r$anova → r$anova_table (SPSS-style overall-model ANOVA tibble) or anova(r) (R's per-term sequential SS table).
• r$model |> predict(...) → predict(r, ...) directly.
• r$model |> broom::tidy() → broom::tidy(r) directly.

Edge cases

• use = "pairwise": no single fitted lm is available, so the result is a custom list with class "linear_regression" only. predict()/anova() etc. raise an informative error pointing at use = "listwise".
• Grouped results (top-level): no single model. predict()/anova() raise an informative error pointing at lapply(r$groups, predict, ...). Each r$groups[[i]] is itself an lm-inheriting object, so per-group generics work directly.

Test Suite

• New test block in test-linear-regression-spss-validation.R verifies that coef(), predict(), anova(), vcov(), confint(), residuals(), fitted(), formula(), nobs(), model.matrix() all dispatch natively, plus the grouped/pairwise error paths.
• 1184/1184 tests pass; R CMD check on built tarball: Status OK.

mariposa 0.6.2

Behavior Change

linear_regression() and logistic_regression(): factor predictor handling

Both regression functions now expose a factors argument controlling how factor predictors enter the model. The new default factors = "dummy" matches base R lm() / glm(): a factor with L levels expands into L - 1 dummy contrasts via stats::model.matrix(). Previous versions silently coerced factor levels to integer codes (SPSS ordinal-as-scale default) with no warning, which surprised users who relied on standard R semantics.

To restore the previous SPSS-style behavior, pass factors = "numeric" explicitly. That mode emits a one-line cli::cli_inform() listing the coerced variables for transparency. The "numeric" mode is required to reproduce SPSS REGRESSION / LOGISTIC REGRESSION output when factor predictors carry ordered meaning (e.g., a 4-level education variable treated as 1–4 ordinal scale).

Behavioral consequences:

• A model with a 3-level factor predictor that previously returned one coefficient row now returns two dummy-contrast rows under the new default.
• For pairwise missing handling (use = "pairwise"), factor predictors are not supported with factors = "dummy"; the function now errors with an actionable message pointing to either factors = "numeric" or use = "listwise".

Migration: scripts that depend on the old SPSS-style coercion should set factors = "numeric" at the call site. The cli_inform() message can be silenced with suppressMessages() if desired.

Source-Code Fixes (weighted regression)

Two more functions joined the Charter §5.1 audit list (the "unrounded sum(w)" weighted-statistics convention previously applied to t_test, oneway_anova, and levene_test):

• linear_regression(): weighted variance, SE, df, F, R², and adjusted-R² now use the unrounded sum(weights) throughout. Earlier versions used n_effective <- round(sum(w)) in df and MS calculations, producing systematic drift from SPSS REGRESSION (off by ~0.001 on F, ~0.01 on adj-R² for typical weights). The displayed N is still round(sum(w)).
• logistic_regression(): pseudo-R² formulas (Cox & Snell, Nagelkerke, McFadden) now use the unrounded sum(weights) in the exponential denominator. The displayed N and rounded classification counts remain integers.

These are bug fixes; weighted results may shift slightly toward closer agreement with SPSS v29.

Other Fixes

• summary.linear_regression(descriptives = TRUE) now actually prints the Descriptive Statistics table (Variable, Mean, SD, N). Previously the parameter was accepted but documented as "Reserved for future use" and produced no output.
• The compact print.linear_regression() no longer crashes on weighted models with non-integer df: the F-statistic line now rounds df for display before formatting with %d.
• Roxygen examples for summary.linear_regression (collinearity = FALSE) and summary.logistic_regression (classification_table = FALSE) referenced parameters that do not exist; corrected to descriptives = FALSE and classification = FALSE respectively.

Test Suite

The linear_regression SPSS validation test suite expanded from 1 scenario (unweighted bivariate) to 6 scenarios covering all four Charter §8 quadrants — Tests 1a, 1c, 2a, 2c, 3a, and 4a from tests/spss_reference/outputs/linear_regression_output.txt. The weighted scenarios (2a, 2c, 4a) verify the Charter §5.1 fix above. New behavioral tests cover the factors argument (dummy expansion, numeric coercion, pairwise + dummy + factor error path). 222/222 assertions pass.

mariposa 0.6.1

Validation

Substantial hardening of the SPSS-compatibility test suite. All 29 SPSS- validation test files were rewritten under a new Validation Charter (see vignette("spss-compatibility")) that defines tolerance tiers (Spec / Display / Exception / Internal), forbids inline tolerance literals, NA placeholders, and expect_true(TRUE) reporting blocks, and requires citation comments linking every reference value to its source line in tests/spss_reference/outputs/.

• 1832+ passing assertions across all 29 validation files, 0 failures.
• New tests/testthat/helper-validation-tolerances.R provides assert_spss() and tol() helpers with explicit tier semantics.
• New tests/testthat/test-validation-discipline.R meta-test lints validation files for Charter-forbidden patterns.
• New vignettes/spss-compatibility.Rmd reports per-function validation status, auto-generated from the test suite.
• New CI workflow .github/workflows/strict-validation.yaml runs the full suite in strict-discipline mode on release tags and weekly.

Source-Code Fixes (weighted statistics)

Three weighted statistical functions were corrected to use unrounded sum(w) per SPSS frequency-weights convention. Earlier versions rounded too early and produced systematic drift from SPSS in weighted scenarios.

• t_test(): weighted variance, SE, and df calculations now use unrounded sum(w) (one-sample and two-sample paths). Welch- Satterthwaite df now derived from unrounded per-group weighted N.
• oneway_anova(): weighted variance divisor is now (sum(w) - 1) (sample formula, not population). Weighted SE uses sqrt(sum(w)), not sqrt(physical n). Weighted CI t-critical-value uses df = sum(w) - 1, not Kish design-effective N. df_within now uses floor(sum(w)) - k (SPSS ONEWAY-specific convention).
• levene_test(): weighted Levene df now uses unrounded sum(w) - k (SPSS T-TEST family convention).

These changes are bug fixes and may slightly shift weighted-scenario results in user code. Differences are small (typically < 0.01 on F or t) and bring mariposa into closer agreement with SPSS v29.

SPSS-Compatibility Vignette

vignette("spss-compatibility") documents the per-function validation status, the four tolerance tiers, and the SPSS-procedure-specific WEIGHT BY conventions discovered during the migration:

• T-TEST family: unrounded sum(w)
• ONEWAY: floor(sum(w))
• UNIANOVA: Type III SS
• NPAR TESTS: WEIGHT BY effectively ignored
• NONPAR CORR (Spearman, Kendall): WEIGHT BY effectively ignored
• CORRELATIONS (Pearson): WEIGHT BY honored
• CROSSTABS, FREQUENCIES, RELIABILITY, FACTOR, REGRESSION: WEIGHT BY honored

DESCRIPTION

• Title shortened to "SPSS-Compatible Statistical Tools for Survey Data" (CRAN soft-limit compliance).
• Suggests cleanup: removed PMCMRplus and survey (no longer needed).

Audit-Driven Math Fixes (post-Phase-1)

A second audit pass identified additional math defects and test fudges, all corrected in this release:

• dunn_test(): SE now includes the Dunn (1964) / Conover (1999) tie correction. Previous versions systematically under-estimated |Z| on tied data (e.g., Likert scales). Baselines regenerated from PMCMRplus::kwAllPairsDunnTest (exact match to 4 decimals).
• friedman_test(): weighted branch now applies the tie correction consistently with stats::friedman.test (unweighted branch). The inconsistency caused weighted chi-squared values to be too low for tied data.
• describe(): weighted skewness and kurtosis now delegate to .calc_skewness() / .calc_kurtosis() in helpers.R (Joanes-Gill Type-2 with Σw substitution), matching w_skew() / w_kurtosis() and SPSS FREQUENCIES exactly. The previous duplicate implementation used a simple weighted moment without bias correction.
• .w_quantile(): weighted quantiles now use Type-6 (HAVERAGE) linear interpolation between cumulative-weight crossings — matches SPSS FREQUENCIES /PERCENTILES. Unweighted quantiles also switched from R default type = 7 to SPSS-compatible type = 6.

Documentation Honesty

Several SPSS-compatibility claims were narrowed to reflect what the code actually does:

• Source comments and test-file headers for the weighted paths of kruskal_wallis(), wilcoxon_test(), and friedman_test() corrected from "design-based" / "Lumley-Scott" to "frequency-weighted approximation". Only mann_whitney() is a genuine Lumley & Scott (2013) implementation; the others substitute sum(w) for n in the standard variance formula.
• mann_whitney() test now includes a permanent cross-check against survey::svyranktest() (skipped when survey is not installed).
• spearman_rho(): weights parameter docstring rewritten to disclose that weights are used only for case filtering (per SPSS NONPAR CORR convention), not in the rank correlation itself.
• pearson_cor(): docstring now warns that the weighted-df convention (n = sum(w)) gives spuriously narrow CIs for raw expansion weights; users with such weights should normalize first.
• logistic_regression(): test file replaced with property-based assertions (Wald formula, Sig from chi-sq, exp(B) vs independent 2x2 odds ratio, Cox & Snell/Nagelkerke/McFadden from textbook formulas, Omnibus from likelihood ratio). No longer a tautological glm-vs-glm self-comparison.

Code Smell Cleanup

• oneway_anova(): removed dead-code overwrite of grand_mean_welch in the weighted Welch path.
• levene_test(): stale comment claiming df2 = floor(sum(w)) - k corrected — the code uses unrounded sum(w) - k (T-TEST family convention).

mariposa 0.6.0

New Functions — Label Management

This release adds 10 label management functions for working with labelled survey data (inspired by sjlabelled, consolidated into a clean, consistent API), plus data transformation, row operations, and data exploration functions.

Variable & Value Labels

• New var_label(): dual-mode function for getting and setting variable labels. var_label(data) returns all variable labels as a named character vector; var_label(data, x = "Age", y = "Gender") sets labels for specific columns. Supports tidyselect for column selection when getting labels.

• New val_labels(): dual-mode function for getting and setting value labels. val_labels(data) returns all value labels as a named list; val_labels(data, x = c("Low" = 1, "High" = 2)) sets labels. Use .add = TRUE to extend existing labels without replacing them.

• New copy_labels(): copies all label attributes (variable labels, value labels, class, tagged NA metadata) from a source data frame to matching columns in the target. Essential for preserving labels after dplyr operations that strip attributes.

• New drop_labels(): removes value labels for values that do not actually occur in the data. Use drop.na = TRUE to also remove labels for tagged NA values.

Type Conversions

• New to_label(): converts haven_labelled vectors to factors, using value labels as factor levels. Supports ordered, drop.na, drop.unused, and add.non.labelled options. Factor levels are ordered by their original numeric codes (not alphabetically).

• New to_character(): converts haven_labelled vectors to character, replacing numeric codes with their label text.

• New to_numeric(): converts factors or labelled vectors to numeric. When use.labels = TRUE, uses value labels if they are numeric; otherwise assigns sequential integers (controlled by start.at).

• New to_labelled(): converts factors, character, or numeric vectors to haven_labelled with proper value labels. Factor levels become value labels automatically.

Missing Value Management

• New set_na(): declares specific numeric values as missing (NA or tagged NA). Supports unnamed values (applied to all numeric columns) and named pairs for per-variable control (e.g., set_na(data, income = c(-9, -8))). With tag = TRUE (default), creates tagged NAs that integrate with na_frequencies(), frequency(), and codebook(). Can be called incrementally to add new missing value codes.

• New unlabel(): strips all label metadata from variables, converting haven_labelled vectors to plain base R types. Removes variable labels, value labels, tagged NA metadata, and format attributes. Tagged NAs become regular NA. Supports tidyselect for selective column unlabelling.

New Functions — Data Transformation

• New rec(): flexible recoding with string syntax (e.g., rec(data, x, rec = "1:2=1 [Low]; 3:5=2 [High]")). Supports value ranges, min/max keywords, copy for unchanged values, and automatic value label generation from bracket syntax. Works with numeric, character, and labelled vectors.

• New to_dummy(): creates dummy (indicator) variables from categorical or labelled vectors. Generates one 0/1 column per unique value with informative column names. Supports tidyselect for multi-variable dummy coding and suffix = "label" to use value labels in column names.

• New std(): z-standardization with four methods ("sd", "2sd", "mad", "gmd"). Supports survey weights, grouped standardization via dplyr::group_by(), and robust = TRUE for median/MAD-based standardization.

• New center(): mean-centering (grand-mean or group-mean). Supports survey weights and dplyr::group_by() for group-mean centering. Returns centered values with the centering value stored as an attribute.

New Functions — Row Operations

• New row_means(): computes row-wise means across selected columns, with min_valid parameter matching SPSS MEAN.x() syntax. Designed for use inside dplyr::mutate(). Replaces the deprecated scale_index().

• New row_sums(): computes row-wise sums across selected columns, with min_valid parameter for minimum valid (non-NA) values.

• New row_count(): counts occurrences of specific values per row. Useful for counting endorsements in multi-item scales (e.g., how many items a respondent agreed with).

New Functions — Data Exploration

• New find_var(): searches variables by name or label using regular expressions. Returns matching variable names with their labels. Useful for exploring large survey datasets with many variables.

Breaking Changes

• scale_index() has been removed and replaced by row_means(), which provides the same functionality with a clearer name. Update existing code: scale_index(data, x, y, z) → row_means(data, x, y, z).

mariposa 0.5.6

New Functions

• New write_spss() function: exports data frames to SPSS .sav format with full tagged NA roundtripping. Tagged NAs are converted back to SPSS user-defined missing values, enabling lossless roundtrips via read_spss() -> processing -> write_spss(). Supports byte, none, and zsav compression.

• New write_stata() function: exports data frames to Stata .dta format. Tagged NAs from any source format are written as Stata extended missing values (.a through .z). Supports Stata versions 8-15.

• New write_xpt() function: exports data frames to SAS transport .xpt format. Tagged NAs are written as SAS special missing values (.A through .Z, ._). Supports transport versions 5 and 8.

Enhancements

• mariposa now provides a unified data import/export platform:
  • Import: read_spss(), read_por(), read_stata(), read_sas(), read_xpt(), read_xlsx()
  • Export: write_spss(), write_stata(), write_xpt(), write_xlsx()
• Cross-format export is supported: data imported from one format can be exported to another (e.g., SPSS to Stata) with automatic missing value type conversion.

mariposa 0.5.5

New Functions

• New read_xlsx() function: reads Excel (.xlsx) files with automatic label reconstruction. When reading back files created by write_xlsx(), variable labels, value labels, and tagged NA metadata are fully restored -- enabling lossless roundtripping of labelled survey data through Excel.

  • Auto-detects mariposa export format (data frame, list, codebook)
  • Reconstructs haven_labelled columns, factor levels, and variable labels
  • Restores tagged NAs with na_tag_map from missing codes in the data
  • Works as a plain Excel reader for non-mariposa files

• New write_xlsx() generic: exports data frames, codebooks, and named lists to Excel (.xlsx) with full support for variable labels, value labels, and tagged NA metadata. Uses openxlsx2 as an optional dependency.

  • write_xlsx(data, "file.xlsx") -- data + "Labels" reference sheet with variable labels, value labels, and missing value codes
  • codebook(data) |> write_xlsx("codebook.xlsx") -- structured codebook workbook with Overview, Codebook, and optional per-variable frequency sheets (frequencies = TRUE)
  • write_xlsx(list(a = df1, b = df2), "multi.xlsx") -- multi-sheet export where each named list element becomes a sheet

Enhancements

• write_xlsx() now preserves tagged NA codes (-9, -11, etc.) as visible values in the data sheet instead of empty cells, enabling perfect roundtripping with read_xlsx(). System NAs remain as empty cells.
• The "Labels" sheet now includes a Column_Type column (haven_labelled or factor) so read_xlsx() can deterministically reconstruct column types.

Dependencies

• Added openxlsx2 as a suggested dependency for Excel import/export.

mariposa 0.5.4

New Functions

• New read_stata() function: reads Stata .dta files and annotates native extended missing values (.a through .z) for use with mariposa's tagged NA system. Stata tagged NAs are preserved automatically by haven; read_stata() adds the na_tag_map attribute for seamless integration with na_frequencies(), frequency(), and codebook().

• New read_sas() function: reads SAS .sas7bdat files with optional catalog file (.sas7bcat) for value labels. Annotates SAS special missing values (.A through .Z and ._) for tagged NA integration.

• New read_xpt() function: reads SAS transport files (.xpt) with tagged missing value support. Transport files are the FDA-approved, platform-independent SAS data format.

• New read_por() function: reads SPSS portable .por files with the same tagged NA support as read_spss(). Shares the SPSS missing value conversion logic internally.

Breaking Changes

• na_frequencies() column spss_code has been renamed to code to reflect multi-format support. The column now contains character values: numeric SPSS codes (e.g., "-9") or native format codes (e.g., ".a" for Stata, ".A" for SAS).

Improvements

• na_frequencies(), untag_na(), and strip_tags() now work universally with data from all supported formats (SPSS, Stata, SAS).

• untag_na() is now format-aware: for Stata and SAS data (where tagged NAs are the native representation with no numeric codes to recover), it warns and falls back to strip_tags() behavior.

• frequency() and codebook() automatically display format-appropriate missing value codes (e.g., -9 for SPSS, .a for Stata, .A for SAS).

mariposa 0.5.3

New Functions

• New read_spss() function: reads SPSS .sav files and preserves user-defined missing values as tagged NAs instead of converting them to regular NA. This allows distinguishing between different types of missing data (e.g., "no answer", "not applicable", "refused") while still treating them as NA in standard R operations. Fixes the sjlabelled::read_spss(tag.na=TRUE) crash on large datasets (e.g., ALLBUS) caused by out-of-bounds letters[] indexing.

• New na_frequencies() function: shows a breakdown of the different types of missing values in a tagged NA variable, with counts, original SPSS codes, and value labels.

• New untag_na() function: converts tagged NAs back to their original SPSS missing value codes (e.g., -9, -8, -42).

• New strip_tags() function: converts all tagged NAs to regular (untagged) NA values, producing the same result as reading with haven::read_sav() directly.

Improvements

• frequency() now displays tagged NAs individually when data was imported with read_spss(). Each missing value type is shown as a separate row with its original SPSS code and label, followed by a "Total Valid" and "Total Missing" summary row.

• frequency() with show.unused = TRUE correctly handles tagged NA labels (no longer shows them as unused with freq=0).

mariposa 0.5.2.2

Convenience

• New fre() shorthand alias for frequency(). Both functions are identical; fre() simply provides a quicker way to call frequency analysis. ?fre shows the same help page as ?frequency.

mariposa 0.5.2

New Features

• New codebook() function: generates an interactive HTML data dictionary displayed in the RStudio Viewer pane. Shows variable ID, name, type, label, empirical values, value labels, and frequencies in a clean, scrollable table. Inspired by sjPlot's view_df() but built natively with htmltools.

• HTML codebook features a subtle-accent design: dark header, alternating row stripes, monospace type badges, and per-value frequency counts displayed as vertical lists aligned across columns.

• Console print() shows a minimal metadata overview (variable count, observations, types). Full details are reserved for the HTML viewer.

• summary() method provides a detailed text-based fallback with toggleable sections (overview, variable_details, value_labels).

• Supports tidyselect variable selection, optional survey weights for weighted frequencies, and sort.by.name ordering.

Dependencies

• Added htmltools as an imported dependency for HTML codebook generation.

mariposa 0.5.1

Three-Layer Output System

• All 13 analysis functions now support summary() for detailed SPSS-style output with toggleable sections. The three-layer pattern works as follows:

  • print() — compact one-line overview (default when typing the object name)
  • summary() — builds a detailed summary object with boolean section toggles
  • print.summary() — renders the full verbose output with all requested sections

• Supported functions: t_test(), oneway_anova(), factorial_anova(), ancova(), chi_square(), mann_whitney(), pearson_cor(), spearman_rho(), kendall_tau(), reliability(), efa(), linear_regression(), logistic_regression().

• Each summary() method accepts boolean parameters to control which output sections are displayed (e.g., summary(result, effect_sizes = FALSE) or summary(result, descriptives = FALSE)).

Internal Helpers

• Added build_summary_object() and format_p_compact() in R/summary_helpers.R as shared infrastructure for all summary methods.

Documentation

• Complete Roxygen2 documentation for all 39 S3 methods (13 print + 13 summary

  • 13 print.summary), each with @description, @param, @return, @examples, and @seealso.

• All 13 main function @examples now demonstrate the three-layer output pattern (result, summary(result), summary(result, toggle = FALSE)).

• Added print.reliability() and print.efa() documentation (previously undocumented).

Bug Fixes

• Fixed print.chi_square() Roxygen2 tag (@keywords internal replaced with correct @method print chi_square).

• Fixed example syntax errors in ancova() and factorial_anova() (formula syntax replaced with correct dv/between interface).

• Fixed incorrect variable name education_level in examples (corrected to education).

Tests

• Added test-summary-methods.R with tests for all 13 summary methods.

• Updated test-print-methods.R to reflect the new three-layer structure.

mariposa 0.5.0

New Functions

• Added factorial_anova() for multi-factor between-subjects ANOVA (up to 3 factors) with Type III Sum of Squares matching SPSS UNIANOVA. Includes main effects, all interaction terms, partial eta squared, R-squared, and Levene's test for homogeneity of variance. Full survey weight support via WLS (matching SPSS /REGWGT). Integrates with existing tukey_test(), scheffe_test(), and levene_test() S3 generics.

• Added ancova() for Analysis of Covariance — tests group differences after controlling for continuous covariates. Matches SPSS UNIANOVA with the WITH keyword. Provides ANOVA table, parameter estimates (B, SE, t, p, partial eta squared), estimated marginal means (adjusted for covariates), and Levene's test. Supports up to 3 factors and multiple covariates with full survey weight support.

SPSS Validation

• Added 612 SPSS validation tests for factorial_anova() across 9 scenarios: unweighted (2-factor, 3-factor, 2-factor with missing data), weighted (2-factor, 3-factor), grouped (2-factor, 3-factor), and weighted+grouped (2-factor, 3-factor).

• Added 579 SPSS validation tests for ancova() across 11 scenarios: one-way ANCOVA, two-way ANCOVA, weighted, grouped, weighted+grouped, multiple covariates, and single factor with single covariate.

• Total test suite: 4,986 tests passing (0 failures).

Technical Details

• Type III Sum of Squares computed via contr.sum contrasts and stats::drop1() — no dependency on the car package.

• Weighted analyses use WLS (stats::lm() with weights), matching SPSS's /REGWGT subcommand behavior exactly.

• Weighted Levene's test uses the SPSS /REGWGT algorithm: z_i = sqrt(w_i) * |y_i - weighted_cell_mean_i| followed by unweighted ANOVA.

• Corrected Model SS computed as Corrected Total - Error (not sum of Type III SS) to correctly handle unbalanced designs.

mariposa 0.4.0

New Functions

• Added fisher_test() for Fisher's exact test of independence in contingency tables. Recommended when sample sizes are small or expected cell frequencies fall below 5 (where chi-square approximation becomes unreliable). Supports survey weights, multi-variable analysis, and group_by().

• Added chisq_gof() for chi-square goodness-of-fit testing. Tests whether the observed frequency distribution of a categorical variable matches an expected distribution (default: equal proportions). Supports custom expected proportions, residual analysis, survey weights, and multi-variable analysis.

• Added mcnemar_test() for testing changes in paired proportions between two dichotomous measurements (e.g., before/after designs). Provides both asymptotic and exact binomial p-values, 2×2 contingency tables, and continuity correction. Supports survey weights.

• Added dunn_test() as an S3 generic for Dunn's post-hoc pairwise comparisons following a significant Kruskal-Wallis test. Identifies which specific group pairs differ using rank-based Z-statistics with adjustable p-value correction (Bonferroni, Holm, BH, etc.). Dispatches on kruskal_wallis result objects.

• Added pairwise_wilcoxon() as an S3 generic for pairwise Wilcoxon signed-rank post-hoc comparisons following a significant Friedman test. Identifies which measurement pairs differ with adjustable p-value correction. Dispatches on friedman_test result objects.

SPSS Validation

• Added SPSS validation tests for all 5 new functions across weighted/unweighted and grouped/ungrouped scenarios.

Improvements

• Extended post-hoc analysis framework: dunn_test() and pairwise_wilcoxon() join tukey_test(), scheffe_test(), and levene_test() as S3 generics that dispatch on their parent test result objects.

mariposa 0.3.1

Enhancements

• efa() now supports Maximum Likelihood (ML) extraction via extraction = "ml". ML extraction provides a goodness-of-fit chi-square test, initial communalities as SMC (squared multiple correlations), and uniquenesses. Uses stats::factanal() with correlation matrix input for seamless survey weight support.

• efa() now supports Promax rotation via rotation = "promax". Like Oblimin, Promax is an oblique rotation that produces Pattern Matrix, Structure Matrix, and Factor Correlation Matrix. Uses stats::promax() (base R, no new dependency).

• Internal refactoring of efa(): extraction logic separated into .efa_extract_pca() and .efa_extract_ml() for cleaner architecture and easier extension with future extraction methods (PAF planned).

mariposa 0.3.0

New Functions

• Added kruskal_wallis() for comparing 3+ independent groups on ordinal data (non-parametric alternative to one-way ANOVA). Supports survey weights, group_by(), and multi-variable analysis. Effect size: Eta-squared.

• Added wilcoxon_test() for comparing two paired measurements without assuming normality (Wilcoxon signed-rank test). Includes rank categories (negative, positive, ties) and effect size r.

• Added friedman_test() for comparing 3+ related measurements on ordinal data (non-parametric alternative to repeated-measures ANOVA). Effect size: Kendall's W.

• Added binomial_test() for testing whether an observed proportion matches an expected value (exact binomial test). Supports multiple binary variables and custom test proportions.

SPSS Validation

• Added 294 new SPSS validation tests across all 4 non-parametric functions, covering weighted/unweighted and grouped/ungrouped scenarios.

• Total test suite: 2,227 tests passing (0 failures, 0 skips).

mariposa 0.2.0

New Functions

• Added reliability() for Cronbach's Alpha with item statistics, including corrected item-total correlations, alpha-if-item-deleted, and inter-item correlation matrix. Genuine implementation with full survey weight support.

• Added efa() for Exploratory Factor Analysis with PCA extraction. Supports Varimax rotation (Base R) and Oblimin rotation (via optional GPArotation package). Includes KMO measure, Bartlett's test, communalities, and sorted factor loading matrix with configurable blank threshold.

• Added scale_index() for creating mean indices across survey items, with min_valid parameter matching SPSS MEAN.x() syntax. Designed for use inside dplyr::mutate().

• Added pomps() for Percent of Maximum Possible Scores transformation, rescaling values to a 0-100 range for cross-scale comparability.

• Added linear_regression() as a wrapper around stats::lm() with SPSS-compatible output: coefficients table (B, SE, Beta, t, p), ANOVA table, model summary (R, R-squared, adjusted R-squared), and standardized coefficients. Supports both formula and SPSS-style (dependent/predictors) interfaces.

• Added logistic_regression() as a wrapper around stats::glm() with odds ratios, Wald statistics, pseudo-R-squared measures (Nagelkerke, Cox-Snell, McFadden), and classification table.

Dependencies

• Added GPArotation as suggested dependency for Oblimin rotation in efa().

• Added MASS as suggested dependency for enhanced regression diagnostics.

Improvements

• All 6 new functions support survey weights and grouped analysis via dplyr::group_by().

• All functions include comprehensive roxygen2 documentation with practical examples, "When to Use" guidance, and "Understanding the Output" sections.

mariposa 0.1.0

Breaking Changes

• gamma() has been renamed to goodman_gamma() to avoid shadowing base::gamma(). The function remains an alias for chi_square() and works identically.

• S3 class names unified: removed _results suffix from all result classes (e.g., chi_square_results -> chi_square, t_test_results -> t_test). Class names now match the function name that created them.

Bug Fixes

• Fixed namespace collisions from triple-defined internal helper functions (.process_variables(), .process_weights(), .effective_n()). These are now defined once in helpers.R and shared across all functions.

• Fixed weighted variance/SD formula inconsistency. All weighted calculations now use the SPSS frequency weights formula: sum(w * (x - w_mean)^2) / (V1 - 1).

• Fixed Gamma ASE (asymptotic standard error) calculation. Replaced empirical magic-number formula with the correct ASE0 formula from Agresti (2002).

• Fixed weighted Cohen's d calculation in t_test(). Previously multiplied values by weights (x * w); now uses proper weighted means and pooled weighted standard deviation.

• Fixed weighted kurtosis formula. Changed from population excess kurtosis (m4/m2^2 - 3) to SPSS Type 2 sample-corrected formula (G2 = ((n+1)*g2 + 6) * (n-1) / ((n-2)*(n-3))), matching SPSS output.

Improvements

• Refactored 9 of 11 w_* functions to use a shared factory pattern (R/w_factory.R). Eliminated ~2,460 lines of duplicated boilerplate (4,204 → 1,740 lines, -58.6%). w_modus and w_quantile remain standalone due to their fundamentally different interfaces.

• Added 83 SPSS validation tests for all w_* functions across 4 scenarios (weighted/unweighted × grouped/ungrouped) in test-weighted-statistics-spss-validation.R.

• Reduced memory usage: result objects now store only the columns needed for post-hoc tests instead of the full input data frame.

• Unified print helper system: all output formatting now uses print_helpers.R. Removed deprecated .print_header(), .print_border(), .get_border(), and .print_group_header() from helpers.R.

• Deduplicated t_test.R print methods: print.t_test_results and print.t_test_result now share a common implementation (~360 fewer lines).

• Added input validation:

  • t_test(): validates conf.level is between 0 and 1
  • t_test(), oneway_anova(): validate that selected variables are numeric
  • chi_square(): warns when expected cell counts < 5

• Migrated error handling from stop()/warning() to cli_abort()/cli_warn() with structured messages, {.arg} and {.var} markup, and pluralization support.

• Added cli as dependency for professional user-facing messages and print output.

• Added @family tags to all 24 exported functions for cross-referencing in documentation (families: descriptive, hypothesis_tests, correlation, posthoc, weighted_statistics).

• Added tests/testthat/helper-mariposa.R with shared test utilities and centralized SPSS validation tolerances.

• Migrated print_helpers.R infrastructure to cli (cli_rule(), cli_bullets(), cli_h2()).

• Extended globals.R with missing NSE variable declarations.

• Fixed "SURVEYSTAT" reference in imports.R.