v1.1.0

New features

• coef(), vcov(), SE(), and confint() methods are now available for all
  survey_result objects produced by get_*() functions. The .survey_result
  attribute carrying degrees of freedom and parameter metadata is now wired into
  every estimation function. (#148, #149)

Bug fixes

• as_survey_replicate() now uses the correct default scale factor for "ACS"
  and "successive-difference" replicate types (4/R per Ash (2014) / Fay &
  Train (1995)), fixing underestimated standard errors for these designs. (#150)

surveycore 1.0.0

First stable release

surveycore 1.0.0 is the first stable API release. The full set of survey
design types (Taylor series linearization, replicate weights, two-phase,
non-probability), variance estimators, analysis functions (means, totals,
frequencies, quantiles, ratios, correlations, regression, t-tests, ANOVA,
effective sample size), the metadata system, and calibration-adjusted variance
are now complete and stable.

New features

• as_survey_nonprob() now accepts jackknife replicate schemes via
  type = "JK1", "JK2", and "JKn". All get_*() estimation functions
  dispatch to jackknife variance when a survey_nonprob design carries
  jackknife replicate weights. print() displays the replicate type and count
  for jackknife designs. (#133–#136)

• survey_glm() now routes survey_nonprob designs that carry bootstrap
  replicate weights through the replicate-weight variance estimator, matching
  the behaviour of all other get_*() functions for such designs. (#138)

• New dataset ca_api_2000: the 2000 California Academic Performance Index
  (API) survey, structured as a simple random sample from California schools.
  Used as a primary example for survey_taylor SRS designs throughout the
  documentation and vignettes. (#146)

Bug fixes

• get_diffs(): fixed a regex metacharacter injection when variable names
  contained special characters (., *, ?, etc.), and fixed pct_change
  returning NA for some group configurations. (#145)

• Fixed six print/dispatch/error bugs (B1–B6) uncovered during a documentation
  audit: print.survey_nonprob no longer errors on designs with no repweights
  set; dispatch errors now surface the correct class name in message text; three
  error classes that were raised but never registered have been registered.
  (#143)

Documentation

• Corrected roxygen text, stale cross-references, contradictory @tags, and
  mismatched @section headings across 40+ R files (D1–D75). (#143–#144)

• The non-probability estimation vignette section has been rewritten with a
  complete bootstrap variance example using as_survey_nonprob() with
  repweights.

v0.9.0

New features

• set_higher_is() and extract_higher_is() store and retrieve a
  direction-of-improvement attribute ("better" or "worse") for survey
  variables. The attribute is used by get_diffs(show_favorability = TRUE)
  to classify differences as favorable or backlash. Supports all three
  calling conventions (named ... args, named vector,
  variable = + direction =), data frames, and tidy-select.

• set_reverse_coded() and extract_reverse_coded() store and retrieve a
  reverse-coded flag for survey variables. Supports tidy-select ... (bare
  names and selection helpers) and the variable = character-vector
  interface. Setting reverse_coded = FALSE removes the flag. Works on both
  survey design objects and plain data frames.

• get_diffs() gains alpha and show_favorability arguments.
  show_favorability = TRUE appends favorable and backlash logical
  columns to the result. A difference is classified as favorable when it is
  statistically significant (p < alpha, default 0.05) and in the
  direction indicated by higher_is metadata set via set_higher_is().
  When no higher_is metadata is set, both columns are all FALSE. The
  adjusted p-value (when pval_adj is supplied) is used for classification.

• as_survey_nonprob() gains replicate-weight (bootstrap) support via six
  new arguments: repweights, type, scale, rscales, mse, and
  reference_sample. When repweights is supplied, the constructor resolves
  and validates the replicate weight columns and stores them in @variables.
  A reference_sample property is also added to the survey_nonprob class
  for provenance tracking.

• All get_*() estimation functions now dispatch to the replicate-weight
  variance engine when a survey_nonprob object has repweights set.
  When no repweights are supplied, a warning is emitted and the SRS
  approximation is used as a fallback. Degrees of freedom for
  survey_nonprob estimates always use the normal approximation (Inf).

• print() for survey_nonprob now shows bootstrap replicate information
  (replicate count, type, and scale) when repweights are set.

surveycore 0.8.4

New features

• get_effective_n() computes the effective sample size of a survey design
  using the Kish (1965) weight approximation (method = "kish") or the full
  design effect for a specified variable (method = "deff"). Supports all
  design types and survey_collection.

CRAN preparation (0.8.3)

• Resubmission addressing CRAN feedback on the 0.8.2 submission. Four
  changes: (1) Added DOI references to the package Description
  (Lumley 2004 for variance estimation; Mannan 2025 for weighted
  polychoric/polyserial correlation). (2) Uncommented three
  as_survey() calls in the @examples blocks of anes_2024,
  gss_2024, and pew_npors_2025, so they now run during
  R CMD check. The ANES and GSS examples (and the corresponding
  prose @details sketches) also gained nest = TRUE, which is
  required by those designs and silences a previously-emitted
  warning. (3) Replaced the \section{Tidy-select} \preformatted{}
  code sketch in as_survey() with prose, and added two runnable
  demonstrations to the @examples block. (4) Replaced globalenv()
  with baseenv() as the formula environment in survey_glm() to
  comply with CRAN policy on .GlobalEnv modification.

surveycore 0.8.0

Breaking changes

• Constructing a survey_collection from member surveys with divergent @groups now errors surveycore_error_collection_group_divergent. Previously, a mixed-grouping collection would dispatch analysis functions per-survey and stitch a patchwork of grouped and ungrouped rows together with bind_rows() — violating the pseudo-data.frame mental model. All members must either share @groups or the caller must supply group = explicitly.
• as_survey_collection()'s .on_missing argument has been replaced by .if_missing_var, and the previously silent no-op behaviour is fixed. .if_missing_var is now stored on the returned collection's @if_missing_var property and is honoured (rather than ignored) by every dispatched get_*(). Callers using the old name will see R's positional-argument-mismatch error.
• The .on_missing named-only argument on every collection-dispatching get_*() (get_means(), get_totals(), get_freqs(), get_ratios(), get_diffs(), get_corr(), get_variance(), get_quantiles(), get_covariance(), get_t_test(), get_pairwise()) has been renamed to .if_missing_var. The default flips from "error" to NULL; NULL resolves to the collection's stored @if_missing_var property, while a non-NULL value overrides it for that call. The .id argument similarly defaults to NULL and resolves to the collection's stored @id. Callers passing .on_missing = ... will silently have the value flow into ... (no behaviour change at the analysis layer); update to .if_missing_var = ... to restore intent.

New features

survey_collection per-call dispatch defaults

• survey_collection gains two new properties:
  • @id (character(1), default ".survey") — column name .dispatch_over_collection() uses when an analysis function is dispatched across the collection without an explicit per-call .id. Validated via the new shared helper; the existing surveycore_error_collection_invalid_id class fires on bad input.
  • @if_missing_var (character(1), default "error", must be one of c("error", "skip")) — controls how dispatched get_*() calls behave when a member survey is missing a requested variable. Validated via the new helper; raises the new surveycore_error_collection_invalid_if_missing_var error class on bad input.
• New exported setters set_collection_id(x, id) and set_collection_if_missing_var(x, if_missing_var) mutate the corresponding property and return the collection invisibly. Both validate via the same shared helpers; both raise surveycore_error_not_survey_collection on non-collection input.
• add_survey() and remove_survey() now propagate the source collection's @id and @if_missing_var onto the returned collection.
• print(survey_collection) renders id: and if_missing_var: lines on every print, regardless of whether they hold the default values.
• .dispatch_over_collection() resolves both .id and .if_missing_var via two-tier precedence: a non-NULL value at the analysis-function call site beats the value stored on the collection's property. The surveycore_error_collection_id_collision hint additionally surfaces set_collection_id() as a fix path when the collision was triggered by the stored @id.

Uniform grouping on survey_collection

• survey_collection gains a @groups property (character(0) by default). Every member survey's @groups is asserted identical() to the collection's value by the class validator — a uniform-grouping invariant that guarantees dispatched get_*() results share a single grouping structure.
• as_survey_collection() gains a group = argument that accepts tidy-select column names (bare, c(), all_of()). Missing or empty-resolved group = (including NULL, character(0), c(), all_of(character(0))) adopts the members' uniform @groups or errors on divergence; a supplied non-empty group = overrides any pre-existing member @groups and emits a typed surveycore_warning_collection_group_overridden per divergent member.
• add_survey() and remove_survey() now preserve coll@groups across mutation: a grouped collection propagates its @groups onto any empty-grouped new member and errors on divergent-grouped members (surveycore_error_collection_group_conflict); removal keeps the collection-level grouping.

Polychoric and polyserial correlation via get_corr(method = ...)

• get_corr() gains a method = "pearson" argument. Setting method = "polychoric" fits a weighted two-step MLE for the correlation between two ordinal variables under a bivariate-normal latent model (Olsson 1979; Mannan 2025); method = "polyserial" fits the analogous MLE for one ordinal + one continuous variable (Cox 1974). Auto-detection of the ordinal / continuous side is handled internally; no new user-facing argument is required. Confidence intervals are constructed on the Fisher-z scale and back-transformed to [-1, 1]. Variance is design-based: Taylor linearization via a perturbation-based influence function on survey_taylor, and a full per-replicate re-fit of both thresholds and rho on survey_replicate. For method != "pearson", df = NA_integer_ and statistic is the z-scale Wald statistic referred to a standard normal distribution. meta(result)$bivariate_normal_cdf is "pbivnorm", and meta(result)$n_failed_replicates_total carries the total count of non-converged replicates when the replicate path observed any. Agreement with polycor::polychor() / polycor::polyserial() on equal-weight fixtures is within 1e-4.
• New package Import: pbivnorm (>= 0.6.0), used as the bivariate-normal CDF for the polychoric / polyserial likelihood.
• Fourteen new typed error / warning classes (PC-1 through PC-14) surface ordinal-type, optimizer, sparse-cell, boundary, and replicate-convergence conditions — see plans/error-messages.md for the full list.

New functions

• get_variance() computes design-based finite-population variance estimates for one or more numeric variables in a survey design, matching survey::svyvar() at tolerance 1e-10 on point estimates and 1e-8 on SEs. Returns a survey_variance tibble with point estimate, SE, CI, CV, MOE, design effect (deff), and cell sizes. Supports grouping (via group = and group_by()), per-variable na_handling = "pairwise" (default) or "listwise", name_style = "broom" renaming, and column-level label attributes for downstream gt integration. Dispatches over survey_taylor, survey_replicate, survey_twophase, survey_nonprob, and survey_collection designs.
• get_covariance() computes design-based finite-population covariance estimates for all unordered pairs drawn from one or more numeric variables in a survey design, matching the off-diagonal entries of survey::svyvar() at tolerance 1e-10 on point estimates and 1e-8 on SEs. Returns a survey_covariance tibble with covariance, SE, CI, CV, MOE, design effect (deff), and pairwise cell sizes. Pearson-only, pairwise-complete NA handling. Supports grouping (via group = and group_by()), redundant = TRUE to include both (x, y) and (y, x) orderings, diagonal = TRUE to include (x, x) self-pairs (which equal get_variance(x) exactly at 1e-10), name_style = "broom" renaming, and column-level label attributes for downstream gt integration. Dispatches over survey_taylor, survey_replicate, survey_twophase, survey_nonprob, and survey_collection designs.

New warning classes

• surveycore_warning_variance_all_na — fired when every row of the active domain is NA on the focal variable.
• surveycore_warning_variance_insufficient_n — fired when the focal variable has fewer than two non-NA observations in the active domain (variance is undefined).
• surveycore_warning_covariance_all_na — fired when every row of the active domain is NA on at least one variable in the pair.
• surveycore_warning_covariance_insufficient_n — fired when a pair has fewer than two pairwise-complete observations in the active domain (covariance is undefined).
• surveycore_warning_covariance_non_numeric — fired when one or more variables passed via x are non-numeric and silently dropped from the pair list.

v0.7.0

Breaking changes

• get_anova()'s first argument is now object and dispatches on class. The former model2 positional argument has been removed — get_anova(fit1, fit2) must now be written get_anova(list(fit1, fit2)). The S3 anova(fit1, fit2) interface is unchanged.

New functions

Design-based group comparisons

• get_t_test() performs a design-based two-sample t-test comparing group means for a numeric outcome across two levels of a by variable. Returns a survey_t_test tibble with estimate, per-group means and cell sizes, CI, t-statistic, df, p-value, and significance stars. Supports optional stratification via group (one row per stratum) and matches survey::svyttest() at tolerance 1e-10 for point estimates and test statistics.
• get_pairwise() computes all k(k−1)/2 pairwise t-tests across the levels of a factor, with multiple-comparison p-value adjustment via any stats::p.adjust() method ("holm" by default, or "none"). Adjustment is applied separately within each group stratum when stratified. Returns a survey_pairwise tibble with one row per pair.

Design-based ANOVA

• get_anova() computes Rao-Scott design-based ANOVA for survey_glm_fit objects, supporting both Wald and LRT tests with F or Chi-squared reference distributions. Three dispatch branches:
  • get_anova(<survey_glm_fit>) — sequential term-by-term anova (matches anova.svyglm() semantics).
  • get_anova(<list<survey_glm_fit>>) — chained pairwise comparison across k nested fits, returning k − 1 rows.
  • get_anova(<survey_base>, formula = ...) — fits the model internally via survey_glm() and runs sequential anova on the fit; extra ... are forwarded to survey_glm().
    Matches survey::regTermTest() at tolerance 1e-8 on statistics and 1e-6 on p-values.
• anova(fit) on a survey_glm_fit now dispatches to get_anova() via a registered S3 method.
• plot() on a survey_glm_fit produces a dot-and-whisker coefficient plot with design-based Wald confidence intervals.

Select-all-that-apply (SATA) metadata

• set_sata() marks one or more variables on a survey design (or data frame) as select-all-that-apply. Accepts either tidy-select ... or a variable character vector; setting sata = FALSE removes the flag. Idempotent on already-flagged variables.
• extract_sata() returns SATA status as a named logical vector (default), a list, or a data frame. fill = FALSE yields a dense view (unmarked variables reported as FALSE); fill = NULL returns only flagged variables.
• classify_question_type() classifies a set of requested variables into "single", "sata", or "battery" by grouping them on shared question_preface metadata and honoring per-variable SATA flags. Group numbers are assigned in order of first appearance. Warns when a lone SATA-flagged variable has no preface mate, or when a preface group has mixed SATA flags.

Survey collections

• survey_collection is a new S7 container holding an ordered, uniquely-named list of survey_base objects — useful for wave-to-wave analyses, panel studies, or any workflow that compares estimates across multiple designs.
• as_survey_collection() constructs a collection from named (wave1 = d1, wave2 = d2) or bare (d1, d2) arguments; duplicate names are repaired by appending _1, _2, … with a warning showing the rename mapping.
• add_survey() and remove_survey() return new collections with surveys appended or removed; the original is unchanged.
• All nine get_*() analysis functions (get_means(), get_totals(), get_freqs(), get_quantiles(), get_ratios(), get_corr(), get_diffs(), get_t_test(), get_pairwise()) now dispatch over a survey_collection, iterating across surveys and returning a single combined tibble. Two new named-only control args on each function: .id = ".survey" names the identifier column, and .on_missing = c("error", "skip") controls behavior when a requested variable is absent from a survey. Regression functions (survey_glm(), get_anova()) do not support collection dispatch and raise an explicit error pointing users to lapply().

Other improvements

• survey_glm() gains a quiet = argument to suppress convergence warnings.
• extract_*() metadata functions now accept tidyselect helpers (starts_with(), all_of(), any_of(), matches()) in place of bare name lists.

Bug fixes

• get_diffs() now correctly computes pct_change when show_means = FALSE is combined with grouped marginal effects and show_pct_change = TRUE (previously returned NA).

surveycore 0.6.0

Breaking changes

• survey_srs class and as_survey_srs() constructor have been removed. SRS
  designs are now created via as_survey() with no ids or strata — this
  produces a survey_taylor with no cluster/strata structure. All estimates are
  numerically identical.

New features

• get_diffs() estimates treatment effects (differences from a reference group)
  via survey-weighted regression. Supports bivariate and multivariate models,
  Gaussian and non-Gaussian families, and optional subgroup analysis. Two
  estimation paths: direct coefficients for simple models, and
  marginaleffects::avg_slopes() / avg_predictions() for models with
  covariates or non-Gaussian AMEs. Returns a survey_diffs tibble with optional
  mean, pct_change, n_weighted columns, significance stars, and p-value
  adjustment. marginaleffects moved from Suggests to Imports.

• as_survey() now supports multi-column FPC for multi-stage designs
  (e.g., fpc = c(fpc_stage1, fpc_stage2)). Each FPC column corresponds to one
  ID stage. Per-stage FPC is validated for NAs, non-positive values, and
  within-cluster constancy.

• print() for survey_taylor now displays per-stage FPC bullets for
  multi-stage designs (e.g., FPC (stage 1): fpc, FPC (stage 2): fpc2).

Bug fixes

• SRS variance estimation now uses Taylor (HT) linearization via
  .build_cluster_matrices(), correct for any weight structure. Previously used
  unweighted sample variance which was incorrect for non-proportional weights.

• survey_glm() now correctly indexes weights when na.action = na.omit drops
  non-contiguous rows.

• get_freqs() now routes survey_nonprob designs through the
  Horvitz-Thompson variance path, consistent with the other five analysis
  functions.

• as_survey_twophase() now accepts survey_replicate and SRS
  survey_taylor objects as the phase-1 design (previously restricted to
  stratified/clustered survey_taylor only).

• as_survey() SRS fallback downgraded from warning to message.

Internal infrastructure

• .build_cluster_matrices() extracts multi-stage cluster, strata, and FPC
  matrix construction into a shared helper, used across the Taylor variance
  engine, analysis cell estimators, and GLM sandwich variance.

surveycore v0.5.0

Breaking changes

• as_survey_replicate() replaces as_survey_repweights(). The constructor
  name now matches the underlying survey_replicate class.

• survey_nonprob and as_survey_nonprob() replace survey_calibrated and
  as_survey_calibrated(). "Calibrated" implies a post-processing step on a
  probability sample; nonprob accurately reflects the design type.

• The positional setter form set_var_label(svy, age, "label") has been
  removed. Use the named form set_var_label(svy, age = "label") instead.

• extract_var_label(), extract_question_preface(), and extract_var_note()
  now return a named character vector. extract_var_label(svy, age) now
  returns c(age = "Age in years") rather than "Age in years".

• extract_val_labels() now returns a named list. extract_val_labels(svy, sex)
  now returns list(sex = c(Male = 1L, Female = 2L)) rather than
  c(Male = 1L, Female = 2L).

• set_variable_labels(), set_value_labels(), set_question_prefaces(), and
  set_variable_notes() have been removed. Use set_var_label(),
  set_val_labels(), set_question_preface(), and set_var_note()
  respectively — all four now accept multiple variables via named ....

New features

• set_universe() and extract_universe() set and retrieve universe
  (eligibility) annotations for survey variables.

• set_missing_codes() and extract_missing_codes() set and retrieve missing
  value code vectors for survey variables.

• extract_metadata() returns all metadata fields (variable_label,
  value_labels, question_preface, note, universe, missing_codes,
  transformations) for one or more variables as a named list.

Enhancements

• All setter functions now support three call conventions: named ...
  (e.g., set_var_label(svy, age = "Age in years")), a single named
  vector/list in ..., or explicit variable = / content-argument pairs.
  All setters also now work on plain data.frames.

• All extractor functions accept multiple variables via ..., support three
  output formats ("named_vector", "list", "data_frame"), and accept a
  fill argument to include variables with no metadata in the output.

surveycore v0.4.0

New features

• survey_glm() fits survey-weighted generalized linear models for all five
  design classes (survey_taylor, survey_replicate, survey_srs,
  survey_twophase, survey_calibrated); returns a survey_glm_fit object
  with design-based (Binder 1983 sandwich) standard errors and degrees of
  freedom.

• clean() converts a survey_glm_fit to a tidy survey_glm_tidy tibble
  with one row per coefficient, design-based confidence intervals, structured
  metadata, and optional reference rows for factor predictors.

• survey_glm_fit objects support 20 S3 methods: print(), summary(),
  coef(), vcov(), predict(), fitted(), residuals(), confint(),
  formula(), terms(), model.matrix(), model.frame(), deviance(),
  df.residual(), nobs(), hatvalues(), logLik(), AIC(), BIC(), and
  update().

• survey_glm_fit integrates with the marginaleffects package; when
  marginaleffects is installed, avg_slopes(), avg_predictions(), and the
  full marginaleffects API work directly on survey_glm_fit objects.

• broom::tidy() is supported for survey_glm_fit objects via a shim that
  delegates to clean().

• as_survey_rep() has been renamed to as_survey_repweights() to avoid a
  namespace clash with the srvyr package.

Bug fixes

• as_survey_twophase() variance estimation (method = "approx" and
  "full") now uses the correct PSU-level Phase 2 stratum sampling fraction
  instead of a row-level fraction, resolving an approximately 2× variance
  underestimation.

v0.3.3

New features

• print() methods for all five survey design classes (survey_taylor, survey_srs, survey_replicate, survey_twophase, survey_calibrated) now display a Domain: <n> of <N> rows line when surveytidy::filter() has been applied. The line appears after the sample size line and before the Groups: line. For two-phase designs, domain counts reflect Phase 2 rows only.