weightflow 0.2.0

New features

• Tidy population totals for step_calibrate(). In addition to the classic margins/totals inputs (which keep working unchanged), calibration targets can now be given as tidy data frames, paired with the new count argument that names the counts column:
  • Post-stratification: a data frame with one or more category columns plus a counts column. Several category columns are crossed automatically, so there is no need to build a collapsed cell variable by hand.
  • Raking: a list of data frames, one per margin.
  • Linear/GREG: a named list matching the formula terms, with a data frame (all categories) for each factor and a single number for each continuous total; weightflow builds the model.matrix totals internally, so the user never drops a reference category or handles the intercept. Calibration also reports clearer diagnostics and warnings: post-stratification flags cells in the sample but missing from the totals (error) or in the totals but absent from the sample (warning); raking warns on mutually inconsistent margins or non-convergence; linear calibration warns when the constraints are not fully satisfied; and calibration variables with missing values raise an informative error.
• Subsampling of more than one person per household in step_select_within(). A new n_selected argument (a single number or an unquoted column) works alongside n_eligible for simple random selection of a subsample: the weight is multiplied by n_eligible / n_selected (equivalent to prob = n_selected/n_eligible). It defaults to 1, so selecting a single person keeps working unchanged.
• External consistency totals for step_model_calibration(). The totals of the x_formula auxiliaries can now be supplied through the new x_totals argument, in the same two shapes as step_calibrate(method = "linear"): the tidy format (a named list with a data frame per factor, paired with count, and a single number per continuous total) or the classic model-matrix vector. This covers the common case where the X control totals come from an external source rather than from the frame; the auxiliaries then need to be present only in the sample, not in population. When x_totals is NULL (default) the X totals are still taken from population, so existing code is unchanged. population remains required, because the model-assisted block predicts each outcome over every population unit. Model calibration now also warns, like linear calibration, when the achieved totals do not fully satisfy the constraints (collinear or ill-conditioned auxiliaries).
• Delete-a-PSU jackknife variance (recipe-aware). jackknife_weights() builds jackknife replicate weights by deleting one PSU at a time and re-running the whole recipe on each replicate, so the replicate weights carry the variability of every adjustment. It is the stratified jackknife (JKn) with strata/psu, the unstratified jackknife (JK1) with strata = NULL, and the delete-one-unit jackknife with psu = NULL. jackknife_estimate() (plus jack_total() / jack_mean()) summarise a statistic with the JKn variance and match survey’s replicate jackknife for totals. as_svrepdesign() now also accepts a jackknife object, so the recipe-aware replicate weights flow into survey/srvyr for any estimand and any domain.
• Domain (partitioned) calibration in step_calibrate(). A new by argument names a domain (partition) column; the weights are then calibrated independently within each domain, each to its own totals (partitioned / domain calibration). The tidy totals carry the domain as a column, and a continuous total becomes a data frame domain, value (one total per domain); the domain variable does not go in the formula/margins. It composes with calfun, bounds, penalty and equal_within_cluster, applied within each domain, and reproduces every domain’s benchmarks. by = NULL (default) calibrates globally, unchanged.
• Exponential (raking) distance for step_calibrate(method = "linear"). calfun now also accepts "raking" (the multiplicative distance g = exp(u)), next to "linear" and "logit". It keeps the calibration weights positive without needing explicit bounds and still satisfies the constraints exactly, and works on mixed categorical and continuous auxiliaries as well as with the integrative option (equal_within_cluster, one weight per cluster). Matches survey::calibrate(calfun = "raking").
• R-indicator of response representativity (automatic diagnostic). When the recipe includes a nonresponse adjustment, summary() and report_weighting() now report the R-indicator (Schouten, Cobben & Bethlehem), R = 1 - 2*S, with S the design-weighted standard deviation of the estimated response propensities over the eligible sample: closer to 1 means a more representative response and less nonresponse-bias risk. The report also shows the unconditional partial R-indicators by auxiliary, pointing to which variable drives the lack of representativity. It is computed on the auxiliaries of the nonresponse step and needs no new function or user action; recipes without a nonresponse step are unaffected.
• Machine-learning response propensities (CART, random forest and gradient boosting via xgboost) for step_nonresponse() and step_model_calibration().
• k-fold cross-fitting (crossfit) to estimate each unit out-of-sample, with folds formed by cluster to avoid leakage.
• Ridge (penalized) calibration (penalty) to keep weights stable with many auxiliaries.
• Potter MSE-optimal trimming (method = "potter"), a data-driven cutoff.
• Quality alerts in prep(). prep() now computes non-fatal quality alerts and stores them on the prepped object ($alerts) and per step: negative or sub-1 weights and g-factors outside the Deville-Särndal bounds [0.1, 10] after calibration, small adjustment cells (new min_cell_n, default 30, following Kalton and Flores-Cervantes 2003) and excessive adjustment factors (new max_factor, default 2.5). Alerts always appear in the HTML report; set prep(warn = TRUE) to also raise them as R warnings.
• Weight distribution and alerts in report_weighting(). The report gains a “Weight distribution (final)” summary (min, p1, median, p99, max, max/min ratio, and counts of negative, sub-1 and extreme weights) and a per-step “Quality alerts” block.
• New disposition column in the sample_one example data. A single factor with the full field disposition (eligible respondent, eligible nonrespondent, household nonresponse, ineligible, unknown eligibility), recoded from the existing indicator columns (which are kept). It gives a tidy single-column view of the dispositions and can be used directly via logical conditions in the steps.
• New vignette “Preparing the sample: eligibility and response before weighting”, on how the input sample should be classified (the disposition tree), how it is sized (eligibility and response inflation) and how the dispositions map to the adjustment steps.

Bug fixes

• The optional machine-learning engines (engine = "forest" via ranger, engine = "boost" via xgboost) now run single-threaded by default, for reproducibility and to respect CRAN’s check limits. Set options(weightflow.num_threads = n) to use n threads.

• report_weighting() now flags calibration steps that did not converge. When a raking, linear or bounded calibration stops without satisfying the requested totals (the same condition that already prints a console warning), the HTML report shows a “Did not converge” alert on that step and no longer states that the step converged. Previously the report always reported convergence, regardless of the actual result.

• step_calibrate(equal_within_cluster = TRUE) now implements the genuine Lemaitre-Dufour (1987) integrative method: each unit’s auxiliaries are replaced by their household mean before a person-level calibration, so the per-household penalty scales with household size. This matches survey’s calibrate(aggregate.stage = ) (Vanderhoeft 2001), ReGenesees and Statistics Canada’s GES. The previous implementation used a household-level distance (summed auxiliaries, uniform per-household penalty), a different (non-standard) method. Integrative-calibration weights will change; totals are still met exactly and weights remain constant within household.

weightflow 0.1.0

First release.

A dependency-free, pipeable API to compute survey weights from design base weights through a chain of hierarchical adjustment stages. Build a recipe lazily, estimate it with prep(), and extract the weights with collect_weights(). Separating define from apply makes the whole process reproducible and auditable, and lets the bootstrap re-run the entire cascade on each replicate.

Adjustment steps

• step_unknown_eligibility(): redistribute the weight of unknown-eligibility cases to the known ones (person- or household-level via cluster).
• step_drop_ineligible(): zero out out-of-scope units.
• step_select_within(): within-household selection (unequal prob or equal n_eligible).
• step_nonresponse(): weighting-class or propensity adjustment, at the person or household level (cluster).
• step_calibrate(): raking, post-stratification and linear/GREG calibration, with bounded (Deville-Särndal) and integrative (one weight per household) cluster options.
• step_model_calibration(): Wu-Sitter model calibration.
• step_trim(), step_trim_weights(), step_round(), step_rescale(): trimming, rounding and rescaling.
• step_assert(): quality checkpoint (deff, weight ratio, effective n).

Inspection and reporting

• summary(), plot() and weight_factors() for per-stage diagnostics.
• design_effect() for the Kish design effect and effective sample size.
• report_weighting() builds a self-contained HTML report with a pipeline diagram, the variables used, per-stage summaries and per-step visuals.

Variance estimation

• bootstrap_weights() resamples PSUs within strata (Rao-Wu rescaling) and re-applies the whole recipe on each replicate, so the replicate weights carry the variability of every adjustment.
• boot_mean() and boot_total() return the estimate, standard error and CI.
• as_svydesign(), as_svrepdesign() and collect_replicate_weights() bridge to the survey and srvyr packages for design-based inference.

Data

• Bundled example datasets population, sample_survey (take-all roster) and sample_one (multistage select-one design), all with stratum, PSU and design weight.