Skip to contents

Rake survey weights to marginal population totals

Source: R/rake.R
rake.Rd

Iterative proportional fitting (raking) that adjusts survey weights to match multiple marginal population totals simultaneously. Supports two algorithms: the "anesrake" method (chi-square variable selection, improvement-based convergence) and the "survey" method (fixed-order IPF, epsilon-based convergence).

Usage

rake(
  data,
  margins,
  weights = NULL,
  wt_name = "wts",
  type = c("prop", "count"),
  method = c("anesrake", "survey"),
  cap = NULL,
  control = list()
)

Arguments

data

A data.frame, weighted_df, survey_taylor, or survey_nonprob. survey_replicate → error. Any other class → error.

margins

Named list or data frame specifying population margin targets.

Format A — named list:

list(
  age_group = c("18-34" = 0.28, "35-54" = 0.37, "55+" = 0.35),
  sex       = c("M" = 0.49, "F" = 0.51)
)

Each element can be a named numeric vector or a data frame with columns level and target (formats can be mixed within the list).

Format B — long data frame with columns variable, level, target:

data.frame(
  variable = c("age_group", "age_group", "sex", "sex"),
  level    = c("18-34", "35-54", "M", "F"),
  target   = c(0.40, 0.60, 0.49, 0.51)
)

Format B is auto-detected and converted to Format A before use. The converted Format A is stored in the weighting history.

weights

<tidy-select> Weight column name (bare name). NULL → auto-detected from weighted_df attribute or survey object @variables$weights. For plain data.frame with weights = NULL, uniform starting weights are used and the output column is named by wt_name (default "wts").

wt_name

Character scalar. Name of the output weight column in the returned weighted_df. Default "wts". Ignored when data is a survey object (survey_taylor or survey_nonprob).

type

Character scalar. "prop" (default): margins values are proportions. "count": margins values are counts.

method

Character scalar. "anesrake" (default): chi-square discrepancy variable selection with improvement-based convergence, as in the anesrake package. "survey": fixed-order IPF cycling through all margins, with epsilon-based convergence, as in survey::rake().

cap

Numeric or NULL. Cap on the weight ratio w / mean(w). Any weight exceeding cap × mean(w) is set to cap × mean(w). Applied after each per-margin adjustment step (not post-hoc). NULL (default) means no cap. Applies to both methods.

control

Named list of algorithm parameters. Merged with method-specific defaults — omitted keys retain their defaults.

method = "anesrake" defaults:

  • maxit = 1000: maximum full sweeps

  • improvement = 0.01: percentage improvement convergence threshold

  • pval = 0.05: chi-square p-value threshold for variable selection

  • min_cell_n = 0L: minimum unweighted observations per cell (0 = no min)

  • variable_select = "total": chi-square aggregation for ranking ("total", "max", or "average")

method = "survey" defaults:

  • maxit = 100: maximum full sweeps

  • epsilon = 1e-7: maximum relative margin error convergence threshold

Passing anesrake-specific keys when method = "survey" (or vice versa) triggers a surveywts_warning_control_param_ignored warning per ignored parameter.

Value

  • data.frame or weighted_df input → weighted_df

  • survey_taylor or survey_nonprob input → same class as input (survey_taylor or survey_nonprob; class is preserved)

The weight column in the output contains raked weights. A history entry with operation = "raking" is appended to weighting_history.

Details

method = "anesrake": At each sweep, variables are sorted by their chi-square discrepancy (controlled by control$variable_select). Variables with any cell below control$min_cell_n unweighted observations are excluded entirely. Variables where the chi-square p-value exceeds control$pval are skipped in that sweep. Convergence is assessed as the percentage improvement in total chi-square between consecutive sweeps. If all variables pass or are excluded in sweep 1, a surveywts_message_already_calibrated message is emitted.

method = "survey": Variables are raked in the fixed order given by margins. All variables participate in every sweep. Convergence is assessed as the maximum relative error across all margin cells falling below control$epsilon.

See also

Other calibration: calibrate(), poststratify()

Examples

df <- data.frame(
  age_group = c("18-34", "35-54", "55+", "18-34", "35-54"),
  sex       = c("M", "F", "M", "F", "M"),
  stringsAsFactors = FALSE
)
margins <- list(
  age_group = c("18-34" = 0.30, "35-54" = 0.40, "55+" = 0.30),
  sex       = c("M" = 0.48, "F" = 0.52)
)
result <- rake(df, margins = margins)