Skip to contents

Iterative proportional fitting (raking)

rake.Rd

Adjusts survey weights so that weighted marginal distributions match known population targets. Supports automatic variable selection, iterative re-raking, and weight bounding.

Usage

rake(
  data,
  targets,
  base_weights = NULL,
  cap = 5,
  bounds = NULL,
  type = c("nolim", "pctlim", "nlim"),
  pctlim = 0.05,
  nlim = 5L,
  choosemethod = c("total", "max", "average", "totalsquared", "maxsquared",
    "averagesquared"),
  na_method = c("exclude", "bucket"),
  iterate = TRUE,
  max_iter = 1000L,
  tol = 1e-06,
  verbose = FALSE,
  diagnostics_every = 0L
)

Arguments

data

A data frame or tibble containing the survey data.

targets

A named list of named numeric vectors specifying target proportions for each raking variable. Names of the list must match column names in data. Each vector's names must match the levels of the corresponding variable. Values should sum to 1 (proportions); if not, they are normalized with a warning.

base_weights

Optional numeric vector of base (design) weights. If NULL (default), uniform weights of 1 are used. Centered to mean 1 before raking.

cap

Maximum weight value (ratio cap). Weights exceeding this value are trimmed and all weights are renormalized. Default 5. Ignored if bounds is specified.

bounds

Optional numeric vector of length 2, c(lo, hi), specifying minimum and maximum weight bounds. Overrides cap.

type

Variable selection method:

  • "nolim" (default): use all variables in targets.

  • "pctlim": use only variables with discrepancy >= pctlim.

  • "nlim": use the nlim most discrepant variables.

pctlim

Discrepancy threshold for type = "pctlim". Default 0.05 (5 percentage points).

nlim

Number of variables for type = "nlim". Default 5.

choosemethod

Method for aggregating per-category discrepancies into a single variable score. One of "total", "max", "average", "totalsquared", "maxsquared", "averagesquared".

na_method

How to handle NA values in raking variables. "exclude" (default): targets are proportions among non-NA cases only; NA cases are invisible to that margin. Matches anesrake. "bucket": NAs become a frozen extra category; their total weight is preserved and the named targets are rescaled to the remaining non-NA mass.

iterate

Logical. If TRUE and type = "pctlim", re-check discrepancies after raking and add newly discrepant variables, repeating up to 10 times. Default TRUE.

max_iter

Maximum number of raking iterations. Default 1000.

tol

Convergence tolerance (max proportional error). Default 1e-6.

verbose

Logical. If TRUE, print iteration progress. Default FALSE.

diagnostics_every

Record per-margin diagnostics every k iterations. 0 means only baseline. Default 0.

Value

An ipf_rake object (S3 class) containing:

  • weights: final raked weight vector

  • data: the input data frame

  • converged: logical

  • iterations: number of iterations

  • max_prop_err: final max proportional error

  • targets: normalized targets used

  • vars_used: character vector of variables raked on

  • base_weights: original base weights

  • type, choosemethod, na_method, cap: settings used

  • deff, n_eff: design effect and effective sample size

  • diagnostics: tibble of per-iteration diagnostics

Examples

data <- data.frame(
  gender = sample(c('M', 'F'), 100, replace = TRUE, prob = c(0.6, 0.4)),
  age = sample(c('young', 'old'), 100, replace = TRUE, prob = c(0.7, 0.3))
)
targets <- list(
  gender = c(M = 0.5, F = 0.5),
  age = c(young = 0.6, old = 0.4)
)
result <- rake(data, targets)
print(result)
#> 
#> ── Raking result (ipf) 
#> Converged: Yes (2 iterations, max prop err = 3.13e-08)
#> Variables raked: "gender" and "age"
#> Missing handling: "exclude"
#> Design effect: 1.074 | Effective n: 93 / 100
#> Weight range: [0.735, 1.55] | Mean: 1 | SD: 0.272