Getting Started with ipf
ipf.RmdOverview
Survey samples rarely match the population perfectly on key
demographics. Raking (iterative proportional fitting)
adjusts case weights so that the weighted sample margins match known
population targets. ipf makes this fast by using Rust to
perform all computations.
This vignette walks through a complete raking workflow using the
bundled anes24 dataset, which is taken from a subset of the
2024 American National Election Study https://electionstudies.org/data-center/2024-time-series-study/.
Setup
The data contains 966 respondents from the ANES 2024 face-to-face sample:
anes24
#> # A tibble: 966 × 7
#> state sex race income education married presidential
#> <chr> <chr> <chr> <chr> <chr> <chr> <chr>
#> 1 CO Female White Over $100k NA Divorced Trump
#> 2 CA Male White Over $100k NA Married Harris
#> 3 NC Female White NA Bachelor's Married Harris
#> 4 MA Male NA Over $100k Some college Married Trump
#> 5 WA Female Black NA Bachelor's Married NA
#> 6 GA Male White $50k-$100k NA NA Trump
#> 7 AR Male Black Under $50k NA NA Harris
#> 8 TX Male Hispanic Under $50k Some college NA Trump
#> 9 OR Male White Under $50k NA Divorced NA
#> 10 AL Female White Under $50k NA Divorced NA
#> # ℹ 956 more rowsInspect data
Before writing targets, inspect the levels in your sample and see where values are missing:
table(anes24$sex, useNA = 'ifany')
#>
#> Female Male <NA>
#> 503 458 5
table(anes24$race, useNA = 'ifany')
#>
#> Asian Black Hispanic Other White <NA>
#> 28 118 114 63 632 11
table(anes24$income, useNA = 'ifany')
#>
#> $50k-$100k Over $100k Under $50k <NA>
#> 300 389 230 47When you define targets, the target names must match the data values
exactly. By default, NA values are ignored for that
variable during raking. If you use na_method = 'bucket',
ipf treats missing values as an implicit extra category and
preserves their total weight while rescaling the named targets to the
remaining nonmissing weight mass.
Define population targets
Targets are a named list of named numeric vectors. Each vector’s names must match the levels in your data, and the values should be proportions summing to 1.
targets <- list(
sex = c(Male = 0.472, Female = 0.528),
race = c(
White = 0.706,
Black = 0.121,
Hispanic = 0.107,
Asian = 0.047,
Other = 0.019
),
income = c(
'Under $50k' = 0.151,
'$50k-$100k' = 0.294,
'Over $100k' = 0.555
)
)If your targets don’t sum to 1, ipf will normalize them
automatically with a warning.
Rake
The main function is rake():
result <- rake(anes24, targets, cap = NULL)
result
#>
#> ── Raking result (ipf)
#> Converged: Yes (5 iterations, max prop err = 3.04e-07)
#> Variables raked: "sex", "race", and "income"
#> Missing handling: "exclude"
#> Design effect: 1.113 | Effective n: 868 / 966
#> Weight range: [0.199, 1.928] | Mean: 1 | SD: 0.337result is an ipf_rake object containing the
weights and diagnostics.
If you want missing values in a raking variable to act like their own
implicit category, use na_method = 'bucket':
bucketed <- rake(anes24, targets, cap = NULL, na_method = 'bucket')
bucketed
#>
#> ── Raking result (ipf)
#> Converged: Yes (6 iterations, max prop err = 5.46e-07)
#> Variables raked: "sex", "race", and "income"
#> Missing handling: "bucket"
#> Design effect: 1.114 | Effective n: 867 / 966
#> Weight range: [0.199, 1.936] | Mean: 1 | SD: 0.337If you already have design weights, pass them through
base_weights:
base_w <- ifelse(anes24$sex == 'Female', 1.1, 0.9)
base_w[is.na(base_w)] <- 1
base_weighted <- rake(anes24, targets, base_weights = base_w, cap = NULL)
base_weighted
#>
#> ── Raking result (ipf)
#> Converged: Yes (5 iterations, max prop err = 3.04e-07)
#> Variables raked: "sex", "race", and "income"
#> Missing handling: "exclude"
#> Design effect: 1.113 | Effective n: 868 / 966
#> Weight range: [0.199, 1.928] | Mean: 1 | SD: 0.337Inspect results
Design effect
The design effect measures how much the weighting inflates variance. A deff of 1.0 means no inflation (uniform weights). Higher values mean less effective data.
design_effect(result$weights)
#> $deff
#> [1] 1.113407
#>
#> $n_eff
#> [1] 867.6076Per-variable diagnostics
summary() shows a full diagnostic report:
summary(result)
#>
#> ── Raking Summary (ipf)
#> ────────────────────────────────────────────────────────────────────────────────
#> ✔ Converged in 5 iterations (max prop err = 3.04e-07)
#> ℹ No base weights (uniform)
#> ℹ Selection: type = "nolim", method = "total"
#> ℹ Missing handling: "exclude"
#> ℹ Variables raked: "sex", "race", and "income"
#> ── Weight Summary ──────────────────────────────────────────────────────────────
#> Min: 0.1993 Q1: 0.6688 Median: 0.9703 Mean: 1 Q3: 1.2794 Max: 1.9275
#> SD: 0.3368 CV: 0.3368
#> ── Design Effect ───────────────────────────────────────────────────────────────
#> Deff: 1.1134 | Effective n: 868 / 966
#> ── Per-Variable Assessment ─────────────────────────────────────────────────────
#>
#> ── sex
#> # A tibble: 3 × 9
#> level target unweighted_n unweighted_pct weighted_n weighted_pct change_pct
#> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
#> 1 Male 0.472 458 0.477 454. 0.472 -0.00459
#> 2 Female 0.528 503 0.523 508. 0.528 0.00459
#> 3 Total 1 961 1 962. 1.00 0.00917
#> # ℹ 2 more variables: residual_disc <dbl>, original_disc <dbl>
#>
#> ── race
#> # A tibble: 6 × 9
#> level target unweighted_n unweighted_pct weighted_n weighted_pct change_pct
#> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
#> 1 White 0.706 632 0.662 674. 0.706 0.0442
#> 2 Black 0.121 118 0.124 116. 0.121 -0.00256
#> 3 Hispanic 0.107 114 0.119 102. 0.107 -0.0124
#> 4 Asian 0.047 28 0.0293 44.9 0.0470 0.0177
#> 5 Other 0.019 63 0.0660 18.1 0.0190 -0.0470
#> 6 Total 1 955 1 955. 1.00 0.124
#> # ℹ 2 more variables: residual_disc <dbl>, original_disc <dbl>
#>
#> ── income
#> # A tibble: 4 × 9
#> level target unweighted_n unweighted_pct weighted_n weighted_pct change_pct
#> <chr> <dbl> <dbl> <dbl> <dbl> <dbl> <dbl>
#> 1 Under $… 0.151 230 0.250 139. 0.151 -0.0993
#> 2 $50k-$1… 0.294 300 0.326 270. 0.294 -0.0324
#> 3 Over $1… 0.555 389 0.423 509. 0.555 0.132
#> 4 Total 1 919 1 917. 1.00 0.263
#> # ℹ 2 more variables: residual_disc <dbl>, original_disc <dbl>The residual discrepancy column shows how close the weighted distribution is to the target.
Tidy output
For programmatic use, the broom-style methods return
tibbles:
# One row per variable-level
tidy(result)
#> # A tibble: 10 × 5
#> variable level target weighted_pct discrepancy
#> <chr> <chr> <dbl> <dbl> <dbl>
#> 1 sex Male 0.472 0.472 1.43e- 7
#> 2 sex Female 0.528 0.528 -1.43e- 7
#> 3 race White 0.706 0.706 -3.36e- 9
#> 4 race Black 0.121 0.121 6.71e- 9
#> 5 race Hispanic 0.107 0.107 -2.10e- 9
#> 6 race Asian 0.047 0.0470 -1.69e- 9
#> 7 race Other 0.019 0.0190 4.44e-10
#> 8 income Under $50k 0.151 0.151 3.05e-16
#> 9 income $50k-$100k 0.294 0.294 1.11e-16
#> 10 income Over $100k 0.555 0.555 -2.55e-15
# One-row summary
glance(result)
#> # A tibble: 1 × 7
#> converged iterations max_prop_err deff n_eff n_obs n_vars
#> <lgl> <int> <dbl> <dbl> <dbl> <int> <int>
#> 1 TRUE 5 0.000000304 1.11 868. 966 3Augmenting the data
To use the weights in downstream analyses, attach them to your data:
weighted_data <- augment(result)
head(weighted_data)
#> # A tibble: 6 × 8
#> state sex race income education married presidential .weight
#> <chr> <chr> <chr> <chr> <chr> <chr> <chr> <dbl>
#> 1 CO Female White Over $100k NA Divorced Trump 1.37
#> 2 CA Male White Over $100k NA Married Harris 1.28
#> 3 NC Female White NA Bachelor's Married Harris 1.07
#> 4 MA Male NA Over $100k Some college Married Trump 1.23
#> 5 WA Female Black NA Bachelor's Married NA 1.12
#> 6 GA Male White $50k-$100k NA NA Trump 0.908The .weight column can then be used in downstream
analyses.
For example, you can compare an estimate before and after weighting:
presidential_data <- subset(weighted_data, !is.na(presidential))
presidential_unweighted <- prop.table(table(presidential_data$presidential))
presidential_weighted <- aggregate(
.weight ~ presidential,
presidential_data,
sum
)
presidential_weighted$weighted_pct <- presidential_weighted$.weight /
sum(presidential_weighted$.weight)
presidential_compare <- tibble::tibble(
presidential = presidential_weighted$presidential,
unweighted_pct = as.numeric(presidential_unweighted[
presidential_weighted$presidential
]),
weighted_pct = presidential_weighted$weighted_pct
)
presidential_compare
#> # A tibble: 2 × 3
#> presidential unweighted_pct weighted_pct
#> <chr> <dbl> <dbl>
#> 1 Harris 0.564 0.575
#> 2 Trump 0.436 0.425Advanced options
Weight bounding
By default, weights are capped at 5. Tighter bounds reduce extreme weights but can leave more residual mismatch.
# Unbounded fit from above
range(result$weights)
#> [1] 0.1993456 1.9275117
design_effect(result$weights)
#> $deff
#> [1] 1.113407
#>
#> $n_eff
#> [1] 867.6076
# Default cap
default_bounded <- rake(anes24, targets)
range(default_bounded$weights)
#> [1] 0.1993456 1.9275117
design_effect(default_bounded$weights)
#> $deff
#> [1] 1.113407
#>
#> $n_eff
#> [1] 867.6076
# Tighter cap
tight <- rake(anes24, targets, cap = 3)
range(tight$weights)
#> [1] 0.1993456 1.9275117
design_effect(tight$weights)
#> $deff
#> [1] 1.113407
#>
#> $n_eff
#> [1] 867.6076
# Or specify both min and max bounds
bounded <- rake(anes24, targets, bounds = c(0.3, 3))
range(bounded$weights)
#> [1] 0.300000 1.950523Variable selection
With many potential raking variables, you can let ipf
select only the most discrepant ones:
targets_many <- list(
sex = c(Male = 0.472, Female = 0.528),
race = c(
White = 0.706,
Black = 0.121,
Hispanic = 0.107,
Asian = 0.047,
Other = 0.019
),
income = c('Under $50k' = 0.151, '$50k-$100k' = 0.294, 'Over $100k' = 0.555),
married = c(
Married = 0.58,
Widowed = 0.06,
Divorced = 0.10,
Separated = 0.01,
'Never married' = 0.25
)
)
# Only rake on variables where discrepancy exceeds 5%
result_pct <- rake(anes24, targets_many, type = 'pctlim', pctlim = 0.05)
result_pct$vars_used
#> [1] "race" "income" "married" "sex"Use type = 'nlim' to select the top N most discrepant
variables, or iterate = TRUE to re-check after each round
and add newly discrepant variables.
Checking discrepancies directly
You can inspect raw discrepancy scores without raking:
find_discrepant_vars(
anes24,
targets_many,
weights = rep(1, nrow(anes24)),
choosemethod = 'total'
)
#> sex race income married
#> 0.009173777 0.123801047 0.263427639 0.483570392