Skip to content

Complex Survey Inference Lab

One sample, several standard errors. The naive s/√n pretends the sample were SRS. The design-based SEs use the strata and the PSUs: units in one PSU resemble each other, so the honest SE is larger. Compare them with the true SE from repeated sampling.

SE design 1.339, naive 0.823

Design tree: strata → sampled PSUs

6 of 40 PSUs per stratum, 8 units per PSU; w = (40/a)(M_i/8)

  • S1 (w̄ = 39.4)
    • PSU 12: n = 8, weighted mean 49.9
    • PSU 26: n = 8, weighted mean 51.5
    • PSU 10: n = 8, weighted mean 41.9
    • PSU 29: n = 8, weighted mean 42.5
    • PSU 20: n = 8, weighted mean 42.2
    • PSU 32: n = 8, weighted mean 49.8
  • S2 (w̄ = 41.1)
    • PSU 123: n = 8, weighted mean 47.2
    • PSU 118: n = 8, weighted mean 45.0
    • PSU 109: n = 8, weighted mean 37.9
    • PSU 116: n = 8, weighted mean 43.0
    • PSU 125: n = 8, weighted mean 46.1
    • PSU 137: n = 8, weighted mean 53.1
  • S3 (w̄ = 30.8)
    • PSU 230: n = 8, weighted mean 42.4
    • PSU 237: n = 8, weighted mean 48.1
    • PSU 207: n = 8, weighted mean 51.8
    • PSU 204: n = 8, weighted mean 53.9
    • PSU 201: n = 8, weighted mean 65.5
    • PSU 238: n = 8, weighted mean 49.1
  • S4 (w̄ = 34.9)
    • PSU 305: n = 8, weighted mean 46.7
    • PSU 318: n = 8, weighted mean 60.0
    • PSU 333: n = 8, weighted mean 52.4
    • PSU 322: n = 8, weighted mean 53.4
    • PSU 314: n = 8, weighted mean 52.5
    • PSU 321: n = 8, weighted mean 62.8

SE of the mean of y by method

bars: naive s/√n, Taylor linearisation (declared design), the replicate method, and the true SE (SD of the estimate over 300 repeated samples from the population)

naive s/√n0.823Taylor (design)1.339 (×1.63)replicate: JKn1.349 (×1.64)true SE (simulation)1.150 (×1.40)
SE of the mean of y by method: naive s/√n 0.823; Taylor (design) 1.339; replicate: JKn 1.349; true SE (simulation) 1.150.

95 % CI coverage over 300 samples

naive: 85.3 %

design (t, PSU − H df): 94.7 %

JKn (stratified delete-one-PSU)

Formula

V̂(ȳ_w) = Σ_h n_h/(n_h−1) Σ_i (z_hi − z̄_h)², z_hi = Σ_k w_hik (y_hik − ȳ_w)/Σw

Method: Seeded synthetic population (H strata × 40 PSUs of 30–60 units, PSU effect SD 4) and a stratified two-stage sample. Taylor linearisation with the stratified ultimate-cluster (with-replacement) formula, optional first-stage fpc; JK1, JKn and BRR replicate weights as survey::as.svrepdesign (replicate-mean centring); domains by zero weights; Rao–Scott as svychisq; regression as svyglm. · Synthetic, seeded data. Illustrative, not an empirical finding.

Validation: Means, Taylor SEs (with and without fpc), deff, JK1, JKn and BRR SEs, domain SEs, both Rao–Scott tests and svyglm coefficients and SEs match R survey 4.2.1 on five exported samples (tests/unit/complex-survey.test.ts); 95 % design intervals cover between 93.5 and 96.5 % in a 1 000-sample simulation (spec target 94–96 %; 96.2 % observed for H = 4, a = 6). · numerically validated