Finite-Population and Complex-Survey Distributions

Introduction and context

Quality audits and supplier studies rarely sample from an effectively infinite, independently and identically distributed population. An auditor may inspect 80 lots from a supplier’s finite quarterly production, select plants first and units second, deliberately oversample high-risk suppliers, or divide the population into geographical and performance strata. These designs improve operational efficiency and risk coverage, but they change the sampling distribution of an estimator.

This chapter develops the statistical foundations required to quantify that change. The central issues are finite-population correction, unequal selection probabilities, survey weights, stratification, clustering, and design effects. The emphasis is not merely on calculating a standard error, but on identifying the population, sampling design, estimator, and source of uncertainty before making a quality decision.

Learning objectives

By the end of this chapter, the reader should be able to:

1. Distinguish a finite-population distribution from a model-based distribution.
2. Apply the finite-population correction when sampling without replacement.
3. Construct and interpret design weights under unequal selection probabilities.
4. Estimate means, totals, proportions, and their uncertainty under stratified designs.
5. Explain why clustering generally increases variance and calculate a design effect.
6. Select defensible variance estimators for complex quality audits and supplier studies.
7. Evaluate practical, ethical, and governance implications of weighting and variance estimation.

1. Finite populations and sampling distributions

A finite population consists of the actual, identifiable units relevant to a study: batches produced in a month, invoices processed in a quarter, or components delivered by a supplier. Let the population contain N units, indexed by i = 1, 2, ..., N, with quality value y_i. The finite-population mean and variance are

Here S² is a descriptive property of the fixed population, not a random quantity. Randomness arises because the sample is selected. Under simple random sampling without replacement (SRSWOR), every subset of size n has equal probability.

The sample mean is

and is unbiased for Ȳ under SRSWOR. Its exact design variance is

where f = n/N is the sampling fraction. The factor

is the finite-population correction (FPC) for a standard error. It is dimensionless and applies when sampling is without replacement from a finite population and the sampling fraction is not negligible.

The commonly taught expression S²/n is an approximation appropriate when the population is very large relative to the sample, or when f is sufficiently small. A practical convention is to examine the FPC when f > 0.05, although the substantive threshold depends on the precision required and the sampling design.

Why the FPC matters

If 50 of 100 production lots are inspected without replacement, the remaining lots are not independent of the inspected lots. Sampling half the population provides more information than 50 observations drawn from a very large population. The FPC reflects this reduction in uncertainty. It does not correct bias caused by a defective sampling frame, nonresponse, measurement error, or convenience selection.

For a finite-population total,

an unbiased estimator under SRSWOR is Ť = Nȳ, with variance

For a binary characteristic, such as nonconforming versus conforming, y_i takes values 0 or 1. The finite-population proportion is P = Ȳ, and the estimated proportion is p̂ = ȳ. Its variance is

when expressed using the finite-population binary variance; in practice, P is replaced by an appropriate sample estimate. For small samples or proportions close to 0 or 1, normal intervals can be poor; Wilson, exact, or design-based methods may be preferable.

2. Design-based reasoning and inclusion probabilities

Complex surveys are best understood through the sampling design. Let π_i denote the first-order inclusion probability: the probability that unit i appears in the sample. A unit with a small inclusion probability represents more population units than one with a large inclusion probability.

The basic design weight is

The Horvitz–Thompson estimator of a population total is

and is design-unbiased when every relevant unit has π_i > 0 and the inclusion probabilities are correctly specified. For a population mean, the corresponding estimator is often the ratio or normalized-weight estimator

The Horvitz–Thompson mean, when N is known, is T̂_HT/N. The normalized estimator is generally not exactly unbiased in finite samples, but often has favorable practical performance and guarantees that weights represent the observed population total.

Weight construction

Operational weights commonly combine several factors:

The design weight is the inverse selection probability. The nonresponse adjustment attempts to compensate for sampled units that do not provide usable observations. Calibration or post-stratification adjusts weighted sample totals to trusted external benchmarks, such as the known number of lots by region or supplier class. These adjustments can reduce bias if their assumptions are credible, but they do not automatically remove nonresponse or coverage bias.

Weights should not be treated as arbitrary multipliers. A weight is meaningful only in relation to a target population, a sampling frame, and a documented selection mechanism. Extreme weights increase variance and can make results highly sensitive to a few observations. Weight trimming may stabilize estimates, but it introduces a bias–variance trade-off and must be reported transparently.

3. Stratified sampling

In stratified sampling, the population is divided into H mutually exclusive and collectively exhaustive strata. Strata may be supplier risk class, product family, geography, or plant size. Let stratum h contain N_h units, of which n_h are sampled. Define W_h = N_h/N. The stratified estimator of the population mean is

where ȳ_h is the sample mean in stratum h. Under independent SRSWOR within each stratum, its variance is

where f_h = n_h/N_h and S_h² is the finite-population variance within stratum h. In practice, S_h² is estimated by the sample variance s_h².

Stratification can lower variance when units within a stratum are relatively homogeneous and the stratum means differ. It also ensures representation of important subgroups. Proportional allocation sets n_h = nW_h. If costs differ or variability differs substantially, a form of Neyman allocation is useful:

This allocation is optimal for a fixed total sample size under equal per-unit costs and known or estimated within-stratum standard deviations. With stratum-specific costs c_h, the cost-adjusted form is proportional to N_h S_h/√c_h. Allocation must also respect minimum sample sizes needed for subgroup reporting.

Worked example: stratified supplier audit

An organization has 1,000 supplier lots: 700 from low-risk suppliers and 300 from high-risk suppliers. It audits 70 low-risk lots and 60 high-risk lots. The observed nonconformance rates are 4% and 10%, and the within-stratum sample standard deviations for the binary indicator are approximated by 0.196 and 0.302.

The weighted estimate is

Thus the estimated population nonconformance rate is 5.8%. The estimated variance is

The two components are approximately 0.0000358 and 0.0000656, giving v̂ ≈ 0.0001014. Therefore

An approximate 95% interval is 0.058 ± 1.96(0.0101), or approximately 3.8% to 7.8%. This interval is a design-based approximation and should be interpreted cautiously if the sample is small, the binary outcome is rare, or the audit process has clustering or nonresponse.

4. Cluster sampling and intracluster correlation

Cluster sampling selects groups—such as plants, suppliers, shifts, or pallets—and then observes units within selected groups. It reduces travel and administrative cost, but observations in the same cluster tend to resemble one another. Treating them as independent usually understates uncertainty.

The intracluster correlation coefficient (ICC), denoted ρ, measures the correlation between two observations from the same cluster. A simple equal-size cluster design effect is

where m is the average number of sampled units per cluster. The effective sample size is approximately

These expressions assume approximately equal cluster sizes, a common outcome variance, and a one-stage or suitably simplified clustered design. They are not universal replacements for a design-specific variance estimator.

For unequal cluster sizes, a useful approximation is

where m̄ is the average cluster sample size and CV_m is its coefficient of variation. Written without ambiguity, the bracketed term is ((1 + CV_m²)m̄ − 1)ρ, so the full expression is 1 + ((1 + CV_m²)m̄ − 1)ρ. Unequal cluster sizes increase the design effect because large clusters contribute many correlated observations.

A practical interpretation

Suppose 120 audited units are drawn from 12 suppliers, with 10 units per supplier, and ρ = 0.20. Then

The effective sample size is only about 120/2.8 ≈ 42.9. The nominal sample size is 120, but the information is comparable to roughly 43 independent observations. Adding more units within the same suppliers may be less effective than adding suppliers.

In multistage designs, variance can be affected by clustering at several levels. For example, units may be nested within lots, lots within plants, and plants within suppliers. A valid analysis should identify the primary sampling unit and account for all material stages, often through Taylor linearization, replicate weights, or a multilevel model used for a clearly stated inferential target.

5. Unequal probabilities and the variance of weighted estimators

Unequal-probability sampling is often appropriate when high-risk suppliers deserve more inspection. However, an unweighted mean then estimates the average of the sampled units rather than the population average. Inverse-probability weighting restores representation under the design.

For a general without-replacement design, the exact Horvitz–Thompson total variance involves second-order inclusion probabilities π_ij, the probability that both units i and j are included:

In an actual sample, unbiased variance estimators require estimable inclusion probabilities and appropriate sample-level formulas. In complex audits these probabilities may be unavailable or unwieldy, so analysts commonly use Taylor linearization, jackknife, balanced repeated replication, or bootstrap methods designed for the sampling structure.

Taylor linearization approximates the variance of a nonlinear estimator, such as a ratio or weighted proportion, by replacing it with the variance of a first-order linearized variable. Replication methods repeatedly recompute the estimator after deleting, resampling, or reweighting primary sampling units. Replicates must preserve strata, clusters, and selection features; an ordinary individual-level bootstrap can be invalid for a clustered survey.

6. Estimation of proportions, means, and quality indices

For a weighted binary quality indicator y_i ∈ {0,1}, the weighted proportion is

If y_i is a measurement such as tensile strength, defect count, or cycle time, the same expression estimates a weighted mean. A total requires weights on the population scale, often T̂ = ∑w_i y_i, whereas a normalized mean divides by ∑w_i. Confusing totals and means is a frequent reporting error.

A confidence interval should reflect the design-based standard error. A generic approximate interval is

where z_1−α/2 is a standard normal quantile. The approximation is most defensible when the effective sample size is adequate and no single weight dominates. For small numbers of clusters, a t reference distribution with degrees of freedom related to the number of independent primary sampling units is often more appropriate than a normal reference. Exact degrees-of-freedom rules depend on the survey design and software method.

7. Integrated case: a supplier quality audit

Background

This illustrative case concerns a manufacturer receiving 4,000 lots from 80 suppliers during one quarter. Suppliers are classified as low, medium, or high risk. High-risk suppliers account for only 10% of lots but historically contribute a disproportionate share of failures.

Problem

Management wants an estimate of the overall lot nonconformance rate and a decision about whether it is below the contractual limit of 6%. The audit selects 12 suppliers, oversampling high-risk suppliers, and then selects 8 lots from each selected supplier. The data therefore contain unequal supplier selection probabilities and clustering by supplier.

Analysis

The analyst first obtains the supplier inclusion probabilities and the conditional lot-selection probabilities. For lot i, the base weight is the inverse of the product of these probabilities:

Nonresponse adjustments are made within risk classes only when response behavior and the adjustment cells are defensible. The weighted nonconformance estimate is 6.4%. An unweighted estimate is 8.1%, because high-risk suppliers were intentionally oversampled.

A naive individual-level standard error is 0.9 percentage points. However, the audit team estimates an ICC of 0.18 among lots within suppliers. With eight lots per supplier, the approximate cluster design effect is

The clustering-adjusted approximation is therefore 0.9√2.26 ≈ 1.35 percentage points, before considering any additional weight variability. The estimate is thus approximately 6.4% with an uncertainty materially larger than the naive analysis suggests. The correct operational conclusion is not that the specification is definitively violated, but that the evidence is insufficiently precise for a confident pass decision. Additional suppliers, rather than merely additional lots within the same suppliers, would improve precision more effectively.

Approach and solution

The organization adopts a survey-aware estimator with supplier as the primary sampling unit, reports weighted and unweighted results separately, and uses a small-sample adjustment based on the number of sampled suppliers. It also publishes the number of suppliers, lots per supplier, weight range, response rate, and variance method. A follow-up audit samples more suppliers and fewer lots per supplier while retaining minimum coverage of high-risk suppliers.

Outcome and learning points

The revised design produces a more defensible estimate and prevents management from treating deliberate oversampling as evidence that the entire supplier population is worse than it is. The case demonstrates five principles: the estimand must be explicit; weights are essential under unequal selection; clusters reduce effective information; the number of independent primary units matters; and variance methods are part of the audit design, not an afterthought.

8. Challenges, limitations, and governance

Coverage and frame error

No variance formula repairs an incomplete sampling frame. If informal suppliers or failed lots are absent from the frame, an unbiased estimator for the frame is not necessarily unbiased for the intended population. Frame construction, duplicate removal, eligibility rules, and time boundaries should be documented.

Nonresponse and missing measurements

Nonresponse may be related to quality. A supplier that withholds records is not equivalent to a randomly missing supplier. Adjustments should use credible response cells, auxiliary data, follow-up procedures, and sensitivity analyses. Imputation must preserve the design and should not conceal uncertainty.

Weight instability

Large weights can dominate estimates. Analysts should inspect the weight distribution, report trimming or calibration decisions, and compare results under reasonable alternatives. Trimming without disclosure can make a result appear more precise than it truly is.

Measurement and audit dependence

Repeated measurements by the same auditor, instrument drift, and correlated inspection rules create additional dependence. Sampling variance is only one component of total uncertainty. A quality system should distinguish sampling error from measurement error, process change, and systematic audit bias.

Confidentiality and fairness

Supplier-level results may affect contracts, reputation, and access to business. Governance should define who can see identifiable results, how high-risk suppliers are selected, and how automated risk scores are validated. AI may assist with frame construction, anomaly detection, or adaptive sample allocation, but selection rules must remain auditable. If an AI model changes inclusion probabilities, those probabilities must be recorded so that valid weights and uncertainty estimates can be produced.

9. Best-practice workflow

1. Define the target population, estimand, time period, and unit of analysis.
2. Build and audit the sampling frame.
3. Select strata, clusters, and units using documented probabilities.
4. Record inclusion probabilities at every sampling stage.
5. Compute design, nonresponse, and calibration weights separately before combining them.
6. Inspect weight variability, cluster sizes, response patterns, and extreme observations.
7. Use a variance estimator that matches the design: FPC for SRSWOR, stratified formulas for independent strata, and cluster-aware methods for multistage designs.
8. Report estimates, standard errors or confidence intervals, effective sample size, design effect, response rate, and limitations.
9. Conduct sensitivity analyses for alternative weighting, nonresponse, and rare-event assumptions.
10. Link the statistical conclusion to the quality decision while distinguishing evidence from managerial judgment.

Chapter Summary

Finite-population inference treats the population as fixed and the sample selection as the source of randomness. Under SRSWOR, the finite-population correction √(1 − n/N) reduces the variance because sampling without replacement supplies more information than independent sampling from an effectively infinite population. Stratified estimators combine stratum estimates using population weights and can improve precision when strata are internally homogeneous. Unequal-probability designs require inverse-probability weights, while nonresponse and calibration adjustments must be justified rather than applied mechanically. Cluster sampling is operationally efficient but usually increases variance because observations within clusters are correlated; the design effect and effective sample size make this loss visible. Valid quality-audit inference requires alignment among the target population, sampling design, weights, estimator, and variance method. Frame error, nonresponse, measurement dependence, and ethical handling of supplier information remain important sources of uncertainty beyond the sampling formulas.

Key Terms

Finite population: A defined, countable set of units to which an inference applies.

Sampling fraction: The ratio n/N of sample size to population size.

Finite-population correction: The factor √(1 − n/N) used for sampling without replacement.

Design-based inference: Inference based on the randomization induced by the sampling design.

Inclusion probability: The probability that a specified unit is included in the sample.

Design weight: The inverse of a unit’s inclusion probability.

Horvitz–Thompson estimator: An inverse-probability estimator of a finite-population total.

Normalized-weight estimator: A weighted estimator divided by the sum of sample weights.

Stratified sampling: Sampling conducted independently within defined population subgroups.

Cluster sampling: Sampling groups of population units rather than units individually.

Primary sampling unit: The first-stage unit selected in a multistage design.

Intracluster correlation: Similarity or correlation among observations in the same cluster.

Design effect: The ratio of complex-design variance to a reference simple-random-sampling variance.

Effective sample size: The size of an independent simple sample with approximately equivalent precision.

Neyman allocation: Allocation proportional to stratum size multiplied by within-stratum variability.

Taylor linearization: A first-order approximation used to estimate variance for nonlinear estimators.

Replication variance: Variance estimated by repeatedly recomputing an estimator on structured replicates.

Calibration weighting: Adjustment of weights so weighted sample totals match known benchmarks.

Weight trimming: Limiting extreme weights to reduce variance, with possible introduced bias.

Coverage error: Bias resulting when the sampling frame omits, duplicates, or misclassifies target units.

Review Questions

Short Answer Questions

1. What does the finite-population correction represent, and when is it applicable?
2. Why is the inverse of an inclusion probability used as a design weight?
3. State the principal difference between stratified sampling and cluster sampling.
4. Define intracluster correlation and explain its effect on variance.
5. Why can an unweighted estimate be misleading when high-risk suppliers are oversampled?

Long Answer Questions

1. Derive the SRSWOR variance of the sample mean conceptually and explain the role of the sampling fraction.
2. Explain how design, nonresponse, and calibration weights may be combined, including the risks of extreme weights.
3. Compare proportional and Neyman allocation in a stratified supplier audit.
4. Discuss Taylor linearization and replication methods as approaches to variance estimation.
5. Explain why a large nominal sample may have a small effective sample size in a clustered audit.

Analytical / Application Questions

1. A population contains 2,000 lots and an SRSWOR audit examines 200. If the estimated population standard deviation is 0.25, calculate the approximate standard error of the sample mean using the FPC. State the assumptions.
2. An audit samples 15 suppliers and 6 lots per supplier. If the ICC is 0.12, calculate the approximate cluster design effect and effective sample size for 90 lots.
3. Design an audit for a population with low-, medium-, and high-risk suppliers. Specify strata, selection probabilities, weights, a variance method, and the information that should be reported to management.

Multiple Choice Questions with Answers

Question 1

Under SRSWOR, which factor is the finite-population correction for a standard error?

A. 1 − n/N

B. √(1 − n/N)

C. √(N/n)

D. n/N

Correct Answer: B. √(1 − n/N)

Question 2

A unit has inclusion probability 0.20. What is its basic design weight?

A. 0.20

B. 1.20

C. 5

D. 20

Correct Answer: C. 5

Question 3

Which estimator combines stratum sample means using population stratum weights?

A. The unweighted cluster mean

B. The stratified estimator ∑W_hȳ_h

C. The within-cluster variance

D. The finite-population total only

Correct Answer: B. The stratified estimator ∑W_hȳ_h

Question 4

If the average cluster size is 6 and ρ = 0.10, what is the approximate equal-size cluster design effect?

A. 1.06

B. 1.50

C. 1.60

D. 2.00

Correct Answer: C. 1.60

Question 5

Which action generally improves precision most when observations are strongly clustered?

A. Add more observations within the same clusters only

B. Reduce the number of sampled clusters

C. Sample more independent clusters

D. Ignore the clustering in analysis

Correct Answer: C. Sample more independent clusters

Question 6

What is the principal purpose of calibration weighting?

A. To remove all measurement error

B. To make weighted sample totals agree with trusted population benchmarks

C. To guarantee zero sampling variance

D. To replace the sampling frame

Correct Answer: B. To make weighted sample totals agree with trusted population benchmarks

Question 7

In a stratified design, the finite-population correction within stratum h is based on which fraction?

A. n/N for the entire population only

B. n_h/N_h

C. H/n_h

D. N_h/n

Correct Answer: B. n_h/N_h

Question 8

What is a major danger of using an ordinary individual-level bootstrap for a clustered survey?

A. It always overestimates the population size

B. It may destroy the dependence structure and understate uncertainty

C. It converts means into totals

D. It prevents calculation of weights

Correct Answer: B. It may destroy the dependence structure and understate uncertainty

Question 9

A weighted estimate of a binary quality indicator is interpreted as what?

A. A weighted population proportion

B. The number of sampled clusters

C. The unweighted response rate only

D. The design effect

Correct Answer: A. A weighted population proportion

Question 10

Which issue cannot be repaired merely by applying a finite-population correction?

A. Sampling without replacement

B. A high sampling fraction

C. An incomplete sampling frame

D. A known population size

Correct Answer: C. An incomplete sampling frame