Acceptance Sampling and Conformance Decisions

Introduction

Statistical process control (SPC) and acceptance sampling are related but fundamentally different quality-management activities. SPC uses data collected during production to detect process change and prevent defects. Acceptance sampling examines a sample from a completed lot and decides whether the lot should be accepted, rejected, or subjected to further action. SPC is therefore primarily a process-improvement tool, whereas acceptance sampling is primarily a conformance-decision tool.

This distinction is strategically important. Inspection cannot create quality in a poorly controlled process, and acceptance sampling does not replace process capability analysis, mistake-proofing, or supplier development. Nevertheless, when inspection of every unit is destructive, expensive, technically difficult, or impractical, a scientifically designed sampling plan can provide a rational basis for lot decisions.

Learning Objectives

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

1. Differentiate process control from lot acceptance.
2. Explain the operating principles of single and double sampling plans.
3. Design and interpret sampling plans using sample size and acceptance number.
4. Explain producer’s risk, consumer’s risk, AQL, LTPD, and OC curves.
5. Evaluate the economic, managerial, ethical, and contractual implications of acceptance decisions.

1. Conformance Decisions and the Unit of Quality

A unit of product is the individual item inspected. A defect is a nonconformity that may affect suitability for use. A defective unit is a unit containing one or more defects and classified as nonconforming under the applicable specification. These terms must not be confused: one defective unit may contain several defects.

A lot is a collection of units produced under sufficiently similar conditions and submitted for acceptance. Lot definition matters because mixing products, shifts, machines, or materials can conceal important sources of variation. A meaningful lot should normally be homogeneous with respect to product, process, time period, and manufacturing conditions.

Acceptance sampling is a decision rule based on a random sample. A single-sampling plan is commonly written as (n,c), where n is the sample size and c is the acceptance number. If d defectives are found in the sample, the lot is accepted when d ≤ c and rejected when d > c.

The plan does not determine the exact quality of an individual lot. It controls the probabilities of accepting or rejecting lots of different quality levels. Consequently, a sampling decision is probabilistic rather than certain.

2. Process Control versus Lot Acceptance

SPC asks: “Is the process stable and predictable?” Acceptance sampling asks: “Does this submitted lot meet the agreed acceptance standard?”

A stable process may still produce output at an unacceptable mean or variation. Conversely, an unstable process may occasionally produce an acceptable lot. The strongest quality system therefore combines process control, capability analysis, supplier controls, and appropriate acceptance procedures.

3. Probability Models for Sampling Plans

3.1 Hypergeometric model: finite-lot sampling

If a lot contains N units, of which D are defective, and a sample of n units is selected without replacement, the probability of finding exactly d defectives is:

Here, C(a,b) denotes the number of combinations of b items selected from a; X is the number of defectives in the sample. The formula applies to random sampling without replacement and requires feasible values of d, n-d, D, and N-D.

For a single plan, the probability of acceptance is:

where P_a is the probability of accepting the lot.

3.2 Binomial approximation

When the lot is large relative to the sample and sampling fractions are small, the binomial model is often a practical approximation. If p is the lot’s defective proportion, then:

and

The approximation assumes independent observations with approximately constant defective probability. A common practical rule is to use it cautiously when the sample is no more than about 10% of the lot; exact hypergeometric calculations are preferable for small lots or high sampling fractions.

For c=0, the binomial expression simplifies to:

This special case is particularly useful for illustrating zero-acceptance plans.

4. Operating Characteristic Curves

The operating characteristic (OC) curve plots P_a against the fraction defective p. It describes how a plan discriminates between lots of different quality. A good plan has a high probability of accepting good lots and a low probability of accepting poor lots, although perfect discrimination is impossible with a sample smaller than the lot.

For a plan (n,c), the OC curve is obtained by evaluating P_a(p) over a range of values of p. As p increases, P_a generally decreases. The curve is not a process-performance curve; it is a decision-performance curve.

Figure: A typical OC curve relating lot quality to probability of acceptance.

5. AQL, LTPD, and Risks

The acceptable quality limit (AQL) is a quality level that, for a specified sampling plan, has a high probability of acceptance. In many systems, the probability of acceptance at the AQL is set approximately to 1−α, where α is the producer’s risk. AQL is not a guarantee that every lot at or below that quality level will be accepted, nor is it necessarily a desirable long-term process target.

The lot tolerance percent defective (LTPD), also called the limiting quality level in some standards, is a poor quality level that the plan is designed to reject with high probability. If the probability of accepting a lot at the LTPD is β, then β is the consumer’s risk.

• Producer’s risk, α: probability of rejecting a lot whose quality is at the specified AQL.
• Consumer’s risk, β: probability of accepting a lot whose quality is at the specified LTPD.

A plan designed around an AQL and an LTPD is often called a producer–consumer risk plan. The two points alone do not uniquely determine every possible plan unless additional constraints are imposed, such as lot size, inspection level, or allowable integer sample sizes.

The terms AQL and LTPD must be used carefully in contracts. They describe probabilistic protection under a stated plan, not an intrinsic property that can be measured without reference to a sampling procedure.

6. Designing a Single-Sampling Plan

A practical design procedure is:

1. Define the quality characteristic, defect classification, lot definition, and measurement method.
2. Establish the producer’s and consumer’s protection points: (p₁,1−α) and (p₂,β).
3. Select a candidate n and c.
4. Calculate or obtain the OC probabilities using the hypergeometric or binomial model.
5. Compare the resulting probabilities with the agreed risks.
6. Evaluate inspection cost, delay, test destructiveness, and the disposition of rejected lots.
7. Document the plan, randomization method, escalation rules, and record requirements.

Worked example: a zero-acceptance plan

Suppose a large lot is sampled using n=100 and c=0. Under the binomial approximation, for a lot containing 1% defectives:

Thus, the lot has only about a 36.6% chance of acceptance. At 5% defectives:

The plan almost always rejects such a lot. However, the plan also rejects approximately 63.4% of lots whose defective proportion is 1%, illustrating why a zero-acceptance rule can impose substantial producer risk at modest defect levels. It should not be interpreted as proving that an accepted lot contains no defects.

7. Double-Sampling Plans

A double-sampling plan permits a second sample when the first sample is inconclusive. It can reduce average inspection when very good or very poor lots are common, while retaining defined protection.

A plan may be represented as (n₁,n₂,c₁,c₂), with a rejection boundary after the first sample. A common structure is:

• Inspect n₁ units.
• Accept if the first-sample defect count d₁ ≤ c₁.
• Reject if d₁ ≥ r₁, where r₁ is the first-stage rejection number.
• Continue to a second sample if c₁ < d₁ < r₁.
• Inspect n₂ additional units and accept if the combined count satisfies d₁+d₂ ≤ c₂; otherwise reject.

The exact acceptance probability is the sum of the probabilities of all acceptance paths:

Under a binomial model, this can be written as:

where Bin(n,d;p)=C(n,d)p^d(1-p)^(n-d), and X₂ is the number of defectives in the second sample. The formula assumes independent samples from a sufficiently large lot; exact finite-lot calculations require a multistage hypergeometric treatment.

Double sampling introduces operational complexity: samples must be traceable, the second-stage rule must be unambiguous, and the organization must prevent selective sampling after observing unfavorable results.

8. Multiple, Sequential, and Continuous Sampling

Multiple-sampling plans extend the logic of double sampling to several stages. Sequential sampling evaluates units one at a time and stops when cumulative evidence crosses an acceptance or rejection boundary. These plans can achieve lower average sample numbers but require disciplined administration and clear statistical documentation.

Continuous sampling plans are useful when product flows continuously rather than in naturally defined lots. They may combine 100% inspection for an initial period with reduced inspection after demonstrated conformance. Such plans must specify switching rules, restoration of full inspection, and treatment of isolated defects.

The selection among single, double, multiple, and sequential plans should consider not only OC performance but also average sample number, inspection cost, testing time, staffing, and the consequences of delayed release.

9. Economic and Managerial Evaluation

Acceptance sampling creates four broad economic outcomes:

1. Accepting a conforming lot.
2. Rejecting a conforming lot, creating producer cost.
3. Accepting a nonconforming lot, creating consumer cost.
4. Rejecting a nonconforming lot, potentially requiring sorting, rework, replacement, or supplier corrective action.

A simplified expected-cost expression is:

where C_s is sampling cost, C_R is the cost associated with rejection of a lot that would otherwise be released, and C_C is the expected cost of accepting nonconforming output. The terms must be defined consistently for the decision context; costs may include warranty, recall, sorting, reputation, regulatory sanctions, and customer disruption.

For rejected lots subjected to 100% screening, the outgoing quality depends on screening effectiveness. If inspection is perfect and all rejected-lot units are screened, rejected lots may be repaired or replaced, but screening is costly and does not eliminate the need for root-cause correction.

Managers should avoid optimizing only the sample size. A smaller sample may reduce inspection cost while increasing customer exposure; a larger sample may improve discrimination while increasing release time and test expense. Total cost of quality, risk severity, and strategic supplier relationships must be considered.

10. Applied Case: Illustrative Medical-Device Connector Lot

Background

A manufacturer receives 10,000 sterile connectors from a qualified supplier. A locking defect can cause leakage in use. The test is destructive, so 100% inspection is undesirable. The purchase agreement specifies a single-sampling plan with n=125 and c=1, with random selection across the lot.

Problem

A sample contains two defective connectors. The receiving department proposes accepting the lot because only 1.6% of the sample is defective, while quality engineering recommends rejection.

Analysis

The rule is based on the acceptance number, not on an informal comparison of percentages. Because d=2 and c=1, the lot must be rejected. The sample proportion is also not a precise estimate of the lot proportion; it is evidence used under the agreed decision rule.

The team then verifies lot homogeneity, sampling randomness, test-system validity, and defect classification. It discovers that the two defects came from the same molding cavity. This raises a process-risk concern not visible from the overall sample percentage.

Approach and outcome

The lot is quarantined. The supplier performs containment and 100% screening, while a corrective-action investigation identifies cavity wear. The customer contract permits resubmission only after documented screening and process verification. Future lots receive intensified sampling temporarily, and the supplier adds cavity-level traceability.

Learning points

The case demonstrates that acceptance sampling is a governance mechanism, not merely a mathematical exercise. Contractual rules, randomization, measurement-system integrity, traceability, defect severity, and corrective action are all essential. A statistically accepted lot could still require escalation if the defect is safety-critical or if evidence indicates a special cause.

11. Limitations, Ethics, and Governance

Sampling plans are vulnerable to biased selection, inadequate lot definition, measurement error, misclassification, and tampering with records. A convenience sample is not a random sample. If operators select visually accessible units or remove suspected defects before sampling, the stated OC performance no longer applies.

Severity-based classification is essential. Critical defects may require zero tolerance, 100% verification, automated detection, or regulatory reporting regardless of an ordinary attribute-sampling result. Sampling must never be used to justify knowingly releasing unsafe products.

Contracts should state the lot definition, quality characteristic, defect classification, sampling standard, randomization method, acceptance and rejection actions, resampling authority, record retention, and escalation requirements. Regulatory and industry standards may prescribe specific plan systems; organizations must use the applicable current standard rather than relying on an informal textbook plan.

Digital inspection systems and artificial intelligence can support randomization, image classification, traceability, and real-time dashboards. However, AI introduces risks of biased training data, model drift, unexplained classifications, cybersecurity compromise, and false confidence. AI-assisted inspection requires validation, human oversight, version control, and periodic performance monitoring. The sampling plan remains a controlled decision rule; an unvalidated algorithm cannot substitute for metrological evidence.

12. Best-Practice Framework

An effective acceptance-sampling program should:

• use SPC and process capability improvement as the primary quality strategy;
• define homogeneous, traceable lots;
• select samples using a documented random method;
• validate measurement systems before collecting acceptance data;
• distinguish defects from defective units;
• set AQL and LTPD values according to risk, not convenience;
• examine the complete OC curve, not only one quality point;
• specify treatment of rejected, reworked, and resubmitted lots;
• include escalation for critical or clustered defects;
• review plans when process capability, technology, supplier performance, or customer risk changes;
• audit records and protect the independence of acceptance decisions.

Chapter Summary

Acceptance sampling is a probabilistic method for deciding whether a lot conforms to an agreed standard. It differs from SPC, which monitors and improves process behavior over time. Single-sampling plans use sample size and acceptance number; double-sampling plans allow an inconclusive first sample to be followed by a second sample. Hypergeometric models are appropriate for finite-lot sampling without replacement, while binomial models provide a useful approximation under large-lot, small-fraction conditions.

OC curves show the probability of acceptance at different quality levels. AQL represents a quality level intended to receive high acceptance probability, while LTPD represents a poor quality level intended to receive low acceptance probability. Producer’s risk is the probability of rejecting acceptable quality; consumer’s risk is the probability of accepting poor quality. Sampling design must also consider inspection cost, rejected-lot disposition, safety, contracts, supplier relationships, measurement systems, and ethical responsibility. Acceptance sampling is strongest when integrated with prevention, process control, capability analysis, and corrective action.

Key Terms

Acceptance number: Maximum number of defectives permitted for lot acceptance.

Acceptance sampling: Statistical procedure for deciding whether to accept or reject a lot.

AQL: Quality level associated with a specified high probability of acceptance.

Binomial model: Probability model commonly approximating sampling from a large lot.

Consumer’s risk: Probability of accepting a lot at the specified poor-quality level.

Defect: A nonconformity in a product or process requirement.

Defective unit: A unit classified as nonconforming because it contains one or more defects.

Double-sampling plan: Plan in which a second sample may be required after an inconclusive first sample.

LTPD: Poor quality level that a plan is designed to reject with high probability.

Lot: Defined collection of product submitted for acceptance.

Lot tolerance percent defective: Another commonly used expression related to LTPD.

OC curve: Graph of probability of acceptance against lot quality.

Producer’s risk: Probability of rejecting a lot at the specified acceptable quality level.

Random sampling: Selection method giving units a controlled and known chance of selection.

Sequential sampling: Sampling that stops when cumulative evidence crosses an acceptance or rejection boundary.

Single-sampling plan: Plan defined primarily by sample size and acceptance number.

SPC: Use of statistical methods to monitor and improve process stability and performance.

Statistical risk: Decision error arising because conclusions are based on a sample rather than a census.

Review Questions

Short Answer Questions

1. What is the principal difference between SPC and acceptance sampling?
2. Define AQL and explain why it is not a guarantee of defect-free lots.
3. What do the parameters n and c represent in a single-sampling plan?
4. When is the binomial model a reasonable approximation to the hypergeometric model?
5. Define producer’s risk and consumer’s risk.

Long Answer Questions

1. Explain the construction and managerial interpretation of an OC curve.
2. Describe a systematic procedure for designing a single-sampling plan.
3. Compare single, double, multiple, and sequential sampling plans.
4. Discuss the economic implications of accepting nonconforming lots and rejecting conforming lots.
5. Explain why acceptance sampling should be integrated with SPC and process capability improvement.

Analytical / Application Questions

1. A large lot is inspected using n=80 and c=0. Calculate the approximate probability of acceptance when the lot contains 2% defectives, and interpret the result.
2. A double-sampling plan has first-stage acceptance, continuation, and rejection regions. Design a clear decision flow for operators and identify two governance controls needed to prevent biased sampling.
3. A safety-critical component has a low observed defect rate but a defect with potentially catastrophic consequences. Evaluate whether ordinary acceptance sampling is adequate and recommend a defensible quality-control strategy.

Multiple Choice Questions with Answers

Question 1

Which statement best distinguishes SPC from acceptance sampling?

A. SPC decides whether a completed lot should be released, while acceptance sampling adjusts machine settings.

B. SPC monitors process behavior over time, while acceptance sampling makes a decision about a submitted lot.

C. SPC is used only for destructive tests, while acceptance sampling is used only for nondestructive tests.

D. SPC eliminates sampling risk, while acceptance sampling always identifies the assignable cause.

Correct Answer: B. SPC monitors process behavior over time, while acceptance sampling makes a decision about a submitted lot.

Question 2

For a single-sampling plan written as (n,c), what does c represent?

A. The lot size

B. The confidence level

C. The maximum number of defectives allowed for acceptance

D. The consumer’s risk

Correct Answer: C. The maximum number of defectives allowed for acceptance

Question 3

Which probability model is most appropriate for sampling without replacement from a finite lot when the sampling fraction is substantial?

A. Normal

B. Poisson

C. Binomial

D. Hypergeometric

Correct Answer: D. Hypergeometric

Question 4

For a zero-acceptance plan with n=100, using the binomial approximation, what is approximately P_a when p=0.01?

A. 0.010

B. 0.366

C. 0.500

D. 0.990

Correct Answer: B. 0.366

Question 5

Producer’s risk is the probability of:

A. Accepting a poor-quality lot at the LTPD

B. Rejecting a lot whose quality is at the specified AQL

C. Finding exactly one defect in a sample

D. Failing to inspect the entire lot

Correct Answer: B. Rejecting a lot whose quality is at the specified AQL

Question 6

What does an OC curve primarily describe?

A. The probability of acceptance at different lot-quality levels

B. The time sequence of machine temperatures

C. The cost of producing each unit

D. The capability index of a stable process

Correct Answer: A. The probability of acceptance at different lot-quality levels

Question 7

In a double-sampling plan, a second sample is normally taken when:

A. The first sample immediately satisfies the acceptance rule

B. The first sample immediately satisfies the rejection rule

C. The first sample result is inconclusive under the stated decision boundaries

D. The lot has already been accepted and shipped

Correct Answer: C. The first sample result is inconclusive under the stated decision boundaries

Question 8

If a sample of 125 units contains 2 defectives under a plan with c=1, what is the ordinary decision?

A. Accept, because the sample defective percentage is below 2%

B. Accept, because two defectives are always permitted in a sample over 100

C. Reject, because the observed count exceeds the acceptance number

D. Defer automatically until a second sample is taken

Correct Answer: C. Reject, because the observed count exceeds the acceptance number

Question 9

Which practice most directly threatens the validity of an OC curve?

A. Randomly selecting units throughout a homogeneous lot

B. Using a validated measurement system

C. Selecting convenient or visibly attractive units instead of a random sample

D. Documenting the lot definition

Correct Answer: C. Selecting convenient or visibly attractive units instead of a random sample

Question 10

Why should acceptance sampling not be treated as a substitute for SPC?

A. Acceptance sampling cannot involve probability calculations

B. Acceptance sampling decides on lots but does not by itself identify or correct process causes

C. SPC is always less expensive than every sampling plan

D. SPC applies only to service organizations

Correct Answer: B. Acceptance sampling decides on lots but does not by itself identify or correct process causes