Does LQAS estimate an exact defect rate?

Its main objective is classification; precise estimation generally needs a design aimed at that purpose.

Can the lot be a geographic area?

Yes, if the target population and sampling process are clearly defined.

Is a smaller cutoff always better?

It reduces acceptance at a given defect rate but can increase rejection of satisfactory lots, so both risks matter.

AI Glossary / Statistics and probability / Intermediate

Lot quality assurance sampling

Definition

Lot quality assurance sampling, or LQAS, is a sampling approach used to classify a lot, service area, or other defined group against quality or performance criteria. A sample is assessed and its result is compared with a predetermined decision rule. The design aims to control the probabilities of incorrect classification at specified quality levels. Its primary purpose is a practical classification decision, rather than a highly precise estimate of every lot's underlying quality rate.

What it means

A “lot” need not be a physical batch. It can be a district, work unit, or supervision area. What matters is that the population and the quality indicator are clearly defined. Each sampled unit is classified according to the chosen criterion, such as whether an item has a specified defect.

The sample size and cutoff work together. Choosing a small sample because it is convenient and inventing a cutoff afterwards does not create a defensible LQAS plan. Good and poor quality reference levels, together with acceptable decision risks, determine what the plan can distinguish. Intermediate quality levels remain an uncertainty region.

How it works

Define the lot, sampling frame, measurement rule, and action that follows each classification. Select two reference quality levels and specify the risks of rejecting a satisfactory lot and accepting an unsatisfactory one. Choose a sample size and decision cutoff whose operating-characteristic curve meets those requirements.

For a simple defect-count plan, randomly inspect n units and accept when the number of defective units is at most c. Under an appropriate binomial model, acceptance probability is the sum of binomial probabilities for counts zero through c. Finite lots sampled without replacement may require hypergeometric calculations instead.

Collect the sample as designed, apply the fixed rule, and document the result. Cluster sampling introduces dependence and may require a different design calculation. Distinguish counting defective units from counting multiple defects on a single unit; these are different statistical outcomes.

Three practical examples

A manufacturer classifies a production batch using randomly selected units and a specified maximum defective-unit count.
A service organization identifies areas that may need additional supervision using a predefined binary service-quality indicator.
An AI data team screens annotation batches for a clearly defined error type, using a validated inspection plan before deciding which batches need rework.

Think of it this way

Think of a checkpoint designed to sort groups into action categories. It does not inspect every member or certify perfection. Its usefulness depends on how often it sends good groups for unnecessary rework and how often it lets poor groups pass. The cutoff must balance these risks before the checkpoint is used.

Worked example

Consider an original teaching plan that samples 30 independent units from a sufficiently large lot and accepts if no more than two are defective. This is an illustrative plan, not a recommended standard for a particular product.

If the true defective proportion is p, the acceptance probability is the sum, for k = 0, 1, and 2, of choose(30,k) times p to the power k times (1 - p) to the power (30 - k).

At p = 0.02, that probability is approximately 0.97828. The rejection probability is therefore about 2.17%. At p = 0.15, acceptance probability is approximately 0.15140, so a lot at this poor-quality reference level would still pass about 15.14% of the time.

If the desired risk of accepting such a poor lot were at most 10%, this plan would fail that design requirement. Observing two defective units would trigger acceptance under the stated rule, but it would not prove that the lot's defective proportion is 2/30 or below a particular threshold. The arithmetic illustrates why sample size, cutoff, and error probabilities must be evaluated together.

Illustrative case study

Fictional case: Summit Labels reviews batches of manually annotated examples before they enter an AI training pipeline. The quality team defines one auditable label error, draws random examples within each batch, and specifies a classification plan before inspection.

An early candidate plan has an unacceptable chance of passing a poor batch, so the team changes the design rather than merely renaming the cutoff. Rejected batches receive documented investigation and rework. Accepted batches remain subject to process monitoring because a sampling decision does not establish zero errors, eliminate reviewer mistakes, or guarantee that every important error category was covered.

Common misconceptions and mistakes

A passing classification is not a guarantee of perfect quality or a precise prevalence estimate. A failing classification is not proof that every item is defective. The risks apply under the design and sampling assumptions, not regardless of how observations were selected.

Do not replace random selection with easy-to-reach units and retain the original error guarantees. Do not reuse a familiar small sample size without checking the required thresholds and risks. Clustering, inaccurate assessments, and an unrepresentative frame can undermine the plan. Changing the decision rule after seeing the sample changes its operating properties.

Business use

LQAS can support efficient prioritization when inspection resources are limited and decisions concern defined groups. Its value is greatest when the organization knows what action follows each classification and can tolerate the quantified error risks.

For AI quality work, specify whether the lot is a batch of labels, generated outputs, or another stable population. Separate error severity categories if they require different decisions. Keep an audit trail of sample selection and adjudication, and assess process trends across lots. Classification sampling complements broader quality management; it does not replace investigating why errors occur.

Check your understanding

1. In the example, is a count of two defects accepted?

Answer: Yes, under the stated rule. 2.

Does the plan meet a 10% acceptance-risk limit at p = 0.15?

Answer: No; its risk is approximately 15.14%. 3.

Can convenience sampling preserve the binomial design guarantee automatically?

Answer: No.

Related terms

Acceptance sampling
Binomial distribution
Cluster Sampling
operating-characteristic curve
producer's risk
consumer's risk
quality control
sampling frame

Learn next

Plot acceptance probability against defect proportion for several sample sizes and cutoffs. Then examine finite-population and cluster designs, keeping classification decisions separate from precise rate estimation.

Acceptance probability

Sources

NIST/SEMATECH, “Choosing a Sampling Plan with a given OC Curve.” https://www.itl.nist.gov/div898/handbook/pmc/section2/pmc232.htm — primary binomial acceptance-probability and decision-risk guidance. “Choosing a Cluster Sampling Design for Lot Quality Assurance Sampling Surveys.” https://pmc.ncbi.nlm.nih.gov/articles/PMC4488393/ — indexed original-research introduction consulted for LQAS classification scope; direct full-page retrieval required browser verification. The worked plan and fictional case are original.

Previous: Lost-in-the-middle effect Next: Loudness