Acceptance Sampling
Acceptance Sampling: Single Sampling Plans, OC Curves and Risks
The decision problem
A receiving department gets a lot of N items. Inspecting every item may be too costly, too slow, or impossible if testing destroys the item. Acceptance sampling inspects a random sample and makes a lot-level decision: accept or reject.
It is a decision rule about lots, not a tool for estimating or improving a process over time (that is the job of process control methods).
Single sampling plans
A plan is written (n, c):
- Draw a random sample of n items from the lot.
- Count the defectives, d.
- If d ≤ c, accept the lot. If d > c, reject it.
The acceptance number c is a count of defective items in the sample, not a percentage.
Probability of acceptance
If the lot's fraction defective is p and the lot is large relative to the sample, the number of defectives in the sample is approximately binomial, so
Pa(p) = Σk=0c C(n, k) pk(1 − p)n−k
- When n is large and p small, the Poisson approximation with λ = np is convenient: Pa ≈ Σk=0c e−λλk/k!
- When the sample is a large fraction of a small lot, the hypergeometric distribution is exact.
Worked example. Plan n = 30, c = 1, lot fraction defective 3%.
- P(0) = 0.9730 = 0.4010
- P(1) = 30 × 0.03 × 0.9729 = 0.3721
- Pa = 0.4010 + 0.3721 = 0.773
The same plan accepts a 10%-defective lot with probability 0.184. A Poisson illustration: for n = 100, c = 3 and p = 0.02, λ = 2 and Pa ≈ e−2(1 + 2 + 2 + 1.333) = 0.857.
The operating characteristic curve
Plotting Pa against p gives the plan's OC curve. It starts at 1 when p = 0 and falls as quality worsens. Its shape shows how well the plan discriminates:
- Increasing n with c fixed makes the curve fall faster (stricter). At p = 0.03, n = 60 with c = 1 accepts with probability 0.459, versus 0.773 for n = 30.
- Increasing c with n fixed makes the plan more lenient: n = 60 with c = 2 gives 0.731 at p = 0.03.
- Increasing both in proportion makes the curve steeper around its midpoint, which means better discrimination between good and bad lots, at a higher inspection cost.
Two quality levels and two risks
Term | Meaning | Associated risk |
|---|---|---|
AQL (acceptable quality level) | A fraction defective the consumer regards as satisfactory as a process average | Producer's risk α = 1 − Pa(AQL) |
LTPD (lot tolerance percent defective), also called RQL | A fraction defective the consumer wants to reject | Consumer's risk β = Pa(LTPD) |
Plans are often designed so that α is about 0.05 and β about 0.10.
Rectifying inspection
When rejected lots are 100% inspected and every defective found is replaced with a good item, the plan changes the quality of what is shipped onward:
- Average outgoing quality: AOQ ≈ Pa · p · (N − n)/N. Accepted lots keep their uninspected defectives; rejected lots leave with none; the sampled items are cleaned either way.
- AOQL: the maximum of AOQ over all p, a worst-case long-run outgoing quality.
- Average total inspection: ATI = n + (1 − Pa)(N − n).
AOQ is small for very good lots (few defectives) and for very bad lots (almost always rejected and screened), so it peaks at an intermediate quality.
A supplier and a customer agree on a sampling plan with an acceptable quality level (AQL) of 1.5% defective and a lot tolerance percent defective (LTPD) of 6% defective.
Which quantity is the producer's risk of this plan?
A single sampling plan uses n = 20 and c = 1. Lots are large, so the binomial model applies.
What is the probability of accepting a lot whose fraction defective is 8%?