Discuss the theoretical basis of the sequential probability ratio test (SPRT) in acceptance sampling. Explain how acceptance and rejection numbers are chosen to satisfy α and β risks.
15 marks · directive: Discuss
The core of a good answer
The SPRT, due to Wald (1945), inspects items one at a time and after m items compares the likelihood ratio λₘ = Π f(xᵢ; p₂)/f(xᵢ; p₁) with constants B ≈ β/(1 − α) and A ≈ (1 − β)/α, accepting the lot when λₘ ≤ B, rejecting when λₘ ≥ A and continuing otherwise; for attributes this reduces to two parallel lines in the (m, d) plane, acceptance number aₘ = −h₁ + sm and rejection number rₘ = h₂ + sm, with h₁, h₂ and s fixed by p₁, p₂, α and β.
An outline written by upsc.help, not an official UPSC answer; every fact checked by a separate review. How we work
What the question asks
Theory: hypotheses H₀: p = p₁ (AQL) against H₁: p = p₂ (LTPD), the likelihood ratio, Wald's boundaries and why they approximately give the stated α and β, then the derivation of the acceptance and rejection lines for a binomial process, plus OC and ASN properties. A worked numerical illustration strengthens the answer.
Answer outline
Introduction
In acceptance sampling the lot is good if p ≤ p₁ and bad if p ≥ p₂. The producer's risk is α = P(reject | p₁) and the consumer's risk is β = P(accept | p₂). Wald and Wolfowitz (1948) proved that the SPRT is optimal: among tests with these error probabilities it needs, on average, the fewest observations at p₁ and p₂.
Step 1: The likelihood ratio
- After inspecting m items with dₘ defectives: λₘ = [p₂^dₘ (1 − p₂)^(m−dₘ)] / [p₁^dₘ (1 − p₁)^(m−dₘ)].
- Rule: accept if λₘ ≤ B; reject if λₘ ≥ A; otherwise inspect another item (B < 1 < A).
Step 2: Choice of A and B from α and β
- Wald's inequalities: A ≤ (1 − β)/α and B ≥ β/(1 − α).
- Taking A = (1 − β)/α and B = β/(1 − α) gives actual risks α′, β′ with α′ + β′ ≤ α + β (and each close to its target), because overshoot of the boundary is small.
Step 3: Acceptance and rejection numbers (binomial case)
- Take logs: dₘ log(p₂/p₁) + (m − dₘ) log[(1 − p₂)/(1 − p₁)] lies between log B and log A.
- Let g₁ = log(p₂/p₁), g₂ = log[(1 − p₁)/(1 − p₂)].
- Acceptance line: aₘ = −h₁ + sm, with h₁ = log[(1 − α)/β]/(g₁ + g₂).
- Rejection line: rₘ = h₂ + sm, with h₂ = log[(1 − β)/α]/(g₁ + g₂).
- Common slope: s = g₂/(g₁ + g₂), lying between p₁ and p₂.
- Accept when dₘ ≤ aₘ (rounded down), reject when dₘ ≥ rₘ (rounded up); plot the two parallel lines and the cumulative defectives as a step path.
Step 4: Illustration (p₁ = 0·05, α = 0·05, p₂ = 0·20, β = 0·10)
- g₁ = log 4 = 1·3863, g₂ = log(0·95/0·80) = 0·1719 (natural logs).
- h₁ = log(9·5)/1·5582 = 1·445, h₂ = log(18)/1·5582 = 1·855, s = 0·1103.
- At m = 20: accept if d ≤ 0, reject if d ≥ 5; at m = 40: accept if d ≤ 2, reject if d ≥ 7.
- The earliest acceptance is at m = 14 with no defectives; the earliest rejection at m = 3 with 3 defectives.
- Wald's approximate ASN is about 21 at p = 0·05 and 17 at p = 0·20, compared with n = 47 for the single sampling plan with the same risks.
Step 5: OC and ASN
- OC: L(p) = (A^h − 1)/(A^h − B^h), where h solves p(p₂/p₁)^h + (1 − p)[(1 − p₂)/(1 − p₁)]^h = 1; L(p₁) ≈ 1 − α, L(p₂) ≈ β, L(s) = h₂/(h₁ + h₂).
- ASN: E(m) = [L(p) log B + (1 − L(p)) log A] / [p g₁ − (1 − p) g₂]; it is largest near p = s.
- In practice the test is truncated at some maximum sample size.
Conclusion
The SPRT rests on the likelihood ratio and Wald's boundaries A = (1 − β)/α and B = β/(1 − α); for attributes these become two parallel lines whose intercepts h₁, h₂ and slope s are set by p₁, p₂, α and β, giving the acceptance and rejection numbers at each stage with, on average, far fewer inspections than single sampling.
Facts to remember
- SPRT boundaries: A ≈ (1 − β)/α, B ≈ β/(1 − α).
- Acceptance number aₘ = −h₁ + sm; rejection number rₘ = h₂ + sm.
- h₁ = log[(1 − α)/β]/(g₁ + g₂), h₂ = log[(1 − β)/α]/(g₁ + g₂), s = g₂/(g₁ + g₂).
- Wald-Wolfowitz: the SPRT minimises expected sample size at both hypotheses among tests with the same α and β.