Describe sequential probability ratio test (SPRT). Find the SPRT for testing H₀ : θ = θ₀ against H₁ : θ = θ₁, using random observations from Poisson (θ). Also find the OC function for the test.
15 marks · directive: Describe
The core of a good answer
For θ₁ > θ₀ the SPRT continues sampling while [log B + m(θ₁ − θ₀)]/log(θ₁/θ₀) < ∑ᵢ₌₁ᵐ xᵢ < [log A + m(θ₁ − θ₀)]/log(θ₁/θ₀), with A ≈ (1 − β)/α and B ≈ β/(1 − α), rejecting H₀ above and accepting below; its OC function is L(θ) = (Aʰ − 1)/(Aʰ − Bʰ), traced parametrically by θ = h(θ₁ − θ₀)/[(θ₁/θ₀)ʰ − 1].
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
Three parts: Wald's SPRT in general (sequential sampling, likelihood ratio, two boundaries and their approximation through α and β), its specialisation to the Poisson as two parallel lines in the (m, ∑x) plane, and the OC function through Wald's h equation E_θ[(f₁/f₀)ʰ] = 1.
Answer outline
Introduction
Observations x₁, x₂, … arrive one at a time. λₘ = ∏ᵢ₌₁ᵐ f(xᵢ; θ₁)/f(xᵢ; θ₀). Strength (α, β).
Step 1: The SPRT
- At stage m: accept H₀ if λₘ ≤ B; reject H₀ if λₘ ≥ A; otherwise take another observation.
- Wald's approximations: A ≈ (1 − β)/α, B ≈ β/(1 − α).
- Wald and Wolfowitz (1948) later proved the SPRT optimal.
Step 2: Poisson log ratio
- log[f(x; θ₁)/f(x; θ₀)] = x log(θ₁/θ₀) − (θ₁ − θ₀).
- log λₘ = log(θ₁/θ₀) ∑xᵢ − m(θ₁ − θ₀).
Step 3: Decision rule (θ₁ > θ₀)
- Continue while a₀ + bm < ∑xᵢ < a₁ + bm, where b = (θ₁ − θ₀)/log(θ₁/θ₀), a₀ = log B/log(θ₁/θ₀), a₁ = log A/log(θ₁/θ₀).
- Accept H₀ when ∑xᵢ ≤ a₀ + bm; reject H₀ when ∑xᵢ ≥ a₁ + bm.
- If θ₁ < θ₀, log(θ₁/θ₀) < 0 and the inequalities reverse.
Step 4: OC function
- L(θ) = P_θ(accept H₀) ≈ (Aʰ − 1)/(Aʰ − Bʰ), where h = h(θ) ≠ 0 solves E_θ[(f(X; θ₁)/f(X; θ₀))ʰ] = 1.
- For the Poisson: E_θ[(θ₁/θ₀)^{hX}] e^{−h(θ₁ − θ₀)} = exp{θ[(θ₁/θ₀)ʰ − 1] − h(θ₁ − θ₀)} = 1.
- So θ = h(θ₁ − θ₀)/[(θ₁/θ₀)ʰ − 1]; letting h run over the real line gives the curve (θ, L(θ)).
Step 5: Check points
- h = 1 ⇒ θ = θ₀, L = (A − 1)/(A − B) ≈ 1 − α.
- h = −1 ⇒ θ = θ₁, L = β.
- h → 0 ⇒ θ = (θ₁ − θ₀)/log(θ₁/θ₀) = b, L = log A/(log A − log B).
- Python simulation with θ₀ = 1, θ₁ = 2, α = β = 0.05: the proportion of runs accepting H₀ was 0.974 at θ = 1, 0.425 at θ = 1.5 (Wald's OC gives 0.417) and 0.035 at θ = 2, in line with the approximations (actual error rates slightly below nominal).
Conclusion
The Poisson SPRT compares the running total ∑xᵢ with two parallel straight lines of slope (θ₁ − θ₀)/log(θ₁/θ₀); the OC curve falls from about 1 − α at θ₀ to about β at θ₁.
Examples and data to cite
- With θ₀ = 1, θ₁ = 2, α = β = 0.05: log A = log 19 = 2.944, log(θ₁/θ₀) = 0.693, so sampling continues while −4.25 + 1.443m < ∑xᵢ < 4.25 + 1.443m.
Facts to remember
- Wald's SPRT (Wald, 1945): boundaries A ≈ (1 − β)/α and B ≈ β/(1 − α).
- Poisson pmf: e^{−θ}θˣ/x!; m.g.f. E(e^{sX}) = exp[θ(e^s − 1)].
- OC function: L(θ) = (Aʰ − 1)/(Aʰ − Bʰ), h(θ) from E_θ[(f₁/f₀)ʰ] = 1.
- Optimality of the SPRT was proved by Wald and Wolfowitz (1948).