A population of N = 4 units has values y₁ = 10, y₂ = 20, y₃ = 30 and y₄ = 40. Consider a sample of size n = 2 without replacement with inclusion probabilities proportional to a size measure xᵢ, where x₁ = 1, x₂ = 2, x₃ = 3, x₄ = 4. The inclusion probabilities are taken as
πᵢ = nxᵢ/∑xᵢ; i = 1, 2, 3, 4
(i) Suppose the selected sample is {2, 4}. Using Horvitz-Thompson estimator, estimate the population total.
(ii) Given the joint inclusion probabilities π₂₄ = 0·3 for units 2 and 4, estimate the variance of Horvitz-Thompson estimator and comment on the possibility of negative variance estimate.
15 marks · directive: Estimate
The core of a good answer
πᵢ = 0.2, 0.4, 0.6, 0.8, so Ŷ_HT = 20/0.4 + 40/0.8 = 50 + 50 = 100; the Horvitz-Thompson variance estimator gives 2000 − 333.33 = 1666.67, while the Sen-Yates-Grundy estimator gives 0 (since y₂/π₂ = y₄/π₄); the HT form can be negative when πᵢⱼ < πᵢπⱼ and the cross term dominates, whereas the SYG form is non-negative when πᵢπⱼ ≥ πᵢⱼ, as here (0.32 ≥ 0.3).
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
Compute the inclusion probabilities, then the HT estimate ∑yᵢ/πᵢ over the sample. For the variance estimate, apply the Horvitz-Thompson unbiased variance estimator (and, for a fixed-size design, the Sen-Yates-Grundy form), and discuss when each can be negative. Marks 7 + 8.
Answer outline
Introduction
∑xᵢ = 10, n = 2: π₁ = 0.2, π₂ = 0.4, π₃ = 0.6, π₄ = 0.8 (sum = 2 = n). Sample s = {2, 4}; π₂π₄ = 0.32, π₂₄ = 0.3.
Part (i): HT estimate
- Ŷ_HT = ∑_{i∈s} yᵢ/πᵢ = 20/0.4 + 40/0.8 = 50 + 50 = 100.
- The true total is 10 + 20 + 30 + 40 = 100.
Part (ii): HT variance estimator
- v_HT = ∑_{i∈s}(1 − πᵢ)yᵢ²/πᵢ² + 2∑_{i<j∈s}[(πᵢⱼ − πᵢπⱼ)/(πᵢⱼπᵢπⱼ)]yᵢyⱼ.
- Unit 2: (0.6/0.16) × 400 = 1500; unit 4: (0.2/0.64) × 1600 = 500.
- Cross term: 2 × (0.3 − 0.32)/(0.3 × 0.32) × 20 × 40 = 2 × (−0.2083) × 800 = −333.33.
- v_HT = 2000 − 333.33 = 1666.67.
Part (ii): Sen-Yates-Grundy estimator
- v_SYG = ∑_{i<j∈s}[(πᵢπⱼ − πᵢⱼ)/πᵢⱼ](yᵢ/πᵢ − yⱼ/πⱼ)² = (0.02/0.3)(50 − 50)² = 0.
Comment
- Here yᵢ = 10xᵢ exactly, so yᵢ/πᵢ = 50 for every unit and every sample of size 2 gives Ŷ_HT = 100: the true variance is 0. The SYG estimate reflects this; the HT-form estimate (1666.67) is unbiased only on average over samples and is unstable.
- v_HT can be negative: its cross terms are negative whenever πᵢⱼ < πᵢπⱼ and yᵢyⱼ > 0, and can outweigh the positive terms.
- v_SYG is non-negative for every sample if πᵢπⱼ − πᵢⱼ ≥ 0 for all pairs (the Yates-Grundy condition); here 0.32 − 0.3 = 0.02 > 0.
- Both estimators require πᵢⱼ > 0 for every pair in the sample.
Conclusion
Ŷ_HT = 100 (exactly the true total). The variance estimate is 1666.67 by the Horvitz-Thompson formula and 0 by the Sen-Yates-Grundy formula; the latter is the sensible one here and cannot be negative under the stated condition. Arithmetic checked in Python.
Examples and data to cite
- One design consistent with these πᵢ and π₂₄ = 0.3 is π₁₂ = 0.09, π₁₃ = 0.10, π₁₄ = 0.01, π₂₃ = 0.01, π₂₄ = 0.30, π₃₄ = 0.49; under it the HT-form estimate is −112500 for the sample {2, 3} and −72500 for {1, 4}, so negative values do occur (computed in Python).
Facts to remember
- Horvitz-Thompson (1952): Ŷ_HT = ∑_{i∈s} yᵢ/πᵢ, unbiased when all πᵢ > 0.
- For a fixed-size design ∑πᵢ = n.
- Sen-Yates-Grundy variance estimator for fixed-size designs; non-negative if πᵢπⱼ ≥ πᵢⱼ.
- If yᵢ ∝ πᵢ, the HT estimator has zero variance.