PPS Sampling, Ratio and Regression Estimators
CSIR NET sampling: PPS selection, Hansen-Hurwitz and Horvitz-Thompson estimators, Yates-Grundy variance, and ratio, product and regression estimators.
- 20 min
On this page · 14 sections
- Prerequisites
- The Hook: Four Farms, One Draw
- Using a Size Measure: PPS Selection
- PPS With Replacement: the Hansen-Hurwitz Estimator
- PPS Without Replacement: the Horvitz-Thompson Estimator
- Ratio Estimators
- Regression Estimators
- Memory Tricks
- Don’t Confuse
- Common Mistakes
- Practice
- Quick Revision Table
- Key Takeaways
- What’s Next
Prerequisites
- Simple Random, Stratified and Systematic Sampling: \( S^2 \), the fpc \( 1 - f \), and inclusion probabilities \( \pi_i, \pi_{ij} \) are used throughout.
- Descriptive Statistics and Exploratory Data Analysis: correlation \( \rho \), coefficients of variation \( C_x, C_y \), and the least-squares slope \( \frac{S_{xy}}{S_x^2} \).
- Sample Space, Independence and Bayes’ Theorem: draw-by-draw probabilities for without-replacement selection.
- The chapter map is on Sampling Theory and Design of Experiments.
The Hook: Four Farms, One Draw
Using a Size Measure: PPS Selection
Each unit carries a known positive size \( X_i \) (area, last census count, turnover). Set \( P_i = \frac{X_i}{X} \), \( X = \sum X_i \). In PPS sampling with replacement, each draw selects unit \( i \) with probability \( P_i \).
Cumulative total method
List the cumulative sizes \( T_i = X_1 + \dots + X_i \). Draw a random integer \( R \) from \( 1, \dots, X \); select unit \( i \) with \( T_{i-1} < R \le T_i \). Unit \( i \) owns exactly \( X_i \) of the \( X \) integers, so it is selected with probability \( \frac{X_i}{X} \). (Sizes are taken as integers, after scaling if needed.)
Lahiri’s method
Cumulating is tedious for large \( N \). Instead, let \( M \ge \max X_i \). Draw a pair \( (i, j) \) with \( i \) uniform on \( 1, \dots, N \) and \( j \) uniform on \( 1, \dots, M \). Accept unit \( i \) if \( j \le X_i \); otherwise reject the pair and draw again.
On one trial, \( P(\text{unit } i \text{ accepted}) = \frac{1}{N}\cdot\frac{X_i}{M} \), so conditional on acceptance the probability is \( \frac{X_i}{\sum X_k} = P_i \). The acceptance probability per trial is \( \frac{\bar X}{M} \), so the expected number of trials per selection is \( \frac{M}{\bar X} \): take \( M = \max X_i \) to waste as few draws as possible.
PPS With Replacement: the Hansen-Hurwitz Estimator
Draw \( n \) units independently with probabilities \( P_i \). With \( y_r, p_r \) the value and selection probability of the unit chosen at draw \( r \),
$$\hat Y_{HH} = \frac{1}{n}\sum_{r=1}^n \frac{y_r}{p_r}.$$Each term \( \frac{y_r}{p_r} \) has expectation \( \sum_i P_i\frac{Y_i}{P_i} = Y \), so \( \hat Y_{HH} \) is unbiased, and the terms are i.i.d., so
$$V(\hat Y_{HH}) = \frac{1}{n}\sum_{i=1}^N P_i\left(\frac{Y_i}{P_i} - Y\right)^2, \qquad v(\hat Y_{HH}) = \frac{1}{n(n - 1)}\sum_{r=1}^n\left(\frac{y_r}{p_r} - \hat Y_{HH}\right)^2,$$the second being unbiased for the first.
- Zero variance when \( Y_i \propto X_i \) exactly, since then \( \frac{Y_i}{P_i} = Y \) for every \( i \).
- PPS can lose when \( Y \) is unrelated to size (or inversely related): a small unit with a large \( Y_i \) gives a huge \( \frac{Y_i}{P_i} \).
- With \( P_i = \frac{1}{N} \), \( \hat Y_{HH} = N\bar y \) and the formula reduces to the SRSWR variance \( \frac{N^2\sigma^2}{n} \).
Hook check. With \( n = 1 \), \( V = 0.1(10)^2 + 0.2(5)^2 + 0.3(6.67)^2 + 0.4(5)^2 = 10 + 5 + 13.33 + 10 = 38.33 \). Equal probability gives \( N^2\sigma^2 = 16 \times 112.5 = 1800 \).
PPS Without Replacement: the Horvitz-Thompson Estimator
For any design without replacement, let \( \pi_i > 0 \) be the first-order and \( \pi_{ij} \) the second-order inclusion probabilities. The Horvitz-Thompson estimator of the total is
$$\hat Y_{HT} = \sum_{i \in s}\frac{y_i}{\pi_i}.$$Writing \( \hat Y_{HT} = \sum_{i=1}^N a_i\frac{Y_i}{\pi_i} \) with \( E(a_i) = \pi_i \) shows it is unbiased. Its variance is
$$V(\hat Y_{HT}) = \sum_{i=1}^N \frac{1 - \pi_i}{\pi_i}Y_i^2 + \sum_{i \ne j}\frac{\pi_{ij} - \pi_i\pi_j}{\pi_i\pi_j}Y_iY_j.$$Fixed sample size identities
If every sample has exactly \( n \) units:
$$\sum_{i=1}^N \pi_i = n, \qquad \sum_{j \ne i}\pi_{ij} = (n - 1)\pi_i, \qquad \sum_{i \ne j}\pi_{ij} = n(n - 1).$$Using them, the variance takes the Yates-Grundy (Sen-Yates-Grundy) form
$$V(\hat Y_{HT}) = \sum_{i < j}(\pi_i\pi_j - \pi_{ij})\left(\frac{Y_i}{\pi_i} - \frac{Y_j}{\pi_j}\right)^2,$$with unbiased estimator \( v_{YG} = \sum_{i < j \in s}\frac{\pi_i\pi_j - \pi_{ij}}{\pi_{ij}}\left(\frac{y_i}{\pi_i} - \frac{y_j}{\pi_j}\right)^2 \) (needs all \( \pi_{ij} > 0 \)). The YG form makes two facts obvious:
- The variance is zero when \( \frac{Y_i}{\pi_i} \) is constant, i.e. \( \pi_i \propto Y_i \). That is why one wants \( \pi_i \propto X_i \) for a size \( X \) roughly proportional to \( Y \).
- \( v_{YG} \ge 0 \) for every sample when \( \pi_i\pi_j \ge \pi_{ij} \) for all pairs.
Under SRSWOR (\( \pi_i = \frac{n}{N} \)), \( \hat Y_{HT} = N\bar y \) and the formula gives back \( N^2(1 - f)\frac{S^2}{n} \).
Worked example: a three-unit design
\( N = 3 \), \( n = 2 \), with sample probabilities \( P(\{1, 2\}) = 0.5 \), \( P(\{1, 3\}) = 0.3 \), \( P(\{2, 3\}) = 0.2 \). Then \( \pi_1 = 0.8 \), \( \pi_2 = 0.7 \), \( \pi_3 = 0.5 \) (sum 2), and \( \pi_{12} = 0.5 \), \( \pi_{13} = 0.3 \), \( \pi_{23} = 0.2 \). Let \( Y = (8, 7, 4) \), total 19.
| Sample | Probability | \( \hat Y_{HT} \) |
|---|---|---|
| \( \{1, 2\} \) | 0.5 | \( 10 + 10 = 20 \) |
| \( \{1, 3\} \) | 0.3 | \( 10 + 8 = 18 \) |
| \( \{2, 3\} \) | 0.2 | \( 10 + 8 = 18 \) |
\( E(\hat Y_{HT}) = 10 + 5.4 + 3.6 = 19 \), and \( V = 0.5(1)^2 + 0.5(1)^2 = 1 \). The YG form agrees: \( \frac{Y_i}{\pi_i} = 10, 10, 8 \), so only the pairs \( (1, 3) \) and \( (2, 3) \) contribute: \( (0.4 - 0.3)(4) + (0.35 - 0.2)(4) = 0.4 + 0.6 = 1 \).
Inclusion probabilities for two draws without replacement
If two units are drawn successively without replacement, the first with probabilities \( p_i \) and the second with probabilities \( \frac{p_j}{1 - p_i} \) among the rest, then
$$\pi_i = p_i\left[1 + \sum_{j \ne i}\frac{p_j}{1 - p_j}\right], \qquad \pi_{ij} = p_ip_j\left[\frac{1}{1 - p_i} + \frac{1}{1 - p_j}\right].$$For \( p = (0.1, 0.2, 0.3, 0.4) \), \( \pi_1 = 0.1(1 + 0.25 + 0.4286 + 0.6667) \approx 0.2345 \). Note \( \pi_1 \ne 2p_1 = 0.2 \): draw-by-draw PPS without replacement does not give \( \pi_i \) exactly proportional to size. Special schemes are needed for that; for the exam, compute \( \pi_i \) from the scheme given.
Ratio Estimators
Now use \( x \) at the estimation stage. Take an SRSWOR of size \( n \), observe \( (y_i, x_i) \), and suppose \( \bar X \) (or \( X \)) is known. Write \( R = \frac{\bar Y}{\bar X} \), \( \hat R = \frac{\bar y}{\bar x} \).
$$\hat{\bar Y}_R = \frac{\bar y}{\bar x}\bar X, \qquad \hat Y_R = \frac{\bar y}{\bar x}X.$$Bias and MSE (to first order)
Put \( \bar y = \bar Y(1 + e_0) \), \( \bar x = \bar X(1 + e_1) \) and expand \( (1 + e_1)^{-1} \). With \( C_y, C_x \) the coefficients of variation (using \( S_y, S_x \)) and \( \rho \) the correlation,
$$B(\hat{\bar Y}_R) \approx \frac{1 - f}{n}\bar Y\left(C_x^2 - \rho C_xC_y\right),$$$$\operatorname{MSE}(\hat{\bar Y}_R) \approx \frac{1 - f}{n}\left(S_y^2 + R^2S_x^2 - 2R\rho S_xS_y\right) = \frac{1 - f}{n}\bar Y^2\left(C_y^2 + C_x^2 - 2\rho C_xC_y\right).$$Efficiency condition. Compare with \( V(\bar y) = \frac{1 - f}{n}\bar Y^2C_y^2 \):
$$\boxed{\hat{\bar Y}_R \text{ beats } \bar y \iff \rho > \frac{C_x}{2C_y} = \frac{1}{2}\cdot\frac{RS_x}{S_y}}$$If \( C_x \approx C_y \), the ratio estimator needs \( \rho > \frac{1}{2} \).
Exact facts worth knowing
- The bias is exact in one identity: \( \operatorname{Cov}(\hat R, \bar x) = E(\bar y) - E(\hat R)\bar X \), so \( B(\hat R) = -\frac{\operatorname{Cov}(\hat R, \bar x)}{\bar X} \). Hence \( \frac{\lvert B(\hat R) \rvert}{\sigma_{\hat R}} \le C_{\bar x} = \frac{\sigma_{\bar x}}{\bar X} \): the bias is negligible relative to the standard error when the CV of \( \bar x \) is small (say below 0.1).
- The bias vanishes if the population points lie on a line through the origin (\( Y_i = RX_i \)), since then \( \hat R = R \) for every sample.
- Model optimality. Under the model \( y_i = \beta x_i + \varepsilon_i \) with \( V(\varepsilon_i) \propto x_i \), the ratio estimator is the best linear unbiased predictor of the total. This is the case where it is “the” right estimator: a line through the origin and variance growing with \( x \).
- Hartley-Ross unbiased estimator. With \( r_i = \frac{y_i}{x_i} \) and \( \bar r \) their mean, \( \bar r\bar X + \frac{n(N - 1)}{N(n - 1)}(\bar y - \bar r\bar x) \) is exactly unbiased for \( \bar Y \) under SRSWOR.
Product estimator
If \( \rho < 0 \), use \( \hat{\bar Y}_P = \frac{\bar y\bar x}{\bar X} \), with \( \operatorname{MSE} \approx \frac{1 - f}{n}\bar Y^2(C_y^2 + C_x^2 + 2\rho C_xC_y) \). It beats \( \bar y \) iff \( \rho < -\frac{C_x}{2C_y} \).
Regression Estimators
$$\bar y_{lr} = \bar y + b(\bar X - \bar x).$$Fixed slope \( b = \beta_0 \)
\( \bar y_{lr} \) is unbiased (each of \( \bar y, \bar x \) is), with
$$V(\bar y_{lr}) = \frac{1 - f}{n}\left(S_y^2 - 2\beta_0S_{xy} + \beta_0^2S_x^2\right).$$This quadratic in \( \beta_0 \) is minimised at \( \beta_0 = B = \frac{S_{xy}}{S_x^2} \) (the population regression slope), giving
$$V_{min}(\bar y_{lr}) = \frac{1 - f}{n}S_y^2\left(1 - \rho^2\right).$$Special cases: \( \beta_0 = 0 \) gives \( \bar y \); \( \beta_0 = 1 \) gives the difference estimator \( \bar y + (\bar X - \bar x) \); \( \beta_0 = R \) reproduces (to first order) the ratio estimator’s MSE.
Estimated slope \( b = \frac{s_{xy}}{s_x^2} \)
Now \( \bar y_{lr} \) is biased (bias of order \( \frac{1}{n} \)), and its large-sample variance is the same \( \frac{1 - f}{n}S_y^2(1 - \rho^2) \).
The three-way comparison (large samples)
$$V(\bar y) - V(\bar y_{lr}) = \frac{1 - f}{n}\rho^2S_y^2 \ge 0, \qquad \operatorname{MSE}(\hat{\bar Y}_R) - V(\bar y_{lr}) = \frac{1 - f}{n}(B - R)^2S_x^2 \ge 0.$$- Regression is never worse than the mean per unit, and equal only when \( \rho = 0 \).
- Regression is never worse than the ratio estimator, and equal only when \( B = R \): the population regression line passes through the origin.
graph TD
A[Auxiliary x known for all units] --> B{Use it where?}
B -->|Design stage| C["PPS: pi_i proportional to X_i, then Hansen-Hurwitz or Horvitz-Thompson"]
B -->|Estimation stage| D{Shape of y on x}
D -->|"Line through origin, rho high"| E[Ratio estimator]
D -->|"Line with intercept"| F[Regression estimator]
D -->|"Negative correlation"| G[Product estimator]Worked example
\( N = 200 \), SRSWOR of \( n = 10 \): \( \bar y = 30 \), \( \bar x = 20 \), known \( \bar X = 22 \).
- Ratio: \( \hat{\bar Y}_R = \frac{30}{20}(22) = 33 \); total \( 200 \times 33 = 6600 \).
- Regression with \( b = 1.2 \): \( \bar y_{lr} = 30 + 1.2(2) = 32.4 \); total 6480.
- Mean per unit: \( \bar y = 30 \); total 6000.
All three agree that the sample drew low-\( x \) units and adjust upward, by different amounts.
Suppose the population has \( C_y = 0.3 \), \( C_x = 0.2 \), \( \rho = 0.8 \). In units of \( \frac{1 - f}{n}\bar Y^2 \):
| Estimator | Relative (MSE) | Efficiency vs \( \bar y \) |
|---|---|---|
| \( \bar y \) | \( C_y^2 = 0.090 \) | 1 |
| Ratio | \( 0.09 + 0.04 - 2(0.8)(0.06) = 0.034 \) | 2.65 |
| Regression | \( 0.09(1 - 0.64) = 0.0324 \) | 2.78 |
| Product | \( 0.09 + 0.04 + 0.096 = 0.226 \) | 0.40 |
Ratio beats \( \bar y \) since \( 0.8 > \frac{0.2}{2(0.3)} = 0.333 \).
Memory Tricks
Don’t Confuse
| This | vs | That |
|---|---|---|
| \( p_i \): selection probability on one draw (PPSWR) | \( \pi_i \): probability unit \( i \) is anywhere in the sample (WOR) | |
| HH estimator \( \frac{1}{n}\sum \frac{y_r}{p_r} \) | HT estimator \( \sum_{i \in s}\frac{y_i}{\pi_i} \) | |
| HT variance (general form, any design) | Yates-Grundy form (fixed sample size only) | |
| Ratio estimator: biased, \( O(\frac{1}{n}) \) bias | regression with fixed \( \beta_0 \): unbiased | |
| Ratio better than \( \bar y \) iff \( \rho > \frac{C_x}{2C_y} \) | regression better than \( \bar y \) iff \( \rho \ne 0 \) | |
| Cumulative total method: one draw, needs cumulative sums | Lahiri: pairs \( (i, j) \), rejection, no cumulation |
Common Mistakes
Practice
All questions below are original practice questions written for this page. They are not from any CSIR NET paper. In Level 3, as in Part C, one or more options may be correct.
Level 1: Foundation
- In Lahiri’s method with \( N = 5 \), \( M = 40 \) and \( X_3 = 25 \), the pair \( (3, 27) \) is drawn. Is unit 3 selected?
- For any fixed-size design without replacement with sample size \( n \), what is \( \sum_{i=1}^N \pi_i \)?
- State the condition under which the ratio estimator of \( \bar Y \) has smaller (first-order) MSE than \( \bar y \).
Answers
- No: \( j = 27 > X_3 = 25 \), so the pair is rejected and a new pair is drawn.
- \( n \), because \( \sum_i \pi_i = E(\text{number of units in the sample}) \).
- \( \rho > \frac{C_x}{2C_y} \).
Level 2: Part B Style
Practice Q1. In a PPSWR sample of size 2 from the farm population of the hook (\( p = 0.1, 0.2, 0.3, 0.4 \)), the draws give farm 3 (\( y = 32 \)) and farm 2 (\( y = 19 \)). The Hansen-Hurwitz estimate of the total yield is closest to
(a) 95.0 (b) 100.8 (c) 106.7 (d) 102.0
Answer: (b)
\( \frac{y}{p} = \frac{32}{0.3} = 106.67 \) and \( \frac{19}{0.2} = 95 \). \( \hat Y_{HH} = \frac{1}{2}(106.67 + 95) = 100.83 \). Its variance estimate is \( \frac{1}{2 \cdot 1}[(5.83)^2 + (5.83)^2] \approx 34.0 \). Option (d), 102, is \( 2(32 + 19) \), the equal-probability estimate \( N\bar y \).
Practice Q2. An SRSWOR of 10 units from 200 gives \( \bar y = 30 \), \( \bar x = 20 \). The population mean of \( x \) is 22. The ratio estimate of the population total of \( y \) is
(a) 6000 (b) 6480 (c) 6600 (d) 7260
Answer: (c)
\( \hat{\bar Y}_R = \frac{30}{20} \times 22 = 33 \); \( \hat Y_R = 200 \times 33 = 6600 \). (a) is \( N\bar y \); (b) is the regression estimate with \( b = 1.2 \); (d) is \( 6600 \times 1.1 \), a double adjustment.
Practice Q3. In a population, \( C_y = 0.3 \) and \( C_x = 0.2 \). The ratio estimator of \( \bar Y \) is more efficient (to first order) than the sample mean if and only if
(a) \( \rho > \frac{1}{3} \) (b) \( \rho > \frac{2}{3} \) (c) \( \rho > \frac{1}{2} \) (d) \( \rho > 0 \)
Answer: (a)
\( \rho > \frac{C_x}{2C_y} = \frac{0.2}{0.6} = \frac{1}{3} \). Option (b) is \( \frac{C_x}{C_y} \) (forgetting the 2); (c) is the rule of thumb for \( C_x = C_y \); (d) is the condition for the regression estimator.
Practice Q4. Two units are drawn successively without replacement from four units with initial probabilities \( p = (0.1, 0.2, 0.3, 0.4) \), the second draw being proportional to the remaining \( p \)’s. The inclusion probability of unit 1 is closest to
(a) 0.200 (b) 0.235 (c) 0.190 (d) 0.100
Answer: (b)
\( \pi_1 = p_1\left[1 + \frac{0.2}{0.8} + \frac{0.3}{0.7} + \frac{0.4}{0.6}\right] = 0.1(1 + 0.25 + 0.4286 + 0.6667) = 0.2345 \) (exactly \( \frac{197}{840} \)). Option (a) is \( 2p_1 \), the trap; (c) is \( 1 - 0.9^2 \), the with-replacement chance of appearing at least once.
Level 3: Part C Style
Practice Q5. A design on \( N = 3 \) units selects \( \{1, 2\}, \{1, 3\}, \{2, 3\} \) with probabilities 0.5, 0.3, 0.2. The values are \( Y = (8, 7, 4) \). Which statements are true for the Horvitz-Thompson estimator \( \hat Y_{HT} \) of the total?
(a) \( \pi_1 = 0.8 \), \( \pi_2 = 0.7 \), \( \pi_3 = 0.5 \). (b) \( \hat Y_{HT} \) takes the value 20 with probability 0.5 and 18 otherwise. (c) \( V(\hat Y_{HT}) = 1 \). (d) If \( Y_3 \) were 5 instead of 4 (other values unchanged), \( \hat Y_{HT} \) would have variance 0.
Answer: (a), (b), (c) and (d)
- (a) True. \( \pi_1 = 0.5 + 0.3 \), \( \pi_2 = 0.5 + 0.2 \), \( \pi_3 = 0.3 + 0.2 \); they sum to \( n = 2 \).
- (b) True. \( \frac{8}{0.8} + \frac{7}{0.7} = 20 \); \( \frac{8}{0.8} + \frac{4}{0.5} = 18 \); \( \frac{7}{0.7} + \frac{4}{0.5} = 18 \).
- (c) True. Mean 19; \( 0.5(1) + 0.5(1) = 1 \). The Yates-Grundy form gives \( (0.4 - 0.3)(10 - 8)^2 + (0.35 - 0.2)(10 - 8)^2 = 0.4 + 0.6 = 1 \).
- (d) True. Then \( \frac{Y_i}{\pi_i} = 10, 10, 10 \) are all equal, so every sample gives \( 20 \), the new total, and the YG variance is 0.
Practice Q6. For a population with \( C_y = 0.3 \), \( C_x = 0.2 \) and \( \rho = 0.8 \), SRSWOR of size \( n \) is used and first-order approximations apply. Which statements are true?
(a) The ratio estimator is about 2.65 times as efficient as \( \bar y \). (b) The regression estimator (with the optimum slope) is about 2.78 times as efficient as \( \bar y \). (c) The ratio and regression estimators are equally efficient here. (d) The product estimator is less efficient than \( \bar y \).
Answer: (a), (b) and (d)
In units of \( \frac{1 - f}{n}\bar Y^2 \): \( V(\bar y) = 0.09 \); ratio \( 0.09 + 0.04 - 0.096 = 0.034 \); regression \( 0.09 \times 0.36 = 0.0324 \); product \( 0.226 \).
- (a) True. \( \frac{0.09}{0.034} = 2.647 \).
- (b) True. \( \frac{1}{1 - \rho^2} = \frac{1}{0.36} = 2.78 \).
- (c) False. They tie only when \( B = R \). Here \( B = \rho\frac{S_y}{S_x} = 0.8 \times \frac{0.3\bar Y}{0.2\bar X} = 1.2R \ne R \). The gap \( \operatorname{MSE}_R - V_{lr} = \frac{1 - f}{n}(B - R)^2S_x^2 \) equals, in the same units, \( \frac{(0.2R)^2(0.2\bar X)^2}{\bar Y^2} = 0.04 \times 0.04 = 0.0016 \), which is exactly \( 0.034 - 0.0324 \).
- (d) True. \( 0.226 > 0.09 \); the product estimator needs negative correlation.
Practice Q7. For the four-farm population \( X = (1, 2, 3, 4) \), \( Y = (11, 19, 32, 38) \), samples of size \( n = 2 \) are drawn with replacement. Which statements are true?
(a) Under PPSWR with \( p_i \propto X_i \), \( V(\hat Y_{HH}) \approx 19.2 \). (b) Under SRSWR, the variance of \( N\bar y \) is 900. (c) If \( Y \) were \( (10, 20, 30, 40) \), the PPSWR estimator would have zero variance. (d) PPSWR with \( p_i \propto X_i \) would remain more efficient than SRSWR for every \( Y \) vector.
Answer: (a), (b) and (c)
- (a) True. For \( n = 1 \) the variance is \( \sum P_i(\frac{Y_i}{P_i} - 100)^2 = 10 + 5 + 13.33 + 10 = 38.33 \); for \( n = 2 \) halve it: 19.17.
- (b) True. \( \sigma^2 = \frac{1}{4}[(11 - 25)^2 + (19 - 25)^2 + (32 - 25)^2 + (38 - 25)^2] = \frac{196 + 36 + 49 + 169}{4} = 112.5 \); \( V(N\bar y) = \frac{N^2\sigma^2}{n} = \frac{16 \times 112.5}{2} = 900 \).
- (c) True. Then \( \frac{Y_i}{P_i} = 100 \) for every unit.
- (d) False. Take \( Y = (40, 30, 20, 10) \) (inversely related to size): \( \frac{Y_i}{P_i} = 400, 150, 66.7, 25 \), and the PPS variance (\( n = 1 \)) is \( 0.1(300)^2 + 0.2(50)^2 + 0.3(33.3)^2 + 0.4(75)^2 = 9000 + 500 + 333 + 2250 \approx 12083 \), far above the SRSWR value \( 16 \times 125 = 2000 \).
Quick Revision Table
| Idea | Fact |
|---|---|
| Cumulative total | random integer in \( 1..X \); unit owning it selected; prob \( \frac{X_i}{X} \) |
| Lahiri | \( (i, j) \), \( i \sim U\{1..N\} \), \( j \sim U\{1..M\} \); accept iff \( j \le X_i \); trials \( \frac{M}{\bar X} \) |
| Hansen-Hurwitz | \( \frac{1}{n}\sum\frac{y_r}{p_r} \); \( V = \frac{1}{n}\sum P_i(\frac{Y_i}{P_i} - Y)^2 \) |
| Horvitz-Thompson | \( \sum_{s}\frac{y_i}{\pi_i} \); unbiased for any design with all \( \pi_i > 0 \) |
| Fixed size | \( \sum\pi_i = n \); \( \sum_{j \ne i}\pi_{ij} = (n-1)\pi_i \) |
| Yates-Grundy | \( V = \sum_{i |
| Ratio MSE | \( \frac{1-f}{n}\bar Y^2(C_y^2 + C_x^2 - 2\rho C_xC_y) \); better iff \( \rho > \frac{C_x}{2C_y} \) |
| Ratio bias | \( \approx \frac{1-f}{n}\bar Y(C_x^2 - \rho C_xC_y) \); \( \frac{\lvert B \rvert}{\sigma} \le C_{\bar x} \) |
| Regression | \( \bar y + b(\bar X - \bar x) \); min \( V = \frac{1-f}{n}S_y^2(1 - \rho^2) \) at \( b = B \) |
| Product | \( \frac{\bar y\bar x}{\bar X} \); better iff \( \rho < -\frac{C_x}{2C_y} \) |
Key Takeaways
What’s Next
- CRD, RBD and Latin Square Designs: from observing a fixed population to controlling an experiment.
- Simple Random, Stratified and Systematic Sampling: revisit the SRS variance that every comparison on this page is measured against.
- Sampling and Design of Experiments: Quick Revision.