Chapter 8 — Sample Surveys and Design of Experiments
Data do not arrive by accident: someone decides which units to observe and which treatments to apply where. This chapter is about making those decisions well, with probability sampling designs for finite populations and randomised designs for experiments, so that estimates are unbiased, their errors can be computed, and a fixed budget buys as much precision as possible.
In 1936 an American magazine, the Literary Digest, mailed straw-poll ballots to millions of people taken from telephone directories and car-registration lists, and received about $2.4$ million replies. It predicted a comfortable victory for the challenger Alf Landon. Franklin Roosevelt won $46$ of the $48$ states. A young pollster, George Gallup, called the result correctly with a sample of a few tens of thousands chosen to mirror the electorate. The lesson has been relearned many times since. A huge sample from the wrong frame measures the frame very precisely, and the population not at all. Two years earlier Jerzy Neyman had shown, in a paper on "the representative method", that only random selection lets the statistician compute the error of an estimate, and that dividing the population into strata and allocating the sample cleverly can make that error much smaller. In India, P. C. Mahalanobis was building large-scale crop surveys on these ideas in the late 1930s, and the National Sample Survey he helped found in 1950 still produces the country's figures on consumption, employment and health.
The second half of the chapter moves from observing to intervening. At the Rothamsted agricultural station in the 1920s, R. A. Fisher faced field trials in which fertility varied from one corner of a field to another. He answered with three principles. Randomise the allocation of treatments to plots, so that no hidden pattern can masquerade as a treatment effect. Replicate, so that the error can be estimated. Use local control (blocking), so that the comparison is made within homogeneous groups. The same principles organise clinical trials, industrial experiments and A/B tests on websites today. Sections 8.1 and 8.2 cover sample surveys: simple random, stratified and systematic sampling, then sampling with probability proportional to size and the ratio and regression estimators. Sections 8.3 and 8.4 cover the design of experiments: the completely randomised, randomised block and Latin square designs, then incomplete block designs, BIBDs and $2^k$ factorial experiments with confounding.
- compute the mean and variance of the sample mean under simple random sampling with and without replacement, use the finite population correction, and determine the sample size for a required precision;
- estimate means, totals and proportions with standard errors and approximate confidence intervals;
- analyse stratified sampling, derive proportional, Neyman and cost-optimum allocations, and decide exactly when stratification gains;
- compute the variance of systematic sampling, compare it with random and stratified sampling for trends and periodic populations, and explain why it has no unbiased variance estimator;
- select units with probability proportional to size (cumulative total and Lahiri's methods), and use the Hansen–Hurwitz and Horvitz–Thompson estimators with their variances;
- derive the bias and mean squared error of the ratio and regression estimators and decide which estimator suits a given population;
- state the principles of experimental design, and analyse completely randomised, randomised block and Latin square designs, including efficiency and a missing plot;
- work with the incidence and C-matrices of block designs, the parameter relations of BIBDs and Fisher's inequality;
- estimate main effects and interactions in $2^2$ and $2^3$ experiments with Yates' algorithm, and construct confounded block arrangements.
How this is examined
CSIR NET draws on this chapter for both parts of the paper. In Part B the typical question asks for one number: the variance of a sample mean with or without replacement, a sample size, a Neyman or proportional allocation, the variance of a systematic sample under a linear trend, a Hansen–Hurwitz estimate or the expected number of trials in Lahiri's method, the first-order MSE of a ratio estimator, the error degrees of freedom in an RBD or LSD, the parameters of a BIBD from three given ones, or a main effect from the treatment totals of a $2^3$ experiment. In Part C, where several options may be correct, the questions test the structure. Is an estimator unbiased under a given design? Is $s^2$ unbiased for $S^2$ or for $\sigma^2$? Can stratification or systematic sampling be worse than simple random sampling? Is a variance estimator always non-negative? When does the ratio estimator beat the sample mean? Which parameter sets can be those of a BIBD? Which interaction is confounded in a given block arrangement? The wrong options are almost always taken from the ⚠️ boxes of the four sections. GATE ST examines the same material, often as numerical-answer questions.
Sections in this chapter
Simple Random, Stratified and Systematic Sampling
Finite populations and inclusion probabilities; SRSWR and SRSWOR, the finite population correction, proportions and sample size; stratified sampling with proportional, Neyman and cost-optimum allocation and the gain in precision; systematic sampling, trends and periodicity.
PPS Sampling; Ratio and Regression Estimators
Probability proportional to size with replacement (Hansen–Hurwitz), cumulative total and Lahiri's methods; inclusion probabilities and the Horvitz–Thompson estimator; the ratio estimator, its bias and MSE; the regression estimator and comparisons.
Completely Randomised, Randomised Block and Latin Square Designs
Randomisation, replication and local control; the CRD, RBD and LSD with their models and ANOVA tables; efficiency comparisons; the missing plot technique.
Block Designs, BIBD and 2ᵏ Factorial Experiments
Incidence matrices, the C-matrix, connectedness, completeness and orthogonality; BIBD parameter relations, Fisher's inequality and symmetric designs; $2^2$ and $2^3$ factorials, Yates' algorithm, total and partial confounding.
From probability you need expectation, variance and covariance of sums of dependent random variables, indicator variables, conditional expectation and the tower property (§1.3, §1.4), and the hypergeometric distribution (§2.1). From inference you need sampling distributions and standard errors (§4.1), bias, mean squared error and the Rao–Blackwell theorem (§4.2), confidence intervals (§4.3), and the central limit theorem and delta method (§2.3). Sections 8.3 and 8.4 rest on the linear model $y=X\beta+\varepsilon$, estimability, the Gauss–Markov theorem and $F$-tests (§6.1), and on the analysis of variance (§6.2). Section 8.2 also uses simple regression (§6.3). Matrix algebra (rank, eigenvalues, generalised inverses) is covered in the sister course CSIR NET Unit 1: Analysis and Linear Algebra. Throughout, population quantities are in capitals ($Y_i$, $\bar Y$, $S^2$) and sample quantities in lower case ($y_i$, $\bar y$, $s^2$), as is standard in survey sampling.