Statistics Paper II: Core Topics
Learning Objectives
- Understand the fundamental concepts of Statistics Paper II: Core Topics
- Apply key formulas and techniques to solve problems
- Practice with exam-level questions to build speed and accuracy
Design of Experiments
1.1 Principles of Experimental Design
The foundation of any scientific investigation lies in the careful planning of experiments so that the effects of treatments can be isolated from extraneous variation. Randomization ensures that each experimental unit has an equal chance of receiving any treatment, thereby distributing unknown confounding factors evenly across treatment groups. Replication provides an estimate of the experimental error and increases the precision of treatment comparisons, while local control — achieved through blocking — reduces variability by grouping homogeneous units together. These three principles, first formalized by Ronald A. Fisher in the 1920s, remain the cornerstone of modern experimental practice and are explicitly required in the UPSC Statistics Paper II syllabus.
1.2 Basic Designs: Completely Randomized, Randomized Block, Latin Square
The completely randomized design (CRD) assigns treatments to experimental units purely by chance, making it the simplest layout but often the least efficient when heterogeneity among units is present. The randomized block design (RBD) improves precision by forming blocks of similar units and randomizing treatments within each block, thereby removing block-to-block variation from the error term. The Latin square design (LSD) extends this idea to two blocking factors, arranging treatments in a square grid so that each treatment appears exactly once in each row and column; it is especially valuable in agricultural field trials where fertility gradients run in two perpendicular directions. The choice among these designs depends on the number of nuisance factors that can be identified and the practical constraints of the experiment.
| Design | Blocking Factors | Degrees of Freedom for Error | Typical Use Case |
|---|---|---|---|
| Completely Randomized Design | None | N − t | Homogeneous laboratory settings |
| Randomized Block Design | One | (b − 1)(t − 1) | Field trials with a single fertility gradient |
| Latin Square Design | Two | (t − 1)(t − 2) | Two‑dimensional heterogeneity such as row and column effects |
1.3 Factorial Experiments and Interaction Effects
Factorial experiments allow the simultaneous study of two or more factors, each at several levels, within a single experiment. The main advantage is the ability to detect interaction effects — situations where the effect of one factor depends on the level of another factor — which cannot be discovered in one‑factor‑at‑a‑time studies. A full factorial design with k factors at l levels each requires l^k treatment combinations; when this number becomes impractically large, fractional factorial designs based on defining relations and generators are employed to estimate a subset of effects while confounding higher‑order interactions assumed to be negligible. The analysis of variance (ANOVA) table for a factorial experiment partitions the total sum of squares into components attributable to each main effect, each interaction, and the residual error, enabling formal F‑tests for significance.
1.4 Analysis of Variance for Designed Experiments
ANOVA provides the statistical machinery to test whether observed differences among treatment means exceed what would be expected from random variation alone. In a CRD the model is Yij=μ+τ_i+εij, where τ_i represents the fixed treatment effect and εij ∼ N(0,σ²). The F‑statistic is the ratio of the mean square for treatments to the mean square for error. For RBD the model adds a block effect β_j, and the error degrees of freedom become (b − 1)(t − 1). In LSD the model includes both row and column effects, and the error degrees of freedom are (t − 1)(t − 2). The assumptions of normality, independence, and homoscedasticity of errors must be checked through residual plots and formal tests such as Levene’s test for equality of variances.
1.5 Missing Plot Techniques and Covariance Analysis
In practice some experimental units may be lost due to unforeseen circumstances, leading to missing observations. The Yates procedure estimates the missing value by minimizing the error sum of squares, iterating until convergence, and then proceeds with the usual ANOVA while adjusting the error degrees of freedom. Covariance analysis (ANCOVA) incorporates one or more concomitant variables measured on each unit before treatment application; by adjusting treatment means for these covariates, ANCOVA increases precision and can correct for initial differences among units. The model extends the ANOVA framework to Yij=μ+τ_i+β(xij−x)+εij, where β is the common regression coefficient assumed homogeneous across treatments.
Econometrics
2.1 Classical Linear Regression Model and Assumptions
The classical linear regression model (CLRM) postulates a linear relationship between a dependent variable Y and a set of explanatory variables X₁,…,X_k: Y = β₀ + β₁X₁ + … + β_kX_k + u. The Gauss‑Markov assumptions — linearity in parameters, random sampling, no perfect multicollinearity, zero conditional mean of the error term (E[u|X]=0), and homoscedasticity (Var(u|X)=σ²) — guarantee that the ordinary least squares (OLS) estimator is the best linear unbiased estimator (BLUE). When the normality assumption (u|X ∼ N(0,σ²)) is added, OLS becomes the maximum likelihood estimator and exact finite‑sample inference using t‑ and F‑distributions is valid.
2.2 Estimation Methods: OLS, GLS, Maximum Likelihood
OLS minimizes the sum of squared residuals and yields closed‑form solutions β̂ = (X′X)⁻¹X′Y. When heteroscedasticity or autocorrelation violates the spherical error assumption, generalized least squares (GLS) transforms the model by a weighting matrix Ω⁻¹ᐟ² so that the transformed errors satisfy the Gauss‑Markov conditions; feasible GLS (FGLS) replaces the unknown Ω with a consistent estimate. Maximum likelihood estimation (MLE) specifies the full distribution of the error term, typically normal, and maximizes the likelihood function; MLE is asymptotically efficient and provides a natural framework for hypothesis testing via likelihood‑ratio, Wald, and Lagrange multiplier tests.
2.3 Diagnostic Testing: Heteroscedasticity, Autocorrelation, Multicollinearity
Diagnostic checks are essential before trusting regression results. The Breusch‑Pagan and White tests detect heteroscedasticity by regressing squared residuals on the explanatory variables or their squares and cross‑products. The Durbin‑Watson statistic and the Breusch‑Godfrey LM test examine serial correlation in the residuals, the latter allowing for higher‑order autoregressive structures and lagged dependent variables. Multicollinearity, the near‑linear dependence among regressors, inflates the variance of OLS coefficients; it is diagnosed through variance inflation factors (VIF) where VIF > 10 signals a serious problem, and through condition indices exceeding 30. Remedial measures include variable dropping, principal component regression, ridge regression, or collecting additional data.
2.4 Dynamic Models: Distributed Lag, Autoregressive Distributed Lag
Economic relationships often involve lagged effects. A finite distributed lag model expresses the current dependent variable as a function of current and past values of an explanatory variable: Y_t = α + Σi=0^q β_i Xt‑i + u_t. The geometric (Koyck) lag assumes β_i = β₀λ^i, yielding a parsimonious autoregressive specification. The autoregressive distributed lag (ARDL) model combines lagged dependent variables with lagged explanatory variables, allowing for both short‑run dynamics and long‑run equilibrium relationships. The bounds testing procedure of Pesaran, Shin, and Smith (2001) provides a unified cointegration test within the ARDL framework irrespective of the integration order of the variables.
2.5 Panel Data Models: Fixed Effects vs Random Effects
Panel data combine cross‑sectional and time‑series dimensions, offering richer variation and the ability to control for unobserved heterogeneity. The fixed effects (within) estimator eliminates time‑invariant individual effects by demeaning each variable within each cross‑sectional unit, yielding consistent estimates even when the unobserved effect is correlated with the regressors. The random effects (GLS) estimator treats the individual effect as a random variable uncorrelated with the regressors, gaining efficiency but risking inconsistency if the orthogonality condition fails. The Hausman test compares the two estimators; a significant test statistic favors fixed effects. Recent extensions include correlated random effects (Mundlak) and dynamic panel GMM estimators (Arellano‑Bond) for models with lagged dependent variables.
| Feature | Fixed Effects | Random Effects |
|---|---|---|
| Assumption on individual effect | Correlated with regressors | Uncorrelated with regressors |
| Estimation method | Within transformation (demeaning) | Generalized least squares |
| Consistency | Always consistent | Consistent only if orthogonality holds |
| Efficiency | Less efficient (uses within variation only) | More efficient (uses both within and between) |
| Time‑invariant variables | Cannot be estimated | Can be estimated |
Multivariate Analysis
3.1 Multivariate Normal Distribution and Its Properties
The multivariate normal distribution generalizes the univariate normal to a p‑dimensional random vector X with mean vector μ and positive‑definite covariance matrix Σ. Its density function is f(x) = (2π)-p/2 |Σ|-1/2 exp{−½ (x−μ)′ Σ⁻¹ (x−μ)}. Key properties include: any linear combination a′X is univariate normal; marginal distributions of subsets of components are multivariate normal; conditional distributions are also multivariate normal with mean and covariance given by the Schur complement; and independence is equivalent to zero covariance. These properties underpin most classical multivariate techniques such as Hotelling’s T², MANOVA, and discriminant analysis.
3.2 Principal Component Analysis and Factor Analysis
Principal component analysis (PCA) seeks orthogonal linear combinations of the original variables that capture maximal variance. The first principal component is the eigenvector associated with the largest eigenvalue of the sample covariance (or correlation) matrix; subsequent components correspond to decreasing eigenvalues and are mutually uncorrelated. PCA is a descriptive tool for dimension reduction and does not assume an underlying statistical model. Factor analysis, by contrast, postulates a latent variable model X = ΛF + ε, where F are common factors, Λ is the loading matrix, and ε are unique factors. The communalities h_i² = Σ_j λij² represent the proportion of variance explained by the common factors. Estimation methods include principal factor, maximum likelihood, and generalized least squares; rotation (varimax, oblimin) aids interpretability.
3.3 Discriminant Analysis and Classification
Linear discriminant analysis (LDA) constructs linear combinations of predictors that best separate two or more predefined groups. For two groups with common covariance Σ, the discriminant function is δ(x) = x′ Σ⁻¹ (μ₁−μ₂) − ½ (μ₁+μ₂)′ Σ⁻¹ (μ₁−μ₂) + log(π₁/π₂), where π_k are prior probabilities. Classification assigns an observation to the group with the highest posterior probability. Quadratic discriminant analysis (QDA) relaxes the equal‑covariance assumption, leading to quadratic decision boundaries. The performance of a classifier is assessed via the confusion matrix, apparent error rate, and cross‑validated error estimates such as leave‑one‑out or k‑fold cross‑validation.
3.4 Canonical Correlation Analysis
Canonical correlation analysis (CCA) explores the linear relationships between two sets of variables, X (p variables) and Y (q variables), by finding pairs of linear combinations U = a′X and V = b′Y that have maximum correlation. The first canonical correlation is the largest eigenvalue of ΣXX−1 ΣXY ΣYY−1 ΣYX; subsequent canonical variates are orthogonal to previous ones. Significance of each canonical correlation is tested using Wilks’ lambda, which follows an approximate χ² distribution under the null of zero correlation. CCA is widely used in psychometrics, genomics, and economics to uncover shared latent structures across variable blocks.
3.5 Cluster Analysis: Hierarchical and Non‑hierarchical Methods
Cluster analysis groups objects so that intra‑cluster similarity is high and inter‑cluster similarity is low. Hierarchical methods build a dendrogram either agglomeratively (starting with singletons and merging) or divisively (starting with all objects and splitting). Common linkage criteria are single, complete, average, and Ward’s minimum variance linkage. Non‑hierarchical methods, notably k‑means, partition the data into a pre‑specified number of clusters by iteratively assigning objects to the nearest centroid and updating centroids. The choice of distance metric (Euclidean, Mahalanobis, correlation‑based) and the determination of the optimal number of clusters (elbow method, silhouette width, gap statistic) are critical practical decisions.
Official Statistics
4.1 Statistical System in India: CSO, NSSO, MOSPI
India’s official statistical architecture is coordinated by the Ministry of Statistics and Programme Implementation (MOSPI). The Central Statistics Office (CSO) is responsible for macroeconomic aggregates such as national accounts, industrial production indices, and price indices. The National Sample Survey Office (NSSO) conducts large‑scale household surveys on employment, consumer expenditure, health, and education. The National Statistical Commission (NSC), established in 2005, provides oversight and ensures adherence to the United Nations Fundamental Principles of Official Statistics. Coordination among these bodies is essential for data consistency, timeliness, and credibility.
4.2 National Accounts Statistics: GDP, GNP, NDP, NNP
Gross domestic product (GDP) measures the market value of all final goods and services produced within the geographic boundary of a country during a given period. Gross national product (GNP) adds net factor income from abroad (NFIA) to GDP, thus reflecting the income earned by residents regardless of location. Net domestic product (NDP) subtracts consumption of fixed capital (depreciation) from GDP, while net national product (NNP) subtracts depreciation from GNP. At factor cost, indirect taxes are removed and subsidies added; at market prices, they are included. The CSO releases quarterly GDP estimates at constant (2011‑12) prices and current prices, enabling real growth analysis.
4.3 Price Indices: CPI, WPI, GDP Deflator
The consumer price index (CPI) tracks the average change in prices paid by urban and rural households for a fixed basket of goods and services; it is the primary inflation gauge for monetary policy. The wholesale price index (WPI) measures price changes at the first commercial transaction stage, covering primary articles, fuel and power, and manufactured products. The GDP deflator, defined as (Nominal GDP / Real GDP) × 100, reflects price changes for all domestically produced final goods and services and automatically updates the basket each period. Differences in coverage, weighting, and base year cause the three indices to diverge, a fact frequently examined in policy analysis.
4.4 Survey Methodology: Sampling Design, Estimation, Non‑response
Large‑scale surveys such as the NSSO’s Employment‑Unemployment Survey employ stratified multi‑stage sampling: first‑stage units are villages or urban blocks selected with probability proportional to size; second‑stage units are households selected systematically within each first‑stage unit. Estimation uses design‑based weights equal to the inverse of the selection probability, calibrated to known population totals (post‑stratification). Non‑response is addressed through weighting adjustments (non‑response adjustment cells) and imputation methods such as hot‑deck or regression imputation. Variance estimation for complex designs relies on Taylor linearization or replication methods (jackknife, bootstrap, balanced repeated replication).
4.5 Quality Frameworks and Dissemination Standards
The quality of official statistics is assessed against dimensions of relevance, accuracy, timeliness, accessibility, interpretability, and coherence, as articulated in the UN’s Generic Statistical Business Process Model (GSBPM) and the IMF’s Data Quality Assessment Framework (DQAF). India subscribes to the Special Data Dissemination Standard (SDDS) and the enhanced General Data Dissemination System (e‑GDDS), which prescribe metadata provision, advance release calendars, and simultaneous release to all users. The NSC’s “Code of Practice” mandates transparency in methodology, revision policies, and user engagement. Adherence to these frameworks ensures that Indian statistics are internationally comparable and trusted by policymakers, researchers, and the public.
The comprehensive coverage of experimental design, econometric modelling, multivariate techniques, and the institutional framework of official statistics equips candidates to answer both theoretical and applied questions in Statistics Paper II with confidence.
Split‑Plot and Repeated Measures Designs
Split‑plot designs arise when two levels of experimental units are present, typically because one factor is hard to randomise completely. The whole‑plot factor is applied to large units (whole plots) and the subplot factor to smaller units nested within each whole plot. The linear model for a two‑factor split‑plot with a completely randomised whole‑plot structure is Yijk = μ + τi + βj + (τβ)ij + εik + δijk, where εik is the whole‑plot error (variance σ²W) and δijk is the subplot error (variance σ²S). Because the two error terms have different variances, the appropriate F‑tests use different denominators: the whole‑plot factor is tested against the whole‑plot mean square, while the subplot factor and interaction are tested against the subplot mean square. Ignoring this hierarchy inflates Type I error for the whole‑plot factor.
Repeated measures designs are a special case of split‑plot where the same experimental unit receives all levels of the within‑subject factor over time. The covariance structure of the repeated observations (compound symmetry, autoregressive, unstructured) must be modelled. The classical univariate approach assumes sphericity (equal variances of differences) and uses the Greenhouse‑Geisser or Huynh‑Feldt corrections when violated. The multivariate approach (MANOVA) avoids sphericity but loses power when the number of time points is large relative to the sample size. Mixed‑model formulations treat time as a fixed effect and subject as a random effect, allowing flexible covariance structures (e.g., AR(1), Toeplitz) and handling missing observations under the missing‑at‑random assumption.
Practical implementation in statistical software requires specifying the correct error terms. In R, aov(Y ~ A*B + Error(Subject/A)) fits a split‑plot with A as whole‑plot factor and B as subplot factor. For repeated measures, lme(Y ~ Time*Group, random = ~1|Subject, correlation = corAR1(form = ~Time|Subject)) fits a linear mixed model with AR(1) correlation. Diagnostics include residual plots at each level, checking normality of whole‑plot and subplot residuals separately, and comparing AIC across covariance structures.
Examiners often test the distinction between the two error strata and the consequences of pooling them. A typical question provides an ANOVA table with only one error term and asks to reconstruct the correct split‑plot table, compute the appropriate F‑statistics, and comment on the validity of conclusions drawn from the pooled analysis.
Unit Root Testing, Cointegration and Vector Autoregression
Time‑series econometrics in Paper II extends beyond static regression to dynamic systems where variables may be non‑stationary. A series yt has a unit root if its autoregressive representation yt = ρ yt‑1 + εt has ρ = 1. The Augmented Dickey‑Fuller (ADF) test regresses Δyt on yt‑1 and lagged differences Δyt‑1…Δyt‑p to absorb serial correlation. The null hypothesis is ρ = 1 (unit root); critical values are non‑standard (Dickey‑Fuller tables). The Phillips‑Perron test corrects for serial correlation and heteroscedasticity non‑parametrically using Newey‑West standard errors. The KPSS test reverses the null (stationarity) and is useful as a confirmatory check.
When two or more I(1) series share a common stochastic trend, they are cointegrated. The Engle‑Granger two‑step procedure first estimates the long‑run equilibrium yt = β xt + ut by OLS, then tests the residuals ut for a unit root using ADF critical values adjusted for the generated regressor. The Johansen maximum‑likelihood approach estimates a Vector Error Correction Model (VECM) ΔYt = Π Yt‑1 + Σ Γi ΔYt‑i + εt, where the rank of Π determines the number of cointegrating vectors. The trace and maximum‑eigenvalue statistics test r = 0, 1,… against the alternative r > 0. Identifying the cointegrating rank is crucial: over‑specifying r leads to spurious error‑correction terms, under‑specifying omits long‑run information.
A Vector Autoregression (VAR) of order p, Yt = A1 Yt‑1 + … + Ap Yt‑p + εt, captures the joint dynamics without imposing theoretical restrictions. Impulse response functions (IRFs) trace the effect of a one‑standard‑deviation shock to one variable on the system over time; orthogonalised IRFs use the Cholesky decomposition of the residual covariance matrix, making results ordering‑dependent. Forecast error variance decomposition (FEVD) attributes the h‑step‑ahead forecast variance of each variable to structural shocks. Structural VAR (SVAR) imposes identifying restrictions (short‑run, long‑run, or sign restrictions) to give economic meaning to shocks.
Common exam traps include applying standard t‑tests on coefficients of a VAR with I(1) variables (spurious regression), mis‑specifying the lag length (use AIC, BIC, HQIC, and check residual autocorrelation), and confusing the null hypotheses of ADF (unit root) and KPSS (stationarity). Candidates must also demonstrate the ability to interpret a VECM output: the α matrix (adjustment speeds) and β matrix (cointegrating vectors) and to explain the economic interpretation of a significant error‑correction term.
Multivariate Analysis of Variance and Profile Analysis
Multivariate Analysis of Variance (MANOVA) extends ANOVA to multiple correlated response variables. The model Y = X B + E has Y (n × p) response matrix, X (n × k) design matrix, B (k × p) parameter matrix, and E (n × p) error matrix with rows i.i.d. Np(0, Σ). Hypotheses are linear restrictions on B, e.g., H₀: C B = 0. Four classic test statistics summarise the eigenvalues λi of H E⁻¹ (H = hypothesis SSCP, E = error SSCP): Wilks’ Λ = ∏(1+λi)⁻¹, Pillai’s trace = ∑ λi/(1+λi), Hotelling‑Lawley trace = ∑ λi, Roy’s largest root = max λi. Pillai’s trace is most robust to violations of covariance homogeneity; Roy’s root is most powerful when a single dimension dominates but sensitive to assumption breaches.
Profile analysis addresses repeated‑measures questions when the within‑subject factor is quantitative (e.g., dose levels, time points). It tests three hypotheses: (1) parallelism – all group profiles are parallel (no group × time interaction); (2) level – assuming parallelism, the average level across groups is equal; (3) flatness – assuming parallelism and equal level, the common profile is flat (no time effect). These are tested by constructing contrast matrices C for the within‑subject factor and applying MANOVA to the transformed responses Y C. The parallelism test is equivalent to the interaction test in a split‑plot MANOVA; the level test uses the between‑subject factor; the flatness test uses a contrast that sums to zero across time points.
Assumptions include multivariate normality, homogeneity of covariance matrices (Box’s M test), and independence. When Box’s M is significant, Pillai’s trace or a robust permutation MANOVA (e.g., using the vegan package) is recommended. Missing data in repeated measures can be handled by maximum‑likelihood estimation under MAR, or by multiple imputation before MANOVA. Effect size measures such as partial η² (for each response) and multivariate η² (based on Pillai’s trace) aid interpretation.
Examiners frequently ask to derive the contrast matrix for a given profile hypothesis, to compute Wilks’ Λ from given eigenvalues, or to decide which test statistic to report when Box’s M rejects homogeneity. A typical applied question provides a 3 × 4 design (three groups, four time points) with summary SSCP matrices and asks to test parallelism, level, and flatness sequentially, reporting the appropriate F‑approximations and degrees of freedom.
Administrative Data, Big Data and SDG Monitoring in Official Statistics
Official statistics increasingly supplement survey data with administrative registers (tax, social security, health records) and big‑data sources (satellite imagery, mobile‑phone metadata, scanner data). Administrative data offer near‑complete coverage, high frequency, and low marginal cost, but suffer from definition mismatches, coverage gaps (e.g., informal sector), and legal access restrictions. The Generic Statistical Business Process Model (GSBPM) integrates these sources in the “Collect” and “Process” phases, requiring robust record linkage (deterministic or probabilistic) and statistical matching when unique identifiers are absent.
Big‑data streams are characterised by volume, velocity, variety, and veracity. For price statistics, web‑scraped price quotes can update the CPI basket more frequently than traditional outlet surveys. Machine‑learning classifiers (e.g., random forests, gradient boosting) assign product codes to unstructured descriptions. However, selection bias (only online retailers) and concept drift (changing product attributes) demand model‑based calibration against a probability‑sample benchmark. The UN’s “Big Data for Official Statistics” framework recommends a quality assurance matrix covering relevance, accuracy, timeliness, accessibility, and coherence.
Sustainable Development Goal (SDG) monitoring requires disaggregated indicators (by sex, age, disability, geography). The Inter‑Agency and Expert Group on SDG Indicators (IAEG‑SDGs) defines a tier system: Tier I indicators have established methodology and data; Tier II have methodology but insufficient data; Tier III lack both. India’s National Indicator Framework (NIF) maps 297 national indicators to global SDG targets, leveraging NSSO surveys, administrative MIS, and the Periodic Labour Force Survey (PLFS). Data gaps are addressed through proxy indicators, small‑area estimation (Fay‑Herriot models), and citizen‑generated data validated by statistical offices.
Key exam topics include the distinction between design‑based (survey) and model‑based (big‑data) inference, the role of the Statistical Data and Metadata Exchange (SDMX) standard for interoperability, and the ethical‑legal framework (India’s Statistics Act 2008, Personal Data Protection Bill) governing data sharing. Candidates should be able to outline a hybrid estimation strategy: use administrative data for frame construction, big‑data for nowcasting, and probability surveys for benchmark calibration, with variance estimation via bootstrap replication that respects the complex design of each component.
Bootstrap and Simulation‑Based Inference for Complex Surveys
Analytic variance formulas for multistage stratified designs with calibration weighting are often intractable. Replication methods — Balanced Repeated Replication (BRR), Jackknife Repeated Replication (JRR), and Bootstrap — provide design‑consistent variance estimates by repeatedly re‑weighting the sample. The Rao‑Wu bootstrap for stratified multistage sampling draws primary sampling units (PSUs) with replacement within each stratum, preserving the original inclusion probabilities. For each replicate r, the replicate weight wi⁽ʳ⁾ = wi × (mh / (mh‑1)) × I(i ∈ PSUh⁽ʳ⁾), where mh is the number of PSUs in stratum h. The bootstrap variance of an estimator θ̂ is v̂boot(θ̂) = (1/R) Σ (θ̂⁽ʳ⁾ – θ̂)².
Calibration (post‑stratification, raking) adjusts weights to known population totals. The bootstrap must replicate the calibration step in each replicate: recompute calibrated weights wi⁽ʳ⁾ using the same control totals. This ensures that the variability from calibration is captured. For non‑response adjustments, the bootstrap replicates the response propensity model (e.g., logistic regression) on each replicate sample, preserving the two‑phase structure.
Simulation‑based inference extends to hypothesis testing and confidence intervals. The percentile bootstrap CI uses the α/2 and 1‑α/2 quantiles of the replicate distribution; the bias‑corrected and accelerated (BCa) interval corrects for bias and skewness. For complex estimators (e.g., Gini coefficient, quantile regression), the bootstrap is often the only feasible variance estimator. In small‑area estimation, parametric bootstrap draws random effects from the fitted Fay‑Herriot model, generates synthetic populations, and re‑estimates the Empirical Best Linear Unbiased Predictor (EBLUP) to obtain mean‑squared‑error estimates.
Exam questions may present a simplified two‑stage design with given PSU counts and ask to write the bootstrap weight formula, compute the number of replicates needed for a target Monte‑Carlo error, or compare BRR (requires Hadamard matrix, works for two PSUs per stratum) versus bootstrap (flexible but computationally heavier). Candidates should also discuss the “finite population correction” in bootstrap: when sampling fractions are large, the standard bootstrap overestimates variance; the mirror‑match bootstrap or the “with‑replacement” bootstrap with a correction factor (1‑f) addresses this.
- Confusing the whole‑plot and subplot error strata in split‑plot ANOVA leads to incorrect F‑tests.
- Applying standard t‑tests on VAR coefficients when variables are I(1) produces spurious significance.
- Using Wilks’ Λ without checking Box’s M can give misleading results under covariance heterogeneity.
- Treating administrative data as a probability sample ignores selection bias and coverage error.
- Forgetting to re‑calibrate weights inside each bootstrap replicate underestimates variance.
- Assuming the percentile bootstrap interval is always adequate; BCa is needed for skewed estimators.
- Review these traps before the exam to avoid common pitfalls.
Optimal Design Theory and Response Surface Methodology
Optimal design theory formalises the choice of experimental runs to maximise statistical efficiency. The Kiefer–Wolfowitz equivalence theorem states that a design is D‑optimal (maximising the determinant of the information matrix X′X) if and only if the maximum of the scaled prediction variance d(x) = p·x′(X′X)−1x over the design space equals the number of parameters p. A‑optimality minimises the trace of (X′X)−1, reducing average parameter variance, while G‑optimality minimises the maximum prediction variance. For linear models, the Fedorov exchange algorithm and coordinate‑exchange methods construct exact D‑optimal designs for a given run size. In nonlinear models, locally optimal designs depend on unknown parameters; Bayesian optimal designs integrate over a prior distribution using Monte Carlo approximations.
Response surface methodology (RSM) extends optimal design to sequential exploration of a region of interest. Central composite designs (CCD) combine a 2k factorial (or fractional factorial) with axial points at distance α and centre points; rotatability requires α = (2k)1/4. Box–Behnken designs avoid extreme corner points, requiring three levels per factor and providing near‑rotatability with fewer runs. For constrained mixture experiments, simplex‑centroid and simplex‑lattice designs respect the summing‑to‑one constraint. Optimal RSM designs (I‑optimal, minimising average prediction variance over the region) are generated by algorithms such as the modified Fedorov or particle swarm optimisation. Model‑robust designs protect against misspecification by optimising a weighted criterion across a set of candidate models. In practice, candidates should be able to write the information matrix for a quadratic model, derive the rotatability condition for CCD, and explain how a D‑optimal design for a given candidate set is verified using the equivalence theorem.
Dynamic Panel Data Models: Arellano–Bond GMM Estimation
Dynamic panels with a lagged dependent variable yit‑1 suffer from Nickell bias when the time dimension T is small relative to N. The Arellano–Bond first‑difference GMM estimator eliminates the fixed effect by differencing: Δyit = αΔyit‑1 + β′Δxit + Δεit. Lagged levels yit‑2, yit‑3, … serve as instruments for Δyit‑1 under the assumption that εit is serially uncorrelated. The moment conditions are E[yis·Δεit] = 0 for s ≤ t‑2. The optimal weighting matrix uses a consistent estimate of the variance of the moment conditions, typically the two‑step estimator with Windmeijer’s finite‑sample correction for the standard errors. System GMM adds the level equation with lagged differences as instruments, exploiting the additional moment condition E[Δyit‑1·(ηi+εit)] = 0 under mean‑stationarity. Validity is tested via the Arellano–Bond test for second‑order serial correlation in Δεit (should be absent) and the Sargan/Hansen test of over‑identifying restrictions. Weak instruments arise when the autoregressive parameter α is close to unity; in that case, the Blundell–Bond system GMM gains efficiency. Candidates must be able to construct the instrument matrix Z for a given T, write the GMM criterion Q(θ) = g′Wg, and explain why the first‑difference estimator loses one time period per lag used as instrument.
Structural Equation Modeling and Confirmatory Factor Analysis
Structural equation modeling (SEM) integrates a measurement model linking latent variables to observed indicators with a structural model specifying causal relations among latents. The measurement model for a factor ξ with indicators y is y = Λξ + ε, where Λ is the loading matrix and ε ∼ N(0, Θ) with Θ diagonal. The structural model is η = Bη + Γξ + ζ, where η are endogenous latents, ξ exogenous latents, B and Γ coefficient matrices, and ζ ∼ N(0, Ψ). Identification requires the order condition (number of knowns ≥ number of unknowns) and the rank condition; the t‑rule (t = p(p+1)/2) gives a necessary count for the covariance matrix of p observed variables. Common estimators: maximum likelihood (ML) under multivariate normality, generalized least squares (GLS), weighted least squares (WLS) and its robust variants (WLSMV, DWLS) for ordinal data. Fit indices include the comparative fit index (CFI) > 0.95, Tucker‑Lewis index (TLI) > 0.95, root mean square error of approximation (RMSEA) < 0.06, and standardized root mean square residual (SRMR) < 0.08. Modification indices (Lagrange multiplier tests) suggest parameter additions; expected parameter change (EPC) quantifies the effect. Multi‑group SEM tests measurement invariance (configural, metric, scalar) by constraining loadings and intercepts across groups. MIMIC models incorporate covariates directly on indicators to detect differential item functioning. Candidates should be able to write the implied covariance matrix Σ(θ) = Λ(I−B)−1(ΓΦΓ′+Ψ)(I−B′)−1Λ′ + Θ, derive the degrees of freedom, and interpret a significant χ2 difference test for nested models.
Small Area Estimation: Fay–Herriot Model and Empirical Best Linear Unbiased Prediction
Small area estimation borrows strength across areas to produce reliable estimates for domains with small sample sizes. The Fay–Herriot area‑level model specifies the direct survey estimate yi = xi′β + vi + ei, where vi ∼ N(0, σv2) is the random area effect and ei ∼ N(0, σe,i2) is the known sampling error variance. The empirical best linear unbiased predictor (EBLUP) of the small area mean θi = xi′β + vi is θ̂i = γ̂iyi + (1−γ̂i)xi′β̂, with shrinkage factor γ̂i = σ̂v2 / (σ̂v2 + σe,i2). Variance components σv2 are estimated by restricted maximum likelihood (REML), method of moments (Prasad–Rao), or Henderson’s method III. Mean squared error (MSE) of the EBLUP is approximated by the Prasad–Rao second‑order unbiased estimator: MSE(θ̂i) ≈ g1i(σ̂v) + g2i(σ̂v) + 2g3i(σ̂v), where g1i captures sampling error, g2i estimation of β, and g3i estimation of σv2. Bootstrap MSE (parametric or wild) resamples the random effects and sampling errors to capture higher‑order terms. Extensions include spatial Fay–Herriot models with a conditional autoregressive (CAR) prior on vi, unit‑level nested error regression models (Battese–Harter–Fuller), and benchmarking to known population totals via constrained EBLUP. In the Indian context, NSSO uses Fay–Herriot for district‑level poverty and employment estimates, with MSE reported via parametric bootstrap. Candidates must derive the EBLUP weight γi, write the REML log‑likelihood for the area‑level model, and explain why the Prasad–Rao estimator includes a term for the variability of σ̂v2.
Multiple Imputation and Missing Data Mechanisms in Complex Surveys
Missing data in surveys are classified as missing completely at random (MCAR), missing at random (MAR), or missing not at random (MNAR). Under MAR, the probability of missingness depends only on observed data, justifying multiple imputation (MI). Rubin’s rules combine m imputed datasets: the overall estimate θ̄ = (1/m) Σ θ̂(k), total variance T = W + (1+1/m)B, where W = (1/m) Σ U(k) is the within‑imputation variance and B = (1/(m−1)) Σ (θ̂(k)−θ̄)2 is the between‑imputation variance. For complex surveys, the imputation model must incorporate survey weights, stratification, and clustering. Fully conditional specification (FCS), also known as chained equations, iteratively imputes each incomplete variable conditional on all others using appropriate models (logistic for binary, predictive mean matching for continuous). Survey‑weighted imputation uses the weights as frequency weights in the imputation regressions or adopts a two‑phase approach: first impute ignoring design, then adjust via propensity score weighting. Variance estimation combines the MI variance with the design‑based variance using the bootstrap or jackknife replicate weights within each imputed dataset, then applying Rubin’s rules to the replicate estimates. Diagnostics include comparing observed and imputed distributions, checking convergence of FCS chains, and assessing the fraction of missing information (FMI = (1+1/m)B/T). When MNAR is suspected, sensitivity analysis via pattern‑mixture or selection models is required. Candidates should be able to write the FCS algorithm steps, derive the total variance formula for a scalar estimand, and explain why ignoring the survey design in imputation leads to biased variance estimates.
- Confusing the order condition for identification in SEM with the rank condition; the order condition is necessary but not sufficient.
- Using the first‑difference GMM estimator without testing for second‑order serial correlation, which invalidates the instruments.
- Assuming the Fay–Herriot model’s sampling variances σe,i2 are known without error; in practice they are estimated and the MSE must reflect that uncertainty.
- Applying standard multiple imputation ignoring survey weights, stratification, or clustering, which underestimates variance and can bias point estimates.
- Treating a D‑optimal design for a linear model as optimal for a nonlinear model without re‑evaluating the information matrix at the true parameter values.
- Overlooking the need to re‑calibrate weights inside each bootstrap replicate when using replication variance estimation for calibrated survey estimators.
- Misinterpreting the Windmeijer correction as a substitute for the two‑step optimal weighting matrix rather than a finite‑sample adjustment to its standard errors.