Statistical Methodologies

3.1 Introduction to Census Statistical Methods

While censuses aim for complete enumeration, statistical methods play essential roles throughout census operations. Sampling reduces costs in data collection and content testing. Weighting adjusts for differential response rates across subpopulations. Imputation fills gaps from item non-response. Error estimation quantifies uncertainty. These methods must balance statistical rigor, computational feasibility, and interpretability for diverse users.

This chapter presents the WIA-SOC-016 standard's framework for census statistical methodologies, covering design considerations, implementation procedures, and quality assessment. Whether conducting a traditional complete enumeration, a register-based census, or hybrid approaches, these methods ensure census data meets high standards of accuracy and reliability.

3.2 Sampling Design for Census Operations

Sampling in census contexts serves multiple purposes: pre-census content testing, quality assurance through post-enumeration surveys, coverage improvement programs, and in some cases, detailed content collection for a sample while conducting basic enumeration for the full population.

3.2.1 Probability Sampling Fundamentals

Probability sampling ensures every member of the target population has a known, non-zero probability of selection, enabling design-based statistical inference. Key probability sampling designs include:

• Simple Random Sampling: Every unit has equal selection probability. Straightforward but may miss rare populations and doesn't leverage auxiliary information.
• Stratified Sampling: Divide population into homogeneous strata, sample independently within each. Ensures representation of all groups and typically improves precision.
• Cluster Sampling: Sample groups of units (e.g., census blocks) rather than individuals. Reduces costs but typically increases variance unless clusters are internally heterogeneous.
• Systematic Sampling: Select every kth unit from an ordered list. Simple to implement and provides implicit stratification if list is ordered by relevant characteristics.
• Multi-stage Sampling: Hierarchical selection (e.g., first sample enumeration areas, then households within areas). Balances cost efficiency with adequate coverage.
• Probability Proportional to Size (PPS): Larger units have higher selection probabilities. Often used in multi-stage designs where first-stage units vary greatly in size.

3.2.2 Sample Size Determination

Sample size calculations balance precision requirements, budget constraints, and design effects from complex sampling. The basic formula for simple random sampling estimates the required sample size n for estimating a proportion p with margin of error E and confidence level Z:

For census quality assurance surveys, sample sizes typically range from 0.1% to 5% of the population, depending on desired precision for small-area estimates, available budget, and design complexity. Larger samples support more detailed geographic and demographic breakdowns.

3.3 Weighting and Calibration

Even in complete enumerations, differential response rates across demographic groups, geographic areas, and housing types can bias census results. Weighting adjusts for these differentials, ensuring census statistics accurately represent the entire population.

3.3.1 Base Weights

For sample surveys, base weights equal the inverse of selection probability. For a person with selection probability 0.05, the base weight is 20, meaning they "represent" 20 people in the population. Base weights may vary across strata if stratified sampling with differential sampling rates is employed.

3.3.2 Non-response Adjustment

Non-response adjustment increases weights for respondents to compensate for non-respondents with similar characteristics. Common approaches include:

• Weighting Class Adjustment: Group units into cells based on auxiliary variables (age, sex, geography), increase weights proportionally to non-response rate within each cell
• Propensity Score Weighting: Model response probability using logistic regression, adjust weights inversely proportional to estimated response probability
• Calibration Weighting: Adjust weights to match known population totals from administrative sources or previous census

3.3.3 Post-Stratification and Calibration

Post-stratification adjusts weights to match known population totals for demographic and geographic groups. If the census counts 50 million males but administrative records indicate 52 million, male weights are increased by factor 52/50 = 1.04. Iterative proportional fitting (raking) extends post-stratification to multiple dimensions simultaneously, adjusting weights to match known marginal totals while preserving multivariate distributions.

3.3.4 Weight Trimming and Smoothing

Extreme weights can arise from very low selection probabilities or response propensities, increasing variance and creating influential cases that dominate estimates. Weight trimming caps extreme weights at reasonable thresholds (e.g., 95th percentile), redistributing excess weight to other cases. Smoothing reduces weight variability within cells through regression models or robust methods.

3.4 Imputation for Item Non-response

Even when households participate in the census, they may skip specific questions (item non-response). Imputation fills these missing values to create complete datasets for analysis. The WIA-SOC-016 standard recommends explicit imputation with documentation over implicit treatment of missingness.

3.4.1 Hot-Deck Imputation

Hot-deck imputation replaces missing values with observed values from "similar" respondents (donors). Sequential hot-deck processes records in random order, using the most recent donor within each imputation class. Nearest-neighbor hot-deck selects donors that are most similar to recipients based on distance metrics incorporating multiple auxiliary variables. Predictive mean matching combines regression modeling with hot-deck donor selection to preserve distributional properties.

3.4.2 Model-Based Imputation

Regression imputation predicts missing values using linear or logistic models estimated from complete cases. Deterministic regression imputation uses predicted values directly but underestimates variance. Stochastic regression imputation adds random residuals to predictions, better preserving distributions. Multiple imputation creates several plausible values for each missing item, propagating imputation uncertainty through analysis.

3.4.3 Edit-Imputation Systems

Census data processing typically combines editing (identifying inconsistent or implausible values) with imputation (correcting errors and filling gaps). The Fellegi-Holt methodology provides a theoretical framework for minimum-change edit-imputation that corrects identified errors while modifying as few values as possible. Modern implementations use mathematical optimization to find minimum-change solutions respecting complex edit rules.

3.5 Coverage Evaluation and Adjustment

Coverage errors occur when census enumeration misses people (undercount) or counts them multiple times (overcount). Net undercount is typically 1-5% in developed countries, higher in developing contexts, and varies substantially across demographic groups.

3.5.1 Dual System Estimation

Dual system estimation (DSE), also called capture-recapture, estimates population size by comparing census enumeration with an independent post-enumeration survey (PES). The estimator assumes both sources are independent: P̂ = (C × S) / M, where C = census count, S = survey count, M = matched count appearing in both sources. The independence assumption is strong and often violated, requiring adjustment methods.

3.5.2 Demographic Analysis

Demographic analysis estimates population by age, sex, and race using historical birth, death, and migration data from administrative sources. Comparing demographic analysis estimates with census counts indicates coverage quality. Unlike DSE, demographic analysis doesn't require assumptions about source independence but relies on complete, accurate vital statistics and migration data.

3.5.3 Coverage Improvement Programs

Proactive coverage improvement identifies and enumerates missed housing units and persons during census operations. Address list quality assurance, non-response follow-up, and targeted outreach to hard-to-count populations all improve coverage. Post-census statistical adjustment for residual coverage errors remains controversial, with debates about whether statistical uncertainty from adjustment exceeds coverage error from non-adjustment.

3.6 Variance Estimation for Complex Surveys

Standard variance formulas assuming simple random sampling substantially underestimate uncertainty from complex census designs. Appropriate variance estimation accounts for stratification, clustering, weighting, and non-response adjustment.

3.6.1 Taylor Series Linearization

Taylor series linearization approximates complex statistics (ratios, regression coefficients) as linear functions of survey variables, enabling variance estimation using stratified cluster sampling formulas. Widely used for smooth statistics but not applicable to non-smooth statistics like medians or domain boundaries.

3.6.2 Replication Methods

Replication methods (jackknife, balanced repeated replication, bootstrap) create multiple partial sample estimates, measuring variance from their distribution. Applicable to any statistic and relatively easy to implement once replicate weights are created. Computationally intensive for large datasets but increasingly feasible with modern computing power.

3.6.3 Design Effect and Effective Sample Size

The design effect (DEFF) quantifies efficiency loss from complex sampling: DEFF = Var_complex / Var_SRS. Effective sample size equals actual sample size divided by DEFF. Reporting effective sample size helps users understand the actual precision of estimates. Design effects vary across statistics, population subgroups, and geographic areas, requiring careful analysis and documentation.

3.7 Quality Indicators and Data Quality Framework

The WIA-SOC-016 standard adopts a comprehensive data quality framework encompassing multiple dimensions. Census agencies should assess, document, and communicate quality across all dimensions.

3.7.1 Quality Dimensions

Dimension	Definition	Key Indicators
Relevance	Degree to which data meets user needs	User satisfaction, consultation coverage, usage statistics
Accuracy	Closeness of estimates to true values	Coverage error, content error, processing error
Timeliness	Delay between reference period and data availability	Release schedule adherence, time lag
Accessibility	Ease of obtaining and using data	Availability of formats, API uptime, documentation quality
Interpretability	Availability of metadata and documentation	Metadata completeness, user support quality
Coherence	Consistency across datasets and over time	Consistency with other sources, time series breaks

3.7.2 Total Survey Error Framework

The total survey error framework decomposes census error into representation errors (coverage, sampling, non-response, adjustment) and measurement errors (validity, measurement, processing). Minimizing total error requires balancing trade-offs; for example, intensive follow-up reduces non-response but may increase measurement error if reluctant respondents provide lower-quality data under pressure.

3.7.3 Quality Reporting

Comprehensive quality reporting includes:

• Response rates by mode, geography, and demographic groups
• Coverage error estimates from PES and demographic analysis
• Imputation rates for each variable
• Variance estimates and confidence intervals for key statistics
• Comparisons with previous census and alternative data sources
• Documentation of deviations from planned methodology
• Known limitations and appropriate use guidance

3.8 Small Area Estimation

Census data users increasingly need estimates for small geographic areas and population subgroups. Direct estimation (using only data from the area) provides insufficient precision when samples are small. Small area estimation (SAE) "borrows strength" from related areas or auxiliary data to improve precision.

3.8.1 Model-Based Small Area Estimation

SAE models express area-level estimates as functions of covariates available from census or administrative sources, plus area-specific random effects. The Fay-Herriot model for area-level data and the nested error model for unit-level data are widely used. Empirical best linear unbiased prediction (EBLUP) provides optimal estimates under model assumptions. Model diagnostics and validation using out-of-sample areas assess model adequacy.

3.8.2 Composite Estimation

Composite estimators combine direct estimates with synthetic estimates (predicted from larger areas) using weights inversely proportional to their variances. Areas with larger samples receive more weight from direct estimation; areas with small samples rely more on synthetic estimation. This balances bias (from potentially incorrect synthetic models) and variance (from small sample direct estimates).

3.9 Seasonal Adjustment and Time Series Analysis

While full censuses are infrequent, census data form baselines for intercensal population estimates and projections. Understanding temporal patterns supports accurate estimates between censuses and planning for future census operations.

3.9.1 Population Projections

Cohort-component models project populations by age and sex based on fertility, mortality, and migration assumptions. Stochastic population projections incorporate parameter uncertainty to quantify forecast intervals. Multi-regional models extend cohort-component approaches to geographic flows. Model validation compares past projections with subsequent census results to assess forecast accuracy.

3.9.2 Intercensal Estimation

Intercensal estimates for years between censuses use administrative data (births, deaths, migration) to update census baselines. Postcensal estimates extrapolate forward from the most recent census. Intercensal estimates are revised after new census data become available, incorporating information from both surrounding censuses. Housing unit method and ratio-correlation methods leverage administrative data on housing construction and population-housing relationships.

The next chapter examines data collection methods, from traditional paper questionnaires to modern digital platforms, building on statistical foundations to ensure high-quality census data collection.