{
  "id": "survey-sampling",
  "title": "Survey Sampling",
  "lang": "en",
  "source_lang": "en",
  "summary": "Survey sampling selects and analyzes subsets of populations to estimate population characteristics while accounting for selection mechanisms and uncertainty.",
  "category": "mathematics",
  "subcategory": "probability-statistics",
  "type": "concept",
  "aliases": [
    "Sample survey methodology",
    "Survey sampling theory"
  ],
  "generated_at": "2026-11-05",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/survey-sampling/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/survey-sampling/",
      "json": "https://aiwiki.page/zh/survey-sampling.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "Survey sampling is the branch of statistics concerned with selecting a subset of a population and using observations from that subset to estimate characteristics of the population. Its scope includes sample selection, the construction of estimators, the measurement of uncertainty, and adjustments for incomplete coverage and nonresponse. Unlike a census, which seeks information about every population unit, a sample survey obtains information from only some units, potentially reducing cost, collection time, and respondent burden.\n\nThe central problem is not merely how many observations to collect, but how the observed units relate to the population about which conclusions are required. A carefully selected sample of a few thousand households can support national estimates, whereas a much larger collection of self-selected responses may have substantial selection bias. Survey sampling supplies the mathematical and operational framework connecting selection, measurement, and inference. It underlies official statistics, establishment surveys, social research, environmental monitoring, and many other forms of population investigation."
    },
    {
      "heading": "Foundations and historical development",
      "text": "A finite population can be represented as \\(U=\\{1,\\ldots,N\\}\\), with a value \\(y_i\\) attached to each unit. Common targets are the population total \\(T_y=\\sum_{i\\in U}y_i\\), the population mean \\(\\bar Y=T_y/N\\), and proportions defined by coding a characteristic as zero or one. Other targets include medians, ratios, distributional measures, and changes over time. The particular quantity being estimated is called the estimand. Its definition must specify the population, reference period, unit, and variable: annual earnings of employees, for example, differ from annual income of households.\n\nIn design-based inference, the population values are treated as fixed. Randomness comes from the procedure that selects a sample \\(s\\) from \\(U\\). A sampling design assigns probabilities \\(p(s)\\) to possible samples, and an estimator is evaluated over repeated selections under that design. This framework permits inference without assuming that the population values follow a particular probability distribution. In model-based inference, by contrast, a statistical model describes how population values arise. Model-assisted methods combine probability sampling with models used to improve estimation.\n\nThe modern theory developed from debates about whether selected observations could represent whole populations. In 1895, Anders Nicolai Kiaer presented his representative method at a meeting of the International Statistical Institute. Arthur Lyon Bowley subsequently helped establish the role of random selection and sampling error in social investigation. A decisive theoretical contribution was Jerzy Neyman’s 1934 paper contrasting stratified random sampling with purposive selection. Neyman developed results on efficient allocation and interval estimation that helped give probability sampling a rigorous foundation.\n\nSubsequent advances addressed increasingly complex designs. Morris Hansen and William Hurwitz developed influential methods for unequal-probability sampling, while Daniel Horvitz and Donovan Thompson introduced their general inverse-inclusion-probability estimator in 1952. Practical requirements from national statistical agencies encouraged the development of multistage household designs, rotating samples, replicated variance estimation, and survey weighting. These developments made clear that representativeness is not a single visible property of a dataset: reliable inference depends on the selection mechanism, the estimator, the quality of measurement, and the treatment of missing observations."
    },
    {
      "heading": "Populations, sampling frames, and operational units",
      "text": "The target population is the collection of units about which information is required. The survey population is the collection actually covered by the survey’s operational procedures. These may differ. A household survey intended to describe all residents may exclude institutional populations, people without conventional addresses, or residents of inaccessible areas. Such exclusions change the scope of inference unless additional procedures cover the omitted groups. A precise population definition therefore precedes sample selection rather than being inferred from the respondents afterward.\n\nA sampling frame is the list or other operational representation from which units are selected. Examples include address registers, business registers, lists of schools, maps divided into geographical areas, and telephone-number ranges. Frames can contain omissions, duplicates, ineligible records, and outdated classifications. A business register may retain closed enterprises while missing recently established ones. An address list may identify dwellings without identifying how many households occupy each dwelling. Frame quality therefore concerns both the presence of eligible units and the accuracy of information needed for selection.\n\nThe sampling unit, reporting unit, and analysis unit need not coincide. A geographical area may be selected first, an address second, and a person third; one household member may then report information about several people. Inclusion probabilities must follow the entire selection path. If an individual can be reached through multiple frame records, the procedure must either remove duplication or account for the multiple opportunities for selection. Treating duplicated records as distinct people generally gives some individuals excessive representation.\n\nFrame construction often involves trade-offs between coverage and cost. An area frame can cover dwellings missing from administrative lists, but field listing is expensive. A list frame may be cheap to use and contain valuable auxiliary information, but require frequent updating. Dual-frame designs combine two frames, such as landline and mobile telephone frames, while explicitly handling overlap. Their estimators must distinguish units covered by only one frame from those covered by both; simply pooling independently weighted samples can count the overlapping population twice.\n\nCoverage also has a temporal dimension. A frame assembled at one date may not match a population defined months later. Births, deaths, migration, business openings, and business closures create discrepancies. Repeated surveys therefore need frame maintenance, rules for newly eligible units, and consistent treatment of changes in eligibility. These operational decisions determine whether an estimator refers to the intended population, even when its mathematical properties are otherwise sound."
    },
    {
      "heading": "Simple random, systematic, and stratified designs",
      "text": "In simple random sampling without replacement, every subset of \\(n\\) units from a population of \\(N\\) units has the same selection probability. Each unit consequently has inclusion probability \\(n/N\\). The unweighted sample mean estimates the population mean without design bias. Writing the finite-population variance as\n\n\\[S_y^2=\\frac{1}{N-1}\\sum_{i=1}^{N}(y_i-\\bar Y)^2,\\]\n\nthe sampling variance is\n\n\\[\\operatorname{Var}_p(\\bar y)= \\left(1-\\frac{n}{N}\\right)\\frac{S_y^2}{n}.\\]\n\nThe factor \\(1-n/N\\) is the finite-population correction. It reflects the fact that observing a large fraction of a finite population leaves less uncertainty than independent sampling with replacement would suggest.\n\nSystematic sampling selects units at regular intervals after a random start. If \\(N=nk\\), an integer start is chosen uniformly from \\(1,\\ldots,k\\), and every \\(k\\)th record is selected. This is operationally convenient and spreads the sample across the list. Its performance depends on list order. Sorting by geography or another useful variable can improve precision through implicit stratification, but periodic patterns aligned with the sampling interval can produce poor samples. With only one systematic random start, some pairs of units cannot occur together, complicating design-unbiased variance estimation.\n\nStratified sampling divides the population into nonoverlapping groups, or strata, and samples separately within each. For stratum \\(h\\), let \\(N_h\\) and \\(n_h\\) be its population and sample sizes. Under independent simple random sampling within strata, the estimated total and its variance are\n\n\\[\\widehat T_{\\mathrm{st}}=\\sum_h N_h\\bar y_h, \\qquad \\operatorname{Var}_p(\\widehat T_{\\mathrm{st}}) =\\sum_h N_h^2 \\left(1-\\frac{n_h}{N_h}\\right)\\frac{S_h^2}{n_h}.\\]\n\nStratification is particularly effective when units within strata resemble one another for the outcome being measured.\n\nAllocation determines how the total sample is distributed across strata. Proportional allocation sets \\(n_h\\) proportional to \\(N_h\\), producing equal sampling fractions. With equal per-unit costs and a fixed total sample size, Neyman allocation sets \\(n_h\\) proportional to \\(N_hS_h\\). With differing per-unit costs \\(c_h\\) and a fixed variable-cost budget, the corresponding allocation is proportional to \\(N_hS_h/\\sqrt{c_h}\\). These formulas require advance information about variability and may be modified by minimum sample sizes or multiple precision targets.\n\nDisproportionate allocation can deliberately oversample small regions or uncommon groups. For example, selecting the same number of businesses from a small industry and a large industry supports separate industry estimates, but requires weights when producing an economy-wide estimate. Stratification therefore serves two related purposes: increasing precision for aggregate estimates and ensuring adequate information for important subpopulations."
    },
    {
      "heading": "Cluster, multistage, and unequal-probability sampling",
      "text": "Cluster sampling selects groups of units rather than selecting every ultimate unit independently. A school survey may select schools and then students; a household survey may select geographical areas, dwellings, and residents. Such designs reduce travel and listing costs and make sampling possible when no complete list of ultimate units exists. The first selected groups are primary sampling units, or PSUs. Later stages can use different selection methods and different sampling fractions.\n\nClustering usually reduces precision at a fixed number of observed individuals because people within the same cluster often resemble one another. Students in one school share institutional conditions, while neighboring households share local environments. For equal cluster sizes \\(m\\), a common approximation to the variance inflation relative to simple random sampling is\n\n\\[\\operatorname{DEFF}\\approx 1+(m-1)\\rho,\\]\n\nwhere \\(\\rho\\) is the intracluster correlation. With \\(m=20\\) and \\(\\rho=0.05\\), this approximation gives a design effect of \\(1.95\\). The expression is not universal: unequal cluster sizes, stratification, weighting, and the particular estimator can substantially alter the result.\n\nProbability-proportional-to-size sampling uses an auxiliary size measure to give larger clusters greater selection chances. Measures might be the number of dwellings in an area or enrollment in a school. With replacement, each draw can select a cluster with probability proportional to its size. Without replacement, specialized algorithms can achieve specified first-order inclusion probabilities, subject to feasibility constraints. Very large units may require certainty selection, with inclusion probability one, rather than an impossible probability above one.\n\nMultistage probabilities are obtained from conditional selection probabilities along the selection path. If area \\(a\\) is selected with probability \\(\\pi_a\\), and household \\(i\\) within that area is selected with conditional probability \\(\\pi_{i\\mid a}\\), then \\(\\pi_i=\\pi_a\\pi_{i\\mid a}\\). Further stages introduce additional factors. In a suitably constructed two-stage design, selecting areas with probabilities proportional to their household counts and then selecting a fixed number of households per selected area can give approximately equal household inclusion probabilities.\n\nCluster sampling and stratification should not be confused. Samples are usually taken from every stratum, but only from selected clusters. Strata are often designed to be internally homogeneous to remove between-stratum variation from the sampling error; clusters are operational groupings whose internal similarity can increase sampling error. Large national surveys commonly combine both approaches, stratifying primary sampling units before selecting them and then subsampling within selected units."
    },
    {
      "heading": "Inclusion probabilities and population estimators",
      "text": "For a population unit \\(i\\), the first-order inclusion probability is \\(\\pi_i=P_p(i\\in s)\\). For units \\(i\\) and \\(j\\), the second-order inclusion probability is \\(\\pi_{ij}=P_p(i,j\\in s)\\), with \\(\\pi_{ii}=\\pi_i\\). These are probabilities of appearing in the final sample, not necessarily probabilities on a single draw. Their distinction matters in sampling with replacement, where the same unit can be selected repeatedly. The base sampling weight \\(d_i=1/\\pi_i\\) expresses how many population units a sampled unit represents in an inverse-probability estimator.\n\nThe Horvitz–Thompson estimator of a total is\n\n\\[\\widehat T_{\\mathrm{HT}} =\\sum_{i\\in s}\\frac{y_i}{\\pi_i}.\\]\n\nProvided every population unit has positive inclusion probability, it is design-unbiased:\n\n\\[E_p(\\widehat T_{\\mathrm{HT}})=T_y.\\]\n\nThis follows because the inclusion indicator for unit \\(i\\) has expected value \\(\\pi_i\\). Equal-probability sampling is a special case, not a prerequisite. The estimator can therefore accommodate stratification, unequal probabilities, and multistage designs.\n\nIts design variance is\n\n\\[\\operatorname{Var}_p(\\widehat T_{\\mathrm{HT}}) = \\sum_{i\\in U}\\sum_{j\\in U} (\\pi_{ij}-\\pi_i\\pi_j) \\frac{y_i}{\\pi_i}\\frac{y_j}{\\pi_j}.\\]\n\nWhen relevant joint inclusion probabilities are positive and known, a corresponding unbiased variance estimator replaces the population double sum with a sample double sum and divides each term by \\(\\pi_{ij}\\). If some pairs can never appear together, this standard construction is unavailable for those terms. Thus, positive individual inclusion probabilities suffice for unbiased total estimation but do not automatically guarantee an unbiased variance estimator of the usual form.\n\nWhen \\(N\\) is known, dividing \\(\\widehat T_{\\mathrm{HT}}\\) by \\(N\\) gives a design-unbiased estimator of the mean. Another common estimator is the Hájek weighted mean,\n\n\\[\\widehat{\\bar Y}_{\\mathrm H} = \\frac{\\sum_{i\\in s}d_i y_i} {\\sum_{i\\in s}d_i}.\\]\n\nThis ratio is generally not exactly unbiased, but it can be more stable because it normalizes the weights. Its denominator estimates population size. Under simple random sampling with fixed sample size, it reduces to the ordinary sample mean.\n\nThese estimators assume that the required values are observed for selected eligible units. Once nonresponse adjustments or calibration factors modify the base weights, exact Horvitz–Thompson unbiasedness does not follow merely from retaining an inverse-weighted form. Properties must instead be assessed under the actual selection and response processes, or under explicit assumptions about the adjustments."
    },
    {
      "heading": "Precision, variance estimation, and sample size",
      "text": "A survey estimate’s standard error is the square root of its estimated sampling variance. A confidence interval commonly takes the form \\(\\widehat\\theta\\pm c\\,\\widehat{\\operatorname{SE}}(\\widehat\\theta)\\), where the critical value \\(c\\) reflects the intended coverage and the approximation used. The familiar value \\(1.96\\) relies on an adequate normal approximation. In multistage surveys with few sampled PSUs, a Student’s t distribution with design-based degrees of freedom may be more appropriate. Large numbers of respondents do not necessarily compensate for very few independent primary sampling units.\n\nNonlinear statistics require additional methods. Taylor linearization approximates a statistic by a linear function of estimated totals. For a ratio \\(\\widehat R=\\widehat T_y/\\widehat T_x\\), uncertainty is approximately governed by residual-like quantities \\(y_i-Rx_i\\), scaled by the denominator total. The variance procedure then applies the actual stratification and clustering to these linearized quantities. Quantiles and other less smooth statistics may require specialized linearization or replication methods.\n\nReplication variance estimation repeatedly perturbs the sample’s contributions and recalculates the statistic. Methods include delete-group jackknife, balanced repeated replication, and survey adaptations of the bootstrap. Their replicate weights must reflect the sample design. Resampling individuals as though they were independent generally fails for clustered data. Replication is especially useful when an estimator includes several weighting stages, since those stages can be repeated or suitably represented within the replicates.\n\nInitial sample-size calculations often begin with a simple random sample approximation. For a proportion \\(P\\), margin of error \\(e\\), and normal critical value \\(z\\), an infinite-population approximation is \\(n_0=z^2P(1-P)/e^2\\). The corresponding finite-population adjustment is\n\n\\[n=\\frac{Nn_0}{N-1+n_0}.\\]\n\nUsing \\(P=0.5\\), \\(z=1.96\\), and \\(e=0.03\\) gives approximately 1,068 completed observations before accounting for finite-population size, complex design, or nonresponse. The value \\(0.5\\) maximizes the Bernoulli variance and is therefore conservative for this calculation when no better information is available.\n\nA design effect compares the variance under an actual design and estimator with that under a specified simple-random-sampling reference. The heuristic effective sample size \\(n/\\operatorname{DEFF}\\) expresses precision in simple-random-sample terms, but is statistic-specific. Planning often inflates the required number of completed interviews for design effects and then divides by an anticipated response rate to obtain an initial sample size. This addresses expected precision and workload; it does not imply that collecting more observations removes nonresponse bias."
    },
    {
      "heading": "Auxiliary information, calibration, and model-assisted estimation",
      "text": "Auxiliary information consists of variables known beyond the responding sample and related to survey outcomes or participation. Examples include census population counts by age and region, payroll totals, administrative business classifications, and frame-based dwelling counts. Such information can guide sample allocation, improve precision, and adjust discrepancies between the respondent sample and known population characteristics. Its usefulness depends on accuracy, conceptual compatibility, and association with the quantities being estimated.\n\nCalibration weighting chooses final weights \\(w_i\\) close to initial weights \\(d_i\\), subject to constraints such as\n\n\\[\\sum_{i\\in s}w_i x_i=X,\\]\n\nwhere \\(x_i\\) is a vector of auxiliary variables and \\(X\\) is its known population total. Closeness is defined through a distance function or optimization criterion. Different choices produce different calibration procedures, and constraints may be added to limit extreme or negative weights. Calibrating an intercept to a known population count makes the weights sum to that count.\n\nPoststratification adjusts weights within cells whose population counts are known. If cells are defined jointly by age group, sex, and region, their weighted sample totals can be made equal to their population totals. Raking instead matches several marginal distributions iteratively, without necessarily matching every joint cell. This is useful when joint population counts are unavailable or joint cells would be too sparse. Neither procedure guarantees correction for differences that remain within adjustment groups, such as an association between survey participation and income among people with the same recorded demographic characteristics.\n\nA generalized regression estimator combines an estimated relationship with a correction based on known auxiliary totals:\n\n\\[\\widehat T_{\\mathrm{GREG}} = \\widehat T_{\\mathrm{HT},y} + \\left(X-\\widehat T_{\\mathrm{HT},x}\\right)^\\top\\widehat\\beta.\\]\n\nHere \\(\\widehat\\beta\\) is commonly obtained from a survey-weighted linear regression. When the auxiliary variables predict the outcome well, the correction can reduce sampling variance. Model-assisted justification seeks good design properties even when the working regression model is not exactly correct; it does not require treating the model as a literal description of every population value.\n\nWeights can become highly variable when selection probabilities are unequal, response propensities are small, or calibration constraints are demanding. Weight trimming limits extreme values but generally introduces a bias–variance trade-off and may require recalibration. The approximation\n\n\\[\\operatorname{DEFF}_{w}\\approx 1+\\operatorname{CV}(w)^2\\]\n\nis sometimes used to describe the precision loss from unequal weighting. It relies on simplifying assumptions and should not replace a design-based variance estimate, especially when weights are associated with outcomes or incorporate efficiency gains from stratification and calibration."
    },
    {
      "heading": "Nonresponse and the total survey error framework",
      "text": "Sampling error is only one component of uncertainty. Total survey error encompasses errors from coverage, nonresponse, measurement, processing, and estimation as well as sample selection. For an estimator \\(\\widehat\\theta\\), mean squared error decomposes as\n\n\\[\\operatorname{MSE}(\\widehat\\theta) = \\operatorname{Var}(\\widehat\\theta) + \\operatorname{Bias}(\\widehat\\theta)^2,\\]\n\nunder the probability framework used to define those quantities. A larger sample reduces some variance components but does not automatically reduce systematic error. A census eliminates sampling error from selecting only part of the population, yet can still suffer from undercoverage and inaccurate responses.\n\nUnit nonresponse occurs when an eligible selected unit supplies no usable survey data; item nonresponse occurs when particular questions lack answers. Their effects depend on the relationship between missingness and the study variables, not simply on the percentage missing. A low response rate can coexist with limited bias for one variable, while a relatively high response rate can conceal substantial bias for another. Response rates remain important operational indicators, but are not direct measurements of nonresponse bias.\n\nWeighting-class and response-propensity adjustments attempt to compensate for unit nonresponse using information available for respondents and nonrespondents. A common assumption is that response is unrelated to the outcome after conditioning on the adjustment variables, with a positive response probability throughout the relevant population. If respondents and nonrespondents differ systematically even after conditioning, this adjustment can remain biased. Extremely low estimated response probabilities also create unstable weights.\n\nItem nonresponse is often treated through imputation, including donor methods, regression prediction, and multiple imputation. Single deterministic imputation can understate uncertainty if filled-in values are subsequently treated as observed. Multiple imputation represents missing-data uncertainty through several completed datasets, but its validity depends on the imputation model and its compatibility with the sampling design and analysis. Neither weighting nor imputation makes the assumptions about unobserved values disappear.\n\nMeasurement error includes misunderstanding, recall failure, interviewer effects, instrument effects, and inaccurate proxy reporting. Question wording and response categories can alter the variable effectively being measured. Telephone, web, and face-to-face modes may reach different respondents and generate different answers. Processing errors arise during coding, linkage, editing, and data transformation. Consequently, a published sampling margin of error ordinarily describes only one component of total error and should not be interpreted as a bound on every possible discrepancy."
    },
    {
      "heading": "Nonprobability samples and model-dependent inference",
      "text": "In a probability sample, selection is governed by a specified random mechanism with known or calculable inclusion probabilities. Nonprobability samples include convenience samples, volunteer web panels, quota samples without random selection within quotas, and many collections of digital traces. Such datasets can be large and informative, but their observed composition does not by itself identify how they relate to a target population. The principal challenge is selection bias, rather than the arithmetic calculation of a standard error.\n\nQuota sampling illustrates the distinction. A sample may match the population’s age and sex distributions while differing sharply in political engagement, working hours, internet access, or willingness to participate. Matching a few margins does not establish randomness within the resulting groups. Similarly, a very large digital dataset may exclude people who do not use a platform or may overrepresent highly active users. As the sample size grows, random variation can become small while systematic selection differences persist.\n\nAdjustment methods include propensity weighting, calibration, outcome modeling, and combinations of these approaches. A reference probability sample can provide information about the target population’s covariate distribution. Participation models may then estimate relative tendencies to enter the nonprobability sample. For population inference, important conditions include adequate overlap and sufficient observed information to account for selection relevant to the outcome. These conditions are substantive assumptions; they do not follow merely from fitting a flexible machine-learning model.\n\nSome combined estimators have a double-robustness property: under additional regularity and identification conditions, consistency can hold if either a specified outcome model or a specified participation model is correct. This does not mean that inference remains valid under arbitrary selection mechanisms. Unmeasured determinants of both participation and outcomes can invalidate both routes. Sensitivity analysis can examine how conclusions change under alternative assumptions about such unmeasured differences, although it cannot identify them from the observed sample alone.\n\nNonprobability studies can report uncertainty intervals, but the interpretation must follow the inferential framework. A model-based or Bayesian interval reflects assumptions about the outcome and selection processes; it is not automatically a conventional probability-sampling margin of error. Conversely, a probability sample with substantial nonresponse also needs assumptions beyond the original sampling design. The important distinction is therefore the basis for identification and uncertainty, not a claim that one label removes every source of error."
    },
    {
      "heading": "Repeated surveys, domain estimation, and responsible analysis",
      "text": "Repeated surveys estimate both population levels and change. Independent cross-sectional samples draw new units on each occasion, while panel surveys follow the same units over time. Rotating designs replace part of the sample at scheduled intervals. Overlap can improve the precision of change estimates because successive estimates are correlated. For estimates \\(\\widehat\\theta_t\\) and \\(\\widehat\\theta_{t-1}\\),\n\n\\[\\operatorname{Var}(\\widehat\\theta_t-\\widehat\\theta_{t-1}) = \\operatorname{Var}(\\widehat\\theta_t) + \\operatorname{Var}(\\widehat\\theta_{t-1}) - 2\\operatorname{Cov}(\\widehat\\theta_t,\\widehat\\theta_{t-1}).\\]\n\nPositive covariance can reduce the variance of the difference, although the benefit depends on the outcome and design.\n\nLongitudinal designs face attrition, changes in eligibility, and conditioning effects from repeated participation. A sample representing a population at an initial date does not necessarily represent the population at a later date, especially when new people, households, or businesses enter. Cross-sectional and longitudinal weights therefore serve different estimands. Refreshment samples can bring in newly eligible units or replenish coverage, while consistent measurement is needed to distinguish genuine change from changes in instruments or collection modes.\n\nDomain estimation concerns subpopulations such as regions, age groups, or industries. A domain need not coincide with a sampling stratum, and its realized sample size may be random. Direct estimates use observations from the domain itself. Variance estimation must retain the original sample design rather than treating a filtered subset as a new simple random sample. For example, respondents in a small domain may be distributed across only a few PSUs, leaving little independent information even when the respondent count appears adequate.\n\nSmall area estimation combines survey observations with auxiliary information through statistical models when direct domain estimates are too imprecise. The Fay–Herriot model, introduced in 1979, links direct area estimates to area-level predictors while representing their sampling variances. Unit-level models use auxiliary information for individual population units. Such methods borrow strength across areas or observations, but precision depends on model adequacy and auxiliary-data quality. Their uncertainty includes estimation of model parameters and prediction of unobserved quantities, not just the original sampling variance.\n\nReleased survey datasets often include final weights, stratum identifiers, PSU identifiers, or replicate weights so that users can reproduce design-appropriate estimates. Confidentiality protections may limit the geographical detail or identifiers available, creating a practical relationship between statistical usability and data privacy. Analyses of regression, group differences, and hypothesis tests must distinguish sampling weights from frequency weights or inverse-variance weights. Documentation of the target population, frame, selection stages, response treatment, calibration, and variance method establishes which population conclusions the released data can support."
    }
  ],
  "links": [
    "statistics",
    "population-census",
    "estimator",
    "probability-distribution",
    "random-sampling",
    "sample-mean",
    "stratified-sampling",
    "correlation",
    "expected-value",
    "standard-error",
    "confidence-interval",
    "students-t-distribution",
    "bootstrap-sampling",
    "sample-size",
    "linear-regression",
    "mean-squared-error",
    "missing-data-imputation",
    "selection-bias",
    "machine-learning",
    "sensitivity-analysis",
    "bayesian-inference",
    "data-privacy",
    "statistical-hypothesis-testing"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}