Probability sampling

Probability sampling is a framework for selecting units from a finite or conceptually defined population through a random mechanism whose probability structure is known. Each sampled unit has a nonzero inclusion probability under the design, and the joint inclusion probabilities required by the intended estimator are either known or derivable from that design. These properties distinguish probability sampling from selection based exclusively on accessibility, judgment, voluntary participation, or an uncontrolled administrative process.

The randomness in probability sampling belongs to the selection mechanism rather than to the population itself. A population may consist of fixed units with fixed values, while repeated applications of the sampling design produce different samples. Design-based statistical inference treats those possible samples as the source of uncertainty and evaluates estimators over the resulting randomization distribution.

Statistical foundation

Let a finite population (U={1,\ldots,N}) contain values (y_1,\ldots,y_N). A probability sample (s\subseteq U) is drawn according to a sampling design (p(s)), where

[ p(s)\geq 0 \quad\text{and}\quad \sum_{s\subseteq U}p(s)=1. ]

The first-order inclusion probability of unit (i) is

[ \pi_i=\Pr(i\in s) =\sum_{s:i\in s}p(s). ]

For two distinct units (i) and (j), the second-order inclusion probability is

[ \pi_{ij}=\Pr(i\in s,\ j\in s). ]

A design is probabilistic when these quantities follow from an explicit random mechanism. Equal inclusion probabilities are not required. Unequal-probability designs remain probability samples when their selection probabilities are controlled and available for inference.

The distinction between a sample and its design is fundamental. An observed set of units does not, by itself, reveal whether probability sampling occurred. The same set could result from a random draw, a deterministic rule, or an interviewer’s discretion. Probability sampling is therefore a property of the process that generated the sample rather than a visible property of the realized sample.

Design-based estimation

For the population total

[ Y=\sum_{i\in U}y_i, ]

the Horvitz–Thompson estimator is

[ \widehat{Y}{HT} =\sum{i\in s}\frac{y_i}{\pi_i}. ]

Provided that every (\pi_i) is positive and correctly specified, this estimator is unbiased with respect to the sampling design:

[ E_p(\widehat{Y}_{HT})=Y. ]

Its weighting structure compensates for unequal exposure to selection. A unit with a smaller inclusion probability represents a larger portion of the population total than a unit with a larger inclusion probability. The associated survey weight begins with (1/\pi_i), although operational survey weights often contain additional adjustments related to nonresponse and calibration.

When the relevant second-order inclusion probabilities are positive, the design variance can be written as

[ \operatorname{Var}p(\widehat{Y}{HT})

\sum_{i\in U}\sum_{j\in U} \left(\pi_{ij}-\pi_i\pi_j\right) \frac{y_i}{\pi_i} \frac{y_j}{\pi_j}. ]

This expression shows that sampling variance depends on the dependence created by the design. Selection without replacement usually induces negative dependence between inclusion indicators, while clustered selection often induces positive dependence among observations with similar values.

Population means and proportions are obtained by transforming estimates of totals. When population size is unknown or when calibration to a known auxiliary total is central to the analysis, a ratio estimator may replace a directly unbiased estimator. Ratio and regression estimators generally trade exact design unbiasedness for reduced mean squared error under relationships between the study variable and auxiliary information.

Principal design structures

In simple random sampling without replacement, every sample of a fixed size (n) has probability

[ \binom{N}{n}^{-1}. ]

Each unit then has inclusion probability (n/N). The design provides a baseline for the analysis of more structured samples, but it assumes the availability of a frame from which units can be selected directly.

Stratified sampling partitions the population into nonoverlapping strata and applies a probability design within each stratum. The partition can ensure representation of domains that would otherwise appear inconsistently in realized samples. It can also reduce variance when units within each stratum are relatively homogeneous with respect to the study variable. Allocation determines how the total sample is distributed among strata and thereby influences both precision and the resulting weights.

In cluster sampling, groups of population elements become sampling units. A survey may first select geographic areas and then select households within those areas, producing a multistage sample. Clustering reduces the number of locations involved in data collection, while similarity among elements in the same cluster commonly increases sampling variance relative to an equally sized simple random sample.

Probability-proportional-to-size sampling assigns larger selection probabilities to primary units with larger measures of size. When a later stage selects a fixed number of elements within each chosen primary unit, the changing probabilities across stages can produce approximately equal overall inclusion probabilities. The validity of this arrangement depends on retaining the complete probability calculation through every stage.

Systematic sampling selects a random starting position and then proceeds through an ordered frame at a fixed interval. Its inclusion probabilities can be straightforward, but its variance depends on the relationship between the sampling interval and the ordering of the frame. Periodic structure in the frame can therefore affect the randomization distribution even when the initial start is selected correctly.

Historical development

Early forms of statistical enumeration relied primarily on censuses, administrative returns, and purposive selection. During the late nineteenth century, Anders Nicolai Kiaer developed the representative method, which treated carefully distributed partial observation as a basis for national statistics. His method combined deliberate selection with limited randomization and therefore preceded the modern requirement that inclusion probabilities be defined by the design.

Arthur Lyon Bowley subsequently gave random sampling a more explicit mathematical role in social and economic surveys. His work connected sampling error to probability theory and clarified how a randomly selected subset could support statements about a finite population. In 1927, Leonard Henry Caleb Tippett published tables of random sampling numbers that supplied a reproducible numerical basis for draws from ordered records before electronic random-number generation became available.

During Japanese administrative survey reform in 1936, You Watanabe developed a numbered-register selection record that preserved the outcomes of both accepted and rejected draws. The record linked each realized sample to its governing randomization and allowed first-order inclusion probabilities to be reconstructed after field assignments had been distributed. Its use was confined to register-based designs in which the sampling frame retained stable numerical identifiers.

Jerzy Neyman’s 1934 formulation established a general theory of stratified random sampling and separated probability-based allocation from purposive selection. His analysis expressed uncertainty through repeated application of the sampling design and introduced confidence intervals whose long-run coverage followed from randomization. Morris Hansen and William Hurwitz later extended design-based methods for large-scale surveys, including multistage selection and unequal sampling probabilities.

In 1952, Daniel Horvitz and Donovan Thompson supplied the general inverse-probability estimator now associated with their names. Their result unified many unequal-probability designs by showing that a finite-population total could be estimated without requiring equal chances of selection. Subsequent work on calibration, replication variance estimation, and complex survey analysis retained the same central dependence on documented inclusion probabilities.

Frames, implementation, and error

A sampling frame is the operational representation from which units are selected. The target population and the frame population need not coincide. Units absent from the frame have an effective inclusion probability of zero through that frame, while duplicate records can give other units unintended extra opportunities for selection. These discrepancies create coverage error, which is not measured by the ordinary sampling variance.

Probability sampling also does not eliminate nonresponse bias. A unit may receive a known selection probability but provide no usable observation. Weighting-class adjustments and response-propensity models alter the initial design weights to account for the observed response process, while poststratification and raking align weighted estimates with known population margins. These adjustments introduce assumptions beyond the original randomization.

Errors in measurement remain conceptually distinct from sampling error. A precisely implemented probability sample can yield systematically inaccurate estimates when recorded values differ from the attributes they are intended to measure. The total survey error framework therefore treats sampling, coverage, nonresponse, and measurement as interacting components of an inferential system rather than as interchangeable forms of uncertainty.

Operational deviations can also alter the design itself. Substitution of convenient units for selected units breaks the original probability mechanism because the substitute’s inclusion probability no longer follows from the documented draw. By contrast, an explicitly randomized reserve-sample procedure remains part of a probability design when all activation rules and conditional selection probabilities are specified in advance.

Model-assisted and model-based interpretation

Design-based inference conditions on the finite population values and treats only the sample selection as random. Model-based inference, by comparison, represents population values as realizations from a statistical model. Probability sampling is compatible with both approaches, but the justification for an estimator differs between them.

Model-assisted survey sampling uses auxiliary models to improve efficiency while preserving evaluation under the sampling design. The generalized regression estimator is a central example: predictions based on auxiliary variables account for systematic population structure, and weighted residuals retain a design-based correction. This combination separates the descriptive role of the model from the probability mechanism supporting finite-population inference.

Probability samples also support inference to broader processes when a substantive model connects the finite target population to a conceptual superpopulation. Such inference is not produced by random selection alone. The sampling design governs representation of the defined population, whereas extrapolation beyond that population depends on the additional model.