Stratified sampling is a method of random sampling in which a population is divided into mutually exclusive, collectively exhaustive groups called strata, and a sample is selected separately within each stratum. In statistics, it is used to improve the precision of population estimates, obtain adequate information about particular subgroups, or organize data collection. Population estimates combine the stratum results using weights that reflect the sampling design; an unweighted analysis is generally inappropriate when different strata have different sampling fractions.
Design and rationale
A stratification is a partition of the population: every population unit belongs to exactly one stratum. Strata may be defined by geographic region, business size, age group, land-cover type, or combinations of characteristics known before sampling.
A basic stratified design involves four steps:
- Define the target population and a sampling frame that identifies its units.
- Assign each unit to a stratum.
- Allocate a sample size to each stratum.
- Select units independently across strata using a specified probability-sampling method.
The most common elementary design uses simple random sampling without replacement within every stratum. More complex designs may use unequal selection probabilities or multiple sampling stages within strata.
The main precision advantage arises when units within each stratum are relatively similar for the quantity being measured, while stratum means differ. Separate sampling fixes the number of observations drawn from each group, reducing variation caused by an accidentally unbalanced sample composition.
Stratification can also serve goals other than improving an overall estimate. A small population subgroup may need a large sampling fraction to support a useful subgroup estimate. Administrative regions may require separate results, or different strata may require different collection methods.
Estimation and sampling weights
Suppose a finite population contains units divided into strata. Let
- be the population size of stratum ;
- be its sample size;
- be its population share;
- be its sampling fraction;
- be its sample mean.
The total sample size is . Under simple random sampling within strata, an estimator of the population mean is
An estimator of the population total is
These estimators are unbiased under the stated sampling design: their expected values over repeated sampling equal the corresponding population quantities.
Within stratum , each unit has inclusion probability
and therefore the base sampling weight is
Each sampled unit represents population units. Because these weights sum to , the weighted sample mean equals .
The unweighted sample mean instead weights each stratum by its sample share, . It agrees with the stratified estimator when those shares equal the population shares, as in exact proportional allocation. Otherwise, it generally estimates a differently weighted quantity.
For a population proportion, the same formulas apply by coding each observation with an indicator function equal to 1 when the characteristic is present and 0 otherwise.
Variance and uncertainty
Let the finite-population variance within stratum be
where is the stratum population mean.
If samples are selected independently across strata, the variance of the stratified mean estimator is
The factor is the finite-population correction. It accounts for sampling without replacement and becomes important when a substantial proportion of a stratum is observed. A completely enumerated stratum, with , contributes no sampling variance.
An unbiased variance estimator replaces with the sample variance :
Its square root is the estimated standard error. Under suitable large-sample conditions, an approximate confidence interval is
where is a standard normal-distribution quantile.
Small samples, highly skewed outcomes, and complex within-stratum designs may require different approximations. A noncensus stratum containing only one sampled unit does not provide a direct within-stratum sample variance estimate, so additional assumptions or specialized variance procedures are needed.
Allocation of the sample
Allocation determines how the total sample is distributed across strata. Different allocations serve different objectives.
Proportional allocation
Proportional allocation sets
Ignoring integer rounding, all strata have the same sampling fraction and all sampled units have the same base weight. The design is then self-weighting.
This allocation preserves population shares in the sample but may provide too few observations for small subgroups.
Equal allocation
Equal allocation assigns approximately observations to every stratum. It can support comparisons among strata, but it oversamples small strata and undersamples large ones relative to their population shares.
Equal allocation is not generally optimal for estimating the overall population mean. Its usefulness depends on the importance attached to subgroup estimates.
Neyman allocation
For a fixed total sample size, equal per-unit costs, and simple random sampling within strata, the allocation minimizing the variance of the overall mean is
This is called Neyman allocation, after Jerzy Neyman, whose 1934 work helped establish the design-based theory of stratified sampling.
The rule assigns more observations to strata that are larger or have greater within-stratum standard deviation. Since the values are usually unknown in advance, planning may use pilot studies, earlier surveys, or related auxiliary information.
Allocation with unequal costs
If the cost per sampled unit in stratum is , a variance-minimizing allocation under a fixed variable-cost budget has the form
Other things equal, expensive strata receive fewer observations than under Neyman allocation.
These formulas describe continuous optima. Actual allocations require integer sample sizes, cannot exceed stratum population sizes, and may impose minimum samples for subgroup reporting. With several important outcomes or precision constraints, allocation becomes a mathematical optimization problem rather than a single universal rule.
Numerical example
Consider a population of 10,000 businesses:
| Stratum | Population size | Sample size | Sample mean revenue |
|---|---|---|---|
| Small businesses | 9,000 | 100 | 100,000 currency units |
| Large businesses | 1,000 | 100 | 1,000,000 currency units |
The estimated population mean revenue is
The unweighted sample mean would be 550,000 because the two strata contribute equal numbers of observations. That calculation gives large businesses half the weight, although they make up only one-tenth of the population.
The base weights are 90 for each sampled small business and 10 for each sampled large business. Estimated total revenue is
The sample means alone do not determine the precision of this estimate; within-stratum variances are also required.
Choosing strata and assessing efficiency
Effective stratification variables are available for the sampling frame before selection and are informative about important survey outcomes. A variable strongly associated with one outcome may be weakly associated with another, so a stratification suitable for estimating income may not be equally effective for estimating travel behavior.
Strata need not have equal sizes or correspond to socially meaningful groups. Continuous variables, such as business employment or farm area, can be divided into intervals. Very large or influential units may form a certainty stratum, in which every unit is selected.
Precision gains depend on both the boundaries and the allocation. Under proportional allocation, stratification reduces the contribution of between-stratum differences to sampling variability; with negligible sampling fractions, this gives a clear variance advantage when stratum means differ. However, an inefficient disproportionate allocation can produce a larger variance than simple random sampling of the same total size.
A common efficiency measure is the design effect:
where the denominator refers to an appropriate simple-random-sampling benchmark. A value below 1 indicates lower sampling variance for that estimator and comparison.
Related methods
Cluster sampling
In cluster sampling, groups of units are selected, and observations are collected from all or some units within the selected groups. In a conventional stratified design, by contrast, a sample is drawn from every stratum.
Stratification often seeks internally homogeneous groups to improve precision. Cluster sampling often uses naturally occurring groups, such as schools or neighborhoods, to reduce collection costs. The two methods can be combined: clusters may be selected separately within geographic strata.
Post-stratification
Post-stratification adjusts weights after sampling so that weighted counts agree with known population totals for specified groups. Unlike design stratification, it does not fix subgroup sample sizes before selection and cannot ensure that every subgroup appears in the sample.
Both methods use auxiliary population information, but their sampling properties and variance calculations differ.
Quota sampling
Quota sampling may also set target numbers for subgroups, but selection within groups need not be random. Meeting subgroup quotas alone does not establish known inclusion probabilities or justify the design-based formulas used for stratified random sampling.
Stratified randomization
In experimental design, treatment assignment may be randomized separately within strata to balance treatment groups on selected characteristics. This concerns treatment assignment rather than selection from a population. Stratified sampling and stratified randomization can occur in the same study but address different sources of variation.
Applications and limitations
Stratified sampling is widely used in household surveys, business statistics, agricultural surveys, environmental monitoring, and quality assessment. It supports both overall estimates and estimates for reporting domains. In Monte Carlo methods, an analogous technique partitions a simulation domain and samples within its parts to reduce numerical estimation variance.
Its main practical limitation is the need for adequate frame information. Missing units, outdated classifications, or incorrect stratum assignments can undermine coverage and efficiency. Creating many strata can also increase administrative complexity and leave too few observations in individual strata for stable variance estimation.
Stratification does not by itself remove measurement error, nonresponse, or selection bias caused by departures from the intended design. Weight adjustments for nonresponse and calibration may be needed, and uncertainty calculations must reflect the actual selection and adjustment procedures. Treating a complex stratified sample as an ordinary independent, identically distributed sample can give incorrect standard errors even when the point estimate uses suitable weights.