Stratified sampling is a cornerstone of rigorous survey design and statistical planning. By dividing a population into homogeneous groups, or strata, and then sampling within those strata, researchers can achieve more precise estimates with potentially smaller samples than simple random sampling. The backbone of this approach is the formula for stratified sampling—the set of rules that determines how many units to sample from each stratum and how to combine those samples into an overall estimate. This article unpacks the theory, the formulas, and the practical steps required to implement stratified sampling effectively in a wide range of contexts, from market research to public health audits.
Understanding the basics: what is stratified sampling?
In stratified sampling, the population is partitioned into non-overlapping subgroups called strata. Each stratum contains units that are similar with respect to the characteristic of interest. For example, a national health survey might stratify by region, age group, or socio-economic status. The goal is to ensure that the sample accurately reflects the diversity of the population, and to reduce the variance of the estimator by exploiting within-stratum homogeneity.
Key terms to know include:
- Population size (N): the total number of units in the population.
- Stratum (h): each subpopulation, indexed by h = 1, 2, …, H.
- Stratum size (N_h): the number of units in stratum h, with N = sum_h N_h.
- Sample size (n): the total number of units selected from the population.
- Stratum sample size (n_h): the number of units sampled from stratum h, with n = sum_h n_h.
- Stratum mean (ȳ_h): the average value of the variable of interest within stratum h.
- Stratum variance (S_h^2): the variance of the variable of interest within stratum h.
- Weights (W_h): typically W_h = N_h / N, the relative size of each stratum in the population.
formula for stratified sampling: the baseline concepts
At the heart of the approach are two practical questions: how many observations should be drawn from each stratum, and how should the information from all strata be combined to produce an estimate for the population parameter of interest? The answers depend on the allocation rule and the estimator used. The most common estimation target is the population mean, though stratified sampling can also be used for totals, proportions, and other statistics.
Proportional allocation: a straightforward starting point
Proportional allocation is often the simplest and most intuitive approach. The idea is to sample in each stratum in proportion to its size in the population. If the population has N_h units in stratum h and the total sample size is n, then the stratum sample sizes are:
n_h = n × N_h / N
Advantages of proportional allocation include simplicity and interpretability. It tends to work well when there is little variation in the variable of interest across strata or when sampling costs are roughly similar across strata. The corresponding estimator for the population mean is a weighted average of the stratum means:
ȳ_st = ∑_h (W_h × ȳ_h), where W_h = N_h / N.
Neyman allocation: optimising precision under cost constraints
When costs differ across strata or when the variation within strata differs substantially, Neyman allocation can improve precision. It allocates more samples to strata with greater product of size and standard deviation, to minimise the variance of the estimator for a given total sample size. The formula for the stratum sample size under Neyman allocation is:
n_h = n × (N_h × S_h) / ∑_k (N_k × S_k)
Here, S_h denotes the standard deviation of the variable of interest within stratum h. The intuition is clear: allocate more observations to strata that are both large and heterogeneous. The estimator remains the same form as in proportional allocation:
ȳ_st = ∑_h (W_h × ȳ_h), with W_h = N_h / N.
Practical considerations: costs, errors, and feasibility
In real-world applications, the pure Neyman or proportional rules may not be fully feasible due to budget, time, or logistical constraints. In such cases, researchers frequently consider constrained or optimised allocations that strike a balance between precision and practical limits. Alternatives include:
- Cost-constrained allocation: allocate within a fixed budget while seeking minimum variance.
- Optimal allocation with minimum sample sizes per stratum to meet a target margin of error.
- Two-stage or multi-stage sampling for complex populations where elements of certain strata are hard to reach.
Estimating population characteristics using the stratified framework
Beyond merely selecting samples, stratified sampling provides efficient ways to estimate population parameters. The two most common goals are estimating the population mean and estimating the population total. The formulas below show how the stratified design contributes to accurate, low-variance estimates.
Estimating the population mean
Assuming the population is partitioned into H strata with known N_h, the estimator for the population mean is:
ȳ_st = ∑_h (W_h × ȳ_h), where ȳ_h is the sample mean in stratum h and W_h = N_h / N.
In practice, ȳ_h is computed from the sampled units in stratum h. If a stratum contains only a few sampled units, its estimate will be more influenced by random variation; this is why appropriate allocation is critical for precision.
Estimating the variance of the stratified mean
The variance of the stratified mean estimator under simple random sampling within each stratum is given by:
Var(ȳ_st) = ∑_h (W_h^2 × (S_h^2 / n_h))
Where S_h^2 is the within-stratum variance and n_h is the number of observations drawn from stratum h. This formula makes explicit how the precision benefits depend on within-stratum homogeneity and the allocation of samples across strata.
Important caveats include:
- If the sampling within strata is done via simple random sampling without replacement (SRSWOR), the finite population correction (FPC) can be incorporated when strata are small relative to the sample.
- With complex sampling within strata, such as systematic sampling or stratified cluster sampling, the variance estimation becomes more intricate and may require design-based variance estimators or replication methods.
ipractical steps to implement the formula for stratified sampling in practice
Turning theory into practice involves a sequence of deliberate steps. Below is a pragmatic checklist that highlights where the formula for stratified sampling plays a central role in each stage of the process.
Step 1: Define the population and strata
Identify the population of interest and choose stratification variables that are strongly related to the variable of interest and logistically feasible to measure. Common stratification dimensions include geography, time, age, income band, or educational attainment. The key is that units within a stratum should be as homogeneous as possible with respect to the target parameter.
Step 2: Acquire or estimate stratum sizes (N_h)
Obtain or approximate the size of each stratum within the population. Accurate N_h values are essential because they determine both the weighting W_h and the allocation of the sample. In some organisations, data from census records or administrative databases provide reliable stratum sizes; in others, researchers may rely on recent estimates or prior surveys.
Step 3: Decide total sample size (n) and allocation strategy
Set a target total sample size that aligns with the study’s precision requirements and budget constraints. Decide whether to use proportional allocation or Neyman allocation, or a customised mix that fits practical considerations. If using Neyman allocation, obtain or estimate S_h, the within-stratum standard deviations. If S_h is unknown, consider pilot studies or existing literature to inform these estimates.
Step 4: Compute n_h using the chosen allocation rule
Using proportional allocation, compute:
n_h = n × N_h / N
Alternatively, using Neyman allocation, compute:
n_h = n × (N_h × S_h) / ∑_k (N_k × S_k)
When designing the sample, ensure that each n_h is an integer and does not exceed N_h. If necessary, round while keeping the total at n and adjusting to maintain feasibility. This step is where the formula for stratified sampling directly shapes the sampling plan.
Step 5: Collect data and compute stratum estimates
Proceed to collect data within each stratum according to the allocated n_h. Calculate ȳ_h, S_h^2, and any other stratum-level statistics. It is important to implement consistent data collection procedures to avoid systematic biases across strata. Well-documented procedures improve the reliability of subsequent estimates derived from:
ȳ_st = ∑_h (W_h × ȳ_h)
Step 6: Estimate the overall population parameter and its precision
Combine the stratum estimates using the weights W_h to obtain the overall estimate. Calculate the variance of the stratified estimator and, if possible, construct confidence intervals. The general form for the variance used in a standard stratified sample is:
Var(ȳ_st) = ∑_h (W_h^2 × (S_h^2 / n_h))
These variance estimates guide interpretation and help you communicate the precision of your results to stakeholders.
Worked example: illustrating the formula for stratified sampling in action
Consider a hypothetical national employee satisfaction survey. The population comprises 1,000 units (N = 1000), divided into two strata by region: Region A with N_1 = 600 and Region B with N_2 = 400. The analyst plans to obtain a total sample size of n = 200. Historical data suggest within-region standard deviations S_1 = 0.90 and S_2 = 1.10.
Step 1: Decide allocation. Suppose we apply Neyman allocation to maximise precision under a fixed sample size.
Step 2: Compute the denominator: ∑_k (N_k × S_k) = (600 × 0.90) + (400 × 1.10) = 540 + 440 = 980.
Step 3: Compute n_h for each region:
n_1 = 200 × (600 × 0.90) / 980 ≈ 200 × 540 / 980 ≈ 110.2
n_2 = 200 × (400 × 1.10) / 980 ≈ 200 × 440 / 980 ≈ 89.8
Rounding to feasible integers, we might allocate n_1 = 110 and n_2 = 90.
Step 4: Collect data and calculate stratum means ȳ_1, ȳ_2, and stratum variances S_1^2, S_2^2 based on the samples. Step 5: Compute the overall estimate:
ȳ_st = (N_1 / N) × ȳ_1 + (N_2 / N) × ȳ_2 = (600/1000) × ȳ_1 + (400/1000) × ȳ_2.
Step 6: Compute the variance of the stratified mean:
Var(ȳ_st) = (W_1^2 × S_1^2 / n_1) + (W_2^2 × S_2^2 / n_2) = (0.6^2 × S_1^2 / 110) + (0.4^2 × S_2^2 / 90).
In this simplified example, the allocation emphasises regions with greater product of size and variability, yielding more precise estimates for the overall mean without inflating the total sample size.
Practical considerations, pitfalls, and best practices
Stratified sampling offers clear advantages, but researchers must address several practical considerations to avoid biases and inefficiencies. The following guidance helps ensure robust applications of the formula for stratified sampling.
Ensuring valid strata
Strata should be mutually exclusive and collectively exhaustive. Each unit belongs to one and only one stratum, and every unit is included in some stratum. Poorly defined strata can blur differences and undermine the benefits of stratification. If strata are poorly constructed or overly similar, variance reduction may be marginal, and the complexity added by stratification may not be warranted.
Dealing with unknown stratum sizes
If N_h is uncertain, the weighting W_h becomes less precise. In such cases, it is prudent to conduct sensitivity analyses, use conservative estimates, or plan a pilot survey to refine stratum sizes before the main data collection.
Balancing precision and cost
While Neyman allocation often improves precision, it can be more costly to sample from certain strata, particularly if fieldwork or laboratory analyses are expensive in some regions. In practice, practitioners often adopt a hybrid approach, imposing minimum sample sizes for hard-to-reach strata and adjusting allocations to stay within budget while preserving overall precision.
Adjusting for non-response and missing data
Non-response can distort stratified estimates if response rates differ across strata. Researchers should anticipate non-response by inflating n_h accordingly or by employing response rate adjustments post-data collection. When missing data are being treated, consider multiple imputation or weighting adjustments within strata to preserve the validity of your formula for stratified sampling.
Variance estimation under complex sampling designs
In real surveys, sampling within strata may involve clustering, unequal selection probabilities, or multi-stage designs. In such cases, standard variance formulas may understate uncertainty. Replication methods (jackknife, bootstrap) or design-based variance estimators are often used to obtain accurate standard errors. Always document the design and the variance estimator used to keep results credible and reproducible.
Related methods and alternatives to stratified sampling
While stratified sampling is powerful, it is part of a broader family of sampling techniques. Understanding the landscape helps researchers select the most appropriate approach for their study.
Cluster sampling vs stratified sampling
In cluster sampling, the population is divided into clusters, often geographically, and a sample of clusters is chosen with all or a subset of units within selected clusters being observed. Clustering can reduce logistical costs but may increase variance if units within clusters are highly similar. Stratified sampling, by contrast, aims to homogenise within-strata and maintain heterogeneity across the population through careful weighting and allocation.
Systematic sampling and stratification
Systematic sampling can be used within strata to simplify fieldwork, especially in constrained environments. For example, after defining strata, a systematic interval within each stratum might be sampled to reduce interviewer burden. Care must be taken to ensure that any periodicities in the population do not bias the estimates.
Post-stratification and calibration weighting
Sometimes the sampling design is fixed, but the final sample deviates from the planned composition due to non-response or eligibility. In such cases, post-stratification or calibration weighting can realign the sample to known population margins, maintaining the intent of the original strata and improving representativeness.
Common questions about the formula for stratified sampling
Below are concise answers to frequent queries that practitioners often raise when designing stratified surveys.
Why use stratified sampling at all?
Stratified sampling can reduce variance and improve estimator precision when strata are internally homogeneous but different from one another. It is particularly advantageous when the attribute of interest varies between subpopulations or when costs differ across strata.
What is the advantage of Neyman allocation?
Neyman allocation minimises the variance of the estimator for a given total sample size by allocating larger samples to strata with greater variability and larger population size. It is especially beneficial when within-stratum variability is heterogeneous and when sample costs differ across strata.
How do I deal with unknown standard deviations S_h?
If S_h is not known before data collection, use pilot studies, historical data, or expert opinion to obtain reasonable estimates. You may also perform interim analyses and adjust allocations as more information becomes available, in line with adaptive design principles.
Can I combine stratified sampling with other sampling methods?
Yes. Hybrid designs may involve stratified sampling with cluster sampling within strata, or a two-stage stratified design where the first stage selects units (e.g., PSUs) and the second stage samples individuals within selected units. This combination can offer both operational efficiency and statistical robustness when executed carefully.
Conclusion: key takeaways on the formula for stratified sampling
The formula for stratified sampling provides a disciplined framework for planning, sampling, and analysing data. By partitioning a population into meaningful strata, researchers can achieve targeted precision gains and cost efficiencies. Whether employing proportional allocation for simplicity or using Neyman allocation to optimise precision under variability and cost constraints, the essence lies in accurately determining stratum sizes, allocating samples appropriately, and combining stratum results through well-defined weights. When carried out with rigorous data collection practices and appropriate variance estimation, stratified sampling offers a robust path to reliable, generalisable insights in a British English context and beyond.
Remember the core steps: define sensible strata, determine N_h, choose an allocation rule (proportional or Neyman), calculate n_h, collect data within each stratum, compute ȳ_h and S_h^2, and combine results using weights W_h to obtain the final population estimates. With careful attention to design, execution, and analysis, the formula for stratified sampling becomes a powerful ally in producing accurate, credible, and actionable results.