Multistage Sampling: Stages, Design, and When to Use It

Multistage sampling is a probability sampling technique that selects a final sample through two or more successive rounds of random selection, each stage narrowing the population further before the next stage draws from what remains. Rather than sampling individuals directly from a full population — which requires a complete list of every individual, rarely available for large or dispersed populations — multistage sampling starts from broad, easy-to-list units (counties, districts, institutions), randomly selects a subset of them, then repeats the process one or more levels down until it arrives at the units actually measured.

This guide covers what distinguishes multistage sampling from single-stage cluster sampling, the standard terminology for each stage’s units, how stratification is layered into a multistage design, two real national surveys built this way, a worked example, how selection probabilities compound across stages, the design-effect cost, and when multistage sampling is the right call versus the alternatives.

What Is Multistage Sampling?

A multistage design proceeds through a defined sequence of stages, each with its own sampling frame and its own selection method:

1. Stage 1 selects primary sampling units (PSUs) — typically large, well-documented units such as counties, school districts, hospitals, or census enumeration areas — from a frame listing every such unit in the population.
2. Stage 2 selects secondary sampling units (SSUs) from within each selected PSU — schools within a selected district, wards within a selected hospital, households within a selected enumeration area — using a frame that only needs to exist for the PSUs already selected, not the whole population.
3. Additional stages, if the design uses them, repeat the same logic until the design reaches its ultimate sampling unit (USU) — the unit actually measured (a student, a patient, an individual respondent).

The unit selected at each stage is only ever drawn from within units that survived the stage before it. A household can only enter the sample if its enumeration area was already selected at Stage 1 and its block was already selected at Stage 2 — selection is conditional and cumulative, not independent at each level.

Every stage needs its own frame, and only the frame for the current stage’s already-selected units needs to be built — a Stage 3 household frame only has to cover the blocks selected at Stage 2, not every block nationwide. This staged construction is the core practical reason multistage sampling exists: it replaces one impossible, population-wide frame (every individual) with a series of small, buildable ones.

Multistage Sampling vs. Single-Stage Cluster Sampling

Multistage sampling and cluster sampling overlap heavily and are frequently used as if interchangeable, but they answer different questions. Cluster sampling describes what is being selected — naturally occurring groups instead of individuals. Multistage describes how many rounds of random selection happen before reaching the final unit. A one-stage cluster sample (every member of each selected cluster is included) is cluster sampling but not multistage. A two-stage or deeper cluster design — clusters selected, then a sub-sample drawn within each — is both cluster sampling and multistage sampling at once, which is why the two terms get used loosely as synonyms in practice.

Where the two genuinely diverge: a multistage design does not require that every stage select a “cluster” in the everyday sense. A national survey might select counties (Stage 1), then census blocks within each county (Stage 2), then individual addresses within each block (Stage 3) — three stages of increasingly fine-grained geographic selection, with no stage ever including “everyone in a group” the way one-stage cluster sampling does. For the mechanics shared by both — the design effect, how clustering inflates standard errors, and one-stage/two-stage terminology — see the fuller treatment in the cluster sampling guide; this page focuses on the general multistage framework, sampling-unit terminology, and how stratification and mixed selection methods layer into it.

Combining Stratification With a Multistage Design

Multistage and stratified sampling are not alternatives — large surveys routinely use both together, and a design described as “stratified multistage” is stratifying at one or more stages, not choosing stratification instead of staging. A common pattern: PSUs (Stage 1) are first sorted into strata — region, urbanicity, facility type — and a fixed number of PSUs is drawn from within each stratum, guaranteeing representation across every stratum before any further staging happens. Stratification improves precision by ensuring known sources of variation are represented rather than left to chance; multistaging is about feasibility, not precision, and the two corrections are applied for different reasons within the same design.

How Units Are Selected at Each Stage

Different stages of the same design can legitimately use different selection methods:

• Simple random sampling or systematic sampling at a stage where every unit is equally easy to survey once selected.
• Probability-proportional-to-size (PPS) sampling at an earlier stage, so that a PSU with more population has a proportionally higher chance of selection — the standard choice for Stage 1 in large national surveys, since it keeps later-stage sample sizes roughly comparable across selected PSUs regardless of how large the PSU itself is.

Mixing methods across stages is normal, not a design flaw, as long as each unit’s selection probability at each stage it passed through is known and recorded — that record is what makes correct weighting possible afterward.

Real-World Multistage Designs

Two well-known examples illustrate how the stages compose in practice, described here at the structural level rather than with survey-cycle-specific numbers, which change between rounds:

• The National Health and Nutrition Examination Survey (NHANES), run by the U.S. National Center for Health Statistics, uses a complex, stratified, multistage probability design: geographic primary sampling units are selected first, then smaller area segments within them, then households within segments, then individuals within households — with strata and oversampling built in at various stages to support reliable estimates for specific subpopulations.
• Demographic and Health Surveys (DHS), conducted across dozens of countries with a shared methodology, typically use a two-stage stratified design: enumeration areas (clusters) are selected first, usually with PPS given wide variation in area population, then a fixed number of households is listed and randomly selected within each chosen area.

Both are cited here as structural illustrations of how real multistage designs compose stratification, PPS selection, and multiple stages — not as a substitute for a specific survey cycle’s own documented sampling report, which is the authoritative source for exact stage counts and sample sizes in any given round.

Worked Example (Illustrative)

The following is an illustrative composite, not a real study or a real institution. A research team wants to survey clinical staff experience across a national hospital network too large to sample individuals from directly.

1. Stage 1 — PSUs: from a frame of all 400 hospitals in the network, 40 are selected using PPS, so a hospital with more staff has a proportionally higher chance of selection.
2. Stage 2 — SSUs: within each of the 40 selected hospitals, departments are listed and 4 are randomly selected per hospital (160 departments total).
3. Stage 3 — USUs: within each selected department, 10 staff members are randomly selected from the department roster (1,600 staff total).

A staff member’s overall selection probability is the product of the probability their hospital was selected, times the probability their department was selected given the hospital was, times the probability they were selected given their department was — and each staff member’s final sampling weight is the inverse of that compounded probability. See the fuller weight-construction walkthrough, including the base-weight formula, in the survey weighting guide and the multistage weighting section of the cluster sampling guide.

Design Effect and the Precision Cost of Staging

Multistage designs introduce a precision cost analogous to cluster sampling’s: because units within the same PSU or SSU tend to be more similar to each other than to units elsewhere (the same intraclass correlation that drives cluster sampling’s design effect), a multistage sample of a given size yields less precise estimates than a simple random sample of the same size. The design effect (DEFF), commonly attributed to Leslie Kish’s Survey Sampling (1965), quantifies this: DEFF = 1 + (m − 1) × ICC for a simple equal-cluster-size case, where m is the average number of USUs per final-stage cluster and ICC is the intraclass correlation. The effective sample size is the nominal sample size divided by DEFF. A multistage design with several stages of clustering compounds this effect across stages, which is one reason national surveys with 3+ stages typically need substantially larger nominal sample sizes than a single-stage design would to achieve the same precision. Software built for complex survey designs (accounting for stage, stratum, and weight) is required to analyze multistage data correctly — treating a multistage sample as if it were simple random understates standard errors and inflates false-positive rates.

Advantages and Disadvantages

• Advantage — no population-wide frame required. Only a frame of PSUs is needed up front; frames for later stages are built only for the units that survived earlier stages.
• Advantage — fieldwork cost control. Concentrating data collection within a manageable number of PSUs and SSUs sharply reduces travel and logistics costs compared to a geographically scattered simple random sample of the same size.
• Advantage — scalable to very large or dispersed populations. National and multi-country surveys are rarely feasible any other way.
• Disadvantage — precision cost. The design effect described above means a multistage sample needs a larger nominal size than a simple random sample to reach the same precision.
• Disadvantage — analytical complexity. Correct analysis requires complex-survey-design software, careful weight construction, and reporting of stage/stratum/weight details that many published methods sections omit.
• Disadvantage — compounding selection error across stages. An error or bias introduced at an early stage (a poorly constructed PSU frame, for instance) propagates through every later stage drawn from it.

When to Use Multistage Sampling

Multistage sampling is the right choice when the population is too large or too geographically dispersed for a population-wide list of individuals to exist or be affordable to build, and when fieldwork cost makes a scattered simple random sample impractical. It is the wrong choice when a full sampling frame already exists and is affordable to sample from directly — in that case, simple random sampling, systematic sampling, or stratified sampling alone will produce more precise estimates for the same sample size, without the design-effect penalty or the analytical overhead multistaging introduces. See Sampling Methods for how multistage sampling fits alongside the other probability and non-probability designs, and the Sampling & Statistical Power hub for the rest of this sub-cluster.

Frequently Asked Questions

What is multistage sampling?

Multistage sampling is a probability sampling method that selects a final sample through two or more successive rounds of random selection — primary sampling units first, then progressively smaller units within each selected unit — instead of sampling individuals directly from the whole population.

What’s the difference between multistage sampling and cluster sampling?

Cluster sampling describes what is selected (naturally occurring groups rather than individuals); multistage describes how many rounds of selection happen. A two-stage-or-deeper cluster design is also a multistage design, but a multistage design doesn’t have to select “clusters” in the everyday sense at every stage — it can select purely geographic or administrative units at each level.

What is a primary sampling unit (PSU)?

The unit selected at Stage 1 of a multistage design — typically a large, well-documented unit such as a county, district, or institution — drawn from a frame listing every such unit in the population before any later-stage selection happens.

Why use multistage sampling instead of simple random sampling?

Simple random sampling requires a complete list of every individual in the population, which is often unavailable or prohibitively expensive to build for large or dispersed populations. Multistage sampling only requires a frame of primary sampling units up front, with later-stage frames built only for units that already survived selection.

How do you calculate selection probability in a multistage design?

Multiply the selection probability at each stage together: a unit’s overall probability of selection equals the product of the probability its PSU was selected, times the probability its SSU was selected given the PSU was selected, times the probability it was itself selected given its SSU was selected, and so on for additional stages. The sampling weight for that unit is the inverse of this compounded probability.

Do real national surveys use multistage sampling?

Yes — large national and multi-country surveys, including the U.S. National Health and Nutrition Examination Survey (NHANES) and the Demographic and Health Surveys (DHS) conducted in dozens of countries, use stratified multistage probability designs specifically because a population-wide list of individuals isn’t available or affordable to build directly.