03 · What You Need to Know
When Does Multistage Sampling Become the Appropriate Design?
What Exactly Is Multistage Sampling?
Multistage sampling is a probability sampling design in which selection occurs through two or more successive stages. At each stage, the units eligible for selection are determined by the units selected at the preceding stage.
For example, a researcher might select provinces at the first stage, schools within selected provinces at the second stage, and teachers within selected schools at the third stage.
The units selected at the first stage are commonly called primary sampling units, or PSUs. Units selected at subsequent stages may be described as secondary sampling units and, where relevant, tertiary sampling units.
The ultimate sampling units are the units whose data are collected for the intended analysis. In a teacher survey, teachers may be the ultimate sampling units even though provinces and schools are selected first.
Multistage sampling is often associated with cluster sampling because higher-stage units frequently contain groups of lower-stage units. However, not every cluster sampling design requires multiple selection stages.
One-stage cluster sampling
Selects clusters and includes every eligible ultimate population unit within the selected clusters.
Multistage sampling
Selects units at successive stages, such as selecting schools and then randomly selecting teachers within those schools.
When Is a Complete Individual-Level Sampling Frame Unavailable?
One of the strongest justifications for multistage sampling is the absence of a practical population-wide frame of individual units.
Consider a regional study of households. A government agency may maintain a reliable list of census enumeration areas but not a current list of every household in the region. Preparing a household listing for every enumeration area would require extensive fieldwork.
A multistage design allows the researcher to select enumeration areas first and then construct household lists only within those selected areas.
The same logic applies to educational research. A national school registry may be available even when teacher or student rosters are maintained separately by individual schools.
In these situations, multistage sampling can substantially reduce the cost of frame construction.
However, the absence of a complete individual-level frame does not remove the need for valid frames. Each stage requires a sufficiently accurate frame of the units eligible for selection at that stage.
When Does Geographical Dispersion Justify Multiple Sampling Stages?
Multistage sampling may be useful when the target population covers a large geographical area and individual-level selection would produce a widely dispersed sample.
Suppose a researcher randomly selects 800 teachers from a nationwide teacher registry. Those teachers could be distributed across hundreds of schools, requiring extensive coordination and potentially costly site visits.
Instead, the researcher could select districts, then schools within districts, and finally teachers within schools. The sample would be geographically concentrated, making data collection more manageable.
This advantage is particularly relevant when researchers must conduct face-to-face interviews, administer supervised assessments, observe classrooms, or obtain institutional permissions.
For a fully online survey with reliable contact information, however, the logistical advantage of geographically concentrating the sample may be much smaller.
The appropriate design therefore depends partly on the data collection mode, not simply on the population's geographical size.
How Many Sampling Stages Are Necessary?
There is no universally appropriate number of stages. The design should use enough stages to make sampling operationally feasible without introducing unnecessary complexity.
| Design |
Selection sequence |
Illustrative application |
| Two-stage |
Schools → teachers |
Select schools, then randomly select teachers within them. |
| Three-stage |
Districts → schools → teachers |
Select districts, schools within selected districts, and teachers within selected schools. |
| Four-stage |
Regions → districts → schools → students |
Select geographical and institutional units successively before selecting students. |
Adding a stage can reduce the number of locations requiring detailed listing or fieldwork. Yet it can also increase the concentration of observations within higher-level units, potentially increasing sampling variance.
A researcher should therefore ask what operational problem each proposed stage solves. If districts can be sampled directly from a complete national district frame, selecting regions first may be unnecessary unless regional selection offers a defensible logistical or analytical advantage.
How Do You Construct Sampling Frames at Different Stages?
Each sampling stage requires a frame appropriate to the units being selected.
For a three-stage educational survey, the researcher might use a district registry at the first stage, school lists for selected districts at the second stage, and verified teacher rosters for selected schools at the third stage.
These frames do not all need to exist before the first selection. A lower-stage frame may be constructed after the relevant higher-stage units have been selected.
Nevertheless, the procedure must ensure that eligible population units have the intended opportunity for selection.
Potential problems include omitted districts, schools assigned to the wrong district, duplicate institutional records, and teachers employed at multiple schools.
Researchers should establish consistent eligibility and membership rules and address duplicate or outdated sampling-frame records before applying the relevant selection procedure.
How Are Units Selected at Each Stage?
Multistage sampling does not prescribe one selection method for every stage. Different probability sampling methods may be combined within a single design.
For example, districts might be selected using probability proportional to size sampling, schools within districts might be selected through simple random sampling, and teachers within schools might be selected systematically.
Each procedure must be justified and implemented according to its own requirements.
When using systematic sampling at a particular stage, the researcher must consider the ordering of the frame and potential periodicity.
When using probability proportional to size selection, the measure of size and its relationship to the intended ultimate population units should be specified.
Different methods can be combined without compromising probability sampling, provided the resulting inclusion probabilities can be determined and the design is implemented correctly.
How Are Overall Selection Probabilities Calculated?
For a unit selected through successive conditional stages, its overall inclusion probability is obtained by multiplying the relevant stage-specific probabilities along its selection path.
Overall selection probabilities are essential for determining design weights and making population-level estimates.
Where different ultimate units have different inclusion probabilities, sampling weights must reflect those differences when required by the estimator.
When Is Probability Proportional to Size Sampling Useful?
Higher-stage units often differ substantially in size. A district may contain 20 schools while another contains 200. A school may employ 15 teachers while another employs 150.
If equal numbers of ultimate units are selected from higher-stage units of very different sizes, their overall selection probabilities may differ considerably.
Probability proportional to size, or PPS, sampling can help manage this issue by assigning larger higher-stage units greater selection probabilities.
For example, schools may be selected with probabilities related to their teacher counts. If the design then selects a suitable number of teachers within each selected school, the combined procedure may produce approximately equal teacher inclusion probabilities.
However, PPS does not automatically create equal probabilities. The result depends on the actual selection algorithm, size measures, treatment of certainty units, and subsequent sampling stages.
Researchers should calculate the overall probabilities rather than assuming that the design is self-weighting.
How Does Multistage Sampling Affect Statistical Precision?
Multistage sampling often concentrates observations within selected higher-level units. Individuals within the same school, district, or community may share characteristics associated with the study outcomes.
Such similarity can increase sampling variance compared with a simple random sample containing the same number of individuals.
The magnitude depends on the number and sizes of selected clusters, intracluster correlations, stratification, unequal weights, and the estimators being used.
Researchers should therefore distinguish the total number of individual observations from the amount of independent information represented by the design.
For many outcomes, selecting more independent higher-stage units may improve precision more effectively than increasing the number of individuals within a small number of clusters.
These issues are closely related to how clustering affects sample size and the design effect used in sample size planning.
What Are the Advantages and Limitations of Multistage Sampling?
Potential Advantages
- Reduces the need to construct a complete individual-level frame for the entire population.
- Can substantially reduce travel and fieldwork coordination costs.
- Supports probability sampling in large, geographically dispersed populations.
- Allows different probability selection methods to be combined across stages.
- Can accommodate stratification and other design features when appropriately planned.
Potential Limitations
- Requires accurate frames and selection records at every stage.
- May increase sampling variance because observations are clustered.
- Can produce unequal inclusion probabilities requiring weighting.
- Requires more complex sample size planning and variance estimation.
- May be vulnerable to institutional nonresponse or frame errors at higher sampling stages.
When Is Multistage Sampling Unnecessary?
Multistage sampling may be unnecessary when a reliable individual-level sampling frame is available and selected units can be contacted efficiently.
For example, a university studying its own students may already have a complete enrollment registry and institutional email addresses. If the survey is administered online, simple random sampling may be easier to implement and analyze.
Likewise, if important population subgroups require deliberate sample allocation, stratified sampling may address the primary objective without requiring additional geographical selection stages.
The need for multistage sampling should emerge from actual frame limitations, costs, or operational requirements rather than the assumption that large-scale studies must use complex sampling designs.
04 · A Practical Example
Designing a Three-Stage Survey of Teachers' AI Readiness
Hypothetical Example
A Regional Educational Technology Survey
A researcher wants to estimate teachers' readiness to use generative AI in a region containing 20 districts. Each district has 50 schools, and each school has 40 eligible teachers.
The total population is therefore 20 × 50 × 40 = 40,000 teachers. The researcher has a district registry but cannot obtain a complete teacher-level frame across the region.
A three-stage probability design is proposed.
Stage 1: Select districts.
Randomly select 5 of the 20 districts without replacement.
Stage 2: Select schools.
Within each selected district, randomly select 10 of its 50 schools.
Stage 3: Select teachers.
Within each selected school, obtain a verified teacher roster and randomly select 12 of its 40 eligible teachers.
The design produces 5 × 10 = 50 selected schools and 50 × 12 = 600 selected teachers.
Under the stated equal-size assumptions, the overall teacher inclusion probability is:
The researcher can now construct teacher rosters for only 50 selected schools rather than all 1,000 schools in the region. This represents a substantial reduction in frame-construction requirements.
However, the sample is concentrated within five districts. Teachers from the same district and school may share contextual influences, so the researcher must account for the multistage structure when estimating uncertainty.
Only five first-stage districts are selected. That limited number may create serious difficulties for precise design-based variance estimation, particularly for subgroup analyses. The researcher should evaluate whether selecting more districts and fewer schools per district would produce a more defensible balance between cost and precision.
The example illustrates that an operationally feasible multistage design is not necessarily an adequately powered or sufficiently precise design. Those requirements must be assessed before data collection.
