Stratified random sampling is a probability sampling method in which researchers divide a population into non-overlapping subgroups called strata and randomly select units from every stratum. It is especially useful when important subgroups differ from one another or when small groups need guaranteed representation in the final sample.
Introduction
A simple random sample is easy to understand: create a list of the population and randomly choose the required number of units. However, this procedure can produce too few participants from a small but important subgroup purely by chance.
Stratified random sampling addresses this problem by dividing the population before selection. Researchers then draw a separate random sample from each subgroup. This makes it possible to guarantee subgroup representation, improve the precision of some population estimates, and produce reliable estimates for specific groups.
This article explains the meaning of stratified random sampling, its main types, formulas, implementation steps, advantages, limitations, analysis requirements, software options, and modern applications.
Key Takeaways
- Stratified random sampling divides a population into mutually exclusive and collectively exhaustive strata.
- A random sample is selected separately from every stratum.
- Proportionate allocation mirrors the population distribution, while disproportionate allocation oversamples or undersamples selected strata.
- Disproportionate designs normally require sampling weights for population-level estimates.
- Stratification can improve precision when the stratification variables are related to the study outcomes.
- Stratification does not automatically correct nonresponse, poor measurement, or an incomplete sampling frame.
What Is Stratified Random Sampling?
Stratified random sampling is a probability sampling technique that separates a target population into distinct groups and then independently selects a random sample from each group.
The groups are called strata, while one group is called a stratum. Strata are normally based on characteristics known before sample selection, such as:
- Academic level
- Geographic region
- Age category
- Employment sector
- School type
- Department
- Income band
- Organization size
For example, a university researcher studying student satisfaction could divide all enrolled students into undergraduate, master’s, and doctoral strata. A random sample would then be selected from each academic level.
The defining feature is not merely that the researcher records participants’ group membership. The population must be divided into strata before selection, and a probability-based selection method must be applied within each stratum.
Core Terms
Target Population
The target population is the complete group about which the researcher wants to draw conclusions. It might consist of all students at a university, all registered nurses in a country, or all businesses in a particular industry.
Sampling Frame
A sampling frame is the operational list or database from which the sample is selected. Examples include a student-registration database, employee roster, household-address file, or business register.
A study can use a technically correct random procedure and still produce biased results when the frame omits part of the target population.
Stratum
A stratum is one non-overlapping subgroup of the population. Every eligible unit should be assigned to exactly one stratum under the chosen stratification scheme.
Stratification Variable
A stratification variable is the characteristic used to form the strata. Researchers should choose variables that are related to:
- Important study outcomes
- Planned subgroup comparisons
- Data-collection costs
- Response rates
- Geographic organization
- Policy-reporting requirements
Sampling Fraction
The sampling fraction in stratum (h) is:
[
f_h = \frac{n_h}{N_h}
]
where:
- (n_h) is the stratum sample size.
- (N_h) is the stratum population size.
Inclusion Probability
Under simple random sampling without replacement within stratum (h), the probability that a unit is selected is:
[
\pi_{hi} = \frac{n_h}{N_h}
]
A unit’s base sampling weight is the inverse of this probability:
[
d_{hi} = \frac{1}{\pi_{hi}} = \frac{N_h}{n_h}
]
How Does Stratified Random Sampling Work?
Stratified random sampling follows four basic principles:
- Divide the population into meaningful strata.
- Determine how much of the total sample to assign to each stratum.
- select units randomly and independently within every stratum.
- Combine the selected units for data collection and analysis.
Suppose a company employs 1,000 people in four departments. Rather than drawing 100 employees from the full staff list, the researcher first divides the list by department. The researcher then randomly selects the required number from each departmental list.
This procedure guarantees that every department contributes to the final sample.
Requirements for Valid Strata
Strata Must Be Mutually Exclusive
Each population unit must belong to only one stratum under the selected classification.
For example, the categories “first-year,” “second-year,” “third-year,” and “fourth-year” are mutually exclusive when each student has one official year level.
Overlapping categories such as “young students” and “undergraduate students” would not be mutually exclusive because one student could belong to both.
Strata Must Be Collectively Exhaustive
Every eligible population unit must be classifiable into one of the strata.
A region-based design that includes north, south, and west but omits the east would not be exhaustive.
Membership Must Be Known Before Selection
Researchers need sufficiently accurate information to place frame units into strata before drawing the sample. If group membership is learned only after data collection, the procedure is not pre-stratified random sampling.
Within-Stratum Selection Must Be Random
Creating subgroups does not by itself make a sample random. The researcher must use simple random sampling, systematic random sampling with a random start, or another valid probability mechanism within each stratum.
When Should Stratified Random Sampling Be Used?
Stratified random sampling is suitable when the population contains identifiable subgroups that matter to the research question or sampling efficiency.
It is particularly useful when:
- Researchers must report results separately for several subgroups.
- A small group might receive too few observations in a simple random sample.
- The population is heterogeneous overall but relatively homogeneous within appropriate strata.
- Administrative lists already contain accurate stratification information.
- Collection costs or response rates differ between groups.
- The research requires controlled representation by region, school type, organization size, or another characteristic.
- Researchers have prior information about within-stratum variability.
For example, a national educational survey may stratify schools by region and school sector to ensure that public, private, urban, and rural institutions are included.
When Is Stratified Sampling Not Appropriate?
It may be unnecessary or impractical when:
- The population is small enough to conduct a census.
- No reliable sampling frame exists.
- Stratum membership is unknown or inaccurate.
- The proposed strata have little relationship to the outcomes.
- There are too many small cross-classified cells.
- Researchers cannot perform random selection within the strata.
- The study is exploratory or qualitative and does not seek probability-based population inference.
- The added administrative complexity offers little statistical benefit.
When researchers recruit whoever is available within each category, the result is not stratified random sampling. It is more accurately described as a stratified convenience or quota-based nonprobability sample.
Types of Stratified Random Sampling
Proportionate Stratified Random Sampling
In proportionate allocation, each stratum receives the same share of the sample that it contributes to the population.
The formula is:
[
n_h = n\left(\frac{N_h}{N}\right)
]
where:
- (n_h) = sample size allocated to stratum (h)
- (n) = total sample size
- (N_h) = population size of stratum (h)
- (N) = total population size
If 30% of the population belongs to one stratum, approximately 30% of the sample is allocated to that stratum.
When the same sampling fraction is used in every stratum, the design is self-weighting at the base-weight level because each unit has the same initial probability of selection.
Disproportionate Stratified Random Sampling
In disproportionate allocation, the fraction sampled differs between strata.
Researchers may deliberately oversample a small group to obtain enough observations for subgroup estimation. For example, doctoral students might constitute only 4% of a university but receive 20% of the sample because the study requires a reliable doctoral-student estimate.
This is statistically valid when:
- Selection remains random within each stratum.
- Inclusion probabilities are recorded.
- Appropriate weights are used for population-level analysis.
- The oversampling decision is disclosed.
Equal Allocation
Equal allocation assigns the same sample size to every stratum:
[
n_h = \frac{n}{L}
]
where (L) is the number of strata.
If a population has four strata and the total sample is 400, equal allocation gives 100 observations to every stratum.
Equal allocation is useful for subgroup comparisons, especially when strata differ greatly in population size. However, the resulting sample is normally disproportionate and requires weighting for estimates of the total population.
Neyman Allocation
Neyman allocation assigns more observations to strata that are larger and more variable on the outcome of interest.
When the cost per observation is similar across strata:
[
n_h =
n
\frac{N_hS_h}
{\sum_{j=1}^{L}N_jS_j}
]
where (S_h) is the estimated standard deviation of the study variable in stratum (h).
Neyman allocation minimizes the sampling variance of a target mean or total for a fixed total sample size under its assumptions (Neyman, 1934; Cochran, 1977).
Its main limitation is that the within-stratum standard deviations must be known or estimated from previous data or a pilot study.
Cost-Optimal Allocation
When data-collection costs differ between strata, allocation can account for population size, variability, and cost.
A commonly used form is:
[
n_h \propto \frac{N_hS_h}{\sqrt{c_h}}
]
where (c_h) is the cost of collecting one observation from stratum (h).
This allocates relatively more observations to strata that are large, variable, and less expensive to sample.
Comparison of Allocation Methods
| Allocation method | Main rule | Best suited to | Weighting needed for overall estimates? |
|---|---|---|---|
| Proportionate | Sample follows population proportions | Straightforward population estimates | Base weights are equal when selection is otherwise equal probability |
| Equal | Same number from every stratum | Direct subgroup comparison | Usually yes |
| Disproportionate | Researcher sets unequal sampling rates | Oversampling small or important groups | Usually yes |
| Neyman | More sample to large and variable strata | Maximizing precision for a fixed sample | Usually yes unless it happens to be proportionate |
| Cost-optimal | Accounts for size, variability, and cost | Surveys with major cost differences | Usually yes |
Main Stratified Sampling Formulas
Let:
- (L) = number of strata
- (N_h) = population size in stratum (h)
- (n_h) = sample size in stratum (h)
- (N = \sum_{h=1}^{L}N_h)
- (n = \sum_{h=1}^{L}n_h)
- (W_h = N_h/N)
- (\bar{y}_h) = sample mean in stratum (h)
Proportionate Allocation Formula
[
n_h = n\frac{N_h}{N}
]
Estimated Population Mean
[
\bar{y}{st} = \sum{h=1}^{L}W_h\bar{y}_h
]
This is a population-share-weighted average of the stratum means.
Estimated Population Total
[
\hat{Y}{st} = \sum{h=1}^{L}N_h\bar{y}_h
]
Estimated Population Proportion
If (\hat{P}_h) is the estimated proportion in stratum (h):
[
\hat{P}{st} = \sum{h=1}^{L}W_h\hat{P}_h
]
Variance of the Estimated Mean
For independent simple random samples without replacement within the strata:
[
\widehat{\operatorname{Var}}(\bar{y}_{st})
\sum_{h=1}^{L}
W_h^2
(1-f_h)
\frac{s_h^2}{n_h}
]
where:
- (s_h^2) is the sample variance within stratum (h).
- (f_h=n_h/N_h) is the stratum sampling fraction.
The estimated standard error is:
[
SE(\bar{y}_{st})
\sqrt{\widehat{\operatorname{Var}}(\bar{y}_{st})}
]
The term (1-f_h) is the finite population correction. It matters most when a substantial fraction of a finite stratum has been sampled.
How to Conduct Stratified Random Sampling
Step 1: Define the Target Population
State exactly who or what is included.
A useful population definition may specify:
- Geographic boundaries
- Eligibility criteria
- Institutional membership
- Observation period
- Inclusion and exclusion rules
Example:
The target population consists of all 5,000 students registered for at least one credit-bearing course at University X during the autumn 2026 semester.
Step 2: Construct or Obtain the Sampling Frame
Obtain a list containing:
- A unique identifier
- Eligibility status
- Stratum membership
- Contact information when required
- Any variables needed for selection
Check the frame for duplicates, missing units, ineligible records, outdated entries, and missing stratum classifications.
Step 3: Select the Stratification Variables
Choose variables connected to the research objectives.
A good stratification variable should be:
- Known for nearly all frame units
- Measured accurately
- Relevant to important outcomes or subgroup reporting
- Stable during sample selection
- Practical to administer
Avoid creating dozens of cells from every available demographic characteristic. Cross-classifying four age groups, three income bands, four regions, and three education categories would create 144 strata, many of which might be too small.
Step 4: Define the Strata
Create non-overlapping classification rules.
For example:
- Arts and humanities
- Business
- Engineering
- Health sciences
Document how ambiguous or missing records are handled.
Step 5: Determine the Total Sample Size
The overall sample size should reflect:
- Desired precision
- Confidence level
- Expected variability
- Population size
- Planned subgroup estimates
- Expected response rate
- Design effect
- Available budget
A generic simple-random-sample calculator can provide an initial value, but the final stratified design must also consider minimum subgroup sample sizes and unequal allocation.
If the study requires 100 completed responses from a stratum and its anticipated response rate is 60%, the number invited from that stratum should be inflated:
[
n_{invited,h} = \frac{n_{completed,h}}{r_h}
]
[
n_{invited,h} = \frac{100}{0.60} \approx 167
]
Step 6: Allocate the Sample Across Strata
Choose proportionate, equal, disproportionate, Neyman, or cost-optimal allocation.
The allocation method should follow the study’s main objective:
- Use proportionate allocation for a straightforward population-representative design.
- Use equal or disproportionate allocation for subgroup comparisons.
- Use Neyman allocation when prior variance information is reliable and overall precision is the priority.
- Use cost-optimal allocation when collection costs differ substantially.
Step 7: Resolve Decimal Allocations
Proportionate calculations often produce decimals, but sample sizes must be whole numbers.
A defensible procedure is the largest remainder method:
- Calculate the unrounded allocation for every stratum.
- Round each value down.
- Count how many sample positions remain.
- Assign the remaining positions to strata with the largest decimal remainders.
- Confirm that the final allocations sum exactly to (n).
The final value for a stratum must not exceed its population size. Very small strata may be designated take-all strata, meaning every eligible unit is selected.
Step 8: Randomly Select Units Within Every Stratum
Possible selection methods include:
- Random-number generators
- Statistical software
- Reproducible scripts
- Systematic selection with a random start
- Secure survey-sampling systems
Use unique identifiers rather than names where possible. Record the random seed, software version, date, allocation, and selection code.
Step 9: Verify the Selected Sample
Confirm that:
- Every selected unit is eligible.
- No duplicate was selected unintentionally.
- Stratum counts match the allocation.
- Selection probabilities are recorded.
- Reserve or replacement procedures were specified before fieldwork.
Researchers should not casually replace a nonrespondent with a convenient participant. Replacement can change inclusion probabilities and introduce bias.
Step 10: Collect and Analyze the Data
Track responses separately by stratum. Calculate final weights when sampling fractions or response rates differ. Use survey-analysis procedures that account for strata, weights, clustering, and finite population corrections where applicable.
Worked Example of Proportionate Stratified Sampling
A university has 5,000 students in four faculties:
| Faculty | Population size (N_h) | Population share |
|---|---|---|
| Arts and humanities | 1,750 | 35% |
| Business | 1,250 | 25% |
| Engineering | 1,000 | 20% |
| Health sciences | 1,000 | 20% |
| Total | 5,000 | 100% |
The researcher wants a sample of 400 students.
The proportionate-allocation formula is:
[
n_h = 400\left(\frac{N_h}{5000}\right)
]
Arts and Humanities
[
n_1 = 400\left(\frac{1750}{5000}\right)=140
]
Business
[
n_2 = 400\left(\frac{1250}{5000}\right)=100
]
Engineering
[
n_3 = 400\left(\frac{1000}{5000}\right)=80
]
Health Sciences
[
n_4 = 400\left(\frac{1000}{5000}\right)=80
]
The final allocation is:
| Faculty | Required sample |
|---|---|
| Arts and humanities | 140 |
| Business | 100 |
| Engineering | 80 |
| Health sciences | 80 |
| Total | 400 |
The researcher creates a separate list for each faculty and randomly selects the stated number of students from every list.
Estimating a Mean from the Worked Example
Suppose the selected students report the following average weekly study hours:
| Faculty | Population weight (W_h) | Sample mean |
|---|---|---|
| Arts and humanities | 0.35 | 12.4 |
| Business | 0.25 | 11.8 |
| Engineering | 0.20 | 14.6 |
| Health sciences | 0.20 | 15.2 |
The estimated population mean is:
[
\bar{y}_{st}
(0.35)(12.4)
+
(0.25)(11.8)
+
(0.20)(14.6)
+
(0.20)(15.2)
]
[
\bar{y}_{st}=13.25
]
The estimated average is 13.25 study hours per week.
Because the allocation was proportionate, the same result would be obtained by calculating the ordinary mean across all 400 observations, provided there were no subsequent differences in nonresponse or weighting adjustments.
Example of Disproportionate Allocation and Weighting
Suppose the researcher instead selects 100 students from each faculty. This produces equal subgroup sample sizes, but it does not mirror the population distribution.
The base design weights are:
| Faculty | Population (N_h) | Sample (n_h) | Base weight (N_h/n_h) |
|---|---|---|---|
| Arts and humanities | 1,750 | 100 | 17.5 |
| Business | 1,250 | 100 | 12.5 |
| Engineering | 1,000 | 100 | 10.0 |
| Health sciences | 1,000 | 100 | 10.0 |
Each sampled arts student represents 17.5 students in the frame, while each sampled engineering or health-sciences student represents 10.
Using the four stratum means from the earlier example, an unweighted average of the means would be:
[
\frac{12.4+11.8+14.6+15.2}{4}=13.5
]
However, the population-share-weighted result is:
[
13.25
]
The unweighted result gives too much influence to the smaller faculties. This demonstrates why disproportionate samples require weighting for population-level estimates.
How to Analyze Stratified Samples
Use the Population Weights
For a stratified mean:
[
\bar{y}_{st}=\sum W_h\bar{y}_h
]
Do not simply average the stratum means unless the strata are equal in population size or the research question intentionally gives every stratum equal importance.
Account for Unequal Selection Probabilities
Start with the base weight:
[
d_{hi}=\frac{1}{\pi_{hi}}
]
Final survey weights may also include:
- Eligibility adjustments
- Nonresponse adjustments
- Trimming
- Calibration
- Raking
- Post-stratification
Use Design-Based Standard Errors
Standard statistical procedures often assume independent and identically distributed observations from a simple random sample. A stratified or multistage design may require survey-specific variance estimation.
Researchers should provide the analysis software with available information about:
- Strata
- Sampling weights
- Primary sampling units
- Finite population corrections
- Replicate weights
Distinguish Subgroup and Population Questions
An unweighted analysis may be acceptable for a purely descriptive comparison of equally sized sample groups in some limited situations. It is not automatically valid for estimating the overall population.
The analysis method must match the estimand—the precise population quantity being estimated.
Stratified Sampling Compared with Other Methods
| Method | How the population is handled | How units are selected | Main purpose |
|---|---|---|---|
| Simple random sampling | Treated as one population | Randomly from the full frame | Simplicity and equal-probability selection |
| Stratified random sampling | Divided into strata; sample drawn from every stratum | Randomly within every stratum | Subgroup representation and possible precision gains |
| Cluster sampling | Divided into clusters; only selected clusters are studied | Clusters are sampled, followed by all or some units within them | Lower travel and administrative costs |
| Systematic sampling | Ordered as one list or within strata | Random start followed by every (k)th unit | Efficient selection from an ordered frame |
| Quota sampling | Divided into quota categories | Normally nonrandom recruitment until quotas are filled | Fast control of sample composition |
Stratified Sampling Versus Simple Random Sampling
A simple random sample selects directly from the complete frame. It does not guarantee a particular number from each subgroup.
A stratified random sample first divides the frame and controls how many selections are made within each group.
Stratification is preferable when subgroup representation or subgroup precision is important. Simple random sampling may be preferable when the population is relatively homogeneous, subgroup estimates are not required, and administrative simplicity matters most.
Stratified Sampling Versus Cluster Sampling
The most useful distinction is:
Stratified sampling selects some units from every stratum, whereas cluster sampling selects some clusters and may leave other clusters entirely unobserved.
Strata are commonly constructed to be relatively homogeneous internally on variables related to the outcome. Clusters are often natural administrative or geographic groups that contain diverse units.
For example:
- Stratified design: sample students from every school type.
- Cluster design: randomly select several schools and survey students within only those schools.
Real surveys can combine both methods. A national study might stratify regions, select districts as clusters within each region, and then sample households within selected districts.
Stratified Random Sampling Versus Quota Sampling
Both methods control subgroup numbers, but their selection mechanisms differ.
In stratified random sampling:
- A sampling frame is normally available.
- Selection probabilities are known or calculable.
- Units are randomly selected within strata.
- Design-based statistical inference is possible.
In quota sampling:
- Interviewers or platforms recruit available people until category targets are filled.
- Individual selection probabilities are normally unknown.
- Convenience and self-selection can influence participation.
- Probability-based margins of sampling error are generally not justified by the quota structure alone.
Stratified Sampling Versus Post-Stratification
Stratified sampling is a design-stage procedure. Population units are divided before sample selection.
Post-stratification is an analysis-stage weighting adjustment. Researchers group sampled observations after selection and adjust weights so that weighted totals match known population totals.
Post-stratification can improve alignment with known distributions, but it does not transform a poorly recruited nonprobability sample into a true stratified random sample.
Advantages of Stratified Random Sampling
Guaranteed Representation of Selected Subgroups
Every stratum contributes observations. This prevents a small but important group from being omitted by chance.
More Reliable Subgroup Estimates
Researchers can allocate enough observations to estimate means, proportions, or relationships separately within important domains.
Potentially Greater Precision
Stratification can reduce sampling variance when units within strata are relatively similar and strata differ on variables related to the study outcome (Cochran, 1977; Lohr, 2021).
The gain is not automatic. It depends on the quality of the stratification and allocation.
Flexible Allocation
Researchers can balance several objectives:
- Overall population precision
- Subgroup precision
- Fieldwork cost
- Policy-reporting requirements
- Rare-group analysis
Administrative Benefits
Strata may correspond to regions, institutions, or operational units, making sample management and fieldwork easier.
Transparent Control of Sample Composition
The planned number selected from each group is known before data collection and can be verified.
Limitations of Stratified Random Sampling
It Requires Population Information
The researcher must know enough about frame units to assign them to strata before selection.
Frame Errors Can Remain
Stratification does not correct an outdated, duplicated, incomplete, or inaccurate frame.
Design and Analysis Are More Complex
Researchers must manage separate frames or classifications, calculate allocation, preserve selection probabilities, and use appropriate weights and variance procedures.
Poor Strata May Provide Little Benefit
A stratification variable unrelated to the study outcomes may not improve precision. Excessive stratification can create many tiny cells.
Misclassification Can Damage the Design
Incorrect stratum labels can distort allocation, weights, and subgroup estimates.
Oversampling Can Be Misinterpreted
Disproportionately sampled data may appear unrepresentative when examined as raw counts. The design is valid, but weighted and unweighted results answer different questions.
Nonresponse Can Disrupt the Planned Design
Even a proportionate selected sample may become disproportionate if response rates differ by stratum. Nonresponse adjustment and careful follow-up may be necessary.
Small Strata Create Practical Problems
A calculated allocation may be too small for variance estimation or too large relative to the stratum. Researchers may need minimum allocations, certainty selection, collapsing rules, or specialized variance procedures.
Common Mistakes
Mistake 1: Calling Any Grouped Sample Stratified Random Sampling
Recording age, gender, or department after recruiting a convenience sample does not make the design stratified random sampling.
Correction: Define the strata and select randomly within them before data collection.
Mistake 2: Using Overlapping Strata
A participant cannot simultaneously belong to two strata under one stratification scheme.
Correction: Use mutually exclusive categories or create a single cross-classified variable.
Mistake 3: Omitting Part of the Population
Strata must collectively cover the eligible population.
Correction: Include a clearly defined residual category when necessary and investigate unclassified records.
Mistake 4: Selecting Convenient Participants Within Each Stratum
Convenience recruitment destroys the probability basis of the design.
Correction: Use a verifiable random or probability-based selection procedure.
Mistake 5: Ignoring Weights After Oversampling
Raw averages from disproportionate samples can give excessive influence to oversampled groups.
Correction: Use inverse-probability or final survey weights for population inference.
Mistake 6: Confusing Strata with Clusters
Sampling entire groups is generally cluster sampling, not stratified sampling.
Correction: Remember that a stratified design samples from every stratum.
Mistake 7: Rounding Each Allocation Independently
Independent rounding can make the stratum allocations sum to more or less than the required total.
Correction: Use a controlled rounding method such as largest remainder.
Mistake 8: Creating Too Many Strata
Excessive cross-classification creates sparse cells, unstable estimates, and operational complexity.
Correction: Prioritize variables related to outcomes and essential reporting domains.
Mistake 9: Assuming Stratification Removes All Bias
Random selection addresses selection at the sampling stage, but it does not eliminate nonresponse, coverage, recall, interviewer, or measurement error.
Correction: Evaluate the full survey-error process.
Mistake 10: Reporting the Method Without Enough Detail
The label “stratified random sampling” is insufficient for replication.
Correction: Report the strata, frame, allocation, selection procedure, sampling fractions, response, weights, and analysis method.
Stratified Random Sampling in Modern Research
Stratified designs remain important in:
- Government household surveys
- Public-health surveillance
- Educational assessment
- Election and public-opinion research
- Business and establishment surveys
- Environmental monitoring
- Clinical and epidemiological research
- Program evaluation
- Online research panels
- Machine-learning model evaluation
Small but policy-relevant groups are often oversampled so that separate estimates can be produced.
A recent U.S. Census Bureau working paper described the use of linked survey and administrative data, together with machine-learning methods, to construct sampling flags for households participating in or eligible for nutrition-assistance programs. The purpose was to improve the targeting of small subpopulations that would otherwise be difficult to sample in sufficient numbers (Eggleston et al., 2025).
This illustrates an important modern development: computational models can help identify efficient strata, but the eventual sample design still requires documented probabilities, verification, weighting, and evaluation.
Digital Tools for Stratified Sampling
Spreadsheets
A spreadsheet can be used for a small study by:
- Adding a stratum column.
- Assigning a random number to every record.
- Sorting within each stratum.
- Selecting the first (n_h) records.
- Saving the original frame, seed-equivalent information, and selected IDs.
Spreadsheet recalculation can change random values, so researchers should convert generated random values to fixed values before selection and preserve an audit copy.
R
R supports both sample selection and complex survey analysis.
The survey package can represent stratification, weights, clusters, finite population corrections, post-stratification, calibration, and design-based variance estimation.
A simplified design declaration may resemble:
library(survey)
design <- svydesign(
ids = ~1,
strata = ~faculty,
weights = ~design_weight,
data = sample_data
)
svymean(~study_hours, design)
Python
For a straightforward frame, pandas can draw observations within groups:
sample = (
frame.groupby("faculty", group_keys=False)
.sample(frac=0.08, random_state=2026)
)
A fixed random_state supports reproducibility. For unequal stratum sample sizes, calculate each (n_h) first and sample each group using the required count.
Python can perform selection efficiently, but researchers must still calculate and preserve inclusion probabilities and use appropriate survey-analysis procedures.
Stata
Stata’s survey commands allow users to declare weights, strata, and sampling units before running means, proportions, regressions, and other models.
A simplified declaration is:
svyset [pweight=final_weight], strata(stratum)
svy: mean study_hours
Machine-Learning Stratification Is Not the Same as Survey Sampling
Machine-learning libraries often offer a stratify option for dividing a dataset into training and testing subsets. This preserves class proportions within an existing dataset.
It does not make the original dataset representative of a wider target population. It solves a data-splitting problem rather than establishing a probability sample from a population.
Appropriate Uses of Artificial Intelligence
AI can assist with:
- Identifying duplicate or inconsistent frame records
- Classifying records into proposed strata
- Predicting response propensity
- Exploring alternative allocation scenarios
- Generating or checking selection code
- Documenting variable definitions
- Detecting unexpectedly small strata
- Comparing weighted and unweighted distributions
AI should not be trusted to make undocumented sampling decisions. Researchers should verify:
- The source and accuracy of the frame data
- Classification error across population groups
- Privacy and data-protection requirements
- Whether protected characteristics are being used lawfully and ethically
- The reproducibility of the allocation and random selection
- The final probabilities and weights
Human methodological oversight remains necessary.
How to Report Stratified Random Sampling in a Research Paper
A complete report should state:
- The target population.
- The source and date of the sampling frame.
- The variables used to create the strata.
- The population size of every stratum.
- The total planned sample size.
- The allocation rule.
- The sample size selected from every stratum.
- The random-selection procedure and software.
- The response rate by stratum.
- The calculation and adjustment of weights.
- The variance-estimation method.
- Important frame, coverage, and nonresponse limitations.
Methodology-Section Template
The target population consisted of [population definition]. The sampling frame was obtained from [source] and contained [N] eligible units. The frame was stratified by [variable] into [L] mutually exclusive strata: [list strata]. A total sample of [n] was allocated using [proportionate/equal/Neyman/other] allocation. Within each stratum, units were selected by [simple random/systematic] sampling using [software and version]. The resulting stratum sample sizes were [list]. Base weights were calculated as the inverse of each unit’s probability of selection. The analysis incorporated [weights, strata, finite population corrections, or other design features]. The main limitations were [frame coverage, nonresponse, classification, or other limitations].
Edit the template to reflect what was actually done. Do not claim random selection when participants were recruited through convenience, volunteer, or snowball procedures.
Practical Design Checklist
Before selecting the sample, confirm that:
- The target population has been defined.
- The frame adequately covers that population.
- Every eligible unit has one stratum classification.
- The strata are relevant to outcomes or reporting needs.
- Minimum subgroup precision has been considered.
- The allocation rule has been justified.
- Decimal allocations sum to the intended total.
- Random selection can be reproduced.
- Inclusion probabilities will be stored.
- A nonresponse plan exists.
- The analysis software can handle the design.
- Reporting will include weights and limitations.
Conclusion
Stratified random sampling is most useful when researchers need dependable representation of important population subgroups or want to improve statistical precision through well-chosen strata. Its value depends on more than dividing a list into categories: the frame must be adequate, selection within every stratum must be random, allocation must match the study objectives, and unequal sampling probabilities must be handled correctly during analysis.
References
Foundational academic references in APA 7th style
- Cochran, W. G. (1977). Sampling techniques (3rd ed.). Wiley. The third edition and publication details are confirmed by Wiley and bibliographic records.
- Kish, L. (1965). Survey sampling. John Wiley & Sons.
- Lohr, S. L. (2021). Sampling: Design and analysis (3rd ed.). CRC Press. The third edition was published in November 2021.
- Neyman, J. (1934). On the two different aspects of the representative method: The method of stratified sampling and the method of purposive selection. Journal of the Royal Statistical Society, 97(4), 558–606. https://doi.org/10.1111/j.2397-2335.1934.tb04184.x
University and statistical-agency sources
- Pennsylvania State University. (n.d.). Stratified sampling: STAT 506—Sampling theory and methods. This lesson provides estimator, variance, proportional-allocation, Neyman-allocation, and cost-allocation formulas.
- Statistics Canada. (2019). Section 2: Stratified sampling. It identifies stratum definition, sample allocation, sample size, stratification variables, and within-stratum selection as factors affecting design efficiency.
- U.S. Census Bureau. (2021). Statistical Quality Standard A3: Developing and implementing a sample design. This source is particularly valuable for frame construction, allocation, probabilities, weights, verification, documentation, and reproducibility.
- Eggleston, J., McBride, L., & Klee, M. (2025). The design of sampling strata for the National Household Food Acquisition and Purchase Survey (CES Working Paper No. 25-13). U.S. Census Bureau. It provides a modern example of machine learning and administrative data supporting the identification of small target populations.
Software documentation
- Lumley, T., Gao, P., Schneider, B., & Kolenkikov, S. (2026). survey: Analysis of complex survey samples [R package]. CRAN. The package supports stratification, unequal weights, clusters, post-stratification, calibration, and design-based variance estimation.
- pandas development team. (2026). DataFrameGroupBy.sample. The function can randomly sample records from groups and preserve a reproducible random state.
- scikit-learn developers. (2026). train_test_split. The documentation confirms that its
stratifyargument performs class-based dataset splitting; this should not be confused with drawing a probability sample from a target population. - StataCorp. (2026). Survey features. Stata permits researchers to declare weights, strata, and sampling-unit identifiers for subsequent survey estimation.
