A stratified random sample is a method of selecting participants (or data points) by first dividing the full population into smaller subgroups based on shared characteristics, then randomly selecting from each subgroup separately. This ensures every important segment of a population is represented in the final sample, rather than leaving representation to chance.
How It Works
The core idea is simple: before you start picking people at random, you sort the entire population into categories called “strata.” These strata are based on a characteristic relevant to what you’re studying. Age brackets, income levels, geographic regions, and gender are common examples. The key rule is that the strata must be mutually exclusive, meaning every person in the population belongs to exactly one group. No overlaps.
Once the population is divided, you draw a random sample from within each stratum independently. This is what separates stratified sampling from plain random sampling. In a simple random sample, you pull names from the entire population at once, which means some subgroups could end up overrepresented or underrepresented purely by luck. Stratified sampling eliminates that problem by guaranteeing each subgroup contributes members to the final sample.
Proportional vs. Disproportional Sampling
There are two main ways to decide how many people to draw from each stratum.
In proportional allocation, you sample each stratum in proportion to its size in the overall population. If 60% of your population is urban and 40% is rural, and you need a sample of 500, you’d draw 300 urban residents and 200 rural residents. The formula is straightforward: the sample size for each stratum equals the total sample size multiplied by that stratum’s share of the population. This is the most common approach and keeps the math simple.
In disproportional (or “optimal”) allocation, you deliberately oversample smaller or more variable groups. This is useful when a subgroup is so small that proportional sampling wouldn’t capture enough members to draw meaningful conclusions. A study of liver cancer screening published in BMC Medical Research Methodology used this approach with electronic health records: researchers allocated 90% of rural residents but only 10% of urban residents to their exploratory sample, guaranteeing at least 50 people in each subgroup defined by race, ethnicity, and rurality. Without that deliberate oversampling, rural minority groups would have been too small to analyze.
Choosing the Right Strata
The variable you use to create strata should have a meaningful relationship to whatever you’re measuring. If you’re studying exercise habits, stratifying by age group makes sense because exercise patterns differ sharply across age. Stratifying by eye color would not, because it has no plausible connection to how much people exercise. Poor strata choices add complexity without improving accuracy.
Researchers also need to be selective about how many stratifying variables they use. Each additional variable multiplies the number of subgroups. Stratifying by gender (2 groups) and age bracket (4 groups) creates 8 strata. Add income level (3 groups) and you’re at 24. More strata means you need a larger overall sample to keep each subgroup big enough to be useful, which drives up cost and effort. The general principle is to stratify on a small number of factors that have the largest influence on the outcome you care about.
Why Use It Instead of Simple Random Sampling
The main advantage is precision. By ensuring each subgroup is properly represented, stratified sampling typically produces estimates with less error than a simple random sample of the same size. This is especially valuable when the population contains subgroups that behave very differently from one another. Polling firms use stratified sampling to make sure their election surveys include the right mix of voters by region, age, and education, because voting patterns vary dramatically across those dimensions.
Stratified sampling also lets researchers make reliable comparisons between subgroups. If you want to compare health outcomes for different racial or ethnic groups, a simple random sample might give you 400 people from the majority group and only 15 from a smaller minority group. Those 15 aren’t enough to draw conclusions. Stratified sampling solves this by setting a minimum sample size for each group from the start.
A Practical Example
Suppose a university wants to survey student satisfaction across its four colleges: Engineering (2,000 students), Business (3,000), Arts (4,000), and Science (1,000). The total population is 10,000 students, and the budget allows for surveying 500.
With proportional allocation, you’d survey 100 engineering students (20% of the population), 150 business students (30%), 200 arts students (40%), and 50 science students (10%). Within each college, those individuals are selected at random. Every student has a chance of being chosen, but the final sample mirrors the actual composition of the university.
If the university specifically wanted to compare satisfaction across colleges, they might use disproportional allocation and sample 125 from each college instead. This gives the smaller Science college enough respondents for a meaningful comparison, even though it overrepresents them relative to their population share.
Limitations Worth Knowing
Stratified sampling requires information about the population before you begin. You need to know which stratum each person belongs to, which means you need a complete list of the population along with the characteristic you’re stratifying on. For some populations, this information simply isn’t available.
It also adds logistical complexity. Managing separate random draws for each stratum, tracking response rates by group, and potentially weighting the results afterward all take more planning than pulling a single random sample. When strata are disproportionally sampled, the final results need to be weighted back to reflect the true population, adding another layer of analysis.
If the stratifying variable turns out to have little relationship to the outcome being measured, the extra effort doesn’t improve accuracy. In that case, a simple random sample would have produced equally good results with less work.