10Min Research - 34. Understanding and Performing Stratified Random Sampling in Social Sciences

Stratified random sampling is designed to prevent a common failure of simple random sampling: when a population contains distinct subgroups, random draws can over-represent large groups and under-represent—or entirely miss—small ones. In social science research, that matters because each subgroup may contribute different perspectives that affect conclusions. The core idea is to split the population into strata (singular: stratum)—for example, Bachelor, Master, and PhD students—and then run simple random sampling within each stratum so every group appears in the final sample.

A higher-education example makes the risk concrete. Suppose a university has 1,550 students total: 1,000 Bachelor students, 500 Master students, and 50 PhD students. If the study needs a sample of 300 people and sampling is done purely at random from the full list, the sample can easily end up dominated by Bachelor and Master students, with a real chance of including zero PhD students. That would be methodologically wrong for studies where PhD input is important for understanding higher-education practices and procedures.

To fix this, the population is divided into three strata, and sampling is performed separately inside each one. The first approach is proportionate stratified random sampling, where each stratum’s share of the sample matches its share of the population. With a sample size of 300 from a population of 1,550, the overall sampling fraction is 300/1,550 ≈ 19.35%. Applying that fraction to each stratum yields target counts of about 193.5 from each group if the same percentage were used directly in the calculation shown; the practical takeaway is that the sample size allocated to each stratum is computed from the stratum’s population size times the overall sampling fraction (expressed as a percentage).

Because exact targets can be awkward and because researchers want to ensure minimum representation, the method can be adjusted by increasing the number of people contacted in each stratum. The transcript illustrates this by moving from the strict proportionate targets to a larger set of contacted counts (e.g., contacting more Bachelor and Master students and a higher number of PhD students) to help guarantee that the final number of responses meets the minimum required.

A second approach is disproportionate stratified random sampling, where the allocation across strata is intentionally changed to secure stronger representation of smaller or more important groups. Instead of using the same proportional allocation, the researcher increases the PhD stratum share (and adjusts others downward) to avoid the “too few PhD responses” problem.

Operationally, each stratum becomes its own sampling frame. With a list of all Bachelor students (say 1,000 entries in an Excel sheet), the researcher uses a random number generator to select the required number of individuals (e.g., 300) by choosing random indices within the allowed range (minimum 1, maximum 1,000). The same process is repeated for Master and PhD lists. The method depends on having access to the full population frame for each stratum; without the ability to identify and select every element, neither simple random sampling nor stratified random sampling can be carried out properly.