On this page
Partition the frame before selection
Strata should be defined so every frame unit belongs to one group under the stated rules. A fictional business survey might use two size classes. Record how borderline firms are classified and whether the group sizes were measured before sampling.
Choose sample counts explicitly
A proportional allocation follows group population counts. A disproportionate allocation can devote more observations to a smaller group. Neither label removes the need to state the within-group selection rule or to handle a group allocated zero observations.
Use the population shares for a population mean
For within-stratum simple random samples, combine stratum sample means with weights N_h/N for the finite population mean. Weighting each group by n_h/n instead answers the sample-composition question when allocation differs from population composition.
| Weight | What it measures |
|---|---|
| N_h/N | Stratum share of population units |
| n_h/n | Stratum share of sampled units |
Evaluate precision with the relevant design
The effect of stratification depends on within-group variation, allocation, costs and the estimator. Do not promise that any chosen partition always improves precision. The next worked example isolates the weighting calculation without making a universal efficiency claim.