Home Statistical Dictionary Stratification Variable
Stratification Variable
The variable used to divide a population into strata for a stratified sampling design, known for every unit in the sampling frame in advance and chosen because it is expected to be correlated with the survey's key outcomes.
In Plain English
The stratification variable is whatever characteristic a stratified sampling design actually splits the population by, geographic region, age bracket, business industry code, school type, whatever the designer believes will meaningfully group units together on the outcomes the survey cares about. The whole benefit of stratified sampling depends on choosing a good stratification variable, one that's known for every single unit in the sampling frame before any sampling happens, and one that's genuinely correlated with the survey's key outcomes, stratifying by something irrelevant to what you're measuring gains essentially nothing over simple random sampling, while a well-chosen stratification variable can meaningfully sharpen the survey's precision for the same total sample size.
Definition
A stratification variable is the characteristic used to partition a population into non-overlapping strata for a stratified sampling design, it must be known, or at least assignable, for every unit in the sampling frame before sampling begins, since stratified sampling requires allocating a specific sample size to each stratum in advance, unlike post-stratification, which can use a variable measured only after the survey is fielded. The value of a chosen stratification variable is directly tied to how strongly it correlates with the survey's outcome variables, effective stratification variables produce strata that are internally homogeneous (similar units within a stratum) and externally heterogeneous (different strata differ meaningfully from each other) on the outcome of interest, since it is exactly that within-stratum homogeneity that removes between-stratum variance from the sampling error, a stratification variable uncorrelated with the outcomes provides little to no efficiency gain over simple random sampling, despite still guaranteeing representation of each stratum.
Properties
- The precision benefit of stratified sampling is entirely contingent on the chosen stratification variable's correlation with the outcome of interest, a well-chosen variable (like region for a survey of regional economic conditions) can substantially improve precision, while a poorly chosen one (stratifying by a characteristic unrelated to what's being measured) gains essentially nothing over simple random sampling of the same size.
- A single survey commonly stratifies by more than one variable simultaneously (for example, cross-classifying by both region and establishment size in a business survey), producing a finer set of strata, this generally increases precision further when both variables are relevant, at the cost of a more complex allocation and, if carried too far, strata that become too small to sample reliably.
- The requirement that a stratification variable be known in advance for the entire sampling frame is what fundamentally distinguishes true stratification from post-stratification, a variable that can only be measured during data collection itself (like a survey response) cannot serve as a stratification variable, no matter how strongly it correlates with the outcome, it can only be used for post-stratification instead.
At a Glance