Statistics Glossary → Stratified analysis (stratum variable)
Stratified analysis (stratum variable)concept
An approach that controls for a confounding or design variable by running the analysis separately within each of its levels (strata) and then combining the results.
Stratification controls for a confounder (e.g. age group, centre) by making the comparison within each stratum rather than by adding the confounder to the model as a coefficient. With categorical data a separate 2x2 table is built for each stratum, and a common odds ratio and test are obtained with the Cochran-Mantel-Haenszel method. In survival analysis the stratified log-rank test compares groups within strata; the stratified Cox model gives each stratum its own baseline hazard while keeping the covariate coefficients common.
Stratification makes no assumption about the form of the confounder's effect (e.g. linearity), but many strata or sparse strata weaken the estimates. Before combining into a common effect, check whether the effect is similar across strata (homogeneity, e.g. the Breslow-Day test); an effect that changes direction across strata may signal Simpson's paradox. In sampling theory, a stratum is a subgroup into which the population is divided before sampling; in stratified sampling the variance is computed within strata and is usually smaller.
An everyday example
A drug's recovery rate is compared across two hospitals. Hospital A admits severe cases and mostly gives the drug; hospital B treats mild cases and rarely uses it. With all data pooled, the drug seems useless; looked at within each hospital, patients who received it recover more often in both.
In MerQur
Appears as 'Stratum' in the Cochran-Mantel-Haenszel test, 'Layer (control)' in crosstabs, 'Strata (STRATA) variable' in Kaplan-Meier analysis, 'Strata variable (STRATA)' in Cox and time-dependent Cox regression, and 'Baseline-hazard stratum' in Survey-PHREG and the cluster-robust Cox model. In weighted (survey) analyses 'Stratum' is the design stratum, propensity score analysis offers subclassification, and age standardization uses the stratum (age group); stratified resampling can be chosen in the bootstrap CI, and in conditional logit the choice set or matched set acts as the stratum.
Analyses where this concept appears
Bootstrap CICochran-Mantel-Haenszel (CMH) TestCrosstabPropensity Score AnalysisWeighted Mean (Survey Means)Weighted Frequency (Survey Freq)Weighted Total (Survey Total)Weighted Linear Regression (Survey Regression)Weighted Logistic Regression (Survey Logistic Regression)Conditional Logit (Choice Model)Kaplan-Meier Survival AnalysisCox Regression (Proportional Hazards Model)Time-dependent Cox RegressionSurvey-PHREG (Weighted Cox)Cluster-Robust Cox (Marginal and Frailty Models)Age Standardization and SMR