Matched Cox Proportional Hazards — Stratified Cox Regression for Matched Designs
Stratified Cox Proportional Hazards Regression for Matched Designs · Also known as: stratified Cox regression, conditional Cox model, matched survival analysis, Cox model for matched pairs
Matched Cox proportional hazards is a survival analysis method that extends the Cox regression model to appropriately handle data arising from matched study designs — matched cohorts or matched case-control studies with time-to-event outcomes. By stratifying the partial likelihood by matched set, the method eliminates confounding from matching factors without estimating their baseline hazard, yielding valid hazard ratio estimates that are free from matching-induced bias.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use matched Cox proportional hazards when you have time-to-event data from a matched cohort or matched case-control study and need to estimate adjusted hazard ratios while properly accounting for the matching structure. It is appropriate when the outcome is a time-to-event variable, matching has been performed to control confounding, and the proportional hazards assumption holds within matched strata. Do not use it when the matched sets are very small and events per stratum are sparse — the partial likelihood loses efficiency in that setting and alternative methods (e.g., conditional logistic for binary endpoints, or frailty models) may be preferable. It is not suited to unmatched data; applying the stratified Cox model without actual matched structure wastes degrees of freedom unnecessarily.
Strengths & limitations
- Correctly accounts for matched-set structure, eliminating bias from matching factors without estimating their effects.
- Retains the semi-parametric flexibility of Cox regression — no distributional assumption for the baseline hazard within each stratum.
- Allows additional covariate adjustment beyond the matching variables, enabling multivariable analysis.
- Hazard ratio estimates are directly interpretable and widely understood by clinical and epidemiological audiences.
- Can handle variable-ratio matching (1:1, 1:2, 1:M) within the same stratified likelihood framework.
- Loses information on the matched confounders themselves — their effects on survival cannot be estimated because they are absorbed into the stratum-specific baseline hazards.
- Efficiency decreases as matched strata become small; strata with only one participant or zero events contribute nothing to the partial likelihood.
- The proportional hazards assumption must hold within each matched stratum, adding complexity to assumption checking in large matched datasets.
- Requires the investigator to correctly identify all matched strata; incorrect stratum assignment invalidates the analysis.
Frequently asked
What is the difference between stratified Cox regression and a frailty model for matched data?
Both address within-cluster correlation in survival data. Stratified Cox regression treats the matched-set effects as nuisance parameters and eliminates them through conditioning, making no distributional assumption about them. Frailty models treat matched-set effects as random variables drawn from a distribution (e.g., gamma or log-normal) and estimate that distribution. Stratified Cox is more robust because it avoids distributional assumptions, but frailty models can be more efficient when the frailty distribution is correctly specified.
Can I use this method with propensity-score matched data?
Yes. Propensity-score matched datasets are a common application. Each matched pair or set forms a stratum, and the stratified Cox model is fitted using matched-set membership as the stratification variable. This properly accounts for the within-pair correlation introduced by the matching and yields valid hazard ratio estimates for the treatment or exposure.
What happens if some matched sets have no events?
Strata without any events contribute nothing to the stratified partial likelihood and are effectively dropped from the estimation. This is expected behaviour and does not bias the estimates, but it does reduce effective sample size and statistical power. If many strata have no events, reconsider the matching ratio or consider whether the event rate is too low for this design.
How do I test the proportional hazards assumption in stratified Cox regression?
The same tools used for standard Cox regression apply within each stratum: Schoenfeld residual tests and log-log survival plots. Most software (R survfit/cox.zph, Stata stphtest) supports stratified models. Interaction terms between covariates and log-time can also be included to test and model non-proportional hazards within strata.
Does matched Cox regression require 1:1 matching, or can it handle variable ratios?
It handles any matching ratio. Each matched set — whether a pair, a triplet, or a 1:M group — forms one stratum. The partial likelihood sums over events within each stratum regardless of stratum size, so variable-ratio and full-matching designs are accommodated without modification.
Sources
- Cox, D. R. (1972). Regression models and life-tables. Journal of the Royal Statistical Society: Series B (Methodological), 34(2), 187–202. DOI: 10.1111/j.2517-6161.1972.tb00899.x ↗
- Thomas, D. C. (1977). Addendum to: Methods of cohort analysis: Appraisal by application to asbestos mining. Journal of the Royal Statistical Society: Series A, 140(4), 483–485. link ↗
How to cite this page
ScholarGate. (2026, June 3). Stratified Cox Proportional Hazards Regression for Matched Designs. ScholarGate. https://scholargate.app/en/epidemiology/matched-cox-proportional-hazards
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Cox proportional hazardsEpidemiology↔ compare
- Kaplan-Meier AnalysisEpidemiology↔ compare
- Matched Cohort StudyEpidemiology↔ compare
- Survival AnalysisResearch Statistics↔ compare