Course contentsShow
← Data Science
Lesson 832 of 2,14521. Survival AnalysisFree lesson

Stratified Cox Models

Handle violations of proportional hazards by stratifying on variables with non-proportional effects.

Stratified Cox Models

What you'll learn: How to handle violations of the proportional hazards assumption by stratifying on variables with non-proportional effects.

The Problem: Non-Proportional Hazards

In the previous lesson, you learned how to test whether the proportional hazards assumption holds. But what happens when a variable fails that test? If the hazard ratio between groups changes over time, including that variable normally in your Cox model will produce biased results.

The Solution: Stratification

Stratified Cox models allow you to account for a variable's effect on survival without assuming proportional hazards for that variable. Instead of estimating a coefficient for the problematic variable, you let each level of that variable have its own baseline hazard function.

Think of it like analyzing separate survival studies for each group (e.g., males vs. females), but still estimating common effects for your other predictors across all strata.

How It Works

When you stratify on a variable (say, treatment center), you're telling the model:

  • "Treatment centers may have fundamentally different baseline risk patterns over time"
  • "But age, treatment type, and other predictors have the same hazard ratios across all centers"

This is like running parallel analyses where the starting risk differs, but the relative effects of other variables remain consistent.

When to Stratify

Stratify when:

  • A variable violates proportional hazards (detected via tests you learned earlier)
  • You want to control for that variable's effect but don't need to interpret its coefficient
  • The variable has relatively few levels (too many strata reduce statistical power)

Key Takeaway: Stratified Cox models let you handle non-proportional hazards by allowing different baseline hazards for each stratum while still estimating common covariate effects across groups—preserving your ability to analyze other predictors correctly.