Testing the Proportional Hazards Assumption
What You'll Learn
You'll learn how to verify whether your Cox model's key assumption—that hazard ratios stay constant over time—actually holds for your data.
Why This Matters
Remember from lesson 826 that the Cox model assumes proportional hazards: the effect of a predictor (its hazard ratio) doesn't change over time. But what if age matters more early on, or treatment effects wear off? You need to test this assumption, not just hope it's true.
Three Main Approaches
1. Schoenfeld Residuals
These special residuals check whether each predictor's effect stays constant. You:
- Calculate residuals for each event time
- Test if they correlate with time
- A significant correlation means the proportional hazards assumption is violated for that variable
Think of it like checking if a player's advantage stays the same throughout a game, or if it changes as time passes.
2. Log-Log Plots
Plot log(-log(survival)) versus log(time) for different groups:
- If hazards are proportional, lines should be roughly parallel
- Crossing or converging lines suggest violation
- Visual, intuitive, but somewhat subjective
3. Time-Dependent Covariates
Include interaction terms between predictors and time (e.g., age × log(time)):
- If the interaction is significant, the effect changes over time
- This directly tests whether hazard ratios shift
When to Worry
Small violations? Often acceptable. Large violations mean you need to stratify by that variable or use time-dependent models instead.
Key Takeaway: Always test proportional hazards using Schoenfeld residuals, log-log plots, or time interactions—never assume the assumption holds without checking.