The proportional hazards assumption

Survival Analysis in Python

Shae Wang

Senior Data Scientist

Use the Kaplan-Meier curves

If the covariate only has a few values, inspect each group's Kaplan-Meier curve.

Curves intersect: fails the proportional hazards assumption.

failed ph assumption due to intersecting curves

Use the Kaplan-Meier curves

If the covariate only has a few values, inspect each group's Kaplan-Meier curve.

Curves have different shapes: fails the proportional hazards assumption.

failed ph assumption due to intersecting curves

Use the Kaplan-Meier curves

If the covariate only has a few values, inspect each group's Kaplan-Meier curve.

Curves have similar shapes and are parallel: satisfies the proportional hazards assumption.

successful proportional hazard assumption km curve

.check_assumptions()

If the covariates are continuous, use the .check_assumptions() method.

Parameters
- training_df: the original DataFrame used in the call to fit the model.
- p_value_threshold: the threshold to use to alert the user of violations (default: 0.01, recommended: 0.05).

.check_assumptions()

model.check_assumptions(training_df, p_value_threshold=0.05)

1. Variable 'A' failed the non-proportional test: p-value is 0.0007.
Advice 1: ...
Advice 2: ...
2. Variable 'B' failed the non-proportional test: p-value is 0.0063.
Advice 1: ...
Advice 2: ...

When the proportional hazards assumption fails

Usually, it's a reasonable assumption and violations do not impact model performance significantly.
If it fails, try other modeling frameworks, such as the Weibull AFT model, and compare their AIC scores.

Let's practice!

Survival Analysis in Python