Course contentsShow
← Data Science
Lesson 830 of 2,14521. Survival AnalysisFree lesson

Testing Coefficient Significance

Use Wald tests, likelihood ratio tests, and score tests to assess whether covariates significantly affect survival.

Testing Coefficient Significance

What you'll learn: How to determine whether a predictor truly influences survival time in your Cox model using statistical tests.

Why Test Coefficients?

After fitting a Cox proportional hazards model, you've estimated coefficients for each predictor. But do these variables actually matter? Testing coefficient significance helps you decide whether a covariate genuinely affects survival or if its apparent effect could just be random noise.

Three Main Testing Approaches

1. Wald Test

Think of this like a t-test for your coefficient. It divides the coefficient estimate by its standard error to produce a z-score. If this score is large (typically |z| > 1.96 for 95% confidence), the predictor is significant. The Wald test is fast and appears in most Cox model output tables alongside p-values.

2. Likelihood Ratio Test

This compares two models: one with your predictor and one without it. If including the predictor substantially improves the model's likelihood (how well it fits the data), the variable is significant. This test is generally more reliable than the Wald test, especially with smaller samples.

3. Score Test

Also called the "log-rank test for covariates," this evaluates whether the gradient of the likelihood is significantly different from zero at the null hypothesis. It's computed without fully fitting the alternative model, making it computationally efficient.

Which Test Should You Use?

All three tests will usually agree when sample sizes are large. The likelihood ratio test is considered the gold standard for accuracy. The Wald test is most convenient since it's automatically provided. The score test is useful for initial screening.

Key Takeaway: Use Wald, likelihood ratio, or score tests to determine if covariates significantly affect survival; all three assess whether a predictor's coefficient is meaningfully different from zero, helping you build parsimonious and interpretable survival models.