1

I am having trouble linking these concepts together:

  1. Here is a Cox PH (Proportional Hazards) Regression Model:

$$h(t|X) = h_0(t) \exp(\beta^T X)$$

where:

  • $h(t|X)$ is the hazard function for an individual with covariate values $X$.
  • $h_0(t)$ is the baseline hazard function, representing the hazard for an individual with $X=0$. -$\beta$ is a vector of regression coefficients.
  1. Here is a Martingale:

A martingale $M_t$ is a stochastic process satisfying the following property:

$$E[M_{t+1} | M_1, M_2, ..., M_t] = M_t$$

  1. Here is the Proportional Hazards assumption for a Cox PH model:

The hazard ratio between any two individuals is constant over time:

$$\frac{h(t|X_1)}{h(t|X_2)} = \exp(\beta^T (X_1 - X_2))$$

where $X_1$ and $X_2$ are two sets of covariate values, and $\beta$ is the vector of regression coefficients. This implies that the effect of covariates on the hazard rate is constant over time.

But what I don't understand is that how/why are martingales useful at validating the proportional hazards assumption in Cox PH?

The only thing which comes to mind is that residuals have to be without trends and a martingale encapsulates this idea via the expectation. But this makes me wonder why martingales are not used in basic regression models (e.g. OLS) to validate assumptions.

In Cox-PH, is it possible to mathematically see:

  • why the Proportional Hazards assumption is important (e.g. what happens when it is not met ... perhaps coefficients become biased, large variance, not consistent, etc.)
  • what residual distribution/properties are indicative of the Proportional Hazards assumption (e.g. Expected Residual at time t+1 equals expected residual at time t given all histories)
  • why martingales are useful at validating if the Proportional Hazards assumption is being met?
  • Welcome to Cross Validated! Could you please edit the question to provide a reference that suggests use of martingales to evaluate how well the proportional hazards (PH) assumption is met? That's usually done with scaled Schoenfeld residuals. For what happens when PH doesn't hold, see this page for example; a Cox-model regression coefficient then is a sort of time-averaged log-hazard estimate that still might be useful if combined with a robust "sandwich" type of coefficient covariance matrix. – EdM Mar 13 '24 at 20:04

0 Answers0