Understanding the Cox Proportional Hazards Model

From Description to Regression: The Need for the Cox Model

When analyzing survival data, we rarely believe that a single factor (like a treatment group) is the only thing influencing the outcome. Age, disease severity, lifestyle factors, and other clinical measurements can all play a role. Trying to analyze these with simple group comparisons becomes impractical and can lead to misleading conclusions due to confounding.

The Cox model allows us to build a multivariable model for the hazard rate, enabling us to estimate the effect of each covariate while adjusting for the effects of others. It helps answer questions like: "What is the impact of this new drug on survival, after accounting for differences in age and sex between the treatment and control groups?"

To see where the Cox model fits within the broader landscape of statistical methods for survival analysis, take a look at the following chart.

As you can see, the Cox model is classified as a semi-parametric method. Let's now delve into what that means by examining the model's structure.

The Cox Proportional Hazards Model Equation

The core of the Cox model is an equation that describes the hazard for an individual at a specific point in time, $t$, as a function of a set of covariates.

To summarize, the Cox proportional hazards model is written as:

Let's break this down:

• $h(t|X)$ is the hazard function for a subject with a specific set of covariate values ($X=(X_1,X_2,\dots ,X_p)$) at time $t$. This is the outcome our model is describing—the instantaneous risk of the event occurring at time $t$, given survival up to that time.
• The model has two main parts multiplied together: $h_0(t)$ and $\exp (\dots )$.

This structure is what makes the model "semi-parametric."

Explaining the Semi-Parametric Nature

The term "semi-parametric" means the model combines a non-parametric component with a parametric one.

1. The Non-Parametric Component: $h_0(t)$
   The term $h_0(t)$ is the baseline hazard function. This is the hazard rate when all covariates ($X_1,X_2,\dots$) are equal to zero. It is a function of time and can take on any shape—it can increase, decrease, or fluctuate. The Cox model does not make any assumptions about the shape of this function or the probability distribution of survival times. This is the non-parametric part of the model. This flexibility is a major advantage, as it makes the model robust; we don't risk being wrong about the underlying time-course of risk.

2. The Parametric Component: $\exp ({\beta }_1{X}_1+\cdots +{\beta }_p{X}_p)$
   This part of the model links the covariates to the hazard. It has a specific functional form (a linear combination of covariates that is then exponentiated), and we estimate the parameters ${\beta }_1,{\beta }_2,\dots ,{\beta }_p$ from the data. This is the parametric part. This term acts as a multiplier on the baseline hazard. The value of $\exp (\sum {\beta }_i{X}_i)$ tells us how the hazard for an individual with covariates $X$ compares to the baseline hazard.

The following reading explains this crucial distinction clearly.

The truly ingenious contribution by Sir David Cox was developing a method called partial likelihood estimation. This method allows us to estimate the $\beta$ coefficients—and thus understand the effects of the covariates—without ever needing to estimate the baseline hazard function $h_0(t)$ at all.

The video below gives a great analogy for this, comparing it to finding the slope of a line without needing to know its intercept.

The "Proportional Hazards" Assumption

The model's name also points to its single most important assumption. Because the time-dependent part ($h_0(t)$) is a common multiplier for everyone, the ratio of the hazards for any two individuals is constant over time.

For example, consider two individuals, A and B. Their hazard ratio is:

Notice that the baseline hazard $h_0(t)$ cancels out. The resulting ratio does not depend on time $t$. This is the proportional hazards assumption. If person A has twice the hazard of person B at the beginning of the study, they are assumed to have twice the hazard at any other point in time during follow-up.

This is a strong assumption, and we will dedicate a future lesson to understanding how to test for it. For now, it's important to recognize that this property is a direct consequence of the model's mathematical structure.

Conclusion

In this lesson, we have formally introduced the Cox proportional hazards model and dissected its structure. You should now be able to state the model's equation and explain why it is considered a semi-parametric method.

Here are the key takeaways:

• The Cox model is a regression model for time-to-event data that allows for the inclusion of multiple covariates.
• The model equation is $h(t|X)=h_0(t)\exp (\sum {\beta }_i{X}_i)$.
• It is semi-parametric because it combines a non-parametric part (the unspecified baseline hazard, $h_0(t)$) with a parametric part (the effect of covariates, $\exp (\sum {\beta }_i{X}_i)$).
• This structure allows us to estimate the effects of covariates without making assumptions about the shape of the underlying hazard over time, which makes the model robust and flexible.
• The model's structure implies the proportional hazards assumption: the ratio of hazards between any two individuals is constant over time.

We have established the "what" of the Cox model. In our next lesson, we will move on to the "so what," focusing on how to interpret the model's outputs—specifically the coefficients ($\beta$) and their exponentiated form, the hazard ratio (HR)—to draw meaningful conclusions from data.