In clinical research and epidemiology, we frequently ask questions about time: How long will a patient survive after a cancer diagnosis? How quickly will an individual develop cardiovascular disease based on their lifestyle?
Answering these questions requires analyzing time-to-event data. Standard statistical tools like linear or logistic regression fall short here. Logistic regression can tell us whether an event happened by the end of a study, but it ignores when it occurred. Linear regression struggles with censoring—the common scenario where a study ends before a patient experiences the event, or a patient drops out, leaving their ultimate survival time unknown.
To overcome these limitations, researchers rely on proportional hazards modeling, most notably the Cox proportional hazards regression model. This framework allows us to determine how specific risk factors or treatments accelerate or decelerate the time to an event while elegantly handling censored data and confounding variables.
The Anatomy of the Hazard Function
The core concept behind this approach is the hazard rate, denoted as $h(t)$. Instead of predicting the exact time an event will occur, the model estimates the instantaneous risk that the event happens at time $t$, given that the subject has survived up to that point.
The Cox model expresses this hazard as a product of two distinct components:
$$h(t | X) = h_0(t) \exp(\beta_1 x_1 + \beta_2 x_2 + \dots + \beta_p x_p)$$
Where:
- $h_0(t)$ is the baseline hazard function. It represents how the risk changes over time for an individual with baseline characteristics ($X = 0$). Crucially, the model makes no assumptions about the mathematical shape of $h_0(t)$—it can spike, fluctuate, or remain flat. This makes the model non-parametric regarding time.
- $\exp(\beta X)$ is the relative risk associated with a set of predictors (covariates). This component shifts the baseline hazard up or down multiplicatively based on individual characteristics. This part is parametric.
Because it combines non-parametric and parametric elements, the Cox model is called a semi-parametric model. The brilliant mathematical insight of this design is that when comparing the risk between two individuals, the baseline hazard $h_0(t)$ cancels out completely:
$$\frac{h(t | X_A)}{h(t | X_B)} = \frac{h_0(t) \exp(\beta X_A)}{h_0(t) \exp(\beta X_B)} = \exp(\beta (X_A - X_B))$$
This ratio is the Hazard Ratio (HR). Because the time-dependent baseline hazard disappears, the Hazard Ratio remains constant over time. This is the foundational proportional hazards assumption: if a risk factor doubles a patient's risk of an event today, it is assumed to double their risk next year as well.
By isolating the hazard ratio from the underlying timeline, researchers can quantify the relative risk reduction or increase of an intervention without needing to specify the exact distribution of survival times. For example, Lisa Scarton et al. utilized this property in Examining the relationship between metformin dose and cancer survival: A SEER-Medicare analysis to calculate hazard ratios that isolate the survival benefits of metformin in cancer patients, ensuring that varying follow-up durations did not skew the comparison.
Isolating Independent Effects through Multivariable Adjustment
In observational clinical studies, patients are rarely assigned to groups randomly. A study looking at a dietary habit or a biomarker might be confounded by age, sex, or pre-existing conditions. Proportional hazards modeling resolves this by allowing multiple covariates to be adjusted simultaneously within the exponential term.
When multiple variables are included, the model estimates the independent effect of each predictor while holding all other factors constant. For instance, in evaluating how early life milestones influence long-term health, Lu Chen et al. applied a multivariable Cox model in Age at Menarche and Menopause, Reproductive Lifespan, and Risk of Cardiovascular Events Among Chinese Postmenopausal Women: Results From a Large National Representative Cohort Study. This allowed the authors to isolate the specific impact of reproductive timelines on cardiovascular event risks, adjusting for potential confounding demographic variables across a large national cohort.
Similarly, the authors of Risk of severe COVID-19 disease with ACE inhibitors and angiotensin receptor blockers: cohort study including 8.3 million people adjusted for extensive sociodemographic factors and concurrent medications across millions of records. By doing so, they isolated the true hazard ratio of severe COVID-19 outcomes associated with specific blood pressure medications, mitigating the risk that underlying comorbidities would distort the findings.
Uncovering Non-Linearities and Interaction Effects
While the basic Cox model assumes that a covariate has a linear, log-proportional effect on risk, biological systems are rarely so simple. Proportional hazards modeling is flexible enough to capture complex relationships, such as U-shaped risk curves or effect modifications, by transforming variables or adding interaction terms.
Consider the "obesity paradox" or biomarker thresholds where both extremes of a measurement carry risk. Xiaodong Wu 0001 et al. utilized multivariable Cox models in Association of body mass index with severity and mortality of COVID-19 pneumonia: a two-center, retrospective cohort study from Wuhan, China to reveal a U-shaped relationship between Body Mass Index (BMI) and mortality. The model demonstrated that while high BMI increased severe complications, underweight states multiplied the hazard of death even more severely. A parallel phenomenon was captured by Wenyuan Gan et al. in Association between serum uric acid and all-cause and cardiovascular-related mortality in hemodialysis patients, where the proportional hazards framework identified a U-shaped mortality curve for uric acid levels, helping the authors determine that the risk at lower thresholds was independently driven by underlying muscle wasting.
Furthermore, risks are rarely uniform across different patient populations. The effect of an exposure might depend entirely on another clinical condition. This is known as effect modification and is evaluated using interaction terms. Kaizhi Bai et al. demonstrated this in Hypertension modifies the associations of body mass index and waist circumference with all-cause mortality among older Chinese: a retrospective cohort study, showing that an individual's hypertension status significantly alters the hazard ratios of anthropometric measures on mortality.
Accounting for the Dimension of Time
A profound advantage of proportional hazards modeling is its capacity to handle time-dependent phenomena. In long-term studies, exposures are rarely static; environmental factors fluctuate, and clinical interventions happen at specific, non-random intervals.
When exposures change dynamically, standard static models break down because they violate the proportional hazards assumption. To address this, time-dependent Cox regression updates covariate values over the course of the follow-up period, evaluating the hazard based on the participant's real-time exposure status. Yuling Bao et al. leveraged this capability in Preliminary investigation of the association between air pollution exposure and childhood asthma hospitalizations from 2015 to 2018 in East China to calculate hazard ratios for asthma hospitalizations based on varying longitudinal pollution levels rather than treating environmental exposure as a fixed baseline value.
Similarly, the timing of a medical intervention itself can drastically alter a patient's trajectory. In Delaying BCG immunotherapy onset after transurethral resection of non-muscle-invasive bladder cancer is associated with adverse survival outcomes, the authors quantified how operational delays multiply the underlying risk of tumor recurrence over time. By modeling the delay as a covariate, the framework proved that when an action is taken matters just as much as what action is taken, providing a mathematical foundation for optimizing clinical timelines.