Proportional Hazards Assumption

In fields from medicine to engineering, understanding not just if an event will occur but when is a fundamental challenge. To compare the risk of an event—a patient's relapse, a machine's failure—between different groups over time, we need a simple yet powerful framework. The celebrated Cox proportional hazards model provides this by resting on a single, elegant idea: the Proportional Hazards (PH) assumption. This assumption proposes that the effect of a given factor, like a new drug, multiplies the underlying risk by a constant amount at every moment in time, allowing us to summarize a complex dynamic with a single number: the hazard ratio.

However, with great simplifying power comes great responsibility. What happens when this assumption doesn't hold true? What if a treatment's effect changes, being beneficial at first but harmful later? Ignoring such dynamics can lead to dangerously misleading conclusions. This article delves into this crucial statistical concept, equipping you to use it wisely. The first section, ​​Principles and Mechanisms​​, will demystify the hazard rate, explain the logic of the PH assumption, and detail the detective work required to test its validity. The second section, ​​Applications and Interdisciplinary Connections​​, will then showcase real-world scenarios from oncology to AI where this assumption is violated, revealing deeper scientific insights and exploring advanced methods to model this complexity.

To understand when things happen—whether it's the spoilage of a strawberry, the failure of a machine, or the onset of a disease—we need a language to talk about risk not as a simple "if," but as a "when." The central character in this story is the ​​hazard rate​​.

Imagine you have an old lightbulb. It has faithfully shone for a thousand hours. The question is not simply if it will fail, but what is its propensity to fail right now, in the very next instant, given that it has survived this long. This instantaneous risk, this "proneness to failure," is what we call the ​​hazard rate​​, often denoted by the Greek letter lambda, $\lambda(t)$.

More formally, the hazard at time $t$ is the probability of an event occurring in a tiny future interval, say from $t$ to $t + \Delta t$, given survival up to time $t$, all divided by the length of that interval: $\lambda(t) = \lim_{\Delta t \to 0^+} \frac{\mathbb{P}(t \le T t+\Delta t \mid T \ge t)}{\Delta t}$ The two crucial parts are the condition given survival up to time $t$ ($T \ge t$) and the division by $\Delta t$. This division makes it a rate—like speed in miles per hour—not a pure probability. And because it’s a rate, a hazard can, perhaps surprisingly, take a value greater than one. A hazard of $12$ per year simply means that, at that instant, the rate is such that you would expect $12$ events in a year if that rate were sustained and you had a large group of people.

The hazard function, $\lambda(t)$, tells a dynamic story. It might be high at the beginning and then decrease, like the risk of complications after a surgery. Or it might be low and increase over time, like the risk of an old car breaking down. The ​​survival curve​​, which shows the probability of remaining event-free past time $t$, is the direct consequence of the hazard experienced up to that point. A period of high hazard corresponds to a steep, downward slope in the survival curve. The two are inextricably linked through the ​​cumulative hazard​​, $H(t) = \int_0^t \lambda(u)du$, with the elegant relationship $S(t) = \exp(-H(t))$.

Now, let's say we want to compare two different worlds. We have strawberries stored in a refrigerator and others left at room temperature. Or patients receiving a new drug versus a placebo. How do their risks compare over time?

We could meticulously compare their hazard functions at every single moment, but this would be incredibly complex. Science often progresses by making simplifying, elegant assumptions. What if we propose the simplest possible relationship? What if one group's hazard curve is just a perfectly scaled version of the other's? Imagine the hazard curve for the refrigerated strawberries has a certain shape over time. What if the hazard for the room-temperature strawberries has the exact same shape, but is just multiplied by a constant factor—say, 4—at every single point in time?

This is the beautiful and powerful ​​Proportional Hazards (PH) assumption​​. It states that the ratio of the hazard functions for two groups is constant over time. This constant is called the ​​Hazard Ratio (HR)​​. $\frac{\lambda_{\text{group 1}}(t)}{\lambda_{\text{group 0}}(t)} = \text{HR (a constant)}$ If the HR is 2, it means an individual in group 1 has twice the instantaneous risk of the event as a comparable individual in group 0, and this is true on day 1, day 100, and day 1000. The beauty of this assumption is that it allows us to summarize the entire, potentially complex relationship between two survival curves with a single, meaningful number: the hazard ratio. This is the foundational idea of the celebrated Cox proportional hazards model.

But is the world we study always so neat and proportional? Often, it is not. Consider a real-world dilemma: comparing an aggressive new cancer surgery with a standard drug therapy.

• The surgery carries a high initial risk from post-operative complications. The hazard is high at the start but drops for those who recover.
• The drug therapy has a low initial risk, but its effectiveness might wane over time, or the cancer could develop resistance. The hazard starts low but may rise later.

In this case, the ratio of the hazards is clearly not constant. The surgery is riskier at the beginning but might be safer in the long run. The hazard curves will cross, and the proportional hazards assumption is violated. We can see this immediately if their mathematical forms are different; the ratio $\frac{A\exp(-kt) + C_1}{Bt + C_2}$ is obviously a function of time $t$.

When the HR changes over time, we have ​​non-proportional hazards​​. A tell-tale sign of this is seeing the survival curves of the two groups cross. Under proportional hazards, if one group has a higher risk (HR 1), its survival curve will always be below the other's; they can't switch places.

A deeper intuition for why this happens comes from the fascinating concept of ​​depletion of susceptibles​​. In the group with a high early hazard (the surgery group), the most frail individuals are "selected out" of the population quickly. The group of survivors becomes, in a sense, a hardier bunch. In the drug therapy group, this early "weeding out" process is less intense. So when we compare the hazards late in the study, we are no longer comparing apples to apples. We are comparing the survivors of an early trial-by-fire to a more general group. It is hardly surprising that their relative risk has changed.

So, what happens if we ignore this violation? What if we fit a standard Cox model anyway, forcing it to produce a single hazard ratio to summarize a non-proportional reality? The number it produces is a type of complex weighted average of the true, time-varying hazard ratio, where periods with more events have a stronger influence on the result.

This can be profoundly misleading. It's like summarizing a student's performance who got an 'A' in their first semester and an 'F' in their second with a final grade of 'C'. Does that 'C' reflect their journey? Not at all.

Let's consider a stark, hypothetical scenario to see the true danger. Imagine a new drug is tested. For the first two years, it's wonderfully protective, cutting the hazard in half ($HR = 0.5$). But in the third year, it becomes toxic, doubling the hazard ($HR = 2.0$). Let's say the duration of these periods is such that, by the end of the three-year study, the total number of people who have had the event is exactly the same in the drug group and the placebo group.

What will a standard Cox analysis report? It will see the equal number of total events, calculate an average HR of 1.0, and likely produce a non-significant p-value, leading to the conclusion that the "drug has no effect". This summary is a complete falsehood. It utterly masks the critical truth of a drug that is initially beneficial and later harmful. A single number has not just been imprecise; it has been qualitatively wrong.

Given that the PH assumption is powerful when true but treacherous when false, we must act like good detectives and rigorously check for clues before trusting our model.

​​Clue #1: Visual Inspection.​​ The first step is to plot the survival curves (often using the Kaplan-Meier method). Do they cross? Do they start parallel and then dramatically diverge or converge? These are strong hints of non-proportionality. But be warned: the absence of crossing does not guarantee proportionality. We need more formal tools.

​​Clue #2: Looking at the Leftovers.​​ The best detectives often find clues by sifting through the trash. In statistics, our "trash" consists of the ​​residuals​​—what's left over, the unexplained error, after our model has been fit. For the Cox model, we have a special tool called ​​Schoenfeld residuals​​.

The intuition is this: at each time an event occurs, the Schoenfeld residual for a covariate (like treatment group) is the difference between the covariate value for the person who had the event and the weighted average of that covariate among everyone still at risk at that moment. If the PH assumption holds, the effect of the covariate is constant, so these residuals should be randomly scattered around zero over time. But if we plot them against time and see a systematic trend—a slope or a curve—it's a clear signal from the data that the effect is changing with time, and our assumption is violated. This can be formalized into a statistical test.

​​Clue #3: The "Sting Operation".​​ A very powerful technique is to challenge the assumption directly. We can fit a more flexible, extended Cox model that explicitly allows the effect to change over time. This is often done by adding a ​​time-by-covariate interaction​​ term, for example, modeling the log-hazard ratio as $\beta(t) = \beta_0 + \beta_1 \log(t)$. We can then perform a statistical test on the $\beta_1$ coefficient. If $\beta_1$ is significantly different from zero, we have found strong evidence against the null hypothesis of a constant effect; the data confesses that the PH assumption is violated.

Finding that hazards are not proportional should not be seen as a failure. It is a discovery! It reveals that the relationship we are studying is more interesting and nuanced than we initially assumed.

When we detect non-proportionality, we don't throw up our hands. We switch to better tools. We can use the very time-interaction models from our diagnostic tests to describe and report how the effect changes over time—a much richer and more honest story than a single, misleading HR can provide. Alternatively, we can use different summary measures that do not depend on the PH assumption at all. One popular choice is the ​​difference in Restricted Mean Survival Time (RMST)​​, which simply compares the average event-free time between groups over a pre-specified clinical timeframe.

The journey from the simple, elegant proportional hazards assumption to the detective work of checking it, and finally to the more sophisticated models that embrace complexity, is a microcosm of the scientific method itself. We begin with a beautiful idea, we test it mercilessly against reality, and when it falls short, we build something better—something that gets us closer to the truth.

The journey into the world of statistics is often a search for elegant simplicities—patterns that allow us to see the forest for the trees. The Cox proportional hazards model offers one such breathtaking simplification. In a world of bewildering complexity, where the risk of an event—a patient's relapse, a machine's failure, a person's diagnosis—can change from moment to moment in unknowable ways, the Cox model hands us a gift. It allows us to ignore the messy, complicated shape of this underlying "baseline" hazard and ask a much simpler, more powerful question: how do specific factors, like a new drug or a lifestyle choice, multiply this risk? The core idea, the proportional hazards (PH) assumption, is that this multiplier—the hazard ratio—is constant. The effect of smoking on lung cancer risk, for example, is assumed to be the same multiplier on day one as it is on day one thousand. This assumption is what gives the model its power, allowing researchers in fields like oncology to estimate the effect of prognostic factors on survival without needing to know the exact shape of survival curves.

But with great simplifying power comes great responsibility. An assumption is a lens through which we view the world, and we must always be prepared to check if that lens is distorting our view. A scientist, in this sense, must also be a detective, constantly probing the validity of their tools. How does one test an assumption as grand as the constancy of risk over time? The principal tool for this detective work is a wonderfully intuitive concept known as ​​Schoenfeld residuals​​. Imagine at every moment a patient has an event, we feel a small "surprise." The residual is the difference between the characteristics of the patient who actually had the event and the expected characteristics based on everyone who was at risk at that moment. If the proportional hazards assumption is true, these little packets of surprise should show no particular pattern over time. They should look like random noise. But if we plot them and see a trend—a steady upward climb or a downward slide—a red flag goes up. The surprise is not random; it's systematic. This suggests the effect we're studying is not constant after all. This formal test, often called the Grambsch-Therneau test, involves regressing these residuals against time and checking if the slope is zero [@problem_id:4906532, 5106017, 4631691].

It is when we find these trends, when the simple assumption of proportionality breaks down, that the most interesting stories in science are often told. A violated assumption isn't a failure; it's a discovery.

Consider the modern battle against cancer. For decades, the workhorse was chemotherapy, a treatment that attacks rapidly dividing cells. Its effect is immediate and often harsh. More recently, a new class of drugs called immune checkpoint inhibitors has revolutionized oncology. These drugs don't attack the cancer directly; they teach the body's own immune system to recognize and fight it. This process takes time. If we compare the two treatments in a clinical trial for a disease like melanoma, we often see a curious pattern: for the first few months, the survival curves for the two groups are nearly identical. The hazard ratio is close to $1$. Then, as the immune system "wakes up" in the immunotherapy group, their curve flattens out, and a dramatic survival benefit emerges. The hazard ratio drops significantly below $1$. This phenomenon of "delayed separation of curves" is a classic, beautiful violation of the proportional hazards assumption. The effect of the treatment is simply not constant over time, and to summarize it with a single number would be to miss the entire biological story.

A similar narrative unfolds in the study of chronic diseases like HIV. A patient's CD4 cell count is a critical marker of immune health; a higher count is strongly protective against progressing to AIDS. But is the protective benefit of a high CD4 count constant throughout the long course of the disease? Studies have shown that it may not be. When investigators examine the Schoenfeld residuals for CD4 count, they can see a drift over time. This suggests that the protective effect may attenuate, or weaken, as the years go by. The log-hazard ratio, which starts strongly negative (protective), drifts closer to zero. The risk associated with a given CD4 count is different in year one than it is in year ten.

The effect doesn't always have to weaken. In epidemiological studies of environmental exposures, the opposite can occur. Imagine a cohort study tracking the effect of residential exposure to high traffic pollution on the risk of developing asthma. A test of the proportional hazards assumption might reveal that the hazard ratio for pollution exposure actually increases over time. This could point to a cumulative damage mechanism, where the longer one is exposed, the greater the impact on their instantaneous risk. In each of these cases—immunotherapy, HIV, and pollution—the failure of the proportional hazards assumption is not a statistical nuisance. It is the signature of a deeper, more dynamic biological or environmental process at work.

When our detective work uncovers a non-constant effect, we don't throw out the analysis. Instead, we adapt, using more flexible models that turn the "problem" into a profound scientific insight. The most common strategy is to incorporate the time-dependence directly into the model. Instead of estimating a single coefficient $\beta$ for our variable of interest, we let the coefficient be a function of time, $\beta(t)$. This is often done by adding a "time-interaction term" to the model, for instance, allowing the effect of a biomarker to change as a function of the logarithm of time. This allows us to move beyond a single, summary hazard ratio and instead describe how the effect evolves, reporting time-specific hazard ratios that tell a much richer story.

Another elegant solution, often used when a confounding variable violates the assumption, is ​​stratification​​. If, for example, we find that the effect of smoking is non-proportional, we can simply split our analysis into two separate groups, or "strata": smokers and non-smokers. We then allow the baseline hazard—that messy, unknown risk profile—to be completely different for each group. This perfectly controls for the variable without making any assumption about the shape of its effect over time. The only catch is that this technique cannot be used for the main exposure we want to study, because by splitting the groups, we lose the ability to estimate a hazard ratio for that exposure directly.

The principles of checking model assumptions have become even more critical in the age of artificial intelligence and "big data." In fields like radiomics, researchers develop complex prediction models based on thousands of features extracted from medical images. Reporting guidelines like TRIPOD (Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis) was established to ensure that the development and validation of these models are transparent and reproducible. A key part of this is reporting on checks of core statistical assumptions, like proportional hazards. Furthermore, if the assumption is violated, it means the model's ability to distinguish between high-risk and low-risk patients may change over time. Reporting a single, time-averaged performance metric like the Concordance index ($C$-index) can be misleading. Instead, good practice dictates reporting time-dependent performance measures that show how well the model works at different time horizons.

This timeless statistical wisdom now guides the very frontier of medical AI. Imagine a system that uses Graph Neural Networks (GNNs) to create a dynamic "patient similarity graph" in a hospital's intensive care unit, updating a patient's risk score for sepsis in real-time based on their evolving clinical data and their similarity to other patients. Can we plug these sophisticated, time-varying embeddings into a classic Cox model? The answer is likely no. The very nature of these dynamic features—which capture a patient's evolving response to treatment—screams "non-proportional hazards." The effect of a particular risk score today is unlikely to be the same tomorrow if a life-saving intervention is started in between. The solution is not to abandon the GNN, but to pair it with a survival model that is built for this complexity, such as a discrete-time hazard model. These models break time into small intervals and estimate the probability of an event in each window, naturally accommodating covariates whose values and effects change over time.

From the oncology clinic to the design of cutting-edge AI, the proportional hazards assumption is more than a technical detail. It is a guiding question that forces us to think deeply about the nature of risk itself. Is it static, or does it evolve? By asking this question, and by knowing how to answer it, we transform a simple statistical model into a powerful engine for scientific discovery.