Competing Risks Analysis

In medical and public health research, understanding the time until an event occurs is a central goal. We often rely on survival analysis to predict outcomes like disease recurrence or death. However, standard methods can falter when reality presents more than one possible outcome. A patient might be at risk of graft failure, but also of a heart attack; a public health intervention might reduce firearm suicides, but individuals remain at risk of suicide by other means. These are not isolated possibilities but competing fates, and analyzing one without acknowledging the others can lead to flawed conclusions.

This article addresses a critical gap in traditional time-to-event analysis: the problem of competing risks. The widely used Kaplan-Meier method, by treating competing events as simple "censoring," often overestimates the true probability of an event, painting a distorted picture of risk. To correct this, we will explore the robust framework of competing risks analysis.

This exploration is divided into two parts. First, under "Principles and Mechanisms," we will dissect the core concepts of cause-specific hazard and the cumulative incidence function, revealing why they answer fundamentally different questions and how they provide a more accurate map of patient journeys. Second, in "Applications and Interdisciplinary Connections," we will see these principles in action, demonstrating their indispensable role in clinical decision-making, public health policy, and the frontier of personalized medicine. By the end, you will understand not just the mechanics of this method, but its power to provide a clearer, more honest understanding of risk in a complex world.

In our journey to understand the world, we often simplify. We imagine a single path from cause to effect, a straight line from A to B. But life, especially in biology and medicine, rarely unfolds along a single track. It is a landscape of branching paths, a network of possibilities and dead ends. To navigate this landscape, we need a more sophisticated map. This is the essence of competing risks analysis.

Imagine you are a physician advising a patient who has just received a kidney transplant. The patient’s most pressing question is simple and profound: “What is the chance my new kidney will fail within five years?” On the surface, this seems like a standard time-to-event question. We could follow a group of similar patients, note when their grafts fail, and use a classical survival analysis technique, like the famous ​​Kaplan-Meier estimator​​, to plot the probability of "graft-failure-free" survival over time.

But there’s a complication. In our group of patients, some may die from a heart attack or a stroke, their new kidney functioning perfectly until the very end. How do we account for them? A traditional approach might be to "censor" them—to treat them as if they simply vanished from the study at their time of death, assuming their risk of graft failure was no different from those who remained.

This is where the illusion of the single path breaks down. Death is not like a patient moving to another city and being lost to follow-up. A patient who dies from a heart attack is no longer at risk of graft failure. Their journey has ended on a different path. To ignore this fact—to treat a competing terminal event as simple censoring—is to make a fundamental error. It’s like trying to calculate the odds of a ship reaching its destination by only looking at the ships that are still at sea, and ignoring those that have sunk for unrelated reasons. It leads to a distorted view of reality.

To build a better map, we must recognize that at any moment, a patient stands at a fork in the road. For our transplant patient, one path leads to graft failure, another to death with a functioning graft. Competing risks analysis provides us with two distinct tools to describe these branching paths: the ​​cause-specific hazard​​ and the ​​cumulative incidence function​​. These two concepts answer two very different, but equally important, questions.

The Rate of Divergence: Cause-Specific Hazard

Imagine you are standing at a specific point in time, say three years after the transplant. You are looking at all the patients who are still alive with a functioning graft. The first question you might ask is: “Right now, at this very instant, what is the force of risk pulling these patients toward the path of graft failure?” This instantaneous rate of failure is the ​​cause-specific hazard (CSH)​​.

Formally, for a specific cause of failure $k$ (like graft failure), its cause-specific hazard $\lambda_k(t)$ is the instantaneous probability of failing from cause $k$ in a tiny interval of time after $t$, given that you have survived all possible events up to time $t$.

$\lambda_k(t) = \lim_{\Delta t \to 0} \frac{\mathbb{P}(t \le T \lt t+\Delta t, \text{Cause}=k \mid T \ge t)}{\Delta t}$

Here, $T$ is the time of the first event, and $T \ge t$ is the crucial condition: the individual is still "event-free." This is an etiological question. It’s about the underlying machinery of the disease process. A clinical researcher might ask this to understand how a risk factor, like high blood pressure, affects the immediate risk of graft failure among those who are currently well. When we calculate this hazard, any patient who experiences a competing event (like death from other causes) is immediately and permanently removed from the "at-risk" population for all subsequent calculations. They have left the road entirely.

The Final Destination: Cumulative Incidence

The second question is different. It’s not about the instantaneous rate, but about the big picture. “Looking forward from the moment of transplant, what is the overall proportion of patients who will have experienced graft failure by the five-year mark?” This is a question about cumulative probability, or ​​absolute risk​​. The function that answers it is the ​​cumulative incidence function (CIF)​​, often denoted $F_k(t)$.

Formally, the CIF is simply the probability of having failed from cause $k$ by time $t$.

$F_k(t) = \mathbb{P}(T \le t, \text{Cause}=k)$

Notice there is no conditioning on survival here. This is a prognostic question. It is what most patients, families, and health-system planners want to know. It tells us the real-world burden of an outcome, accounting for the fact that some people will be removed from risk by competing events. The sum of the CIFs for all possible events tells us the total probability that any event has occurred by time $t$.

Let’s now see, with the stark clarity of numbers, why the old Kaplan-Meier approach is so misleading. Imagine a hypothetical scenario in a study of older adults where the risk of developing a certain disease (event 1) competes with the risk of death from other causes (event 2). Let's assume for simplicity that the cause-specific hazards are constant over time: the disease hazard is $\lambda_1 = 0.04$ per year, and the death hazard is $\lambda_2 = 0.08$ per year. We want to know the 5-year risk of getting the disease.

A naive analysis, treating deaths as non-informative censoring, would focus only on $\lambda_1$. The estimated 5-year risk would be $1 - \exp(-\lambda_1 \times 5) = 1 - \exp(-0.04 \times 5) = 1 - \exp(-0.2) \approx 0.181$, or $18.1\%$.

But this ignores that a substantial number of people are being removed from the at-risk pool by the competing risk of death. The correct approach is to calculate the CIF. The overall hazard of any event is $\lambda_{total} = \lambda_1 + \lambda_2 = 0.12$. The probability of remaining event-free at time $t$ is $S(t) = \exp(-0.12t)$. The true cumulative incidence of the disease is found by integrating the cause-specific hazard weighted by this survival probability:

$F_1(5) = \int_0^5 \lambda_1 S(u) \, du = \int_0^5 0.04 \exp(-0.12u) \, du$

This integral works out to be $\frac{\lambda_1}{\lambda_1 + \lambda_2} (1 - \exp(-(\lambda_1 + \lambda_2) \times 5)) = \frac{0.04}{0.12} (1 - \exp(-0.6)) \approx 0.150$, or $15.0\%$.

The difference is not trivial. The naive method overestimates the true risk by more than 3 percentage points. It gives a false picture of the disease burden because it fails to acknowledge that death "steals" individuals who might otherwise have developed the disease. The Kaplan-Meier estimate tells you the risk in a fictional world where death doesn't exist; the CIF tells you the risk in the world we actually live in.

This brings us to a beautiful and counterintuitive point. Competing risks are not independent actors on a stage; they are dancers in a tightly choreographed performance. The link that connects them is the overall survival probability, $S(t)$. Changing the rate of one event ripples through the entire system and affects the cumulative probability of the others.

Let’s return to our disease example, with recovery ($\lambda_1$), disability ($\lambda_2$), and death ($\lambda_3$) as three competing outcomes. Suppose we introduce a powerful new treatment that dramatically cuts the death hazard $\lambda_3$ in half, but has no direct effect on the hazards of recovery or disability. What happens to the five-year cumulative incidence of disability, $F_2(5)$?

Our first intuition might be that nothing changes. The "force" pulling people toward disability, $\lambda_2$, is unchanged. But this is wrong. The cumulative incidence of disability is given by the integral: $F_2(t) = \int_0^t \lambda_2(u) S(u) \, du$. By halving the death hazard, we have reduced the overall hazard. This means that the overall survival probability, $S(t)$, will be higher at every point in time. More people are alive and event-free for longer.

Because more people are alive and "at risk" for disability at any given time, more people will ultimately end up on the disability path. So, counterintuitively, a treatment that reduces mortality will increase the observed cumulative incidence of disability. This is not a paradox; it is a profound demonstration of the interconnectedness of risks. You cannot alter one part of the system without affecting the whole. This is a crucial concept for public health: solving one problem can unmask or even increase the burden of another.

After this journey into complexity, it is fair to ask: is the simple, single-path view ever correct? The answer is yes, but only in specific, well-defined circumstances.

• ​​Overall Survival​​: If your endpoint of interest is "death from any cause," then there are no competing risks. All paths eventually lead to this single outcome. Standard Kaplan-Meier analysis is the perfect tool for this question.

• ​​Composite Endpoints​​: Often in clinical trials, researchers define a composite endpoint, such as "event-free survival," where the event is the first occurrence of relapse, a new malignancy, or death. In this case, the analysis is not about the risk of relapse, but the risk of any of these events. For this combined endpoint, there are no competing risks, and Kaplan-Meier is again the appropriate method.

The moment you narrow your focus to just one component of that composite—say, relapse only—while other events like death can still occur, you are back in the world of competing risks, and the principles we have discussed become essential. Choosing the right analytical tool is not about finding the most complex one; it is about honestly matching your tool to the nature of your question and the structure of reality.

In our last discussion, we uncovered a fundamental truth about time and chance: when multiple futures are possible, you cannot understand the probability of one without accounting for all the others. An event doesn't happen in a vacuum; it competes for its chance to occur. This might sound like an abstract philosophical point, but its consequences are profoundly practical. It changes how we predict a patient's future, evaluate public health policies, and even design the next generation of personalized medicines. Now, let's leave the world of pure principles and venture out to see this idea at work, shaping decisions that affect our health and lives every day.

Imagine you are a doctor. A 50-year-old patient sits before you, worried about their risk of a fatal heart attack in the next five years. How do you give them a realistic answer? You could look at studies on cardiovascular disease (CVD) and quote a risk figure. But this patient, like all of us, is at risk of many things—cancer, accidents, infections. A fatal stroke next year would, in a very final way, make the risk of a heart attack in two years completely irrelevant. The non-cardiovascular causes of death are in a constant race with the cardiovascular ones.

To give a true estimate, we must acknowledge this race. We must calculate the probability of dying from CVD in the presence of all other competing ways to die. This is not just an academic nicety. Using the tools of competing risks analysis, we find that the true 5-year risk of a CVD death is the cumulative result of the moment-by-moment risk of a CVD event, constantly discounted by the probability that the patient has survived everything else up to that moment. For instance, in a large cohort, if the instantaneous risk (the cause-specific hazard) of CVD death is $\lambda_{\text{cvd}} = 0.018$ per year and for non-CVD death is $\lambda_{\text{noncvd}} = 0.012$ per year, the actual 5-year risk of CVD death isn't simply related to $0.018$ alone. It is calculated by integrating the CVD risk over time, while accounting for the fact that the pool of people still "at risk" is constantly shrinking from both causes.

This principle becomes dramatically important when dealing with populations who face high rates of competing events, such as the elderly. Consider an older patient diagnosed with Merkel cell carcinoma, a type of skin cancer. A researcher might naively try to estimate the probability of dying from this cancer by following a group of patients and simply treating deaths from other causes (like stroke or heart failure) as if those patients just "dropped out" of the study—a statistical technique known as right-censoring.

But this is a profound error in logic. A patient who dies of a stroke is not "missing" in the same way as a patient who moves to another city and is lost to follow-up. Their death is a definitive event. It informs us, with certainty, that their chance of dying from Merkel cell carcinoma is now zero. Treating this informative event as non-informative censoring violates a core assumption of the standard Kaplan-Meier survival analysis and will always lead to an overestimation of the cancer-specific mortality. The competing risks framework corrects this by properly partitioning the possibilities: a patient can die of cancer, die of something else, or survive. By doing so, it provides a realistic, and often lower, estimate of the true burden of the disease.

The "events" in this race don't always have to be death. In oncology, a common question is whether a slow-growing cancer will transform into a more aggressive form. For a patient with follicular lymphoma, the future holds several possibilities: their disease could transform, they could pass away from another cause before transformation ever occurs, or they could remain in their current state. These are competing fates. Knowing that the 5-year cumulative probability of transformation is $0.15$ and the 5-year cumulative probability of dying without transformation is $0.10$ tells us a great deal. It means that after 5 years, the total probability of something happening is $0.15 + 0.10 = 0.25$. Consequently, the probability of a patient remaining alive and untransformed is $1 - 0.25 = 0.75$. We can also say that among those who did experience an event, the proportion whose event was transformation is $\frac{0.15}{0.25} = 0.60$. This level of nuanced understanding is impossible without treating these outcomes as competitors.

This framework is the daily language of many high-stakes medical fields. In hematopoietic cell transplantation for leukemia, physicians track patients for the devastating complication of graft-versus-host disease (GVHD). But patients are also at risk of their original cancer relapsing or dying from treatment-related toxicity. These are all competing events. To understand the true risk of GVHD, a clinic must count the number of patients who develop GVHD and divide it by the total number of patients who started, not just by those who didn't relapse or die first. This simple but rigorous accounting is the essence of calculating the cumulative incidence in the face of competing risks.

The logic of competing risks extends beyond the individual patient to the health of entire populations, and it can reveal subtle truths about human behavior.

Consider a major public health campaign aimed at reducing suicide by firearms, perhaps by promoting safe storage. Suppose the campaign is successful and the instantaneous risk (the hazard) of firearm suicide drops significantly. Does this guarantee a drop in the total number of suicides? A skeptic might argue for "method substitution": that people prevented from using one method will simply switch to another.

Competing risks analysis provides the tools to dissect this question with stunning clarity. Let's imagine a world where the campaign works, and the intrinsic danger of firearm suicide is cut in half. We also assume, for the sake of argument, that the intrinsic danger of suicide by other means does not change at all—that is, there is no behavioral substitution. What happens? The number of firearm suicides plummets, as expected. But what about suicides by other methods? Here lies the surprise: the number of people dying by other methods might slightly increase.

How can this be? It's not because people are switching methods. It's because by preventing firearm suicides, the campaign has allowed more people to survive longer. By surviving, they remain "at risk" for all other causes of death, including suicide by other means, for a longer period. This tiny increase in the cumulative incidence of other-method suicides is not a sign of the campaign's failure but a mathematical echo of its success in preventing the primary target. The key is to recognize that the cause-specific hazard for other methods remained unchanged, which is the true measure of whether behavioral substitution occurred. The net result in this scenario is a large, life-saving reduction in total suicides. This is a beautiful example of how the framework protects us from drawing false conclusions from raw numbers.

A similar logic applies when evaluating treatments for opioid use disorder. A patient might be offered methadone, buprenorphine, or naltrexone. We want to know which treatment best prevents a fatal overdose. However, a major challenge with these therapies is treatment dropout. Overdose and dropout are competing fates. A therapy like extended-release naltrexone might have a very low overdose hazard for patients who adhere to it, but it also has a notoriously high dropout rate early on. Another therapy, like methadone, might have a higher overdose hazard but much better patient retention.

Which is better? You cannot simply compare the overdose hazards. The high dropout rate for naltrexone means many patients are quickly removed from its protective effect, becoming vulnerable again. A competing risks analysis integrates both the risk of overdose while on treatment and the risk of dropping out. It can reveal a counter-intuitive result: a therapy with a slightly higher "on-treatment" overdose risk might actually result in a lower overall overdose incidence in a real-world population because it does a much better job of keeping people in care. The best treatment is not just the one that works best in theory, but the one that people can actually stick with.

Today, we stand at a new frontier where competing risks analysis is merging with genomics and data science to usher in an era of personalized medicine.

Imagine a patient who has just had a heart attack. They need antiplatelet medication to prevent another clot, but these drugs also increase the risk of major bleeding. There's a trade-off. Now, add another layer: the most common drug, clopidogrel, works poorly in the 30% of people who carry a specific genetic variant in the CYP2C19 gene, leaving them with a higher risk of clotting. An alternative drug works well for everyone but carries a slightly higher risk of bleeding.

What should we do? Should everyone get clopidogrel (Policy U)? Or should we do a genetic test and give the alternative drug only to the carriers (Policy G)? Competing risks analysis allows us to model this choice precisely. For each subgroup (carriers and non-carriers) and each policy, we can sum the hazards for the two competing events—thrombosis and bleeding—to find the total hazard of any adverse event. By calculating the weighted average of the event probabilities across the whole population, we can see which policy leads to a better overall outcome. The analysis shows that the genotype-guided policy, despite increasing bleeding risk for the carrier subgroup, provides such a large benefit in reducing their clotting risk that the overall incidence of adverse events in the population falls. We can quantify the exact benefit of a personalized medicine strategy.

This same logic is powering breakthroughs in cancer treatment. Patients on modern immune checkpoint inhibitors face a risk of life-threatening immune-related adverse events (irAEs), but they might also discontinue treatment for other reasons, like cancer progression. To estimate the true probability of an irAE, one must use a proper competing risks estimator, such as the Aalen-Johansen method.

Furthermore, if we want to build a model that predicts the risk of an irAE based on a patient's biomarkers, like their tumor mutational burden (TMB), we need even more sophisticated tools. We cannot simply model the cause-specific hazard of an irAE and assume it tells us the whole story about the probability of that event. The effect of a biomarker on the final probability depends on how it affects all competing events. To directly model the probability itself—the cumulative incidence function (CIF)—statisticians have developed the elegant Fine-Gray subdistribution hazard model. This model is specifically designed to tell us how a covariate like TMB affects a patient's ultimate chance of experiencing an irAE, correctly accounting for the tangled web of competing possibilities.

The ultimate challenge comes when we move from a few biomarkers to thousands of them, as in genomic (RNA-seq) data. Imagine trying to find which of 20,000 genes predict a patient's risk of dying from cancer, when non-cancer death is a major competing risk. The number of features ($p$) is far greater than the number of patients ($n$). This is a "high-dimensional" problem. The solution is a beautiful synthesis of old and new: we combine the principled logic of the Fine-Gray model (to target the right quantity, the CIF) with modern machine learning techniques like LASSO penalization, which can sift through thousands of potential predictors to find the few that truly matter. To ensure our final model is accurate, we use advanced validation techniques, like the Brier score adjusted for censoring, to measure how well our predicted probabilities match reality. This is the absolute cutting edge: a seamless integration of classical biostatistical reasoning with high-dimensional data science, all resting on the simple, powerful idea that you must respect the competition.

From a simple question about a patient's five-year risk to the complex task of building a genomic predictor from a mountain of data, the principle of competing risks provides a clear and honest lens. It forces us to see the world not as a set of isolated cause-and-effect chains, but as an interconnected system of possibilities. By embracing this complexity, we gain a truer, more powerful understanding of the dynamics of life, disease, and health.