When competing risks exist, the cumulative incidence function gives the true probability that the event of interest occurs by a certain time, accounting for other events that can prevent it. Kaplan-Meier can overestimate risk, while the Fine-Gray model helps explore covariate effects on CIF.

Multiple Choice

In cohort studies with competing risks, which method should be used to estimate the probability of the event of interest over time?

When other events can prevent the event of interest from happening, you need a method that directly gives the probability of that specific event occurring by each time point, accounting for those competing risks. The cumulative incidence function does exactly this: it provides the probability that the event of interest has occurred by time t, in the presence of competing events. Using a Kaplan-Meier estimator treats competing events as if individuals were simply censored, which assumes they could still experience the event of interest if followed longer. This inflates the estimated probability and misrepresents the true risk under competing risks. The log-rank test is a tool for comparing survival curves, not for estimating the actual probability of a particular event when competing risks are present. A Cox model without competing risks adjustment estimates hazard differences for the cause of interest but does not yield the true cumulative probability by time, because it ignores how competing events alter the risk set. So, the cumulative incidence function is the appropriate choice for estimating the probability of the event of interest over time in cohorts with competing risks. If you later want to assess covariate effects on that probability, you would move to models like the Fine-Gray subdistribution hazard model, which is built for CIF interpretation.

Cohort studies with competing risks: what you really need to know about estimating probability over time

If you’ve ever wandered into the world of survival analysis, you’ll quickly learn that not all events march to the same drumbeat. In many real-world scenarios, people can experience several different outcomes, and one outcome can block another from ever happening. Think of patients at risk for a particular complication, but some may die from another cause before that complication occurs. In this setting, we don’t just want to know “how long until X happens.” We want to know the probability that X has occurred by a given time, while recognizing that other events can stand in the way. That, in a nutshell, is the heart of competing risks in cohort studies.

Let me explain the core idea with a simple mental model. Imagine you’re tracking a rare heart-related event over a span of years. Some individuals will experience the event, some will experience a different fatal outcome, and some will never experience either during the study period. If you pretend those other fatal outcomes simply vanish when someone dies from something else, you’re misreading the landscape. The probability you care about—the event of interest—depends on who’s actually still at risk at each moment in time. Enter the cumulative incidence function (CIF).

What the CIF gives you

The cumulative incidence function is a precise way to estimate the probability that a specific event has occurred by time t, taking into account that competing events can prevent that event from ever happening. It’s not just about “how many people haven’t had the event yet.” It’s about the real, practical chance that the event has already occurred, given that some folks are, in fact, removed from risk by competing outcomes.

  • Intuition matters: If death from another cause happens, that person can’t develop the event of interest later. The CIF directly incorporates that reality.

  • Time, with context: The CIF isn’t a snapshot. It’s a curve showing how the probability evolves as time passes, which helps you compare risk across groups, patterns, or different treatment strategies without pretending everyone could still be at risk forever.

  • Practical interpretation: You can say, “By year five, there’s a 12% probability of event X happening, accounting for competing risks,” rather than giving a number that’s inflated because you treated competing events as if they didn’t exist.

Why not just use the Kaplan-Meier estimator?

The Kaplan-Meier (KM) estimator is the granddaddy of survival analysis, and it’s fabulous in the right context. But it has a blind spot when competing risks are present. KM treats competing events as censored observations—essentially saying, “We’ll pretend they could still experience the event of interest if we kept watching.” That’s a bold assumption when, in reality, those competing events permanently remove people from the risk set.

Because of this, the KM estimate tends to overstate the probability of the event of interest. It’s like saying, “If you ignore the fire in the kitchen, you’ll get a higher chance of your cake turning out,” which is charming in a metaphor, but not useful for accurate risk estimation in cohorts where other outcomes can preempt the event you care about.

And then there’s the log-rank test

If you’re comparing survival curves between two groups, the log-rank test is a handy tool. It tells you whether the overall survival experiences differ across groups. But it’s not designed to quantify the exact probability of the event of interest in the face of competing risks. It’s a comparison instrument, not a probability calculator under competing risk conditions. So you’ll still want the CIF when your research question is about the actual likelihood of the event over time.

Where the Cox model fits in

The Cox proportional hazards model is a workhorse for examining how covariates relate to hazard. However, a standard Cox model without adjustments for competing risks doesn’t yield the true cumulative probability by time for the event of interest. It estimates hazard differences and can imply relationships that don’t translate cleanly into real-world probability when competing events are at play. If your aim is to understand how covariates influence that specific probability under competing risks, you’d move toward a model designed for that interpretation.

That’s where the Fine-Gray subdistribution hazard model enters the scene

If you want to dig into how covariates affect the cumulative incidence of the event of interest, the Fine-Gray model is a natural next step. It’s built specifically to handle competing risks and to interpret effects directly on the CIF. It helps you translate hazard-like relationships into changes in the actual probability curve that clinicians, policymakers, or researchers can relate to.

A practical picture: choosing the right tool for the right question

  • Your question is about the probability that event X has occurred by time t, in the presence of competing events. Use the CIF. It’s the most faithful reflection of real-world risk in this scenario.

  • Your question is about comparing overall survival or time-to-event curves between groups, without a specific focus on the probability of a particular event in the presence of competing risks. A log-rank test could be appropriate for the comparison, but remember it won’t tell you the event-specific probability in the competing risks setting.

  • Your question is about whether and how a covariate shifts the CIF. Consider the Fine-Gray model, which targets that subdistribution hazard and yields covariate effects on the cumulative incidence.

How this plays out in practice

Let’s ground this with a relatable example. Suppose researchers are studying the incidence of a particular post-transplant complication. Some patients might develop the complication, but others may die from infection, organ rejection, or other causes before it ever occurs. If you report the probability that the complication occurs by year two, you want that number to reflect the fact that some patients can’t experience the complication after death from other causes. The CIF gives you that precise probability.

Now, if you want to know whether a treatment lowers that risk, you’d look at covariate effects within a Fine-Gray framework. You’d interpret how the treatment shifts the CIF over time, while staying honest about the competing risks.

Common misconceptions, cleared up

  • Misconception: If I use Kaplan-Meier, I’ll get the right probability for the event of interest even with competing risks.

Reality: KM tends to overestimate the probability because it treats competing events as censored, which assumes those individuals could have experienced the event later—an assumption that can be plainly untrue.

  • Misconception: The log-rank test tells me how likely the event of interest is to occur by a certain time.

Reality: It’s about comparing overall survival curves, not the exact event-specific probability in the presence of competing risks.

  • Misconception: Cox models always give the right interpretation under competing risks.

Reality: A standard Cox model ignores the way competing events shrink the risk set, so its hazard estimates don’t translate directly into the actual cumulative probability.

A friendly note on interpretation and reporting

When you report CIFs, you’re offering a clear, interpretable narrative: the probability that the event of interest has happened by a given time, with competing risks accounted for. People reading your work won’t have to hunt for caveats about assumptions; the CIF communicates that probability directly. If you go further and apply the Fine-Gray model, you’ll land on covariate effects that speak to how those probabilities change with different factors, while still respecting the reality that some patients cannot experience the event because another event occurred first.

Touching on broader ideas without losing focus

Cohort studies sit at the intersection of design and analysis. The guiding principle is simple: let the data speak in a way that mirrors the real world. When events can block other events, the most truthful story is told by methods that acknowledge that blocking. It’s a bit like tracking traffic flow with road closures. If you pretend the road stays open for everyone, you’ll misread the travel time and congestion patterns. The CIF gives you a map that respects those closures, showing you where the traffic actually goes.

In the end, choosing the right method isn’t about chasing the latest statistic. It’s about clarity, honesty, and relevance. The CIF isn’t just a statistical tool; it’s a lens that helps researchers, clinicians, and students see the true choreography of events over time. It answers: what’s the chance that this specific outcome has happened by a certain moment, given that other outcomes can alter who’s still in the running?

A practical takeaway

  • If you’re studying time-to-event data with competing risks, start with the cumulative incidence function to estimate the probability of the event of interest by time t.

  • If you need to understand how covariates influence that probability, consider the Fine-Gray model, which is designed for subdistribution hazards and CIF interpretation.

  • Use Kaplan-Meier and log-rank in their traditional arenas—survival estimation without competing risks or broad curve comparisons—recognizing their limitations in the competing risks setting.

A final thought, to keep the concept approachable

Competing risks aren’t a nuisance to be tucked away; they’re a reality to embrace. They remind us that real life isn’t a neat line where one thing happens after another. It’s a tangle of possibilities, some of which preclude others. The CIF helps you trace that tangle with clarity, so the numbers you report reflect the genuine probabilities people face. And that, in turn, makes the research more useful, more trustworthy, and a little bit closer to how things actually unfold in the world.