The world of clinical trials often feels like a race to prove one word: effective. Yet patients don’t live in single moments of success—their lives are a stream of events: hospital visits, aches, hospital stays, and sometimes the final, terminal event. Traditional analyses tend to spotlight the first major event and then stop, as if the rest of the story doesn’t matter. A new paper from Yale and Dana-Farber researchers turns that idea on its head. For the team led by Xi Fang, with co-authors Fan Li and Hajime Uno, the question isn’t just whether a treatment slows the clock on the big outcomes, but how it changes the entire burden of illness while people are still alive. The result is a method called the while-alive regression, a way to quantify how covariates influence the daily rhythm of recurrent events and death together over time.
Think of it as shifting from a single snapshot to a long-exposure photograph of health. The researchers extend this idea to composite survival endpoints—endpoints that combine more than one kind of event, including fatal and non-fatal outcomes. They show how to model time-varying effects, so a treatment might help a patient early on, but the benefits could grow, fade, or even reverse as time goes by. And they don’t stop at simple, individual data. The framework also handles clustered data, which is how many real-world trials operate when patients are grouped by clinics or hospitals. Their toolkit is implemented in an R package called WAreg, making this live, flexible approach accessible to researchers and clinicians who want to follow the full arc of a patient’s experience rather than a single turning point.
Two real-world trials anchor the paper’s drama. The first is HF-ACTION, a large heart-failure study conducted across dozens of sites; the second is STRIDE, a cluster-randomized trial aimed at preventing falls in older adults. In both cases the researchers show that the story the data tells with the while-alive lens can diverge from what a traditional first-event analysis would reveal. It’s not just a statistical curiosity: it’s a different way of thinking about treatment impact—one that foregrounds the lived experience of patients, not just the countdown to the first major event.
The study is a collaboration among the Yale School of Public Health, the Yale Department of Biostatistics, and the Dana-Farber Cancer Institute, with the authors highlighting how their approach opens up a clearer path to estimating how covariates influence the burden of disease over time. The work also emphasizes practical considerations: how to choose the splines (the time-varying building blocks), how to balance recurrent events against death, and how to handle data that are missing or censored. The authors even provide a practical way to pick model complexity through cross-validation, a nod to real-world constraints in busy clinical research settings.
A fresh lens on composite endpoints
Composite endpoints bundle several outcomes into one summary, a strategy designed to boost the chances of detecting a treatment effect when events are rare or when different outcomes tell different parts of the story. But not all outcomes are equal in clinical importance. A heart attack and a hospitalization in the same patient don’t carry the same emotional or medical weight, and traditional models often implicitly treat every event as equally important. This mismatch has grown more obvious as trials began to count multiple events per person over long periods.
The authors’ response is to shift the focus from simply counting events to measuring the rate at which events occur while a patient is alive. They call this the while-alive loss rate. It combines two ideas: (1) the total, weighted burden of recurrent events (such as repeated hospitalizations or injuries) and (2) the terminal event (death) that ends the clock. The denominator is the expected amount of time a patient spends alive up to a chosen horizon, captured by the restricted mean survival time. In practical terms, the method asks: given how long a patient is alive, how many weighted events accumulate? The answer is a rate that reflects both frequency and time. It’s a more nuanced, patient-centered measure than counting only the first event.
The mathematical engine underneath this idea is a regression framework that links covariates to the while-alive loss rate via a smooth, time-varying function. The researchers show how to flexibly model how a treatment or a patient characteristic affects the burden at different moments in time, rather than forcing a single, time-invariant effect. They achieve this using splines—basically flexible curves stitched together across time—so the model can capture complex patterns, like effects that start small and grow as months turn into years. They also demonstrate how to handle censoring (when a patient leaves the study or is lost to follow-up) and how to extend the method from isolated individuals to clusters of patients treated within the same clinics.
Two trials, one idea
The HF-ACTION trial sits at the heart of the authors’ motivation. It randomized people with heart failure to an exercise program in addition to standard care or standard care alone. The composite endpoint—death or hospitalization—was already known to be informative, but the standard analysis treated death and hospitalization as if they carried equal weight and stopped at the first event. By applying the while-alive regression, Fang and colleagues were able to reveal how the effects of exercise duration and patient characteristics changed over time. In the high-risk subgroup they spotlighted, the treatment did more than reduce the first event; it reduced the burden of events experienced while patients were alive, with the effect becoming more pronounced after roughly two years. The upshot is that the benefits of exercise accumulate, a pattern that would have been easy to miss with a one-time snapshot.
The STRIDE trial, a cluster-randomized study in primary care practices, offers a complementary perspective. Here the focus is not a single patient but a whole practice environment designed to prevent falls and related injuries. The data persistently show a beneficial tilt: fewer recurrent injuries relative to the time lived, with death weighted more heavily than injuries. Yet the gains aren’t always statistically crystal clear. In the STRIDE analysis, the estimated reductions in the while-alive loss rate drifted into the negative territory (a hint of benefit) after year one, and the pattern persisted, but the confidence bands repeatedly included zero. The global test for the treatment effect was modest in strength. What matters, though, is the conceptual shift: the analysis reframes the trial outcomes as a narrative of living with risk over time, not merely a tally of the first bad event.
Taken together, the HF-ACTION and STRIDE applications illustrate a broader point. A time-varying, while-alive lens can reveal meaningful patterns that a fixed, one-shot metric might miss. It also makes the analysis directly relevant to patients and clinicians: it asks not only whether a treatment works on average, but how it modulates the daily burden of illness as time unfolds. The appendices of the paper walk through practical choices—how many time-splines to use, how to align stacking points with the time grid, and how to diagnose and address censoring. And there’s a palpable sense of translation here: the methods are implemented in WAreg, an R package that turns theory into a tool researchers can actually wield in ongoing studies.
Why this matters and what comes next
The methodological advance sits squarely at the intersection of clinical relevance and statistical sophistication. It aligns with the estimand framework that has become central to modern trial design, particularly the idea that researchers should clearly define what they want to estimate and how intercurrent events like death should influence the interpretation. The while-alive estimand—an approach that keeps death as part of the story rather than treating it as a censoring or nuisance event—offers a natural complement to existing strategies for recurrent events. It doesn’t throw away the complexity; it embraces it. In doing so, it gives clinicians and trial designers a richer vocabulary for describing how treatments affect patients’ lives over the course of months and years.
Beyond academic curiosity, there are practical implications. For drug developers, regulators, and health-care providers, a time-varying, while-alive view can influence how trials are planned and how results are communicated. It can highlight when a therapy’s benefits emerge, grow, or wane, which in turn affects decisions about dosing, patient selection, and duration of treatment. It can also improve the way we communicate risk to patients, offering a more complete picture of the trade-offs they face over time.
Of course, no statistical method is a panacea. The authors are careful about caveats. The choice of weights for recurrent versus terminal events is inherently subjective and context-dependent. The knot placement for splines and the stacking times for estimation introduce modeling decisions that can influence inference, though the authors provide data-driven strategies to guide those choices. In clustered settings, the paper adheres to a working-independence framework for estimation, which is robust in many situations but could be refined in future work to exploit more complex within-cluster correlations. The authors themselves point to exciting avenues for future research, including more principled knot-selection approaches that carry their own uncertainty into the inference, and deeper exploration of how to optimally weight different component outcomes in a way that’s agreed upon by clinicians, patients, and regulators.
Still, the collaboration across Yale and Dana-Farber—anchored by lead researchers Xi Fang, Fan Li, and Hajime Uno—represents a meaningful shift in how we interpret trial data. It’s a reminder that the stories clinical data tell aren’t just about whether a drug works in a vacuum, but about how people live with illness over time. The development of WAreg and the open questions the paper raises invite a broader community of researchers to experiment with time-varying, patient-centered endpoints. If we, as a field, can keep elevating the narrative from “first event wins” to “events accumulate while alive,” we edge closer to understanding not just whether treatments work, but how they shape the daily lives of those who rely on them.
As the authors put it in spirit if not exact words: the while-alive approach is more than a new statistical trick. It’s a different lens for looking at health, one that foregrounds the endurance of living, the rhythm of risk, and the slow, meaningful arc of improvement that can emerge only when we measure the whole journey, not just the first milestone.
