TL;DR
Standard methods for estimating the Cumulative Incidence Function (CIF) in competing risks models often fail to sum to 1 at the end of a study, leading to logically inconsistent results. This paper introduces a new estimator ("Method 3") that guarantees the total probability across all event types equals 1 at the final observation, achieving this logical consistency without sacrificing statistical accuracy or power.
What Problem Does This Paper Address?
In survival analysis, "competing risks" occur when a subject is at risk of several different types of events, but the occurrence of one prevents the others (e.g., in a clinical trial, a patient might die from the disease under study or from an unrelated accident). Researchers use the Cumulative Incidence Function (CIF) to estimate the probability of a specific event type occurring over time.
When using the popular Cox proportional hazards model to account for covariates (like age or treatment group), a subtle but annoying problem arises: the sum of the estimated CIFs for all possible events at the end of the study often doesn't equal 1. Depending on the estimator used, the total probability might be 0.85 or 1.10—results that are physically impossible if we assume every subject eventually experiences one of the events. This paper tackles this lack of "logical consistency" in existing statistical software and literature.
Methods
The authors compare three approaches for estimating the CIF under a cause-specific Cox model:
- Method 1: A standard Breslow-type estimator.
- Method 2: A product-integral estimator (commonly used in software like R's
survivalpackage). - Method 3 (The Innovation): A newly proposed estimator inspired by the Kalbfleisch-Prentice survival estimator.
The core of Method 3 is an algebraic construction that ensures the "sum-to-one" property. While Methods 1 and 2 treat the baseline hazards as separate entities, Method 3 links them in a way that forces the estimated total distribution to drop to zero (and thus the cumulative distribution to reach one) at the time of the last observed event, provided there are no ties in the data.
Datasets
The authors validated their new method using two primary sources:
- Extensive Simulations: They tested 36 different scenarios varying hazard shapes (increasing, decreasing, up-and-down), sample sizes (75 and 150), and censoring rates (0% and 50%).
- Program Comprehension Data: A real-world dataset from a 2019 study by Ajami et al. involving 222 professional programmers. The "events" were whether a programmer interpreted a code snippet correctly or incorrectly. With a very low censoring rate (under 1%), this dataset provided a perfect test case for seeing how different methods handled the "total probability" at the end of the experiment.
Key Insights and Ideas
- Consistency for Free: The most significant "aha moment" is that Method 3 provides logical consistency (summing to 1) without a "performance tax." It performs just as well as—and in some cases better than—standard methods in terms of bias and variance.
- The Small Sample Trap: The paper highlights that while Methods 1 and 2 work fine for very large datasets, they can deviate substantially from 1 when sample sizes are small or when specific covariate values are analyzed.
- Confidence Interval Reliability: In simulations, Method 3 provided more reliable confidence interval coverage. Methods 1 and 2 occasionally showed coverage rates as low as 80% for a 95% interval, whereas Method 3 remained robust.
- Learning Effects: In the code comprehension example, the authors found a "substantial learning effect"—programmers became more likely to answer correctly as they progressed through the snippets. Method 3 was able to model these shifting probabilities while maintaining a total probability of 1.
Key Results
- Logical Anchoring: In the real-world data analysis, Methods 1 and 2 produced total CIFs ranging from 0.79 to 1.04 depending on the covariates. Method 3 was anchored at exactly 1.0.
- Bias and Variance: Across 1,000 simulation replications, Method 3 showed maximum mean bias and standard deviation comparable to the established methods, proving it is a mathematically sound alternative.
- Software Availability: To encourage adoption, the authors provided the R code for their simulations and Method 3 on GitHub.
Limitations and Future Work
- The "No Ties" Assumption: The mathematical guarantee of summing to 1 assumes there are no tied event times. While ties are rare in truly continuous data, they are common in discrete datasets. The authors suggest that small random perturbations can handle ties, but a more formal extension for tied data could be a future refinement.
- Censoring Sensitivity: While the method works well with censoring, its primary advantage (the sum-to-one property) is most visible and necessary in datasets with low censoring or when researchers want to extrapolate to the "end of life" for a cohort.
- Model Scope: The study focuses on the Cox cause-specific hazard model. Future work could explore whether similar "anchoring" techniques can be applied to other competing risks frameworks, such as the Fine-Gray subdistribution hazard model.
Summary updated