Abstract
The proportional hazards (PH) model is arguably one of the most popular models used to analyze time to event data arising from clinical trials and longitudinal studies, among many others. In many such studies, the event time of interest is not directly observed but is known relative to periodic examination times; i.e., practitioners observe either current status or interval-censored data. The analysis of data of this structure is often fraught with many difficulties. Further exacerbating this issue, in some such studies the observed data also consists of instantaneous failures; i.e., the event times for several study units coincide exactly with the time at which the study begins. In light of these difficulties, this work focuses on developing a mixture model, under the PH assumptions, which can be used to analyze interval-censored data subject to instantaneous failures. To allow for modeling flexibility, two methods of estimating the unknown cumulative baseline hazard function are proposed; a fully parametric and a monotone spline representation are considered. Through a novel data augmentation procedure involving latent Poisson random variables, an expectation-maximization (EM) algorithm was developed to complete model fitting. The resulting EM algorithm is easy to implement and is computationally efficient. Moreover, through extensive simulation studies the proposed approach is shown to provide both reliable estimation and inference.
Introduction
In survival analysis, the Proportional Hazards (PH) model is the standard tool for modeling the time until an event occurs. However, real-world clinical data often breaks the model's basic assumptions in two significant ways. First, interval censoring occurs when the exact moment of an event is unknown, only that it happened between two checkups. Second, instantaneous failures occur when study units fail the moment the study begins ($t=0$), such as immediate surgical complications or deaths during a first treatment.
Standard software often struggles with the "spike" at zero. Researchers frequently use ad-hoc fixes, like adding a tiny constant to zero-times, but these methods lead to biased results and inaccurate survival curves. This paper introduces a formal mixture model framework to handle both interval censoring and $t=0$ spikes simultaneously under the PH assumption.
Methodology
The authors propose a mixture model that splits the probability into a discrete jump at $t=0$ and a continuous survival function for $t > 0$. To provide modeling flexibility, they offer two approaches for estimating the cumulative baseline hazard function:
- Parametric Representation: Utilizing fixed functional forms like linear or logarithmic curves.
- Monotone Spline Representation: A flexible, data-driven approach using piecewise polynomials constrained to increase over time.
To solve the complex optimization problem, the authors employ a data augmentation strategy using latent Poisson random variables. This "Poisson link" allows the non-linear PH model to be solved via an Expectation-Maximization (EM) algorithm. This approach breaks the problem into simple, closed-form steps that are computationally efficient and numerically stable.
Experimental Evaluation
The model was validated through extensive simulation studies involving 16 different data-generating scenarios. These scenarios varied the hazard shape (logarithmic vs. linear), covariate effects, and observation processes. Each scenario was tested using 500 independent datasets with $n=100$ observations.
While the primary validation is simulation-based, the work is grounded in real-world applications such as end-stage renal disease registries and intent-to-treat clinical trials for hernia repairs. The simulations confirmed that the monotone spline approach is highly robust; even when the true hazard shape is unknown, it performs nearly as well as a perfectly specified parametric model.
Key Results and Insights
- Unbiased Estimation: The proposed spline-based model showed negligible bias and achieved nominal 95% coverage for confidence intervals, providing reliable inference where traditional methods fail.
- Efficiency: Despite the mathematical complexity, the EM algorithm is fast, averaging approximately one second per model fit on standard hardware.
- Failure of Ad-hoc Methods: The study confirmed that ignoring the $t=0$ structure or using "small constant" workarounds leads to inaccurate survival curves that misrepresent initial risk.
- Software Availability: The methodology is integrated into the R package
ICsurv, making it accessible for immediate use by practitioners.
Limitations and Future Work
While the model is robust, some challenges remain. The "Outer Product of Gradients" (OPG) method for calculating standard errors occasionally results in numerical issues (singular matrices) for specific datasets, requiring manual adjustment of spline knots. Future research could expand this framework to handle more complex "mixed-case" censoring patterns or competing risks, where patients might leave a study for multiple different reasons.
Summary updated