Fine-Gray Subdistribution Hazard Model

The defining feature of the Fine-Gray model is that it models the subdistribution hazard rather than the cause-specific hazard. This allows researchers to directly assess the effect of a covariate on the absolute risk of an event.

• Subdistribution Hazard: The instantaneous risk of failure from the event of interest in subjects who have not yet experienced that specific event, including those who have experienced a competing event.
• Risk Set: The risk set is unconventional; subjects who experience a competing event remain in the risk set indefinitely rather than being censored.
• Interpretation: A subdistribution hazard ratio (SHR) of 1.5 means a 50% increase in the cumulative incidence of the primary event, holding other variables constant.

The Fine-Gray model uses a modified partial likelihood with time-dependent weights to account for the competing risk structure. This weighting scheme is the mathematical core that distinguishes it from a standard Cox model.

• Inverse Probability Censoring Weighting (IPCW): Subjects who experience a competing event are retained in the risk set but down-weighted based on the probability of not being censored.
• Weight Calculation: The weight for a subject at time t is inversely proportional to the Kaplan-Meier estimate of the censoring distribution.
• Robust Variance: Standard errors must be calculated using a robust sandwich estimator because the weights are estimated, not known constants.

Like the Cox model, the Fine-Gray model assumes that the effect of a covariate on the subdistribution hazard is constant over time. This is a distinct assumption from the proportional cause-specific hazards assumption.

• Diagnostic Tools: Modified Schoenfeld residuals can be plotted against time to test this assumption for the subdistribution hazard.
• Violation Handling: If the assumption fails, one can include time-varying coefficients or use alternative models like direct binomial regression.
• Clinical Relevance: A constant SHR implies that the relative effect on the cumulative incidence is stable throughout the follow-up period.

The Fine-Gray model requires the assumption of random right-censoring for the validity of its weights. The handling of censoring is critical to avoid biased estimates.

• Administrative Censoring: Standard end-of-study censoring is generally valid.
• Informative Censoring: If subjects are censored for reasons related to their prognosis, the IPCW weights become biased, leading to incorrect SHR estimates.
• Left-Truncation: The model can be extended to handle left-truncated or delayed entry data by adjusting the risk set definitions accordingly.

A critical distinction exists between the Fine-Gray model and a standard Cox model applied to cause-specific hazards. They answer fundamentally different etiological questions.

• Cause-Specific Hazard: Models the biological rate of an event in a hypothetical world where competing events are removed. Useful for understanding disease etiology.
• Subdistribution Hazard: Models the actual risk of an event in the real world where competing events exist. Useful for predictive modeling and patient counseling.
• Numerical Divergence: A covariate can have a null effect on the cause-specific hazard but a significant effect on the subdistribution hazard if it strongly influences the competing event.

The foundational framework that the Fine-Gray model extends. Competing risks analysis accounts for events that preclude the primary event of interest—such as death from other causes preventing cancer relapse.

• Uses the Cumulative Incidence Function (CIF) rather than 1-Kaplan-Meier to avoid overestimation
• Distinguishes between cause-specific hazards and subdistribution hazards
• Essential when multiple event types are mutually exclusive

Example: In a stem cell transplant study, death in remission competes with relapse, requiring CIF estimation rather than standard survival curves.

The semiparametric workhorse of survival analysis that the Fine-Gray model was designed to complement. The Cox model estimates cause-specific hazard ratios by treating competing events as censored observations.

• Assumes proportional hazards across covariate levels
• Provides hazard ratios interpretable as instantaneous relative risk
• Appropriate when the research question concerns etiologic associations rather than absolute risk prediction

Key Distinction: Cox models the rate in those still at risk; Fine-Gray models the rate in the total population including those who have failed from competing causes.

The estimand directly modeled by the Fine-Gray subdistribution hazard approach. The CIF estimates the probability of experiencing a specific event by time t while accounting for competing events.

• Defined as the integral of the subdistribution hazard over time
• Unlike 1-Kaplan-Meier, the sum of all cause-specific CIFs equals the total event probability
• Provides clinically interpretable absolute risk estimates

Formula: CIF_k(t) = ∫₀ᵗ h̄_k(u) S(u) du, where h̄_k is the cause-specific hazard and S is overall survival.

The instantaneous rate of a specific event type among those still at risk for that event. This is the estimand from standard Cox regression applied to competing risks data.

• Answers the question: "What is the immediate risk among those who haven't yet failed?"
• Treats competing events as independent censoring
• Directly interpretable for etiologic research on disease mechanisms

Contrast with Fine-Gray: Cause-specific hazards address biological causation; subdistribution hazards address patient-facing prognostic questions about absolute risk.

Diagnostic tools adapted to test the proportional subdistribution hazards assumption in the Fine-Gray model. These residuals assess whether covariate effects remain constant over time on the subdistribution hazard scale.

• Computed as the difference between observed and expected covariate values at each event time
• Plotted against time with a smoothing spline to detect systematic trends
• A significant non-zero slope in the Grambsch-Therneau test indicates violation

Remedy: If proportional subdistribution hazards fail, consider time-varying coefficients or stratified Fine-Gray models.

An alternative approach to competing risks that reweights the standard Cox partial likelihood rather than modifying the risk set. IPCW addresses dependent censoring by the competing event.

• Weights each uncensored observation by the inverse of its probability of remaining uncensored
• Does not require the proportional subdistribution hazards assumption
• Can be combined with pseudo-observation methods for direct CIF modeling

Comparison: Fine-Gray modifies the risk set; IPCW modifies the weights. Both target the subdistribution hazard but with different technical assumptions.