Adaptive Competing Risks Analysis
- Introduced by
- Fine & Gray (subdistribution hazard, 1999); adaptive extensions by Beyersmann, Schumacher and colleagues
- Year
- 1999 (foundational Fine-Gray model); adaptive extensions 2000s–2010s
- Type
- Statistical survival analysis with adaptive interim monitoring
- Family
- Clinical / epidemiology
What does it do?
Adaptive competing risks analysis combines the Fine-Gray subdistribution hazard framework — which models the cumulative incidence of one cause of failure in the presence of other mutually exclusive causes — with adaptive or group-sequential interim monitoring rules.
This allows a clinical trial or observational study to be modified mid-course (e.g., sample size reassessment, early stopping) based on accumulating competing-risk data while maintaining pre-specified type I error control.
The intuition behind the method
In many clinical studies, patients can fail for more than one reason: cancer death, cardiovascular death, and treatment withdrawal are all possible outcomes. Standard survival analysis cannot distinguish these well; competing risks analysis models each cause separately by estimating the cumulative incidence function (CIF) rather than just the overall survival curve. Adding an adaptive layer means that, after examining interim data, the study team can adjust sample size or stopping boundaries in a statistically rigorous way — without the unplanned looks that inflate false-positive rates. Think of it as building an adjustable plan into your study from the start, with guardrails that ensure the final answer remains valid.
History
The theoretical foundation for competing risks analysis in survival data traces back to actuarial science, but the modern statistical framework was established by Fine and Gray in their landmark 1999 JASA paper, which introduced the proportional subdistribution hazards model and its associated K-sample test statistic. Adaptive and group-sequential extensions for competing risks settings were subsequently developed by Beyersmann, Schumacher, and colleagues, with key contributions also from Schmoor and Schwarzer. The method has gained traction in oncology trials, cardiovascular outcomes research, and transplantation medicine, where multiple failure types are the clinical norm rather than the exception.
When to use it
Use adaptive competing risks analysis when (1) the primary outcome is a time-to-event endpoint subject to mutually exclusive competing events (e.g., disease-specific death alongside all-cause mortality or non-fatal events), (2) uncertainty at the design stage makes mid-course adaptations valuable (e.g., uncertain event rates, effect size, or accrual), and (3) the study is large enough — typically a multi-site randomized trial or registry study — to justify the infrastructure for interim monitoring. Do NOT use when (a) competing events are negligible or all-cause mortality is the only endpoint (standard adaptive survival analysis suffices), (b) the sample is too small for interim monitoring to yield meaningful conditional power estimates, or (c) the adaptive rules cannot be pre-specified and locked before unblinding, which would invalidate type I error control.
Limitations and considerations
- The subdistribution hazard ratio from the Fine-Gray model does not have as direct a biological interpretation as the cause-specific hazard ratio; it mixes the effects of covariates on occurrence and on competing event survival.
- Requires pre-specification of all adaptive rules, alpha-spending functions, and interim look schedules before data unblinding — post-hoc adaptations invalidate inference.
- Interim data review by an unblinded statistician or independent data monitoring committee adds logistical and regulatory complexity.
- Small to moderate sample sizes may not provide sufficient conditional power at interim looks to make adaptation decisions reliable.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Correctly models the probability of failure from a specific cause without the bias introduced by treating competing events as independent censoring.
- The adaptive layer allows sample size reassessment or early stopping, improving efficiency and reducing exposure of participants to an inferior treatment.
- Maintains strict type I error control through pre-specified alpha-spending functions even when the study design is modified mid-course.
- Produces clinically interpretable outputs — cumulative incidence functions — that directly inform treatment decisions better than cause-specific hazard ratios alone.
- Applicable to both randomized clinical trials and large observational registry studies with event-time data.
Applications
- Oncology trials comparing cancer-specific survival in the presence of deaths from other causes, infections, or treatment discontinuation.
- Cardiovascular outcomes studies where the primary endpoint is major adverse cardiac events alongside competing risks of non-cardiovascular death.
- Bone marrow transplantation research modeling relapse incidence in the presence of competing transplant-related mortality.
- Registry-based pharmacovigilance studies assessing drug-attributable adverse events when patients can drop out or die from other causes.
- COVID-19 clinical trials monitoring time to mechanical ventilation while accounting for early discharge or death from non-respiratory causes.
How it works
How it works
Define the primary cause of interest (e.g., cancer-specific mortality) and enumerate all competing events (e.g., death from other causes, withdrawal). Pre-specify the subdistribution hazard model covariates, the number and timing of interim looks, and the adaptive rules (sample size reassessment boundaries, futility or superiority stopping rules) along with the alpha-spending function (e.g., O'Brien-Fleming) to control family-wise type I error.
At each interim and final analysis, estimate the cumulative incidence function for each cause using the Fine-Gray proportional subdistribution hazards model. Subjects who have experienced a competing event remain in the risk set for the cause of interest with weight reduced to zero — a key distinction from cause-specific hazard models. The subdistribution hazard ratio exp(β) quantifies covariate effects on the CIF directly.
Formula
\hat{F}_k(t) = 1 - \exp\!\left(-\int_0^t \exp(\beta^\top Z)\, d\Lambda_{0k}(s)\right)Compute interim test statistics (e.g., Gray's K-sample test or score statistics from the Fine-Gray model) and compare them against the pre-specified spending-function boundaries. Assess whether the cumulative alpha spent to date exceeds the spending boundary for early stopping (superiority or futility).
If pre-specified adaptive rules are met — for example, conditional power falls below a futility threshold or interim effect size suggests a larger sample is needed — modify the design accordingly. Permitted adaptations include sample size reassessment, arm dropping, or covariate-adjustment revisions. All adaptations must follow the pre-registered plan to preserve type I error integrity.
At the final analysis, compute the adjusted p-value and confidence interval for the subdistribution hazard ratio using the combination-test or inverse-normal method to account for adaptive modifications. Report cause-specific CIFs with confidence bands, and present both the primary endpoint and competing event rates to give clinicians a complete picture of the benefit-risk profile.
Common pitfalls
- Treating competing events as independent censoring in a standard Kaplan-Meier or Cox model — this overestimates the cumulative incidence of the primary event and inflates apparent treatment benefits.
- Conducting unplanned interim analyses outside the pre-specified alpha-spending framework, which inflates the type I error rate and can invalidate regulatory submissions.
- Ignoring the competing events entirely when reporting results — a treatment that reduces cancer death but increases cardiovascular death may appear beneficial if only one endpoint is shown.
- Using adaptive rules that depend on the unblinded treatment effect without an appropriate combination test, leading to invalid p-values and confidence intervals at the final analysis.
- Confusing cumulative incidence functions with (1 − Kaplan-Meier) estimator values; only the CIF is the correct estimator when competing risks are present.
Frequently asked questions
What is the difference between the Fine-Gray model and a cause-specific hazard model?
The cause-specific hazard model estimates the instantaneous rate of failure from a specific cause among those still event-free — competing events are treated as censored. The Fine-Gray subdistribution hazard model keeps subjects who experience a competing event in the risk set (with decreasing weight), so it directly models the cumulative incidence function. For predicting absolute risk in a population, the Fine-Gray CIF is more appropriate; for understanding biological mechanisms of a single failure type, cause-specific hazards may be preferable.
How do I choose the alpha-spending function for adaptive interim monitoring?
The O'Brien-Fleming spending function is conservative early and spends more alpha late — suitable when early stopping for efficacy is unlikely. The Pocock spending function allocates alpha more evenly, making early stopping for efficacy easier but requiring a lower final threshold. The choice should be pre-specified in the protocol and statistical analysis plan; regulatory agencies (FDA, EMA) require this pre-specification before any unblinded interim analysis.
Can I add covariates to the Fine-Gray model in an adaptive design?
Yes. Covariate adjustment in the Fine-Gray model is common for baseline imbalances, prognostic stratification factors, or increasing precision. In an adaptive design, the covariates included in the model should be pre-specified; adding or removing covariates after viewing interim data requires a pre-planned covariate-selection rule or risks inflating the type I error.
Is adaptive competing risks analysis accepted by regulatory agencies?
Adaptive trial designs more broadly are accepted by the FDA (see their 2019 guidance on adaptive designs for clinical trials) and EMA, provided the adaptive rules, alpha-spending functions, and operating characteristics are pre-specified and the trial maintains blinding during interim reviews by an independent data monitoring committee. Competing risks endpoints are also standard in oncology and cardiovascular submissions, though the analysis plan should clarify whether the primary endpoint is the CIF or the cause-specific hazard.
What software supports adaptive competing risks analysis?
The R package cmprsk implements the Fine-Gray model and Gray's test. The mstate package supports multistate extensions. For adaptive design operating characteristics, gsDesign and rpact support group-sequential frameworks; combining these with competing risks models typically requires custom simulation code or purpose-built packages such as PASS (NCSS) for sample size calculations under competing risks.
Related methods
Sources
Fine, J. P., & Gray, R. J. (1999). A proportional hazards model for the subdistribution of a competing risk. Journal of the American Statistical Association, 94(446), 496–509.
Open DOI referenceCopied.Beyersmann, J., Allignol, A., & Schumacher, M. (2012). Competing Risks and Multistate Models with R. Springer.
ISBN: 978-1461420767Copied.