Competing Risk Analysis in Cardiovascular Outcome Trials: A Simulation Comparison of Cox and Fine-Gray Models

arXiv:2602.16031 · stat.ME, stat.AP · Submitted 2026-02-17 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: We're looking at the paper: "Competing Risk Analysis in Cardiovascular Outcome Trials".

Jane: Here's the gist of the paper. Competing Risk Analysis in Cardiovascular Outcome Trials: A Simulation Comparison of Cox and Fine-Gray Models is a simulation study comparing Cox and Fine-Gray models across various clinically relevant scenarios to understand when they yield concordant versus divergent estimates, providing practical guidance in cardiovascular outcome studies<ref:2602.16031.arXiv:2602.16031v1

stat.ME: 5>.

Tom: Methodological Setup The study conducts Monte Carlo simulations generating correlated time-to-event data using bivariate copula models<ref:2602.16031.arXiv:2602.16031v1

stat.ME: 5>.

Jane: The joint survival function is defined by the equation S(t1, t2Z = z) = exp(-

(θz 1 · λ 1 · t 1)α + (θz 2 · λ 2 · t 2)α: 1/α)<ref:2602.16031.arXiv:2602.16031v1

stat.ME: 5>.

Tom: Data generation involves randomization, generating correlated uniform variables via Gumbel-Hougaard copula, and transforming these into event times T1i and T2i<ref:2602.16031.arXiv:2602.16031v1

stat.ME: 5>.

Title and authors: Tom: We just talked about the title and authors of this paper, "Competing Risk Analysis in Cardiovascular Outcome Trials: A Simulation Comparison of Cox and Fine-Gray Models," and how it sets up a simulation to compare the Cox model against the Fine-Gray model to see when their results agree or disagree.

Jane: That comparison is key because in cardiovascular trials, we often have two things happening at once, like a patient dying from their heart disease and then dying from another cause while they're still being watched. The title tells us the paper is about figuring out which statistical model to trust when that happens.

Lu: The authors are Tuo Wang and Yu Du from Global Statistical Sciences at Eli Lilly and Company, Indiana, United States. They set up a Monte Carlo simulation using bivariate copula models to generate correlated time-to-event data to test these two approaches across various clinically relevant scenarios.

Meng: It’s interesting that they are simulating the varying competing event rate between zero point five percent and five percent annually, which gives us a wide range of risk levels to test against their models.

Lalam: They specifically vary the treatment effect on the competing event, theta two, from zero point five reduction to one point five increase in hazard. That means they are testing scenarios where treatment helps with the competing event and scenarios where it hurts it too.

Jane: So, when we look at this simulation study, what is the main takeaway for us? Is there a clear answer about which model to use in a real trial right now?

Tom: The main message is that the choice between Cox and Fine-Gray models matters primarily when competing event rates are high and treatment effects are directionally discordant. It’s not some simple rule for every situation.

The paper's summary: Tom: So, summarizing the study, they set up this comparison to see if we get consistent results when we change the inputs—like how fast the competing events happen or how strong the treatment effect is on those events.

Jane: They found that in most standard cardiovascular trials, especially those with about one percent annual non-CV death rate, both Cox and Fine-Gray models give very similar estimates on average.

Lu: That similarity is what they report when a typical CVOT setting has non-CV death rates of approximately one percent annually, the two methods produce very similar estimates on average.

Meng: It means that for the vast majority of standard cardiovascular trials, you can use the simpler Cox proportional hazards model treating competing events as censored without major issues for your primary analysis.

Lalam: They also mentioned that they tested whether the correlation between primary and competing events was a driver of divergence, and it turns out that correlation matters when those other factors are already pushing things away.

Jane: So, the comparison isn't just about which model is "better," but figuring out exactly what causes those two models to give different answers.

The paper's improvements: Tom: The authors suggest that for supplementary descriptive analysis, cumulative incidence using the Aalen-Johansen estimator can provide transparency about the impact of the competing risk.

Lu: They also point out that when marginal causal effects are the target of inference, alternative approaches are motivated when both standard methods fail to recover consistent estimates.

Meng: Specifically, they suggest things like inverse probability of censoring weighting or multistate models as options for complex analyses when seeking that marginal causal treatment effect.

Jane: And they also mentioned including all-cause mortality in the composite endpoint as another alternative that sidesteps competing risk complications entirely by capturing the full mortality benefit.

Tom: But we need to be careful about what these alternatives do, because neither Cox nor Fine-Gray consistently recovers the true marginal hazard ratio under most scenarios with non-negligible competing risks.

Lu: The bias is minimal when the treatment effect on the competing event theta two is similar to the treatment effect on primary events theta one, which is zero point eight zero regardless of competing event rate or correlation structure.

Meng: So, if you're aiming for that true marginal hazard ratio of zero point eight zero, you have to be careful because it only works under very specific conditions.

Conclusion: Jane: So to wrap up this whole discussion on the "Competing Risk Analysis in Cardiovascular Outcome Trials: A Simulation Comparison of Cox and Fine-Gray Models," what’s the final word?

Tom: The concordance of Cox and Fine-Gray methods in typical CVOT settings—where competing event rates remain at approximately one percent—provides reassurance that analytical approach selection has limited impact on trial conclusions under those conditions.

Lu: That reassurance is important, but the paper also says that pre-specified Cox models should not be abandoned post-hoc simply because competing risks appear relevant after unblinding data.

Meng: Also, the Fine-Gray model doesn't constitute proper sensitivity analyses to Cox models per ICH E9(R1) because they target different estimands rather than testing assumptions of the same estimand.

Lalam: And they strongly support continuing the use of interpretable Cox proportional hazards models as the primary analytical approach for most CV outcome trials.

Jane: So, to summarize, the choice between methods matters when competing event rates are high and treatment effects are directionally discordant; otherwise, sticking with the simpler Cox model is usually the best path forward.

Tom: That’s right. This study gives us evidence to inform trial design, protocol development, statistical analysis planning, and regulatory discussions around competing risk methodology.

Lu: And for future work, they suggest looking into inverse probability of censoring weighting or multistate models when those standard methods fail to recover consistent estimates for marginal causal treatment effects.

Meng: And we’re always watching that high competing event rate scenario—exceeding five percent annually combined with highly discordant treatment effects where theta two is one point zero—because that’s the tipping point.

Lalam: That's all we have from this paper for now on this topic.

Tuo Wang, Yu Du

Global Statistical Sciences, Eli Lilly and Company

stat.ME, stat.AP

Submitted: 2026-02-17

Updated: 2026-11-06

Comments: 18 pages, 6 figures

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 90/100

The gist: Competing Risk Analysis in Cardiovascular Outcome Trials: A Simulation Comparison of Cox and Fine-Gray Models is a simulation study comparing Cox and Fine-Gray models across various clinically…

Key concepts

Competing Risk Analysis
This statistical approach is used when more than one event can happen at the same time, and only one outcome is of interest. In cardiovascular trials, this means a patient could die from the primary condition or a competing condition (like cancer) during follow-up. The models help estimate the risk of the main event while accounting for these other possible events.
Cox Model
The Cox proportional hazards model is a standard survival analysis tool used to assess how treatment affects survival time. It treats competing events as censored, meaning it assumes that if a patient experiences a competing event, they are no longer at risk for the primary event. It provides clear and interpretable results in many trial settings.
Fine-Gray Model
The Fine-Gray model is an alternative survival analysis method specifically designed to handle competing risks by estimating the subdistribution hazard. This model is useful when you want to analyze the risk of a specific event occurring before any other event, rather than just censoring patients due to a competing risk.
Marginal Hazard Ratio
This refers to the true effect of treatment on the primary outcome if there were no competing risks. The study found that neither the Cox nor Fine-Gray models consistently recovered this true value under most conditions, suggesting caution when interpreting results in complex scenarios.

Terminology

Summary

Competing Risk Analysis in Cardiovascular Outcome Trials: A Simulation Comparison of Cox and Fine-Gray Models is a simulation study comparing Cox and Fine-Gray models across various clinically relevant scenarios to understand when they yield concordant versus divergent estimates, providing practical guidance in cardiovascular outcome studies<ref:2602. p. 016031.arXiv:2602. p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 016031v1[stat. p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0 p. 0ME] 5>.

Methodological Setup

The study conducts Monte Carlo simulations generating correlated time-to-event data using bivariate copula models<ref:2602.16031.arXiv:2602.16031v1[stat.ME] 5>. The joint survival function is defined by the equation S(t1, t2Z = z) = exp(-[(θz 1 · λ 1 · t 1)α + (θz 2 · λ 2 · t 2)α] 1/α)<ref:2602. p. 416031.arXiv:2602.16031v1[stat.ME] 5>. Data generation involves randomization, generating correlated uniform variables via Gumbel-Hougaard copula, and transforming these into event times T1i and T2i<ref:2602. p. 416031.arXiv:2602.16031v1[stat.ME] 5>. Administrative censoring is generated to mimic staggered uniform enrollment of 2 years and maximum follow-up of 5 years<ref:2602. p. 416031.arXiv:2602.16031v1[stat.ME] 5>.

Simulation Parameters

The simulation systematically varies three parameters across scenarios to test model robustness<ref:2602.16031.arXiv:2602.16031v1[stat.ME] 5>. These parameters include:

  • The copula dependence parameter α takes values of 1.0, 1.2, 1.5, and 2.0, corresponding to Kendall’s τ of 0 (independence) to 0. p. 5500 (strong positive correlation)<ref:2602. p. 416031.arXiv:2602.16031v1[stat.ME] 5>.

  • The competing event rate λ 2 ranges from 0.5% to 5% annually, spanning from very low to moderately high competing risk<ref:2602. p. 516031.arXiv:2602.16031v1[stat.ME] 5>.

  • The treatment effect on competing events θ 2 ranges from 0.5 to 1.5, representing three clinically relevant scenarios: treatment reduces competing events (θ 2 1)<ref:2602. p. 516031.arXiv:2602.16031v1[stat.ME] 5>.

Comparative Findings

Across all simulation scenarios, Cox and Fine-Gray methods generally produce concordant results<ref:2602.16031.arXiv:2602.16031v1[stat.ME] 7>. In typical CVOT settings with non-CV death rates of approximately 1% annually, the two methods produce very similar estimates on average<ref:2602. p. 616031.arXiv:2602.16031v1[stat.ME] 8>. The divergence between Cox and Fine-Gray estimates is primarily driven by three factors: the competing event rate, the treatment effect on competing events, and the correlation between primary and competing events<ref:2602. p. 616031.arXiv:2602.16031v1[stat.ME] 7>. Substantial divergence emerges only when competing rates exceed 5% annually combined with highly discordant treatment effects (θ 2 1. p. 80)<ref:2602. p. 016031.arXiv:2602.16031v1[stat.ME] 8>.

Marginal Hazard Recovery

Neither the Cox model nor the Fine-Gray model consistently recovers the true marginal hazard ratio (θ 1 = 0.80) under most scenarios with non-negligible competing risks<ref:2602. p. 716031.arXiv:2602.16031v1[stat.ME] 9>. Bias is minimal when the treatment effect on competing events (θ 2) is similar to the treatment effect on primary events (θ 1 = 0. p. 780), regardless of competing event rate or correlation structure<ref:2602. p. 716031.arXiv:2602.16031v1[stat.ME] 9>. Both estimators can provide values close to the true marginal hazard ratio under specific conditions, such as when events are independent or when treatment effects are identical across both event types<ref:2602. p. 816031.arXiv:2602.16031v1[stat.ME] 9>.

Practical Recommendations

For the majority of CV outcome trials, simple Cox proportional hazards models treating competing events as censored provide valid primary analyses with clear interpretability<ref:2602.16031.arXiv:2602.16031v1[stat.ME] 14>. In typical CVOT settings where competing event rates remain at approximately 1%, the simple Cox model remains the preferred approach for primary analysis due to interpretational challenges associated with subdistribution hazards<ref:2602. p. 216031.arXiv:2602.16031v1[stat.ME] 8, 9>. The paper concludes that pre-specified Cox models should not be abandoned post-hoc simply because competing risks appear relevant after unblinding data<ref:2602. p. 116031.arXiv:2602.16031v1[stat.ME] 8>. Furthermore, the Fine-Gray model does not constitute proper sensitivity analyses to Cox models per ICH E9(R1), as it targets different estimands rather than testing assumptions of the same estimand<ref:2602. p. 116031.arXiv:2602.16031v1[stat.ME] 8, 9>.

Alternative Approaches

When marginal causal effects are the target of inference, alternative approaches are motivated when both standard methods fail to recover consistent estimates<ref:2602.16031.arXiv:2602.16031v1[stat.ME] 9>. These alternatives include inverse probability of censoring weighting (IPCW) methods and multistate models which explicitly model transitions between health states<ref:2602. p. 116031.arXiv:2602.16031v1[stat.ME] 9, 8>. Including all-cause mortality in the composite endpoint is another alternative that sidesteps competing risk complications entirely by capturing the full mortality benefit<ref:2602. p. 116031.arXiv:2602.16031v1[stat.ME] 9>.

Study Limitations

The simulation study has several limitations, including the use of the Gumbel-Hougaard copula which restricts attention to positive associations between events<ref:2602.16031.arXiv:2602.16031v1[stat.ME] 9>. The data generation assumes proportional hazards for both event types throughout follow-up, whereas real trials may exhibit nonproportionality due to treatment lag effects or waning effects over time<ref:2602.16031.arXiv:2602.16031v1[stat.ME] 9>. The simulation focuses on the typical CVOT setting where competing event rates are relatively low<ref:2602. p. 716031.arXiv:2602.16031v1[stat.ME] 9>. These findings may not apply to clinical trials in other therapeutic areas such as oncology, where competing event rates can be substantially higher<ref:2602. p. 716031.arXiv:2602.16031v1[stat.ME] 9>. The paper is based on simulated data generated through Monte Carlo methods using publicly available statistical software<ref:2602. p. 016031.arXiv:2602.16031v1[stat.ME] 5>.

Conclusion

The concordance of Cox and Fine-Gray methods in typical CVOT settings—where competing event rates remain at approximately 1%—provides reassurance that analytical approach selection, while important to pre-specify, has limited impact on trial conclusions under these conditions<ref:2602.16031.arXiv:2602.16031v1[stat.ME] 5>. The choice between methods matters primarily when competing event rates are high and treatment effects are directionally discordant, scenarios uncommon in typical CV outcome trials<ref:2602. p. 816031.arXiv:2602.16031v1[stat.ME] 9>. These evidence-based findings should inform trial design, protocol development, statistical analysis planning, and regulatory discussions around competing risk methodology<ref:2602. p. 216031.arXiv:2602.16031v1[stat.ME] 9>. The paper supports the continued use of interpretable Cox proportional hazards models as the primary analytical approach for most CV outcome trials<ref:2602. p. 316031.arXiv:2602.16031v1[stat.ME] 9>.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved AI system can perform primary analysis using Cox proportional hazards models treating competing events as censored when competing event rates remain low, as this approach remains appropriate for primary analysis in typical CVOT settings.

  2. Improved AI system can provide supplementary descriptive analysis using cumulative incidence with the Aalen-Johansen estimator to offer transparency about competing risk impact.

  3. Improved AI system can identify scenarios where method choice substantially impacts trial interpretation, specifically when competing event rates exceed 5% annually combined with highly discordant treatment effects (theta2 1.0).

  4. Improved AI system can determine that neither the Cox model nor the Fine-Gray model consistently recovers the true marginal hazard ratio under most scenarios with non-negligible competing risks, such as when "treatment reduces competing events more strongly than primary events (theta2 < 0.7), both Cox and Fine-Gray show modest positive bias."

  5. Improved AI system can prioritize pre-specified analytical approaches, as the paper states that pre-specified Cox models should not be abandoned in favor of competing risk methods and that pre-specified Cox models should not be abandoned post-hoc simply because competing risks appear relevant after unblinding data.

  6. Improved AI system can distinguish between dependent and independent competing risks to decide when to use specialized modeling, noting that When competing events are truly independent, standard Cox models are appropriate regardless of competing event rates.

  7. Improved AI system can implement alternative methods like inverse probability of censoring weighting (IPCW) or multistate models for complex analyses when seeking the marginal causal treatment effect, addressing the limitations where neither the Cox model nor the Fine-Gray model consistently recovers the true marginal hazard ratio.

Related papers