TL;DR
This paper tackles the challenge of analyzing time-to-event outcomes in clinical trials when "intercurrent events" (like a patient stopping treatment or experiencing a related but different event) occur. It provides a rigorous framework for defining, estimating, and testing treatment effects under five strategies proposed by the ICH E9 (R1) guideline, making these complex concepts applicable to time-to-event data.
What Problem Does This Paper Address?
In clinical trials with time-to-event outcomes (e.g., time to disease recurrence, time to death), unexpected events can complicate the interpretation of treatment effects. These "intercurrent events" can either modify the risk of the primary outcome (semi-competing events) or prevent the outcome from occurring altogether (competing events). The ICH E9 (R1) guideline offers five strategies to handle these events, but their direct application to time-to-event data for causal inference has been challenging. This paper bridges that gap by developing methods to define, estimate, and test causal treatment effects under these five strategies specifically for time-to-event outcomes.
Methods
The authors extend the potential outcomes framework to time-to-event data, defining "causal estimands" for each of the five ICH E9 (R1) strategies:
- Treatment Policy Strategy: Compares treatment policies that include the natural occurrence of intercurrent events.
- Composite Variable Strategy: Aggregates the primary outcome and intercurrent event into a single composite outcome.
- While on Treatment Strategy: Measures outcomes only up to the point of an intercurrent event.
- Hypothetical Strategy: Envisions hypothetical scenarios where intercurrent events are manipulated (e.g., removed or controlled).
- Principal Stratum Strategy: Focuses on a subpopulation defined by their potential to experience or not experience intercurrent events.
They derive the mathematical formulations for these estimands and propose nonparametric estimation methods, including the calculation of asymptotic variances and the construction of hypothesis tests (specifically log-rank tests for some strategies).
Datasets
The methods are illustrated using data from the LEADER Trial, a large clinical trial investigating the effect of liraglutide on cardiovascular outcomes in patients with type 2 diabetes. This trial provides a real-world example of time-to-event outcomes (cardiovascular events) and intercurrent events (non-cardiovascular death).
Key Insights and Ideas
- Bridging the Gap: The paper successfully adapts the ICH E9 (R1) strategies, originally developed for simpler outcomes, to the complexities of time-to-event data, which inherently involves censoring and time-dependent risks.
- Causal Interpretation is Key: The focus is on defining estimands that allow for clear causal interpretations, moving beyond mere association.
- Strategy Matters: Each of the five strategies answers a distinct scientific question, leading to different estimands and requiring careful consideration of the research objective.
- Nonparametric Power: The proposed nonparametric methods offer flexibility and robustness, avoiding strong parametric assumptions.
Key Results
The paper demonstrates how to:
- Define the mathematical formulations for causal estimands under the five ICH E9 (R1) strategies for time-to-event outcomes.
- Develop nonparametric estimation methods and construct confidence intervals for these estimands.
- Implement hypothesis tests to assess treatment effects.
- Apply these methods to the LEADER Trial data, showing how different strategies yield different estimates of liraglutide's effect on cardiovascular outcomes, highlighting the importance of choosing the right strategy for the research question.
Limitations and Future Work
- Principal Stratum Identification: The paper notes that identifying the principal stratum estimand often requires strong, untestable assumptions (like monotonicity), and sensitivity analyses are recommended.
- Extension to Observational Studies: While the paper focuses on randomized controlled trials (RCTs), the supplementary material discusses extensions to observational studies, which introduces further complexities like covariate adjustment and propensity score methods.
- Model Misspecification: The authors acknowledge that semiparametric estimators can still be vulnerable to model misspecification, and robust estimation techniques are explored.
- Future Work: The authors suggest extending their framework to accommodate more complex scenarios, such as multiple intercurrent events or multiple endpoints, and exploring semiparametric estimation approaches for greater efficiency and robustness.
Summary updated