Introduction: The Challenge of Heterogeneity in Heart Failure Trials
In the landscape of clinical cardiology, heart failure (HF) remains one of the most complex conditions to study due to the vast diversity in patient presentations and clinical trajectories. Whether categorized by reduced ejection fraction (HFrEF) or preserved ejection fraction (HFpEF), patients within these cohorts exhibit significant risk heterogeneity. This variation in baseline risk—driven by comorbidities, age, and disease severity—poses a substantial challenge for biostatisticians and clinicians attempting to measure the true efficacy of new therapeutic interventions.
For decades, the Cox proportional hazards (PH) model has served as the gold standard for analyzing time-to-event outcomes in cardiovascular trials. However, emerging evidence suggests that in populations with high risk heterogeneity, the Cox model may provide biased estimates of treatment effects. A recent study by Myte et al., published in Circulation: Heart Failure, explores a compelling alternative: the survival proportional odds (PO) model. By applying this model to data from the landmark DAPA-HF and DELIVER trials, researchers have demonstrated its potential to minimize bias and improve the interpretation of treatment benefits in heart failure.
Background: Why Traditional Cox Models May Fall Short
The Bias of Selection in Risk-Heterogeneous Populations
The fundamental assumption of the Cox PH model is that the hazard ratio remains constant over time. While this assumption is often reasonable, it is frequently violated in trials involving heterogeneous populations. When a treatment is effective, high-risk individuals in the control group tend to experience events earlier than high-risk individuals in the treatment group. As the trial progresses, the control group becomes increasingly composed of lower-risk ‘survivors,’ while the treatment group still contains a mix of risk profiles.
This disproportionate reduction of high-risk patients in the control group leads to a phenomenon known as risk heterogeneity bias. It results in a diminishing hazard ratio over time, even if the biological effect of the drug remains constant. Consequently, the Cox model may underestimate the average treatment effect and lose statistical power, potentially masking the true value of a life-saving therapy.
Study Design and Methodology: Evaluating the Survival Proportional Odds Model
To address these limitations, Myte and colleagues evaluated the survival proportional odds (PO) model. Unlike the Cox model, which focuses on the ratio of instantaneous event rates, the PO model focuses on the odds of being event-free at any given time. This approach is inherently more robust to the shifting risk profiles of a trial population.
Data Sources: DAPA-HF and DELIVER
The researchers utilized clinical data from two pivotal trials of the SGLT2 inhibitor dapagliflozin:
1. DAPA-HF (Dapagliflozin in Patients with Heart Failure and Reduced Ejection Fraction): A population generally considered more homogeneous in terms of pathophysiological risk.
2. DELIVER (Dapagliflozin in Heart Failure with Mildly Reduced or Preserved Ejection Fraction): A population known for greater phenotypic diversity and risk heterogeneity.
Simulation and Reanalysis
The study employed a two-pronged approach. First, a simulation study was conducted to compare the performance of Cox regression and the survival PO model across varying degrees of risk heterogeneity. Second, a reanalysis of the DAPA-HF and DELIVER trials was performed to determine if the PO model would yield different insights into the efficacy of dapagliflozin.
Key Findings: Robustness and Power in DAPA-HF and DELIVER
Heterogeneity and Nonproportionality in HFpEF vs. HFrEF
The study confirmed that nonproportional hazards were a more significant issue in the DELIVER trial (HFpEF) than in the DAPA-HF trial (HFrEF). The inherent complexity of HFpEF—often involving older patients with multiple metabolic and renal comorbidities—creates a high-heterogeneity environment where traditional statistical models are most vulnerable to bias.
Simulation Results: Comparing PO and Cox Regression
In the simulation phase, the survival PO model demonstrated clear advantages in high-heterogeneity settings:
1. Robustness: The PO model was less affected by the bias introduced by risk heterogeneity compared to the Cox model.
2. Statistical Power: In populations with high heterogeneity, the PO model achieved higher statistical power. This means that for a given sample size, the PO model is more likely to detect a statistically significant treatment effect if one truly exists.
3. Stability: In more homogeneous populations, the PO model performed similarly to the Cox model, suggesting that there is little downside to using it as a primary or sensitivity analysis tool.
Reanalysis of Clinical Trial Data
When the researchers applied the PO model to the actual trial results, the findings were striking. In the DELIVER trial, the survival PO model provided consistently higher statistical power than the original Cox analysis. In the DAPA-HF trial, where heterogeneity was lower, the PO model and the Cox model yielded similar results. These findings reinforce the idea that the PO model is particularly valuable in the most challenging and diverse patient populations.
Clinical Implications: A New Lens for Interpreting Treatment Benefit
One of the most significant advantages of the survival PO model is its intuitive interpretation for clinicians. While a ‘hazard ratio’ can be difficult to explain to a patient, the ‘odds of being event-free’ is a direct and relatable metric. For example, a result from a PO model can be stated as: ‘Patients treated with this medication have 30% higher odds of remaining free from heart failure hospitalization or cardiovascular death compared to those on placebo.’
Furthermore, the robustness of the PO model ensures that the treatment effect reported at the end of a long-term trial is not artificially deflated by the early loss of high-risk patients in the control group. This provides a more accurate reflection of the drug’s long-term protective value.
Expert Commentary and Methodological Considerations
Methodologists have long argued for a move toward ‘estimands’ that are more resilient to the nuances of trial dynamics. The survival PO model aligns with this shift. However, experts note that the transition from Cox to PO models requires a change in how clinical trialists think about data.
While the PO model solves the problem of risk heterogeneity bias, it is not a panacea. It still requires careful consideration of model fit and the nature of the event being studied. Nevertheless, as the field of cardiology moves toward more personalized medicine and investigates increasingly diverse patient groups (such as those in HFpEF trials), the need for statistical tools that can handle heterogeneity is paramount.
Conclusion: Shaping the Future of Cardiovascular Trial Design
The study by Myte et al. provides a compelling case for integrating the survival proportional odds model into the standard toolkit of cardiovascular research. By demonstrating that this model is more robust to risk heterogeneity and provides higher statistical power in complex populations like HFpEF, the researchers have offered a pathway to more reliable and interpretable trial results.
As we look toward future trials in heart failure and other chronic cardiovascular conditions, the survival PO model should be considered not just as a sensitivity analysis, but as a viable alternative for primary outcome assessment. Its ability to provide a clear, intuitive, and statistically sound estimate of treatment benefit makes it a powerful asset in the quest to improve patient care through evidence-based medicine.
Funding and ClinicalTrial.gov
The original trials analyzed in this study were funded by AstraZeneca. The DAPA-HF trial is registered at ClinicalTrials.gov with the identifier NCT03036124, and the DELIVER trial is registered with the identifier NCT03619213.
References
Myte R, Mattsson A, Poole M, Little DJ, Nyström P, Henderson A, Claggett BL, Gasparyan SB, Solomon SD, McMurray JJV. Survival Odds to Minimize Risk Heterogeneity Bias in Heart Failure Trials: Application to Dapagliflozin. Circ Heart Fail. 2025 Dec;18(12):e013496. doi: 10.1161/CIRCHEARTFAILURE.125.013496. Epub 2025 Oct 31. PMID: 41170566; PMCID: PMC12704667.