Converging or Crossing Curves: Untie the Gordian Knot or Cut it? Appropriate Statistics for Non-Proportional Hazards in Decitabine DACO-016 Study (AML)

Abstract

Introduction: Among patients with acute myeloid leukemia (AML), the DACO-016 randomized study showed reduction in mortality for decitabine [Dacogen® (DAC), Eisai Inc., Woodcliff Lake, NJ, USA] compared with treatment choice (TC): at primary analysis the hazard ratio (HR) was 0.85 (95% confidence interval 0.69-1.04; stratified log-rank P=0.108). With two interim analyses, two-sided alpha was adjusted to 0.0462. With 1-year additional follow-up the HR reached 0.82 (nominal P=0.0373). These data resulted in approval of DAC in the European Union, though not in the United States. Though pre-specified, the log-rank test could be considered not optimal to assess the observed survival difference because of the non-proportional hazard nature of the survival curves. Methods: We applied the Wilcoxon test as a sensitivity analysis. Patients were randomized to DAC (N=242) or TC (N=243). One-hundred and eight (44.4%) patients in the TC arm and 91 (37.6%) patients in the DAC arm selectively crossed over to subsequent disease modifying therapies at progression, which might impact the survival beyond the median with resultant converging curves (and disproportional hazards). Results: The stratified Wilcoxon test showed a significant improvement in median (CI 95%) overall survival with DAC [7.7 (6.2; 9.2) months] versus TC [5.0 (4.3; 6.3) months; P=0.0458]. Conclusion: Wilcoxon test indicated significant increase in survival for DAC versus TC compared to log-rank test. Funding: Janssen-Cilag GmbH.