Estimating Transition Probabilities From Published Oncology Survival Data: Comparison of Methods and Impact on Cost-Effectiveness

OBJECTIVES: Economic evaluations, conducted using state-transition models, rely on probabilities that govern movement between health states. When individual-level patient data (IPD) are not available to estimate transition probabilities, alternative methods using published Kaplan-Meier curves of overall survival (OS) and progression-free survival (PFS) are often used. However, the structural relationship between disease progression and survival is often not accounted for. We compared methods to estimate transition probabilities from published survival data to determine their impact on health-economic outcomes and decision making.

METHODS: Utilizing publicly-available simulated IPD, based on an oncology data set, we recreated summary OS and PFS data. We report relative differences in total costs, quality-adjusted life years (QALYs) and two- and five-year mortality, applying a three-state Markov model with states progression-free, progression, and death. Benchmark transition probabilities were calculated from IPD using competing risks, and compared with four estimation methods: 1) fitting parametric curves to OS and PFS for all transitions, 2) assuming natural mortality for transitions from progression-free to death, 3) estimating exponential transition rates using parameterized survival curves and regenerated IPD, 4) competing risks with regenerated IPD. We present an algorithm to approximate linked IPD from mutually exclusive Kaplan-Meier curves, utilized in methods 3 and 4. We discuss the strengths and weaknesses of this algorithm, along with each estimation method.

RESULTS: Costs, QALYs, and 5-year mortality were closest to the benchmark analysis for method 4 (relative differences 6.9%, 7.7%, and -13.8% respectively), 2-year mortality was closest for method 3 (-17.4%), and method 2 performed the worst (costs: 15.0%, QALYs: 15.7%, 2-year mortality: -73.1%, 5-year mortality: -36.8%).

CONCLUSIONS: Without IPD, alternative methods for deriving transition probabilities result in varying bias to cost-effectiveness outcomes. Careful consideration is required for each of the outlined methods, and sensitivity and scenario analyses would help provide insights into decision uncertainty when IPD are not available