THE INHERENT BIAS FROM USING PARTITIONED SURVIVAL MODELS IN ECONOMIC EVALUATION
Author(s)
Coyle D1, Coyle K2
1University of Ottawa, Ottawa, ON, Canada, 2Brunel University, Uxbridge, ON, Canada
OBJECTIVES: Increasingly, economic evaluations of progressive diseases have adopted the approach of partitioned survival analysis. In three state cancer models (pre-progression, post-progression and death) the proportion of patients in each state are often obtained from survival functions for progression free (PFS) and overall survival (OS). The proportion in the pre-progression state is estimated from the PFS curve and the proportion in the post-progression state is the difference between OS and PFS. Based on simulated data reflecting the results from recent randomized clinical trials, this study explored the accuracy of this method. METHODS: Three clinical scenarios were considered based on varying OS. Clinical trial data sets were simulated for a standard and a novel treatment assuming substantive benefit with the new treatment in terms of slowing progression but no impact of treatment on mortality rates within health states. Costs and utility values for each state were assumed and represent plausible values. Analysis identified the difference between actual results based on complete follow up and results based on curtailed follow up of various durations using both traditional Markov modelling and partitioned survival analysis RESULTS: In the moderate survival scenario (median OS ~12 months), the ICUR for the new treatment based on the raw simulated data was $119,600. With trial follow up of 9 months the ICUR for new treatment was $120,600 with Markov modelling and $87,500 with partitioned survival analysis. With 18 months follow up, the figures were $122,500 and $103,000 respectively. Results with different durations of follow up found a consistent pattern as did results for both the short and long term survival scenarios. CONCLUSIONS: Analyses based on partitioned survival analysis have an inherent bias in favour of treatments which impact disease progression, not within health state mortality. They should not be considered an appropriate basis to facilitate reimbursement decisions.
Conference/Value in Health Info
2014-05, ISPOR 2014, Palais des Congres de Montreal
Value in Health, Vol. 17, No. 3 (May 2014)
Code
PRM74
Topic
Methodological & Statistical Research, Study Approaches
Topic Subcategory
Confounding, Selection Bias Correction, Causal Inference, Modeling and simulation
Disease
Oncology