Article type
Year
Abstract
Background: There appears to exist a tension between evidence-based medicine (EBM) and decision analysis (DA), the two fields devoted to rational decision making. EBM decision making insists on using reliable empirical evidence at the bedside and in policy making. DA, however, typically models decision making over the time horizon beyond empirically available evidence (usually over a lifetime). A little work exists assessing whether there is a discrepancy in the two approaches.
Methods: We took advantage of a recently published individual patient data meta-analysis by the Early Breast Cancer Trialists' Collaborative Group (EBCTG)1 which regularly updates their results at five year intervals. The Markov model was built to compare empirical results with those modelled beyond the time horizon for which data were available. We first modelled breast cancer mortality rates comparing the effects of tamoxifen vs. no treatment. We then modelled the 15-year probability of death from breast cancer using five and 10 year data reported in the overview. Finally, we estimated the effect on life expectancy in 40 to 70 year old women with breast cancer.
Results: Verification of the model produced identical empirical five, 10 and 15 year data reported in the EBCTG overview. Modelling 15 years survival based on five and 10 year empirical data resulted in an average error estimate in breast cancer mortality of 0.75% [range: -0.83 to 2.19%]. Although modelling of differences in life expectancies ranged from an underestimate of -7.93% to an overestimate of 12.64%, in absolute terms this means that over the span of 15 years, there is a loss of 18 days or a gain of 40 days of life, respectively. Regardless of the assumptions used in the model, the decision remained stable: tamoxifen remained the better choice over no therapy.
Conclusions: We conclude that reliable data generated at five or 10 years may be used to produce credible results at 15 years. Whether five year data can be employed over the lifetime horizon remains to be demonstrated.
References
1. Early Breast Cancer Trialists' Collaborative Group (EBCTG). Effects of chemotherapy and hormonal therapy for early breast cancer on recurrence and 15-year survival: an overview of the randomised trials. Lancet 2005; 365:1687-1717.
Methods: We took advantage of a recently published individual patient data meta-analysis by the Early Breast Cancer Trialists' Collaborative Group (EBCTG)1 which regularly updates their results at five year intervals. The Markov model was built to compare empirical results with those modelled beyond the time horizon for which data were available. We first modelled breast cancer mortality rates comparing the effects of tamoxifen vs. no treatment. We then modelled the 15-year probability of death from breast cancer using five and 10 year data reported in the overview. Finally, we estimated the effect on life expectancy in 40 to 70 year old women with breast cancer.
Results: Verification of the model produced identical empirical five, 10 and 15 year data reported in the EBCTG overview. Modelling 15 years survival based on five and 10 year empirical data resulted in an average error estimate in breast cancer mortality of 0.75% [range: -0.83 to 2.19%]. Although modelling of differences in life expectancies ranged from an underestimate of -7.93% to an overestimate of 12.64%, in absolute terms this means that over the span of 15 years, there is a loss of 18 days or a gain of 40 days of life, respectively. Regardless of the assumptions used in the model, the decision remained stable: tamoxifen remained the better choice over no therapy.
Conclusions: We conclude that reliable data generated at five or 10 years may be used to produce credible results at 15 years. Whether five year data can be employed over the lifetime horizon remains to be demonstrated.
References
1. Early Breast Cancer Trialists' Collaborative Group (EBCTG). Effects of chemotherapy and hormonal therapy for early breast cancer on recurrence and 15-year survival: an overview of the randomised trials. Lancet 2005; 365:1687-1717.