Tian Yuan, Hassmiller Lich Kristen, Osgood Nathaniel D, Eom Kirsten, Matchar David B
Program in Health Services & Systems Research, Duke-NUS Graduate Medical School Singapore, Singapore (YT, KE, DBM)
Department of Health Policy & Management, University of North Carolina at Chapel Hill, NC, USA (KHL)
Med Decis Making. 2016 Nov;36(8):1043-57. doi: 10.1177/0272989X16643940. Epub 2016 Apr 18.
As health services researchers and decision makers tackle more difficult problems using simulation models, the number of parameters and the corresponding degree of uncertainty have increased. This often results in reduced confidence in such complex models to guide decision making.
To demonstrate a systematic approach of linked sensitivity analysis, calibration, and uncertainty analysis to improve confidence in complex models.
Four techniques were integrated and applied to a System Dynamics stroke model of US veterans, which was developed to inform systemwide intervention and research planning: Morris method (sensitivity analysis), multistart Powell hill-climbing algorithm and generalized likelihood uncertainty estimation (calibration), and Monte Carlo simulation (uncertainty analysis).
Of 60 uncertain parameters, sensitivity analysis identified 29 needing calibration, 7 that did not need calibration but significantly influenced key stroke outcomes, and 24 not influential to calibration or stroke outcomes that were fixed at their best guess values. One thousand alternative well-calibrated baselines were obtained to reflect calibration uncertainty and brought into uncertainty analysis. The initial stroke incidence rate among veterans was identified as the most influential uncertain parameter, for which further data should be collected. That said, accounting for current uncertainty, the analysis of 15 distinct prevention and treatment interventions provided a robust conclusion that hypertension control for all veterans would yield the largest gain in quality-adjusted life years.
For complex health care models, a mixed approach was applied to examine the uncertainty surrounding key stroke outcomes and the robustness of conclusions. We demonstrate that this rigorous approach can be practical and advocate for such analysis to promote understanding of the limits of certainty in applying models to current decisions and to guide future data collection.
随着卫生服务研究人员和决策者使用模拟模型解决更棘手的问题，参数数量及相应的不确定性程度都有所增加。这常常导致人们对这类复杂模型用于指导决策的信心下降。
展示一种将敏感性分析、校准和不确定性分析相联系的系统方法，以增强对复杂模型的信心。
将四种技术整合并应用于一个针对美国退伍军人的系统动力学中风模型，该模型旨在为全系统干预和研究规划提供信息：莫里斯方法（敏感性分析）、多起点鲍威尔爬山算法和广义似然不确定性估计（校准）以及蒙特卡罗模拟（不确定性分析）。
在60个不确定参数中，敏感性分析确定了29个需要校准的参数、7个虽无需校准但对关键中风结果有显著影响的参数，以及24个对校准或中风结果无影响的参数，这些参数被设定为最佳猜测值。获得了1000个经过良好校准的替代基线，以反映校准的不确定性，并将其纳入不确定性分析。退伍军人的初始中风发病率被确定为最具影响力的不确定参数，应针对该参数收集更多数据。即便如此，考虑到当前的不确定性，对15种不同预防和治疗干预措施的分析得出了一个可靠的结论，即对所有退伍军人进行高血压控制将在质量调整生命年方面带来最大收益。
对于复杂的医疗保健模型，采用了一种混合方法来检验围绕关键中风结果的不确定性以及结论的稳健性。我们证明这种严谨的方法具有实用性，并倡导进行此类分析，以促进对将模型应用于当前决策时确定性限度的理解，并指导未来的数据收集工作。
