A Bayesian decision-theoretic framework to evaluate and optimize decision making for mastitis control in the UK Mastitis Control Scheme.

奶牛乳房炎是奶牛最主要的地方性传染病，对英国和世界乳业来说仍然是一个重大挑战。虽然已经有许多研究报告了奶牛和牛群患乳房炎的风险因素，但还没有研究评估在不同养殖场环境下的最佳决策；这是乳房炎控制中的一个关键未知因素。这项研究的目的是使用贝叶斯决策分析框架来调查这样一种假设，即在给定的农场环境下，通过采用“最佳策略”可以改善乳房炎的控制。为了进行这项研究，可以从我们的工业合作伙伴DairyCo开发的最近推出的全国乳房炎控制计划-DairyCo乳房炎控制计划(www.mastitiscontrol plan.co.uk)获得详细数据。我们将使用来自至少600个牧场的丰富数据，包括详细的牛群规模、牧场设施和人力、管理干预以及临床和亚临床型乳房炎的详细记录。在最初的统计分析中，样本量估计表明，临床乳房炎和亚临床乳房炎5%的差异将在个别或组管理干预措施中检测到，这被认为是临床上重要的效果大小。将构建一个贝叶斯决策分析框架，使国家计划数据的初始多变量数据分析与经济和生产信息相结合，以明确表示与乳房炎控制相关的决策过程。贝叶斯框架提供了一种结构，可以综合多种信息源，也可以评估不同干预措施的成本效益中的不确定性(风险)。这种方法通常被称为概率敏感性分析，现在国家临床卓越研究所(NICE)要求对人类医疗干预措施进行评估，但很少用于动物健康。在第一年，将对全国乳房炎控制计划中600个养殖场的数据进行整理和初步分析。在第二年，将构建贝叶斯框架，以量化不同奶牛乳房炎预防策略的相对重要性和不确定性。我们将具体预测不同干预措施(和干预措施组)在不同农场环境下对每种策略的‘增量净收益’(净财务回报)的影响。因此，鉴于乳房炎的养殖场模式(国家计划中的养殖场根据单位上的主要乳房炎模式被分成四类)和一套可能的干预措施，我们将预测最佳预防策略和相关的财务回报的不确定性。我们将使用单一的集成贝叶斯程序，使用马尔科夫链蒙特卡罗，这具有允许所有联合参数不确定性通过模型传播的数学优势，并且很重要，因为评估与不同决策相关的不确定性是研究的关键领域。在第三年，模型的预测价值将使用模型内和模型外的后验预测进行评估(使用国家计划中的新农场)，在第三年的第二个6个月，学生将与产业合作伙伴一起工作。在第4年，决策模型的结果将被合并到DairyCo国家乳房炎控制计划软件中(与DairyCo合作伙伴QMMS有限公司合作)，以便为农场决策提供信息。该软件将允许计划参与者确定对特定农场最有可能提供最大健康和经济利益的管理干预措施，并将确保研究对改善英国奶牛的乳房炎产生立竿见影的效果。

Bovine mastitis is the foremost endemic infectious disease of dairy cattle and remains a major challenge to the UK and worldwide dairy industries. Although there have been numerous studies reporting cow and herd risk factors for bovine mastitis, no research has been conducted to evaluate optimal decision making in different farm circumstances; this is a critical unknown element in mastitis control. The purpose of this research is to use a Bayesian decision analytic framework to investigate the hypothesis that mastitis control can be improved by adopting a 'best strategy' in given farm circumstances. To conduct this research, detailed data are available from a recently launched national mastitis control scheme, the DairyCo Mastitis Control Plan (www.mastitiscontrolplan.co.uk) that has been developed by our industrial partner DairyCo. We will use the rich data from at least 600 farms in the scheme which includes detailed information on herd size, farm facilities and manpower, management interventions and detailed records of clinical and subclinical mastitis. For the initial statistical analysis, sample size estimates indicate that differences in clinical and subclinical mastitis of 5% will be detectable for individual or groups of management interventions, and this is deemed to be a clinically important effect size. A Bayesian decision analytic framework will be constructed to allow synthesis of the initial multivariable data analysis of the National Scheme data with economic and production information, to explicitly represent the decision process associated with mastitis control. The Bayesian framework provides a structure that will allow synthesis of multiple sources of information and also for uncertainty (risk) to be evaluated in the cost benefit of different interventions. Such an approach, often termed probabilistic sensitivity analysis, is now required by the National Institute of Clinical Excellence (NICE) for the evaluation of human medical interventions, but is rarely used in animal health. In Year 1, collation and initial analysis of the data from 600 farms in the National Mastitis Control Scheme will occur. In Year 2, the Bayesian framework will be constructed to quantify the relative importance of, and uncertainty in, different preventive strategies for bovine mastitis. We will specifically predict the consequences of different interventions (and groups of interventions), in different farm circumstances on the 'Incremental Net Benefit' (net financial return) of each strategy. Therefore, given farm patterns of mastitis (farms in the national scheme are grouped into four categories according to the predominant pattern of mastitis on the unit) and a set of possible interventions, we will predict the optimum prevention strategy and the associated uncertainty of a financial return. We will employ a single integrated Bayesian procedure, using Markov chain Monte Carlo, which has the mathematical advantage of allowing all joint parameter uncertainty to be propagated through the model and is important because evaluation of the uncertainty associated with different decisions is a key area of investigation. In Year 3, the predictive value of models will be evaluated using both 'within' and 'out of model' posterior predictions (using new farms in the national scheme) and in the second 6 months of Year 3, the student will work with the industrial partner. In Year 4, results from the decision models will be incorporated into the DairyCo National Mastitis Control Scheme software (in collaboration with DairyCo partner QMMS Ltd) to inform on farm decision making. This software will allow scheme participants to identify the management interventions that, for specific farms, are most likely to provide the greatest health and financial benefits, and will ensure that the research has an immediate impact to improve mastitis in UK dairy cows.