关于传染性负二项发病率的贝叶斯推断.pdfVIP

摘 要 众所周知，在流行病学领域中，当所研究疾病的发病率很小时，传统方法 为采用负二项分布即逆抽样方法对发病率进行研究，并假定发病率固定不变。 负二项分布通常描述相互独立事件，而在实际的传染性事件中，每个事件 的发生或多或少都会影响下一事件发生的概率。如果忽略了这一点，那么对现 实数据的分析结果可能失真，这是因为传染性因素会导致后面的发病率变动， 这样，传统方法不能够有效地处理这种情况，所以在研究具有传染性的疾病蔓 延时，我们通常需要考虑疾病的蔓延性这个特性。 在这种情况下，需要将发病率考虑为一个随机变量。文献中通常采用分层 贝叶斯或方差成分模型对传染性疾病的发病率进行分析。 传统的方法采用标准 beta分布作为二项及负二项发病率的先验，但标准beta分布在拟合上不稳定。 故本文将采用即拥有 beta 的基本性质又拥有较好的拟合稳定性的广义第二类 beta分布作为发病率的先验，从而结合负二项分布构建一个新的传染性负二项 分布，而后对发病率的后验分布做相应的统计推断。然后，运用 Monte Carlo 来 研究我们提出的新方法，最后将本文的理论及方法应用到实际数据的分析中去。 关键词：传染性负二项分布；广义第二类Beta分布；贝叶斯推断；拟合优 度检验 Abstract It is well known that in the area of epidemiology, when the incidence of a disease is rare, the conventional method to deal with such situation is to make use of negative binomial distribution or inverse sampling, and the probabilities of the occurrence of some diseases are usually assumed to be constant. While considering the situation that if the disease is contagious, the probability may vary from time to time, due to some reasons, which make the classical technique couldn’t solve it effectively. Under this situation, we need to consider the probability as a variable with a distribution. Usually, the bayesian analysis of hierarchical models or variance component models was employed in the literature. Traditionally，we always use the standard beta distribution as a prior distribution for the binomial or negative binomial model. Although it has some advantages, the distribution is generally fast rising with very thin tail. This causes some problems in fitting some sets of prior beliefs and without robustness. Also, the degree of skewness is highly dependent on the mean. The generalized beta distribution of II-kind distribution not only has the same advantages as standard bet