Bayesian Methods for Latent Class Models

我的研究计划集中在开发贝叶斯方法来估计潜在类别模型。这些模型的动机是涉及疾病诊断测试或不存在完美测试的条件的问题。这是许多常见疾病的情况，例如肺炎、儿童结核病、前列腺癌。缺乏一种完美的检测方法给研究人员带来了方法学上的挑战，研究人员对测量这种疾病的患病率或评估这种疾病的诊断测试感兴趣。目前的提案描述了将在三个不同主题下开发的这种研究的新方法：*i)在单个研究的背景下进行潜在类别分析的方法：*在该主题下，我们将研究比较竞争的潜在类别模型的方法、估计新测试的增量价值的方法以及确定在实践中使用的最佳测试序列的方法。*ii)评估诊断测试所需样本大小的方法：*在该主题下，我们将描述设计新研究的方法，其中新测试相对于旧测试的增量值是感兴趣的统计量。我们还将开发设计研究的方法，其中感兴趣的结果是可靠性而不是效度。*iii)在潜在类别元分析模型中建模条件依赖和测量异质性的方法：*在这个主题下，我们将描述在元分析环境中对条件依赖进行建模的方法，以及在元分析中报告研究间异质性的方法。*将始终使用贝叶斯推理框架。对于拟议的研究，这种方法的一个特别优点是，它可以用于估计当观察到的不完美测试的数量很少(少于三到四个)时经常出现的不可识别模型。在这种情况下，需要用关于未知参数子集的先验信息(例如，已建立但不完美的诊断测试的灵敏度和特异度)来扩充观察数据。此外，贝叶斯方法在概念上简单易懂，适用于上述主题下将出现的复杂模型。*实施贝叶斯方法将涉及使用蒙特卡洛马尔可夫链方法。传统上，推广贝叶斯方法的一个重要障碍是缺乏软件。因此，将开发用户友好的软件，以配合每一种拟议的方法。每个主题将产生几个适合研究生或博士后研究员的子项目。这里描述的方法也将在其他出现误测数据的领域得到应用。

My research program is focused on the development of Bayesian methods for estimation of latent class models. These models are motivated by problems involving diagnostic tests for diseases or conditions for which no perfect test exists. This is the case for many common diseases e.g. pneumonia, tuberculosis in children, prostate cancer. The lack of a perfect test presents a methodological challenge to researchers interested in measuring the prevalence of the disease or in the evaluating diagnostic tests for the disease. The current proposal describes new methods for such studies that will be developed under three different themes: ******i) Methods for latent class analysis in the context of a single study: ***Under this theme, we will study methods for comparing competing latent class models, methods for estimating the incremental value of a new test and methods for determining the optimal sequence of tests to be used in practice.******ii) Methods for evaluating the sample size required for evaluating diagnostic tests: ***Under this theme, we will describe methods for designing a new study where the incremental value of a new test over an older test is the statistic of interest. We will also develop methods for designing studies where the outcome of interest is reliability rather than validity.******iii) Methods for modeling conditional dependence and measuring heterogeneity in latent class meta-analysis models: ***Under this theme, we will describe methods for modeling conditional dependence in the context of a meta-analysis and methods for reporting the between-study heterogeneity in a meta-analysis.******A Bayesian inferential framework will be used throughout. One particular advantage of this approach for the proposed research is that it can be used for estimation of the non-identifiable models that frequently arise when the number of observed, imperfect tests is small (less than three or four). In such cases, the observed data needs to be augmented with prior information on a subset of the unknown parameters (e.g. the sensitivity and specificity of a well established but imperfect diagnostic test). Further, the Bayesian approach is conceptually straightforward to apply to the complex models that will arise under the above themes. ******Implementation of the Bayesian methods will involve use of Monte Carlo Markov Chain methods. Traditionally, an important impediment in disseminating Bayesian methods has been the lack of software. Therefore, user-friendly software will be developed to accompany each of the proposed methods. Each theme will give rise to several sub-projects suitable for graduate students or post-doctoral fellows. The methods described here will also find application in other areas where mis-measured data arise.