A Multivariate Probit Model for Health Services Research

描述(由申请人提供)：精神健康服务研究的一个中心主题是服务利用模式的特征及其决定因素的确定。例如，医疗保健提供系统的变化对住院、门诊、急诊室和家庭精神卫生保健服务的利用的影响引起了相当大的兴趣。虽然在这一领域已经进行了大量的统计工作(见Gibbons 2001的概述)，但很少有工作(如果有的话)纳入卫生服务研究数据的纵向和多变量性质。虽然一些研究人员提出了在横断面数据中分析服务使用的多变量模式的方法，另一些人提出了纵向方法来分析随时间重复测量的单一服务使用情况，但在纵向数据分析方面所做的工作很少；所有测量的多变量服务利用数据。简单的零散单变量分析忽略了服务利用数据的相关性和往往是补偿性的(例如，导致急诊室使用量减少的卫生保健系统变化可能导致门诊治疗使用量的增加)。在这项研究中，我们将建立一个通用的混合效应多变量Probit回归模型，用于同时分析重复测量的多变量二进制数据。将单个时间点多个二元测量之间的相关性建模为因子分析过程，将重复测量之间随时间的相关性建模为随机效应过程。最终结果是，我们现在可以模拟设计变量(例如，医疗保健提供系统的变化)和病例组合变量(例如，年龄、性别和种族)对多变量利用模式的影响。该模式的概括将包括，顺序响应数据的扩展(例如，不使用、适度使用、中等使用、高使用或O访问、I、访问、2访问、3或更多访问)、离散和连续响应的混合(例如，服务利用和成本的联合分析)，以及对多变量Logistic回归模型的扩展。该项目的一个组成部分将是探索和开发用于似然估计(定点和自适应求积、拉普拉斯近似和蒙特卡罗积分)、参数估计(牛顿·拉夫森、费舍尔评分和EM算法)和假设检验的替代方法。将进行大规模的模拟研究，以研究一般模型和各种替代公式的统计特性。最后，该模型将用于分析波多黎各大学Margarita Aregria博士收集的关于医疗改革对纵向精神卫生服务利用的影响的数据。除了发展统计理论和估计程序外，我们还建议开发一个基于WINDOWS的免费软件计算机程序MIXMVP，该程序将从MIXREG/MIXOR网站www.uic.edu/Labs/BioStat分发。目前还没有通用的多变量概率回归软件。

DESCRIPTION (provided by applicant): A central theme in mental health services research is the characterization of patterns of service utilization and the identification of their determinants. For example, there is considerable interest in the effects of changes in the health care delivery system on utilization of inpatient, outpatient, emergency room, and home mental health care services. Although considerable statistical worlc in this area has been conducted (see Gibbons 2001 for an overview), little work, ifany, has been done to incorporate both the longitudinal and multivariate nature of health services research data. While some investigators have proposed approaches to the analysis of multivariate patterns of service use in cross-sectional data, and others have proposed longitudinal approaches for the analysis of a single service use repeatedly measured over time, very little if any work has been done on the analysis of longitudinall; measured multivariate service utilization data. Simple piecemeal univariate analyses ignore the correlated and often compensatory nature of service utilization data (eg. health care system changes that lead to decreases in emergency room use may lead to increase in outpatient treatment use). In this study, we will develop a general mixed-effects multivariate probit regression model for the simultaneous analysis of repeatedly measured multivariate binary data. Correlations between multiple binary measures at a single point in time are modeled as a factor analytic process, and correlation among the repeated measurements over time are modeled as : random-effects process. The net result is that we can now model the effects of design variables (e.g, changes in the health care delivery system) and case mix variables (e,g., age, sex, and race) on multivariate utilization patterns. Generalizations of the mode will include, extension to ordinal response data (e.g., no use, mild use, moderate use, high use, or O visits, I, visit, 2 visits, 3 or more visits), mixtures of discrete and continuous responses (e.g., the joint analysis of service utilization and cost), and extension to a multivariate logistic regression model. An integral part of the project will be to both explore and develop alternative approaches to likelihood evaluation (fixed-point and adaptive quadrature, Laplace approximation, and Monte Carlo integration), parameter estimation (Newton Raphson, Fisher scoring, and the EM algorithm), and hypothesis testing. A large-scale simulation study will b conducted to study the statistical properties of the general model and various alternative formulations. Finally, the model will be applied in the analysis of data collected by Dr. Margarita Alegria at the University of Puerto Rico on the effects of health care reform on longitudinal mental health services utilization. In addition to development of the statistical theory and estimation procedure, we propose to develop a WINDOWS based freeware computer program, MIXMVP, to be distributed from the MIXREG/MIXOR web site www.uic.edu/labs/biostat. No general multivariate probit regression software is currently available.