Robust Classification Methods for Categorical Regression

描述（由申请人提供）：改进统计方法，为分类回归提供更好的分类性能和新的分析能力，这对医学和卫生保健研究界来说是非常宝贵的。分类回归模型（二元逻辑，多项逻辑）被广泛用于识别酒精相关症状的模式，定义精神疾病的标准，并评估酒精管制政策。然而，许多这样的模型是在没有充分的自动化支持的情况下开发的，无法充分分析和利用其结果的内在概率性质。这是至关重要的，因为研究人员、临床医生和卫生保健管理人员多次面临使用分类回归模型的分类决策，以i）识别高风险个体或群体，ii）进行临床评估，或iii）制订政策及治疗指引市面上可购得的统计软件并无提供自动化程序，以系统地估计及测试决策门槛的稳健性对于分类回归环境中的分类此外，完全忽略了在决策阈值上估计鲁棒置信区间、比较竞争分类器或评估分类器误指定的存在的能力。 Martingale Research将开发统计分析工具，以提供自动化支持，专门解决分类回归建模的分类方面。本I期研究将使用代表NIAAA数据库的数据集证明，所提出的统计方法将1）估计分类决策阈值，2）提供决策阈值的稳健置信区间，3）应用高级模型选择测试来比较竞争分类器并分析分类器质量。这些结果将证明进一步的第二阶段调查所需的基本技术可行性，并为开发商业软件提供基础。

DESCRIPTION (provided by applicant): Improving statistical methods to provide better classification performance and new analytical capabilities for categorical regression would be invaluable to the medical and health care research communities. Categorical regression models (binary logistic, multinomial logistic) are used extensively to identify patterns of alcohol-related symptoms, define criteria of psychiatric disorders, and assess policies regulating alcohol. However, many such models are developed with inadequate automated support to fully analyze and exploit the intrinsically probabilistic nature of their results. This is of critical importance as researchers, clinicians, and health-care administrators are many times faced with classification decisions using categorical regression models to i) identify high risk individuals or groups, ii) make clinical assessments, or iii) establish policy and treatment guidelines Commercially available statistical software provides no automated procedures to systematically estimate and test the robustness of decision threshold(s) for classification within the context of categorical regression Moreover, the capability to estimate robust confidence intervals on decision threshold(s), compare competing classifiers, or assess the presence of classifier misspecification is completely ignored. Martingale Research will develop statistical analysis tools to provide automated support that specifically addresses the classification aspects of categorical regression modeling. This Phase I study will demonstrate using datasets representative of NIAAA databases that the proposed statistical approach will 1) estimate classification decision threshold(s), 2) provide robust confidence intervals on decision threshold(s), and 3) apply an advanced model selection test for comparing competing classifiers and analyzing classifier quality. These results will demonstrate the essential technical feasibility required for further Phase II investigation and provide the foundation for developing commercially available software.