Statistical Methods for Ordinal Variables in HIV/AIDS Studies

描述（由申请人提供）：HIV/AIDS生物医学研究中的许多重要变量是有序分类的。一些例子包括世卫组织临床阶段、教育水平、冲洗频率、宫颈病变阶段、自我报告使用安全套和双等位基因型。虽然序数变量很常见，但缺乏解释其有序性质的统计方法，特别是当序数变量是预测变量时。大多数标准方法要么将有序预测器视为分类（忽略顺序信息），要么将其视为连续（进行线性假设）。本建议发展统计方法，说明有序变量的有序性质，而不做线性假设。这些方法解决了预测变量(X)是有序分类的情况；结果变量(Y)是连续的、离散的、计数的、时间到事件的或重复的测量；有多个协变量(Z)。一般的方法是拟合适当的Y对Z和X对Z的回归模型，然后检验这两个模型的残差之间的相关性。因此，该方法依赖于有序分类数据的残差的新定义。评估了该残差的统计特性，以及它在模型诊断中的应用。计算了基于残差的检验统计量的渐近性质，推导了与其他方法的关系，开发了实现这些方法的用户友好软件，并通过模拟和实际数据研究了新方法的优点。这些方法被应用于两项HIV研究：第一项研究使用边缘结构模型评估了冲洗频率对一组青春期女性性传播感染的影响。第二个数据应用程序寻找与药物血浆水平、病毒学失败和开始以依非韦伦或阿巴卡韦为基础的抗逆转录病毒治疗方案的患者毒性相关的人类遗传多态性。

DESCRIPTION (provided by applicant): Many important variables in biomedical studies of HIV/AIDS are ordered categorical. A few examples include WHO clinical stage, level of education, frequency of douching, stage of cervical lesions, self- reported condom use, and biallelic genotypes. Although ordinal variables are common, statistical methods that account for their ordered nature are lacking, particularly when the ordinal variable is a predictor variable. Most standard methods either treat the ordinal predictor as categorical (ignoring the order information) or continuous (making linearity assumptions). This proposal develops statistical methods that account for the ordered nature of ordinal variables without making linearity assumptions. The methods address situations when a predictor variable (X) is ordered categorical; the outcome variable (Y) is continuous, discrete, counts, time-to-event, or repeated measures; and there are multiple covariates (Z). The general approach is to fit appropriate regression models of Y on Z, and X on Z, and then to test for correlation between the residuals from these two models. The methods therefore rely on a new definition of residual for ordered categorical data. Statistical properties of this residual are evaluated, as well as its use in model diagnostics. Asymptotic properties of the residual- based test statistics are computed, relationships with other methods are derived, user-friendly software that implements these methods is developed, and the advantages of the new methods are studied using simulated and real data. The methods are applied to two HIV studies: The first assesses the effect of the frequency of douching on sexually transmitted infections among a cohort of adolescent females using marginal structural models. The second data application looks for human genetic polymorphisms associated with drug plasma levels, virologic failure, and toxicities for patients initiating an efavirenz- or abacavir-based antiretroviral regimen. PUBLIC HEALTH RELEVANCE: The new methods are put into computer software and applied to studies of 1) the effect of douching on sexually transmitted infections among adolescent females, and 2) which patients may be able to better respond to specific HIV-treatments based on genetic patterns.