Mixture Modeling of Patterns of Substance Use: Taking into Account Measurement Bias

 描述(申请人提供)：青春期药物滥用是一种流行病，对社会造成重大损失，并对青少年及其家庭产生有害的身体、心理和社会影响。因此，近年来的大量研究广泛地集中在寻找青春期药物滥用的社会和行为决定因素之间的模式，特别是根据药物滥用行为发现临床上相关的青少年亚群的目标。这一目标反映在最近的NIDA计划公告PA-15-003中，题为药物滥用流行病学；该公告列出了毒品使用的轨迹和与药物使用障碍的共病模式，这两个领域经常使用子分组策略进行探索，作为感兴趣的领域。形成有意义的行为亚型的目标在很大程度上是通过使用混合模型来实现的，混合模型是一类广泛的模型，它试图将样本划分为有意义的子组，即被试的潜在类别。在过去的十年里，混合模型在物质滥用研究中得到了广泛的使用，行为分组与成像数据、生物标记物和遗传信息有关。然而，尽管它们被广泛使用，但几乎所有当前混合模型的应用都没有考虑到行为测量的一个关键方面，即差异项目功能(DIF)。给定一项旨在测量某一基础结构的项目，当具有相同水平的基础结构的受试者对测量该结构的项目的反应因其 性别、种族、年龄或任何其他个人特征。当DIF发生并且不被考虑时，它可能会使对基础结构水平的估计有偏差，并且关于组之间差异的推断无效。尽管诸如项目反应理论(IRT)和验证性因子分析(CFA)等测量连续潜在变量的技术已经非常重视DIF作为对推理有效性的根本威胁，但DIF在混合模型中潜在类别的测量中还没有系统的研究，在物质滥用文献中也没有已知的DIF分析在混合模型结果中的应用。这项拟议的研究试图通过严格调查混合模型中的DIF来纠正这一缺陷，重点是药物滥用的行为研究。拟议项目的四个目标是：(1)分析确定DIF在混合模型中的表现方式；(2)开发混合模型中DIF的灵活测试；(3)使用计算机模拟来测试(A)未建模的DIF在混合模型中的后果和(B)目标2中开发的测试的有效性；以及(4)使用DIF建模技术对过渡到成人的药物滥用进行实证研究。通过对这些目标的追求，本项目将使混合模型能够更好地考虑个体差异，从而增强这些结果在行为药物滥用研究中的概括性。

 DESCRIPTION (provided by applicant): Substance abuse in adolescence is an epidemic which takes a significant toll on society and has deleterious physical, psychological, and social effects on adolescents and their families. As such, a great deal of research in recent years has focused broadly on finding patterns among social and behavioral determinants of substance abuse in adolescence, and specifically on the goal of finding clinically relevant subgroups of adolescents according to substance abuse behaviors. This goal is reflected by a recent NIDA program announcement, PA-15-003, entitled Epidemiology of Drug Abuse; this announcement lists trajectories of drug use and patterns of comorbidity with substance use disorders, two areas frequently explored using subgrouping strategies, as areas of interest. The goal of forming meaningful behavioral subtypes has largely been pursued using mixture models, a broad class of models which seeks to divide the sample into meaningful subgroups, known as latent classes, of subjects. Mixture models have seen extensive use in substance abuse studies in the past decade, with behavioral groupings being linked to imaging data, biomarkers, and genetic information. However, despite the their widespread use, virtually all current applications of mixture models have failed to take into account one critical aspect of behavioral measurement known as differential item functioning (DIF). Given an item intended to measure some underlying construct, DIF occurs when subjects with the same level of that underlying construct differ systematically in their responses to an item measuring that construct due to their gender, race, age, or any other individual characteristic. When DIF occurs and is not accounted for, it may render estimates of the level of the underlying construct biased, and inferences about differences between groups invalid. Though techniques for measuring continuous latent variables such as item response theory (IRT) and confirmatory factor analysis (CFA) have given serious attention to DIF as a fundamental threat to the validity of inferences, there has not yet been systematic study of DIF in the measurement of latent classes in mixture models, and there are no known applications of DIF analyses to mixture model results in the substance abuse literature. The proposed research seeks to rectify this shortcoming by rigorously investigating DIF in mixture models with a focus on behavioral studies of substance abuse. The four aims of the proposed project are to: (1) analytically define the ways that DIF can manifest in mixture models; (2) develop a flexible test of DIF in mixture models; (3) use a computer simulation to test (a) the consequences of un-modeled DIF in mixture models and (b) the efficacy of the test developed in Aim 2; and (4) to use DIF modeling techniques in an empirical study of substance abuse in the transition to adulthood. Through the pursuit of these aims, this project will allow fo mixture models to better consider individual differences, thereby enhancing the generalizability of these results in behavioral substance abuse research.