Heterogeneity Among Unobserved Subpopulations

描述（由申请人提供）：拟议的研究项目是一位年轻研究者首次提交的 R01 申请。拟议项目的目标是通过开发和研究新模型来解释未观察到的（基础）亚群之间的异质性，从而在统计进展和心理健康研究实践之间架起桥梁。心理健康研究中经常提出的一个研究问题是，目标人群中是否存在在结果分布、背景特征、发展轨迹和对干预治疗的反应方面存在差异的亚组。考虑亚人群差异通常会导致对研究结果的解释产生重大差异。当亚总体成员完全或部分未被观察到时，就会出现统计挑战。由于共存的统计挑战，解释潜在亚群（潜在类别）之间异质性的统计方法可能会进一步复杂化。拟议的项目将研究更广泛的统计模型框架，这些框架可以反映更现实的环境，同时考虑到未观察到的亚群之间的异质性。通用潜变量（GLV）建模将被用作灵活的分类工具，可捕获连续和离散的异质性谱。该提案围绕三个具体目标进行组织，以应对心理健康研究中出现的常见并发症：首先，研究估计治疗对未观察到的亚群的差异效应的方法。其次，研究利用未观察到的亚群之间的异质性信息来模拟缺失数据机制的方法。第三，研究对考虑多级数据结构的未观察子群体之间的异质性进行建模的方法。将采用三种策略来实现这些目标：首先，对新的统计模型进行数学研究。其次，通过深入的模拟研究评估这些模型的保真度。最后，通过心理健康研究中的实证例子证明新模型的适用性和实用性。实证例子中展示的统计建模特征不仅会影响结果分析，还会影响心理健康研究的研究设计策略。

DESCRIPTION (provided by applicant): The proposed research project is a first submission of an R01 application by a young investigator. The goal of the proposed project is to bridge statistical advances and mental health research practice by developing and investigating new models to account for heterogeneity among unobserved (underlying) subpopulations. A research question often raised in mental health research is whether there are subgroups within the target population that differ in outcome distributions, background characteristics, developmental trajectories, and response to intervention treatments. Considering subpopulation differences often leads to major differences in the interpretation of research findings. Statistical challenges arise when subpopulation membership is completely or partly unobserved. Statistical methods to account for heterogeneity among latent subpopulations (latent classes) can be further complicated due to co-existing statistical challenges. The proposed project will investigate broader statistical modeling frameworks that can reflect more realistic settings while accounting for heterogeneity among unobserved subpopulations. General latent variable (GLV) modeling will be utilized as a flexible classification tool that captures both the continuous and the discrete spectrum of heterogeneity. The proposal is organized around three specific aims formulated in response to common complications that arise in mental health research: First, investigate methods to estimate differential effects of treatments for unobserved subpopulations. Second, investigate methods to model missing-data mechanisms using information on heterogeneity among unobserved subpopulations. Third, investigate methods to model heterogeneity among unobserved subpopulations accounting for multilevel data structures. Three strategies will be employed in pursuing these aims: First, perform mathematical investigations of new statistical models. Second, evaluate the fidelity of these models through intensive simulation studies. Finally, demonstrate applicability and practicality of new models through empirical examples in mental health research. Statistical modeling features demonstrated in empirical examples will have implications not on y in outcomes analysis, but also in study design strategies for mental health research.