Differential Item Functioning (DIF) Testing with Estimation of the Latent Densities

项目反应理论(IRT)的一个重要应用是识别差异项目功能(DIF)。当一个项目对一组人具有不同于另一组人(例如，男性与女性)的测量属性时，就会发生差异，而不考虑潜在变量上的组平均值差异。通常，在基于IRT的DIF检验中，假设潜在变量在人群中呈正态分布。然而，对于某些变量，这一假设可能站不住脚。本研究开发了基于IRT的DIF测试的统计方法和软件，其中每组人的潜在分布与项目参数同时被估计。本研究填补了现有方法的一个空白，旨在提高数据与用于评估DIF的统计模型之间的匹配。测量潜在变量的问卷可能是非正态分布的，被用作制定心理和医学诊断和治疗计划的一部分，以及企业或大学的招聘或录取决定。以多数群体个人为基础构建的评估工具可能会对代表性不足的群体提供误导性的结果。拟议的方法可以应用于这些领域，以促进公平、准确和包容性的衡量。作为支持调查和统计方法研究的联合活动的一部分，这项研究得到了方法学、测量和统计计划、里斯计划和一个联邦统计机构联盟的支持。

One important application of item response theory (IRT) is for identification of differential item functioning (DIF). DIF occurs when an item has different measurement properties for one group of people than another (e.g., men versus women), irrespective of group-mean differences on the latent variable. Typically, in IRT-based DIF testing, the latent variables are assumed to be normally distributed in the populations of people. However, this assumption may be untenable for some variables. The present research develops statistical methodology and software for IRT-based DIF testing in which the latent distribution for each group of people is estimated simultaneously with the item parameters.This research fills a gap in extant methodology and is aimed at improving the match between data and the statistical models used to assess DIF. Questionnaires measuring latent variables which are potentially nonnormally distributed are used as part of making, for example, psychological and medical diagnoses and treatment plans, and hiring or admission decisions in businesses or universities. Assessment instruments constructed based on majority-group individuals may provide misleading results for underrepresented groups. The proposed methodology could be applied in these areas to promote fair, accurate, and inclusive measurement. The research is supported by the Methodology, Measurement, and Statistics Program, the REESE Program, and a consortium of federal statistical agencies as part of a joint activity to support research on survey and statistical methodology.