CAREER: Expanding the Applicability, Utility and Popularity of Item Response Theory Models for Unfolding

这项研究计划将在项目反应理论(IRT)模型的展开领域开发和应用新的测量方法。展开的IRT模型是一种概率模型，它使用对测试或问卷的回答来同时估计每个受访者的潜在特征以及每个项目的特征。与传统的IRT模型不同，传统的IRT模型采用累积项目反应函数来测量能力和熟练程度等特征，而展开IRT模型结合了单峰、非单调的项目反应函数来测量特定发展过程中分阶段发生的态度、偏好和个人位置等结构。这项研究计划将扩大一种特殊的多分支反应展开IRT模型的适用性、实用性和普及性，该模型被称为广义分级展开模型(GGUM)。这些目标将通过一系列综合的教育和研究项目来实现。研究项目将包括：1)调查可与GGUM一起使用的模型、项目和个人匹配的替代指数--结果将使从业者能够更好地识别GGUM何时适合于一组给定的项目回答。2)探索从马尔可夫链蒙特卡罗方法导出的GGUM参数的贝叶斯估计--该技术将潜在地提高具有稀疏数据的情况下GGUM参数估计的准确性，并将更好地表示那些参数估计中固有的不确定性。3)GGUM的新的多维扩展的开发-该模型将允许测量许多实质性领域的社会科学家感兴趣的多维潜在特征，例如心理学、营销学、政治学、4)基于GGUM反应函数的新混合模型的开发--该混合模型将估计有责任心的受访者的潜在特征，同时从算法的测量部分概率地剔除无责任心的受访者。该计划的教育部分包括关于GGUM模型系列的应用和好处的一系列介绍性和强化研讨会。将举办研讨会，并向美国和欧洲的应用测量从业者介绍。改进和随后分发估计GGUM参数的免费、用户友好的计算机软件将补充这些教育活动。最近的心理测量学研究表明，对典型态度和偏好问卷的反应更适合用展开模型来描述，而不是累积模型。事实上，在这些情况下使用累积模型可能会导致衡量标准无效。开展IRT模型将在这些情况下促进更有效的测量，同时也提供通常与累积IRT模型相关的其他好处。这些能力包括建立题库以维持共同的测量尺度的能力，估计每个受访者潜在特质估计的精确度的能力，以及使用计算机化的适应性测试方法衡量态度、偏好和社会科学中其他结构的能力。因此，这项研究的结果可以提高在许多社会科学学科中开发的测量质量，同时以技术上合理和实用的方式扩展应用测量实践。该计划的教育部分将为测量从业者提供必要的知识和工具，以有效地使用这一日益增长的方法。

This research program will develop and apply new measurement methodology in the domain of item response theory (IRT) models for unfolding. An unfolding IRT model is a probabilistic model that simultaneously estimates each respondent's latent trait along with the characteristics of each item using the responses to a test or questionnaire. In contrast to traditional IRT models that implement cumulative item response functions to measure traits like ability and proficiency, unfolding IRT models incorporate single-peaked, nonmonotonic item response functions to measure constructs like attitudes, preferences and individual locations within certain developmental processes that occur in stages. This research program will expand the applicability, utility and popularity of a particular unfolding IRT model for polytomous responses known as the generalized graded unfolding model (GGUM). These goals will be achieved through an integrated sequence of educational and research projects. Research projects will include:1) An investigation of alternative indices of model, item and person fit that can be utilized with the GGUM - the results will enable practitioners to better identify when the GGUM is and is not appropriate for a given set of item responses. 2) An exploration of Bayesian estimates of GGUM parameters derived from a Markov chain Monte Carlo method - this technique will potentially improve the accuracy of GGUM parameter estimates in situations with sparse data and will better represent the uncertainty inherent in those parameter estimates.3) The development of a new multidimensional extension of the GGUM - this model will allow for the measurement of multidimensional latent traits that are of interest to social scientists in many substantive areas like psychology, marketing, political science, etc. 4) The development of a new mixture model that is based on the GGUM response function - this mixture model will estimate latent traits of conscientious respondents while probabilistically culling unconscientious respondents from the measurement portion of the algorithm.The educational component of this program includes a series of both introductory and intensive workshops on the application and benefits of the GGUM family of models. Workshops will be developed and presented to applied measurement practitioners across the U.S. and Europe. The refinement and subsequent distribution of free, user-friendly computer software that estimates GGUM parameters will complement these educational activities. Recent psychometric research suggests that responses to typical attitude and preference questionnaires are more appropriately described by unfolding models rather than cumulative models. Indeed, the use of cumulative models in these situations can lead to invalid measures. Unfolding IRT models will promote more valid measurement in these situations while also providing other benefits commonly associated with cumulative IRT models. These include the ability to build item banks that maintain a common scale of measurement, the ability to estimate the precision of each respondent's latent trait estimate, and the ability to measure attitudes, preferences and other constructs in the social sciences using computerized adaptive testing methods. Consequently, the results from this research can improve the quality of measurements that are developed in many social science disciplines while simultaneously expanding applied measurement practices in ways that are both technically sound and practical. The educational component of this program will provide measurement practitioners with the requisite knowledge and tools to effectively use this growing methodology.