Differential Item Functioning in item response models:new estimators and diagnostic instruments

项目反应模型建立了测量表现和智力的方法学背景。其基本概念是，一个项目的解决方案是由两个潜在的特征决定的，即人的能力和项目的难度。项目响应模型本质上是高维的。该项目的目的是开发正则化方法，能够捕获特定项目与模型假设结构的偏差……正则化估计器的潜力特别用于诊断差异项目功能。差异项目功能（DIF）描述了一种现象，即一个项目的难度可能因人而异，这取决于他们的社会或文化背景，从而产生对个人能力的偏见估计。计划开发惩罚似然和增强方法，以满足建模任务的复杂性。由于当包含潜在的dif诱导变量时，参数的数量会强烈增长，因此所有传统的估计都失败了。这些方法应该能够识别负责DIF的项目，并找到负责DIF的预测器。此外，还将对这些方法进行扩展，以适应具有有序响应的项目的模型。待开发的方法肯定超出了现有的方法。在机器学习社区和当代统计学中首次提出的方法的适应和修改才刚刚开始，以应对DIF在项目响应模型中的挑战。新方法有望极大地扩展现有诊断仪器的工具箱。

Item response models build the methodological background for the measurement of performance and intelligence. The basic concept is that the solution of an item is determined by two latent traits, the ability of the person and the difficulty of the item. Item response models are intrinsically high-dimensional. The aim of the project is to develop regularization methods that are able to capture the deviation of specific items from the structure assumed by the model..The potential of regularized estimators is in particular used for the diagnosis of differential item functioning. Differential item functioning (DIF) describes the phenomenon that the difficulty of an item may vary over individuals depending on their social or cultural background yielding biased estimates of person abilities. It is planned to develop penalized likelihood and boosting methods that meet the complexity of the modeling task. Since the number of parameters is growing strongly when the potentially DIF-inducing variables are included all conventional estimators fail. The methods should be able to identify the items that are responsible for DIF and also find the predictors that are responsible for DIF. In addition, the methods are to be extended to fit models for items with ordered response. The methods to be developed are definitely beyond the methods that are available. The adaptation and modification of methods that have been first proposed in the machine learning community and in contemporary statistics to meet the challenges of DIF in item response models has just started. New methods are expected that crucially extend the tool box of available diagnostic instruments.