Mathematical Sciences: Inference for Nonparametric Regresssion

本文研究了参数回归和非参数回归的欧班克推断问题。通过比较参数估计和非参数估计，提出了参数回归模型的拟合优度检验方法。这对于包括广义线性模型和时间序列模型的相当一般的模型来说是可以实现的。通过使用中间渐近有效性，发展了用于这些非参数类型检验的分析性、渐近比较的新方法。这些测试思想被扩展到更一般的拟合优度问题，例如测试可加性，其中零模型甚至是非参数的。文中还研究了其他推断问题，包括构造具有良好有限样本性质的非参数回归的渐近有效可信区间的几个建议。在许多科学领域，如生物学、工程学、心理学等，人们通常有理由相信，正在收集的数据是由两个组成部分产生的：一个非随机成分，代表所有个人或受试者共同的自然特征；以及一个随机成分，说明个体差异。非随机分量代表正在研究的现象的可重现或可预测的方面，因此具有相当大的兴趣。在许多情况下，根据理论考虑或过去的经验，一个人可以假定一个数学框架或模型，它被认为描述了非随机分量。当这些模型正确时，它们可以提供有用的总结和预测工具。然而，错误的模型可能会导致误导性的结论和对未来事件的不准确预测。因此，重要的是要有方法论来评估数据假设模型的准确性。这项研究的重点是为此目的开发新的工具，这些工具依赖于非随机成分的两种估计量的比较：即在假设模型正确的假设下计算的估计量和不采用这一假设的非常灵活的估计量。这项工作涉及的理论统计问题包括制定比较两个估计量的客观标准，以及确定这些标准的值，这些标准表明数据不支持假设的模型。考虑的其他问题包括使用数据来构造具有已知的、指定的、包含未知的、非随机的数据分量的概率的区间。

DMS 96-25496 Eubank Problems of inference for parametric and nonparametric regression are investigated. Tests of goodness-of-fit for parametric regression models are proposed that derive from the comparison of parametric and nonparametric estimators. This is accomplished for rather general models that include generalized linear and time series models. New methods are developed for analytic, asymptotic comparison of these nonparametric type tests through the use of intermediate asymptotic efficiency. These testing ideas are extended to more general goodness-of-fit problems, such as testing for additivity, where the null model is even nonparametric. Other inference problems are also studied, including several proposals for the construction of asymptotically valid confidence intervals for nonparametric regression that have good finite sample properties. It is often reasonable to believe in many areas of science, such as biology, engineering, psychology, etc., that the data being collected is produced by two components: a nonrandom component, representing a characteristic of nature common to all individuals or subjects, and a random component that accounts for individual variation. The nonrandom component represents the reproducible or predictable aspect of the phenomenon being studied and is therefore of considerable interest. In many cases, from theoretical considerations or past experience, one can postulate a mathematical framework or model that is believed to describe the nonrandom component. When such models are correct they can provide useful summary and predictive tools. However, an incorrect model can lead to misleading conclusions and inaccurate predictions of future events. Thus, it is important to have methodology for assessing the accuracy of postulated models for data. This research focuses on the development of new tools for this purpose that rely on the comparison of two types of estimators for the nonrandom component: namely, an estimator that is computed under the assumpt ion that the postulated model is correct and a very flexible estimator that does not employ this assumption. The theoretical statistical issues that are addressed in this work include the development of objective criteria for comparing the two estimators and the determination of values for these criteria which indicate that the data does not support a postulated model. Other problems that are considered include the use of the data to construct intervals that have a known, specified, chance of containing the unknown, nonrandom component of the data.