EFFECTS OF MISSPECIFICATION IN LONGITUDINAL MODELS

流行病学随访研究在过去的5-10年里变得很常见。 本提案的目的是评估稳健性， 开发健壮的程序，并检测错误说明 用于分析这类纵向数据的回归模型 学习。由于最近出现了许多多元回归方法 与不平衡的数据一起使用建模的这些方面，这些方面很重要 对于实际应用，还没有得到充分的探索。 模型错误说明的两个常见示例是省略的协变量和 对关系的数学形式的错误假设。这些 问题密切相关，因为省略协变量可能会导致 回归方程的形式。相反，例如二次项 在回归方程中可视为省略的协变量。这两种形式的 违反模式可能会导致个人之间和内部的差异 回归系数，这是一个经常观察到的现象。 这种情况的一个例子是通过让 感兴趣的协变量x与遗漏的混杂因子之间的相关性 Z不同于个体之间的差异。这样的情况很常见。 (例如，在影响的流行病学研究中的队列和周期效应 老化)。此模型结果的扩展构成了 拟议的调查。数据生成结构将被推广 对纵向数据分析的影响将是 检查过了。对现有和现有的各种不同的敏感性进行了调查 对于具体的违规行为，新颖的分析方法将起到指导作用 选择和进一步发展稳健的方法。由于受到 模型错误说明因分析方法而异，这也将是 作为一种潜在的检测模型违规的方法被研究。 本提案中开发的方法将应用于统计 对几项正在进行的流行病学跟踪研究的数据进行分析。

Epidemiologic follow-up studies have become common in the last 5-10 years. The aims of this proposal pertain to the evaluation of robustness, development of robust procedures, and detection of misspecification in regression models for the analysis of longitudinal data arising from such studies. Because of the recency of many multivariate regression methods used with unbalanced data these aspects of modelling, which are important for practical application, have not yet been fully explored. Two common examples of model misspecification are omitted covariates and incorrect assumptions on the mathematical form of the relationship. These problems are closely related since omitted covariates can lead to change in the form of regression equations. Conversely, for example quadratic terms in regression equations can be viewed as omitted covariates. Both forms of model violations may lead to differences in across and within individual regression coefficients, an often observed phenomenon. One example of such a situation has been modelled by letting the correlation between the covariate of interest x, and an omitted confounder z be different across than within individuals. Such situations are common (e.g. cohort and period effects in epidemiologic studies of the effect of aging). Extension of the results for this model form the basis for the proposed investigation. The data generating structure will be generalized in several directions, and the impact on longitudinal data analysis will be examined. The investigation of the sensitivity of various existing and novel analysis methods to specific model violations, will serve as a guide to choice and further development of robust methods. Since the impact of model misspecification differs by analysis approach, this will also be investigated as a potential method for detecting model violation. The methods developed in this proposal will be applied in the statistical analysis of data from several ongoing epidemiologic follow-up studies.