STATISTICAL METHODS FOR MISMEASURED OR MISSING DATA

工作总结：错误测量或缺失的数据可能导致错误或 误导性的结论，并提出了一个普遍的问题，在各种 生物医学领域，包括环境流行病学。 这个项目 旨在开发新的统计方法，并将现有方法应用于 应对以下因素对有效统计估计和检验提出的挑战： 错误测量或缺失数据。继续在三个领域开展工作：（1） 描述了保证剂量-反应信号 趋势在测量误差面前保持不变，（2） 病例对照研究中缺失数据的插补值 无偏估计和正确的标准误差，但采用相对 简单的技术，和（3）故意省略一些策略 以节省研究成本的方式收集数据，同时保留大部分 如果收集了所有数据，本应提供的信息。 我们已经发现了至少适用于一个受限制的类的条件 回归模型，在该模型下，剂量反应趋势的符号为 当数据被错误地测量时，我们试图看到 这些结果可能适用。 我们已经看到，将对照组中观察到的平均反应 病例和对照组数据缺失的受试者可以提供有效的 当使用超额相对风险模型时，我们正在努力 使这些结果适应更常见的逻辑模型。 早期的结果表明，当暴露评估成本很高时（例如， 测试血清中的毒性代谢物），测量来自 多学科可以在不丢失大量信息的情况下降低成本。

Summary of Work: Mismeasured or missing data can lead to false or misleading conclusions and present a widespread problem in a variety of biomedical fields, including environmental epidemiology. This project seeks to develop new statistical methods and to apply existing methods to address challenges to valid statistical estimation and testing posed by mismeasured or missing data. Work continues in three areas: (1) delineating conditions that guarantee that the sign of dose-response trend is preserved in the face of measurement error, (2) methods to impute values for missing data in case-control studies that provide unbiased estimates and correct standard errors yet employ relatively simple techniques, and (3) strategies for intentionally omitting some data in ways that save on study costs while preserving most of the information that would have been available had all data been collected. We have discovered conditions that apply at least to a restricted class of regression models under which the sign of a dose-response trend is preserved when data are mismeasured; we are trying to see how broadly these results may apply. We have seen that imputing the mean response observed among control subjects to missing data for both cases and controls can provide valid inference when an excess relative risk model is used; we are working to adapt these results to the more common logistic model. Early results indicate that when an exposure assessment is costly (e.g., testing serum for toxic metabolites), measuring pooled samples from multiple subjects can cut costs without losing much information.