New Statistical Methods for Cox Regression with Measurement Errors in Cancer and Nutritional Epidemiology

项目摘要/摘要 癌症和营养流行病学的最大挑战之一是协变量的不确定性 测量，这是旨在阐明两者之间因果关系的研究中的主要偏差来源 饮食与癌症发病率和死亡率的关系。饮食摄入量通常是使用自我报告的食物频率来估计的 问卷调查，这种膳食评估往往会有很大的测量误差。这一问题 暴露的不确定性可能导致对健康影响的估计存在相当大的偏差，掩盖了我们的能力 来发现真实的联系。在流行病学实践中，右审查的生存结果很常见，而且 考克斯回归模型通常用来对这种结果进行建模。回归校正是一种简单的 COX回归中的测量误差调整方法，常用于癌症和营养领域 流行病学。然而，当测量误差程度较大时，其回归系数 容易出错的协变量很大，回归校正因暴露而不能满足偏差校正的要求 不确定性。我们将开发一种改进的回归校正方法，考虑到 在校准过程中同时测量误差和易出错的协变量系数。 期望最大化(EM)算法已经显著地应用于各种情况 不完整的数据问题。然而，EM在处理测量误差方面并没有受到太多的关注 生存数据中的问题。我们将在主要/验证研究中开发EM的完整治疗方案，以填补 文献中的重要空白。用户友好的公开可用的软件开发将是一个主要特征 伴随着所有待开发的新方法。

Project Summary / Abstract One of the greatest challenges in cancer and nutritional epidemiology is the uncertainty in covariate measurements, which is a major source of bias in research aimed at elucidating causal relationships between diet and cancer incidence and mortality. Dietary intake is usually estimated using a self-report food-frequency questionnaire, and this dietary assessment is often subject to substantial measurement error. The issue of exposure uncertainty lead to the potential for considerable bias in estimated health effects, masking our ability to detect true associations. Right-censored survival outcomes are common in epidemiologic practice, and the Cox regression model is typically used to model such outcomes. Regression calibration is a simple measurement error adjustment method in Cox regression, and is often used in cancer and nutritional epidemiology. However, when the degree of measurement error is large and the regression coefficient of the error-prone covariate is large, regression calibration is unsatisfactory for bias correction due to exposure uncertainty. We will develop an improved regression calibration method considering the degree of measurement error and the coefficient of the error-prone covariate simultaneously in the calibration procedure. The Expectation-Maximization (EM) algorithm has been remarkably applied for a wide variety of situations for incomplete data problems. However, EM does not receive much attention to deal with measurement error problems in survival data. We will develop a complete treatment of EM in main / validation studies to fill an important gap in the literature. User-friendly publicly available software development will be a central feature accompanying all new methods to be developed.