MEASUREMENT ERRORS IN ENVIRONMENTAL EPIDEMIOLOGY

暴露测量误差是几乎所有环境和流行病学研究中的一个问题。在环境卫生研究中通常收集的数据的标准分析中，这是造成偏差和统计能力丧失的主要原因。为了消除或至少减少偏差，并提高分析的统计能力，将开发统计方法，其动机是应用于三个环境流行病学研究。这些方法也应推广到其他环境流行病学设置。由于这些方法的适当使用取决于所做的假设，我们将仔细调查假设的有效性。当数据中出现违反假设的情况时，我们将在可能的情况下修改方法以适应偏差。如果这是不可能的，我们将通过分析手段和刺激，探索基本方法的敏感性假设违反和发展方法的经验验证的假设。将发展三种统计方法。第一项将涉及修正回归校准，以适应在三个研究中收集的数据。在第二和第三部分，我们将使用全参数极大似然方法和半参数估计方程方法，拟合利用所有数据的合适模型。这些方法使我们能够正确地解释非随机验证抽样，如果手头的数据中有证据证明这一点，并且半参数方法还将允许我们探索由测量误差模型的错误规范引起的偏差的影响。将考虑Berkson测量误差模型，并在相关时纳入缺失数据方法。

Exposure measurement errors is a problem in nearly all environmental and epidemiological studies. It is a major source of bias and loss of statistical power in standard analysis of data commonly collected in environmental health research. To eliminate or at least reduce the bias, and to increase the statistical power of the analysis, statistical methods will be developed, motivated by application to three environmental epidemiology studies. These methods should be generalizable to other environmental epidemiology settings as well. Because appropriate use of these methods depends on the assumptions made, we will carefully investigate the validity of the assumptions. When any assumption appears to be violated in the data, we will, when possible, modify the methods to accommodate the departures. If this is not possible, we will, through analytic means and stimulation, explore the sensitivity of the basic methods of assumption violations and develop methods for empirical verification of the assumptions. Three statistical methods will be developed. The first will involve modification of regression calibration to suit the data collected in the three studies. In the second and third, using fully parametric maximum likelihood methods and semi-parametric estimating equations methods, we will fit suitable models which utilize all of the data. These methods permit us to properly account for non-random validation sampling, should there be evidence on this in the data at hand, and the semi-parametric methods will, additionally, allow us to explore the impact of bias induced by mis-specification of the measurement error model. The Berkson measurement error model will be considered, and missing data methodology will be incorporated when relevant.