Statistical Methods for Environmental Data Subject to Detection Limits

描述（由申请人提供）：当研究人员调查癌症与暴露于环境化学品（如微量元素、农药和二恶英）之间的关系时，他们经常发现浓度低于被认为足够可靠的限值，可以报告为数值。在某些研究中，检测限（DL）可能是一个固定的数字，但在其他研究中，它也可能因样品而异。对于后者，DL可能与暴露水平相关，如在肯塔基州的结肠癌研究中观察到的。受数据指标制约的数据对数据分析和解释提出了挑战。在这个建议中，我们集中在两个重要的统计问题，在分析数据的环境流行病学研究：（a）估计的化学品分布在一个特定的群体;（B）群体之间的分布比较。对于这两个问题，特设，参数和非参数的方法已经提出。特别方法是不明智的，除非有相对较少的测量值低于DL;参数方法可能会导致显着偏倚的结果时，参数模型是错误的。非参数方法由于其稳健性，近年来受到越来越多的关注。然而，目前的非参数方法只是借用了常用的右删失生存数据的方法，并没有考虑到以下两个独特的特点，环境暴露数据与DL：（a）它是没有意义的定义风险函数的暴露测量;（B）DL值是可观察的所有科目，包括那些实际暴露水平被检测到。此外，目前的非参数方法不允许抽样权重，这通常存在于调查数据，如国家健康和营养检查调查（NHANES）。由于这些问题，目前的非参数方法可能导致以下四个问题，用于分析具有DL的环境暴露数据：（a）缺乏有意义的解释;（B）低效的结果;（c）无法处理 （d）无法行程以抽样权重计算的调查数据。为了解决上述问题，我们将开发统一和有效的非参数估计和检验方法，可以（a）处理暴露水平和DL之间可能存在的相关性;（B）纳入抽样权重。我们将使用最先进的方法来处理删失生存数据，并将其调整为具有DL的环境暴露数据。所提出的方法将应用于最近在肯塔基州进行的结肠癌病例对照研究、在肯塔基州进行的肺癌病例对照研究和NHANES的数据。1

DESCRIPTION (provided by applicant): As researchers investigate the relationship between cancer and exposure to environmental chemicals such as trace elements, pesticides, and dioxins, they often find concentrations that are lower than limits deemed reliable enough to report as numerical values. The detection limit (DL) may be a fixed number in some studies, but it can also vary widely from sample to sample in other studies. For the latter, the DL may be correlated with the exposure level, as observed in a colon cancer study in Kentucky. The data subject to DLs present challenges for data analysis and interpretation. In this proposal we focus on two important statistical problems encountered in the analysis of data from environmental epidemiologic studies: (a) estimation of the chemical distribution in a specific group; and (b) comparison of distributions among groups. For these two problems, ad hoc, parametric, and nonparametric methods have been proposed. Ad hoc methods are ill-advised unless there are relatively few measurements below DLs; and parametric methods can lead to markedly biased results when the parametric model is misspecified. Nonparametric methods have received increasing attention in recent years because of their robustness. However, current nonparametric methods simply borrow the commonly used methods for right-censored survival data, and do not take into account the following two unique characteristics of environmental exposure data with DLs: (a) it is not meaningful to define the hazard function for an exposure measurement; and (b) DL values are observable for all subjects including those whose actual exposure levels are detected. In addition, current nonparametric methods do not allow for sampling weights, which are typically present in survey data such as the National Health and Nutrition Examination Survey (NHANES). Due to these issues, current nonparametric methods may lead to the following four problems for the analysis of environmental exposure data with DLs: (a) lack of meaningful interpretation; (b) inefficient results; (c) inability to deal with the situation that the exposure level and DL are correlated; and (d) inability to handle survey data with sampling weights. To address the aforementioned problems, we will develop unified and efficient nonparametric estimation and testing methods that can (a) deal with possible correlation between the exposure level and DL; (b) incorporate sampling weights. We will utilize state-of-the-art methods for censored survival data and tailor them to environmental exposure data with DLs. The proposed methods will be applied to data from a recently conducted colon cancer case-control study in Kentucky, an ongoing lung cancer case-control study in Kentucky, and the NHANES. 1