MEASUREMENT ERRORS IN EPIDEMIOLOGY--BREAST/COLON CANCER

正在研究的风险因素中的测量误差是 癌症流行病学中不精确和偏差的重要来源。一个 我们小组关注的例子是人与人之间的差异 在病例对照中收集的关于酒精使用的问卷数据 乳腺癌发病率的研究。最近又出现了一场新的 对研究处理这个问题的统计方法感兴趣。 在某种程度上，这种兴趣可以追溯到越来越多的人使用More 用于确定暴露和结果的复杂技术。 在测量误差方法方面的新研究主要集中在 关于流行病学家常用的技术，如二进制 回归和病例对照研究。然而，这样的情况相对较少 方法已经被引入到癌症的实践中 流行病学。 此应用程序的长期目标是鼓励 适当使用最近开发的测量误差方法。 这将通过开发和记录统计数据来实现 函数形式的方法，用于分析 流行病学数据。测量的成功应用-- 误差方法的实际研究取决于误差方法的可用性 建议算法在计算机中的可靠实现 流行病学研究人员熟悉的软件格式。 统计文献中反复出现的主题之一是关于 二元回归中的测量误差是难以做到的 最大似然估计当似然包含一个 不能解析计算的积分。一批 作者，包括参与调查的调查人员 建议，提出了近似的推论方法 消除这种计算困难，包括有效的分数 广义线性模型的检验，最大值的逼近 Logistic回归的可能性，以及可积模型，如 具有正常测量误差的概率回归。这些方法 将被提炼和调整以用于病因学和 预防乳腺癌、结肠癌和皮肤癌。主要的重点将是 被放置在有助于确定适当的 流行病学调查的样本量 风险因素中的测量误差。

Errors of measurement in the risk factors under study are an important source of imprecision and bias in cancer epidemiology. An example of concern to our group is the within-person variability of questionnaire data on alcohol use being collected in a case-control study of breast cancer incidence. Recently there has been a renewed interest in studying statistical methods for treating this problem. In part, this interest can be traced to the growing use of more sophisticated techniques for determining exposures and outcomes. The new research in measurement error methods has primarily focused on techniques commonly used by epidemiologists such as binary regression and case-control studies. Yet relatively few of these methods have been introduced into the practice of cancer epidemiology. The long term objective of this application is to encourage the appropriate use of recently developed measurement-error methods. This will be accomplished by developing and documenting statistical methods in the form of functions for use in the analysis of epidemiologic data. The successful application of measurement- error methods to practical research depends on the availability of reliable implementations of the suggested algorithms in a computer software format familiar to researchers working in epidemiology. One of the recurrent themes in statistical literature regarding measurement errors in binary regression is the difficulty of doing maximum likelihood estimation when the likelihood contains an integral that cannot be evaluated analytically. A number of authors, including the investigators participating in this proposal, have suggested approximate methods of inference that eliminate this computational difficulty, including efficient score tests for generalized linear models, approximations to maximum likelihood for logistic regression, and integrable models, such as probit regression with normal measurement errors. These methods will be refined and adapted for use in studies of the etiology and prevention of breast, colon and skin cancer. A major emphasis will be placed on procedures useful for determining the appropriate sample size for epidemiologic investigations with substantial errors of measurement in risk factors.