Gradient, Cambridge, Massachusetts 02138, USA.
Crit Rev Toxicol. 2011 Sep;41(8):651-71. doi: 10.3109/10408444.2011.563420.
Both classical and Berkson exposure measurement errors as encountered in environmental epidemiology data can result in biases in fitted exposure-response relationships that are large enough to affect the interpretation and use of the apparent exposure-response shapes in risk assessment applications. A variety of sources of potential measurement error exist in the process of estimating individual exposures to environmental contaminants, and the authors review the evaluation in the literature of the magnitudes and patterns of exposure measurement errors that prevail in actual practice. It is well known among statisticians that random errors in the values of independent variables (such as exposure in exposure-response curves) may tend to bias regression results. For increasing curves, this effect tends to flatten and apparently linearize what is in truth a steeper and perhaps more curvilinear or even threshold-bearing relationship. The degree of bias is tied to the magnitude of the measurement error in the independent variables. It has been shown that the degree of bias known to apply to actual studies is sufficient to produce a false linear result, and that although nonparametric smoothing and other error-mitigating techniques may assist in identifying a threshold, they do not guarantee detection of a threshold. The consequences of this could be great, as it could lead to a misallocation of resources towards regulations that do not offer any benefit to public health.
经典和 Berkson 暴露测量误差都可能导致拟合暴露反应关系出现偏差，这些偏差大到足以影响风险评估应用中暴露反应关系的解释和使用。在估计个体接触环境污染物的过程中，存在多种潜在测量误差源，作者回顾了文献中对实际中普遍存在的暴露测量误差的大小和模式的评估。统计学家都知道，独立变量（如暴露反应曲线中的暴露）值的随机误差可能会导致回归结果出现偏差。对于递增曲线，这种效应倾向于使本来更陡峭、更可能是曲线或甚至是带有阈值的关系变得平坦和线性化。偏差的程度与独立变量的测量误差的大小有关。已经表明，适用于实际研究的偏差程度足以产生错误的线性结果，尽管非参数平滑和其他减轻误差的技术可能有助于识别阈值，但它们并不能保证检测到阈值。这可能会产生重大后果，因为这可能导致资源错误分配给对公共健康没有任何益处的法规。
