Institute for Health Care Research and Improvement, Dallas, TX 75206, USA.
Ann Epidemiol. 2010 Jul;20(7):562-7. doi: 10.1016/j.annepidem.2010.03.012.
To quantify the impact of ignoring misclassification of a response variable and measurement error in a covariate on statistical power, and to develop software for sample size and power analysis that accounts for these flaws in epidemiologic data.
A Monte Carlo simulation-based procedure is developed to illustrate the differences in design requirements and inferences between analytic methods that properly account for misclassification and measurement error to those that do not in regression models for cross-sectional and cohort data.
We found that failure to account for these flaws in epidemiologic data can lead to a substantial reduction in statistical power, over 25% in some cases. The proposed method substantially reduced bias by up to a ten-fold margin compared to naive estimates obtained by ignoring misclassification and mismeasurement.
We recommend as routine practice that researchers account for errors in measurement of both response and covariate data when determining sample size, performing power calculations, or analyzing data from epidemiological studies.
量化在回归模型中忽略因变量分类错误和协变量测量误差对统计功效的影响，并开发一种用于样本量和功效分析的软件，以解决流行病学数据中的这些缺陷。
采用基于蒙特卡罗模拟的方法，说明在正确考虑分类错误和测量误差的分析方法与未正确考虑这些因素的分析方法之间，在设计要求和推论上的差异，包括横断面和队列数据的回归模型。
我们发现，在流行病学数据中不考虑这些缺陷会导致统计功效大幅降低，在某些情况下降低超过 25%。与忽略分类错误和测量错误的简单估计相比，所提出的方法通过最多十倍的幅度降低了偏差。
我们建议研究人员在确定样本量、进行功效计算或分析流行病学研究数据时，常规考虑对因变量和协变量数据的测量误差。
