The logistic analysis of epidemiologic prospective studies: investigation by simulation.
The logistic analysis of epidemiologic prospective studies: investigation by simulation.
复制标题
流行病学前瞻性研究的逻辑分析：模拟调查。
DOI：
10.1002/sim.4780040213
复制
发表时间：
1985
影响因子：
2
通讯作者：
中科院分区：
文献类型：
--
作者：
作者声称，通过模拟证明，用于分析流行病学前瞻性研究结果的Logistic回归产生的回归系数估计“既不准确也不可靠”。他们声称，从有限总体抽样的模拟结果表明，随着样本量的增加，使用Logistic回归得到的参数估计不会收敛于总体参数，并且这些参数估计的估计标准差是不一致和不可靠的。作者们错了。作者感兴趣的是从大小为N的有限总体中抽样时的推断。他们说，他们的结果是值得注意的，因为被检查的样本大小和总体大小是可比的。对于Framingham使用的参数估计，以及其他研究。首先，他们错误地认为通过Logistic回归得到的参数估计不一致。他们使用Logistic回归模型生成了10，00O的人口，但得到的人口10，00O的参数仅在预期上等于用于产生10，00O的人口的参数。作者将发现，从他们的模拟的参数估计是一致的，从某种意义上说，当样本量n增加到10，00时，它们收敛于总体参数(这些值是通过将Logistic模型拟合到10，00O的总体而获得的值)。其次，有理由相信作者的模拟程序中存在错误，并且该程序没有对估计的方差应用有限总体校正。应用有限总体校正将导致将其表七和表八中的条目乘以{N/(N-n)}‘/*，从而产生接近1的比率值(在模拟的精度范围内)。的确有
The authors claim to demonstrate via simulation that logistic regression as used to analyse the results of epidemiologic prospective studies produces estimates of regression coefficients that are ‘neither accurate nor reliable’. They claim that the results of their simulation, sampling from a finite population, demonstrate that the parameter estimates obtained using logistic regression do not converge to the population parameters as the sample size increases and that the estimated standard deviations of these parameter estimates are inconsistent and unreliable. The authors are wrong. The authors’ interest is in inference when sampling from a finite population of size N. They state that their results ‘are noteworthy as the sample sizes and population size examined are comparable... to those used in Framingham, as well as other studies’.First, they are wrong about the inconsistency of the parameter estimates obtained through logistic regression. They generated the population of 10, OOO using a logistic regression model, but the resulting parameters of the population of 10, OOO are equal only in expectation to those used to generate the population of 10, OOO. The authors will find that the parameter estimates from their simulations are consistent, in the sense that they converge, as the sample size n increases to lO, OOO, to the population parameters (which are those values obtained by fitting a logistic model to the population of l0, OOO). Secondly, there is reason to believe that there is an error in the authors’ simulation program and that the program did not apply the finite population correction to the estimated variances. Application of the finite population correction would lead to multiplication of the entries in their Tables VII and VIII by {N/(N-n)}’/* which produces values for the ratios near 1 (within the precision of the simulation). There is