Department of Biostatistics, Fielding School of Public Health, University of California, Los Angeles, California, USA.
Laboratory of Immunoregulation, NIAID, NIH, Baltimore, Maryland, USA.
Biometrics. 2022 Sep;78(3):908-921. doi: 10.1111/biom.13472. Epub 2021 May 3.
A method for generalized linear regression with interval-censored covariates is described, extending previous approaches. A scenario is considered in which an interval-censored covariate of interest is defined as a function of other variables. Instead of directly modeling the distribution of the interval-censored covariate of interest, the distributions of the variables which determine that covariate are modeled, and the distribution of the covariate of interest is inferred indirectly. This approach leads to an estimation procedure using the Expectation-Maximization (EM) algorithm. The performance of this approach is compared to two alternative approaches, one in which the censoring interval midpoints are used as estimates of the censored covariate values, and another in which the censored values are multiply imputed using uniform distributions over the censoring intervals. A simulation framework is constructed to assess these methods' accuracies across a range of scenarios. The proposed approach is found to have less bias than midpoint analysis and uniform imputation, at the cost of small increases in standard error.
描述了一种广义线性回归的区间截断协变量方法，扩展了以前的方法。考虑了一种情况，其中感兴趣的区间截断协变量被定义为其他变量的函数。该方法不是直接对感兴趣的区间截断协变量的分布进行建模，而是对决定该协变量的变量的分布进行建模，并间接地推断感兴趣的协变量的分布。这导致了一种使用期望最大化（EM）算法的估计程序。该方法的性能与两种替代方法进行了比较，一种方法是使用截断区间中点作为截断协变量值的估计值，另一种方法是使用截断值在截断区间上均匀分布的多次插补。构建了一个模拟框架来评估这些方法在一系列场景下的准确性。研究发现，与中点分析和均匀插补相比，该方法的偏差更小，但标准误差略有增加。
