A new framework for managing and analyzing multiply imputed data in Stata
A new framework for managing and analyzing multiply imputed data in Stata
复制标题
DOI：
10.1177/1536867x0800800104
复制
发表时间：
2008-01-01
期刊：
影响因子：
4.8
通讯作者：
Royston, Patrick
中科院分区：
文献类型：
--
作者：
Carlin, John B.;Galati, John C.;Royston, Patrick
本文描述了一套新的工具，用于对数据集集合进行分析，这些数据集包括原始数据的多个副本，并根据多重输入方法的要求对缺失值进行输入。这些工具取代了最初由作者开发的工具。它们基于一种简单的数据管理范式，其中输入的数据集与原始数据一起存储在一个垂直堆叠格式的单个数据集中，正如Royston在他的ice和micomine命令中提出的那样。与单独存储数据集相比，堆叠成单个数据集简化了输入数据集的管理。使用一个新的前缀命令mim来执行堆叠数据集的分析和操作，该命令可以容纳通过任何方法输入的数据，只要在创建输入数据时遵循一些简单的规则。mim可以有效地拟合Stata中大多数可用的回归模型，对多个输入数据集进行拟合，给出根据Rubin结果计算的参数估计和置信区间，用于多个输入推断。特别注意将可用的后估计命令限制为那些已知在多重imputation上下文中有效的命令。但是，用户可以灵活地覆盖这些默认值。这些新工具的特性使用两个先前发布的示例进行说明。
A new set of tools is described for performing analyses of an ensemble of datasets that includes multiple copies of the original data with imputations of missing values, as required for the method of multiple imputation. The tools replace those originally developed by the authors. They are based on a simple data management paradigm in which the imputed datasets are all stored along with the original data in a single dataset with a vertically stacked format, as proposed by Royston in his ice and micombine commands. Stacking into a single dataset simplifies the management of the imputed datasets compared with storing them individually. Analysis and manipulation of the stacked datasets is performed with a new prefix command, mim, which can accommodate data imputed by any method as long as a few simple rules are followed in creating the imputed data. mim can validly fit most of the regression models available in Stata to multiply imputed datasets, giving parameter estimates and confidence intervals computed according to Rubin's results for multiple imputation inference. Particular attention is paid to limiting the available postestimation commands to those that are known to be valid within the multiple imputation context. However, the user has flexibility to override these defaults. Features of these new tools are illustrated using two previously published examples.