Abstract

Mixed models are frequently used in social and economic analysis. Typically, the analyzed data are gathered through a survey sample from the population of interest. The underlying survey design may be non-ignorable, leading to possible bias in the regression parameters if not accounted for. To circumvent this problem, survey weights should be included in the estimation process. We propose a survey weighted generalized linear mixed model allowing for unit-specific survey weights and a flexible random effects structure. An estimation procedure is proposed and evaluated in two simulation studies under different scenarios. The first simulation study strongly encourages the use of survey weighted generalized linear mixed models if the survey design is non-ignorable. The second one replicates a previous simulation study in order to study the competitivity of the proposed method with established approaches and software. Further, the proposed method is used to re-estimate a regression on reading proficiency using US PISA data from 2000. Supplementary files for this article are available online.

摘要

混合模型经常用于社会和经济分析。通常，分析数据是通过来自感兴趣人群的调查样本收集的。基础调查设计可能是不可忽视的，如果不加以考虑，可能会导致回归参数出现偏差。为了避免这个问题，调查权重应该包括在估计过程中。我们提出了一个调查加权广义线性混合模型，允许特定单位的调查权重和灵活的随机效应结构。在不同场景下的两个模拟研究中提出并评估了估计程序。如果调查设计不可忽略，第一项模拟研究强烈鼓励使用调查加权广义线性混合模型。第二个复制了以前的模拟研究，以研究所提出的方法与已建立的方法和软件的竞争力。此外，所提出的方法用于使用 2000 年的美国 PISA 数据重新估计阅读能力的回归。本文的补充文件可在线获取。