Scholarlink

<p>Propensity score (PS) methods are widely used in quasi-experimental and observational studies with clustered data in health and social science settings (e.g., students nested within schools). However, methodological research on their performance in causal mediation analyses with clustered data remains limited. Through simulation studies, we compare four PS models–single-level, fixed-effect, random-effect, and random-effect with cluster means–for estimating the natural (in)direct effects when unmeasured cluster-level confounders exist. Simulation results indicate that incorporating clusters in PS models yielded more accurate effect estimates than ignoring clusters (i.e., the single-level model). The methods are also illustrated using data from the National Longitudinal Study of Adolescent to Adult Health. Our study provides novel insights into using PS methods in causal mediation analyses with clustered observational data.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Evaluating propensity score models for estimating causal mediation effects under unmeasured cluster-level confounding

  • 期刊论文

  • 发表日期：

    2025年8月27日

查看全文
  • Cameron McCann,
  • Xiao Liu

摘要

Propensity score (PS) methods are widely used in quasi-experimental and observational studies with clustered data in health and social science settings (e.g., students nested within schools). However, methodological research on their performance in causal mediation analyses with clustered data remains limited. Through simulation studies, we compare four PS models–single-level, fixed-effect, random-effect, and random-effect with cluster means–for estimating the natural (in)direct effects when unmeasured cluster-level confounders exist. Simulation results indicate that incorporating clusters in PS models yielded more accurate effect estimates than ignoring clusters (i.e., the single-level model). The methods are also illustrated using data from the National Longitudinal Study of Adolescent to Adult Health. Our study provides novel insights into using PS methods in causal mediation analyses with clustered observational data.