Scholarlink

<p>Model estimation for SEM analyses in commonly used software typically involves iterative optimization procedures, which can lead to nonconvergence issues. In this paper, we propose using random starting values as an alternative to the current default strategies. By drawing from uniform distributions within data-driven lower and upper bounds (see De Jonckere et al. (2022) <i>Structural Equation Modeling: A Multidisciplinary Journal,</i> <i>29</i>(3), 412–427), random starting values are generated for each (free) parameter in the model. Through three small simulation studies, we demonstrate that incorporating such bounded random starting values significantly reduces the nonconvergence rate, resulting in increased convergence rates ranging between 87% and 96% in the first two studies. In essence, bounded random starting values seem to offer a promising alternative to the default starting values that are currently used in most software packages.</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

A note on using random starting values in small sample SEM

  • 期刊论文

  • 发表日期：

    2025年1月14日

查看全文
  • Julie De Jonckere,
  • Yves Rosseel

摘要

Model estimation for SEM analyses in commonly used software typically involves iterative optimization procedures, which can lead to nonconvergence issues. In this paper, we propose using random starting values as an alternative to the current default strategies. By drawing from uniform distributions within data-driven lower and upper bounds (see De Jonckere et al. (2022) Structural Equation Modeling: A Multidisciplinary Journal, 29(3), 412–427), random starting values are generated for each (free) parameter in the model. Through three small simulation studies, we demonstrate that incorporating such bounded random starting values significantly reduces the nonconvergence rate, resulting in increased convergence rates ranging between 87% and 96% in the first two studies. In essence, bounded random starting values seem to offer a promising alternative to the default starting values that are currently used in most software packages.