Misbegotten Methodologies and Forgotten Lessons From Tom Swift's Electric Factor Analysis Machine: A Demonstration With Competing Structural Models of Psychopathology
Misbegotten Methodologies and Forgotten Lessons From Tom Swift's Electric Factor Analysis Machine: A Demonstration With Competing Structural Models of Psychopathology
复制标题
DOI：
10.1037/met0000465
复制
发表时间：
2022-01-06
影响因子：
7
通讯作者：
Eaton, Nicholas R.
中科院分区：
文献类型：
--
作者：
Greene, Ashley L.;Watts, Ashley L.;Eaton, Nicholas R.
验证性因子分析（Confirmatory factor analysis， CFA）是涉及心理构念双因素结构的模型比较研究中最常用的方法。然而，在构建和测试双因素CFA模型时，存在一些明显的缺陷，这些缺陷会扭曲结果，限制推广，阻碍理论发展。本研究的总体目标是为应用研究人员提供一个概念清晰的路线图来评估潜在的因素结构，对验证性和探索性因素分析场景中出现的理论和方法上的挑战进行全面的核算，并强调两者结合使用的好处。这些一般问题是在一个应用的例子中说明的，该例子具有在定量精神病理学领域中经常使用的数据集。在双因素CFA和EFA模型中出现了一些统计和概念上的相似性。结果位于精神病理学结构模型的更广泛的文献中，并对p因子理论的含义进行了讨论。本研究的历史叙述追溯了一个世纪以来普遍因素辩论的反复出现的主题，以告诫读者不要在单因素分析解决方案的基础上发展理论。验证性因子分析及其双因子模型在心理构念因素结构的实证研究中非常流行。CFA提供了直接的假设检验，但也有明显的缺陷，比如强加严格的假设（即简单的结构），掩盖了未建模的复杂性。由于双因子CFAs的局限性，它们在样本和研究中产生了异常结果，表明模型规范错误（例如，蒸发特定因子和意外负载）。我们建议使用探索性因子分析（EFA）来评估CFA解决方案的结构有效性-无论是在更严格的CFA模型估计之前还是之后-以(a)识别可能导致异常估计的模型错误规范和(b)通过检查假设结构是否在有限的研究者输入下出现来确认CFA模型。我们评估了沿着探索性-验证性连续体的主要因素结构在不同背景下不变的程度，并展示了糟糕的方法选择如何扭曲结果并阻碍理论发展。所有CFA模型都很适合，但在可复制性和实质性可解释性方面存在许多差异。双因素CFA和EFA模型之间出现了一些相似之处，包括过度提取的证据，特定因素对一般因素的崩溃，以及随后如何定义一般因素的变化。我们将这些方法学上的缺陷置于更广泛的精神病理学结构模型文献中，阐明了由因子分析产生的理论（如p因子）的含义，概述了在进行探索性双因子分析时遇到的问题的几种补救措施，并提出了验证性双因子模型的替代规范。
Translational Abstract Confirmatory factor analysis (CFA) is the most popular method in model comparison studies involving bifactor structures of psychological constructs. However, notable pitfalls are associated with myopic approaches to building and testing bifactor CFA models, which can distort results, limit generalizability, and impede theory development. The overarching aims of this study are to provide applied researchers with a conceptually clear roadmap for evaluating latent factor structures, present a comprehensive accounting of theoretical and methodological challenges that arise in both confirmatory and exploratory factor analytic scenarios, and highlight the benefits of their combined use. These general issues are illustrated in an applied example featuring a frequently used dataset in the field of quantitative psychopathology. Several statistical and conceptual similarities emerged across bifactor CFA and EFA models. Results are situated within the broader literature on structural models of psychopathology, and implications for p-factor theories are discussed. The historical narrative that is woven into this study traces recurring themes of the century-old general factor debate to caution readers against the development of theories on the basis of a single factor analytic solution.Confirmatory factor analysis (CFA) and its bifactor models are popular in empirical investigations of the factor structure of psychological constructs. CFA offers straightforward hypothesis testing but has notable pitfalls, such as the imposition of strict assumptions (i.e., simple structure) that obscure unmodeled complexity. Due to the limitations of bifactor CFAs, they have yielded anomalous results across samples and studies that suggest model misspecification (e.g., evaporating specific factors and unexpected loadings). We propose the use of exploratory factor analysis (EFA) to evaluate the structural validity of CFA solutions-either before or after the estimation of more restrictive CFA models-to (a) identify model misspecifications that may drive anomalous estimates and (b) confirm CFA models by examining whether hypothesized structures emerge with limited researcher input. We evaluated the degree to which predominant factor structures were invariant across contexts along the exploratory-confirmatory continuum and demonstrate how poor methodological choices can distort results and impede theory development. All CFA models fit well, but there were numerous differences in replicability and substantive interpretability. Several similarities emerged between bifactor CFA and EFA models, including evidence of overextraction, the collapse of specific factors onto the general factor, and subsequent shifts in how the general factor was defined. We situate these methodological shortcomings within the broader literature on structural models of psychopathology, articulate implications for theories (such as the p-factor) that are borne out of factor analysis, outline several remedies for problems encountered when performing exploratory bifactor analysis, and propose alternative specifications for confirmatory bifactor models.