Chen Yi-Kai, Yang Tong-Rong, Chen Li-Ting, Hsieh Cheng-Yu, Cheng Che, Wu Po-Ju, Peng Chao-Ying Joanne
Department of Psychology, National Taiwan University, Taipei City, Taiwan.
Department of Educational Studies, University of Nevada, Reno/MS 299, Reno, NV, 89557, USA.
Behav Res Methods. 2025 Sep 8;57(10):279. doi: 10.3758/s13428-025-02715-1.
g is a design-comparable effect size that has been recommended by the What Works Clearinghouse since 2020 to assess an intervention effect in single-case studies and for meta-analyses. Yet, no research has systematically studied how g's performance could be impacted by its non-convergence, and how g's non-convergence and performance could be improved by increasing the case size (m) and measurement size (N). This study expanded on the work of Pustejovsky et al. (Journal of Educational and Behavioral Statistics, 39(5), 368-393, 2014) and Chen et al. (Behavioral Research Methods, 56, 379-405, 2024) to investigate the impact of a wide range of m and N, data distribution, autocorrelation, within-case reliability, and ratio of variance components on g's non-convergence rate and performance in multiple-baseline designs. g's performance was assessed by relative bias, relative bias of variance, and coverage rate of 95% symmetric CIs. Findings revealed that g's performance was improved by convergence, especially when data were non-normal. In addition, g's convergence improved by increasing m, within-case reliability, and ratio of variance components. When data distribution was normal, converged g improved with large m and large within-case reliability. When data distribution was mildly non-normal, converged g improved with medium to large m and small within-case reliability. When data distribution was moderately non-normal, converged g improved with small to medium m and small within-case reliability. Optimal m depended on data distribution and within-case reliability. N had a trivial impact on converged g. In sum, our findings demonstrated the importance of g's convergence and optimal m to improve its application.
g是一种设计可比效应量，自2020年以来一直被有效证据中心推荐用于评估单案例研究和元分析中的干预效果。然而，尚无研究系统地探讨g的非收敛性如何影响其性能，以及如何通过增加案例数量（m）和测量次数（N）来改善g的非收敛性和性能。本研究在Pustejovsky等人（《教育与行为统计学杂志》，39(5)，368 - 393，2014）以及Chen等人（《行为研究方法》，56，379 - 405，2024）的工作基础上进行拓展，以研究在多基线设计中，广泛的m和N、数据分布、自相关性、案例内信度以及方差成分比例对g的非收敛率和性能的影响。通过相对偏差、方差相对偏差以及95%对称置信区间的覆盖率来评估g的性能。研究结果表明，收敛可改善g的性能，尤其是在数据非正态分布时。此外，增加m、案例内信度和方差成分比例可提高g的收敛性。当数据分布呈正态时，收敛的g在m大且案例内信度高时得到改善。当数据分布为轻度非正态时，收敛的g在m为中等到大且案例内信度低时得到改善。当数据分布为中度非正态时，收敛的g在m为小到中且案例内信度低时得到改善。最佳的m取决于数据分布和案例内信度。N对收敛的g影响微不足道。总之，我们的研究结果证明了g的收敛性和最佳m对于改善其应用的重要性。
