Oral and Maxillofacial Radiology, Applied Oral Sciences and Community Dental Care, Faculty of Dentistry, The University of Hong Kong, Hong Kong, China.
PLoS One. 2024 Sep 3;19(9):e0309813. doi: 10.1371/journal.pone.0309813. eCollection 2024.
In current neuroimaging studies, the mainstream practice is to report results corrected for multiple comparisons to control for false positives. In 2009, Lieberman and Cunningham published a highly cited report that promotes the use of uncorrected statistical thresholds to balance Types I and II error rates. This paper aims to review recent studies that cited this report, investigating whether the citations were to justify the use of uncorrected statistical thresholds, and if their uncorrected thresholds adhered to the recommended defaults.
The Web of Science Core Collection online database was queried to identify original articles published during 2019-2022 that cited the report.
It was found that the majority of the citing papers (152/225, 67.6%) used the citation to justify their statistical threshold setting. However, only 19.7% of these 152 papers strictly followed the recommended uncorrected P (Punc) < 0.005, k = 10 (15/152, 9.9%) or Punc < 0.005, k = 20 (15/152, 9.9%). Over half (78/152, 51.3%) used various cluster-extent based thresholds with Punc, with the predominant choices being Punc < 0.001, k = 50 and Punc < 0.001, k = 10, mostly without justifying their deviation from the default. Few papers matched the voxel size and smoothing kernel size used by the simulations from the report to derive the recommended thresholds.
This survey reveals a disconnect between the use and citation of Lieberman and Cunningham's report. Future studies should justify their chosen statistical thresholds based on rigorous statistical theory and study-specific parameters, rather than merely citing previous works. Furthermore, this paper encourages the neuroimaging community to publicly share their group-level statistical images and metadata to promote transparency and collaboration.
在当前的神经影像学研究中，主流做法是报告经过多重比较校正以控制假阳性的结果。2009 年，Lieberman 和 Cunningham 发表了一篇高引用率的报告，提倡使用未经校正的统计阈值来平衡 I 型和 II 型错误率。本文旨在回顾引用该报告的近期研究，调查这些引用是否是为了证明使用未经校正的统计阈值的合理性，以及他们的未经校正的阈值是否符合推荐的默认值。
通过 Web of Science Core Collection 在线数据库查询，确定 2019-2022 年期间发表的引用该报告的原始文章。
发现大多数引用论文（152/225，67.6%）使用该引文来证明其统计阈值设置的合理性。然而，只有 19.7%的这 152 篇论文严格遵循了推荐的未校正 P（Punc）<0.005，k=10（15/152，9.9%）或 Punc<0.005，k=20（15/152，9.9%）。超过一半（78/152，51.3%）使用各种基于簇扩展的 Punc 阈值，其中最主要的选择是 Punc<0.001，k=50 和 Punc<0.001，k=10，大多没有为偏离默认值提供正当理由。很少有论文与报告中模拟得出的推荐阈值的体素大小和平滑核大小相匹配。
这项调查揭示了 Lieberman 和 Cunningham 报告的使用和引用之间存在脱节。未来的研究应该根据严格的统计理论和研究特定的参数来证明他们选择的统计阈值的合理性，而不仅仅是引用以前的工作。此外，本文鼓励神经影像学界公开共享他们的组级统计图像和元数据，以提高透明度和合作。
