证据合成中报告与解释统计学非显著结果的方法:一项系统综述
该研究系统检索了2000年1月1日至2025年1月30日的多个数据库及指南网络,纳入37篇文献,评估证据合成中报告与解释具有临床相关性的统计学非显著结果的方法及其方法学质量和实证验证程度。结果显示多种方法被提出但尚无广泛共识,且均未在利益相关者中实证检验,作者认为需系统测试并制定循证指导。
要点
- 证据合成中解释非显著结果的方法多样但无共识。
- 不确定性区间与统计量结合是最常见方法(15/37,41%)。
- 所有方法均未在利益相关者中实证检验。
核心结论
- 虽已提出多种方法改进统计学非显著结果的呈现与解释,但尚无广泛共识且未系统测试其实用性和有效性。
- 作者认为需在利益相关者中系统测试所提方法并制定循证指导以支持非显著结果的恰当解释。
研究设计与样本
- 系统综述,检索Ovid MEDLINE ALL、Scopus、PsycInfo、Library of Guidance for Health Scientists、MathSciNet及EQUATOR Network,纳入2000年1月1日至2025年1月30日发表的英文文献。
- 共纳入37篇文献,包括15篇社论或观点文章、9篇方法学文献、8篇横断面或混合方法研究、4篇期刊指导文件和1篇系统综述。
- 采用混合方法评价工具、横断面研究评价工具及Joanna Briggs研究所专家意见和系统综述清单进行质量评价,至少两名评价者独立完成,大语言模型辅助数据提取和质量评价。
主要发现
- 不确定性区间及其与各种统计量结合为15/37(41%),贝叶斯因子为6/37(16%),似然比为3/37(8%),效应转换量为2/37(5%),等效性检验为2/37(5%),改良Fisher检验为1/37(3%),反向脆弱指数为1/37(3%)。
- 报告实践中存在有问题的「无效接受」语言(14/37,38%),部分文献劝阻基于非显著结果不当声称无效应(9/37,24%)。
- 27篇文献符合其研究设计或发表类型对应的评价工具质量标准,10篇不符合,包括1篇系统综述、2篇社论或观点文章和7篇横断面研究。
- 所有提出的方法均未在利益相关者中进行实证检验。
关键数值与推荐
- 不确定性区间与统计量结合:15/37(41%)。
- 贝叶斯因子:6/37(16%)。
- 似然比:3/37(8%)。
- 效应转换量:2/37(5%)。
- 等效性检验:2/37(5%)。
临床怎么用
- 原文未据此给出推荐。
局限性
- 所有提出的方法均未在利益相关者中进行实证检验。
- 尚无广泛接受的共识,因为这些方法尚未被系统测试其实用性和有效性。
- 一些方法缺乏指导结果解释的阈值。
- 作者认为需要在利益相关者中系统测试所提方法并制定循证指导。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
OBJECTIVES: To systematically review approaches for reporting and interpreting statistically nonsignificant findings with clinical relevance in evidence synthesis and to assess their methodological quality and the extent of their empirical validation. STUDY DESIGN AND SETTING: We searched Ovid MEDLINE ALL, Scopus, PsycInfo, Library of Guidance for Health Scientists, and MathSciNet for published studies in English from January 1, 2000, to January 30, 2025, for (1) best practices in guidance documents for evidence synthesis when interpreting clinically relevant nonsignificant findings, (2) statistical methods to support the interpretation, and (3) reporting practices. To identify relevant reporting guidelines, we also searched the Enhancing the QUAlity and Transparency Of health Research Network. The quality assessment applied the Mixed Methods Appraisal Tool, Appraisal tool for Cross-Sectional Studies, and checklists for expert opinion and systematic reviews from the Joanna Briggs Institute. At least two reviewers independently conducted all procedures, and a large language model facilitated data extraction and quality appraisal. RESULTS: Of the 5332 records, 37 were eligible for inclusion. Of these, 15 were editorials or opinion pieces, nine addressed methods, eight were cross-sectional or mixed-methods studies, four were journal guidance documents, and one was a systematic review. Twenty-seven records met the quality criteria of the appraisal tool relevant to their study design or publication type, while 10 records, comprising one systematic review, two editorials or opinion pieces, and seven cross-sectional studies, did not. Relevant methodological approaches to evidence synthesis included utilization of uncertainty intervals and their integration with various statistical measures (15 of 37, 41%), Bayes factors (six of 37, 16%), likelihood ratios (three of 37, 8%), effect conversion measures (two of 37, 5%), equivalence testing (two of 37, 5%), modified Fisher's test (one of 37, 3%), and reverse fragility index (one of 37, 3%). Reporting practices included problematic "null acceptance" language (14 of 37, 38%), with some records discouraging the inappropriate claim of no effect based on nonsignificant findings (nine of 37, 24%). None of the proposed methods were empirically tested with interest holders. CONCLUSION: Although various approaches have been proposed to improve the presentation and interpretation of statistically nonsignificant findings, a widely accepted consensus has not emerged, as these approaches have yet to be systematically tested for their practicality and validity. This review provides a comprehensive review of available methodological approaches spanning both the frequentist and Bayesian statistical frameworks and identifies critical gaps in empirical validation of some approaches, namely the lack of thresholds to guide the interpretation of results. These findings highlight the need for systematic testing of proposed methods with interest holders and the development of evidence-based guidance to support appropriate interpretation of nonsignificant results in evidence synthesis. PLAIN LANGUAGE SUMMARY: This review looked at how to best report results that are not statistically significant because some of these findings can still be important to inform clinical care or health policy. We searched databases for studies published between 2000 and 2025. Out of more than 5000 records, 37 studies were relevant. These studies showed that there is no single best way to report nonsignificant findings.
引用
- Sharifan A, Dobrescu A, Harrod C, et al. Approaches for reporting and interpreting statistically nonsignificant findings in evidence syntheses: a systematic review. J Clin Epidemiol. 2025. PMID: 41276091. doi:10.1016/j.jclinepi.2025.112083
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。