Ioannidis John P A
Department of Hygiene and Epidemiology, University of Ioannina School of Medicine, Ioannina, Greece.
J Eval Clin Pract. 2008 Oct;14(5):951-7. doi: 10.1111/j.1365-2753.2008.00986.x.
Statistical tests of heterogeneity and bias, in particular publication bias, are very popular in meta-analyses. These tests use statistical approaches whose limitations are often not recognized. Moreover, it is often implied with inappropriate confidence that these tests can provide reliable answers to questions that in essence are not of statistical nature. Statistical heterogeneity is only a correlate of clinical and pragmatic heterogeneity and the correlation may sometimes be weak. Similarly, statistical signals may hint to bias, but seen in isolation they cannot fully prove or disprove bias in general, let alone specific causes of bias, such as publication bias in particular. Both false-positive and false-negative signals of heterogeneity and bias can be common and their prevalence may be anticipated based on some rational considerations. Here I discuss the major common challenges and flaws that emerge in using and interpreting statistical tests of heterogeneity and bias in meta-analyses. I discuss misinterpretations that can occur at the level of statistical inference, clinical/pragmatic inference and specific cause attribution. Suggestions are made on how to avoid these flaws, use these tests properly and learn from them.
异质性和偏倚的统计检验，尤其是发表偏倚，在荟萃分析中非常常见。这些检验采用的统计方法，其局限性往往未被认识到。此外，人们常常不恰当地自信地认为，这些检验能够为本质上并非统计学性质的问题提供可靠答案。统计异质性只是临床和实际异质性的一个相关因素，而且这种相关性有时可能很弱。同样，统计信号可能暗示存在偏倚，但单独来看，它们一般无法完全证明或反驳偏倚的存在，更不用说偏倚的具体原因，尤其是发表偏倚了。异质性和偏倚的假阳性和假阴性信号都可能很常见，基于一些合理的考虑，可以预期它们的发生率。在此，我将讨论在荟萃分析中使用和解释异质性和偏倚的统计检验时出现的主要常见挑战和缺陷。我将讨论在统计推断、临床/实际推断以及具体原因归因层面可能出现的误解。并就如何避免这些缺陷、正确使用这些检验以及从中吸取经验教训提出建议。
