Chalmers T C, Lau J
Harvard School of Public Health, Boston, MA.
Stat Methods Med Res. 1993;2(2):161-72. doi: 10.1177/096228029300200204.
The advent of meta-analysis, especially when performed cumulatively, raises many questions about how best to approach the conduct of clinical trials in the evaluation of new treatments. We need to be assured that bias is minimized by proper experimental procedures and that clinical data, on the whole and in subgroups, are presented so that they can be effectively combined in meta-analysis. We need to re-examine the idea that we should not start a randomized control trial unless sufficient patients are available to avoid reasonable type I and II errors. Meta-analyses will come to the rescue, provided trials continue to be published at the present rate. We need to perform meta-analyses before undertaking each additional trial, and we need to base estimates of trial size on past data as well as the expected control rates and the differences we do not want to miss. In clinical trials of new interventions attempting to disprove the null hypothesis may be inappropriate because past data so often suggest or even establish that it is not true. Furthermore we need to recognize that trends (p > 0.05) can be both clinically and statistically important, and we must abandon the notion that if p is not < 0.05, the treatment is ineffective. In performing meta-analyses we need to worry about minimizing bias and error and consider the differences between the random and fixed effects models and between reporting results as an odds ratio versus difference in risk, with the control rates given. Experiences with cumulative meta-analysis have required that we think about all of these problems.
荟萃分析的出现，尤其是累积进行时，引发了许多关于在评估新疗法时如何以最佳方式开展临床试验的问题。我们需要确保通过适当的实验程序将偏差降至最低，并且临床数据在整体和亚组层面都能以可在荟萃分析中有效合并的方式呈现。我们需要重新审视这样一种观点，即除非有足够数量的患者以避免合理的I型和II型错误，否则不应启动随机对照试验。如果试验继续以目前的速度发表，荟萃分析将发挥作用。我们需要在进行每一项额外试验之前进行荟萃分析，并且我们需要根据过去的数据以及预期的对照率和我们不想错过的差异来估计试验规模。在试图推翻无效假设的新干预措施的临床试验中，这样做可能不合适，因为过去的数据常常表明甚至证实无效假设并不成立。此外，我们需要认识到趋势（p>0.05）在临床和统计上都可能很重要，我们必须摒弃那种认为如果p不小于0.05，治疗就是无效的观念。在进行荟萃分析时，我们需要关注将偏差和误差降至最低，并考虑随机效应模型和固定效应模型之间的差异，以及在给出对照率的情况下，将结果报告为比值比与风险差异之间的差异。累积荟萃分析的经验要求我们思考所有这些问题。
