医学论文中成组t检验P值错误及其原因分析

P-value errors in two-sample t-test of medical papers and reason analysis

【目的】调查医学论文中成组t检验的P值错误,分析错误原因,并提出相应措施。【方法】选取236种医学期刊,每种期刊选取1项成组t检验,核验其P值,应用χ~2检验、Mantel-Haenszel法、二项式logistic回归分析P值错误。【结果】236项成组t检验中,50项存在P值错误。单因素分析结果显示,核心期刊与非核心期刊相比,P值错误发生率差异具有统计学意义(χ~2=4.871,P=0.027);给出具体P值组与未给出具体P值组相比,P值错误发生率差异具有统计学意义(χ~2=15.440,P<0.0001)。将是否给出具体P值作为混杂因素,比较核心期刊与非核心期刊P值错误发生率,差异无统计学意义(χ~2=2.703,P=0.100)。多因素分析结果显示,是否方差齐(OR值为0.470,95%CI为0.230～0.961)、是否给出具体P值(OR值为5.459,95%CI为2.311～12.895)具有统计学意义。【结论】医学论文成组t检验P值错误较多。为及时发现P值错误,期刊编辑应当重视对统计学方法应用条件的审查,要求作者给出统计描述以及统计推断的具体结果,能够利用简单易学的统计学软件核实P值。

[Purposes]This paper investigates P-value errors in two-sample t-test of medical papers,analyzes the causes of errors,and puts forward the corresponding measures.[Methods]We reviewed 236two-sample t-test cases from 236 medical journals.The P-value was verified and the P-value errors were analyzed by χ^2 test,Mantel-Haenszel method and binomial logistic regression.[Findings]Among the 236cases,50cases have P-value errors.The univariate analysis shows that there is a significant difference between core journals and non-core journals in incidence of P-value errors (χ^2=4.871,P =0.027),and there is a significant difference between cases with concrete P-value and cases without concrete P-value in incidence of P-value errors (χ^2=15.440,P<0.0001).When we regard whether the concrete P-value is described as the confounding factor,no significant difference in the incidence of P-value errors is found between core journals and non-core journals (χ^2=2.703,P =0.100).The multivariate analysis shows that whether the variance is equal (OR =0.470,95% CI:0.230-0.961)and whether the concrete P-value is described (OR =5.459,95% CI:2.311-12.895)are significant variables.[Conclusions ]There is a high incidence of P-value errors in two-sample t-test of medical papers.In order to discover the P-value errors in time,editors should pay attention to the examination of the application conditions of statistical methods,ask the authors to give statistical description and statistical inference result,and verify P-value by simple statistical software.