定量数据考虑基线与否的几种方差分析模型的模拟比较

目的比较几种考虑基线与否的方差分析模型的统计性能。方法应用Monte Carlo技术,在基线均衡和不均衡情况下,比较以下方差分析模型:以基线为协变量的变化量协方差分析(ANCOVA)、变化率协方差分析(PCS-ANCOVA)和对数变化率协方差分析(logPCS-ANCOVA);不考虑基线的变化量方差分析(ANOVA)、变化率方差分析(PCS-ANOVA)和对数变化率方差分析(logPCS-ANOVA)。以I类错误与检验效能评价各种方法的统计性能。结果在基线均衡的情况下,PCS-ANCOVA和ANOVA均可很好地控制I类错误,且检验效能都较高;在基线不均衡的条件下,若基线对因变量无影响,ANCOVA与ANOVA均可以较好地控制I类错误,此时ANOVA的检验效能高于ANCOVA;若基线对因变量有影响时,只有ANCOVA可以很好地控制I类错误,且检验效能较高,其他方法效果不佳。结论考虑到实际应用中绝大部分情况是基线对因变量有影响,即相关,建议优先采用以基线为协变量的协方差分析或变化量的协方差分析,无论基线是否均衡。用变化率做方差分析或协方差分析,有可能冒着比值的分布不满足参数方法条件的风险,应用时应慎重。

A Simulation Comparison of Several Models of Analysis of Variance for Quantitative Data Considering Baseline or Not

Objective To compare the statistical performance of several ANOVA models considering baseline or not.Methods Monte Carlo technique was used to compare the following analysis of variance models in baseline equilibrium and imbalance:variation covariance analysis(ANCOVA),percent of change score with covariance analysis(PCS-ANCOVA),and logarithm of percent of change score with covariance analysis(logPCS-ANCOVA);Analysis of variance(ANOVA),percent of change score with analysis of variance(PCS-ANOVA),and logarithm of percent of change score with analysis of variance(logPCS-ANOVA).The statistical performance of those methods was evaluated by type I errors and power.Results In the case of baseline equilibrium,both PCS-ANCOVA and ANOVA can control type I errors well,and their powers are high under the condition of baseline imbalance,if the baseline has no effect on the dependent variable,both ANCOVA and ANOVA can control type I errors well,and the power of ANOVA is higher than that of ANCOVA;if the baseline has an effect on the dependent variable,only ANCOVA can control type I errors well,and have a high power,while other methods are not effective.Conclusion Considering that most of the actual clinical trials are that the baseline has an effect on the dependent variable,that is,relevant,it is recommended to use analysis of covariance with the baseline as a covariate or the analysis of covariance with change score,regardless of whether the baseline is balanced.Using the percent of change score for analysis of variance or analysis of covariance,there may be risk that the distribution of the ratio does not meet the condition of parameters method,and the application should be cautious.