逆概率权重法在诊断试验评价证实偏倚中的应用

Implication of inverse-probability weighting method in the evaluation of diagnostic test withverification bias

如何估计和校正筛查或诊断试验中存在的证实偏倚，文中通过宫颈癌筛查实例，采用逆概率权重法R软件CompareTests校正其灵敏度和特异度，利用随机抽样方法生成新数据，将逆概率权重法计算的灵敏度和特异度，与传统计算方法以及最大似然估计方法计算得到的灵敏度和特异度进行比较。结果表明HPV自检法的真实灵敏度和特异度分别为83．53％（95％CI：74．23．89．93）和85．86％（95％CI：84．23～87．36）。随机抽样结果显示，传统方法计算的灵敏度和特异度分别为90．48％（95％C1：80．74～95．56）和71．96％（95％CI：68．71．75．00），：果用逆概率权重法校正后的灵敏度和特异度分别为82．25％（95％C1：63．11～92．62）和85．80％（95％C1：85．09～86．47）；采用最大似然估计法校正后的灵敏度和特异度分别为80．13％（95％CI：66．81～93．46）和85．80％（95％CI：84．20～87．41）。表明在复杂抽样情况下，逆概率权重法能够有效校正；存在证实偏倚的灵敏度和特异度。

To evaluate and adjust the verification bias existed in the screening or diagnostic tests. Inverse-probability weighting method was used to adjust the sensitivity and specificity of the diagnostic tests, with an example of cervical cancer screening used to introduce the Compare Tests package in R software which could be implemented. Sensitivity and specificity calculated from the traditional method and maximum likelihood estimation method were compared to the results from Inverse-probability weighting method in the random-sampled example. The true sensitivity and specificity of the HPV self-sampling test were 83.53% （95% CI： 74.23-89.93） and 85.86% （95% CI： 84.23-87.36）. In the analysis of data with randomly missing verification by gold standard, the sensitivity and specificity calculated by traditional method were 90.48% （95% CI： 80.74-95.56） and 71.96% （95%CI： 68.71-75.00）, respectively. The adjusted sensitivity and specificity under the use of Inverse-probability weighting method were 82.25% （95% CI： 63.11-92.62） and 85.80% （95% CI： 85.09-86.47）, respectively, whereas they were 80.13% （95% CI： 66.81-c^3.46） and 85.80% （95% CI： 84.20-87.41 ） under the maximum likelihood estimation method. The inverse-probability weighting method could effectively adjust the sensitivity and specificity of a diagnostic test when verification bias existed, especially when complex sampling appeared.