缺失数据下一致性系数 AC₁ 不同处理方法的比较

打开文本图片集

Abstract：Objective Toexplore the impactof differentmissing data handling methodson AC1 coefficient estimation through simulation studies.Methods Monte Carlo simulation was used to generate evaluation data under diferent missing mechanisms.Theparameters generated includedthe number ofraters，categories，sample size，diseaseprevalence，random rating probability，and mising proportion.Four missing data handling methods，by excluding subjectswith zero ratings， excluding subjectswithncompleteatings，atermodeimputationandsubjectmodeimputation，werecomparedusingbias andmeansquarederror （MSE）as metrics.Results Whendisease prevalencewas balancedorthe misingdata mechanism was missing completelyat random （MCAR）or at random （MAR），excluding subjectswith zero ratings showed the best performance，with bias andMSE close to zero at a missing proportionbelow 30% . Under skewed prevalence and missing not atrandom（MNAR），subjectmode imputationwassuperiorfor AC1 coefficientestimation，resultinginabiaswithin ±0.10 and an MSE below 0.o9; for a sufficient sample size and a missing proportion ⩽30% ，theMSE of this method was nearly zero.Rater modeimputation showed the worst performanceacrossallthese scenarios.Excluding subjectswith incompleteratings resultedinanacceptableerroronlyinrelativelysimplesettings（tworatersandtwocategories）withlowamissingproportion underMCAR/MAR，butshowed a poor stabityinotherscenarios.Conclusion No universallyoptimal method exists for handling mising data inACestimation. Werecommend excluding subjects with zero ratings for balanced prevalence or MCAR/MAR， and subject mode imputation for skewed prevalence under MNAR. Researchers should report AC1 estimates from multiple methods to allow assessment of result sensitivity.

Keywords： agreement evaluation; nominal ratings; AC1 coefficient; missing data

在临床环境中，常需对患者的疾病状态进行分类，例如，心理疾病的诊断、影像学肿瘤分型等。（剩余11223字）