纵向临床试验中患者报告结局缺失数据的处理:一项模拟研究
该研究通过蒙特卡洛模拟比较了多重插补、机器学习方法和完整病例分析在纵向患者报告结局缺失数据中的表现。结果显示MICE和随机森林在多数条件下偏倚最小且精度最高,而完整病例分析在高比例非随机缺失时产生严重偏倚,KNN始终最不准确。作者建议采用MICE和随机森林等稳健方法并进行敏感性分析。
要点
- 完整病例分析在高比例非随机缺失时偏倚严重,应避免默认使用。
- MICE和随机森林在多数模拟条件下偏倚最小且精度最高。
- 推荐采用稳健插补方法并进行敏感性分析以确保PRO研究有效性。
核心结论
- 默认使用完整病例分析等简单方法存在产生显著偏倚的高风险。
- 推荐采用MICE和随机森林等稳健方法并进行敏感性分析框架。
研究设计与样本
- 蒙特卡洛模拟研究,共生成36,000个数据集。
- 样本量设为100、200、300,缺失比例设为20%、40%、60%。
- 缺失机制包括随机缺失(MAR)和非随机缺失(MNAR)。
- 比较四种方法:链式方程多重插补(MICE)、随机森林(RF)、K近邻(KNN)和完整病例分析(CCA)。
主要发现
- MICE和RF在多数条件下提供偏倚最小且精度最高的估计。
- CCA在高比例(60%)MNAR缺失下产生严重且临床显著的偏倚(偏倚 > -3.9分)。
- KNN始终是准确性最差的方法。
临床怎么用
- 在纵向PRO研究中应避免默认使用完整病例分析。
- 优先考虑MICE或随机森林处理缺失数据,并开展敏感性分析。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
OBJECTIVES: To compare the performance of multiple imputation (MI), machine-learning (ML) methods, and complete case analysis (CCA) for handling missing data in longitudinal patient-reported outcomes (PROs) under varying missing data mechanisms. STUDY DESIGN AND SETTING: A Monte Carlo simulation study was conducted with 36,000 datasets, varying sample sizes (N = 100, 200, 300), missingness proportions (20%, 40%, and 60%), and mechanisms (missing at random [MAR] and missing not at random [MNAR]). Four imputation methods were compared: Multivariate imputation by chained equations (MICE), random forest (RF), k-nearest neighbors (KNN), and CCA. RESULTS: MICE and RF provided the least biased and most precise estimates across most conditions. Conversely, CCA produced severe and clinically significant bias under high (60%) MNAR missingness (bias > -3.9 points), while KNN was consistently the least accurate method. CONCLUSION: Defaulting to simple methods like CCA carries a high risk of substantial bias. To ensure the validity of findings in PRO research, we recommend a sensitivity analysis framework that employs robust methods like MICE and RF. PLAIN LANGUAGE SUMMARY: When patients do not complete all questionnaires in clinical trials, simple methods like ignoring incomplete cases can lead to wrong conclusions. Advanced methods like MI or ML provide more reliable results, especially when sicker patients are more likely to miss follow-ups.
引用
- Khan H. Handling missing patient-reported outcomes in longitudinal clinical trials: a simulation study. J Clin Epidemiol. 2026. PMID: 41905724. doi:10.1016/j.jclinepi.2026.112254
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。