文献速递流行病学与生物统计医学人工智能与信息学

罕见事件荟萃分析中超几何正态随机效应模型似然比推断的自举校准

Bootstrap Calibration of Likelihood-Ratio Inference for the Hypergeometric-Normal Random-Effects Model in Rare-Event Meta-Analysis.
PubMed 文献 · Stat Med · 2026-01-01

针对罕见事件荟萃分析中正态近似和固定方差假设易被违反、零单元格校正可能引入偏倚的问题,探讨超几何正态(HGN)随机效应模型的似然比(LR)推断,并提出两种参数自举校准方法。模拟显示未校正LR置信区间已具稳定覆盖,自举均值校准在极端罕见事件下校准最一致,并应用于罗格列酮与心肌梗死荟萃分析。

要点

  • HGN模型可实现基于精确条件似然的推断,避免零单元格校正偏倚。
  • 未校正LR置信区间在罕见事件模拟中已显示稳定覆盖。
  • 自举均值校准在极端罕见事件设定下提供最一致的额外校准。

核心结论

  • HGN随机效应模型的LR推断在罕见事件条件下表现良好,自举均值校准可进一步改善极端罕见事件下的有限样本校准。

研究设计与样本

  • 方法学研究,通过广泛模拟研究评估罕见事件条件下LR推断及两种参数自举校准程序的性能,并应用于罗格列酮与心肌梗死荟萃分析。

主要发现

  • 未校正LR置信区间在罕见事件模拟中已提供稳定覆盖。
  • 自举均值校准在最具挑战的极端罕见事件设定下通常提供最一致的额外校准。
  • 直接自举推断提供了基于分布的另一替代方法。

临床怎么用

  • 在罕见事件荟萃分析中可考虑采用HGN模型的LR推断,并在极端稀疏数据下使用自举均值校准以改善置信区间校准。

局限性

  • 普通Wald型推断在小样本和稀疏数据设定下可能仍不准确。
  • 零单元格时Haldane型(+0.5)校正可能进一步偏倚估计效应。

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

Conventional random-effects models in meta-analysis rely on normal approximations and fixed within-study variance assumptions that are often violated in rare-event settings. When zero cells occur, Haldane-type (+0.5) corrections can further bias the estimated effects. To overcome these limitations, likelihood-based approaches such as the hypergeometric-normal (HGN) model have been developed to enable exact conditional likelihood-based inference. However, ordinary Wald-type inference may still be inaccurate in small and sparse data settings. In this article, we investigate likelihood-ratio (LR)-based inference for the HGN random-effects model and propose two parametric bootstrap procedures for improving its finite-sample calibration: (1) bootstrap mean calibration of the LR statistic and (2) bootstrap inference based directly on the empirical distribution of the LR statistic. Extensive simulation studies under rare-event conditions demonstrate that the uncorrected LR-based confidence interval already provides stable coverage, reflecting the favorable properties of the exact conditional likelihood. Bootstrap mean calibration generally provided the most consistent additional calibration under the most challenging extremely rare-event settings, whereas direct bootstrap inference provided an alternative distribution-based approach. An application to a meta-analysis of rosiglitazone and myocardial infarction is also presented.

来源:Stat Med,查看原文。版权归原作者所有。

引用

  1. Noma H, Evrenoglou T, Maruo K, et al. Bootstrap Calibration of Likelihood-Ratio Inference for the Hypergeometric-Normal Random-Effects Model in Rare-Event Meta-Analysis. Stat Med. 2026. PMID: 42869622. doi:10.1002/sim.70776

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。