Multivariate zero-inflated generalised poisson data generation methods for simulating counts of adverse events

Counts of maximum-grade adverse events collected in clinical trials are important measures of treatment toxicity and tolerability. Analyzing the frequencies and correlations of adverse event counts by type, treatment cycle, and grade provides deeper insight into toxicity profiles. A prerequisite for developing such inferential methods is the ability to generate multivariate count data with specified event rates and correlation structures. In this article, we present three methods for simulating multivariate count data with zero-inflated generalised Poisson (ZIGP) distributions. The methods accommodate arbitrarily specified pairwise correlations within the feasible range for the target distribution. We develop them under the Normal-to-Anything (NORTA) and Sample-Iterate (SI) simulation frameworks. Simulation studies show strong performance in reproducing desired rates, scales, zero-inflation, and correlation matrices. We apply the methods to simulate AE counts based on the NCCTG N9741 multicenter randomised phase III colorectal cancer trial. We also illustrate broader applicability by simulating hospital visit counts using data from the National Medical Expenditure Survey.

临床试验中收集的最高级别不良事件（adverse event, AE）计数，是评估治疗毒性与耐受性的重要指标。按事件类型、治疗周期及不良事件等级分析其计数的频率与相关性，可深入揭示治疗毒性谱特征。开发此类推断方法的核心前提，是能够生成具备指定事件发生率与相关性结构的多元计数数据。本文提出三种可模拟服从零膨胀广义泊松（zero-inflated generalised Poisson, ZIGP）分布的多元计数数据的方法。此类方法可在目标分布的可行取值范围内，适配任意指定的两两相关性参数。本文基于正态到任意分布（Normal-to-Anything, NORTA）与样本迭代（Sample-Iterate, SI）两种模拟框架完成了上述方法的开发。模拟研究结果显示，该方法在复现预设发生率、尺度参数、零膨胀特性与相关矩阵方面表现优异。我们将所提方法应用于基于NCCTG N9741多中心随机Ⅲ期结直肠癌临床试验的不良事件计数模拟。此外，我们借助国家医疗支出调查（National Medical Expenditure Survey）的数据集模拟医院就诊次数，以此展示该方法更广泛的应用场景。