How various design decisions on matching individuals in relationships affect the outcomes of microsimulations of sexually transmitted infection epidemics

Microsimulations are increasingly used to estimate the prevalence of sexually transmitted infections (STIs). These models consist of agents which represent a sexually active population. Matching agents into sexual relationships is computationally intensive and presents modellers with difficult design decisions: how to select which partnerships between agents break up, which agents enter a partnership market, and how to pair agents in the partnership market. The aim of this study was to analyse the effect of these design decisions on STI prevalence. We compared two strategies for selecting which agents enter a daily partnership market and which agent partnerships break up: random selection in which agents are treated homogenously versus selection based on data from a large German longitudinal data set that accounts for sex, sexual orientation and age heterogeneity. We also coupled each of these strategies with one of several recently described algorithms for pairing agents and compared their speed and outcomes. Additional design choices were also considered, such as the number of agents used in the model, increasing the heterogeneity of agents’ sexual behaviour, and the proportion of relationships which are casual sex encounters. Approaches which account for agent heterogeneity estimated lower prevalence than less sophisticated approaches which treat agents homogeneously. Also, in simulations with non-random pairing of agents, as the risk of infection increased, incidence declined as the number of agents increased. Our algorithms facilitate the execution of thousands of simulations with large numbers of agents quickly. Fast pair-matching algorithms provide a practical way for microsimulation modellers to account for varying sexual behaviour within the population they are studying. For STIs with high infection rates modellers may need to experiment with different population sizes.

微观模拟（microsimulation）正越来越多地被用于估算性传播感染（sexually transmitted infections, STIs）的流行率。此类模型以代表性活跃人群的智能体（agent）作为基本建模单元。将智能体匹配为性关系伴侣的过程计算量极大，同时也给建模者带来了诸多棘手的设计决策：包括如何选择终止智能体间伴侣关系的对象、哪些智能体应进入伴侣匹配市场，以及如何在该市场中完成智能体的配对。本研究旨在分析上述各类设计决策对性传播感染流行率的影响。我们对比了两种用于筛选进入每日伴侣匹配市场的智能体，以及终止智能体伴侣关系的策略：一种是将智能体视为同质个体的随机选择策略，另一种则基于德国大型纵向数据集构建的选择策略——该数据集充分考虑了性别、性取向与年龄的异质性特征。此外，我们将上述两种策略分别与近年提出的多款智能体配对算法相结合，并对比了各组合的运行速度与模拟结果。本研究还考量了其他多项设计选择，例如模型中智能体的总数量、提升智能体性行为异质性的设置，以及伴侣关系中临时性性接触的占比。考虑智能体异质性的建模方案，其估算出的流行率低于将智能体视为同质个体的简易建模方法。此外，在采用非随机智能体配对的模拟场景中，当感染风险升高时，模拟发病率随智能体数量的增加反而下降。我们所开发的配对算法可快速完成针对大规模智能体群体的数千次模拟实验。高效的配对匹配算法为微观模拟建模者提供了可行途径，使其能够在所研究的人群中纳入性行为的异质性特征。针对感染率较高的性传播感染，建模者可能需要针对不同的人群规模开展模拟实验。