Prestige Em, Coletti Pietro, Backer Jantien, Davies Nicholas G, Edmunds W John, Jarvis Christopher I
Centre for Mathematical Modelling of Infectious Diseases, London School of Hygiene & Tropical Medicine, United Kingdom.
Data Science Institute, I-BioStat, Hasselt University, Hasselt, Belgium; Université catholique de Louvain, Institute of Health and Society (IRSS), Brussels, Belgium.
Epidemics. 2025 Jun;51:100830. doi: 10.1016/j.epidem.2025.100830. Epub 2025 Apr 23.
During the COVID-19 pandemic, aggregated mobility data was frequently used to estimate changing social contact rates. By taking pre-pandemic contact matrices, and transforming these using pandemic-era mobility data, infectious disease modellers attempted to predict the effect of large-scale behavioural changes on contact rates. This study explores the most accurate method for this transformation, using pandemic-era contact surveys as ground truth. We compared four methods for scaling synthetic contact matrices: two using fitted regression models and two using "naïve" mobility or mobility squared models. The regression models were fitted using the CoMix contact survey and Google mobility data from the UK over March 2020 - March 2021. The four models were then used to scale synthetic contact matrices-a representation of pre-pandemic behaviour-using mobility data from the UK, Belgium and the Netherlands to predict the number of contacts expected in "work" and "other" settings for a given mobility level. We then compared partial reproduction numbers estimated from the four models with those calculated directly from CoMix contact matrices across the three countries. The accuracy of each model was assessed using root mean squared error. The fitted regression models had substantially more accurate predictions than the naïve models, even when models were applied to out-of-sample data from the UK, Belgium and the Netherlands. Across all countries investigated, the linear fitted regression model was the most accurate and the naïve model using mobility alone was the least accurate. When attempting to estimate social contact rates during a pandemic without the resources available to conduct contact surveys, using a model fitted to data from another pandemic context is likely to be an improvement over using a "naïve" model based on mobility data alone. If a naïve model is to be used, mobility squared may be a better predictor of contact rates than mobility per se.
在新冠疫情期间，汇总的出行数据经常被用于估计不断变化的社会接触率。通过获取疫情前的接触矩阵，并利用疫情时期的出行数据对其进行转换，传染病建模者试图预测大规模行为变化对接触率的影响。本研究以疫情时期的接触调查作为基本事实，探索这种转换的最准确方法。我们比较了四种缩放合成接触矩阵的方法：两种使用拟合回归模型，两种使用“简单”出行或出行平方模型。回归模型是使用2020年3月至2021年3月英国的CoMix接触调查和谷歌出行数据进行拟合的。然后，使用这四种模型，利用英国、比利时和荷兰的出行数据对合成接触矩阵（一种疫情前行为的表示）进行缩放，以预测在给定出行水平下“工作”和“其他”场景中预期的接触次数。然后，我们将从这四种模型估计的部分再生数与直接从三个国家的CoMix接触矩阵计算出的再生数进行比较。使用均方根误差评估每个模型的准确性。即使将模型应用于来自英国、比利时和荷兰的样本外数据，拟合回归模型的预测也比简单模型准确得多。在所有调查的国家中，线性拟合回归模型最准确，仅使用出行的简单模型最不准确。在没有资源进行接触调查的情况下试图估计疫情期间的社会接触率时，使用基于另一个疫情背景数据拟合的模型可能比仅基于出行数据的“简单”模型有所改进。如果要使用简单模型，出行平方可能比出行本身更能预测接触率。
