January 9, 2024

expert reaction to a study claiming that epidemic modelling needs to be overhauled to include social networks

A study published in the Journal of Physics: Complexity looks at knowledge of social networks in epidemic modelling. 

Dr Marc Baguelin, Senior Lecturer at Imperial College London, Associate Professor at London School of Hygiene and Tropical Medicine, SPI-M member and lead of the Imperial College UK real-time modelling team during the COVID-19 pandemic, said:

“This paper discusses well-established aspects of epidemic spread on theoretical networks. The findings are common knowledge within the epidemiological modelling community and are integral to introductory modelling courses. Moreover, the methods discussed in the paper were familiar to the modellers who advised the UK government, with some actively contributing to the development of the theory surrounding epidemics on networks over the last 20 years.

“While acknowledging the inherent trade-off in models between complexity and explanatory power, it is difficult to see what this particular study brings to the understanding of the spread of SARS-Cov-2, particularly the reasons behind the occurrence of multiple transmission waves. The author’s model uses scale-free networks, which may not align well with the social interactions relevant to respiratory pathogens. It also implies a complete rewiring of the network as the catalyst for a new wave – akin to a sudden overhaul of one’s social circles (family, school, friends and work colleagues). Finally, this approach disregards crucial mechanisms such as immunity dynamics, behavioural changes resulting from factors like lifting restrictions, or the emergence of new SARS-Cov-2 variants.”

Prof Graham Smith, Lead Scientist for Wildlife, Animal and Plant Health Agency (APHA), said:

“It has long been known that disease model heterogeneity is important to the overall outcome of an epidemic. My experience is primarily within wildlife and livestock disease modelling, where heterogeneity in space and contact structure have been included in many models. Thus, the press release “Models used by scientists to predict how epidemics will spread have a major flaw” starts with a large over-statement. Indeed, other authors have already described the effect of including various types of heterogeneity in COVID models – for example this one published in 2021 https://doi.org/10.1515/cmb-2020-0115.

“The difficulty with human disease models is not only that this heterogeneity is much more complex (think about meeting people at work and leisure, as well as school contacts for children) but that this original network would be adjusted through non-pharmaceutical interventions (e.g. lockdowns) and personal choice (avoiding or reducing contacts by choosing to not going to events such as parties even when permitted, or choosing to go to large rallies protesting about the various interventions). Such changes cannot be easily or accurately predicted, so in many cases would have to be accounted for by discussing the uncertainties and known inaccuracies in any model used for policy. How much this was done during the COVID assessments I do not know.”

Prof Mike Tildesley, Professor of Infectious Disease Modelling at the University of Warwick and the Director of the Mathematics for Real World Systems Centre for Doctoral Training, said:

This paper presents a theoretical analysis that concludes with a result that epidemiologists have known for decades – that variability between contacts has an impact upon how an epidemic will spread. This is an extremely well researched area and as such the results in this paper are not novel. However, what is more concerning is the suggestion in the press release that the models used during the pandemic “have a major flaw” owing to a lack of consideration of variability in contact structure. This shows a significant lack of understanding of the models that were used to guide policy. Many of these models, whilst built upon a compartmental framework, incorporated variability in risk in several ways such as age, geographical region and in some instances contact surveys were used to parameterise these models. The models were also typically fitted and refitted to data as the pandemic progressed, to ensure that they were taking into account the most up to date information regarding the spread of disease. There have also been several papers published both during and in the aftermath of the pandemic to demonstrate how well these models performed in terms of predicting future epidemic behaviour.  Whilst it is important to recognise that there is uncertainty in any model predictions (and that this is duly communicated to decision makers), those other studies would suggest that the statement in the press release that the COVID models “have a major flaw” is extremely misleading.”

Prof Matt Keeling, Director of Zeeman Institute: SBIDER (Systems Biology & Infectious Disease Epidemiology Research), University of Warwick, said:

“This is a reasonable piece of work, that highlights a well-known issue in epidemiology and infectious disease modelling: that the number of social contacts matter. As the author states in the opening section “It has been known for decades that network topology has important effects on spreading processes”. The paper considers the impact of variation in the number of contacts per person, but ignores other network factors such as the strength and duration of contacts and the clustering of contacts (such as the interconnected network within a household). The paper also ignores the observation that those individuals with lots of social connections can only spend a short time with each contact (on average) which greatly reduces the strength of the results described.

“While there is nothing wrong with the work, to suggest that this calls for a radical rethink of all epidemiological models (including those used for COVID-19) is stretching the conclusions way too far. All the groups that worked on COVID-19 modelling in the UK, have all previously studied such network models before and are well aware of the consequences and implications.  In summary is this an interesting theoretical exercise, but it adds little to the field, and certainly does not impact on applied public health modelling.”

Prof Adam Kucharski, Professor in Infectious Disease Epidemiology, London School of Hygiene & Tropical Medicine (LSHTM), said:

“Understanding the relationship between dynamic social contacts and disease transmission is an important question. But unfortunately, this paper seems to reiterate a common misunderstanding about COVID dynamics, which has been made several times by others during the pandemic already (including by some groups who claimed in summer 2020 that there wouldn’t be a second wave).

“Specifically, the model assumes that there are a very small proportion of highly connected individuals who are responsible for superspreading events, and once these individuals are infected (and hence immune) the epidemic ends with only a small proportion of the population infected. There are two key problems with this assumption:

“1. Human social contacts vary a lot from person to person, but crucially a person with a lot of contacts today won’t necessarily be a person with a lot of contacts tomorrow. COVID superspreading events often occurred at events like weddings and social gatherings (e.g. https://wwwnc.cdc.gov/eid/article/26/9/20-1469_article) – however, there aren’t a subset of individuals who attend all the weddings, while others attend none. Analysis of real-life networks has found that ‘highly connected’ individuals on a given day are generally only highly connected for short time periods (https://www.medrxiv.org/content/10.1101/2023.11.22.23298919v2), so we can’t model the accumulation of immunity against acute respiratory infections by assuming that the network is static for prolonged periods.

“Notable exceptions to this unpredictability in contacts is among certain age groups and occupations (e.g. healthcare workers), where a subset of the population routinely make more contacts than others, and households where per-contact transmission risk is much higher (https://www.nature.com/articles/s41586-023-06952-2). This is why COVID disease models generally include variation in age-specific contacts by default, and there were also separate modelling efforts focused on settings with high transmission potential (e.g. households, certain events or workplaces).

“Variation in the number of contacts people make is also important for estimating the likely effort required for individual-targeted measures like contact tracing, and numerous models therefore accounted for this variation and implications for control (e.g. https://www.thelancet.com/article/S2214-109X(20)30074-7/fulltext & https://www.thelancet.com/article/S1473-3099(20)30457-6/fulltext & https://www.nature.com/articles/s41591-020-1036-8 & https://www.nature.com/articles/s41467-022-29522-y). There is lots more to be done on understanding the relationship between contacts and disease transmission, but to be effective, such analysis should focus make best use of available data (much of which is open acce, so free to analyse).

“2. There was evidence in 2020 that a relatively high percentage of individuals got infected with COVID in hotspots, from Austria (https://www.medrxiv.org/content/10.1101/2020.08.20.20178533v1) to New York (https://www.nytimes.com/2020/08/19/nyregion/new-york-city-antibody-test.html) to Brazil (https://www.medrxiv.org/content/10.1101/2020.08.28.20180463v1) to Peru (https://www.researchgate.net/publication/343414173_Seroprevalence_of_anti-SARS-CoV-2_antibodies_in_the_city_of_Iquitos_Loreto_Peru). Since the emergence of Omicron and lifting of control measures, we have also seen the vast majority of the population get infected in the UK (https://epiforecasts.io/inc2prev/paper). Any modelling analysis of herd immunity thresholds (i.e. the point at which the reproduction number drops below 1 under normal behaviour patterns) should therefore focus on trying to explain both these infection results as well as observed contact patterns (e.g. https://www.medrxiv.org/content/10.1101/2023.10.05.23296586v2).”

‘Epidemic modelling requires knowledge of the social network’ by Samuel Johnson was published in Journal of Physics: Complexity at 9:00 UK time on 9^th January 2024.

Declared interests

Prof Graham Smith: No conflicts of interest.

Prof Kucharski: I was a member of SPI-M-O during 2020-22.

Prof Matt Keeling is a Professor at the University of Warwick. He has no financial interests to declare, but is a member of both SPI-M and NERVTAG. 

For all other experts, no reply to our request for DOIs was received.