Key Questions for Modelling COVID-19 Exit Strategies

• The listed keywords are COVID-19, SARS-CoV-2, exit strategy, mathematical modelling, epidemic control and uncertainty.Source: KEYWORDS · p. 3

INTRODUCTION

• Lockdowns reduced transmission, but their full socio-economic costs—including job losses and potential long-term mental health effects—remain unassessed.Source: INTRODUCTION · p. 4
• Exit strategies aim to lift restrictions and alleviate economic, social and other health costs, while active transmission or reintroductions can lead to renewed disease incidence.Countries with successfully suppressed community transmission still face the risk of transmission being reestablished.Source: INTRODUCTION · p. 4
• Estimating the effects of removing individual measures is challenging because intervention packages were introduced in quick succession, making modelling important for planning exits.Source: INTRODUCTION · pp. 4–5

• Workshop participants actively modelling COVID-19 discussed challenges across countries and disciplines, including the interplay between epidemiological models and other fields.The workshop followed four months of model development and data analysis.Source: §27 · p. 5
• The article aims to highlight knowledge gaps hindering predictions and provide a roadmap for scientists contributing to solutions.Source: §27 · p. 5
• Because the evidence base was changing rapidly and systematic review was challenging, researchers gathered expert views and relevant literature from varied countries and disciplines to reduce bias.The evidence is presented in a policy-neutral manner.Source: §27 · pp. 5–6
• The roadmap calls for improving epidemiological parameter estimates, understanding population heterogeneity, and addressing intervention data needs, particularly in LMICs.These areas are interdependent and should be pursued concurrently.Source: §27 · p. 6, Figure 1 · p. 7
• Improved understanding of model resolution and data needs can inform each other; following the roadmap could sharpen predictions while limiting the risk of substantial further transmission.The proposed precision could help policy-makers choose exit strategies that relax interventions while controlling transmission risk.Source: §27 · pp. 6–7

1.1 HOW CAN VIRAL TRANSMISSIBILITY BE ASSESSED MORE ACCURATELY?

• Real-time R(t) estimation is difficult because incomplete data cause reporting delays, changing testing or capacity alters case counts indistinguishably from transmission, estimates are inaccurate at low counts, and serial intervals change over time.COVID-19 evidence suggests significant individual-level variation in SARS-CoV-2 transmission, with some transmission events leading to large numbers of new infections.Source: §1.1 · pp. 9–11

1.2 WHAT IS THE HERD IMMUNITY THRESHOLD AND WHEN MIGHT WE REACH IT?

• The COVID-19 herd-immunity threshold remains uncertain because immunity, basic reproduction numbers, and population heterogeneity are not well established.The threshold is where R(t) falls below one in an otherwise uncontrolled epidemic, but reaching it does not end the epidemic or eliminate further infection risk. Better estimates of immunity and transmission differences within and between populations are needed to assess how disease-acquired immunity affects the threshold.Source: §1.2 · pp. 11, 13

1.3 CAN SEROPREVALENCE SURVEYS PROVIDE INSIGHT INTO HERD IMMUNITY AND TRANSMISSION DYNAMICS?

• Seroprevalence surveys can reveal infections missed by routine surveillance, but their use to estimate population protection and inform exit strategies remains uncertain.Single surveys do not establish when infections occurred, and test sensitivity, specificity, immune-response interpretation, and people with few or no detectable antibodies complicate inference; 10–20% of symptomatically infected people were reported to have few or no detectable antibodies.Source: §1.3 · pp. 14–15

1.4 IS GLOBAL ERADICATION OF SARS-COV-2 A REALISTIC POSSIBILITY?

• Epidemic extinction can be slow, with a long, fluctuating tail of cases driven in part by spatial differences across locations.The passage identifies differences in weather and access to treatment as potential drivers of long epidemic tails (Figure S1).Source: §1.4 · p. 16
• Past experience suggests SARS-CoV-2 may not be driven to complete extinction in the foreseeable future, even if a vaccine and vaccination campaigns are available.The virus could persist in populations with limited healthcare access or circulate in seasonal outbreaks as it evolves.Source: §1.4 · p. 17
• SARS-CoV-2 could become endemic by persisting in populations with limited healthcare access or circulating in seasonal outbreaks as it evolves.The paper stresses that this possibility, including that the virus may never be eradicated, should be communicated to the public and policy-makers.Source: §1.4 · p. 17
• Assessing whether SARS-CoV-2 is extinct is difficult because it may persist in individuals who do not report disease.Source: §1.4 · p. 17

2.1 HOW MUCH RESOLUTION IS NEEDED WHEN MODELLING HUMAN HETEROGENEITIES?

• Models must balance complexity with parsimony and speed, while accounting for spatial contact structure and health vulnerabilities whose effects on exit strategies remain uncertain.Reported risk factors may be proxies for underlying causes, and modelling results communicated to the public could produce behavioural feedback that helps or hinders exit strategies.Source: §2.1 · pp. 18–20
• Spatially detailed models need appropriate calibration across data collected at different scales, and modelling can assess whether targeted interventions justify their operational costs and implementation challenges.Policies often apply to whole administrative populations, although targeting specific groups may be more effective.Source: §2.1 · pp. 18–19

2.2 WHAT ARE THE ROLES OF NETWORKS AND HOUSEHOLDS IN SARS-COV-2 TRANSMISSION?

• Contact-network measures may miss transmission routes or include low-risk contacts, and the relationship between observed social-network proxies and transmission—and its variation across regions—remains uncertain.Contact tracing may work better in clustered than equivalent non-clustered networks, while whether particular people are super-spreaders or any infected person can generate such events is debated.Source: §2.2 · pp. 20–22
• Household and multilayer network models can explore how networks merge as restrictions ease, but whether sustained household infection chains amplify infections between households remains an open question.Comparisons across countries and First Few X studies could examine household structures, policy costs, and epidemiological characteristics including asymptomatic and presymptomatic transmission.Source: §2.2 · pp. 21–23

2.3 WHAT IS THE ROLE OF CHILDREN IN SARS-COV-2 TRANSMISSION?

• Evidence on age-specific susceptibility and infectiousness is mixed, although emerging evidence suggests children may be less susceptible than adults and infected children often have milder disease.The mechanism behind lower susceptibility is unknown and may not generalise across settings; children may also be more likely to have subclinical infections.Source: §2.3 · pp. 23–24
• The effect of school reopening on population transmission is unclear, and models must account for changes in contacts across schools, households, and the wider community.Closure effects were difficult to isolate because they coincided with other interventions; decisions about which students return and how they are grouped must weigh transmission among children, staff, and households with differing risks of severe outcomes.Source: §2.3 · pp. 24–25

2.4 THE PANDEMIC IS SOCIAL: HOW CAN WE MODEL THAT?

• Agent-based models can represent individual traits, social interactions, and societal influences, but national-scale models are unlikely to be built in time for the current pandemic and require careful parameterisation and uncertainty analysis.They can incorporate disease-state transitions from compartmental models and need not be all-encompassing to serve as policy “thought experiments.”Source: §2.4 · p. 26
• For future pandemics, the authors call for more data on people’s ordinary behaviour, including what they do each day and whom they meet.Time-use diaries and, with consent, mobile phone data are proposed as sources.Source: §2.4 · p. 27
• Modeling should account for how people understand and act on government regulation and how social media and broadcast information influence them.Source: §2.4 · p. 27
• A large, modular agent-based model calibrated as a social “digital twin” could represent population heterogeneity and enable virtual policy experiments.Source: §2.4 · p. 27
• Developing these models now could help plan exit strategies in future situations.Source: §2.4 · p. 27

3.1 WHAT ARE THE ADDITIONAL CHALLENGES OF DATA LIMITED SETTINGS?

• Confirmed case numbers may not represent epidemic trajectories in underserved populations, where limited diagnosis and absent historical baselines constrain surveillance.The passages identify clinically diagnosed cases as a possible alternative measure.Source: §3.1 · p. 28
• In LMICs, lockdown decisions must weigh COVID-19 control against hunger and disruption to routine health care, while relevant health-system and socioeconomic data are often unavailable.The passage notes that these lockdown-related harms can cost lives in the short and long term.Source: §3.1 · p. 29
• Models for data-limited settings should prioritize policies robust to missing information, value-of-information analyses, and methods that use metadata to interpret epidemiological patterns.When additional collection is unaffordable, available metadata can include testing and reporting rates, sample backlogs, and syndromic surveillance.Source: §3.1 · p. 29

3.2 WHICH DATA SHOULD BE COLLECTED AS COUNTRIES EMERGE FROM LOCKDOWN, AND WHY?

• Estimating the effects of lockdown components is needed to decide how and in what order to release interventions, but their impacts on mixing remain uncertain.Models for policy often make broad assumptions about effects of social-distancing measures.Source: §3.2 · pp. 30–31
• Cross-country comparisons may help estimate component effects, but importations and differences in mixing, behavior, households, occupations, age, and comorbidities complicate attribution.The passages also identify weather and climate effects as an open question whose transmission impact has not been established conclusively.Source: §3.2 · pp. 31–32
• Coordinated longitudinal and case-control studies, standardized protocols, and shared field models can support locally parameterized models and early detection of resurgence.The authors connect these efforts with identifying when interventions may need to be reintroduced.Source: §3.2 · p. 33

3.3 HOW SHOULD MODEL AND PARAMETER UNCERTAINTY BE COMMUNICATED?

• Uncertainty includes model assumptions, parameter values, data, and initial conditions; sensitivity analysis and decision analysis can help assess its implications for policy choices.Sensitivity analyses compare predictions across repeatedly sampled plausible parameter values.Source: §3.3 · pp. 34–35
• Model comparison is difficult when models address different questions, while past predictive performance can inform real-time comparisons and ensemble weighting.The passages state that weighting models by past performance typically improves point and variance estimates.Source: §3.3 · pp. 34–35

SUMMARY AND DISCUSSION

• Models should be revised as evidence grows, critically assess data, and bring together evidence from multiple fields because fully validated models cannot wait for complete data.The workshop roadmap was intended to be feasible within the pandemic.Source: SUMMARY AND DISCUSSION · pp. 36–37
• The roadmap calls for improved epidemiological parameter estimates, better understanding of population heterogeneity, and focused data collection, particularly in LMICs.Its steps are interdependent rather than linear.Source: SUMMARY AND DISCUSSION · pp. 37–38
• Improving parameter estimates requires close collaboration between modellers and data collectors, with each side's caveats and assumptions clearly understood.The authors argue that acknowledging simplifying assumptions allows their impacts to be assessed and models to be validated or replaced as evidence emerges.Source: SUMMARY AND DISCUSSION · p. 37
• Models can represent heterogeneity through agent-based approaches, age or other population strata, spatial structure, and networks that capture epidemiological and social interactions.The authors also identify opportunities to collaborate with fields such as the social sciences.Source: SUMMARY AND DISCUSSION · pp. 37–38
• In LMICs, models validated with global data can provide cost-effective guidance, while modellers should share expertise with those who need it most.The authors call for attention to country-specific data requirements and communication with governments, public health authorities, and the public.Source: SUMMARY AND DISCUSSION · p. 38
• Models can support reliable inferences about policy impacts across countries, where death tolls have sometimes been used to attribute success or failure without convincing evidence.The authors call for clearer communication of rapidly changing knowledge and its multiple sources of uncertainty.Source: SUMMARY AND DISCUSSION · p. 38
• International collaboration is essential for data collection and novel research during the pandemic and may also improve responses to future emerging infectious disease outbreaks.The modelling effort drew on years of outbreak experience and engagement with public health agencies and policy-makers.Source: SUMMARY AND DISCUSSION · pp. 36, 38, 39

Funding

• Authors acknowledged support from the Isaac Newton Institute, Christ Church, the Wellcome Trust, BBSRC, NERC, and UK MRC and DFID centre funding.The acknowledgments also name support from the UK NIHR, HDR UK, and Netherlands funding organisations.Source: Funding · pp. 53–54
• Additional acknowledgments name support from the Bill and Melinda Gates Foundation, Wellcome Trust, Royal Society, SACEMA, CNPq, FAPERJ, and Vetenskapsrådet.The funding section also lists La Trobe University and the Leverhulme Trust.Source: Funding · p. 54

Ethics statement

• The authors declared that no ethical concerns exist.Source: Ethics statement · p. 54

Data accessibility statement

• No new data were created or analysed, so data sharing was not applicable.Source: Data accessibility statement · p. 54

Competing interests statement

• The authors declared no competing interests.Source: Competing interests statement · p. 54

Authors’ contributions

• All authors attended the workshop, contributed to discussions and manuscript sections, and edited and approved the final version.RNT compiled the manuscript; RNT, TDH, VI, HH, DM, and OR organised the workshop and designed the study.Source: Authors’ contributions · p. 55