Skip to main content

Our take —

This study, covering over 17 million adults in England, confirms previously observed associations with mortality: higher risks were seen with older age, male sex, chronic comorbidities, socioeconomic deprivation, and nonwhite ethnicity. Although causality cannot be inferred for any given risk factor owing to the study design, the size of the analysis provides a striking replication of results observed elsewhere in smaller studies.

Study design

Retrospective Cohort

Study population and setting

This study analyzed electronic medical records from 17,278,392 adults aged 18 years and over from the National Health Service in England, representing approximately 40% of the adult population in England, to estimate risk factors for COVID-19 mortality. Registered active adult patients with >1 year of prior follow-up and with recorded age, sex, and “deprivation” (a measure related to poverty, derived from area of residence) were eligible for inclusion; the study period was from February 1 to May 6, 2020. Risk factors for COVID-19 mortality were assessed with Cox proportional hazards models 1) adjusted for age and sex; and 2) adjusted simultaneously for all included risk factors.

Summary of Main Findings

There were 10,926 deaths (0.06%) attributed to COVID-19 in the study population. In the fully adjusted model, the risk of mortality increased with every decade of age: relative to the reference category aged 50-59 years, the hazard ratio (HR) for death was 20.61 (95% CI: 18.72-22.70) for those aged 80 years and older. Men had a higher hazard of mortality relative to women (HR: 1.59, 95% CI: 1.53-1.65). Mortality risks for Black (HR: 1.48, 1.30 to 1.69), South Asian (HR: 1.44, 1.32-1.58), and mixed ethnicities (HR: 1.43, 1.11-1.85) were higher relative to white ethnicity. The most deprived quintile (from a geographically-defined index of socioeconomic status) had 1.80 times (1.69-1.91) the mortality hazard of the least deprived quintile. Multiple comorbidities were associated with increased mortality risk, including obesity, diabetes, chronic heart disease, chronic liver disease, kidney dysfunction, recently diagnosed cancer, and severe asthma. Although smoking and hypertension were associated with higher mortality in the age- and sex-adjusted model, they were not in the fully adjusted model. Post-hoc analyses of smoking and hypertension indicated that controlling for chronic respiratory disease (for smoking) and for diabetes and obesity (for hypertension) explained much of the changes in hazard ratios.

Study Strengths

The primary strength of this study is its size; it was able to cover a significant proportion of the adult population of England.

Limitations

Estimated hazard ratios for covariates were derived from a model that adjusted simultaneously for all covariates; results for a given risk factor cannot be interpreted causally due to possible mediation, unmeasured confounding, and collider stratification bias. There were high proportions of missing data for ethnicity (26%), obesity (22%), and hypertension (10%). Some deaths may have been missed due to misclassification. Although the dataset was very large, it may not have been representative of the population of England (e.g., only 17% of health care practices in London were included). The study does not address risk factors for non-fatal but nonetheless severe manifestations of COVID-19.

Value added

This extremely large study confirms associations seen in other studies and provides further evidence for the role of structural inequalities in COVID-19 impacts.

Our take —

This study found that for surveillance purposes, analytical sensitivity of tests are far less important than testing frequency and result turnaround time. Tests with sensitivity significantly less than the typical qPCR should be considered when they may allow for more frequent testing (e.g., every 1-3 days) and rapid results delivery (e.g., less than one day) These results will help motivate regulatory authorities on the utility of lower sensitivity tests and will help those designing surveillance programs decide between different types of tests given budget, staffing and logistic constraints.

Study design

Modeling/Simulation

Study population and setting

This study combined a model of within-host viral growth and decay with transmission models to estimate the potential impact of analytical test sensitivity (limit of detection of 1,000 vs. 100,000 RNA copies), frequency (every 1, 3, 7 or 14 days) and turnaround time (0, 1 or 2 days) on the effectiveness of a population-wide surveillance program. Individuals were assumed to self-isolate either upon receiving a positive test or upon the start of symptoms, until they were no longer infectious. Transmission simulations were conducted using two different models, with different levels of complexity, including an agent-based model with age-specific and household patterns of potentially infectious contacts based on data from New York City; and a university-like setting, where individuals mix at random and infections are constantly introduced from an outside source.

Summary of Main Findings

Testing daily or every three days with effective isolation after positive test results both reduced the effective reproduction number below one regardless of the test sensitivity, and as long as results were returned in fewer than two days. For weekly testing, if results were returned on the same day, a higher sensitivity test led to less than half the cases than when a low sensitivity test was used; however, even slight delays in reporting test results reduced this margin between the tests significantly. Fortnightly testing was ineffective in all cases. Results were similar across the two different models similar to the city and university settings.

Study Strengths

The study simulated a range of viral load trajectories in infected individuals, so it could explicitly relate the limit of detection to whether a case would be detected. Results were replicated across two different models and were robust to a range of assumptions about the relationship between viral load and infectiousness, and rate of imported infections.

Limitations

The study assumed that the whole population took part in the surveillance program, there was 100% compliance in self-isolation, and that tests detected all cases where the viral load was above the limit of detection. Although the supplementary material provides methods to adjust for false negatives and test refusal (which is mathematically the same as only testing a proportion of the population), numerical results were not shown. The study also assumed that a self-isolating individual could not transmit at all; however, in some settings (e.g., within households), individuals may not be able to self-isolate perfectly. The study did not consider self-isolation of contacts of confirmed cases; including this would increase effectiveness of all the surveillance programs, but the relative increase in effectiveness for each program is unknown. The maximum delay considered, 3 days, is still shorter than that currently available in many places. Although the city setting considered different types of contacts, superspreading events due to an individual having a much-greater-than-average number of contacts was not considered.

Value added

Workplaces, schools and universities are planning for reopening, and tests with different performance characteristics, cost and turnaround time are becoming available. This study helps designers of surveillance programs to understand the tradeoff between test frequency, sensitivity and turnaround time in reducing the number of cases. The number of cases with and without surveillance for different model assumptions can be calculated using an interactive web-application (https://larremorelab.github.io/covid-calculator3).

Our take —

Herd immunity is the threshold of population immunity to a disease that must be achieved in order to bring an epidemic under control and limit onward transmission in the population. In this paper, researchers demonstrated the importance of real-world assumptions like age-dependent susceptibility to infection, age-dependent probability of transmission, and age-specific contact patterns on the calculation of herd immunity thresholds. After accounting for realistic variation in age and contact patterns, they conclude that the herd immunity threshold that must be achieved to prevent a second large wave of COVID-19 may be lower than projections from models that do not account for this age-specific variability in disease transmission. However, results are not meant to be interpreted as exact values, and should be interpreted qualitatively.

Study design

Modeling/Simulation

Study population and setting

Authors used a Susceptible-Exposed-Infectious-Recovered (SEIR) mathematical model in a generic population that included six age cohorts, three social activity levels (high, normal, low), and “preventive measures” to 1) demonstrate the effectiveness of the interventions; and 2) how herd immunity to SARS-CoV-2 is affected by different population structures. Authors varied the effectiveness of the preventive measures, but did not specify explicit interventions. Assuming an R0 (basic reproductive number) of 2.5 and introduction of the virus on February 15, 2020, authors modeled differences in effectiveness and disease-induced herd immunity assuming homogenous and heterogenous (varying population mixing by age only, activity levels only, or both) populations.

Summary of Main Findings

Models demonstrated that herd immunity levels were lower among a non-homogenous population (i.e., one of the mixed population structures described above) compared to a homogenous population; mixing both age and activity levels decreased the herd immunity level by 17 percentage points (60% vs 43%). Mixing only activity levels resulted in a greater reduction in disease-induced herd immunity levels compared to mixing age groups only (46.3% vs 55.8%). Assuming preventive measures are put in place one month after virus introduction (March 15, 2020) and lifted almost two months later (June 30, 2020), preventive measures reduced the size and delayed the timing of the peak. In the scenario with the most restrictive preventive measures, lifting these resulted in a clear second wave of infections.

Study Strengths

Authors explored different types of heterogeneous mixing, making the model adaptable to regions or locations with known age structures and activity levels.

Limitations

Although authors considered age structures and varying activity levels in their models, they did not consider more complex parameters, such as contact matrices and differences in daily contacts based on location (e.g., household vs work). Authors made several assumptions in the model that contribute to uncertainty. First, authors assumed that preventive measures would decrease all contact rates proportionally, however, most interventions target older adults and persons at highest risk. Furthermore, they assumed that these measures would be lifted simultaneously, when in reality most countries have been lifting restrictions in stages. Authors also assumed immunity to SARS-CoV-2 after infection for an “extended period of time”. However, the true duration of infection-acquired immunity is not currently known, and more work is needed in this area.

Value added

This study estimates the impact population heterogeneity can have on disease parameters, such as herd immunity, and illustrates the importance of incorporating real-world assumptions (i.e., that populations are not homogenous) when developing mathematical models.

Our take —

The study estimated the cumulative burden of hospitalization and ICU utilization for COVID-19 across 3142 US counties. It found the 315 counties estimated to be in the 90th percentile of COVID-19 hospitalizations had relatively larger populations over 60 years. 247 counties were in the 90th percentile for hospitalization-per-bed (indicating a high burden of healthcare system utilization), and 136 counties were in the 90th percentile for ICU-per-bed. These counties were scattered across the US, but primarily in the western US. Their model made a number of simplifying assumptions that limit the interpretability of their results, but the counties projected to have the greatest burden was robust to widely varying assumptions about transmission.

Study design

Ecological; Modeling/Simulation

Study population and setting

The study objective was to estimate the burden of hospitalization and ICU admissions for COVID-19 across the US in 3,142 counties as identified in the 2010 US census. The study used data from the 2018 American Community Survey to estimate the age-specific demographic distribution in each county, and the American Hospital Association 2018 annual survey to estimate the number of hospital beds and ICUs available in a given county. Counties without beds or ICUs were then distributed to the nearby counties with greater capacity. The study used a susceptible-exposed-infected-recovered (SEIR) transmission model, and assumed a constant population in the counties included. Authors compared outcomes from two models with varying assumptions about intensity of transmission: 1) a pessimistic transmission model which assumed the reproductive number was 5; 2) an optimistic model in which it was assumed to be 2.

Summary of Main Findings

The model estimated 315 counties at or above the 90th percentile of per-capita hospitalization in the optimistic scenario, and 308 were also at or above the 90th percentile in the pessimistic scenario. Counties with relatively larger populations over the age of 60 were most likely to experience more hospitalizations from COVID-19. The counties with the greatest estimated burden were, for the most part, far from urban centers. In assessing the cumulative burden of hospitalizations, 247 counties were at or above the 90th percentile in both transmission scenarios for cumulative hospitalization per bed and were scattered across the US, and were mostly from the west and the northern midwest. 136 counties were at or above the 90th percentile for ICU admission per bed to estimate ICU utilization in comparison to healthcare system capacity, though these counties were not statistically more likely to be rural.

Study Strengths

The study used two nationwide surveys in order to inform the model for each county, and was able to assume cases in areas with reduced hospital and ICU capacity sought care in nearby counties with capacity, as might happen in a real-world scenario. Many of the study’s results regarding the highest burden areas were robust to both optimistic and pessimistic transmission scenarios. They estimated cumulative hospitalization/ICU utilization in order to circumvent issues of temporality and inaccuracies in estimating a specific transmission peak.

Limitations

The study made numerous assumptions, which was necessary for their models but may reduce the applicability of the findings. For instance, their optimistic reproductive number of 2 does not take into account interventions such as shutdowns and stay-at-home orders, that may further reduce the reproductive number. They also assumed 20% of the population in each county eventually becomes infected, which may not be a reasonable assumption, given areas that are urban are likely to experience higher transmission than very rural areas with lower contact between individuals. Finally, the study used data on hospital and ICU bed capacity from 2018, which does not account for field hospitals or other stockpiling of ventilators and other equipment to increase hospital capacity which may occur in response to the pandemic.

Value added

This study estimated the cumulative hospitalization and ICU-related health system burden due to COVID-19 across the United States, and demonstrated that many rural counties will likely be unable to cope with the demand for hospitalizations without efforts to increase capacity, which has important policy implications for resource allocation and planning during the pandemic.

Our take —

This study based on seroprevalence data and modeling from Geneva Switzerland, available as a preprint and thus not yet peer-reviewed, confirms previous studies which have shown that the infection fatality ratio increases with age, and increases sharply after 65 years of age.

Study design

Modeling/Simulation

Study population and setting

The authors estimated the overall and age-specific infection fatality risk (IFR) for Geneva Switzerland using daily case and death reports between February 24 and June 2, 2020, combined with five weekly seroprevalence estimates starting April 6, 2020. The authors linked the sources of information to be able to infer IFR, and used Bayesian methods to account for the lag between infection and seroconversion, and between infection and death.

Summary of Main Findings

There were 286 total COVID-related deaths from February 2 to June 2, 2020 in Geneva. The IFR was low among those under 50 years of age at between 0.00032 and 0.0016% but this increased to 0.14% (95% Credible Interval (CrI) 0.096-0.19) among those 50-64 years, and to 5.6% (95% CrI 4.3-7.4) for those 65 years or older. The authors estimated the population-wide IFR to be 0.64% (95% CrI 0.38-0.98). When cases and deaths among care home residents were excluded from the analysis, the estimated IFR was 2.7% (95% CrI: 1.6 – 4.6) for those 65 years or older.

Study Strengths

The study was focused on a specific geographic area (Geneva) which allowed for more complete data of cases and deaths. This was matched with data from sequential seroprevalence surveys from a representative population.

Limitations

The focus on a distinct geographic location may limit the generalizability of the findings outside of Geneva. In the older population (65 years +), half of the deaths were among nursing home residents, but such residents are likely underrepresented in the sero-survey data used to create the estimates, which could bias the results, likely by overestimating the IFR in this group. Finally, if antibody responses differ by disease severity with milder cases having a weaker and shorter response, authors’ estimates of IFR may be biased upwards.

Value added

This study provides additional information on the modeled infection fatality risk (IFR) underpinned by seroprevalence data to add to the small but growing literature on IFR. A diversity of IFR estimates will be needed given that it is likely to differ from one population to another. It also highlights the disproportionate number of deaths in vulnerable settings such as care homes.

Our take —

This study provides credible estimates that by May 4, 2020, the 11 European countries in this study had achieved control of the COVID-19 epidemic by bringing the reproduction number below 1, after 3.2% to 4.0% of the population had become infected (far more than has been reported). The authors further estimate that country-wide non-pharmaceutical interventions (e.g., lockdowns and school closures) prevented over 3 million deaths. Results suggest the need for suppression measures to continue for a long time, since susceptible individuals are still estimated to represent a large fraction of the European population. However, this implication relies on model assumptions that preclude the ability of other factors, including individual behavior changes, to reduce transmission.

Preprint Review

This expert summary is for the peer-reviewed article linked above. We also summarized this paper before it underwent peer-review.  You can find the original review of the preprint by clicking here.

Study design

Ecological; Modeling/Simulation

Study population and setting

A semi-mechanistic Bayesian hierarchical model was fit to observed COVID-19 deaths in 11 European countries, and was used to estimate the number of cases and the reproduction number (Rt) of the infection. The model used partially pooled data across countries, along with the timing of country-specific non-pharmaceutical interventions, to estimate the impact of these interventions on Rt (with particular attention to whether Rt has been driven below 1) and on the number of infections and deaths in these countries up to May 4, 2020.

Summary of Main Findings

The reproduction number of the virus is estimated to have been reduced from 3.8 (95% credible interval: 2.4 to 5.6) to below 1 in all countries. Many times more people are estimated to have been infected by SARS-CoV-2 than have been confirmed: across all 11 countries, estimated cases totaled 12 to 15 million for an overall attack rate of 3.2% to 4.0%. Estimates of country-specific attack rates ranged from 0.46% (0.34% to 0.61%) in Norway to 8.0% (6.1% to 11.0%) in Belgium. 3.1 million deaths (2.8 million to 3.5 million) are estimated to have been averted by all non-pharmaceutical interventions by May 4. Of the categories of intervention, only lockdown had an identifiable impact.

Study Strengths

The model is fit to observed deaths, which are likely to be more reliable than case counts or hospitalizations. Estimates of case counts are supported by evidence from serologic surveys. The model reproduces observed data up to May 4th, 2020 very well. Uncertainty from various sources is appropriately handled by the model explicitly and by the discussion implicitly. Prior distributions and parameter values are chosen based on current, best available data.

Limitations

The model did not allow for other factors, such as changes in individual risk-avoiding behavior, to affect Rt. The infection fatality ratio, which is key to estimation of the number of infections, is treated as a fixed value by the model; there remains considerable uncertainty about the true value of this parameter. The timing of country-specific interventions makes it difficult to distinguish between the effects of specific interventions. Interventions are assumed to have the same impact across countries and time. Effect estimates in the model are heavily influenced by countries with a high number of deaths that implemented interventions earlier.

Value added

This study is the most thorough model-based estimate of the impact of European country-wide non-pharmaceutical interventions so far.

Our take —

This study modeled the impact of several non-pharmaceutical interventions on the number of new cases, hospitalizations, ICU bed needs, and deaths related to COVID-19 in the United Kingdom. Using a stochastic model stratified by five-year age groups, authors determined that although interventions would reduce the number of cases by an estimated 20-30% and delay the peak by 3-8 weeks, these reductions would not be enough to control the epidemic or prevent exceeding ICU bed capacity.

Study design

Modeling/Simulation

Study population and setting

Authors used a stochastic model (i.e., one that takes randomness into account) stratified into 5-year age groups to estimate the effectiveness of non-pharmaceutical interventions (NPIs) in preventing new cases, hospitalizations, ICU bed demand, and deaths from COVID-19 in Wales, Scotland, and Northern Ireland. Over 66 million people aggregated to county-level administrative units were included in the model and a baseline R0 (basic reproductive number) of 2.7 was used. Interventions evaluated through the model were school closures, physical distancing, shielding of older persons, self-isolation of symptomatic persons, a combination of these four, and more locally-focused interventions, including a reduction in leisure events (e.g., spectator sports). Finally, authors estimated the impact increased childcare by grandparents would have on the effectiveness of these interventions, namely school closures.

Summary of Main Findings

In the absence of NPIs, the case-fatality rate was estimated to be 1.5% (95% PI 1.3-1.7), and the infection-fatality ratio was estimated to be 0.63% (95% PI 0.45-0.79); this was estimated to result in a peak number of required ICU beds of 200,000 and 350,000 total deaths. When implemented in the middle of the peak of the unmitigated epidemic, on average, each intervention delayed the peak by 3-8 weeks and decreased the total number of cases by 20-30%. Model results suggest that when implemented locally, the effect of school closures, physical distancing, shielding of older persons, and self-isolation of symptomatic cases only modestly reduced the total number of cases and deaths compared to implementing them nationally. Suspending leisure activities such as spectator sports by 75% averted 1.9 million cases. Model results indicate that over a 3-month time period, the beneficial effects of school closures would be virtually eliminated in terms of the number of deaths and peak ICU bed needs due to the increased contact per weekday between children younger than 15 years and older persons (i.e., grandparents).

Study Strengths

The model estimated both case-fatality and infection-fatality ratios, which allowed it to account for asymptomatic and undiagnosed infections in addition to symptomatic cases. Authors also ran simulations using a distribution of R0 values, SARS-CoV-2 introduction dates, and timing of intervention implementation, which provided robust uncertainty bounds for the model projections. The modeling framework (i.e., location-based contact matrices) makes the scenarios well-oriented for policy making.

Limitations

Projections of the relative effect of the NPIs considered in this study are estimates only and derived from measurements taken in 2006. Per contact matrix (home, work, school, other) and intervention combination (e.g., home contacts assuming school closures), authors assumed decreases in contacts would be uniform between each five year age group pairing. However, the model did not structure individuals by household, which limits the ability to evaluate the impact of these interventions on household contacts. Challenges in comparing the results from this study to empirical UK data exist since implemented interventions differ from the scenarios evaluated here.

Value added

Going beyond the growing body of evidence estimating effectiveness of non-pharmaceutical interventions (NPIs) for individual countries, this study evaluated the impact of NPIs on different spatial scales, and compared their effectiveness when implemented at the local versus national level. This study also considered how the impact of certain NPIs could be negated by resulting changes to contact patterns, such as grandparents being responsible for childcare during school closures.

Our take —

Using daily laboratory-confirmed case counts, this study evaluated the association between control measures implemented in China and the spread of COVID-19 beyond Wuhan. Control measures such as the Wuhan travel ban and implementation of Level 1 Emergency response were associated with a delay in spread to other Chinese provinces and an overall decline in incident cases by mid-February 2020. Restrictions were especially effective when implemented prior to confirmation of COVID-19 cases. This study was unable to disaggregate the impact of travel restrictions from other containment measures. The combination of centrally coordinated measures may not be possible in other locations.

Study design

Modeling/Simulation

Study population and setting

This study used laboratory-confirmed case reports and human mobility data (mobile phone records) from 34 Chinese provinces to estimate effects of transmission control measures during the first 50 days of the COVID-19 epidemic (December 31, 2019 – February 19, 2020). Authors used a linear regression model to evaluate the association between the Wuhan travel ban and spread of COVID-19 to other Chinese provinces. They then used a Susceptible-Exposed-Infected-Recovered (SEIR) model fitted to the number of newly confirmed cases each day from each province to estimate the effect of implemented control measures on the trajectory of the epidemic outside of Wuhan.

Summary of Main Findings

Authors found that the total number of cases detected in each province was strongly associated with the number of travelers estimated from Wuhan. Banning travel in/out of Wuhan slowed COVID-19 transmission to other cities by an estimated 2.91 days (95% CI: 2.54-3.29). Cities that implemented control measures before having any confirmed cases had 33.3% fewer confirmed cases during the first week of their outbreak (13.0 cases, 95% CI 7.1-18.8) compared to cities that implemented controls after detecting cases (20.6 cases, 95% CI 14.5-26.8). Model results estimated that R (the effective reproductive number) decreased from 3.15 to 0.04 after the Level 1 response measures were 95% implemented. Authors estimated that, implemented alone, the Wuhan travel ban or Level 1 response would not have been enough to result in the observed decrease in incident cases by February 19, 2020. However, when combined, these control measures were estimated to have prevented 96% of COVID-19 cases that could have been expected to occur in the absence of all interventions.

Study Strengths

This study leverages mobile phone data for a detailed examination of how travel restrictions affected human mobility. Combining these data with case reports and the timing of control measures allowed multiple approaches to estimating the impact of control measures on transmission. The SEIR model was fitted to empirical case reports, meaning that the model was adapted to be most appropriate for the observed data.

Limitations

Results were generated from statistical and mathematical models, and strong statistical associations between implementation of control measures and observed delays and decreases in incident cases do not indicate a causal relationship. Multiple containment measures were implemented at once, and so this study is unable to disaggregate the impact of some containment measures (e.g., isolation of patients, closure of schools).

Value added

This study adds to the growing evidence regarding the effectiveness of the Wuhan travel ban and other high-level containment and suppression measures implemented in response to the COVID-19 pandemic. As with other studies, this modeling study suggests that these strict mobility interventions delayed the spread of COVID-19 within mainland China, and may have contributed to the observed reduction in cases by mid-February 2020.

Our take —

Results from this study, available as a preprint and thus not yet peer reviewed, suggest that even relatively small changes in the timing of stay-at-home measures at the beginning of the COVID-19 epidemic in the United States could have made a large difference in morbidity and mortality. However, the aggregated, non-specific treatment of non-pharmaceutical interventions in this study limits the utility of its conclusions. Applicability of results are also largely limited to densely populated, metropolitan areas of the country.

Study design

Modeling/Simulation

Study population and setting

Using US county-level reported cases and human mobility data, the authors developed a metapopulation model (which segments populations of interest into sub-populations, usually based on location) to estimate changes in COVID-19 transmission following stay-at-home measures introduced between March 15 and May 3, 2020. Authors used age-stratified infection fatality rates in the model and generated daily confirmed cases and deaths by county. There were 311 counties with over 400 cases by May 3, 2020; these were modeled separately from the remaining counties, which were further segmented into 16 smaller groups. The authors simulated case numbers and fatalities if stay-at-home measures had been implemented one to two weeks earlier.

Summary of Main Findings

If stay-at-home measures had been implemented one week earlier, the authors estimated that 56.5% (95% CI: 48.1% to 65.9%) of cases and 54% (95% CI: 43.6% to 63.8%) of deaths in the US would have been avoided. Under a scenario in which stay-at-home measures were implemented two weeks earlier, the estimate of cases averted rose to 84.0% (95% CI: 78.7% to 88.4%), and the estimate of deaths averted rose to 82.7% (95% CI: 76.1% to 87.6%).

Study Strengths

Authors calibrated the model against known county-level incidence and mortality data from the month prior to and including their timeframe of interest (February 21 to May 3, 2020). In general, metapopulation models allow for the inclusion of more realistic spatial transmission dynamics in epidemic models.

Limitations

The changes in population behavior modeled by the authors are treated as if they are a monolithic response to a broad aggregate of non-pharmaceutical interventions by federal, state, and local governments; they also assume that all behavior changes are a consequence of these NPIs rather than individual decision-making. Assuming that stay-at-home measures could have been implemented one to two weeks earlier with the same effect on infection dynamics is problematic. First, the public may not have complied as readily with stay-at-home measures earlier, as the perceived threat from COVID-19 may have been lower. Second, the timing of control measure implementation was influenced by other considerations, such as economic concerns. Furthermore, assumptions on inter-county movement were based on 2011-2015 data and may not reflect mobility between counties today. Due to the number of parameters that needed to be estimated as part of the model, the authors assumed a fixed value for the latency period, how long an individual was contagious, the infectiousness of unreported infections, and the proportion of movements that were not work-related. Estimated daily death rates incorporated variation in age structure by county but, given the potential impact of sex and race on case-fatality proportions, it is unclear why other demographic factors were not included. Finally, authors focused on metropolitan areas (e.g., New York, Chicago) and study results may not be applicable to less densely populated US counties.

Value added

As our knowledge about the incidence and transmission of COVID-19 evolves, this study contributes to the growing body of literature assessing intervention methods at finer spatial scales. Quantifying the impact of earlier implementation of stay-at-home measures provides a compelling case for early, aggressive intervention.

Our take —

This study, available as a preprint and thus not yet peer-reviewed, highlights the importance of real-world social contact data and networks for evaluating outbreak control measures. Tracing contacts of contacts can reduce outbreak size more than tracing only contacts, but requires high test capacity and results in a large proportion of the population in quarantine. The network model used may not scale to larger cities or for longer outbreaks, however is nevertheless valuable in exploring and optimizing potential control strategies.

Study design

Modeling/Simulation

Study population and setting

Authors used detailed contact data collected from 468 individuals during the 2017/2018 citizen-science BBC Pandemic Project in Haslemere, Surrey, UK. These high-resolution GPS data were collected to be specifically relevant for infectious disease modelling, and were used to construct a “social network” that represented real-world close contacts that could result in SARS-CoV-2 transmission such as “face-to-face” contacts. COVID-19 outbreaks were simulated using a mathematical model across this social network testing different network structures on how individuals within the network may be connected to each other. Authors allowed factors such as the proportion of asymptomatic individuals and their infectiousness to vary. Four control scenarios and their impact on the epidemic were assessed: 1) no control, where no individuals are isolated or quarantined; 2) case isolation, where individuals isolate when they start having symptoms; 3) primary contact tracing with quarantine, where individuals isolate once they have symptoms and traced contacts are quarantined upon their infector’s symptom onset (after a delay); and 4) secondary contact tracing which is the same as (3) but includes tracing of the contacts of contacts. The model assumed that isolated and quarantined individuals were all isolated for 14 days and explored how well contact tracing was conducted (% of contacts traced) impacted the effect of control. Finally authors explored a range of “test and release” strategies by varying the time from quarantine to being tested, and also looked at two social distancing scenarios whereby individuals either reduce or completely stop contacts with people they spend the least time with (e.g. outside the household) and instead spend this time with frequent contacts (e.g. within the household).

Summary of Main Findings

Compared to an estimated 12% of the Haslemere network being infected after 70 days in an uncontrolled epidemic scenario, case isolation of symptomatic individuals (control scenario 2); primary contact tracing (control scenario 3); and contact tracing contact of contacts (control scenario 4) resulted in a median 9.3%, 9%, and 7.3% of the population being infected respectively. Tracing contacts of contacts led to the greatest reduction in epidemic size, but also the highest proportion (29%) of individuals under quarantine at any one time. All control scenarios reduced the overall size and growth rate of the simulated outbreaks, with more efficient contact tracing (higher % of contacts traced) also leading to smaller outbreaks. Increasing test capacity, allowing more isolated cases and non-infectious contacts to be released from quarantine, led to substantial reductions in the number of quarantined cases but only a very small increase in outbreak size in both contact tracing scenarios. However this test and release strategy required high test capacity. Social distancing measures combined with contact tracing substantially reduced the number of tests required and the number of individuals in quarantine at any one time.

Study Strengths

This study uses very detailed app-based social contact data which were not limited by recall bias (often the case for other self-reported contact surveys), or restricted to a certain setting (such as schools or workplaces). By examining transmission dynamics in a social network that better captures real world interactions, authors were able to examine how different intervention methods may impact local epidemics. The study highlights how specific social structures should be considered in the design of control measures.

Limitations

The social contact data did not include children <13 years and were collected in a small geographic area over a short period of time. Therefore the social network explored here may not extrapolate to larger populations or over longer time periods. A short delay between isolation/quarantine and testing was assumed (48 hours) and so these testing strategies may only apply when there is sufficient testing capacity. Although social distancing was considered this was not in the context of specific strategies such as workplace closures. Therefore conclusions on what would be an effective social distancing intervention cannot be drawn.

Value added

This study highlights the value of real-world social contact data in epidemic modelling and the optimization of different control measures. It also shows the potential trade-offs between level of control (tracing contacts of contacts leading to smallest outbreaks), resources required (high testing capacity required), and the feasibility of intervention strategies (results in almost 30% of the population in quarantine).