Model structure
We built upon a stochastic agent-based model of the epidemic of COVID-19 in France that previously showed adequate calibration and validation.15 Briefly, this model includes (i) a realistic synthetic population generated with demographic characteristics, medical comorbidities and household structure representative of the French general population,17–20 (ii) a social contact network among the individuals, each with a geolocalized activity sequence over the day, taking into account co-location probability and duration, including contacts with family members, extended family members or friends (at home or at bars and restaurants), contacts at school or at work, and during public transport or grocery shopping or cultural activities, and (iii) a disease model, which translates the edge weights in the social contact network into infection probability of the edge over the day. The model parameters are summarized in eTable 1. We updated the contamination risk and proportion of undiagnosed cases of our initial model15,21 and included data on SARS-CoV-2 seroprevalence in May 2020 in France.22
Outcomes
Outcomes included cumulative incidence, mortality, and number of hospital admissions.
The probabilities of hospital admission or death were stratified by age and adjusted for comorbidities, including, obesity, diabetes, chronic cardiac diseases, and chronic respiratory diseases, based on hazard ratios calculated using data from Institut Pasteur23 and from the OpenSAFELY cohort study.16 To reflect improved care of patients with COVID-19, we reduced the risk of death by an average of 10% in the model, regardless of age, starting July 1, 2020 to fit observational data.24,25 Delays between infection, symptom onset, hospital admission, death or recovery were based on prior reports.23,26,27
Vaccine Efficacy
COVID-19 vaccine efficacy was assessed using published results for the BNT162b2 mRNA COVID-19 Vaccine from Pfizer/BioNTech.10 Based on these data, the efficacy of two doses of the BNT162b2 mRNA COVID-19 Vaccine is expected to be 95.6% (95% CI: 89.4%–98.6%) in individuals aged 16 to 55 years, 93.7% (95% CI: 80.6%–98.8%) in those aged 55 to 65 years, and 94.7% (95% CI: 66.7%–99.9%) in those aged 65 to 75 years. For individuals aged over 75 years, we assumed a similar efficacy as in those aged 65 to 75 years. The efficacy of two doses of COVID-19 Vaccine mRNA-1273 from Moderna9 is expected to be very similar, with an estimated rate of 94.5% (95% CI: 86.5%-97.8%) in individuals aged 18 years and over. Finally, because the efficacy reported for two doses of the ChAdOx1 nCoV-19 vaccine from AstraZeneca/Oxford8 was substantially lower, i.e., 70.4% (95% CI: 54.8%-80.6%), in individuals aged 18 years and over, we conducted sensitivity analyses using efficacy data of this vaccine.
Vaccine uptake
Estimated COVID-19 vaccine acceptance was based on a discrete choice experiment conducted in a large sample representative of the French population aged 18-64 years.11 This study showed that vaccine uptake would be expected to assume an inverted U-shape relationship with advancing age. We assumed in our model that individuals accepting the vaccine would be vaccinated. In the absence of specific data, we also assumed that vaccine acceptance in the population aged 65 years and over would be similar to that reported for individuals aged 55 to 64 years.
Statistical Analysis
The stochastic agent-based microsimulation model of the COVID-19 epidemic in France was run using C++ from March 1st, 2020, until August 1st, 2021, on 500,000 individuals with an average of 200 simulations. The results were extrapolated to the French population, which comprises about 67 million people. We provided uncertainty measures by using 100 bootstrap samples based on the random variation of all parameters simultaneously, excluding vaccination acceptance to facilitate interpretation, either within their 95% confidence interval for parameters estimated from the literature or within a +/- 20% interval if the parameter was assumed.15,21 All results are presented per 100,000 inhabitants to facilitate international comparisons.
We examined whether the model had adequate calibration, i.e., whether it was able to adequately reproduce retrospectively the course of the epidemic until December 20th, 2020, based on R² and Normalized Root Mean Squared Error (NRMSE) for weekly mortality and hospital admissions, and visual comparison between model-predicted and observed mortality and hospital admissions.
All scenarios included a full population lockdown between March 17th, 2020 and May 11th, 2020, followed by a progressive return to 75% of the pre-pandemic social contacts level until July 1st, 2020, except at schools, which remained closed during that period, and a 30% rate of workers using telework. Following prior epidemiological trends,28 we assumed that SARS-CoV-2 infection was associated with a 23% reduction in disease transmission due to warm weather between July 1st, 2020 and September 25th, 2020. This latter date was chosen because it marked a significant drop in temperature in France and was quickly followed by a significant increase in the number of cases. On September 1st, 2020, schools reopened for all students and telework use decreased to 16% based on Google Mobility Reports for France.24 Based on data from Santé Publique France,24 we considered that face mask use at work, in public transport, during grocery shopping and for cultural events increased from 15% to 70% between April 4th and September 1st, 2020, and remained at this level hereafter. We assumed a limited use of face mask (i.e., 30%) in households or with friends or extended family members during this period and hereafter. Curfew was instated on October 17th, 2020, which has led to cancelation of all cultural events and was assumed to reduce social contacts with friends and extended family members by 50%. We considered that this curfew would last until January 15th, 2021. A second less stringent lockdown was instated between October 30th and December 15th, 2020, with schools and workplaces remaining opened. We assumed that 50% of individuals worked remotely from home during this second lockdown period and that this rate will be of 30% during the curfew after the second lockdown lifting.
To examine whether any vaccination scenario could allow for a lifting of NPIs, we assumed that from January 15th, 2021, social behaviors would return to those observed before the COVID-19 epidemic, with full discontinuation of all NPIs, and examined associated cumulative incidence, mortality, and number of hospital admissions. In our model, we considered that a vaccination strategy would allow for the discontinuation of NPIs if it was associated with (i) a cumulative number of deaths lower than 17 per 100,000 and (ii) a cumulative number of hospital admissions below 240 per 100,000, between December 27th, 2020 and August 1st, 2021. The first threshold corresponds to the mean plus two standard deviations of the total number of deaths observed in France between January 15th, 2020 and August 1st for the years 2015 to 2019, representing the threshold above which an increase in death could be considered significant. Given that hospital-bed capacity is 600 per 100,000 inhabitants in France24 and that the mean duration of a hospitalization for COVID-19 is about 21 days,24 the second threshold corresponds to a maximum hospital-bed occupancy rate for COVID-19 of 5% between December 27th, 2020 and August 1st, 2021. We chose this threshold because current hospital-bed occupancy by patients with COVID-19 is currently estimated at 6.3% (25,000/400,000) of the total number of hospital beds in France.24
Given the limited production and distribution capabilities for COVID-19 vaccines, it is expected that vaccinating the full non-immunized French population aged 18 years or older would probably require several months, even with three vaccines. Because our main aim was to examine whether different vaccination strategies would allow for lifting NPIs, rather than make questionable assumptions on time for vaccinating the population, we considered that vaccination in each scenario would be achieved by January 15th, 2021 and calculated the number of vaccine doses needed (considering 2 doses per individual) in each scenario.
Next we examined the impact of different vaccination scenarios according to the choice of the non-immunized adult populations to prioritize for vaccination: (i) no vaccination, (ii) vaccination of the full population, (iii) vaccination of adults aged less than 65 years, (iv) vaccination of adults aged more than 45 years, (v) vaccination of adults aged less than 35 years or more than 65 years, (vi) vaccination of adults aged more than 65 years, (vii) vaccination of adults aged more than 55 years with mandatory vaccination of adults aged more than 65 years (assuming that it would lead to a 90% vaccination rate in this population), and (viii) vaccination of individuals at higher risk for severe SARS-CoV-2 infection (i.e. adults aged more than 65 years and those with medical conditions associated with increased risk of severe COVID-1916). Two factors drove the choice of these scenarios: (i) the substantial risk of severe disease in adults aged more than 65 years and in those with medical conditions, and (ii) the higher expected vaccination uptake in younger than in older adults (Table 1).
We performed several sensitivity analyses for the four following scenarios: (i) vaccination of the full adult non-immunized population, (ii) vaccination of adults aged more than 45 years, (iii) vaccination of adults aged more than 55 years with mandatory vaccination of adults aged more than 65 years, and (iv) vaccination of at-risk individuals. First, we considered a 10% lower rate of vaccine uptake than that expected.11–14 Second, we examined the impact on our results of a lower efficacy of the vaccine in preventing COVID-19 among individuals aged more than 75 years (i.e. 50% instead of 94.7%),29 since very few COVID-19 cases were reported in Polack et al. in this population.10 Third, we reproduced the analyses while considering efficacy data of the ChAdOx1 nCoV-19 vaccine instead of the BNT162b2 mRNA COVID-19 vaccine. Fourth, we examined the impact of delaying the vaccination on the course of the epidemic by considering that the target population of each scenario was vaccinated by April 14th instead of January 15th. In this scenario, we assumed that NPIs present on December 15th would be maintained until April 15th and then discontinued after that date. Fifth, given the uncertainty of the effect of the vaccine on virus transmission (e.g. data from the ChAdOx1 nCoV-19 vaccine8 suggest an efficacy on carriage that is 60% lower than the immune response, as previously seen in other vaccines,28,30–32 and data from BNT162b2 mRNA and mRNA-1273 did not include carriage as an endpoint9,10), we tested a scenario where the vaccine would only decrease by 50% the virus transmission of vaccinated individuals with immune response, instead of 100% of those with immune response as in the main analyses. Finally, we examined the robustness of our results by evaluating the impact on outcomes of varying simultaneously all individual parameter values by ±20% for the scenario ‘vaccination of the full adult non-immunized population’.