Adherence and sustainability of interventions informing optimal control against the COVID-19 pandemic.

Background: After one year of stop-and-go COVID-19 mitigation, in the spring of 2021 European countries still experienced sustained viral circulation due to the Alpha variant. As the prospect of entering a new pandemic phase through vaccination was drawing closer, a key challenge remained on how to balance the efficacy of long-lasting interventions and their impact on the quality of life.
Methods: Focusing on the third wave in France during spring 2021, we simulate intervention scenarios of varying intensity and duration, with potential waning of adherence over time, based on past mobility data and modeling estimates. We identify optimal strategies by balancing efficacy of interventions with a data-driven "distress" index, integrating intensity and duration of social distancing.
Results: We show that moderate interventions would require a much longer time to achieve the same result as high intensity lockdowns, with the additional risk of deteriorating control as adherence wanes. Shorter strict lockdowns are largely more effective than longer moderate lockdowns, for similar intermediate distress and infringement on individual freedom. Conclusions: Our study shows that favoring milder interventions over more stringent short approaches on the basis of perceived acceptability could be detrimental in the long term, especially with waning adherence.

Introduction

The emergence of the SARS-CoV-2 Alpha variant in December 2020[1,2] disrupted the management of COVID-19 pandemic in Europe. The alert arrived as some governments were lifting interventions that had been applied to curb the second wave. Some countries, such as the UK and Ireland, were forced to rapidly implement strict lockdowns to control the explosion of cases due to the variant. Others maintained or strengthened their restrictions because of concerns over the new variant[3].

Few months after, with vaccination lagging behind (25% of the population of the European Union with a first dose on May 1, 2021 vs. 44% in the US, 51% in the UK, and 62% in Israel[4]) and a third wave due to the Alpha variant, continental Europe faced the challenge of relying once again on heavy restrictions to reduce sustained viral circulation and improve the epidemic situation approaching the summer. But what is the optimal strategy, given vaccination rollouts, the epidemic conditions, and the sustainability of long-lasting restrictive policies? On one side, limited available options remain beyond high intensity interventions, once milder layers of social distancing have been accumulated, strengthened, and extended over time (e.g., curfew, closure of restaurants and bars, closure of schools). On the other side, the efficacy and long-term sustainability of the adopted policies are potentially threatened by loss of adherence and policy-induced fatigue[5,6], affecting the quality of life of the population.

Building on observed adherence waning and introducing a data-driven measure capturing the limitations on individual freedom resulting from restrictions, we compared intervention scenarios of varying intensity and duration, and examined the role of adherence and sustainability on optimal epidemic control. The study is applied to the third wave in Île-de-France—the Paris region, the most populated of France and heavily hit by the pandemic—accounting for vaccination rollout plans, seasonality, and plans for the phasing out of restrictions.

We show that long-lasting interventions of moderate stringency achieve the same reductions in viral circulation and healthcare burden of shorter but higher stringency restriction, but at the expense of a higher distress in the population. This is exacerbated if adherence to policy wanes over time.

Methods

Data

Hospital surveillance data

We used regional daily hospital admission data, collected in the SIVIC database[7]. The database includes the number of admissions of COVID-19 confirmed patients to regular hospital or intensive care units. Hospital data are corrected for notification delays and do not suffer changes in detection or sampling, unlike the number of detected cases. As such, they provide a robust data source and have been used throughout 2020 in France for pandemic assessment and response[8-11].

Mobility data

Mobility reductions shown in Fig. 1 were extracted from two different data sources. Overall mobility was reconstructed from mobile phone data provided by Orange Business Service Flux Vision[12,13]. Data included origin-destination travel flows of mobile phone users among 1436 geographical areas in France. Each area corresponds to a group of municipalities, defined according to the 2018 EPCI level (Établissements Publics de Coopération Intercommunale[14]). Mobility reduction in a given week was computed as the relative variation of the number of trips with respect to the prepandemic baseline. Estimated presence at workplaces was obtained from Google Mobility Reports[15]. This dataset provides the relative change in the daily number of visitors to places of work compared to a prepandemic baseline, based on Google location-history data.

Fig. 1

COVID-19 pandemic waves in Île-de-France, with associated mobility reductions, social distancing, risk perception, and psychosocial burden.

a–c Weekly hospital admissions in Île-de-France during the first (a; weeks 10–20, March 2–May 17, 2020), second (b; weeks 41–52, October 5–December 27, 2020), and third (c; weeks 6–16, February 8–April 25, 2021) pandemic wave. Dots refer to data; filled dots correspond to the data used to fit the model, void dots correspond to data outside the inference window. Curves and shaded areas correspond to median fitted trajectories and 95% probability ranges, obtained from n = 250 independent stochastic runs. Horizontal dashed lines refer to the peak of the first and second wave in the region. d–f Mobility reduction in Île-de-France during the first (d), second (e), and third (f) pandemic wave. Yellow histograms represent the variation of mobility with respect to prepandemic levels, based on the number of trips extracted from mobile phone data[12]. Blue curves show the estimated change in presence at workplace locations over time with respect to prepandemic levels based on Google location-history data[15]. Shaded rectangles in the plots of the first two rows correspond to social distancing measures (strict lockdown in the first wave, moderate lockdown in the second wave, strengthened measures in the third wave). The second week of the second lockdown and the third week of the strengthened measures against the third wave have lower mobility and presence at workplaces due to bank holidays in the week. Vertical dotted gray lines correspond to school holiday periods. g–i Percentage of individuals avoiding crowded public places[16] (g), percentage of individuals scared to contract COVID-19[16] and prevalence of anxiety in the context of COVID-19 epidemic (h)[17] as functions of time; scattered plot between the percentage of individuals scared to contract COVID-19 and the percentage of individuals avoiding crowded places (i) in the time period October 2020–April 2021 (full time period shown in Supplementary Fig. S5), with the results of a Pearson correlation test (effect size 0.71, p-value < 10−3). Results for these indicators refer to the national scale. Shaded rectangles in panels g, h correspond to social distancing measures as in panels a–f.

Indicators of social distancing, risk perception, mental health

Several initiatives collect data over time through surveys to explore individual behaviours in response to COVID-19 pandemic. Here we use data from YouGov[16] and Santé publique France[17]. Surveys gather self-reported data, tracking compliance with preventive measures (e.g., avoiding social gatherings or contacts with other people, frequency of the use of masks), as well as risk perception and mental health indicators (e.g., fear to contract the virus, anxiety, depression). Indicators for specific social distancing behaviors (avoiding gatherings, use of masks) are used in addition to mobility data described above. YouGov surveys cover multiple countries and provide data at least every 2 weeks. Santé publique France polls collect data at the national level at least every month.

Ethics statement

Orange Business Service Flux Vision aggregated mobility travel flows were previously anonymised in compliance with strict privacy requirements, presented to and audited by the French data protection authority (CNIL, Commission Nationale de l’Informatique et des Libertés). They were accessed under license for this study. The study did not require an ethical approval as it involved review of publicly available documents, involved analyses that were based on previously published studies, involved aggregated and anonymous data, did not involve evaluation of experimental or patient data.

SARS-CoV-2 two-strain transmission model

We used a stochastic discrete age-stratified two-strain transmission model, integrating data on demography[18], age profile[18], social contacts[19], mobility[15], and adoption of preventive measures[17]. The model accounts for the co-circulation of two strains and vaccination. Four age classes are considered: [0-11], [11-19], [19-65], and 65+ years old (children, adolescents, adults and seniors respectively). Transmission dynamics follows a compartmental scheme specific for COVID-19 (Supplementary Fig. 1) where individuals are divided into susceptible, exposed, infectious, hospitalized, and recovered. The infectious phase is divided into two steps: a prodromic phase (Ip) and a phase where individuals may remain either asymptomatic (Ias, with probability pa = 40%[20]) or develop symptoms. We distinguished between different degrees of severity of symptoms, ranging from pauci-symptomatic (Ips), to individuals with mild symptoms (Ims), or severe symptoms (Iss) requiring hospitalization[11,21]. The duration of the infectious period was computed from the estimated mean generation time of 6.6 days[22] (Supplementary Fig. 2). Prodromic, asymptomatic and pauci-symptomatic individuals have a reduced transmissibility[23]. A reduced susceptibility is considered for children and adolescents, along with a reduced relative transmissibility for children, based on available evidence[24-27]. We assume that infectious individuals with severe symptoms reduce of 75% their number of contacts because of the illness they experience. Parameter values and corresponding sources are reported in the Supplementary Table 1. Sensitivity analysis on the probability of being asymptomatic, the susceptibility of younger age classes and transmissibility of children was performed in previous works[8,9,28].

Contact matrices are parametrized over time to account for behavioral response to social distancing interventions and adoption of preventive measures. Contacts at school, work and on transports are considered according to the French school calendar, school closures, and presence at workplaces estimated by Google. Physical contacts are reduced based on data from regular large-scale surveys conducted by Santé Publique France[8]. Contacts engaged by seniors are subject to an additional reduction of 30%, to account for evidence of a higher risk aversion behavior of the older age class compared to other age classes[8].

Alpha variant

Genomic and virological surveillance to identify specific mutations are in place in France since the start of 2021 to monitor variants over time. The first large-scale genome sequencing initiative (called Flash1 survey) was conducted on January 7–8 and analyzed all positive samples provided by participating laboratories[29]. The proportion of the Alpha variant in Île-de-France was estimated to be 6.9%, compared to the national estimate of 3.3%, making Île-de-France the region with the highest penetration registered in the country. Flash surveys are performed on average every two weeks on a sample of sequences. Starting week 6, 2021 a new protocol for virological surveillance was implemented to provide more timely estimates on the weekly frequency of detected viruses with specific mutations. It was based on second-line RT-PCR tests with specific primers that allow the detection of the main mutations that characterize the variants of concern. They must include at least the N501Y mutation and allow to distinguish the Alpha variant from the Beta or the Gamma variants. The frequency of the Alpha variant over time in Île-de-France is reported in the Supplementary Fig. 3.

We considered the co-circulation of the Alpha variant together with the historical strains, assuming complete cross-immunity. An increase in transmissibility of 59% (95% credible interval: 54–65%)[29] was considered for the Alpha variant compared to the historical strains. This early estimate was obtained from the Flash1 and Flash2 survey in France, and it is in line with other estimates[1,2]. To account for uncertainty in the transmission advantage and possible changes due to restrictions, we also show for sensitivity the results assuming 40% of increase in transmissibility, i.e., the lower estimate provided by ref. [2]. (Supplementary Fig. 11). We considered a 64% increase in hospitalization rates, following evidence of an increased risk of hospitalization after infection due to the Alpha variant compared with other lineages[30,31]. The frequency of the Alpha variant was initialized in the model on January 7, 2021 using the estimates of the first large-scale nationwide genomic surveillance survey (Flash1). The model was validated against virological and genomic surveillance data[10] on prevalence of Alpha variant over time. The Alpha variant was estimated to become dominant in the region by mid-February 2021[10] (Supplementary Fig. 3).

Vaccination rollout campaign

Administration of vaccines was included in the model according to the vaccination rhythm adopted in France starting January 2021. We considered the administration of 100,000 doses per day (including first and second doses) at the national level from the end of January (w04), accelerated to 200,000 first doses per day starting the beginning of March (w10), and 300,000 first doses per day starting April (w13). Rollout plans were expressed in terms of first administrations from March on to follow the objectives of authorities, delaying the administration of the second dose to reach a higher coverage in a smaller timeframe. Higher vaccination paces (400,000–800,000 doses/day) were also tested (Supplementary Fig. 9). Paces are defined at the national level, and the number of doses is proportionally distributed to the region according to the population eligible for the vaccine. Vaccination is prioritized to the older age class, assuming 80% coverage, and then shifted to adults considering 50% coverage, according to surveys on vaccine hesitancy[32]. Vaccination to healthcare personnel and patients in long-term care facilities, performed at the start of the vaccination program, could not be explicitly included.

We considered 75% vaccine efficacy against infection[33] and 65% vaccine efficacy against transmission[34], estimated after the first injection. We further considered 80% vaccine efficacy against symptoms given infection, computed from the estimated vaccine reduction of symptomatic disease[34,35] estimated at 95% after the second dose, and found to be similar after the first dose[36]. As the landscape for vaccine efficacy rapidly evolves, we also tested vaccine efficacy against transmission equal to 40%[37] (Supplementary Fig. 13). We assumed efficacy to start 3 weeks after the first injection, and tested a delay of 2 weeks for sensitivity (Supplementary Fig. 14).

Inference framework

The model is fitted to daily hospital admission data through a maximum likelihood procedure, by estimating the transmission rate in each pandemic phase. More precisely, prior to the first lockdown and in absence of intervention (period January–March 2020), we estimated {β, t} where β is the transmission rate per contact and t is the date of the start of the simulation. Then, in each phase we estimated α, i.e., the scaling factor of the transmission rate per contact specific to the pandemic phase under study (e.g., lockdown, exit from lockdown, summer, start of second wave, second lockdown, etc.). The transmission rate per contact in each phase is then defined as the transmission rate per contact in the pre-lockdown phase β multiplied by the scaling factor α. A pandemic phase is defined by the interventions implemented (e.g., lockdown, curfew, and other restrictions) and activity of the population (school holidays, summer holidays, etc.). The effective reproductive number is derived from the estimated transmission rate through the next generation matrix approach[38]. The likelihood function is of the formwhere Θ indicates the set of parameters to be estimated, H(t) is the observed number of hospital admissions on day t, H(t, Θ) is the number of hospital admissions predicted by the model on day t using parameter values Θ, Poiss(⋅∣H(t, Θ)) is the probability mass function of a Poisson distribution with mean H(t, Θ), and [t1, t] is the time window considered for the fit.

For Île-de-France, we seeded the model with 140 infected individuals to reduce the strong fluctuations associated with fitting the rapid increase and the high peak of hospitalizations observed in the first wave (the region was one of the areas mostly affected by the epidemic in early 2020). Simulations progress throughout 2020 to build immunity in the population. The model was validated against the estimates of three independent serological surveys conducted in France[8]. We used 250 stochastic simulations to compute median values and associated 95% probability range for all quantities of interest.

First lockdown, second lockdown, curfew

French authorities implemented two national lockdowns in 2020 to face the rapid surge of COVID-19 cases observed in the first and second wave. The first lockdown started on March 17, 2020 and lasted 8 weeks. It involved strict mobility restrictions outside home, together with closure of schools and non-essential activities. A less stringent lockdown was implemented for 6 weeks, starting on October 30, 2020. Schools remained open and a larger number of job sectors were allowed to operate. Measures were relaxed in the last two weeks of the lockdown, with the reopening of all retail for Christmas shopping. The second lockdown was lifted in mid-December with the application of a curfew starting at 8 pm, then anticipated in January 2021 to 6 pm to face increasing SARS-CoV-2 spread. Starting March 20, 2021, strengthened measures were additionally put in place in the region of Île-de-France to curb the third wave. These measures included mobility restrictions for trips exceeding 10 km, closure of business and of schools (1 week for primary schools, 2 weeks for middle and high schools in addition to 2-week school holidays in April). Values of the stringency index according to the timeline of interventions applied in France can be found in the Supplementary Fig. 4.

Loss of adherence

We used mobility data during the second lockdown and estimates of mobility reductions over time to assess if adherence to adopted policy waned over time, given unchanged restrictions. Focusing on the second lockdown, we compared the mobility reduction and reproductive number estimated in the first 3 weeks of lockdown implementation (w45–47, November 2–November 22, 2020) with respect to the following week. We considered the average over the first-3-week period to smooth out the effect of the national holiday on November 11, altering mobility and presence at work with respect to a regular week.

Lockdown scenarios

Starting from week 12, 2021 (March 22, 2021), we compared a scenario assuming unchanged curfew conditions—as estimated in week 11 (curfew scenario)—with the trajectories resulting from the application of a lockdown for a duration of 2–8 weeks. We modeled the effect of a strict lockdown and a moderate lockdown based on measured mobility reductions and estimated transmissibility conditions during the first and second lockdowns, respectively, before relaxation emerged. The delay from the date of implementation of lockdown and the peak of hospitalizations was estimated to be 9 days during the first lockdown in the region, and varied between 7 and 12 days across regions[8]. In our scenarios we assumed a 7-day delay, and tested 10 days for sensitivity (Supplementary Fig. 12). We also tested lockdown scenarios with different starting dates, ranging from w11 to w15, 2020 (Supplementary Figs. 6-S7).

For lockdowns longer than 2 weeks, we compared scenarios assuming full adherence with situations characterized by a loss of adherence over time. We modeled the loss of adherence throughout interventions by a relative increase in the reproductive number, according to estimates from the second lockdown. We applied it after 2 weeks from implementation of interventions (to model a faster dynamics of adherence waning compared to the one observed in the second lockdown), and considered it limited in time (one drop) or continuous (repeated drops every two weeks).

Distress index

In order to quantify the infringement on individual freedom associated with lockdowns and provide a measure of the policy impact on the quality of life, we introduced a quantity called distress index. This measure takes into account both the duration and the intensity of restrictions. It is defined as the sum of the absolute values of weekly mobility reductions, over the number of weeks in which each restriction is maintained, and normalized to a scale from 0 to 10 (10 representing a strict 8-weeks lockdown and 0 the absence of restrictions). In case of a strict or moderate lockdown without loss of adherence, we considered the mobility reductions recorded during the two interventions in 2020, respectively, and varied durations from 2 to 8 weeks. Loss of adherence is computed with a variation of the mobility reduction after 2 weeks (limited loss) and repeated every 2 weeks (continuous loss), according to estimates from the second lockdown. We took the end of January 2021 (w04) as reference for the mobility reduction associated with curfew.

Seasonality

Multiple studies have investigated the relationship between SARS-CoV-2 transmission and weather factors, including temperature, humidity, ultraviolet radiation[39], suggesting that summer conditions may help in reducing transmission of the virus. Seasonal factors and simultaneous social distancing interventions are difficult to disentangle; however, containment measures are estimated to have a larger impact on the epidemic compared to seasonal effects only. Considering the estimated dependence of the reproductive number on UV radiation[40] and temperature[41], we extracted data on downward UV radiation at the surface and daily temperature recorded in Paris, in Île-de-France, in the last three years (2018–2020)[42] to derive an approximate estimate of the reduction in the transmission rate induced by climate factors for the region under study.