We conducted a phone survey during March 31–April 7, 2020, to estimate the number of social contacts and age mixing of the population on a weekday during the lockdown and on the same day of the week before the pandemic, during mid-January 2020, by using contact diaries (Appendix Figure 1). Participants provided oral informed consent. We defined contact as either skin-to-skin contact or a 2-way conversation with >3 words spoken in the physical presence of another person (6). For each contact, we recorded information on the contact person’s age and location of the contact, such as home, school, workplace, transportation, leisure, or other. We planned to recruit 600 participants of all ages residing in Athens by using proportional quota sampling and oversampling among persons 0–17 years of age.

We estimated the average number of contacts for the prepandemic and lockdown periods. We defined 6 age groups to build age-specific contact matrices, adjusting for the age distribution of the population of Greece, by using socialmixr in R software (R Foundation for Statistical Computing, https://www.r-project.org).

To estimate the course of the epidemic, we first estimated the basic reproduction number (R₀), the average number of secondary cases 1 case would produce in a completely susceptible population in the absence of control measures. Then, we used social contacts matrices to assess the effects of physical distancing measures on R₀. Finally, we simulated the course of the epidemic using a susceptible-exposed-infectious-recovered (SEIR) model.

We used a SEIR model to simulate the outbreak from the beginning of local transmission until April 26, 2020, the day before the originally planned date to ease lockdown measures. Susceptible persons (S) become infected at a rate β and move to the exposed state (E) as infected but not infectious. Exposed persons become infectious at a rate σ, and a proportion p will eventually develop symptoms (p = 80%) (21). To account for asymptomatic transmission during the incubation period, we introduce a compartment for infectious presymptomatic persons (I_pre). I_pre cases become symptomatic infectious (I_symp) cases at a rate of σ_s. We assumed that infectiousness can occur 1.5 days before the onset of symptoms (22–24). The remainder (1 – p) will be true asymptomatic or subclinical cases (I_asymp)_. We assumed that the infectiousness of subclinical cases relative to symptomatic cases was q = 50% (24). Symptomatic cases recover (R) at a rate of γ_s, and asymptomatic cases recover (R) at a rate of γ_asymp (Table 1; Figure 2; Appendix).

We derived the transmission rate β from R₀ and parameters related to the duration of infectiousness (Appendix). We incorporated uncertainty in R₀ by drawing values uniformly from the estimated 95% CI (2.01–2.80). We modeled the effect of measures by multiplying β by the parameters δ₁ and δ₂; in which δ₁ corresponds to the reduction of R₀ in the period of initial social distancing measures, where δ₁ was drawn from a normal distribution with a mean of 42.7% (SD 1.7%); and δ₂ corresponds to the reduction of R₀ during lockdown, for which δ₂ was drawn from a normal distribution of 81.0% (SD 1.6%) estimated from the bootstrap on the contact data. To account for the uncertainty in R₀, δ₁, and δ₂, we performed 1,000 simulations of the model and obtained median estimates and 95% CrIs.

We obtained the infection fatality ratio (IFR) and the cumulative proportion of critically ill patients by dividing the reported number of deaths and of critically ill patients (25) by the total number of cases predicted by the model. We used a lag of 18 days for deaths and 14 days for critically ill patients based on unpublished data on hospitalized patients from the National Public Health Organization in Greece. To validate our findings, we used a reverse approach; we applied a published estimate of the IFR (26) to the number of infections predicted by the model and compared the resulting cumulative and daily number of deaths to the observed deaths (Appendix Table 3).

Because multiple social distancing measures were implemented simultaneously, to delineate the effects of each measure on R₀, we used information from the contacts reported on a regular weekday in January 2020 and mimicked the impact of each intervention by excluding or reducing subsets of corresponding social contacts (16,17,19,20) (Appendix). We also assessed scenarios with less disruptive social distancing measures (Appendix). In addition, we evaluated the increase in effective reproduction number (R_t) for varying levels of infection control measures (hand hygiene, use of facemasks, and maintaining distance >1.5 m) when social distancing measures are partially lifted after lockdown (Appendix).

In total, 602 persons provided contact diaries and reported 12,463 contacts before the pandemic and 1,743 during lockdown (Table 2). The mean daily number of contacts declined from 20.7 before to 2.9 during lockdown; when adjusted for the age distribution of the population, the reduction was 19.9 before and 2.6 during lockdown (86.9%).

We noted a change in age-mixing patterns in the contact matrices (Figure 3, panel A). In the prepandemic period, the diagonal of the contact matrix depicts the assortativity by age; participants tended to associate more with people of similar age (Figure 3, panel A). When social distancing measures were put into effect, the assortativity by age disappeared and contacts occurred mainly between household members (Figure 3, panels B–D).

Before lockdown, the estimated R₀ was 2.38 (95% CI 2.01–2.80). During the first period of social distancing measures, in which schools, entertainment venues, and shops were closed, R₀ was estimated to decrease by 42.7% (95% CrI 34.9%–51.3%); under lockdown, R₀ decreased by 81.0% (95% CrI 71.7%–86.1%). Thus, the cumulative measures implemented during lockdown would have reduced R₀ to <1.0 even if the initial R₀ had been as high as 5.3 (95% CrI 3.5–7.2). Estimated R_t was 1.13 (95% CrI 1.38–1.61) during the period of the initial measures but was 0.46 (95% CrI 0.35–0.57) during lockdown (Figure 4, panel A).

We assessed the effect of each measure separately and in combinations (Figure 5). During lockdown, the estimated reduction in R₀ attributed to each measure was 10.3% (95% CrI 5.2%–20.3%) for the decline in work contacts, 18.5% (95% CrI 10.7%–26.3%) for school closures, and 24.1% (95% CrI 14.8%–34.3%) for the decline in leisure activity contacts. Thus, each measure separately would have reduced R₀ to <1.0 if the initial R₀ had been as high as 1.11 for the decline in work contacts, 1.23 for school closures, and 1.32 for the decline in leisure activity contacts. A combination of measures could be effective if the initial R₀ had been as high as 1.78 for interventions reducing work and school contacts, 1.72 for reducing work and leisure contacts, and 1.43 for reducing school and leisure contacts.

We assessed alternative scenarios with less disruptive social distancing measures. A 50% reduction in school contacts, such as smaller class sizes; 20% in work contacts, such as teleworking for part of the population or rotating weekly schedules in which employees telework some days and work onsite other days; and 20% in leisure activities could reduce R₀ to <1.0 for initial levels as high as 1.32 (95% CrI 1.27–1.38). An even larger decline in leisure activities (50%) could successfully reduce an initial R₀ as high as 1.48 (95% CrI 1.35–1.62).

Finally, we assessed the increase in R_t when measures were partially lifted after lockdown. To mimic the measures implemented after lockdown in Greece, we assumed that contacts at work would return to levels 50% lower than pre-pandemic, school to 50%, and leisure to 60%. For instance, class sizes were reduced 50% when schools reopened in May. Under this scenario, R_t would remain <1.0 assuming >20% reduction in susceptibility as a result of infection control measures, including hand hygiene, use of face masks, and maintaining physical distances >1.5 meters (Figure 6). Under milder social distancing measures, infection control policies would need to be much more effective (Appendix Figure 2).

By April 26, 2020, Greece had 2,517 diagnosed COVID-19 cases, 23.0% of which were imported, and 134 deaths (Figure 1) (25). The corresponding naive case-fatality ratio (CFR) was 5.3%. Based on our SEIR model, the cumulative number of infections during February 15–April 26 would be 13,189 (95% CrI 6,206–27,700) (Figure 4, panel B), which corresponds to an attack rate (AR) of 0.12% (95% CrI 0.06%–0.26%). The estimated case ascertainment rate was 19.1% (95% CrI 9.1%–40.6%). By the end of April, 25 (95% CrI 6–97) new infections per day and 329 (95% CrI 97–1,027) total infectious cases were estimated (Figure 4, panels C, D).

On the basis of the number of deaths and critically ill patients reported in Greece by April 26, and using the number of infections obtained from the model as denominator, we estimated the IFR to be 1.12% (95% CrI 0.55%–2.31%) and the cumulative proportion of critically ill patients to be 1.55% (95% CrI 0.75%–3.22%). As a validation, we estimated the number of deaths by applying a published age-adjusted estimated IFR to the number of infections predicted by the model (Appendix Table 3). The predicted number of deaths was 137 (95% CrI 66–279) compared with the reported number of 134 deaths (Appendix Figure 3). As a sensitivity analysis, we simulated the epidemic and calculated IFR and AR assuming a shorter mean serial interval of 4.7 days. We obtained similar results for the AR and the IFR as when the serial interval was 6.67 days (Appendix Figure 4).