Importance of non-pharmaceutical interventions in the COVID-19 vaccination era: A case study of the Seychelles

www.jogh.org • doi: 10.7189/jogh.11.03104 1 2021 • Vol. 11 • 03104 The Republic of Seychelles is an archipelago of 115 islands in the Indian Ocean with a population of approximately 98 000. As of June 28, 2021, the Seychelles was one of only a dozen countries that had succeeded in fully vaccinating more than half of their population against COVID-19 [1]. The Seychelles began its vaccination campaign on January 13, 2021, with two-dose Sinopharm and AstraZeneca vaccines. Both vaccines reportedly have at least 78% efficacy against symptomatic disease 14 or more days after the second dose [2,3]. With various non-pharmaceutical interventions (NPIs) in place (Figure S1 in the Online Supplementary Document) and mounting vaccination coverage, the Seychelles suppressed its incidence to an average of 42 daily cases from January 1 to April 15, 2021. By May 5, over 61% of the population was fully vaccinated. Despite the high vaccination coverage, the country experienced a surge of COVID-19 infections soon after most NPIs were lifted in mid-April, reporting the world’s highest number of daily cases per capita and raising concerns about the efficacy of the vaccines [4]. To understand the determinants of the recent surge, and the impact of the interplay between vaccination and NPIs, we used a previously established data-driven dynamic model [5] and calibrated it to reported cases and vaccination rollout in the Seychelles (Appendix S1 in the Online Supplementary Document).


Back-calculation of incidence
We used a Bayesian non-parametric approach to back-calculate the times series of infections based on the daily reported cases of COVID-19 in the Seychelles from October 21, 2020 to May 24, 2021. 1 Letting represent the number of infections in the time interval, the reported cases on day , , can satisfy the convolution equation: 2,3 where is the probability that the time between infection and identification is . Considering the incubation period as a proxy for time from infection to identification, can be directly calculated from the incubation period distribution. Here we make a simplifying assumption that the distribution of the incubation period does not change over time.
We employed a Bayesian approach to estimate posterior densities of the expected incidence of infections in time in a random walk simulation process. The observed number of infections at time (the calendar day ) was modelled by a nonhomogeneous Poisson process: where prior distributions for the were uninformative given by , and , for in the truncated distribution for . The standard deviation, , was set to 10 for our analyses to maintain some variation. Back calculation was implemented in a Bayesian Markov Chain Monte Carlo (MCMC) setting using Nimble, with the R statistical environment acting as the front end. MCMC simulations were run in 5 independent chains, each consisting of 25,000 iterations, with a burn-in period of 10,000 iterations and a thinning factor of 10. To assess convergence, we inspected the trace plots and applied the Gelman-Rubin convergence test by computing the potential scale reduction factors (PSRF) of posterior densities. Figure S1 shows the simulation results for the cumulative incidence of infection with the 95% credible interval.

Agent-based modelling
We fitted an agent-based model of COVID-19 transmission to incidence derived from backcalculation from October 21, 2021, to January 6, 2021. In this period, there were no school or workplace closures, and non-pharmaceutical interventions for stay-at-home, and cancelation of gatherings remained mainly at the "Recommended" level. Border restrictions included a ban on high-risk regions for International travelers ( Figure S1). We determined disease transmissibility by fitting the model to incidence data during this period while accounting for mask-wearing and recommended measures to reduce contact patterns to 80% of the pre-pandemic behaviour. The model simulated scenarios of COVID-19 incidence without vaccination, and when vaccination was implemented on January 13, 2021.

Model structure
We adopted our previous agent-based model 4 with the natural history of COVID-19, including epidemiological classes for susceptible; latently infected (not yet infectious); asymptomatic (and infectious); pre-symptomatic (and infectious); symptomatic (and infectious) with either mild or severe illness; recovered; and dead. The model population was stratified into six age groups of 0-4, 5-19, 20-49, 50-64, 65-79, and 80+ years based on demographics of the Seychelles. 5 Daily contacts between individuals were sampled from a negative-binomial distribution parameterized using an empirically-determined age-specific contact network. 6,7

Disease dynamics
We parameterized the transmissibility of asymptomatic, mild symptomatic, and severe symptomatic stages to be 26%, 44%, and 89% relative to the pre-symptomatic stage. [8][9][10] The incubation period was sampled from a log-Normal distribution [LogN(shape: 1.434, scale: 0.661)] with a mean of 5.2 days. 11 For those who developed symptomatic disease, the presymptomatic stage was sampled from a Gamma distribution [Gamma(shape: 1.058, scale: 2.174)] with a mean of 2.3 days. 9,12 The infectious period following the onset of symptoms was sampled from a Gamma distribution [Gamma(shape: 2.768, scale: 1.1563)] with a mean of 3.2 days. 9,13 Those who did not develop symptomatic disease remained asymptomatic until recovery, with an infectious period that was also sampled from a Gamma distribution [Gamma(shape: 5, scale: 1)] with a mean of 5 days. 13,14 We assumed that recovery from infection conferred immunity against reinfection for at least one year. We also assumed that all symptomatic cases who were not hospitalized self-isolated within 24 hours of symptom onset, and reduced their number of daily contacts by an average of 75%.

Model implementation
We simulated the model with a population of 98000 individuals, assuming no pre-existing immunity. Results were generated by averaging 500 independent Monte-Carlo realizations in each scenario, and credible intervals (CrI) were generated using the bias-corrected and accelerated bootstrap method (with 500 replications). The model was implemented in Julia, which is an open-source, high-performance, dynamic programming language. The simulation codes are available at: https://github.com/thomasvilches/Seychelles Figure S2. The cumulative mean of posterior distributions for infections (black) back-calculated by fitting the convolution equation to daily reported cases (red). The 95% credible interval is represented by a blue shaded area. Simulated outbreaks using the agent-based model (fitted to cumulative infections till January 6, 2021) are shown for the cumulative number of infections without vaccination (orange curve), and with a two-dose vaccine series (green curve). Shaded orange and green areas represent the 95% credible intervals of simulated outbreaks. Grey area represents the vaccination era.