COVID-19 Pandemic Response Simulation: Impact of Non-pharmaceutical Interventions on Ending Lockdowns

Model overview

We developed an agent-based simulation based on the open-source model FRED (a Framework for Reconstructing Epidemic Dynamics)¹. Our current UWFRED model considered King County’s population distribution and daily contact pattern, COVID-19 transmission using SEIR model, current NPIs and was simulated for 365 days. We estimated new infection per day to measure the effects of easing social-distancing restrictions.

Baseline assumptions include:

• First infection in King County occurred on January 15^th, 2020. Although the man lived in Snohomish County, he commuted to Seattle and thus might have impact on King County². We also assumed this is the only non-local transmission. All other transmissions are community transmissions in King County.

• We divided patients into non-severe and severe groups. We assumed asymptomatic and mild symptomatic patients show similar patterns and considered them as non-severe cases. Severity of symptom depends on age³, and degree of severity does not influence infectivity.

• We assumed full immunity for people who recovered from COVID-19.

Calibration procedure

We calibrated model parameters including COVID-19 transmissibility, daily contact patterns in household, neighborhood, school, and workplace, and probability of staying home when sick, targeting the basic reproduction number (R₀) of COVID-19 which is known to be 2 to 3⁴. We used Latin hypercube sampling to form 1,000 candidate parameter sets. For each parameter set, we replicated the epidemic simulation 50 times to consider stochasticity of the simulation. After selecting the best calibrated parameter sets based on a goodness-of-fit criterion, we clustered them using complete linkage and Euclidean distance method to group results, determining six representative parameter sets (see Appendix 1).

Interventions

We modeled three most common NPIs in the U.S. See Table 1.

Other possible NPIs, such as contact tracing and voluntary home quarantine, were not considered as they were either not being practiced in large scale in King County, or current social distancing already reflects such interventions.

Scenarios for ending lockdowns

We simulated combinations of eight different levels of social distancing strategies that could be practiced after May 4^th, the current final date of Washington government’s “Stay Home, Stay Healthy” emergency order, and a ‘do nothing’ scenario. We set interventions before May 4^th based on Washington state orders (Appendix 4). As extending school closure to the next school year is practically difficult, we assumed schools will reopen in September. See Table 2 and Figure 1.

• Download figure
• Open in new tab

Figure 1.

Diagram of Simulated Scenarios

Calibration result

Out of 1,000 Latin hypercube parameter sets, 58 sets targeted R₀ between 2 to 3. After clustering, six clusters had high similarity and low distance level (Appendix 1). Therefore, for each scenario, we used the centroid of the clusters and replicated 50 times for each centroid, totaling 300 simulations per scenario.

Scenarios outcomes

Comparing Scenario 0 to Scenarios 1-8 shows that NPIs activated on mid-March were effective in flattening the curve and lowering total infection. As expected, Scenario 1, the most powerful and long-lasting social distancing, was most effective in reducing total infections from 41% to 15% of the population, compared to Scenario 0. New incidence peaked on March 18^th in Scenario 0 with 37,304 newly infected persons whereas other scenarios peaked on March 11^th, with 18,881 persons. The ranking by effectiveness in reducing transmission is Scenario 1, 2, 5, 3, 6, 4, 7, 8, 0. See Figure 2.

• Download figure
• Open in new tab

Figure 2.

Simulated Scenarios Projections

However, Scenarios 1 to 8 all yielded a second wave. Even in Scenario 1, the best case, whose daily new infection dropped to 2 on June 4^th and remained around 20 per day until August, as soon as the social distancing strategies were relaxed and school opened, its incidence peaked again to 874 cases per day without import cases. It is noteworthy to look at the total infection since May 4^th for Scenarios 2 and 5. Although Scenario 2 totally reopened society from September 1^st, the total infection is similar to Scenario 5 which retained low social distancing from September until the end of the simulation. This implies that extending one more month of strong social distancing until June 4^th could be as effective as retaining low social distancing in September. See Figure 3.

• Download figure
• Open in new tab

Figure 3.

Projected Second Waves

In addition, the infections by place in Scenarios 1 and 2 (Figure 4) suggest that when social distancing is totally relaxed, peak workplace infection would double and neighborhood infection would more than triple compared to a low level of social distancing.

• Download figure
• Open in new tab

Figure 4.

New Infections by Place