Predicting effectiveness of countermeasures during the COVID-19 outbreak in South Africa using agent-based simulation

The paper addresses how the timing and strictness of non-pharmaceutical interventions (NPIs) influence COVID-19 transmission in South Africa, with a case study of Nelson Mandela Bay Municipality (NMBM). Given limited healthcare capacity, socioeconomic disparities, and potential comorbidities (e.g., HIV, TB), South Africa faces challenging policy trade-offs. The study’s purpose is to evaluate the epidemiological impact of different lockdown levels and durations, especially the risks of easing measures too early, using agent-based simulations that capture spatial and behavioral heterogeneity. The research asks which combinations of lockdown intensity and duration can suppress or slow transmission sufficiently to avoid overwhelming healthcare services and whether observed declines in cases in mid-2020 could be sustained under different policy scenarios.

The authors situate their work within research on NPIs’ timing and effectiveness (e.g., Chinazzi et al., Dignum et al., Martin-Calvo et al., Rocklöv, Sugishita, te Vrugt), and highlight trade-offs between public health and economic activity (Dignum, Silva). They note particular South African and Sub-Saharan African determinants (younger demographics, high TB/HIV prevalence, spatial and economic inequalities). They contrast classic compartmental models (SIR/SEIR) with agent-based models (ABMs), emphasizing ABMs’ ability to incorporate spatial structure, heterogeneous contacts, and stochasticity (Horni et al., Hackl & Dubernet, Müller et al.). Prior ABM COVID-19 work informs their approach, but they introduce a transportation-based ABM (MATSim + Episim) to approximate contact patterns via simulated daily activities and mobility.

Design overview: A two-step approach combines transportation simulation (MATSim) to generate daily activity and movement patterns of a synthetic population with an epidemic simulation (Episim) to model contact-based infections and disease progression. Calibration aligns simulated infections with reported case data through a single infection scaling parameter. Transportation simulation (MATSim): - Version: 12.0-2019w48-SNAPSHOT. - Study area: Nelson Mandela Bay Municipality (NMBM), South Africa. - Population: 10% sample (114,346 agents; 32,597 households; average household size 3.51; mean age 30.27 years) synthesized from the 2011 Census and 2004 travel survey data. - Network: Derived from OpenStreetMap; agents perform activities over a simulated day of 108,000 s (30 h). - Activities: Home, Work, Primary education, Higher education, Shopping, Leisure, Other. Activity prevalence indicates high counts for Home and Other, with relatively fewer Work activities due to synthesis classification. - Modes: Teleportation (walk, bicycle, passenger) and network modes (car, public bus/train, minibus taxis). Given suspended formal bus services, minibus taxis represent the primary public transport. A Demand Responsive Transport (DRT) module approximates minibus operations. - Minibus taxis: Modeled with 2,374 DRT vehicles, each with 15 seats, after calibration for occupancy (avg ~3.01 persons/vehicle), travel distance, rejection rate, and ratio of empty runs (Table 1). Due to data limitations, DRT operates door-to-door rather than stop-based; adjustments ensure realistic contact durations and trip lengths. Epidemic simulation (Episim): - Contacts occur within “containers” corresponding to specific activity locations and vehicles. Only agents co-present in the same container at the same time can transmit infection. - Infection probability per susceptible agent n from infectious agent m at day t: P_{n,t} = 1 - exp[-θ q_{m,t} i_{m,t} τ_{m,n,t}] where q is shedding rate (assumed equal across contacts), i is activity-specific infectivity, τ is contact duration, and θ is a calibration parameter. - Infectivity parameters (activity-specific): Higher for minibus taxis (20) and home (6), adjusted from a Berlin baseline to reflect NMBM conditions; other activities per Table 3. Max secondary infections per agent capped at 3. Disease progression (SEIQR extension): - States: Susceptible (S), Exposed (E), Infectious (I), Infectious in Quarantine (IQ), Seriously sick (Is), Critical (Ic), Recovered (R). - Assumptions and parameters: Non-infectious for first 4 days post-infection; 80% asymptomatic; symptomatic individuals self-quarantine for 14 days with complete contact blocking (including within-household); ~4.5% become seriously sick at day 10 post-infection; 25% of seriously sick become critical the next day; infectious recover at day 16; severe cases recover at day 23. All infected ultimately recover and acquire immunity for the simulation horizon. Policy interventions and scenarios: - South African “risk-adjusted strategy” (Levels 1–5) mapped to quantitative activity reduction factors by category (e.g., Work, Leisure, Education, Shopping, Minibus taxis) as in Table 3. Level 5 is most stringent; Level 1 least. - Simulations start 12 May 2020 with Level 4 in place, then Level 3 from 1 June, tightened on 27 July (school closures). From 26 August 2020, simulated lockdowns of different levels and durations: 30, 60, 90 days, and continuous until 31 Dec 2020. Calibration: - θ calibrated to reported cases (national data scaled to NMBM by population share) from early June to 27 July 2020, prioritizing conservative underestimation. Final θ = 1.5 × 10^-7 (supported by SSE and regression slope diagnostics across daily and cumulative series). - Initial conditions: 10 randomly infected agents to reduce variance; ensemble of 100 runs per scenario; sensitivity checks with ±5–10% θ variations shown. Outputs: - Time series of active infections and cumulative recoveries; distribution of infection sites by activity class across scenarios and durations (Table 6).

- Stringency and duration matter: Strict and prolonged lockdowns substantially reduce active infections; lenient or short-lived measures lead to continued exponential growth or a pronounced second wave. - Timing of relaxation: Ending measures (even after temporary declines) consistently returns infection trajectories to exponential growth. On-and-off strategies risk resurgence unless measures are sufficiently strict and long. - Scenario-specific outcomes: - 30-day additional measures from 26 Aug 2020: Reducing to Level 1–2 yields ~tenfold increase in active cases in 30 days; maintaining stricter Level 3 reduces growth with delayed visible impact. All scenarios show renewed acceleration once measures end on 25 Sep 2020. Peaks occur Oct–Nov 2020, with peak infected and recovered at the intersection around 200,000–300,000 each (approximately half to two-thirds of NMBM’s population). - 60-day measures: Stronger mitigation; strict scenarios show trend reversal during the intervention, but lifting on 25 Oct 2020 still leads to later peaks (shifted into early 2021). - 90-day measures: More infection chains die out during the measure period in Level 4–5; after lifting on 24 Nov, remaining chains resume growth, but overall burden reduced and peak delayed further. - Continuous measures until 31 Dec 2020: Under Level 5, active infections are eradicated in most simulations by year-end; Level 3 strict and above drive sustained declines, whereas Level 1–2 still see increases despite prolonged measures. - Infection locations: Home dominates infection sites across scenarios (≈86–96%), followed by minibus taxis and primary education, reflecting high infectivity parameters and activity prevalence. Stricter and longer lockdowns shift an even larger share of infections to the home due to increased time at home and larger household sizes. - Policy threshold: Measures between strict and lax Level 3 are the minimum in simulations to produce decreasing case numbers; Level 2 is insufficient during rising phases. Given conservative calibration, real-world containment may require Level 4–5. - Initial SA lockdown: The 27 March 2020 Level 5 lockdown was likely sufficient to slow growth, but subsequent relaxations risked a severe second outbreak in NMBM. - Health system implications: Without sufficiently strict measures, sustained exponential growth or a large second peak threatens to overwhelm limited healthcare capacities, increasing mortality among severe and critical patients. - Calibration diagnostics: θ = 1.5×10^-7 provided balanced fit without overestimation; SSEs and regression slopes indicate acceptable correspondence with reported daily and cumulative cases in June–July 2020 under a conservative approach.

The simulations indicate that continuous, strict lockdowns effectively flatten curves, delay peaks, and in the most stringent, sustained scenarios can eliminate transmission chains. In contrast, lenient measures permit exponential spread that outpaces contact reductions. The epidemic’s slowdown in lenient scenarios correlates with high proportions of recovered individuals approaching herd-immunity-like thresholds (roughly 50–66%), consistent with published estimates; however, achieving such levels via natural infection would impose substantial health burdens. The required policy intensity to achieve declining case counts lies around strict Level 3 or higher in the model; because calibration intentionally underestimates transmission, real-world containment likely requires Level 4–5 under rising incidence. Any complete termination of measures ultimately leads to resurgence unless prevalence is driven very low and chains are extinguished. Thus, reliance on short on-off measures is risky and of limited benefit compared to sustained suppression. Methodologically, the transportation-based ABM captures spatial mixing and activity-specific contacts, enabling nuanced testing of targeted restrictions (e.g., education, public transport). This approach provides policy-relevant insights when granular contact data are scarce, though parameter choice and calibration remain challenging and introduce uncertainty.

The study demonstrates that, epidemiologically, strict and prolonged lockdowns are the most effective countermeasures to suppress COVID-19 transmission in NMBM, substantially reducing infections and alleviating pressure on constrained healthcare resources. Intermediate combinations of measures near a strict Level 3 can stabilize or reduce active cases, whereas Level 2 is insufficient during growth periods. The initial South African lockdown likely curbed early growth, but premature relaxation can precipitate a severe second wave. While the model conservatively underestimates transmission, the results cautiously align with the contemporaneous claim that South Africa may have passed a peak in mid-2020, contingent on maintaining adequate controls. Future research could refine activity-specific infectivity and compliance parameters using empirical mobility and contact data; incorporate mortality, waning or partial immunity, vaccination, and variable quarantine adherence; improve representation of informal transport operations; and calibrate against local (sub-municipal) surveillance and seroprevalence data to reduce uncertainty.

- Conservative calibration to reported cases likely underestimates true transmission, delaying simulated peaks and extending epidemic duration relative to reality. - Assumptions: All infected recover and become immune for the study horizon; perfect 14-day quarantine blocks all contacts (including intra-household); uniform shedding rate; maximum of three secondary infections per agent; full compliance with policies—these may not hold in practice. - Initialization with 10 infected agents reduces variance but distorts early dynamics. - DRT-based approximation of minibus taxis (door-to-door) due to limited stop/timetable data; capacity and vehicle numbers adjusted to match macro indicators. - Population is a 10% synthetic sample; some activities ("Other") aggregate heterogeneous settings with low infectivity assigned, potentially misrepresenting risks. - Calibration uses national data scaled to NMBM by population share due to lack of local case data; unreported infections not explicitly modeled. - Environmental factors, behavioral adaptations, and seasonality not explicitly represented.