Evaluating targeted COVID-19 vaccination strategies with agent-based modeling

Abstract

We evaluate approaches to vaccine distribution using an agent-based model of human activity and COVID-19 transmission calibrated to detailed trends in cases, hospitalizations, deaths, seroprevalence, and vaccine breakthrough infections in Florida, USA. We compare the incremental effectiveness for four different distribution strategies at four different levels of vaccine supply, starting in late 2020 through early 2022. Our analysis indicates that the best strategy to reduce severe outcomes would be to actively target high disease-risk individuals. This was true in every scenario, although the advantage was greatest for the intermediate vaccine availability assumptions and relatively modest compared to a simple mass vaccination approach under high vaccine availability. Ring vaccination, while generally the most effective strategy for reducing infections, ultimately proved least effective at preventing deaths. We also consider using age group as a practical surrogate measure for actual disease-risk targeting; this approach also outperforms both simple mass distribution and ring vaccination. We find that quantitative effectiveness of a strategy depends on whether effectiveness is assessed after the alpha, delta, or omicron wave. However, these differences in absolute benefit for the strategies do not change the ranking of their performance at preventing severe outcomes across vaccine availability assumptions.

Fig 1. Model disease states and spatial structure.

(a) Progression of the disease states in the model: susceptible (S) individuals may become exposed (E) to the virus, then progress to being infectious (initially asymptomatic [I_A], possibly progressing to mild [I_M], severe [I_S] or critical [I_C]), and finally recovering (R) or dying (D). I_C individuals can either die or revert to I_S, in which case they will eventually recover. Recovered individuals have strain-specific immunity that changes over time. (b) Nightime lights satellite image [16, 17] of Marion County, FL, the region used for the model’s spatial structure. (c) Locations of the 161.5K model households (orange dots). Roads [18] are shown for reference but are not modeled. Nighttime lights data is available from https://eogdata.mines.edu/nighttime_light/monthly/v10/2024/202401/vcmcfg/ under the Creative Commons Attribution 4.0 International License https://eogdata.mines.edu/files/EOG_products_CC_License.pdf. Map base layer data is available from https://www.openstreetmap.org under the Open Data Commons Open Database License https://opendatacommons.org/licenses/odbl/, copyright by OpenStreetMap contributors.

Fig 2. Individual interactions and behaviors in the model.

Interactions occur when people in the model are in the same location at the same time, and may occur in households, workplaces, schools, hospitals, and long-term care facilities (not shown). Households have an inherent risk tolerance (indicated by color), and probabilistically have inter-household connections homophilously based on that tolerance. The population overall has a time-varying perception of risk of COVID-19 infection that may be different from the actual risk. (a) When the societal perception of risk is lower than a household’s risk tolerance, household members engage in all their normal activities, including socializing with specific other households and patronizing specific high-transmission-risk workplaces like restaurants and bars. (b) When the societal perception of risk exceeds a given household’s risk tolerance, the household will cease high-risk activities (indicated with greyed arrows), while maintaining more essential activities like going to work, school, and patronizing low risk workplaces (e.g., grocery stores). (c, d) Employees and patrons interact in some workplace types, with interactions between employees more likely to result in transmission. (d) When perceived risk is high, risk-intolerant (blue) employees of high-transmission-hazard workplaces still go to work, while risk-intolerant patrons stop visiting these locations (and thus are grayed-out).

Fig 3. Time-varying model inputs and indicators of model performance.

Panels (a-d) show model inputs, and (e-i) compare model outputs to observed data. In (e-i), points and cross-hairs indicate observed values, solid lines median trends, and faded lines sample trajectories. Horizontal gridline values are plotted above October 2020. (a) Seasonal forcing has a 6-month period, peaking in January and July each year; we also considered an alternative model with no seasonal forcing, see Section E in S1 Text. for details. (b) Detection and reporting probabilities by disease outcome. (c) Simulated first, second and third vaccine doses distributed statewide in Florida, used to calibrate the model (but not for evaluating strategies). (d) Societal risk perception, which drives personal protective behaviors in the model, is fitted so that cumulative reported cases in the model match empirical case data for FL (black dots in panel e). For approximately the month of April 2020, non-essential businesses were closed in the state, and thus are closed during this period in the model (gray “lockdown” shaded region). Not shown: schools in the model close during the summers and during spring 2020, and are 50% and 80% open during the 2020–2021 and 2021–2022 school years, respectively. (e-i) Simulated data closely track empirical data for incidence of reported cases (e), daily hospital admissions (f), excess deaths (g), seroprevalence (h), and the fraction of infections that occurred in vaccinees (i). Results in (e-g) are scaled to show values per 10,000 individuals, and VOC waves are labeled as alpha (α), delta (δ) and omicron (o). For empirical seroprevalance data in (h), horizontal bars indicate the dates covered by each data point and vertical lines indicate the 95% CI).

Fig 4. Cumulative vaccine doses administered per 10k over time, by supply level and distribution strategy.

For each combination of the four supply levels (columns) and quarantine policy (dashed lines), we considered four vaccination strategies: ring vaccination (i.e., infection-risk prioritization) (orange), risk prioritization (blue), age prioritization (green), and a standard mass vaccination (black). For low (LS) and middle (MS) supply levels, all strategies use all available doses. For high (HS) and USA supply levels, the strategies sometimes differ in doses delivered due to shortages of individuals eligible for revaccination; only risk- and age-based strategies always use all available doses.

Fig 5. Cumulative incidence of infection and death per 10k people, by supply level and distribution strategy.

Columns represent vaccine supply scenarios. Rows represent infection (top) and death (bottom) outcomes. Median values and 90% interquantile range are shown as bold lines and shaded ribbon, respectively. For infections, the major effects are supply level (columns) and the policy of quarantining (dashed lines) or not quarantining (solid lines), whereas the four vaccination strategies perform similarly. For deaths, supply level and quarantine are again the strongest factors. However, a strong effect of vaccination strategy also emerges: relative to a standard vaccine roll-out (black), risk-based vaccination (blue) and age-based vaccination (green) are more effective at preventing deaths, whereas ring vaccination (orange) is less effective. See the text for further explanation.

Fig 6. Non-Quarantining Strategy Dominance.

As in Fig 5, columns represent vaccine supply scenarios and rows represent disease outcomes. Each row of pixels in a panel corresponds to one simulation replicate. Color indicates which strategy would prevent the most cumulative incidence had it been pursued up to that day. Replicates are ordered first by which strategy has the highest advantage in the most replicates, then by the magnitude of that advantage. Advantage is calculated each day as the additional prevented deaths versus the next-best strategy, and the overall advantage of a strategy within a replicate is the sum across all days. Relatively low advantage is mostly transparent in this figure, high advantage is mostly opaque. Strain dominance is indicated in the date band at bottom for alpha, delta, and omicron variants. Against infections (top row), the ring-based approach is typically preferable over the time period considered, with a brief interval during omicron where random vaccine distribution would have been preferred for high-supply scenarios. For deaths, the risk-prioritized strategy generally dominates within a few months of introduction, though for low-supply settings other strategies are preferred in a minority of replicates. See Figs K-M in S2 Text for strategies with quarantining, as well as without any seasonal forcing, but in summary: the same qualitative trends are present, i.e., the vaccine distribution strategy preferences suggested by the model are independent of both quarantining and seasonality assumptions.

Fig 7. Cumulative effectiveness after variant waves.

Columns depict vaccine supply scenarios and rows separate infection and death results; these are effectively “snapshot” values of the effectiveness curves at 2021-05-07 (post alpha), 2021-11-26 (post delta), and 2022-03-07 (post omicron) (see Figs J and N in S2 Text). “Waves” are defined generally as the time from when a VOC is introduced to when a new VOC is introduced (however the alpha period starts at the beginning of the simulation and omicron period ends at the end of the simulation). The non-quarantining, standard strategy is used as the baseline for all comparisons.