Quantifying the Benefits of Targeting for Pandemic Response

1.1. Contributions

Our framework overcomes two practical challenges. First, a rigorous quantification of the value of targeting must be grounded in formal optimization. To that end, we embed an optimization framework within a traditional SEIR epidemiological model that differentiates policies based on both population groups and activities, and balances the lost economic value with the cost of deaths. Since embedding complex controls into the highly non-linear, non-convex, SEIR epidemic model renders exact optimization methods not applicable, we focus on designing an algorithm that can tractably produce approximate solutions. The algorithm is a linearization and optimization procedure that we call Re-Optimization with Linearized Dynamics (ROLD), and that is inspired by model predictive control (Bemporad 2006, Camacho and Alba 2013) and trust region methods (Yuan 2015). ROLD sidesteps the lack of convexity in the original problem by repeatedly solving optimization problems that rely on a linearized approximation of the SEIR dynamical system, where each optimization problem is reduced to a linear program.

The second contribution of our paper is a real-life implementation of the aforementioned framework, which overcomes the data availability challenge associated with calibrating a finely targeted model like ours. Our calibration leverages publicly available data on (i) real-time hospitalization, (ii) community mobility, (iii) social contacts, and (iv) socio-economic measures collected during the COVID-19 pandemic in Île-de-France, a region of France encompassing Paris with a population of approximately 12 million. Our implementation combines a complex, multi-group epidemiological model with an economic model that allows quantifying trade-offs between economic activity and epidemiological impact via the ROLD algorithm.

We next leverage the framework to address the core research questions that we posed earlier. We calibrate our optimization model up to October 21, 2020, and then apply our optimization framework to derive optimal intervention policies for subsequent horizons of 90 and 360 days. We draw the following conclusions from these experiments:

1. Dual targeting leads to significant Pareto improvements. Specifically, the optimal ROLD policy that differentially targets both on age groups and activities Pareto dominates the optimal ROLD policies that can only target on age groups, activities, or neither, i.e., it leads to lower economic costs without increasing the number of pandemic deaths. The gains from targeting are also super-additive: targeting on both dimensions gains more than the sum of gains from unilateral targeting. In addition, age group and activity targeted ROLD Pareto-dominates a number of benchmark policies resembling those implemented in practice during COVID-19. Finally, although not an explicit objective of the optimization, the dual-targeted ROLD policy generally results in a reduction in confinements for all population groups, relative to less fine-grained policies.

2. Optimized dual-targeted policies have an interpretable structure, imposing less confinement on group-activity pairs that generate a relatively high economic value prorated by activity-specific social contacts. This allows ROLD to create complementary confinement schedules for different groups that reduce both the number of deaths and economic losses, with the important added benefit that they do not require completely confining any age group. To rigorously explore the interpretability of the policies produced by ROLD, we generate a training dataset by running it on an ensemble of instances, and then use CART to train trees that predict its decisions based on simple, transparent features. The resulting trees fit well and reveal that ROLD’s targeting mechanism is interpretable. The key feature that explains how the targeting is done is the “econ-to-contacts-ratio”, i.e., the ratio of marginal economic value to total contacts generated for a given (group, activity) pair; this provides policy makers with a transparent metric to explain ROLD’s decisions.

3. Since targeted policies could lead to increased disparities in the confinements of distinct groups and thus be perceived as unfair, we incorporate fairness considerations into our framework through “limited disparity” constraints that restrict the allowed difference in confinements faced by distinct groups, and we quantify their impact on efficiency (measured in terms of health and economic costs). While the experiments suggest that the cost of limited disparity may be high, satisfactory intermediate trade-offs may be achievable through limited targeting.

Targeting

Paralleling our aforementioned real-life examples, several papers have studied targeted interventions. Kucharski et al. (2020), Prem et al. (2020), Di Domenico et al. (2020) recognize the importance of heterogeneity in the social contacts generated through activities and examine several interventions limiting them. Though some of the models here are age differentiated, targeting only happens through activities. Population group targeting, either through confinements, testing or vaccinations, has been investigated in Bastani et al. (2021), Acemoglu et al. (2020), Matrajt et al. (2020), Goldstein et al. (2021), Bertsimas et al. (2020), Favero et al. (2020), Birge et al. (2020), Chang et al. (2020), Evgeniou et al. (2020), Giordano et al. (2021). By enforcing stricter confinements for higher risk groups (e.g., older populations when considering mortality risk or younger populations when considering the risk of new infections), such targeted policies have been shown to generate potentially significant improvements in health outcomes, and even in economic value if optimally tailored (Acemoglu et al. 2020). A potential benefit of our approach is that, by exploiting complementarities between group and activity targeting, these higher risk groups may experience less confinement for the same level of aggregate deaths and economic losses.

Optimization of interventions

Our proposed optimization algorithm differs in several ways from existing approaches to assessing interventions via SEIR epidemiological models. A number of papers simulate a small number of candidate interventions, e.g. full lockdown versus school-only lockdown (Kucharski et al. 2020, Prem et al. 2020, Di Domenico et al. 2020, El Housni et al. 2020, Favero et al. 2020), restrict the candidates to a simple parametric class for which exhaustive search is computationally feasible (e.g., trigger policies based on hospital admissions as in Duque et al. 2020 or confirmed cases as Ahn et al. 2021), or use global optimization methods such as simulated annealing (Dutta et al. 2021). When considering a more complex policy space like in our targeting model, such approaches can lead to significantly sub-optimal results and misleading conclusions. Another stream borrows from the optimal control literature to design non-targeted control policies for single-group and single-activity SEIR dynamics (Bose et al. 2021, Pataro et al. 2021, Morris et al. 2021). Although optimization-based, the models there are simpler than our own and do not capture targeting. Birge et al. (2020) use formal optimization for location-based targeting, but in a one-shot model that does not differentiate age groups or activities and does not account for time in the calculation of health or economic impact. The paper that is most related to ours is Bertsimas et al. (2020), which also proposes a multi-group SEIR formulation, in conjunction with an iterative coordinate descent algorithm to optimize vaccine allocations in a differentiated fashion. Although taking different approaches to doing so, both the algorithm proposed there and the ROLD heuristic crucially depend on solving linearized versions of the true SEIR dynamics which are tractable via commercial solvers. However, the model of Bertsimas et al. (2020) focuses on vaccine allocation decisions, whereas ours captures the dynamics of differential confinements and also allows activity-based targeting.

SEIR model calibration

Our work is also related to several other papers that have estimated SEIR parameters, particularly in the COVID-19 context. A number of papers estimate SEIR epidemiological parameters stratified by age groups – we use the estimates from the Île-de-France study in Salje et al. (2020). Another stream of papers, focusing on forecasting COVID-19 spread, estimate when and how the underlying SEIR parameters evolve in response to government interventions and changes in individual behavior, such as Perakis et al. (2021), Li et al. (2021). Lastly, we relate to the literature that has used Google and other mobility data to inform COVID-19 response strategies or to estimate the realized reductions in social contacts during COVID-19 (Dutta et al. 2021, Ilin et al. 2021, Wellenius et al. 2020, Cot et al. 2021, Xiong et al. 2020), as well as the literature on social contacts estimation (Béraud et al. 2015, Prem et al. 2017).

3.1. Some Notation

We denote scalars by lower-case letters, as in v, and vectors by bold letters, as in v. We use square brackets to denote the concatenation into vectors: v := [v₀, v₁]. For a time series of vectors v₁, …, v_n, we use the notation v_i:j := [v_i, …, v_j] to denote the concatenation of vectors v_i through v_j. Lastly, we use v^⊤ to refer to the transpose of v.

3.2. Epidemiological Model and Controls

We rely on a modified version of the discretized SEIR (Susceptible-Exposed-Infectious-Recovered) epidemiological model (Anderson and May 1992, Prem et al. 2020, Salje et al. 2020) with multiple population groups that interact with each other. In our case study we use nine groups g ∈ 𝒢 determined by age and split in 10-year buckets, with the youngest group capturing individuals with age 0-9 and the oldest capturing individuals with age 80 or above. Time is discrete, indexed by t = 0, 1, …, T and measured in days. We assume that no infections are possible beyond time T.

Compartmental Model and States

Figure 1 represents the compartmental model and the SEIR transitions for a specific group g. For a population group g in time period t, the compartmental model includes states S_g(t) (susceptible to be infected), E_g(t) (exposed but not yet infectious), I_g(t) (infectious but not confirmed through testing and thus not quarantined), (infectious and confirmed through testing and thus quarantined; this state is further subdivided into for j ∈ {a, ps, ms, ss} to model different degrees of severity of symptoms: asymptomatic, paucisymptomatic, with mild symptoms, or with severe symptoms; we assume that an infectious individual in group g will exhibit symptoms of degree j with probability p_j,g), R_g(t) (recovered but not confirmed as having had the virus), (recovered and confirmed as having had the virus), and D_g(t) (deceased). Individuals with severe symptoms will need hospitalization, either in general hospital wards (H_g(t)) or in intensive care units (ICU_g(t)). All the states represent the number of individuals in a compartment of the model in the beginning of the time period.

• Download figure
• Open in new tab

Figure 1 Compartmental SEIR model for a specific group g with transition rates.

Susceptible individuals get infected and transition to the exposed state at a rate determined by the number of social contacts and the transmission rate β(t). Exposed individuals transition to the infectious state at a rate σ and infectious individuals transition out of the infectious state at a rate µ. An infectious individual will need to be hospitalized in general hospital wards or ICU with probability and , respectively, where . On average, patients who are treated in general hospital wards (ICU) spend days in the hospital (ICU). An infectious individual with severe symptoms in group g will decease (recover) with probability .

We keep track of all living individuals in group g who are not confirmed to have had the disease N_g(t) := S_g(t) + E_g(t) + I_g(t) + R_g(t), and let X_t = S_g(t), I_g(t), …, D_g(t)] _g∈𝒢 denote the full state of the system (across groups) at time 0≤ t≤ T. We denote the number of compartments by |𝒳|, so the dimension of X_t is |𝒢||𝒳| ×1.

3.3. Resources and Constraints

We use K^H(t) (K^ICU(t)) to denote the capacity of beds in the general hospital wards (in the ICU) on day t. When the patient inflow into the hospital or the ICU exceeds the remaining number of available beds, then the policy maker needs to decide how many patients to turn away from each group. Although in principle our framework allows optimizing over such decisions, we choose to not consider this dimension of targeting because it can be extremely contentious in practice. Instead, we implement a proportional rule that allocates any remaining hospital and ICU capacity among patients from all age groups proportionally to the number of cases requiring admission from each group. More formally, with and denoting the number of patients from age group g who are denied admission to general hospital wards and ICU, respectively, in period t, the proportional rule is: where and denote the inflow of patients from group g into the hospital and ICU in period t, respectively. We assume that all patients that are denied admission die immediately.

We further assume a given capacity for viral tests each day, which we denote by K^Vtest(t) on day t. We assume that viral tests used to test individuals with severe symptoms that enter the ICU or hospital, as well as viral tests that test hospitalized individuals to confirm they have recovered, come from a different pool of tests and do not consume the capacity for viral testing in the non-hospitalized population.

We can now write a complete set of discrete dynamical equations for the controlled SEIR model ((EC.1)-(EC.11) in Appendix EC.1) and summarize these using the function where ΔX_t := X_t+1 − X_t. Additionally, we can also write the following constraints for the optimization problem: We denote by 𝒞 (X_t) the feasible set described by (5)-(9) for the vector of decisions u_t at time t.

3.4. Objective

Our objective captures two criteria. The first quantifies the total deaths directly attributable to the pandemic, which we denote by Total Deaths(u_{0:T −1}) := ∑_g∈𝒢 D_g(T) to reflect the dependency on the specific policy u_{0:T −1} followed.

The second criterion captures the economic losses due to the pandemic, denoted by Economic Loss(u_{0:T −1}). These stem from three sources: (a) lost productivity due to confinement, (b) lost productivity during the pandemic due to individuals being quarantined, hospitalized, or deceased, and (c) lost value after the pandemic due to deaths (as deceased individuals no longer produce economic output even after the pandemic ends).

To model (a), we assign a daily economic value v_g(l) to each individual in group g that depends on the activity levels l := [l_g]_g∈𝒢 across all groups and activities. For the working age groups, v_g(l) comes from wages from employment and is a linear function of group g’s activity level in work and of the average activity levels in leisure, other and transport for the entire population (equally weighted). This reflects that the value generated in some industries, like retail, is impacted by confinements across all these three activities. For the school age groups, v_g(l) captures future wages from employment due to schooling and depends only on the group’s activity level in school . For (b), we assume that an individual who is in quarantine, hospitalized, or deceased, generates no economic value. For (c), we determine the wages that a deceased individual would have earned based on their current age until retirement age under the prevailing wage curve, and denote the resulting amount of lost wages with .

The overall economic loss is the difference between the economic value that would have been generated during the time of the pandemic under a “no pandemic” scenario, denoted by V, and the value generated during the pandemic, plus the future economic output lost due to deaths. All the details of the economic modelling are deferred to Appendix EC.2.

To allow policy makers to weigh the importance of the two criteria, we associate a cost χ to each death, which we express in multiples of French GDP per capita. Our framework can capture a multitude of policy preferences by considering a wide range of χ values, from completely prioritizing economic losses (χ = 0) to completely prioritizing deaths (χ → ∞).

3.5. Optimization Problem

The optimization problem we solve is to find control policies (for confinement and testing) that minimize the sum of mortality and economic losses² subject to the constraints that (i) the state trajectory follows the SEIR dynamics, and (ii) the controls and states respect the capacity and feasibility constraints discussed above. Formally, we solve:

4.1. Linearization and Optimization

The key idea is to solve the problem in a shrinking-horizon fashion, where at each time step k = 0, …, T we linearize the system dynamics and objective (over the remaining horizon), determine optimal decisions for all k, …, T, and only implement the decisions for the current time step k.

We first describe the linearization procedure. Recall that the true evolution of our dynamical system is given by (4). The typical approach in dynamical systems is to linearize the system dynamics around a particular “nominal” trajectory. More precisely, assume that at time k we have access to a nominal control sequence and let denote the resulting nominal system trajectory under the true dynamic (4) and under . We approximate the original dynamics through a Taylor expansion around : where ∇_XF_t and ∇_uF_t denote the Jacobians with respect to X_t and u_t, respectively. Note that these Jacobians are evaluated at points on the nominal trajectory, so (14) is indeed a linear expression of X_t and u_t. By induction, every state X_t under dynamic (14) will be a linear function of u_τ for τ < t, and all the constraints will also depend linearly on the decisions.

In a similar fashion, we also linearize the objective (11). Since v_g(l(t)) is linear in u_t for all t = 0, …, T − 1, the objective contains bilinear terms and can be written compactly as: for some matrix M with dimensions |𝒢||𝒰| × |𝒢||𝒳|, and vectors γ and η of dimensions |𝒢||𝒳| ×1 (detailed expressions are available in Appendix EC.3). By linearizing this using a Taylor approximation, we consider the following objective instead: which depends linearly on all the decisions u₀,…, u_{T −1}.

5.1. Parametrization and Calibration

We parametrize the epidemiological model using the confidence regions for SEIR parameters reported in Salje et al. (2020) for the Île-de-France region. We calibrate the model using public data on (i) community mobility, (ii) social contacts, (iii) health outcomes, and (iv) economic output. In particular, we use Google mobility data (Google 2020) to approximate the mean effective lockdowns for all activities during the horizon of interest. Based on these, as well as social contacts data for France (Béraud et al. 2015, Wille et al. 2020), we simulate our SEIR model to generate several potential sample paths, which we use in conjunction with real data from the French Public Health Agency (French Government 2020) on hospital and ICU utilization and deaths in Île-de-France to generate a fitting error metric. Lastly, we estimate values of all parameters of interest by minimizing the sample-average-approximation of the error metric. We calibrate our economic model using data for France, and where available Île-de-France, on full time equivalent wages and employment rates from the French National Institute of Statistics and Economic Studies, and sentiment surveys on business activity levels during confinement from the Bank of France. We provide all the details for calibration and parameter specification in Appendix EC.4. We report experimental results from sensitivity and robustness analyses on the fitted parameters in Appendix EC.6.

5.2. Experimental and Optimization Setup

We run experiments over a range of values for the parameters of our model. Table 1 summarizes the values we use for each parameter in our experimental setup, as well as the details of the optimization setup. For parameters for which multiple values are used in our experiments, the “Baseline Value” column reports the values of the parameters used in our baseline setting, as reported in results in the main paper and the Appendix, unless specified otherwise. In particular, we use a baseline capacity of 2900 beds for ICU in Île-de-France, and also experiment with ICU capacities that range from 2000 to 3200 beds.⁴ We use an infinite capacity for general hospital wards. All results reported in the main paper are obtained under a testing capacity of zero, so only confinement decisions are optimized and compared. (Appendix EC.6.5 discusses the additional benefits of targeted administration of viral tests across age groups.) We optimize decisions starting on October 21 2020, using an optimization horizon of T = 90 days in the experiments reported in the main paper, and allowing up to T = 360 days in additional experiments reported in Appendix EC.6.4. We allow the confinement decisions to change every two weeks.

To quantify the benefits of targeting, we consider several ROLD policies that differ in the level of targeting allowed, which we compare over a wide range of values for χ, from 0 to 990× the annual GDP per capita in France.⁵ For each χ value, we calculate all the ROLD policies of interest, and we record separately the economic losses and the number of deaths generated by each policy. The four versions of ROLD we consider are no targeting whatsoever (NO-TARGET), targeting age groups only (AGE), targeting activities only (ACT), or targeting both (AGE-ACT, or simply ROLD when no confusion can arise). To obtain each of the four variants (NO-TARGET, AGE, ACT, AGE-ACT), we run constrained versions of the optimization problem, imposing the following additional constraints: We impose no additional constraints for AGE-ACT. All four variants are allowed to change the confinement policy through time. For each variant, ROLD gets initialized at the solution of a gradient descent algorithm subject to the corresponding constraints.

6.1. How large are the gains from dual targeting?

To isolate the benefits of each type of targeting, we compare the four versions of ROLD that differ in the level of targeting allowed, as described in Section 5.2. Figure 2a records each policy’s performance in several problem instances parameterized by the cost of death χ. A striking feature is that each of the targeted policies actually Pareto-dominates the NO-TARGET policy, and the improvements are significant: relative to NO-TARGET and for same number of deaths, economic losses are reduced by EUR 0-2.9B (0%-35.9%) in AGE, by EUR 0.4B-2.1B (4.5%-49.8%) in ACT, and by EUR 3.3B-5.3B (35.7%-80.0%) in AGE-ACT. This Pareto-dominance is unexpected since it is not explicitly required in our optimization procedure, and it underlines that any form of targeting can lead to significant improvements in terms of both health and economic outcomes.

• Download figure
• Open in new tab

Figure 2

The total number of deaths and the economic losses generated by ROLD policies with different levels of targeting and by the benchmark policies. Panel (a) compares the four versions of ROLD that differ in the level of targeting allowed. Panel (b) compares the ROLD policy that targets age groups and activities with the benchmark policies. Each marker corresponds to a different problem instance parametrized by the cost of death χ. We include 128 distinct values of χ from 0 to 990×, and panel (b) also includes a very large value (χ = 10¹⁶×).

When comparing the different types of targeting, neither AGE nor ACT Pareto-dominate each other, and neither policy dominates in terms of the total loss objective (Figure 3a). In contrast and crucially, AGE-ACT Pareto-dominates all other policies, which implies its dominance in terms of the total loss objective. Moreover, targeting both age groups and activities leads to super-additive improvements in almost all cases: for the same number of deaths, AGE-ACT reduces economic losses by more than AGE and ACT added together (Figure 4). This suggests that substantial complementarities may be unlocked through targeting both age groups and activities, which may not be available under less granular targeting. These results are robust under more problem instances (Appendix Figure EC.2).

• Download figure
• Open in new tab

Figure 3

Total losses generated by ROLD policies with different levels of targeting and by the benchmark policies, at different values of the cost of death χ. Panel (a) compares the four versions of ROLD that differ in the level of targeting allowed. Panel (b) compares the ROLD policy that targets age groups and activities with the benchmark policies. Each marker corresponds to a different problem instance parametrized by the cost of death χ. We include 128 distinct values of χ from 0 to 990×

• Download figure
• Open in new tab

Figure 4

The super-additivity of ROLD AGE-ACT. The figures compare the improvement from AGE-ACT with the sum of the improvements from AGE and ACT. All improvements are with respect to NO-TARGET. Panel (a) compares the improvements in economic losses for the same number of deaths. Panel (b) compares the improvements in total losses for the same cost of death χ.

To confirm the significance of these gains, we also compare ROLD AGE-ACT with various practical benchmark policies in Figures 2b, 3b. Benchmarks ICU-t and Hybrid-t AND/Hybrid-t OR mimic implementations in the U.S. Austin area (Duque et al. 2020) and, respectively, France (Lehot and Borgne 2020). These policies switch between a stricter and a relaxed confinement level based on conditions related to hospital admissions and occupancy and the rate of new infections (Appendix EC.5). We also consider two extreme benchmarks corresponding to enforcing “full confinement” (FC) or remaining “fully open” (FO); these can be expected to perform well when completely prioritizing one of the two metrics of interest, with FC minimizing the number of deaths and FO ensuring low economic losses.

ROLD Pareto-dominates all these benchmarks, decreasing economic losses by EUR 5.3B-16.9B (71.0%-82.6%) relative to Hybrid-t AND, by EUR 7.1B-11.6B (62.2%-82.8%) relative to Hybrid-t OR, and by EUR 5.4B-11.6B (62.2%-78.0%) relative to ICU-t for the same number of deaths. Additionally, ROLD meets or exceeds the performance of the two extreme policies: for a sufficiently large χ, ROLD exactly recovers the FC policy, resulting in 890 deaths and economic losses of EUR 27.6B; for a sufficiently low χ, ROLD actually Pareto-dominates the FO policy, reducing the number of deaths by 16,688 (76.7%) and reducing economic losses by EUR 1.6B (65.3%). The latter result, which may seem surprising, is driven by the natural premise captured in our model that deaths and illness generate economic loss because of lost productivity; thus, a smart sequence of confinement decisions can actually improve the economic loss relative to FO. Among all the policies we tested, ROLD AGE-ACT was the only one capable of Pareto-dominating the FO benchmark, confirming that dual targeting is critical and powerful.

Another possible benefit of finer targeting is that it could reduce time in confinement for all groups, even though this is not an explicit objective of ROLD. To check this, we calculate the fraction of time spent by each age group in confinement under each ROLD policy, averaged over the activities relevant to that age group (Appendix EC.6.2). The results are visualized in Figure 5, which depicts boxplots for the fractions of time in confinement across all problem instances parameterized by χ.

• Download figure
• Open in new tab

Figure 5

Average time in confinement for the ROLD policies with different targeting types. Each boxplot depicts the fraction of time the age group spends in confinement under the respective policy averaged over the activities relevant to that age-group, for different problem instances parameterized by χ.

The dual-targeted AGE-ACT policy is able to reduce the confinement time quite systematically for every age group, relative to all other policies. Specifically, it results in the lowest confinement time for every age group in 70% of all problem instances when compared with NO-TARGET, in 60% of instances when compared with AGE, in 83% of instances when compared with ACT, and in 50% of instances when compared with all other policies. Moreover, the fraction of confinement time achieved by AGE-ACT is within 5% (in absolute terms) from the lowest confinement time achieved by any policy for every age group, in 76% of all instances; within 10% in 80% of the instances; and within 14% in all instances. Thus, even when the dual-targeted policy confines certain age groups more, it does not do so by much. These outcomes are quite unexpected as they are not something that the ROLD framework explicitly optimizes for, but rather a by-product of a dual-targeted confinement policy that minimizes the total loss objective (11).⁶

It is worth noting that although ROLD AGE-ACT reduces confinements for every age group compared to less targeted policies, it does not do so uniformly. Instead, it can lead to a larger discrepancy in the confinements faced by different age groups: those aged 10-59 are generally more confined than those aged 0-9 or 60+.

6.2. How do gains arise from dual targeting?

We examine the structure of the optimal ROLD AGE-ACT confinement decisions. Figure 6 visualizes the optimized confinement policy for the value χ = 50×, which is in the mid-range of estimates used in the economics literature on COVID-19 (Alvarez et al. 2020) and is representative of the behavior we observe across all experiments.

• Download figure
• Open in new tab

Figure 6

The optimized ROLD AGE-ACT policies for a problem instance with a 90-day optimization horizon starting on October 21, 2020, and with cost of death χ = 50×. (See Appendix EC.6.4 for optimized ROLD policies with a 360-day optimization horizon.) From top to bottom, the seven panels depict the time evolution for the occupation of hospital and ICU beds, the number of actively infectious individuals and the cumulative number of deceased individuals in the population, and the confinement policy imposed by ROLD in each age group and activity. In panels 3-7, the values correspond to the activity levels allowed and are color-coded so that darker shades capture a stricter confinement.

Generally, the ROLD policy maintains high activity levels for those groups with a high ratio of marginal economic value to total social contacts in the activity, i.e., a high where v_g(l(t)) is the economic value created by an individual in age group g when activity levels are l(t). For example, in work, ROLD completely opens up the 40-69 y.o. groups, while confining the 20-39 y.o. groups during the first two weeks and the 10-19 y.o. group for the first ten weeks. This is explainable since the 40-69 y.o. groups produce the highest econ-to-contacts-ratio in work, while younger groups have progressively lower ratios. Similarly, ROLD prioritizes activity in transport, then other, then leisure, in accordance with the relative econ-to-contacts-ratio of these activities.

To understand how complementarities arise in this context, note that the ability to separately target age groups and activities allows the ROLD policy to fully exploit the fact that distinct age groups may be responsible for the largest econ-to-contacts-ratio in different activities. As an example, the 20-69 y.o. groups have the highest ratio in work, whereas the 0-19 y.o. and 70+ y.o. groups have the highest ratio in leisure. Accordingly, we see that ROLD coordinates confinements to account for this: groups 20-69 y.o. remain more open in work but face confinement in leisure for up to the first ten weeks; on the contrary, the 10-19 y.o. group is confined in work for a long period while remaining open in leisure, and the 70+ y.o. groups also remain open in leisure. These complementary confinement schedules allow ROLD to reduce both the number of deaths and economic losses, with the important added benefit that no age group is completely confined.

To generalize the above insights and gain a better understanding of the ROLD policy, we also take a different approach and train interpretable machine learning models (regression trees) to predict the optimal ROLD confinement decisions as a function of several features that depend on salient epidemiological and economic parameters.

More specifically, we generate 26,140 problem instances, with (i) randomly drawn values for the transmission rate β, the probabilities of an infectious individual having severe symptoms {p_ss,g}_g∈𝒢 and the parameter of the social mixing model α₁ = α₂ = α_1,2, and (ii) a finite set of values for the cost of death χ, the ICU capacity, and the sensitivity of economic value on confinement in transport, leisure and other, ν^{other activities}. We solve these instances using ROLD and then we fit a decision tree to predict the ROLD-optimized activity levels based on 22 features. We include as features several important problem parameters (e.g., β, α₁ = α₂ = α_1,2, ICU capacity, economic parameters) as well as derived features based on parameters and SEIR state values (e.g., econ-to-contacts-ratio, ICU utilization and admission rate, infection rates, etc.). Some of these features are targeted, meaning they take different values at a given time step for different age groups or activities (e.g., econ-to-contacts-ratio), whereas others are non-targeted, such as the transmission rate β, the time t or the cost of death χ. The details of the data generation and fitting procedures can be found in Appendix EC.6.3, and the full set of features is in Tables EC.8 and EC.9.

Figure 7 displays the depth-four tree obtained by fitting using the entire feature set, together with the resulting root-mean-squared-error (RMSE). This simple tree predicts the optimal ROLD activity levels quite well (RMSE of 0.15) and it confirms our core insight that the econ-to-contacts-ratio is the most salient feature when targeting confinements, as it is used as a split variable in the root node of the tree and also subsequently used for splits in two sub-trees, with higher ratios leading to higher activity levels. The tree also confirms that time and the cost of death χ are relevant: the optimal ROLD policy tends to enforce stricter confinements in earlier periods and subsequently relaxes these through time, and the confinements become stricter for higher values of χ. The only other feature appearing in the tree is R-perc, which quantifies the total number of individuals in a recovered state, as a fraction of the overall population; this is a natural measure of the level of herd immunity, and the ROLD policy generally seems aligned with it in an intuitive way, enforcing stricter confinements at lower levels of herd immunity.

• Download figure
• Open in new tab

Figure 7

Decision tree of depth four approximating the optimized ROLD confinement decisions trained on a total of 26,140 problem instances with a horizon of T = 90 days (a total of 8,232,210 samples). The nodes are color-coded based on the activity level, with darker colors corresponding to stricter confinement.

We also quantify the importance of the various features by calculating their permutation importance scores, a commonly used metric that measures importance as the increase in model prediction error after the values of the respective feature are randomly permuted. For the tree in Figure 7, the scores for econ-to-contacts-ratio, time, R-perc, and χ are respectively 0.436, 0.019, 0.004, 0.002, and all the other features have scores of zero. The econ-to-contacts-ratio feature thus has by far the largest permutation importance score of all features. Figure EC.4 in the Appendix confirms similar results for a tree of depth 10, and Figure EC.5 documents an additional exercise that further shows econ-to-contacts-ratio as the most salient of all targeted features.

The importance of econ-to-contacts-ratio provides increased transparency into how targeted confinement decisions could be made. A policy maker can explain optimized confinement decisions based on characteristics relating to economic value and public health risk, which is much less contentious than justifying confinement policy based merely on age or activity – to that point, note that the depth-four tree does not split on either the age or the activity.

The Appendix contains additional results to complement this section, including a theoretical justification for the salience of the econ-to-contacts-ratio derived in a simplified model (Appendix EC.7), a more detailed discussion of the optimal ROLD confinement policy (Appendix EC.6.4), and a quantification of the value of targeted testing (Appendix EC.6.5).

6.3. The impact of limited disparity requirements

That targeted policies confine some age groups more than others could be perceived as unfair treatment, so it is important to quantify how an intervention’s effectiveness is impacted when requiring less differentiation across age groups. To capture fairness considerations, we embed a set of “limited disparity” constraints in ROLD that allow the activity levels of distinct age groups to differ by at most Δ in absolute terms, in each activity and at any time (see definition in Appendix EC.6.6). Δ = 0 corresponds to a strictly non-discriminatory policy, whereas a larger Δ allows more targeting, and Δ = 1 captures a fully targeted policy. For every Δ, we record the total loss incurred by a ROLD policy with the limited disparity constraints and the increase in total loss relative to a fully targeted ROLD policy. We repeat the experiment for different problem instances parametrized by χ, and Figure 8 depicts boxplots of all the relative increases in total loss, as a function of Δ.

• Download figure
• Open in new tab

Figure 8

The impact of limited disparity requirements. The plot shows the relative (%) increase in total loss generated by a ROLD policy, compared to a fully targeted policy, as a function of the disparity parameter Δ that measures the maximum allowed difference in activity levels for distinct age groups. The experiments are run using several values of χ, which are used to generate each of the boxplots.

The results suggest that limited disparity requirements may be costly: on average, completely eliminating disparity in confinements would increase the total losses by EUR 1.2B (21.6%) and produce an additional 506 deaths (16.6%) and an extra EUR 0.5B of economic losses (18.9%) compared to a fully targeted policy. In certain instances, the increase in total loss could be as high as 63%. The high losses persist even when some limited discrepancy is allowed, dropping at an initially slow rate as Δ increases from 0 and eventually at a slightly faster rate as it approaches 1. This suggests that to fully leverage the benefits of targeting, a high level of disparity must be accepted, but reasonable trade-offs can be achieved with some intermediate disparity.

Why consider optimized dual-targeted interventions?

The first reason are the significantly better health and economic outcomes: for the same or a lower number of deaths, dual-targeted confinements can reduce economic losses more than any of the simpler interventions that uniformly confine age groups or activities. Furthermore, the super-additive gains imply that significant synergies can be generated through finer targeting, with the ability to target along activities improving the effectiveness of targeting along age groups, and vice-versa. Dual-targeting also may enable all age groups to remain more active, resembling normal life more closely compared to less differentiated confinements. This could result in more socially acceptable restrictions, and a more appealing policy intervention overall.

The second reason is the interpretable nature of the optimized targeted confinement policy, which is consistent with a simple “bang-for-the-buck” rule: impose less confinement on group-activity pairs that generate a relatively high economic value prorated by (activity-specific) social contacts. This simple intuition combined with the reliance on just a few activity levels are appealing practical features, as they provide transparency into how targeted confinement decisions could be made.

Lastly, we note that although dual targeting allows and can result in differences in confinements across age groups, such interventions are actually not far from many real-world policy implementations, which have been more or less explicit in their age-based discrimination. Dual targeting can arise implicitly in interventions that only seem to target activities. As an example, France implemented a population-wide 6 p.m. to 6 a.m. curfew during the first few months of 2021 (Reuters 2021), while maintaining school and work activities largely de-confined. This is effectively implementing restrictions similar to ROLD AGE-ACT: since a typical member of the 20-65 y.o. group is engaged in work until the start of the curfew, their leisure and other activities are implicitly limited; moreover, since most individuals aged above 65 are not in active employment, they are not that restricted in these last two activities.

These examples show that some amount of targeting of activities and age groups is already in place and is perhaps unavoidable for effective pandemic management. Given this state of affairs, our framework highlights the significant benefits in explicitly and transparently modelling targeting and identifying the interventions that rigorously optimize overall societal welfare, given some allowable amount of differentiation across age groups.

Challenges and limitations

An immediate practical challenge is data availability. Social contact matrices by age group and activity may be available from surveys on social behavior, which have been conducted in a number of countries; however, further data collection might be required to obtain these matrices for more refined population group or activity definitions. Similarly, economic data is reported by industry activities, but we are not aware of a dataset that splits economic value into separate (group, activity) contributions. Disparate data sources may be difficult to align: for example, social contact surveys and Google mobility reports use different activity categories, which requires non-trivial fitting (Appendix EC.4).

Availability of data also constrains our model’s structure in several ways. Social contacts between age groups only depend on confinements in the same activity, since the available social contacts dataset (Béraud et al. 2015) only reports interactions in the same activity. However, contacts occur as individuals are engaged in different activities (e.g., a services industry professional interacts during work with individuals who are engaged in leisure activities). A more refined contact mixing model that captures such interactions would be more appropriate for this study, provided that relevant data are available. Our economic model similarly ignores cross effects, such as young age groups engaged in school producing value in conjunction with educational staff engaged in work.

Another challenge with targeted interventions is the perception that they lead to unfair outcomes, as certain age groups face more confinement than others. Such discrimination does arise in the optimized dual-targeted policies; our framework partially addresses these concerns through explicit constraints that limit the disparities across groups. Our requirement that limited disparity hold for every time period and every activity is quite strict, and a looser requirement based, e.g., on time-average confinements could lead to smaller incremental losses. Alternatively, one can impose fairness requirements based on the intervention’s outcomes, e.g., requiring that the health or economic losses faced by different groups satisfy certain axiomatic fairness properties Young (1994).

Although we focus on confinement policies, a direction for future research is to investigate how these can be optimally combined with other types of targeted interventions. Appendix EC.6 reports experiments where we optimize a targeted policy based on confinements and randomized testing and quarantining. The framework is sufficiently flexible to accommodate interventions such as contact tracing and also vaccinations, although a careful implementation would require work beyond the scope of this article.

From a more technical standpoint, future research could be devoted to directly learning an optimal intervention from epidemiological and economic outcomes, bypassing the estimation-optimization procedure we adopted here. Work could also be devoted to deriving a theoretical justification for the importance of the econ-to-contacts-ratio by generalizing our simple analysis in Appendix EC.7. Lastly, future work could focus on deriving tractable upper bounds for the performance of controlled SEIR models that could be used to benchmark various interventions applied in practice or proposed in the academic literature.