The unequal effects of the health-economy trade-off during the COVID-19 pandemic

The study investigates whether mandated non-pharmaceutical interventions (NPIs) and spontaneous behavioural changes during COVID-19 produce similar economic and public health outcomes and how these effects differ across socioeconomic groups. Early pandemic responses included closures of non-essential, customer-facing activities and widespread work-from-home policies, which unevenly affected exposure risk and employment across industries and occupations. Concurrently, individuals voluntarily reduced contacts and patronage of customer-facing services due to fear of infection, but the magnitude and distributional impacts of such behaviour relative to NPIs remain debated. To address these questions, the authors develop a fine-grained, mechanistic, agent-based model that jointly simulates epidemic spread and economic activity at individual, occupational, industry, and income-group levels, focusing on the New York-Newark-Jersey City metro area during the first COVID-19 wave.

Prior work on COVID-19 policy effects often treats epidemic and economic dimensions separately or at aggregated levels, limiting insight into heterogeneity by socioeconomic groups. Some integrated models and agent-based approaches exist, but many are calibrated to a few aggregate indicators and used qualitatively. The authors build on earlier epidemic network models and economic input-output frameworks, extending them into a unified agent-based model with granular socioeconomic attributes (age, income, occupation, work-from-home possibility) and mobility-informed contact networks. This addresses gaps by enabling detailed analysis of differential impacts across industries, occupations, and income groups. The study situates itself within behavioural epidemiology literature, modelling risk-driven behaviour change (fear of infection) as a functional response to reported deaths.

Design: A data-driven, joint epidemic-economic agent-based model (ABM) for a synthetic population of 416,442 individuals representative of the New York-Newark-Jersey City, NY-NJ-PA metro area, with heterogeneity in household composition, age, income, occupation, and work-from-home (WFH) capability. Industries are modelled at two-digit NAICS (20 industries) via a representative-firm approach and linked through a regionalized input–output (IO) network (Flegg Location Quotient applied to BEA national tables to split NY MSA vs rest of US). Epidemic module: Transmission occurs on a multilayer contact network (household, school, workplace, community). Community and workplace layers are reconstructed probabilistically from privacy-preserving GPS mobility data (Cuebiq) and Foursquare places, matched to synthetic individuals by census tract. A stochastic, discrete-time compartmental model extends SLIR with pre-symptomatic (P_s), symptomatic infectious (I), and asymptomatic infectious (I_a) states, with age-specific symptomatic probabilities and empirical IFR for deaths, and reporting delays. Per-contact infection probabilities depend on setting-specific weights; indoor/outdoor differentiation in community is included. Epidemiological parameters are calibrated to ancestral SARS-CoV-2. Economic module: A dynamic macro framework tracks industry employment, output, consumption, and final demand. Steps each day: (1) Industries choose workforce needs based on past employment, demand, and restrictions; hiring/firing is random conditional on worker eligibility; WFH productivity is assumed unchanged; (2) Households set consumption demand by age and income and adjust with employment changes and epidemic conditions; fear reduces demand only in customer-facing industries (entertainment, accommodation-food, other services, retail, transportation, health, education); (3) Total final demand = household consumption + intermediate demand (IO linkages) + other components (government, investment, net exports); (4) Production meets demand subject to labor constraints; shortages rationed pro-rata. Prices are assumed constant over the short horizon. Coupling: Modules are coupled with one-day lag. Economic module takes reported deaths D_{t-1} to compute behaviour change (fear of infection) and reduces customer-facing consumption accordingly; epidemic module takes employment status to update workplace/community contacts (fired workers lose workplace contacts; newly hired gain them). Fear acts via an exponential response Λ(φ, D_{t-1}) = 1 − exp(−φ D_{t-1}). In the economic module, reduction in consumption for customer-facing industries is Λ_ECO = Λ(φ_ECO, D_{t-1}) T_k, where T_k indicates customer-facing. Community contacts are reduced proportionally with φ_EPI = φ_ECO/φ and Λ_EPI = Λ(φ_EPI, D_{t-1}) T_k; implemented by scaling contact weights. Optional spontaneous WFH (absent mandates) reduces workplace contacts among WFH-capable workers according to a similar functional form; no productivity loss assumed. Calibration and timeline: Simulations run daily from 12 Feb 2020 to 30 Jun 2020. Protective measures begin 16 Mar and relax 15 May (schools closed and WFH persist). Parameters, including fear strength, are calibrated using Approximate Bayesian Computation to match key epidemic (deaths) and economic statistics (employment, GDP, consumption by sector). Uncertainty reflects stochastic transmission/hiring-firing and parameter posteriors. Interventions and counterfactuals: NPIs include (1) closures (empirical non-essential closures; variants: only customer-facing closures; no closures), (2) mandated WFH for capable workers, and (3) school closures. Behavioural change scenarios vary fear magnitude: baseline (estimated; implies ≈14% peak customer-facing consumption reduction from fear), low (0.1× baseline; ≈1%), high (10× baseline; ≈77%). Start dates of protective measures vary (early: 17 Feb; baseline: 16 Mar; late: 30 Mar). An age-specific fear variant assigns stronger fear to older vs younger individuals.

Model validation: The model reproduces targeted economic indicators from 2019Q4 to 2020Q2, including a larger drop in employment than GDP and sharper reductions in customer-facing consumption versus non-customer-facing. It matches deaths over March–June 2020. Out-of-sample validations include: industry-level employment changes with Pearson r = 0.82 (P = 2×10^-9), exceeding supply-shock-only correlation r = 0.69 (P = 4×10^-3); community contact reductions across customer-facing industries (r = 0.75, P = 0.05); workplace contact reductions (r = 0.88, P = 5×10^-10). The model reproduces that low-income individuals experienced larger employment losses but smaller consumption cuts than high-income individuals. Behaviour change vs NPIs: Both stronger fear of infection and stricter closures reduce deaths while increasing unemployment, indicating a similar health–economy trade-off from demand- and supply-side channels. Example comparisons: with baseline fear, moving from all-open to non-essential closures increases unemployment by 64% and reduces deaths by 35% (relative to the all-open baseline-fear scenario). With empirical closures, increasing fear to high raises unemployment by 40% and reduces deaths by 50% (relative to the empirical scenario). Effects are stronger for low-income workers concentrated in customer-facing, in-person occupations. Timing and scope: Closing only customer-facing industries outperforms closing all non-essential activities in terms of deaths averted per job lost. In an ‘all open’ setting, age-specific fear (higher for older individuals) modestly improves outcomes compared to uniform fear, reducing deaths by about 6% and unemployment by about 5%. By contrast, closing customer-facing industries reduces deaths by 28% and raises unemployment by 22%. Delaying protective measures can substantially increase fatalities and, in high fear scenarios, also increase unemployment; early measures with targeted closures are most effective. Age-specific dynamics and consumption: Under age-specific fear, older adults reduce contacts/consumption more, especially after the epidemic peak, leading to ≈30% fewer out-of-household infections among older individuals around the peak compared to uniform fear; household transmission limits gains. Consumption shifts: larger decreases in health and accommodation-food consistent with age spending patterns; reallocation increases non-customer-facing demand (e.g., finance) more among older groups.

The study directly addresses whether self-organized behavioural changes can substitute for, or complement, mandates in balancing health and economic outcomes. Quantitatively, both stronger behavioural responses and stricter NPIs save lives at the cost of higher unemployment, producing qualitatively similar trade-offs through different mechanisms (demand vs supply shocks). These effects are uneven: low-income workers, overrepresented in customer-facing, in-person jobs, experience larger employment losses but also greater reductions in workplace infections when activity declines, whether due to policy or fear-driven avoidance. Targeting matters: shutting non-customer-facing sectors (e.g., manufacturing, construction) causes substantial job losses with limited mortality benefits, whereas focusing on customer-facing sectors yields better health outcomes per unit of economic loss. Timing matters: behavioural responses tend to lag reported deaths by weeks, reducing their effectiveness early in a wave, whereas NPIs can be enacted immediately to quickly suppress transmission and enable faster consumption rebound. Even when fear is age-specific, improvements over uniform behavioural change are modest due to household transmission and delayed response to rising deaths. Overall, the results suggest that substantial behavioural change can mimic stringent closures in aggregate outcomes, but well-timed, targeted NPIs are more controllable and can be more efficient in reducing deaths with fewer economic costs, especially when complemented by support for affected low-income workers.

This work introduces and validates a granular, data-driven joint epidemic–economic ABM that integrates mobility-informed contact networks with an IO-based economic model at the industry and household levels. Applied to the New York metro area’s first COVID-19 wave, the model reproduces key epidemic and economic patterns and quantifies the health–economy trade-off under both behavioural change and mandated NPIs. Key contributions include: (1) demonstrating that behavioural changes and NPIs can produce similar qualitative trade-offs; (2) revealing pronounced unequal impacts across income groups; (3) highlighting the importance of targeting customer-facing industries and early action; and (4) showing that age-specific behavioural responses yield only modest improvements. Future research directions include extending the framework beyond the first wave to incorporate masks, testing, tracing, quarantine, variants, vaccination, waning immunity, and heterogeneous per-contact risks across occupational settings; integrating differential severity by socioeconomic status; refining behavioural response mechanisms (information processing, media effects, heterogeneity in risk perception); modelling firm-level spatial heterogeneity for policy targeting; and exploring dynamic, threshold-based control policies with practical implementation constraints.

Scope is limited to the first wave in a single metro area; results may differ with later-pandemic elements (masks, testing and tracing, quarantine, variants, vaccination, waning immunity). The matching of mobility traces to synthetic individuals is probabilistic and based on census tracts, lacking user-level socioeconomic attributes. The epidemic model assumes equal per-contact infection risks across occupational settings and does not model differential severity by socioeconomic status. The production side aggregates to industry-level representative firms without firm/location heterogeneity; spatially targeted policies are not evaluated at fine geographic production granularity. Prices are held constant; childcare-related productivity losses from school closures are not modelled. Intermediate input shortages are excluded (assumed second-order early in the pandemic). Behavioural change (‘fear’) is a simplified, time-invariant exponential function of reported deaths and may not capture delayed information processing or media amplification; heterogeneity in perception beyond age is not included.