Impacts of COVID-19 interventions: Health, economics, and inequality

Coronavirus disease 2019 (COVID-19) is exacerbating inequalities in the US. We build an agent-based model to elucidate the differential causal effects of nonpharmaceutical interventions on different communities and validate the results with US data. We simulate viral transmission and the consequent deterioration of economic condi-tions on socioeconomically disadvantaged and privileged popula-tions. As found in data, our model shows that the trade-off between COVID-19 deaths and deaths of despair, dependent on the lockdown level, only exists in the socioeconomically disadvantaged population. Moreover, household overcrowding is a strong predictor of the infection rate. The model also yields new insights that ﬁll in the gaps of our data analysis. While subsidisation narrows the socioeconomic gap in deaths of despair, the combination of testing and contact tracing alone is effective at reducing disparities in both types of death. Our results contribute to policy modelling and evaluation for reducing inequality during a pandemic.

show a pruned tree here to illustrate the method and provide the full tree in Extended Data Fig. 1. The x and y-axes of each scatterplot are the feature used for the split and the number of deaths per 100, 000 people, respectively. ZCTAs are divided into two subsets at the vertical lines so that the death rates are close to the average (marked by horizontal lines) within each group. b, We compute the importance of a feature in the decision tree as the normalized total reduction of the mean squared error that is attributable to the feature. Fig. 1a shows a pruned tree that is fitted to the ZCTA-125 level data (see the complete decision tree in Extended Data 126 Fig. 1). The top scatterplot contains all ZCTAs in the dataset. 127 Income is identified as the feature that best splits the set 128 with a threshold at US$122, 200. The percent of 65-and-129 older population is the best variable to further split the lower-130 income group (at 17.85%), whereas the percent of household 131 overcrowding is chosen to divide the higher-income group 132 (at 3.72%). The decision tree is built iteratively this way. 133 Although our goal with the dataset is to evaluate feature 134 importance rather than predict the death rate, the decision 135 tree sheds light on the link between regional characteristics and 136 local health outcomes. High COVID-19 death rates are often 137 associated with low income, a large population of seniors and 138 racial minorities, lack of health insurance, high eviction rates, 139 household overcrowding, commuting, and uncommonness of 140 working from home. An exception to this pattern is the first 141 appearance of overcrowding in the decision tree as shown 142 in Fig. 1a. Surprisingly, the ZCTAs with more household 143 overcrowding had lower death rates. It turns out that these 144 ZCTAs are mostly in Lower and Midtown Manhattan where 145 single young professionals with high salaries tend to live. 146 We also compare the best and worst segments in the deci-147 sion tree and find economic inequality in addition to health 148  disparities. Not only did the worst segment have a higher un-149 employment rate (3.03%) than the best one (2.88%) in 2019, 150 but the former group also had a steeper increase (5.69%) in 151 2020 than the best segment (4.22%). The 2020 unemployment

165
We further investigate the effects of income on regional 166 health and economic outcomes. We compare the poorest and 167 richest counties in the US as measured by median income 168 and find that the widely believed health and economic trade-169 offs of lockdowns only exist in poor counties (Fig. 2) for rich counties, with the latter possibly due to residual effects 185 from the first wave of COVID-19 ( Fig. 2a,b) (Fig. 2a,b). Specifically, we estimate the 202 total number of newly unemployed workers in each county 203 using the size of the labour force and the increase in the un-204 employment rate in 2020 compared to 2019. We then use 205 the all-cause mortality rate from 2019 of each county to cal-206 culate the mortality rate of the newly unemployed workers. 207 Finally, we perform linear regression of the projected death 208 rate associated with unemployment on the mobility change. 209 As shown in Fig. 2a,b, the unemployment shock affects poor 210 counties more than the rich ones. One explanation is that the 211 reduction in mobility was significantly more in wealthier areas 212 than poorer areas during the pandemic (9), which indicates 213 that the affluent can weather the economic repercussions of 214 lockdowns partially because their jobs allow for flexibility in 215 terms of working remotely. Prior work has drawn similar 216 conclusions that excess mortality is disproportionately high in 217 disadvantaged groups such as African Americans and people 218 with low educational attainment (26, 51). 219 Fig. 2a,b suggest that the widely believed health and eco-220 nomic trade-offs of lockdowns only exist in poor counties. 221 Fig. 2c illustrates this trade-off by summing regression es-222 timates of COVID-19 deaths and projected excess deaths 223 attributable to unemployment. Our findings confirm marked 224 differences in the way that social distancing and lockdown 225 measures impact different groups. 226 We also explore the association between household over-227 crowding and regional health outcomes. Household overcrowd-228 ing is the condition where there is more than one person per 229 room (52), which may accelerate the spread of respiratory 230 diseases such as COVID-19. We use the Comprehensive Hous-231 ing Affordability Strategy (CHAS) data prepared by the US 232 Census Bureau for the 2013-2017 period (52). We focus on 233 the largest four states for the number of urban counties, which 234 are California, Florida, New Jersey, and New York. A county 235 is urban if at least 95% of the population live in urban areas. 236 The rurality data is published by the US Census Bureau for 237 the year 2010 (53). We restrict the death data (44) to the end 238 of July 2020 to take into account roughly the first six months 239 since the first recorded US case. The qualitative results re-240   Our findings imply an underlying mechanism at play that 248 causes worse health and economic outcomes for poorer commu-   population are vulnerable to severe illness, exclusive to the 285 poor community. Once infected, vulnerable people are more 286 likely to experience worsening symptoms than an average 287 person. 288 We incorporate in our model random graphs to simulate 289 virus transmission and economic activities. In consideration of 290 the high transmission rate in households (55, 60), we construct 291 a collection of complete graphs to represent households where 292 any pair of members in the same household are connected. 293 To capture socioeconomic disparities, we assume that 90% of 294 the population are poor and the rest are rich in expectation. 295 A rich person is characterized by a high output and a small 296 household size. In addition, we overlay the household network 297 with an economic network that represent economic activities 298 which rely on in-person contact (Extended Data Fig. 3b). We 299 generate economic networks using the Watts-Strogatz random 300 graph (61), a classic model that produces the small-world 301 phenomenon as observed in many real-world networks.

302
Our model considers dynamics at both household and ag-303 gregate levels, which include deaths of despair, recession, and 304 undertreatment. We take into consideration deaths of despair 305 that are linked to financial stressors. Specifically, the proba-306 bility that an individual dies from despair is a function that 307 decreases with per capita output in the household. At the 308 aggregate level, with government subsidies taken into account, 309 a drop in the total output leads to more workers becoming 310 economically inactive. In addition, our model incorporates the 311 scenario in which hospitals are overwhelmed and poor patients 312 are undertreated. Undertreatment increases the chance of 313 deterioration in patients.

314
Impacts of NPIs on Inequality. It has been widely accepted by 315 now that there is a trade-off between saving lives from the 316 pandemic and saving lives from recession. What has been less 317 scrutinized, however, is how this trade-off varies in different 318 communities and under various policies (24-26). As we have 319 observed in US data, poorer counties not only have had more 320 COVID-19 deaths but also will see more recession-induced 321 deaths. We investigate the effects of four NPIs on inequal-322 ity, which are lockdown, testing along with contact tracing, 323 government subsidisation, and housing provision. Our model 324 suggests that, for most NPIs considered, the poor community 325 suffers significantly more than the rich counterpart in terms 326 of both types of death. poor community, which is consistent with our conclusion from 347 US data (Fig. 2) ing among susceptible, asymptomatic, and presymptomatic 358 individuals. Once someone tests positive, the person will 359 self-isolate at home until recovery. The person's household 360 members and other contacts will subsequently be prioritized 361 in testing, with the latter being found by contact tracing with 362 a probability of 0.7. Fig. 4b suggests that, even without any 363 other NPI, the combination of testing, contact tracing, and 364 home isolation alone is effective at reducing disparities in both 365 types of death. Our findings corroborate the conclusion in (64) 366 that increased testing and contact tracing capacity enables 367 reopening at a larger scale. 368 We consider government subsidies that are given to anyone 369 in need regardless of socioeconomic status. On each day of 370 simulation, the model looks for and gives money to low-output 371 people who may die from despair. The subsidy is measured as 372 a fraction of an economically active poor individual's personal 373 output. Fig. 4c indicates that need-based subsidies no less 374 than 0.3 effectively eliminate the gap in deaths of despair. We 375 also explore the efficacy of greedy subsidisation subject to 376 budget constraints. Specifically, given a budget, individuals 377 with the lowest outputs are the ones that are most likely to 378 be impacted by economic volatility and hence prioritized for 379 payment. The budget level is measured as the fraction of the 380 population that can be supported if each subsidy is 0.3. Fig. 4d 381 suggests that increasing the budget level reduces disparities 382 in the total death rate and enables stricter lockdown before 383 economic consequences exceed marginal health benefits. 384 We investigate the effects of household overcrowding by 385 varying the maximum size of poor households. The configura-386 tions of rich households are kept at a maximum size of three and 10% of the population. For ease of mathematical analysis, 388 lockdown starts at initialization, and simulation runs for 60 389 days. As shown in Fig. 4e, a larger difference in household 390 size leads to higher inequality in COVID-19 deaths. This re-391 sult confirms the causal link between household overcrowding 392 and the COVID-19 death rate suggested by US data (Fig. 3). 393 Inspired by (16), we quantify the dependence of the infection 394 rate on household size using mean-field approximation. We

442
There are several important implications from this work.

443
Our results underline the importance of intervention design 444 in a pandemic as socioeconomically disadvantaged popula-445 tions bear the brunt of suboptimal policies, which will worsen 446 existing inequalities. Our study has several limitations. First, the conclusions 474 drawn from our analysis rely on aggregate data at the ZCTA 475 and county levels. Ideally, comprehensive data at the individ-476 ual or household level which encompass many aspects such as 477 socioeconomic status, medical conditions, and behaviour in 478 response to COVID-19 are used to infer the differential causal 479 effects of NPIs on different demographic groups. In practice, 480 such granular data rarely exist due to challenges in collec-481 tion and privacy. The limitations of the data are partially 482 addressed by our work on agent-based modelling. Second, 483 our simulations are based on a stylized model that captures 484 key elements to the topic studied, including socioeconomic 485 status, age-dependent risks, and household transmission, but 486 leaves out other details. We have chosen to build a medium-487 sized model in order to obtain qualitative insights. Detailed 488 agent-based models that typically require high-performance 489 computing are needed for drawing quantitative conclusions. 490 Finally, we have only considered lockdowns and testing that 491 are conducted uniformly across the population. In reality, 492 low-income areas across the US have faced obstacles to testing 493 and physical distancing (9, 14, 15). For this reason, the so-494 cioeconomic gap in COVID-19 deaths identified by our model 495 is a conservative estimate.

496
There are several interesting directions for future research. 497 One extension is to investigate interventions that are adjusted 498 over time according to feedback and how such adaptive mea-499 sures affect inequalities. Another interesting avenue of research 500 is exploring how to incentivise safe behaviour that can lessen 501 the need for drastic lockdowns. Given the national variations 502 in the vaccine rollout strategy, it is also urgent to understand 503 how to design vaccine programmes that reduce inequalities. 504 Additionally, it is important to take into consideration fiscal 505 and logistical constraints for the task of policy evaluation.
These questions are not only of much practical relevance to 507 COVID-19 but also fascinating research problems that call 508 for multidisciplinary efforts. Progress towards these goals will 509 have a lasting impact on policy responses to future pandemics.

608
Let I 0 be the number of infections at initialization. Let I p t be the number of newly infected poor individuals at time step t. We suppose that, each day, an infected person who has no symptoms spreads the virus to any of her connections from a different household with probability ρ > 0. Given the high risk of household transmission, we suppose that, once someone is infected, everyone else in the same household is immediately infected. Thus, the effective number of initial infections is npI 0 . We consider that everyone stays at home with probability 1 − α, independently of any other event. In order for an infected person to infect someone from a different household, both persons need to leave home, which occurs with probability α 2 . Under the assumption that each person's economic connections are from different households, every infected individual spreads the disease to α 2 kρd economic connections in one time step on average. Infected connections then immediately infect their household members. Moreover, we suppose that a fraction 0 < < 1 of these new infections further spread the disease to their economic connections and hence their household. Thus, I p 1 is roughly equal to npI 0 Φ(1 + Φ) on average where Φ = npα 2 kρd. Several assumptions such as immediate household transmission and uniqueness of economic connections' households make our estimated number of infections an overestimation, which becomes more marked as time goes on. We adjust for the error by taking into account the susceptible population that shrinks over time.   Fig. 3 | Schematic diagrams of the agent-based model. a, Once infected, an individual progresses stochastically from asymptomatic or presymptomatic, to symptomatic, hospitalized, admitted to the ICU, and deceased, with the possibility of recovery at any stage if not deceased. While staying at home, a susceptible individual may still be infected by people in the same household. Once symptomatic, the infected individual quarantines at home until recovery unless hospitalization becomes necessary. An individual is economically inactive during hospitalization and at death. Moreover, an individual loses connection output while in quarantine or staying home, b, Each blue circle corresponds to a complete graph that represents a household. The economic network is generated using the Watts-Strogatz random graph. Fig. 4 | Epidemiological parameter definitions, baseline values, and sources. Time between different stages of infection is sampled uniformly at random from the corresponding intervals listed.

Baseline value Source
Asymptomatic rate 35% (56) Probability of hospitalization conditional on symptomatic infection NPIs on socioeconomic inequality. Each household comprises members from the same age group. All qualitative observations remain the same as those with multigenerational households (Fig. 4). The fatality rate is calculated within each socioeconomic group. Since the rate of death of despair is close to zero for the rich community, we only show COVID-19 deaths for this group. a, The trade-off between COVID-19 deaths and deaths of despair only exists in the poor community. b, The combination of testing and contact tracing alone is sufficient for eliminating socioeconomic disparities in both types of death. c, Increasing subsidies effectively reduces the gap in deaths of despair.