A Macroeconomic SIR Model for COVID-19
AA MACROECONOMIC SIR MODEL FOR COVID-19
ERHAN BAYRAKTAR, ASAF COHEN, AND APRIL NELLIS
Abstract.
The current COVID-19 pandemic and subsequent lockdowns have highlighted the closeand delicate relationship between a country’s public health and economic health. Macroeconomicmodels which use preexisting epidemic models to calculate the impacts of a disease outbreak aretherefore extremely useful for policymakers seeking to evaluate the best course of action in sucha crisis. We develop an SIR model of the COVID-19 pandemic which explicitly considers herdimmunity, behavior-dependent transmission rates, remote workers, and indirect externalities oflockdown. This model is presented as an exit time control problem where the lockdown endswhen the population achieves herd immunity, either naturally or via a vaccine. A social plannerprescribes separate levels of lockdown for two separate sections of the adult population - thosewho are low-risk (ages 20-64) and those who are high-risk (ages 65 and over). These levels aredetermined via optimization of an objective function which assigns a macroeconomic cost to thelevel of lockdown and the number of deaths. We find that, by ending lockdowns once herd immunityis reached, high-risk individuals are able to leave lockdown significantly before the arrival of a vaccinewithout causing large increases in mortality. Additionally, if we incorporate a behavior-dependenttransmission rate which represents increased personal caution in response to increased infectionlevels, both output loss and total mortality are lowered. Lockdown efficacy is further increased whenthere is less interaction between low- and high-risk individuals, and increased remote work decreasesoutput losses. Overall, our model predicts that a lockdown which ends at the arrival of herdimmunity, combined with individual actions to slow virus transmission, can reduce total mortalityto one-third of the no-lockdown level, while allowing high-risk individuals to leave lockdown wellbefore vaccine arrival. Introduction
The current COVID-19 global pandemic has led to massive lockdowns to slow the spread ofthe virus. Now, policymakers face a dilemma - extended periods of lockdown have put strain onthe economy, but returning to “normal” too quickly could result in an equally troubling secondwave of infections. The task is therefore to find the optimal balance between public health andeconomic growth. Models such as those proposed by Alvarez et al. [AAL20] and Acemoglu etal. [ACWW20] have used a macroeconomic approach and variations on the Susceptible-Infectious-Recovered (SIR) epidemic model proposed by Kermack et al. [KMW27] to solve an optimizationproblem determining the lockdown policy that minimizes both loss of life and effects on output.We consider a variation on these models which incorporates several new concepts and gives a widerpicture of the overall situation. • We formulate an exit time control problem where lockdown measures are lifted when thepopulation reaches herd immunity, even if this occurs before a vaccine is developed. • We incorporate a transmission rate that captures how individuals reacts to current infectionlevels, as discussed in [Coc20]. This “behavior-dependent” transmission rate seeks to modelindividual behaviors that occur independently of lockdown. For example, individuals mightwear masks, practice social distancing, and take other precautions to reduce their risk asinfection numbers go up, even in the absence of official lockdown measures.
Date : July 1, 2020.2010
Mathematics Subject Classification.
Key words and phrases.
Epidemic modeling, COVID-19, planner problem, value iterations, exit time controlproblem. a r X i v : . [ phy s i c s . s o c - ph ] J un E. BAYRAKTAR, A. COHEN, AND A. NELLIS • We consider the costs of indirect deaths attributed to adverse mental and physical effects oflockdown and of continued unemployment after the lockdown has ended, and the positiveimpact of workers who are able to work remotely during lockdown. • We add a penalty for overwhelming intensive care unit (ICU) capacity and a term thatcaptures the future impacts of missed health screenings during the pandemic.The planning problem developed in [AAL20] is the basis of the one used in our model. Thepaper references the SIR method of epidemic modeling, which is also used in subsequent papers,to represent population dynamics. Many of its parameters, like level of obedience of lockdown,are also used in our model. The death rate is calculated as a function of the number of infectedindividuals in order to model the effects of hospital overcrowding and encourage “flattening thecurve”. The objective function quantifies the economic and social impacts of both the pandemicand the resulting lockdown measures, and develops an optimization problem for a planner to solve.The cost of lockdown is represented by the income that is lost by those who are in quarantine andso are unable to work, while the cost of death is calculated as value of statistical life. Their modelexamines the role of the population’s level of obedience, as well as the effect of being able to testthose who are recovered and exclude them from lockdown. They also investigate the results ofdifferent values of statistical life. The authors conclude that being able to test and return recoveredindividuals to the workforce has a large positive effect on outcomes and that these outcomes aresensitive to the fatality rate and its elasticity with respect to infection level.A subsequent paper, [ACWW20], takes this model and extends it by considering the possibilityof different optimal lockdown measure for different groups. In their case (and in ours) the groupsare differentiated by age, since the severity of COVID-19 infection varies widely based on age.The paper also explicitly considers the number of infected individuals admitted to the ICU at eachpoint in time, which is then used to calculate the death rate. The authors use Pareto curves createdby varying the non-pecuniary value of life to show that targeted lockdown measures unilaterallyperform better than uniform lockdowns, regardless of whether one seeks to prioritize reducingoutput loss or reducing mortality. In fact, while they consider three age groups (20–49, 50–64, and65+), their results show that it is sufficient to consider a “semi-targeted” policy which prescribesone lockdown policy to those aged 20–64 and another policy for those over 65 years of age. Due tothis result, we also split the working population into two groups, one aged 20–64 and one aged 65and over.We are also influenced by the work of Cochrane [Coc20], which discusses a Behavioral SIR(BSIR) model wherein individual behaviors affect infection transmission rates. This reflects thetendency of individuals to be more careful as infection levels rise in their community. In this model,the dynamics eventually reach a stable equilibrium with a virus reproduction rate equal to 1 anda nonzero constant rate of infections and deaths per day. To handle this, he considers the role oftechnologies like testing and tracing, which could allow people to reduce transmission levels whilestill maintaining economic activity. We adopt the idea of behavior-dependent transmission basedon infection levels, but instead rely on lockdown and the arrival of herd immunity or a vaccine todrive infection numbers to zero.These works develop a solid framework for our model, but we undertake the task of increasingaccuracy and realism through herd immunity as an exit time, behavior-dependent transmissionrates, deaths indirectly due to lockdown, additional costs of lockdown, and the portion of thepopulation that is able to work remotely. These additions affect the model in various ways, butoverall our augmented model concludes that the high-risk group can be released from lockdownbefore a vaccine arrives without large adverse effects (though of course, this is not to say that avaccine is not necessary or useful). • When the expected vaccine arrival time is 1.5 years after the start of the outbreak, our modelrecommends less than 7 months of lockdown for the high-risk group (instead of locking down
MACROECONOMIC SIR MODEL FOR COVID-19 3 for the full 1.5 years until the vaccine). Additionally, lockdowns for the low-risk group are6 weeks shorter. • The addition of a behavior-dependent virus transmission rate contributes to these shorterlockdowns and decreases mortality. In an extreme situation where individuals can takemeasures that decrease transmission by 95% when infections reach 30%, less than a monthof lockdown is prescribed for the low-risk group. In the more moderate benchmark case,where individuals are able to reduce their transmission by 25% when infections reach 30%of the population, herd immunity arrives a month earlier than in a situation with a constantdisease transmission rate. In both cases, we also observe lower output loss due to shorterlockdown and fewer deaths due to slower transmission. • Increasing the level of remote work reduces the impact of COVID-19 by decreasing bothmortality and output loss, even though a longer lockdown is imposed. This supports theintuitive idea that increased remote work reduces infection risk without sacrificing economicactivity. • Increasing the predicted length of future unemployment and the predicted rate of lockdown-related deaths both decrease lockdown length in a similar manner, but also have negativeimpacts on outcomes. Adjusting the length of future unemployment and the predictednumber of indirect deaths due to lockdown lead to trade-offs between output and mortality.Running the model with different initial conditions shows that higher pre-lockdown infectionlevels lead to earlier onset of herd immunity but higher death tolls, highlighting the risksof infection spikes. Future impacts of current missed health screenings and a penalty foroverfull ICUs are revealed to have little impact on the optimal lockdown policy, at least inour formulation.The main body of the paper presents our model and its numerical results. In Section 2, welay out the SIR dynamics used to model the transmission of the virus and discuss certain modeladditions, especially the addition of deaths indirectly caused by lockdown and a behavior-dependingtransmission rate. In Section 3, we introduce the exit time control problem that ends when thepopulation reaches herd immunity and discuss the terms in the objective function. In Section 4, wediscuss our numerical model, which discretizes the problem and is solved through value iterations.We calibrate it with the results of [ACWW20] and [AAL20] and compare these results to ouraugmented model using death rates on the same scale. Then, we update death rates to match morerecent data from [CDC20] and adjust the non-pecuniary value of life. We present and discuss ourresults and perform some parameter robustness analysis. These experiments serve to illustrate thegeneral mechanisms of the model and to present planners with an idea of our model’s potential. Ifa planner wishes to use our model, parameter values can be changed in our code, found at , toaccurately reflect a specific planner’s current situation.2. Population Dynamics
As in [ACWW20], we consider policies which assign different lockdown strategies to populationgroups with different responses to infection and lockdown. Influenced by their results, we divideadults into one group aged 20-64, called“low-risk” and indexed by j = 1, and one group aged65 and over, called “high-risk” and indexed by j = 2. We will only consider adults older than20, so the low-risk group makes up 82% of the population of interest, while the high-risk groupmakes up 18% [HM10]. The second group can also include individuals of any age who are morelikely to contract severe cases of COVID-19 and experience complications due to immunodeficiency,respiratory weakness, or other preexisting conditions. These individuals are considered separatelyfrom the general working population. We denote the population of group j as a proportion, N j , of https://github.com/april-nellis/COVID19-BSIR E. BAYRAKTAR, A. COHEN, AND A. NELLIS the total population and lay out the following relationship: S j ( t ) + I j ( t ) + R j ( t ) + D j ( t ) = N j , j ∈ { , } , where (cid:88) j N j = 1 . Individuals move from susceptible ( S j ( t )) to infected ( I j ( t )) to recovered ( R j ( t )), with addi-tional flows from all three to death ( D j ( t )). The dynamics are given by˙ S j ( t ) = − ˙ I j ( t ) − γI j ( t ) − ξ ( L j ( t )) S j ( t ) , ˙ I j ( t ) = S j ( t )(1 − θL j ( t )) (cid:88) k β kj ( t )(1 − θL k ( t )) I k ( t ) − γI j ( t ) , ˙ R j ( t ) = γI j ( t ) − φ ( I j ( t )) I j ( t ) − ξ ( L j ( t )) R j ( t ) , ˙ D j ( t ) = φ ( I j ( t )) I j ( t ) + ξ ( L j ( t ))( S j ( t ) + R j ( t )) = − ˙ N j ( t ) . (2.1)The number of new infections depends on the size of the susceptible and infected populations,as well as the lockdown levels, L j ( t ), and transmission rates between groups, β ij . As others havesuggested ([ACWW20], [AAL20]), a portion of the population will disregard lockdown orders. Thislevel of obedience is represented by θ . As in [ACWW20], we set θ = 0 .
75, though this parameter isdifficult to quantify exactly. Patients move out of the infected category with rate γ in accordancewith the expected recovery time of 18 days. Deaths due to COVID-19 occur at rate φ ( I j ( t )) andother deaths occur at rate ξ ( L j ( t )). For convenience, the parameters that appear in (2.1) and inthe objective function (3.1), along with their levels, are listed in Table 1 of the appendix.2.1. Deaths.
One unique element of COVID-19 is its significantly different death rates for differentgroups. Therefore, the base death rate δ j for each group is set individually. In addition, as thenumber of infected individuals increases and hospitals become more crowded, death rates increaseas a function of the total number of infected patients as in [AAL20]. We represent this increase inthe death rate as δ j and follow [ACWW20] in assuming that an infection level of 30% increases thedeath rate by a factor of five. Therefore, the death rate due to viral infection is φ ( I j ( t )) = δ j + δ j (cid:88) j I j ( t ) . Additionally, as lockdowns stretch on concerns have been raised regarding “deaths of despair”due to the impacts of lockdowns on mental health ([ZWRW20], [EMS20]). Hospitals have also shutdown many departments to accommodate the increased need for ICU units for COVID-19 patients.Many non-elective surgeries and routine health checks have also been cancelled or rescheduled([SHM + ξ ( L j ( t )), which is written as ξ ( L j ( t )) = α L L j ( t ) . This represents the number of deaths indirectly caused by the lockdown, and scales with L j ( t ).We argue that, in the absence of any way to verify immunity, indirect deaths occur in both thesusceptible and recovered populations. So, the total number of deaths is given by (cid:88) j φ ( I j ( t )) I j ( t ) + ξ ( L j ( t ))( S j ( t ) + R j ( t )) . Behavior-dependent disease transmission.
The basic transmission rate of COVID-19 isapproximately 0.2 ([AAL20], [ACWW20]). This means that about 20% of those who come incontact with an infected individual will become infected themselves. However, we incorporate atransmission rate that decreases as infections increase due to increased caution between people, asdiscussed in [Coc20]. In addition, we consider an inter-group interaction factor, ρ , as in [ACWW20]. MACROECONOMIC SIR MODEL FOR COVID-19 5
This reflects a lower rate of interactions between groups. For example, working people aged 20-64will interact more with their peers than with those in the high-risk group. Combining these twoideas, we represent the transmission rate as β kj ( t ) = (cid:40) ρβ e − α I I ( t ) if k (cid:54) = j,β e − α I I ( t ) if k = j. The scale factor α I is chosen such that the rate of transmission can be decreased by a factor of e − . α I when 30% of the population is infected with the virus.3. Objective Function
The questions of how long and how severely to lock down the population during a pandemiccan be thought of as a planning problem. Given the above dynamics, we model the optimizationproblem that must be solved by a social planner using the following objective function, whichrepresents the overall societal costs of a given lockdown policy:min L ∈ Λ (cid:90) σ e − ( r + ν ) t (cid:16) (cid:88) j (cid:104) ω j L j ( t ) (cid:0) S j ( t ) + I j ( t ) + pR j ( t ) (cid:1) (1 − h )+ ( χ + ω j r (1 − e − r ∆ j )) φ j ( I ( t )) I j ( t ) + ω j r (1 − e − r ∆ j ) ξ ( L j )( F + S j ( t ) + pR j ( t ))+ α E ω j L j ( t ) (cid:0) S j ( t ) + I j ( t ) + pR j ( t ) (cid:1)(cid:105) + Ω( t ) (cid:17) dt. (3.1)3.1. Attainable Lockdown Levels.
There are certain jobs that must be done even during apandemic, preventing the population from attaining full lockdown. These essential professionsinclude healthcare workers, grocery store employees, delivery workers, and the postal service, amongothers. Because of this, we set an upper limit on the possible lockdown level, denoted ¯ L j , and theset of possible lockdown policies is written as Λ = [0 , ¯ L ] × [0 , ¯ L ]. We set ¯ L at 0.7 to account foressential workers in the low-risk group. On the other hand, ¯ L is set at 1 since we assume that thehigh-risk group does not work.3.2. Herd Immunity.
Previous models considered either an infinite time horizon with stochasticvaccine arrival or a fixed horizon with deterministic vaccine arrival when formulating their objectivefunction. We contribute a new approach that sets reaching herd immunity as the end of the problem.This can be reached either naturally via infection spread and recovery (which confers immunity)or via the arrival of a vaccine. We assume that those who have been infected with COVID-19once will remain immune for the rest of the outbreak. Additionally, we assume that if a vaccineis approved for distribution, vaccination levels will be high enough to produce herd immunity. Wedefine σ as the time at which herd immunity is reached. We assume a herd immunity threshold of60% recovered , so we set σ = σ (60) = inf { t ≥ (cid:80) j R j ( t ) ≥ . } .3.3. Output Loss.
The most noticeable result of lockdown measures is economic slowdown. Manyworkers who are not deemed essential and cannot work remotely have found themselves jobless ascompanies lose revenue. As in [ACWW20], we take the average wage of a full-time worker andnormalize it to 1 and assume that on average, those in the high risk group do not earn any wages.We do not assume the existence of an “immunity passport” given to those who have recovered andare immune, so we set a flag parameter p = 1 (this can be set to 0 to be consistent with the cases of In [CDC20] Table 1 (Scenario 5: Current Best Estimates), the basic reproduction number of COVID-19 is R = 2 .
5. Herd immunity is calculated as 1 − /R = 0 . E. BAYRAKTAR, A. COHEN, AND A. NELLIS [AAL20] and [ACWW20]). On the other hand, we do consider some proportion h of the workforcewho are able to work from home. Therefore, we denote the purely salary-based cost of lockdown as ω j L j ( t )( S j ( t ) + I j ( t ) + R j ( t ))(1 − h ) . When presenting our numerical results, we refer to the output loss due to lockdown. This is notthe value of the objective function presented in (3.1), but rather the losses in output caused byrequiring people to stay home and not work. The output loss is represented by (cid:90) σ e − ( r + ν ) t (cid:88) j ω j L j ( t ) (cid:0) S j ( t ) + I j ( t ) + pR j ( t ) (cid:1) (1 − h ) dt and is compared to annual “normal” output. This baseline output is calculated as the amount ofoutput produced until the expected vaccine arrival time, 1 /nu , if there is no lockdown, annualizedusing the expected vaccine arrival time. This is given by ν (cid:90) ∞ e − ( r + ν ) t (cid:88) j w j N j dt = νr + ν (cid:88) j w j N j . Cost of Death.
We calculate the cost of a COVID-19 death in group j in the same manneras in [ACWW20]. Here, χ is the non-pecuniary cost of life, which we consider as a measure of the public impact of deaths due to COVID-19. This can be thought of as a measure of the planner’spriorities. Lower values of χ lead to prioritizing output loss minimization, while higher values arechosen to encourage longer lockdowns and decrease mortality at the expense of output. To ensurethat this cost is on the same order of magnitude as wages, we scale by the interest rate when wechoosing χ , similarly to [AAL20]. Note that χ = 0 . /r is consistent with [ACWW20] where χ = 20and r = 0 .
01. ∆ j is the number of years left in an average individual’s career. We set ∆ = 20 and∆ = 0. Therefore, the cost per death due to COVID-19 is given by χ + ω j r (1 − e − ∆ j r ) . Deaths indirectly caused by the lockdown are not explicitly categorized and counted, and socan be considered “invisible deaths”. For this reason, we do not include χ in the cost of thesedeaths and only count lost productivity. We also account for similar future deaths due to lack ofpreventative healthcare using a constant F (the number of indirect deaths in the future relative tothose that occur during lockdown). These deaths do not appear in the dynamics, as they have notyet occurred, but they are considered when calculating the costs of lockdown. For this reason, F appears in (3.1) but not in (2.1), and the total cost of indirect deaths is given by ω j r (1 − e − r ∆ j ) ξ ( L j )( F + R j ( t ) + S j ( t )) . Future Loss of Employment.
Another addition to the model acknowledges the long-lastingeconomic impacts of a period of economic slowdown. Already some large corporations like JCPenneyand Hertz have filed for bankruptcy ([Mon20]), and there are certainly more companies, bothlarge corporations and small business, who are under financial strain. Federal stimulus measuresmay alleviate some of this burden, but they cannot completely compensate for current drops inconsumption. Effects may manifest in a variety of ways, but we choose to express them as a “futureloss in employment”, in which 1 day in lockdown results in some α E days of lost employment (onaverage) after lockdown ends. We set this to be 0.42 (reflects current 14.7% unemployment [BLS20]and average 3 days of unemployment for one day of lockdown based on median unemploymentduration of 25.2 weeks (6 months) in 2010 [Cun18]). The cost of future unemployment is modeledby α E ω j L j ( t ) (cid:0) S j ( t ) + I j ( t ) + pR j ( t ) (cid:1) . MACROECONOMIC SIR MODEL FOR COVID-19 7
ICU Overcapacity.
A major incentive for lockdown measures is “flattening the curve” –slowing the spread of the virus so that hospitals and ICUs will not get overwhelmed by a flood ofpatients in need of ventilators and other specialized medical equipment. This is already reflectedin the death rate which increases as infections increase, but we add an additional penalty on topof that. We assume that a fixed proportion of infected patients, ι j , require ICU care. We set thislevel to be 2.6% for people without underlying conditions and 7.4% for high-risk groups [CDC20].Then, we incorporate a penalty η (representing a daily penalty scaled by the level of overcapacity)for hospitalizations exceeding the estimated average ICU capacity, which is 30 beds per 100,000people [PW12]. This is done via the functionΩ( t ) := η (cid:104) (cid:88) j ι j I j ( t ) − ICU (cid:105) + × max j ω j . Numerical Results
Numerical Method.
We used value iterations, first introduced in [Bel57], to solve the op-timization problem presented by our model. The model is discretized using first-order Taylorapproximations and the value function is calculated over a regular grid. Because the change inpopulation due to deaths is very small, we follow the precedent set in [AAL20] and iterate over afour-dimensional ( S , S , I , I ) grid to determine optimal lockdown policy instead of the larger andmore computationally expensive (but more accurate) six-dimensional grid ( S , S , I , I , R , R ).By this, we mean that instead of separately keeping track of the recovered and dead populations,they are considered together as one unit when determining the optimal lockdown policy. Sincethe vast majority of this “non-susceptible” group is recovered, this simplification, which removestwo state variables, has a small effect on accuracy but a large effect on computational complexity.And, when determining the pandemic trajectory for given initial conditions and a given lockdownpolicy, total deaths can still be calculated via the population dynamics shown in (2.1). We choose∆ S = 0 . I = 0 . Calibration.
To test the validity of our numerical models, we use the parameter values of[AAL20] and [ACWW20] and compare our model’s recommendations to their results. A full listof the parameter values used in this section is presented in Table 1. In Figure 1, we compare aone-group version of our model with the one presented in [AAL20], whose recommended optimallockdown reaches 70% lockdown after about one month, and then slows reduces in intensity untillockdown is lifted approximately 140 days (4.5 months) after the outbreak begins. Our version ofthis model maintains the maximum lockdown of 70% for slightly longer and ends slightly later.Interestingly, the ending of lockdown nearly coincides with the population reaching herd immunity,though we did not add any such considerations when running this example. In Figure 2, we setup our model to mimic the semi-targeted policy from [ACWW20] and find similar levels of outputloss, total deaths, and general lockdown recommendations. Namely, the optimal strategy keeps thehigh-risk group in lockdown until the arrival of a vaccine, while the low-risk group is able to emergeand return to work after approximately 200 days of lockdown have elapsed. In this figure, notethat the population reaches herd immunity well before the arrival of a vaccine, implying that thelockdown on the high-risk group could have been ended earlier.Now, we investigate the results of our new model using comparable parameter levels. We incor-porate herd immunity, deaths indirectly due to lockdown, ability to work remotely, and behavior-dependent transmission rates. Additionally, we consider the possibility of lost employment afterthe end of the pandemic, as well as the costs of missed health screenings and a monetary penaltyfor exceeding ICU capacity. To allow comparisons with previous works, we use death rates of a
E. BAYRAKTAR, A. COHEN, AND A. NELLIS similar magnitude to those in [AAL20] and [ACWW20], but we change some parameters to betterfit the current situation. Interest rates have dropped significantly, so we use a 0.001% interest rate,instead of the 5% used by [AAL20] or the 1% used by [ACWW20]. Note that since the interestrate is extremely low, there is little to no discounting applied to wages. We also lengthen the pro-jected average career length in the low-risk group. Finally, we adjust the population distributionslightly from 21% high-risk to 18% high-risk, based on data from the 2010 United States Census[HM10]. The results of our model using the parameters listed in Table 1 is shown in Figure 3.Most noticeably, lockdown rates for both groups fall to 0 after the entire population reaches herdimmunity, which is explicitly imposed by our model. Additionally, the lockdown for the low-riskgroup is slightly shorter but more intense. Unsurprisingly, incorporating deaths of despair increasesthe total number of deaths due to the epidemic, but this effect is kept small by the shorter lock-downs. Long-term costs of lockdown (an additional penalty for ICU overcrowding, future deathsdue to current health negligence, and future unemployment beyond the lockdown) increase outputloss while not directly contributing to deaths during lockdown. However, these output losses areoffset by the proportion of the population that is able to work remotely from home and the shorterlockdown periods.4.3.
Realistic Death Rates.
Recent CDC reports [CDC20] indicate that the death rates are muchlower than those used in Subsection 4.2. To increase the realism of our model, we update the modeldeath rates according to this newer data and use them for all subsequent results. These deathrates are listed in Table 2 and the result, shown in Figure 4, predicts total mortality of 0.4464%and total output loss of 0.0013%. The negligible output loss is due to the negligible lockdown forthe low-risk group. However, lockdowns have already been imposed for both groups (and indeedwe might desire a death rate lower than 0.4464%), so we increase the non-pecuniary value of life, χ , and observe how the model changes. By increasing χ from 0 . /r to 10 /r , Figure 5 shows thatboth groups experience levels of lockdown similar to that of Figure 3, but with an output loss of4.8984% and a lower total death toll of only 0.3266%. We designate this the benchmark situation,which uses death rates from Table 2 and χ = 10 /r but keeps all other parameter values consistentwith those in Table 1. We also compare the results of the optimal lockdown to those generatedby an uncontrolled scenario with the same parameters, shown in Figure 6 and Table 4. Withoutlockdown there is no output loss, but final mortality numbers are approximately twice as high.4.4. Varying Initial Conditions.
Since we are currently in the middle of the pandemic, weinvestigate at how different initial conditions change the recommended lockdown levels. We modela situation where the pandemic is ongoing and lockdown measures have been lifted, but a suddenspike in infections occurs which prompts new lockdown measures. We consider a case where 20%of the population has recovered and 0.2% has died, similar to estimates of the current situation inNew York City [Sta20]. In Figure 7a, a small infection spike affects 5% of the population beforelockdown measure are put in place. In this case, we see additional deaths of 0.2452%. In Figure 7b,a large infection spike affects 25% of the population, causing 0.3346% additional deaths. Note thatthe lockdown is actually shorter for larger infection spikes, since the larger infection level (whichoccurs before lockdowns are imposed) moves the population closer to herd immunity. The price ofthis shorter lockdown, though, is higher mortality rates.4.5.
Parameter Robustness.
It is natural to ask how changes in other parameters affect theoptimal controls. In general, changes in parameters create the expected changes in lockdown We use the data in [CDC20] Table 1 (Scenario 5: Best Current Estimates). To calculate δ , we construct aweighted average of the Symptomatic Case Fatality Ratio for 0-49 year olds and for 50-64 year olds using 2010Census data [HM10] and multiply by 0.65, since the CDC estimates that 35% of cases are asymptomatic. For thesame reason, we multiply the Symptomatic Case Fatality Ratio for the 65+ group by 0.65 to determine δ . We set δ j such that a 30% infection level causes a five-fold increase in deaths, as in [ACWW20]. MACROECONOMIC SIR MODEL FOR COVID-19 9 length and intensity, output loss, and mortality. The more interesting question asks about the level of impact of various parameters. The effects of non-pecuniary value of life ( χ ) have already beendiscussed and are displayed in Figures 4 and 5 and Table 3, but now we discuss other parameters.In Table 5, we list the lockdown levels, mortality, and output loss for various other configurationsof parameter choices.There are some elements of the model which do not have large impacts on the results. TheICU overcapacity constraint barely affects results, likely due to a combination of low infection ratesand a sufficiently high number of ICU beds (on average) in the United States. This allows ICUadmittance rates to remain at or below the threshold. The expected vaccine arrival date also doesnot have much effect on the optimal lockdown levels, since the population is expected to reachherd immunity well before its introduction. Based on the current situation, it seems extremelyunlikely that a vaccine will be developed less than 8 months after the start of the outbreak, so weconsidered expected vaccine arrival times of 1 year ( ν = 1) and 8 months ( ν = 1 .
5) and found thatneither adjustment had much effect. It can also be seen from the table that removing F , the costrepresenting future deaths due to current lack of health maintenance, lengthens the lockdown onlyslightly, and has little effect on output loss and mortality.In contrast, loss of future employment, ability to work remotely, indirect death rate in lockdown,inter-group interaction, and behavior-dependent infection transmission have significant effects onlockdown length and severity. Adding output savings from remote work, future employment loss dueto lockdown, and indirect deaths of lockdown affects the wider population and so produce similarchanges in outcomes. The model produces uniformly better outcomes when the level of remote work, h , is increased, though lockdowns do last longer as seen in Figure 8. Intuitively, this follows fromthe idea that more people working remotely helps to maintain economic activity without increasingrisk of infection. Changing α I and α L in the opposite direction of remote work creates similareffects on the optimal lockdown policy, though outcomes move differently. Increasing the lengthof projected future unemployment, α E , leads to a shorter lockdown and less output loss, whiledeaths increase. This suggests that α E influences the trade-off between output loss and mortality.And, when we look at varying values of α L , we can see what happens when the optimization triesto minimize deaths due to COVID-19 while also trying to avoid deaths due to lockdown. When α L = 0 and the model doesn’t take indirect deaths into account, the lockdown extends for longerand has a larger effect on output, but we see much lower mortality levels. However, when α L isincreased to 50 deaths per 100,000 individuals at full lockdown, the model dramatically shortensthe lockdown, which decreases output loss and indirect deaths but which leads to higher deathsdue to COVID-19.We now discuss the effects of behavior-related parameters. Interestingly, these are the onlyparameters we discuss which have also an effect on the uncontrolled outcomes. If the level ofinteraction between groups, ρ , is lowered, there is less interaction between the low-risk and high-risk groups and so lockdowns are more effective. In Figure 10, when ρ goes from 0.75 to 0.5,the low-risk group is able to begin easing the lockdown earlier since there is less worry abouttransmission to high-risk individuals. However, the lockdown lasts longer overall, since it is harderto reach herd immunity. This increase in output is offset by a drop in mortality. With the optimallockdown, mortality decreases to 0.2552% compared to the benchmark of 0.3166% when ρ = 0 . ρ isincreased to 1, meaning that the groups mix freely. Herd immunity arrives sooner due to increasedinter-group transmission, however mortality increases to 0.3599% with lockdown and to 0.6891%without it. This suggests that it is beneficial for high-risk individuals to exercise extra caution intheir interactions with members of the low-risk group. The other parameter that reflects individualbehavior, α I , also has a notable impact on optimal lockdown policies. This parameter determinesthe efficacy of personal actions taken to slow transmission of COVID-19. In the benchmark case, α I = 1. If transmission rates are constant ( α I = 0), then the lockdown lasts longer due to the increased likelihood of transmission and mortality increases to 0.3394%. In the uncontrolled case,the mortality increases too, to 0.7586%. But, if α I is set very high (say 10, which implies thatpersonal caution can reduce the transmission rate by 95% when infections reach 30%), then alockdown is barely necessary, as shown in Figure 9. In this case, uncontrolled mortality is a mere0.2581%. This second scenario is perhaps too optimistic, but it demonstrates the potential powerof social distancing. If we examine α I = 6 and α I = 8, we see that the lockdowns increase as α I decreases. If we consider the more modest change from α I = 0 to α I = 1 and decrease theinteraction level between groups from ρ = 0 .
75 to ρ = 0 .
5, then imposing the optimal lockdowndecreases overall mortality from 0.7586% to 0.2552%.If the population is less obedient and disregards lockdown measures, for example if θ decreasesfrom 0.75 to 0.6, we see that the lockdown is shorter because herd immunity is reached sooner.This decreases output loss, but leads to more deaths. If the population is more obedient, however,for example if θ = 0 .
85, the effect of a given level of lockdown is larger with respect to the samelevel of output loss so lockdowns are less intense but last longer. This slows the arrival of herdimmunity, leading to higher output loss, but has the benefit of lower mortality rates.Finally, in Figure 11 we investigate the effect of increasing the herd immunity threshold σ ( x ) =min { t ≥ R t ≥ x } . In all of our examples, the low-risk group is able to leave lockdown beforethe arrival of herd immunity, so moving this threshold does not have much impact on outputloss. However, this parameter has important implications for the high-risk group. From Figure11, we see that there is a clear change in dynamics for thresholds of 75% and above, where thelockdown policy becomes comparable to that of [ACWW20]. The length of lockdown for the low-risk group increases by about 20 days, but the high-risk group remains in lockdown until thevaccine arrives – an increase of almost 300 days. This abrupt change in strategy arises becauseeventually the population reaches a steady state with very low infections. The number of susceptibleindividuals decreases very slowly and is driven largely by deaths due to lockdown, so will not achievethresholds of 75% and over before the vaccine is expected to arrive. In this case, the pandemichas been neutralized even though the population has not reached the threshold prescribed by theparameters. These experiments demonstrate that there is not much benefit to be found from aplanner overestimating the herd immunity threshold. If the herd immunity threshold is 100%, thisis equivalent to removing the herd immunity exit time. The impact of σ can therefore be observedas an decrease in output loss and an increase in mortality, while decreasing lockdown length for thehigh-risk group by 340 days and for the low-risk group by 45 days.5. Appendix - Tables and Figures
MACROECONOMIC SIR MODEL FOR COVID-19 11
Parameter Description [AAL20] [ACWW20]
Our Model ¯ L Maximum attainable lockdown 0.7 [0.7, 0.7, 1] [0.7, 1] γ Recovery rate 1/18 1/18 1/18 δ j Base mortality 0.01 γ [0 . γ, . γ, . γ ] [0 . γ, . γ ] δ j Rate of mortality increase based oninfection levels 0.05 γ if I = 30% thenmortality rates are 5times the base rates [0 . γ, . γ ] ι j Rate of ICU admittance N/A σ (unknown) [0.026, 0.074] ICU
ICU capacity as proportion ofoverall population (based onbeds/100,000 individuals) N/A N/A 0.0003 η Scale factor for cost of ICU over-capacity N/A N/A 10 β Initial transmission rate 0.2 0.2 0.2 ρ Interaction level between groups N/A 1 0.75 ν Intensity for vaccine/cure arrival 0.667/365(1.5 yrs) 0.667/365 0.667/365 ω j Normalized individual daily pro-ductivity 1 [1, 1, 0] [1,0] h Proportion of workforce which canwork remotely 0 0 0.4 r Yearly interest rate 5% 1% 0.001% χ Non-pecuniary cost of death 0 20 0.2/r∆ j Years left in career ∞ [15, 7.5, 0] [20, 0] θ j Obedience of lockdown 0.5 0.75 0.75 α L Scaling factor for indirect deathsdue to lockdown 0 0 0.00001 α I Scaling factor for decrease in β t due to personal social distancingmeasures (masks, etc.) 0 0 1 α E Scale factor for decreasing po-tential career length/increasingchance of bankruptcy as lockdownlengthens N/A 0 0.01 F Future cost of missing health main-tenance during lockdown 0 0 1 p Flag for immunity passportp=1 = ⇒ no passportp=0 = ⇒ passport 0 0 1 Table 1.
Full List of Parameters
Figure 1.
Recreation of [AAL20] model (no groups or herd immunity), parameters fromTable 1Output Loss: 13.4232%, Total Deaths: 1.1754%
Figure 2.
Recreation of [ACWW20] model (two groups and no herd immunity), parametersfrom Table 1Output Loss: 8.9676%, Total Deaths: 1.3121%
MACROECONOMIC SIR MODEL FOR COVID-19 13
Figure 3.
Our model (two groups, herd immunity), parameters from Table 1Herd Immunity: 211 days, Output Loss: 7.8767%, Total Deaths: 1.8873%
Figure 4.
Our model ( χ = 0 . /r , death rates from Table 2, all others from Table 1)Herd Immunity: 80 days, Output Loss: 0%, Total Deaths: 0.4464% Group δ δ Age 20-64 0 . × γ . × γ Age 65+ 0 . × γ . × γ Table 2.
Death rates based on [CDC20], γ = 1 /
18 is recovery rate as listed in Table 1
Figure 5.
Benchmark – Herd Immunity: 207 days, Output Loss: 7.3439%, Total Deaths:0.3266%Benchmark Parameters: χ = 10 /r , r = 0 . α E = 0 . h = 0.4, α L = 10 − , α I = 1, ρ = 0 .
75, F = 1, θ = 0 . ν = 0 . η = 10, Death rates from Table 2, Herd Immunity =60% Non-pecuniaryValue ofLife OutputLoss TotalDeaths(All) COVID-19Deaths(All) TotalDeaths(20-64) COVID-19Deaths(20-64) TotalDeaths(65+) COVID-19Deaths(65+) χ = 0 . /r
0% 0.4464% 0.4335 %0.1433% 0.1433% 0.3017% 0.2902% χ = 10 /r (Bench-mark) 7.3439% 0.3266% 0.2544% 0.1201% 0.0855% 0.2066% 0.1689% Table 3.
A comparison of mortality rates for different non-pecuniary value of life, deathrates from Table 2
MACROECONOMIC SIR MODEL FOR COVID-19 15
Situation Output Loss Total Deaths
No Lockdown 0% 0.6189%Optimal Lockdown 7.3439% 0.3266%
Table 4.
Comparison of optimal lockdown policy to no lockdown, using benchmark parameters
ParameterValues Avg.Lock-down(20-64) Length(20-64) (days)
Avg.Lock-down(65+) Length(65+) (days)
OutputLoss (%)
TotalDeaths (%)
COVID-19Deaths (%)Benchmark 0.3188 161 0.8819 207 7.3439 0.3266 0.2544 α E = 0 .
21 0.3123 205 0.8878 239 8.9667 0.3265 0.2388 α E = 0 .
84 0.2864 126 0.8663 167 5.2962 0.338 0.2848 α L = 0 0.3083 192 0.8863 227 8.3612 0.2456 0.2456 α L = 5 × − α I = 0 0.3582 214 0.9057 249 10.6699 0.3712 0.2707 α I = 6 0.2173 144 0.7758 219 4.5236 0.2651 0.2106 α I = 8 0.1503 84 0.4962 188 1.9019 0.259 0.2328 α I = 10 0.041 24 0.3852 189 0.1558 0.245 0.2314 h = 0 0.2874 126 0.8631 168 8.8592 0.338 0.2846 h = 0 . ρ = 0 . ρ = 1 0.3316 147 0.891 194 7.0292 0.3657 0.2975 α F = 0 0.3065 175 0.855 211 7.6457 0.3324 0.2588 ν = 1 0.3177 160 0.8803 206 6.8007 0.3264 0.2548 ν = 1 . η = 0 0.3188 161 0.8819 207 7.3439 0.3266 0.2544 η = 10 θ = 0 . θ = 0 .
85 0.2736 266 0.8815 304 9.7995 0.3075 0.2023 σ (0 .
65) 0.2982 185 0.8973 242 7.8084 0.2948 0.2135 σ (0 .
7) 0.2983 186 0.8919 279 7.8496 0.2724 0.1852 σ (0 .
75) 0.2793 202 0.8633 Vaccine 7.967 0.2732 0.1444 σ (0 .
8) 0.2775 204 0.9909 Vaccine 7.9881 0.2834 0.1416 σ (0 .
9) 0.2754 206 1.0 Vaccine 8.0011 0.2832 0.1405 σ (1) 0.2754 206 1.0 Vaccine 8.0011 0.2832 0.1405 Table 5.
Parameter Robustness Results (Note: lockdown for 65+ ends at herd immunity)Benchmark Parameters: χ = 10 /r , r = 0 . α E = 0 . h = 0.4, α L = 10 − , α I = 1, ρ = 0 .
75, F = 1, θ = 0 . ν = 0 . η = 10, Death rates from Table 2, Herd ImmunityThreshold = 60% Figure 6.
Comparison of optimal lockdown policy to no lockdown, using benchmark pa-rameters, Optimal Lockdown Deaths: 0.3266% vs Uncontrolled Deaths: 0.6189%
MACROECONOMIC SIR MODEL FOR COVID-19 17 (a) S = 74 . , I = 5% R = 20% , D = 0 . (b) S = 54 . , I = 25% R = 20% , D = 0 . Figure 7.
Results for varied initial conditions using benchmark parameters (a) h = 0, Lockdown: 168 days, Output Loss = 8.8592%, Total Deaths = 0.338% (b) h = 0 .
60, Lockdown: 235 days, Output Loss = 5.842%, Total Deaths = 0.3251%
Figure 8.
Robustness results for h (percentage of workforce that can work remotely) ρ = 0 . χ = 10 /r , r = 0 . ν = 0 . α L = 10 − , α I = 1, α E = 0 . η = 10, F = 1 MACROECONOMIC SIR MODEL FOR COVID-19 19 (a) α I = 0, Lockdown: 249 days, Output Loss: 10.6699%, Total Deaths:0.3712% (b) α I = 6, Lockdown: 219 days, Output Loss:4.5236%, Total Deaths: 0.2651% (c) α I = 8, Lockdown: 188 days, Output Loss:1.9019%, Total Deaths: 0.259% (d) α I = 10, Lockdown: 189 days, Output Loss = 0.1558%, Total Deaths= 0.245% Figure 9.
Robustness results for α I (scale factor for individual carefulness in response tocurrent levels of infection) ρ = 0 . χ = 10 /r , r = 0 . ν = 0 . α L = 10 − , α E = 0 . η = 10, F = 1, h = 0 . (a) ρ = 0 .
5, Lockdown: 233 days, Output Loss = 8.118%, Total Deaths = 0.2753% (b) ρ = 1, Lockdown: 194 days, Output Loss = 7.0292%, Total Deaths = 0.3657% Figure 10.
Robustness results for ρ (inter-group interaction level) χ = 10 /r , r = 0 . ν = 0 . α L = 10 − , α I = 1, α E = 0 . η = 10, F = 1, h = 0 . MACROECONOMIC SIR MODEL FOR COVID-19 21 (a)
Herd Immunity = 65%, Output Loss: 7.8084%, Total Deaths: 0.2948% (b)
Herd Immunity = 70%, Output Loss = 7.8496%, Total Deaths =0.2724% (c)
Herd Immunity = 75%, Output Loss: 7.967%, Total Deaths: 0.2732%
Figure 11.
Robustness results for σ (arrival of herd immunity) ρ = 0 . χ = 10 /r , r = 0 . ν = 0 . α L = 10 − , α I = 1, α E = 0 . η = 10, F = 1, h = 0 . References [AAL20] Fernando E Alvarez, David Argente, and Francesco Lippi. A simple planning problem for covid-19 lock-down. Working Paper 26981, National Bureau of Economic Research, , April 2020.[ACWW20] Daron Acemoglu, Victor Chernozhukov, Ivan Werning, and Michael D Whinston. Optimal targetedlockdowns in a multi-group SIR model. Working Paper 27102, National Bureau of Economic Research, , May 2020.[Bel57] Richard Bellman. A markovian decision process.
J. Math. Mech. , 6:679–684, 1957.[BLS20] BLS. The employment situation – April 2020. ,2020.[CDC20] CDC. Covid-19 pandemic planning scenarios. , 2020.[Coc20] John Cochrane. An SIR model with behavior. https://johnhcochrane.blogspot.com/2020/05/an-sir-model-with-behavior.html , 2020.[Cun18] Evan Cunningham. Great recession, great recovery-trends from the current population survey.
MonthlyLab. Rev. , 141:1, , 2018.[EMS20] Timon Elmer, Kieran Mepham, and Christoph Stadtfeld. Students under lockdown: Assessing changein students social networks and mental health during the covid-19 crisis, 2020.[HM10] Lindsey Howden and Julie Meyer. Age and sex composition: 2010. , 2010.[KMW27] William Ogilvy Kermack, A. G. McKendrick, and Gilbert Thomas Walker. A contribution to the math-ematical theory of epidemics.
Proceedings of the Royal Society of London Series A , 115(772):700–721,1927.[Mon20] Paul R. La Monica. Beware of bankrupt stocks like jcpenney and hertz. , 2020.[PW12] Meghan Prin and Hannah Wunsch. International comparisons of intensive care: informing outcomes andimproving standards.
Current opinion in critical care , 18(6):700–706, 2012.[SHM +
20] K. Soreide, J. Hallet, J. B. Matthews, A. A. Schnitzbauer, P. D. Line, P. Lai, J. Otero, D. Callegaro, S. G.Warner, N. N. Baxter, C. Teh, J. Ng-Kamstra, J. G. Meara, L. Hagander, and L. Lorenzon. Immediateand long-term impact of the covid-19 pandemic on delivery of surgical services, 2020. Advance onlinepublication.[Sta20] New York State. Amid ongoing covid-19 pandemic, governor cuomo announces re-sults of completed antibody testing study of 15,000 people showing 12.3 per-cent of population has covid-19 antibodies. ,2020.[ZWRW20] Stephen X. Zhang, Yifei Wang, Andreas Rauch, and Feng Wei. Unprecedented disruption of lives andwork: Health, distress and life satisfaction of working adults in china one month into the covid-19outbreak.
Psychiatry Research , 288:112958, 2020. (E. BAYRAKTAR) DEPARTMENT OF MATHEMATICS, UNIVERSITY OF MICHIGAN, ANN AR-BOR, MICHIGAN 48109, UNITED STATES
E-mail address : [email protected] (A. COHEN) DEPARTMENT OF MATHEMATICS, UNIVERSITY OF MICHIGAN, ANN ARBOR,MICHIGAN 48109, UNITED STATES
E-mail address : [email protected] (A. NELLIS) DEPARTMENT OF MATHEMATICS, UNIVERSITY OF MICHIGAN, ANN ARBOR,MICHIGAN 48109, UNITED STATES
E-mail address ::