A simple tool for comparing benefits and ‘costs’ of COVID-19 exit strategies

Abstract

Background

Governments and health policymakers are now looking for strategies to lift the COVID-19 lockdown, while reducing risk to the public.

Methods

We propose the population attributable risk (PAR) as an established epidemiological tool that could support decision-making through quickly estimating the main benefits and costs of various exit strategies.

Results

We demonstrate the feasibility of use of PAR using pandemic data, that were publicly available in mid-May 2020 from Scotland and the US, to estimate the proportion of COVID-19 hospital admissions which might be avoided, and the proportion of adverse labour market effects – for various scenarios – based on maintaining the lockdown for those of certain ages with and without comorbidities.

Conclusion

These calculations could be refined and applied in different countries to inform important COVID-19 policy decisions, using routinely collected data.

Highlights

• •Many countries have tried to exit COVID-19 lockdown. Typically, trade-offs are required.
• •A simple epidemiological tool -- population attributable risk -- can help quantify these trade-offs.
• •This paper quantifies the proportion of COVID-19 hospitalisations averted, versus the proporition of the labour force locked down, for various combinations of age- and comorbidity-related lockdown exit strategies.
• •We show that both continuing lockdown for all persons over 65, and those with comorbidities over 50, have similar effects on hospitalisations.
• •Other considerations, such as equity and acceptability of these policy options, are therefore critical in these final policy decisions.

Introduction

Many countries are now struggling to identify optimal exit strategies from the COVID-19 pandemic lockdown. As with previous pandemics, responding to COVID-19 has been characterised by ‘uncertainty, high potential loss, time pressure and competing values’ – all of which are challenges to adopting an evidence-based policy response.^{1 , 2} Weible et al.³ and Xu and Basu⁴ have retrospectively reviewed how decisions have been made in various countries, with various degrees of success. However, watching and waiting for the optimal approach to be identified in another country is not a viable option and therefore various decision-making approaches have been advocated in the literature. These include principalism,^{5 , 6} risk-based decision-making,⁷ experimentation⁸ and analytic modelling.9, 10, 11 Within each of these approaches, data are important, including understanding the limitations of the available data when ideal data are not available.¹²

As more data about the pandemic are available, a variety of methods are being applied to understanding what has happened and predicting what might happen next. Stedman et al.¹³ used trend analysis to analyse the state of the pandemic in the UK, with the intent of informing future policy actions. In the city of Honghu in Hubei, China, cloud-based systems were used to monitor the pandemic.¹⁴ Tsay et al.¹⁵ applied more traditional infectious disease models compartmentalising: susceptible, exposed, infectious, recovered (SEIR) people in data from the USA, Germany, Italy and Spain. Li et al.¹⁶ extended the SEIR model to model the impact of mass influenza vaccination and public health interventions. The computational power available today means that these types of models can be developed and run quickly. However, as Rhodes and Lancaster¹⁰ discuss the multitude of models can become problematic, complicated and confusing rather than supporting evidence-based decision-making. Subsequently, there is a need for more easily understood and transparent models, especially when they can help decide between competing interests.¹¹

We describe a traditional and remarkably simple epidemiological tool, rarely applied to infectious diseases, which can be used to quickly estimate and compare the potential main benefits and ‘costs’ (i.e. negative consequences) of various exit strategies. That tool is Levin's population attributable risk (PAR)¹⁷ – the proportion of disease burden which is attributable to any given risk factor, such as age or the presence of one or more chronic diseases, in the case of COVID-19.

$$PAR=\frac{p\left(RR-1\right)}{1+p\left(RR-1\right)}$$

where p is the proportion of the population exposed to the risk factor and RR is the relative risk of the outcome related to the risk factor.

We demonstrate how this method can be easily applied to compare two exit strategies:

• A.continuing lockdown for all older adults (based on two age cut-offs: 50+ and 65+);
• B.continuing lockdown only for adults with one or more chronic disease risk factor/condition (CDRF), also stratified by these same age groups, compared with continuing lockdown for all adults aged 20–49 years.

Methods

These calculations are applied to Scotland in May 2020 (when lockdown relaxation measures were just beginning to be considered) using data readily available at that time:

• •the current Scottish population structure;¹⁸
• •recent Scottish Health Survey¹⁹ prevalence estimates for common chronic conditions known to increase the risk of severe COVID-19 (CDRFs): ‘any cardiovascular disease or diabetes’; obesity; asthma or Chronic Obstructive Pulmonary Disease (COPD) (Table 2.3 of the Survey report); and
• •early-pandemic (March 2020) USA age-specific rates of COVID-19 hospitalisation,²⁰ as well as characteristics of hospitalised cases, in terms of both age and CDRF status.²¹

We chose COVID-19–related hospitalisation as our outcome because hospitalisation is both serious for patients and costly to society; there are also fewer threats to the validity of these data than case counts/incidence rates in settings with incomplete and rapidly changing testing regimens –as is typical of many COVID-19–affected countries to date. The US COVID-19 admission rate data were collected from the large COVID-NET hospital network, with statistically stable admission rates through to April 25th. Because only relative risks (RRs) will be used, these absolute population-based admission rates need not be generalisable beyond that US setting; they just need to be based on consistent admitting practices over time, by age group and risk factor status, which is more credible inside a single set of hospitals so early in the pandemic in the USA. It is appreciated that some features of the US pandemic during this period were different from those that pertain to the subsequent few months' situation in the UK/Scotland. However, we believe that the probable continuing low cumulative incidence of COVID-19 infections in the UK (very few antibody-survey estimates have been above 10% of the general national population, through to early May^{22 , 23}) means that options for exiting lockdown at that time in the UK were likely to carry similar RRs, by age and CDRF status, to the remarkably stable age-specific admission rates over time seen in the USA COVID-NET network through to mid-April. This is especially credible because that was during a period when suitable hospital beds in most of that country were not yet filled with cases, and strong lockdown measures were not yet in place (only 32 states had even started lockdown by the end of March).²⁴

Scenario A – age-based restrictions (Table 1)

Table 1.

Calculating the population attributable risk related to age-based restrictions.

Adult age groups (years)	Step 1: calculate RR		Step 2: calculate p	Step 3: calculate PAR
	Absolute hospitalisation rate (per 100,000)4	Relative risk	Percent of Scottish Population in 20202	Population attributable risk
20a–49	22.6	1 (reference)	38.8%	Reference group (lowest adult admission rate)
50–64	69.3	3.1	21.0%	30.6%
65+	142.7	6.3	19.4%	50.6%

PAR, population attributable risk.

^aNote that the COVID-NET tabulation of cumulative hospitalisation rates (per 100,000 population) for the youngest adult age group includes 18- and 19-year-olds, whereas all the other statistics used here include only those aged 20+ years; we have ignored this discrepancy, noting that COVID-19 hospitalisation rates at ages 18 and 19 are trivially small.

The first step in calculating PAR based solely on age is to calculate the age-specific RRs of COVID-19 hospitalisation. These RRs based on the COVID-NET data are shown in Table 1 (step 1) alongside the absolute hospitalisation rates from which they were calculated.²⁰ Then, we tabulate the proportion of the general population (Scotland, 2011 census) in each of the age groups: p (step 2).¹⁸ Finally, we use the figures from the first two steps to calculate the PARs for each of the three age groups’ using the equation given previously (step 3). The resulting PARs are the proportion of hospitalisation of people aged 50–64 years or older than 65 years out of the total number of hospitalisations. Therefore, the PARs give an indication of the potential reduction in COVID-19 hospitalisation from of continuing lockdown for: (i) those aged older than 50; (ii) those aged older than 65. These results can be interpreted as indicating that continuing effective lockdown for only those aged 65+ would theoretically reduce the total adult COVID-19 hospitalisations by 50.6% and affect the quality of life of about one in five of the Scottish population, the vast majority of whom are retired. Continuing lockdown for those aged older than 50, on the other hand, while massively reducing adult hospitalisations by (30.6 + 50.6) = 81.2%, would interfere with the lives of (21.0% + 19.4%) = 40.4% of the entire Scottish population, of whom more than half are younger than 65 years, with potentially significant economic effects.

Scenario B – age- and comorbidity-based restrictions (Table 2)

Table 2.

Calculating the population attributable risk related to age- and condition-based restrictions.

Adult age groups (years)	Step 1 – calculate odds ratios as an approximation of RR			Step 2 – calculate p	Step 3 – calculate PAR	Step 4 – weight the PARs relative to the whole population
	Proportion of admissions4,5		Odds ratio	Percent of Scottish Population in 2011 with any CDRF3	Population attributable risk	Proportion of US adult hospitalisations4,5	Weighted PAR
	Cases with 1+ CDRF	Controls with 0 CDRF
20–49	85%	30%	13.2	30%	78.5%	24.9%	19.7%
50–64	85%	55%	4.6	55%	66.4%	31.3%	20.7%
65+	94%	64%	8.8	64%	83.3%	43.7%	36.3%
All 20+	90%	45%	11.0	45%	81.8%	100.0%	76.7%

PAR, population attributable risk; CDRF, chronic disease risk factor/condition.

In this scenario, we are estimating the population risk attributable to the combination of two risk factors; age and health, represented by the presence of one or more CDRFs. The American Geriatrics Society has explicitly called for comorbidities to be considered in policy decision, avoiding solely age-based criteria.²⁵ Subsequently, the steps are slightly different. First, we estimate the RR of COVID-19 admission for persons with one or more CDRF, compared with the healthy population, within each of the three age groups (Table 2 , step 1). A rapid way to approximate these RRs is to perform a case-control analysis based on Centres for Disease Control and Prevention (CDC) Atlanta's summary of the proportion of a large series of COVID-NET hospitalisations,²¹ in all of March 2020, who have at least one CDRF compared with ‘controls’ of the same age group in the general population (from Scottish Health Survey¹⁹); to reduce confounding by age, use age-stratification. Note the non-linearity of the relationship between RR and age, indicating a strong effect of comorbidity in the youngest age group, and also in the elderly – effectively an interaction effect. Then, we tabulate the proportion (p) of each of each adult Scottish age group who have at least one of the following CDRFs: heart disease; COPD or asthma; diabetes; obesity; hypertension (step 2). Some interpolation is required because the Scottish Health Survey¹⁹ reports prevalence separately for these common (self-reported) chronic conditions. To prevent double-counting of persons with more than one condition, the estimates in Table 2 are totals of: the full age-specific prevalence of ‘any Cardiovascular Disease (CVD)/diabetes’ added to half of each of the age-specific prevalences of the other three conditions. Better estimates can be readily derived from comorbidity studies in primary care. As in scenario A, we then calculate the analogous PARs for each of the three age-specific subpopulations’ members with at least one CDRF (step 3), modelling a policy of continuing lockdown only for that high-risk group of adults, across the three age-strata compared with no restrictions for that subpopulation. The PARs calculated in step 3 are the proportion of admissions in each age group attributable to having one or more CDRF. Therefore, to estimate the overall population impact, we need weight the PARs by the proportion of admissions from each age group. This fourth step is achieved by weighting these PARs across the three age-strata, by the proportion of US COVID-NET adult admissions^{20 , 21} in each age group (cf. step 3), giving the overall PARs shown in Table 2.

This analysis of scenario B tells us that restricting the activities of persons with at least one CDRF, in all three adult age groups, should reduce the overall COVID-19 hospitalisation rate by over three-quarters (compared with no relaxation of any restrictions for any adults), but at a very high ‘cost’ of interfering in the lives of about 30% of 20-to-49-year-olds (a very large group, demographically speaking), 55% of 50-to-64-year-olds and 64% of those aged 65+ (cf. step 4 aforementioned) – with the added concern that the two younger age groups are typically active in the labour market. Alternatively, by restricting the activity of those aged older than 50 years with CDRFs, we could expect to reduce the hospitalisation rate by 57% (20.7% + 36.3%) and interfering in the lives of the same proportions by 55% and 64% of those aged 50–64 and 65+ years, respectively. Because these persons are typically already aware of their CDRF status, their willingness to continue lockdown may be higher than for restrictions based on age alone, to minimise the personal risk based on their medical conditions.

Conclusion

As shown in Table 3 , all five policy options are less than ideal, with only two carrying reasonable benefits, in terms of substantially reduced COVID-19 hospitalisations, without removing large numbers of people from the labour force: policies #2 and #4. There is not much to choose – in terms of epidemiologically estimated reductions in COVID-19 hospitalisations – between restricting the activities of all persons aged older than 65 years compared with restricting all persons aged older than 50 years with CDRFs. However, the economic effects of the former policy would be much less than those of the latter because the latter would affect a significant proportion of the active labour force; advocates for the elderly, on the other hand, are likely to be concerned about the ‘discriminatory nature’ of purely age-based restrictions.²⁵ Policy options #1 and #3 would prevent a substantially larger proportion of future COVID-19 admissions – but only by continuing to lockdown much larger numbers of adults – almost half the entire adult population in the case of policy option #3, including many younger and middle-aged adults in the active labour force.

Table 3.

Summary of the potential benefits (reduced COVID-19 hospitalisations) compared with the proportion of the population required to maintain lockdown for the policy options examined.

Policy scenario	Proportion of adult hospitalisations reduced	Proportion of adult population affected
AAge-based
1. Restrict all persons aged older than 50 years	81%	40%
2. Restrict only persons aged older than 65 years	51%	19%
BAge- and comorbidity-based
3. Restrict adults of all ages with CDRFs	77%	45%
4. Restrict all those with CDRFs aged older than 50 years	57%	21%a
5. Restrict all those with CRDFs aged older than 65 years	36%	11%a

CDRF, chronic disease risk factor/condition.

^aThese are the age group–specific prevalences of one or more CDRFs in the most recent Scottish Health Survey, calculated as population-weight weighted averages of the prevalences in narrower age bands.

We recognise that there are multiple other factors that governments need to consider, when assessing the options for easing the lockdown, including indirect effects on transmission dynamics, and the varying likelihoods of being able to work from home in these different subpopulations. Yet, using only publicly available data, it is possible through calculating PAR to gain an insight into the trade-off between protecting the public and maintaining the economy. Furthermore, compared with the mathematical models being used to model the pandemic, the arithmetic necessary to calculate PAR can be quickly carried out using any computer or calculator. Therefore, we believe that PAR is a relatively simple and transparent tool that can be used to provide useful data to quickly and easily compare the potential benefits and crude societal ‘costs’ (adverse consequences) of various exit policy options from the COVID-19 lockdown.

Author statements

Ethical approval

None sought.

Funding

None declared.

Competing interests

None declared.

Contributors and sources

JF originally conceptualised the paper, and made the initial rough calculations, which were then carefully reviewed and double-checked (including the data-sources used) by AJW. Both authors reviewed and approved the final manuscript version for submission. JF is the guarantor.

Patient involvement

No patient or public involvement occurred during the preparation of this manuscript, because it is a relatively technical paper for consideration by public health experts and officials, created without external funding, under time-pressure for submission while it is still relevant to COVID-19 exit strategy decision-making.