Association of Simulated COVID-19 Policy Responses for Social Restrictions and Lockdowns With Health-Adjusted Life-Years and Costs in Victoria, Australia

This economic evaluation determines the optimal policy response to the COVID-19 pandemic in Victoria, Australia, using a net monetary benefit approach for policies ranging from aggressive elimination and moderate elimination to tight suppression and loose suppression.

eTable 2. Conceptualisation and specification of the triggers to shift between stages by policy scenario eTable 3. Parameter estimates and 'agent' characteristics most relevant to current paper used in the agent-based model (for full details see source code and ODD protocol in footnote to this table) eTable 4. Key input parameters by level of policy stringency in the ABM eTable 5. Input parameters to PMSLT (excluding those from ABM) and GDP costs (inputs in italics only used in sensitivity analyses) eTable 6. Outputs from ABM, and estimated GDP loss, for 'best' scenario: 12 months intervention or ABM time horizon (i.e., assumed vaccination available in 12 months); 1% probability per day of incursion of infected person into Victoria in elimination strategies eTable 7. Estimate incremental health loss (HALYs) loss compared to BAU (i.e., no SARS-CoV-2 pandemic) and additional health expenditure (3% discount rate) eTable 8. Sensitivity analyses incremental health loss (HALYs) compared to BAU (i.e., no SARS-CoV-2 pandemic) and additional health expenditure and GDP loss (3% discount rate; in US$ millions), using the median infection rate across 100 simulations (% in parentheses are change relative to baseline) eFigure 1. Stages and triggers for tight suppression eAppendix 1. Estimates of GDP loss by stage, from Australian and Victorian Treasuries eTable 9. Estimates of per week GDP loss by stage, relative to no restrictions. Scope included GDP losses caused by Victorian restrictions, but borne beyond Victoria

Seeded cases
An initial volume of 2400 active cases were seeded into the model on day 0. This was followed by 7 days of 80 cases per day. COVID-Safe Electronic App Uptake (normal distribution) m = 30%

Agent Characteristics Definition
Infection status Susceptible, Infected, recovered, deceased

Time now
The number of days (integer) since an infected person first became infected with SARS-CoV-2 Age-range The age-bracket (categorical) of the person, set to census data deciles from 0 to 100. Used in this simulation to capture differences in exposure risk through school closures and workforce status.

Risk of death
The overall risk of death (float) for each person based on their ageprofile. Purely used in this simulation to remove the agents dying during the 100-day simulation time.

Location
Agents interact in over a 2-dimensional plane with their location recorded at each time-step via an x/y coordinate system.

Span
The distance the person moves around the environment away from their home location -longer distances result in higher likelihood of close contact with novel other people (agents) in the model.

Heading / Distance
The direction and extent of travel of the person at the current time-step. The heading and speed variables combine to create local communities and control interaction between and across communities. At higher lockdown stringency levels, agents are restricted to movement in areas closer to their home location.

Contacts
A count (integer) of contacts the person (agent) had interacted with in the past day as they moved within the model's environment. This is used in estimation of contacts with transmission potential each time-step and calculation of individual reproduction numbers at the end of infectious periods.
Code for ABM at: https://github.com/JTHooker/COVIDModel (last accessed 23 August 2020). ODD protocol at: https://github.com/JTHooker/COVIDModel/blob/master/ODD%20Protocol%20Aus%20NZ%20COVID19%20model.pdf (last accessed 23 August 2020). ¥ Assumed parameter based on expert opinion in conjunctions with available public data sources such as Google COVID-19 mobility reports. ¥¥ 10% of the population potentially transmit infections widely through occasional travel to random locations. ¥¥¥ The source paper reports an adjusted odds ratio of 0.15 for a systematic review of observational studies. Given possible residual confounding, and to be conservative, we used 80% rather than 85%. *This reports all cases known to the model user on day 6 of their infection. In alternative modes, code also allows for underreporting under extreme pressure on the track and trace system (e.g., in unmitigated scenarios). ₤ % mask wearing is fixed part of scenario, therefore no uncertainty. % reduction in contact tracing time due to COVID-Safe App, when both people have the App 50% 50% 50% 50% 50% † The range of movement is in a two-dimensional plane, meaning the relative difference in number of destinations is a function of the quadratic, e.g., for Stage 5 c.f. Stage 1, 25 to 4 relative difference.
‡ For this paper, all children <18 years treated the same (but can be stratified in extensions to modelling) € At each time-step, both the proportion of people who complied with social distancing measures and the proportion of time they complied declined by 1 unit to the baseline level set at each stage. ₤ A recent systematic review found a pooled OR for reduced transmission of 0.85 (or 85%), in mostly clinical studies and some community studies. 7 This probably overestimate effectiveness in real-life. We therefore specified a beta distribution 24.3 and 8.08, giving mean 0.75, SD 0.075, 95% uncertainty interval 0.590 to 0.881. eTable 5: Input parameters to PMSLT (excluding those from ABM) and GDP costs (inputs in italics only used in sensitivity analyses)

Input Specification Uncertainty Comment and source
Population counts Estimated usually resident population 2020 Nil UN World Population Prospects for Jul 2020 † All-cause mortality rates Single year of age mortality rates, generated from GBD five-year age group rates using interpolation on log scale.
Log normal approximation to GBD published 2.5 th and 97.5 th percentiles. ‡

IHME GHDx
All-cause morbidity rates Single year of age prevalent years of life with disability (YLD) proportions Log normal approximation to GBD published 2.5 th and 97.5 th percen les. ‡

IHME GHDx
Cause-specific mortality rates (road traffic crash) GBD five-year age group mortality rates.
(Nil -only used in sensitivity analyses as expected values)

IHME GHDx
Cause-specific morbidity rates (road traffic crash non-fatal injuries, depression, anxiety) Single year of age prevalent years of life with disability (YLD) proportions for these for conditions.
(Nil -only used in sensitivity analyses as expected values)

IHME GHDx
Forecast annual percentage change (APC) in all-cause mortality rates APC by sex by five-years age-groups for GBD mortality rates 1980-2017, used to forecast mortality rates to 2035 -then no change.
Nil IHME GHDx  16.3% (11.4% -23.5%) † Victorian Gross State Product (GSP) was Aus $454.59 billion in 2019, or US $308.20 billion (using OECD purchasing power parity). There were large economic stimulus packages in Australian, most notably an Aus $507 billion Federal Government stimulus (KPMG, last updated 18 Nov 2020, accessed 2 May 2021; https://home.kpmg/xx/en/home/insights/2020/04/australia-government-and-institution-measures-inresponse-to-covid.html) -of which perhaps a quarter to a third was directed to Victoria. Whilst not all of this would have flowed through in the year post-ceding the Victorian second wave, it does explain why the pre-stimulus estimates in this table show a percentage loss in GDP of greater magnitude than that actually observed (e.g. a 1.1% fall for calendar year 2020 compared to 2019 for all of Australia; https://www.abs.gov.au/statistics/economy/national-accounts/australian-national-accounts-national-income-expenditure-and-product/latestrelease, accessed 2 May 2021). That is, our estimates are do not include the offsetting stimulus impacts on GSP and GDP. -107 (-175, -41.7) † HALYS for the Victorian population (3% discount rate) over the remainder of their lifetime in BAU were 127 million, and health expenditure (also 3% discount rate) was $US 1021 billion. Thus, by way of comparison, the HALY loss as percentage of BAU HALYs was 0.0002% from aggressive and moderate elimination, 0.0015% from tight suppression was 0.0075% from loose suppression.
. eTable 8: Sensitivity analyses incremental health loss (HALYs) compared to BAU (i.e., no SARS-CoV-2 pandemic) and additional health expenditure and GDP loss (3% discount rate; in US$ millions), using the median infection rate across 100 simulations (% in parentheses are change relative to baseline) Sensitivity

: Estimates of GDP loss by stage, from Australian and Victorian Treasuries
State and Commonwealth treasuries in Australia provided estimates of the impact of COVID-19. We prioritised the use of such government estimates since these are usually bolstered by a wide array of near real-time indicators, for instance, income and sales tax collection data.
The Victorian Treasury in its July 2020 update, before the announcement of Stage 4 restrictions, estimated that a combination of approximately six weeks of Stage 3 and six weeks of various Stage 2 lockdowns in the April to June and July to September quarters would reduce GDP by 11 per cent (or roughly 1 billion dollars a week compared to expected GDP had there been no COVID-19). 9 For each stage, the broader Australian economy had a smaller impact from equivalent restrictions, in part due to a smaller reliance outside of the State of Victoria on hospitality and the resilience of iron ore prices. The Australian Department of Finance in its Mid-Year Economic and Financial Update estimated that Stage 3 restrictions between Mar 30 and mid-May had an estimated Australia-wide cost of AUD 4 billion per week, around, roughly 11 per cent of the weekly Australian GDP of AUD 36.3 billion dollars (~AUD1.89 trillion/52). It estimated the increment between Stage 3 and the 'unlocked' economy to be around AUD 2bn per week or around 5.5% of GDP. Finally, the prime minister of Australia announced on 6 Aug 2020 10 , that the cost of six weeks of Victorian Stage 3 restrictions in the Jul-Sep quarter would be around AUD 3.3 billion, and the cost of six weeks of Stage 4 restrictions, incremental to Stage 3, was AUD 7 to 9 billion. Approximately 80 per cent, or $6 billion to $7 billion, was expected to arise from businesses and activity in Victoria, while the remainder cost was borne by the rest of Australia due to spill-over effects.
Based on these announcements and using a scope of including GDP losses caused by Victoria even if borne beyond Victoria, we estimate the impact of the COVID-19 control strategies to be approximately as shown below in Supplementary Table 9.

eAppendix 2: Net Monetary Benefit
We estimated the monetary benefit (NMB) 1 approach for each of the 100 runs:

= × −
Where: -i indexes the 100 iterations -j indexes the WTP -k indexes the four policy scenarios -and Cost is the net health expenditure for the health system perspective analyses, and from the societal perspective adds GDP costs to health system costs.
Within each iteration i and WTP j, the policy scenario with the highest NMB is selected. Across all 100 iterations, each policy response k will have a probability of having the highest NMB, and the policy option with the highest probability is deemed 'optimal' at that WTP. Finally, these outputs can be shown as cost effectiveness acceptability curves.

eAppendix 3: Average citizen annual health expenditure
Consistent with recommended practice in cost effectiveness analyses 13,14 , in the USA 15 and the Netherlands 16 , we included 'unrelated disease costs' in the economic evaluation. This means that in addition to including the costs of SARS-CoV-2 cases per se (Appendix 3), knock-on changes in health system expenditure are also included. For SARS-CoV-2, this means that if someone dies due to SARS-CoV-2 infections, their reduced health expenditure in the future is included (leading to a potentially net negative expenditure depending on the balance of costs, age and discount rate). In a simulation model, this is easy to incorporate, by including an expenditure reward per cycle in the model for diseases not explicitly modelled elsewhere -which in the case of SARS-CoV-2 modelling, is simply the expected annual (or monthly) average health system expenditure.
Data were extracted from the Australian Institute of Health and Welfare (AIHW) report 'Disease expenditure in Australia, which separates the total expenditure by sex and age. 8 The data are from the 2015-16 financial year, where the total health expenditure totalled $170.4 billion $AU (2016). The AIHW attributed $106.857 billion of this spending to age and sex related health spending (62.7% of total health expenditure), with data provided as total expenditure by age and sex subgroup. 17 We extracted population demographics from the Australian Bureau of Statistics (ABS) 2016 population report, 17 and the total health expenditure for each subgroup was then divided by the corresponding population numbers for these subgroups. Thus the 2016 health expenditure is expressed as per capita expenditure, by age and sex. 18 The AIHW estimates variable health expenditure at 94% of total health expenditure, 8 whilst New Zealand variable expenditure is estimated at 91% total expenditure. 19 We elected to assume variable expenditure was 90%, allowing for fixed costs in running services Noting the above 62.7% of total health expenditure captured by AIHW estimates, we therefore multiplied all age by sex empirical estimates by a factor of 90/62.7 to generate the estimated predicted Australian variable health expenditure per capita, by age and sex.
Next, we inflation adjusted these expenditures from 2016 $AU to 2019 $AU using Australian CPI adjustment factors (OECD rates; 18 https://data.oecd.org/price/inflation-cpi.htm). Finally, we converted to 2019 USD using the AUD-USD 2019 purchasing power from the OECD.

eAppendix 4: SARS-CoV-2 parameters
For each monthly cycle, the number of SARS-CoV-2 infections were split into the following categories for all modelled cases, equivalent to all notified and confirmed cases 1 : A. Asymptomatic B. Symptomatic, not admitted to hospital C. Symptomatic, admitted to hospital D. Symptomatic, admitted to hospital and ICU E. Die (may come from anyone of B, C and D) This is slightly different from our previous model 20 as there is now sufficient within-Australia data (i.e. for Victoria, from the Victorian Department of Health and Human Services (Vic DHHS)) to estimate probabilities of hospitalisation, ICU admission and death directly. A large fraction of people dying did not get admitted to ICU, dying on a general ward or in community care -especially elderly people with a do not resuscitate order. We therefore estimated proportions of cases into four mutually exclusive categories (A, B, C and D) for the quantification of morbidity and health expenditure, and one additional category for the quantification of HALYs lost due to death (E).
In this Appendix we describe in order: -The epidemiological parameters to split each month's SARS-CoV-2 infections into the five above categories. -The excess health expenditure assigned to each of the three symptomatic SARS-CoV-2 categories (B, C and D). -The morbidity-loss assigned to each of the three symptomatic SARS-CoV-2 categories (B, C and D).

Epidemiological parametrisation
Supplementary Table 11 (below) shows the number of cases, hospitalisations, ICU admissions and deaths in Victoria. The dates are deliberately different, so that average time lags are allowed for: up to 14 days from notification to death; subsuming 10 days from notification to ICU admission; subsuming 7 days from notification to hospitalisation. Supplementary Figure 2 shows the ln odds of hospitalisation, ICU admission and deaths for observed data when the number of events is 5 or more, and from simple predictive logistic regression models on the same data. For the latter regressions, main effects were included for sex and age as a continuous variable, and additional age-dummies due to non-linearity on the ln odds scale for: -Hospitalisation: 0-9, 10-19, 80-89 and 90+ year olds -ICU admission: 80-89 and 90+ year olds -Deaths: 90+ year olds. eFigure 2: Ln odds of hospitalisation, ICU admission, and death for confirmed cases for observed data when number of events >5 and from a logistic regression prediction otherwise Error bars are 95% confidence intervals We elected to use the observed ln odds when the number of events was five or more, otherwise use the logistic regression predicted ln odds. These estimates, and their standard errors, are shown in Supplementary Table 12.
Sequentially, the process to estimate the actual disaggregation of SARS-CoV-2 cases (outputted by the ABM) by category of morbidity and mortality was: -Estimate the monthly number of deaths (E) by sex and age; propagate through the PMSLT increasing mortality rates (no change in morbidity) -Estimate the number of ICU admissions (D) by sex and age; propagate through the PMSLT increasing morbidity and health expenditure (no change in mortality) -Estimate the number of hospitalisation admissions (B) by sex and age, subtracting off ICU admissions; propagate through the PMSLT increase morbidity and health expenditure (no change in mortality) -Estimate number of asymptomatic cases (A).

Morbidity loss by SARS-CoV-2 category
An 'average' incremental morbidity impact due to SARS-CoV-2 for each month was estimated as the weighted sum of the morbidity impact for each of the five symptomatic SARS-CoV-2 categories (weighted by proportionate distribution of SARS-CoV-2 infections by category -which varied iteration to iteration given the uncertainty described above). We measured morbidity impacts using disability rates (DR), 21 according to severity of acute infection (mild, moderate and severe). For ICU admissions, DR were based on severe chronic obstructive pulmonary disease (COPD) to reflect Acute Respiratory Distress Syndrome (ARDS). We assume all survivors return to their baseline health status (pre-SARS-CoV-2) (DR: 0) following a specified recovery period as described below.
Morbidity loss for the four categories of symptomatic SARS-CoV-2 infection include: -Morbidity for people admitted to ICU, but surviving, assuming a mean duration from symptom onset to recovery of 6 weeks was based on the higher range of the median time from onset to clinical recovery for patients with severe or critical disease. 22 We applied a DR for moderate acute infectious episode for 1 week of 0·051 (0·032-0·074), plus DR for severe acute infection for 2 weeks of 0·133 (0·088-0·190), and ICU admission for 1 week of 0.408 (0.273-0.556), plus a return to baseline health (DR: 0) over 2 weeks (equivalent to 50% probability of ARDS for 2 weeks). -Morbidity for people admitted to hospital but not requiring ICU, assuming a mean duration from symptom onset to recovery of 4 weeks, based on the lower range of the median time from onset to clinical recovery for patients with severe or critical disease. We applied a DR for moderate acute infectious episode for 1 week of 0·051 (0·032-0·074), plus severe infectious episode for 2 weeks of 0·133 (0·088-0·190), plus return to baseline health (DR: 0) over 1 week (equivalent to 50% probability of severe infectious episode for one week). -Morbidity for people diagnosed with symptomatic disease but not admitted to hospital, assuming a mean duration from symptom onset to recovery of 2.5 weeks, based on data from the WHO-China Joint Mission report on median time from onset to clinical recovery for mild cases of approximately 2 weeks. 23 We assume half of people with symptomatic disease who are not admitted to hospital have mild symptoms, and the other half have moderate symptoms. 22 Therefore we applied a DR of 0·051 (0·032-0·074) for moderate acute infection for 50% of this group for 2.5 weeks and a DR of 0·006 (0·002-0·012) for mild acute infection for 2.5 weeks for the remaining 50% of this category. We assumed those who died from SARS-CoV-2 in the community have the same morbidity as ICU deaths.

Excess health expenditure by SARS-CoV-2 category
We estimated the excess health expenditure by SARS-CoV-2 category using an ingredients approach. For each patient category, we modelled the expected patient pathway through the health system, based on available SARS-CoV-2 data from China, Australia, and the UK (Supplementary Tables below). For each patient subgroup category described in the paper, we calculated total health expenditure by first estimating resource use required for a typical patient (e.g., hospital or outpatient visits, or drugs), and multiplying by unit costs for each of the specified resources.
Supplementary  24 As it currently unknown how many patients have died without receiving hospital treatment, we assume that end-of-life care within these facilities will require similar levels of health care as hospitalisation.
The SARS-CoV-2 pandemic is expected to add additional costs to hospital operations, adjusting for complexity of patients and added infection control required including the need for isolation of patients, staff time for proper fitting of personal protective equipment, and enhanced cleaning regiments. As a result, inpatient and ICU hospital costs have been scaled up by 20% to account for these extra costs. This 20% estimate is based on the Coronavirus Aid, Relief and Economic Security (CARES) Act in the United States that provides a 20% add-on payment for COVID-19 patients. 25 We expect that this loading will be a moderate estimate, and likely underestimate the true hospital costs during a pandemic outbreak. 25 Total costs for each patient category described above is found in Supplementary Table 15. COVID-19 restrictions may result in lower road traffic deaths by reducing the number of people leaving their homes. Simultaneously, such restrictions have been linked to higher car and bike usage on roads, compared to public transport use as individuals aim to socially distance. Similarly, lower road congestion may be linked to higher speeds and potentially higher road traffic deaths. To estimate the effect of COVID-19 restrictions on road traffic deaths in Victoria, we obtained COVID-19 mobility data for January to October 2020 from Apple, which describes the activity of users seeking driving routing directions. This data was supplemented by the data on the road traffic deaths on Victorian roads from January through September 2020, from the Australian Bureau of Infrastructure and Transport Research Economics. The data was used to calculate the association of a change in the driving mobility index on road traffic deaths using a Poisson regression. A 1-point increase in the mobility index causes a 0.45% increase in road traffic deaths, although the result is not statistically significant. The results of the regression are presented in Supplementary Table 19. A limitation of the methodology based on using observational data is that the initial response to Stage 3 restrictions was stronger than later in July, this may be due to the lower national cases of COVID-19 when the second wave. The effect of Stage 3 on RTC cannot be disentangled from this behavior response to the initial wave and may therefore be overestimated.