A Net Benefit Approach For The Optimal Allocation of A COVID-19 Vaccine

medRxiv preprint doi: https://doi.org/10.1101/2020.11.30.20240986; this version posted December 2, 2020.
The copyright holder for this preprint

(which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.
All rights reserved. No reuse allowed without permission.
A Net Benefit Approach for the Optimal

Allocation of a COVID-19 Vaccine
Erin Kirwin MA1,2,*, Ellen Rafferty PhD1, Kate Harback PhD1, Jeff Round PhD1,3, Christopher
McCabe PhD1,4
1
Institute of Health Economics
2
Health Organisation, Policy, and Economics, School of Health Sciences, The University of Manchester
3
School of Public Health, University of Alberta
4
Department of Emergency Medicine, Faculty of Medicine and Dentistry, University of Alberta
*Corresponding author. [email protected], Institute of Health Economics #1200, 10405 Jasper Avenue Edmonton, AB, Canada T5J 3N4
Competing Interests
The authors have no competing interests to declare.
Funding
The IHE is funded in part by the Alberta Ministry of Health.
1
NOTE: This preprint reports new research that has not been certified by peer review and should not be used to guide clinical practice.
medRxiv preprint doi: https://doi.org/10.1101/2020.11.30.20240986; this version posted December 2, 2020. The copyright holder for this preprint
Abstract
Coronavirus disease 2019 (COVID-19) is a severe, novel virus that has spread globally. The implementation of a
combination of public health interventions is required to reduce viral spread and avoid overwhelming acute
care systems. Once available, an effective vaccination will further mitigate the impact of the COVID-19
pandemic. However, decision makers will initially need to prioritise access to limited vaccine stockpiles as
these will be insufficient to vaccine the whole population.
The aim of this study is to identify optimal vaccine allocation strategies defined by age and risk target groups,
coverage, effectiveness, and cost of vaccine, within a dynamic context where other public health responses
and population behaviour change. In this study we use an epidemiological model of COVID-19 that has been
enhanced to produce expected costs and Quality Adjusted Life Year results as well as total cases,
hospitalisations, deaths, and net monetary benefit. We use the model to simulate hypothetical scenarios
where vaccine is allocated beginning on October 15, 2020 with vaccine assumptions ranging from moderately
optimistic to ‘worst-case scenario’. Net monetary benefit is used as the objective for optimisation.
In a scenario with a sterilizing vaccine that is 80% effective, a stockpile sufficient for 40% population coverage,
and prioritisation of those over the age of 60 at high risk of poor outcomes, active cases are reduced by 29.2%
and net monetary benefit increased by $297 million dollars, relative to an identical scenario with no vaccine.
The relative impact of prioritisation strategies varies greatly depending on concurrent public health
interventions, for example, polices such as school closures and senior contact reductions have similar impacts
on incremental net monetary benefit when there is no prioritisation given to any age or risk group (147 vs. 120
million, respectively), but when older and high risk groups are given priority, the benefit of school closures is
much larger than reducing contacts for seniors (iNB 122 vs. 79 million, respectively). Results demonstrated
that rank ordering of different prioritisation options varied greatly by prioritisation criteria, with different
vaccine effectiveness and coverage, and by concurrently implemented policies.
The results of this paper have three key policy implications: (i) that optimal vaccine allocation will depend on
the public health policies, and human behaviours in place at the time of allocation; (ii) the outcomes of vaccine
allocation policies can be greatly supported with interventions targeting contact reduction in critical sub-
populations; and (iii) the identification of the optimal strategy depends on which outcomes are prioritised.
Keywords
COVID-19, Cost-effectiveness, Net Benefit, Vaccine, Immunisation, Mathematical Modelling, Epidemiologic
Modelling, Risk
Word Count
Main body, excluding tables and figures: 5,896
2
Introduction
Background. Coronavirus disease 2019 (COVID-19) is a severe, novel virus that has spread globally. On 30
January 2020, the World Health Organization (WHO) declared a Public Health Emergency of International
Concern and declared a Pandemic on 11 March 2020 [1]. As of 27 November 2020, there have been over 60
million confirmed cases and over 1.4 million deaths worldwide. In Alberta, Canada, the first presumptive
COVID-19 case was identified on 5 March 2020; as of 27 November, there have been over 51 thousand
confirmed cases and 510 deaths in the province [2].
Key public health interventions for COVID-19 mitigation include hand hygiene, mask wearing, physical
distancing in various locations, travel restrictions, testing, contact tracing, isolation of cases, and quarantine of
close contacts. Public health policies and individual decisions combine to produce changes in human
behaviours, including reductions in contact. The implementation of a combination of public health
interventions is required to reduce viral spread and avoid overwhelming acute care systems. Decision makers
are challenged with choosing the optimal combination of interventions to mitigate public health risks of
COVID-19 while meeting other population needs including access to the health system for diagnosis and
treatment of other conditions, economic activity, and social well-being. Once available, additional effective
interventions such as vaccination and antiviral treatment will further mitigate the impact of the COVID-19
pandemic.
Over 200 COVID-19 vaccination products are currently under development, several of which have reached the
phase III trial stage [3]. These vaccines may reduce disease burden by preventing infection in exposed
individuals, reducing severity of infection in infected individuals, and/or by preventing secondary infections.
Internationally, decision makers will need to prioritise access to limited vaccine stockpiles [4-6]. Many
countries have produced decision frameworks [7], and researchers have developed commentaries on ethical,
economic, and epidemiologic justifications for different allocation objectives and strategies [8-14].
Mathematical models with health economic outputs can assist in public health decision-making by assessing
outcomes associated with various public health interventions. These models can simulate a range of vaccine
allocation scenarios, and demonstrate outcome dynamics with changing vaccine characteristics, allocation,
public policy, and population behaviours. Hundreds of models have been developed internationally to support
decision making and public health messaging. A handful of studies evaluating vaccine allocation strategies
have been published to preprint servers [13, 15-17]. Only one of the studies evaluates scenarios with
concurrent public health mitigation strategies [13]. None of the studies consider interventions targeting
individuals by both age and risk group, with well-defined vaccine allocation policies or health economic
outputs, limiting their value to decision makers.
Aim and Objective. The aim of this study is to identify optimal vaccine allocation strategies defined by age and
risk target groups, coverage, effectiveness, and cost of vaccine, within a dynamic context where other public
health responses and population behaviour change. In this study we demonstrate the use an epidemiological
model of COVID-19 that has been enhanced to produce expected cost and Quality Adjusted Life Year (QALY)
outcomes. Model outputs include total cases, hospitalisations, deaths, and net monetary benefit (NMB) [18].
We use the model to simulate hypothetical scenarios where vaccine is allocated beginning on October 15,
2020 with vaccine characteristic scenarios ranging from moderately optimistic to ‘worst-case scenario’, and
NMB is used as the objective for optimisation. This range of scenarios is designed to highlight key
considerations and demonstrate outcome trade-offs.
3
Methods
We developed the COVID-19 Risk Assessment Model (CRAM), a transmission dynamic ordinary differential
equation compartmental susceptible-exposed-infected-recovered (SEIR) model. We calibrated the CRAM to
the time variant state of public health interventions and human behaviour, using data from Alberta, Canada.
The model is designed to expand upon the capabilities of existing models for vaccine allocation by: (i) including
an expanded set of age groups and dividing these further into two risk groups; (ii) disaggregating physical
distancing interventions into workplace, school, home, and other locations; and (iii) including additional
interventions such as case isolation. To evaluate the risks and outcomes associated with each strategy,
scenarios are characterised by a combination of public health interventions that change over time, in
comparison to a baseline scenario. CRAM estimates epidemiologic outcomes such as active cases,
hospitalisations, and total infections over time, and produces health economic outputs including total cost,
QALYs, and NMB. CRAM was programmed in Wolfram Mathematica, version 12.1.
Model Structure
CRAM is stratified into 16 age groups (five-year ranges, from 0 to 75 years). Each of these 16 groups is divided
further into two risk groups (high-risk and not high-risk), by risk of poor outcome due to COVID-19. An
additional group is used to represent the Long-Term Care (LTC) population, who are all defined as high risk
individuals aged 75 years and older. This results in 34 age-risk groups. CRAM is based on the structure of an
existing model developed for the province of Alberta [15]. CRAM has five overarching compartments:
Susceptible (which includes Susceptible, No Distancing, Vaccine, and Vaccine Failure), Exposed, Infected
(composed of Pre-Symptomatic, LTC, Community, Hospital, and Pre-Isolation), Isolation, and Recovered
(comprised of Immunised, Recovered, and Dead) (see Figure 1). The model structure was developed to
produce outcomes of interest to decision-makers and to leverage the highest quality data available in real-
time during the COVID-19 pandemic.
A fraction of the population is initially allocated to the Exposed compartment, with the remaining population
allocated to the Susceptible compartment. Susceptible individuals are vaccinated based on program start date,
vaccine prioritisation policy, number of doses remaining, and the daily maximum number of vaccines possible.
Vaccine prioritisation model scenarios provide vaccine access to different age and risk groups on different
dates. Once vaccinated, after the period required to develop antibodies, individuals move either to the
Immunised compartment, or to the Vaccine Failure compartment (and remain susceptible) as a function of the
vaccine effectiveness. If the number of individuals who are eligible for vaccine on a given day exceeds the
maximum vaccine capacity, the available doses are prorated across eligible age and risk groups by group size.
Susceptible individuals are exposed to the virus as a function of the force of infection which is determined by
the transmissibility of the virus, seasonality, number of persons infected, and rate of contact between
individuals across age groups. Susceptible individuals exposed to the virus move to the Exposed compartment
and following the latent period they move from the Exposed to Infected.
Depending on public health guidance, some individuals may move from the Susceptible compartment to the
Susceptible, No Distancing compartment, which can be used to separate segments of the population following
different distancing measures. It can be used, for example, to simulate in-person schooling. Children who are
not high risk move to the Susceptible, No Distancing compartment with higher rates of contact due to in-
person education.
Upon arrival in the Infected compartment, individuals are Pre-Symptomatic, defined as cases in the community
that are shedding virus but not yet experiencing or displaying symptoms. Once the pre-symptomatic period
4
has elapsed, individuals are allocated between the Community and Hospital compartments, dependent on
their age and risk group. If a case isolation policy is in place, a subset of the Community compartment will
move to the Pre-Isolation and subsequently the Isolation compartments, where they remain for the duration
of the remaining infectious period. All cases contribute to the force of infection prior to moving to the Isolation
compartment, but individuals in the Isolation compartment do not contribute to the force of infection. In
CRAM the Isolation compartment represents the isolation of cases but not the isolation of close contacts.
Entry into LTC and Hospital varies by age group, with high risk individuals over the age of 75 moving from the
pre-symptomatic to the LTC compartment. Individuals move from the LTC compartment at the end of the
infectious period to either the Recovered or Dead compartments in proportions dependent upon their risk
group. Entry into the Hospital compartment is a function of age and risk, fit by an exponential regression on
admissions by age, and adjusted for risk by an odds ratio. The Hospital compartment includes both general
inpatient cases, and those admitted to intensive care units (ICUs). At the end either the infectious period or
the hospital length of stay (LOS), individuals move from the Community, LTC or Hospital compartment to
either the Recovered, or Dead compartment, at a proportion dependent upon their risk group.
We used a synthetic contact matrix disaggregated into four locations: home, workplace, school, and other,
where other comprises all remaining possible locations [19]. Disaggregating the total contact matrix by
location allows for adjustments in contact rates at locations to align with policy interventions. For example, we
can apply work-based interventions to specific age groups by adjusting the work contact matrix at various
points in time. The total contact matrix is generated by summing the contacts for each age group across each
of the matrices. Contact reductions for seniors aged 75 years and older and for high-risk individuals are applied
directly to the total contact matrix. Complete details of model equations and assumptions are available in
Appendix 1.
Figure 1. CRAM Structure
Note: Figure intended for colour printing.

Epidemiologic Inputs
CRAM was parameterised wherever possible with data from Alberta, including estimates for the total
population by age group and risk of severe outcome, hospitalisation rate by age group, and LOS for
hospitalised cases. Odds ratios for the risk of hospitalisation were estimated by risk group using logistic
regression. Active cases in the model include both diagnosed and undiagnosed cases, fitted using
seroprevalence data from 20 May 2020 provided by Alberta Health. Details on all model inputs are provided in
Table 1.
Data from the Google Community Mobility report [20] were used as key inputs for contact reductions in
workplaces and other locations. A rapid expert elicitation exercise was conducted to obtain estimates for
unknown model parameters. The initial rapid elicitation activity surveyed experts in Alberta on reductions in
contacts for seniors aged 75 years and older, high-risk individuals, and reduction in contacts in home locations.
A second rapid elicitation exercise determined likely return-to-school rates and in-school contact reductions by
age group. We adapted elicitation methods and expert training materials from the SHELF (Sheffield ELicitation
Framework) tool [21], the results of which are included in Appendix 2.
Table 1. Epidemiologic Parameters
Symbol Parameter Baseline value Description and

name source
Transmission Fitted.
Number of 10 cases, distributed between age groups 35–40 years to ≥75 years Assumed.
individuals on model day 1.
exposed at
model
initialisation
,
Seasonal A standard function from mathematical epidemiology [22] was Assumed.
transmission modified, as given in Appendix C.
adjustment
season can take on values between 0 and 1 and is centred to reach
its minimum value on the mid-date between the hottest day of the
year in Edmonton and Calgary.
In this report, = 0.3, which solves to 0.4 at its minimum,
meaning that transmission at the seasonal minimum is 40% of
transmission at the seasonal maximum.
Susceptibility by 0.489 if aged <10 years, 1 otherwise. Estimates taken
age group from a modelling
study [23].
,, , Contact A 100% reduction in the contact rate is applied after 15 March, Rapid expert
reduction in which is maintained for the duration of the modelling period in the elicitation
schools baseline scenario. exercise.
The following values were elicited from the expert panel for the fall
return to school (median and 90% CI):
Age group Contact reduction Proportion

returning
5–9 years 0.692 [0.358, 0.869] 0.760 [0.399, 0.933]
10–14 years 0.573 [0.314, 0.846] 0.831 [0.399, 0.941]
15–19 years 0.522 [0.278, 0.861] 0.860 [0.634, 0.994]
6

, Contact A 41% reduction in the contact rate is applied after 17 March. Rapid expert
reduction in elicitation
homes at time t exercise.
Home refers to
all residential
locations, so that
reductions in
visits to
secondary
households are
reflected in
these contact
rates.
, , Contact Reduction in the contact rate relative to normal (100), given in a Based on
, reduction in weekly interval. Future weeks are expected to remain constant at reduction in time
workplaces and the rate of the most recent week. Day 1 is 6 March, all dates prior spent in
other locations have an assumed contact reduction of 100, indicating normal workplaces, or
at time t contact rates. combined time
spent at retail,
Start day Other Workplace grocery, and
1 100 100
transit locations
from Google
8 86 80 Community
15 62 57
Mobility data
[20].
22 54 51
29 56 50
36 55 45
43 60 53
50 62 53
57 66 55
64 68 58
71 71 55
78 75 64
85 78 67
92 78 68
99 81 70
106 84 71
113 81 65
120 84 69
127 84 71
134 84 70
141 85 71
148 83 64
, Contact A 61% reduction in the contact rate for seniors in all locations is Rapid expert
reduction for applied after 20 March. Seniors defined in this model as those aged elicitation
seniors 75 and above. exercise.
7
Contact A 64% reduction in the contact rate is applied to reflect physical Rapid expert
reduction for distancing for high-risk individuals after 15 March. elicitation
high-risk exercise.
individuals
, , Proportion of User input day for scenario analysis, assumed 0 for baseline. User/scenario
susceptible input.
Can be used for scenarios with different contact rates in a subset
group by age
of the population.
and risk
changing to new
contact rates at
time t
, , Proportion of User input day for scenario analysis. User/scenario
susceptible input.
group by age
and risk
immunised at
time t
Proportion of User input day for scenario analysis, 0.169 based on User/scenario
community seroprevalence studies conducted in Alberta. Note that the input.
infections Isolation compartment represents case isolation only.
isolated
Latent period Mean 3.3 days. PHAC Modelling
Evidence
summary [24].
Incubation Mean 1.3 days. (Gamma shape 17.916, rate 3.908) PHAC Modelling
period Evidence
summary [24].
Estimated by applying a gamma distribution to the median and
th
maximum values identified in the PHAC report (4.5 and 6.5, as 50
th
and 95 percentiles, respectively) and subtracting the latent
period.
Infectious Mean 20.3 days. (Gamma shape 11.608, rate 0.537) PHAC Modelling
period Evidence
summary [24].
Estimated by applying a gamma distribution to the median and
maximum values identified in the PHAC report (21.0 and 33.0, as

th th
50 and 95 percentiles, respectively) and subtracting the
incubation period.
Days in hospital Mean 8.08 days. Alberta Health
data.
Days prior to 5 Assumed.
isolation of
cases

Days prior to 14 Assumed.
immune
response from
immunisation
OR of hospital 6.304 (95% CI [4.973, 7.991]). Alberta Health
for high-risk data.
individuals
OR of death for 42.384 (95% CI [22.449, 80.018]). Alberta Health
high-risk data.
individuals
Fitting term for Fitted.
hospitalisation
probability
8
Hospitalisation Hospitalisation risk by age estimated from cumulative counts Alberta Health
risk by age converted to proportion by age and analysed in an exponential data.
regression for model input.
Fitted values: Max[ 6.346 0.3093" #, 0$

, Probability of Function of %& '(
, , and the fitted value of &) .
hospitalisation
LTC probability
*, Probability of Function of &) .
LTC admission
+, Probability of Fitted.
death from
community
+- Probability of Function of ., '( , /2.
death from LTC
+1 Probability of Function of ., '( , /3.
death from
hospital
deaths from
LTC, hospital,
and ICU
2- LTC proportion 116/474 Alberta Health
dying data.
21 Hospital 93/460 Alberta Health
proportion data.
dying
3, Population by Total population is spilt by 5-year age groups up to age 75 years. Alberta Health
age and risk of The population is split into two risk groups, by risk of severe data.
poor outcome outcome related to COVID-19.
High-risk is defined as individuals who have met an administrative
data case definition for ischemic heart diseases (ICD-10 codes: I20–
I25), respiratory diseases including asthma (ICD-10 codes: J45, J46)
and COPD (ICD-10 codes: J41–J44).
A subset of the high-risk population aged ≥70 years is shifted to a

th
17 age group representing LTC residents, with the same contact
rates as the high-risk ≥75 years age group, and additional LTC
isolation policies applied.
4
5 45, Contact matrix School, home, workplace, other. Canadian
and Adjusted for different interventions and applied to the 7, and
estimated
4
5 456, 78, compartments, respectively.
contact matrices.
Transformed to
symmetric [19].
Note: CI: confidence interval; COPD: chronic obstructive pulmonary disease; ICD-10: International Statistical Classification
th
of Diseases and Related Health Problems, 10 revision; ICU: intensive care unit; LTC: long-term care; OR: odds ratio; SD:
standard deviation
Economic Evaluation
9
The study is conducted from a health system payer perspective. Outcomes are tabulated from February 5th,
2020 to January 1st, 2021, spanning 332 days. Health system costs include hospitalisation, incremental LTC
costs for infection prevention measures, and the cost of purchasing and administering a vaccine. Health utility
losses include disutility from infection, hospitalisations and LTC infections (during stay and following
discharge), and deaths. We assessed economic outcomes using the NMB framework [18]. We apply the health
opportunity costs estimated for Canada [25], using the value of $30,000 per QALY that is recommended for all
Canadian provinces. All costs are converted to 2020 Canadian dollars, and discounted at an annual rate of
1.5%, according to CADTH guidelines [26].
As no health utility estimates are currently available for COVID-19 disease, we identified utility impacts of
similar conditions. We estimated disutility for community COVID-19 infections from an EQ-5D survey of
individuals with lab confirmed influenza B infections, stratified by age group [27]. Because hospitalised COVID-
19 cases are often diagnosed with pneumonia, we used disutility estimates for viral pneumonia inpatients [28]
to estimate in-facility treatment losses, as well as utility losses for the year following facility discharge. We
estimated utility losses from premature death using the life table approach developed by Briggs et al. [29],
adjusting for baseline utility, increased relative mortality rates, and discounted. All disutility values are
converted to QALYs. Costs related to community infections are not included. We estimated incremental direct
costs associated with COVID-19 hospitalisation and ICU admission treatment based on published Alberta per-
diem costs, as well as the incremental cost of infection prevention and control for LTC cases. Cost and disutility
estimates are presented as incremental values in Table 2. These incremental costs and disutility values are
applied to the relevant compartments, and the resulting NBM is estimated for each scenario.
Fitting
We fit five parameters: virus transmissibility, the probability of hospitalisation, infections in LTC, and the
probability of death from either facility or community infections. Using a deterministic version of the model
defined with mean values of each input parameter we fit observed surveillance data for daily hospital census
counts, as well as cumulative daily death and LTC counts from 6 April to 13 August. Given the interdependence
in the fitting parameters, we used simultaneous fitting methods by optimizing model fit over a Latin hypercube
generated from plausible values for each parameter. Optimal parameter values are identified when the sum of
the normalised difference in model and observed outcomes are minimised.
Outputs
The CRAM sampled from the probability distributions of model inputs, as either 10% of results from the model
fitting, or as estimated from the parameter source (Tables 1 & 2), to account for outcome uncertainty over 200
iterations, and then generated model outputs from the sampled values. Outputs include cumulative estimates
for the mean number of active cases, total infections, and hospitalisations, with credibility intervals to present
results from all simulations.
The economic outputs include expected NMB, a summary statistic calculated as (benefit * cost-effectiveness
threshold)-cost, so that comparisons without the use of ratios can be made [18]. For each of the vaccine
scenarios, incremental net benefit (iNB) is estimated relative to the baseline with no vaccine.
Table 2. Economic Parameters
10
Parameter Name Baseline Value Source and Description
Disutility, During infectious period, QALY loss QALY loss from viral pneumonia, as estimated by [28].
Hospitalisation, ICU 0.58. QALY loss at admission multiplied by duration in the
admission, LTC case infectious compartment. SE or other measures can not
Assume uniform distribution +/- 25%. be calculated because data is not available for paired
t-test.
Disutility, community QALD loss, uniform distribution (min- Influenza B estimate from [27]. Multiplied by duration
case max) by age group: 0-15 (0.58-3.27), in the infectious compartment and adjusted to QALY.
16-65 (00-10.83), 65+ (1.84-7.17).
Disutility, facility Beta distribution, mean 0.10, SE QALY loss from viral pneumonia, as estimated by [28].
discharge 0.0158
Disutility, death Age Mean Range [29], modified with Canadian population norms and
0-9 68.96 (61.45-76.47) standard deviations. Values represent net present
10-19 59.02 (51.51-66.54) value of total QALYs lost.
20-29 50.11 (43.28-56.95)
30-39 41.64 (35.98-47.3)
40-49 33.27 (28.49-38.04)
50-59 25.40 (21.83-28.98)
60-69 18.22 (15.88-20.56)
70-79 11.64 (10.17-13.11)
80-90 6.46 (5.64-7.27)
90-100 3.14 (2.74-3.53)
Incremental cost, Gamma distribution, Alberta Health Interactive Health Data Application,
Hospitalisation and Shape 2234.18, rate 1.72. CMG+ 2018 Grouper 133, Infectious/Parasitic Disease
ICU admission of Respiratory System [30]. Inflation adjusted from

th
Per diem median $1381.64, 90 2018 to 2019 using Canada All-Item inflation, CANSIM
percentile $1781.53. Table 18-10-0005-01 [31].
Incremental cost, LTC Gamma distribution, per diem median Cost from [32] converted to CAD using average 2015
case
th
$216.92, 90 percentile $801.57. exchange rate, inflation adjusted from 2015 to 2019
using CANSIM Table 18-10-0005-01 and fit to a gamma
distribution.
Vaccine cost per dose, $40 Assumed.
including cold chain

and administration
Note: LTC: long term care; QALY: quality-adjusted life year, QALD: quality-adjusted life day
Scenarios
11
In the baseline scenario (‘No Vaccine Baseline’), no vaccine is available, and all contact matrixes and time
variant inputs are held constant at the last observed value from model fitting. Specifically, we maintained
contact reductions at the values determined through the expert elicitation for ‘home’ locations, as well as for
seniors and high-risk individuals, and a fall return to school. Moreover, we kept the most recent observation of
reduction in time spent in workplaces and ‘other’ locations from the Google Community Mobility reports
consistent over the model run. In all scenarios, the vaccine is ‘sterilizing’, meaning that the vaccine protects
individuals from infection.
Vaccine scenarios represent different vaccine allocation strategies, as well as various vaccine characteristics
(i.e., effectiveness, coverage), as described in Table 3. Vaccine scenarios are characterised by the maximum
number of doses that can be delivered per day, the start date of the vaccine program, the definition of the
priority group, the interval (if any) between access for the priority group and the general population, the
effectiveness of the vaccine, and vaccine coverage as a proportion of the total population.
Table 3. Scenarios
Scenario Vaccine Priority Group Vaccine Effectiveness Population dose

Coverage
No Vaccine Baseline N/A N/A N/A
Vaccine Baseline Age and Risk Based 80% 40%
Vaccine Coverage Comparison Age and Risk Based 80% 20%
40%
70%
Vaccine Effectiveness Comparison Age and Risk Based 40% 40%
60%
80%
60% above 60 years
old, 80% otherwise
Vaccine Priority Group Comparison None 60% 40%
Age Based
Risk Based
Age and Risk Based
Worst-Case Age and Risk Based 40% 20%
Note: Maximum doses per day is assumed to be 80,000, with access starting on October 15, 2020 for the priority group,
and 30 days later for everyone else.
If the number of individuals eligible for immunisation exceeded the maximum number of doses that could be
delivered on a given day, the maximum number of doses were pro-rated across each eligible age and risk
group daily. This means that a group representing 5% of the eligible population would receive 5% of the total
number of doses available on a given day. Each of the scenarios begins on October 15, 2020, before the
anticipated distribution of a vaccine in Canada.
To examine the impact of variations in vaccine effectiveness, we included a scenario where effectiveness
varies by age, so that the vaccine is 60% effective over the age of 60, and 80% effective otherwise. In all other
scenarios, effectiveness is constant across age groups. The scenario with 80% effectiveness approximates
12
preliminary results from three phase three trials which have reported efficacy ranging from approximately 70
to 90% [33].
In each of the scenarios, we assumed that the vaccine cost per dose inclusive of delivery is $40, and there was
a 30-day interval between access for different priority groups, starting on October 15, 2020. Extrapolating
from the maximum estimate of vaccines per day in Alberta from [34], we assumed that up to 80,000
individuals can be immunised daily. Each of the comparison scenarios contains the ‘Vaccine Baseline’ scenario
with age and risk-based prioritisation, 80% effectiveness, and 40% coverage, as a common reference point
across scenarios. Scenarios that prioritise based on age give first access to those aged 55 and over, while
programs that prioritise high-risk individuals prioritise all high-risk individuals irrespective of age.
We also defined a worst-case scenario with low vaccine effectiveness (40%) and coverage (20%). Under this
worst-case scenario we modelled additional public health measures, including a 50% reduction in contacts for
seniors, and complete school closures (each policy implemented on November 1, 2020). These simulations
allow model outcomes to compare vaccine policies in combination with other mitigation strategies.
Results
We used the CRAM to produce outputs for each vaccine scenario. The mean and 90% credibility intervals for
epidemiologic and health economics outcomes are presented in the tables and figures below. In the no vaccine
scenario, total cases were shown to remain relatively low in September 2020, but to increase rapidly through
November 2020 to January 2021. In the Vaccine Baseline scenario, active cases are reduced by 29.2%, Table 4
and in Figure 2. The impact of population vaccine coverage on outcomes demonstrates improved outcomes
with greater population vaccine coverage, as expected (Table 4, Figure 2). When coverage is 70%, there is
sufficient herd immunity to suppress the pandemic. The impact of vaccine effectiveness on outcomes behaved
in the expected manner, where NMB increases alongside effectiveness (Table 4, Figure 2). With 40%
effectiveness, the availability of a vaccine reduces cumulative cases by 17.1%, while with 60% effectiveness,
cumulative cases decrease by 23.7%.
Vaccine prioritisation strategies demonstrated dynamic results. Prioritizing both high risk and individuals over
the age of 55 had the smallest impact on cumulative cases and produced the second lowest value for iNB.
Equal prioritisation to all age and risk groups minimised cumulative cases and had the second greatest iNB
value. Risk prioritisation had the highest iNB value. This demonstrates the need for decision makers to pre-
determine which outcomes to target when determining vaccine allocation strategies.
The results of the worst-case scenarios demonstrate several important relationships. First, the relative ranking
of each prioritisation scenario varied depending on the outcome of interest. For example, in both worst-case
scenarios with seniors reducing contacts, and school closures, age-based prioritisation had the lowest iNB, but
risk-based prioritisation performed had the highest number of cumulative cases (Figure 3, Table 5).
Second, the impact of prioritisation strategies varies greatly depending on concurrent public health
interventions – for example, school closures and senior contact reductions have similar impacts on iNB when
there is no prioritisation given to any age or risk group (Table 5, iNB 147 vs. 120 million, respectively), but
when older and high risk groups are given priority, the benefit of school closures is much larger than reducing
contacts for seniors (Table 5, iNB 122 vs. 79 million, respectively).
Third, we found that the relative ranking of the prioritisation scenarios based on iNB changed dynamically
between the baseline scenario (50% coverage, 80% effectiveness, Table 4) and worst-case scenarios (20%
coverage, 40% effectiveness, Table 5). In the worst-case scenarios, giving equal priority to all age and risk
groups maximised iNB (iNB 147, and 121, for the school closure, and senior contact reduction scenarios
13
respectively). Conversely, prioritizing high-risk individuals had the best iNB outcome in the baseline scenario
(iNB 316 million). Age-based prioritisation, and age and risk-based prioritisation resulted in the lowest and
second lowest iNB respectively in both the baseline and worst-case scenarios.
14
Coverage, Effectiveness, and Prioritisation Scenario Results
Table 4: Scenario Results
Scenario Cumulative Cases Cumulative Hospitalisations NMB, Millions iNB*, Millions

Mean (SD) Mean (SD) Mean (SD) Mean (SD)
Reference Scenarios
No Vaccine 249,048.58 (54,172.81) 4,800.92 (1106.8) -1,025.72 (483.39) N/A

Vaccine Baseline† 176,314.71 (27,396.61) 2,740.22 (274.35) -728.67 (433.75) 297.04 (50.20)
Coverage Scenarios
Coverage 20% 205,361.22 (37,490.49) 3,529.32 (566.79) -836.49 (452.82) 189.22 (30.90)
Coverage 40%† 176,314.71 (27,396.61) 2,740.22 (274.35) -728.67 (433.75) 297.04 (50.20)
Coverage 70% 105,918.85 (48,96.808) 2,131.25 (66.170) -688.64 (421.53) 337.08 (63.14)
Effectiveness Scenarios
Effectiveness 40% 206,513.76 (37,950.19) 3,545.06 (573.79) -856.80 (453.91) 168.92 (29.82)
Effectiveness 60% 190,129.21 (32,135.95) 3,097.63 (401.96) -788.36 (443.46) 237.35 (40.39)
Effectiveness 60% age 60
and above, 80% otherwise 184,284.61 (30,055.56) 2,979.30 (356.44) -764.04 (441.09) 261.68 (42.84)
Effectiveness 80%† 176,314.71 (27,396.61) 2,740.22 (274.35) -728.67 (433.75) 297.04 (50.20)
Prioritisation Scenarios
None 107,326.29 (5,878.80) 2,398.65 (156.93) -717.94 (431.69) 307.78 (53.21)
Age 145,508.16 (18,053.95) 2,612.94 (236.95) -744.62 (431.26) 281.10 (52.90)
Risk 156,969.28 (20,239.08) 2,725.69 (242.07) -709.74 (434.56) 315.97 (49.61)
Age and Risk† 176,314.71 (27,396.61) 2,740.22 (274.35) -728.67 (433.75) 297.04 (50.20)
Note: SD = Standard Deviation, NMB = Net Monetary Benefit, iNB = Incremental Net Benefit, N/A = Not Applicable. Prioritisation Scenarios are defined as follows: None = equal
priority given to all age and risk groups, Age = priority given to those aged 55 and over, Risk = priority given to high risk individuals, Age and Risk = prioritisation for who are high
risk or older than 55. Maximum doses per day is assumed to be 80,000, with access starting on October 15, 2020 for the priority group, and 30 days later for everyone else. *For
all scenarios, iNB is estimated relative to the scenario with no vaccine. †Represent the same vaccine baseline scenario, with 40% coverage, 80% effectiveness, and age and risk
prioritisation.
15
Prioritisation Scenarios
Figure 2. Scenario Results - Active Cases and Incremental Net Benefit
16
Effectiveness Scenarios
For scenario definitions, see main text.

Coverage Scenarios
Note:
Worst-Case Scenarios: Concurrent Public Health Responses
Table 5. Worst-Case Scenario Prioritisation Policy Results
Concurrent public Prioritisation Cumulative Cases Cumulative Hospitalisations NMB, Millions iNB*, Millions
health interventions Mean (SD) Mean (SD) Mean (SD) Mean (SD)
School Closure None 166,495.86 (25,484.01) 3,473.25 (564.78) -878.85 (458.63) 146.86 (25.69)
Age 194,788.46 (35,115.05) 3,742.41 (666.09) -915.22 (461.44) 110.50 (22.34)
Risk 198,389.01 (35,880.49) 3,814.94 (672.77) -898.82 (463.00) 126.90 (20.77)

Age and Risk 192,932.31 (34,204.52) 3,739.95 (657.02) -904.16 (462.27) 121.55 (21.59)
Reduce senior None 184,031.27 (31,363.92) 3,683.29 (651.12) -905.32 (461.50) 120.39 (22.58)
contacts 50% Age 227,946.85 (46,166.87) 4,102.08 (811.38) -960.36 (466.11) 65.364 (17.42)
Risk 231,658.41 (46,913.09) 4,176.43 (816.36) -940.04 (467.74) 85.68 (15.84)
Age and Risk
225,300.54 (44,939.66) 4,094.08 (799.20) -946.78 (466.91) 78.93 (16.66)
Note: SD = Standard Deviation, NMB = Net Monetary Benefit, iNB = Incremental Net Benefit. Prioritisation Scenarios are defined as follows: None = equal priority given to all age
and risk groups, Age = priority given to those aged 55 and over, Risk = priority given to high risk individuals, Age and Risk = prioritisation for who are high risk or older than 55.
Maximum doses per day is assumed to be 80,000, with access starting on October 15, 2020 for the priority group, and 30 days later for everyone else. *For all scenarios, iNB is
estimated relative to the scenario with no vaccine.
17
Figure 3. Worst-Case Scenario Prioritisation Policy Results – Active Cases and Incremental Net Benefit
Reduce Senior Contacts by 50%
18
For scenario definitions, see main text.
School Closure
Note:
Discussion
A SEIR model, CRAM, was developed to simulate the current COVID-19 pandemic in Alberta, Canada. The model
was used to predict the impact of various vaccine strategies on total cases, hospitalisations, and deaths, as well as
estimating NMB from a health system payer perspective. To our knowledge, this is the first study to demonstrate
vaccine allocation over both age and risk groups, with consideration of concurrent public health interventions, and
population behaviours, and is the first study presenting NMB and iNB outcomes.
It is critical to consider the time-variant epidemiological conditions in which a vaccine stockpile is allocated, in
addition to specific vaccine program characteristics such as effectiveness, coverage, population prioritisation, and
date of availability. This need is highlighted through the results of the worst-case scenarios with different
concurrent strategies – the cumulative cases and iNB values vary greatly depending upon which concurrent
strategies are in place (Table 5 and Figure 3), and the main scenarios demonstrate the importance of vaccine
program characteristics (such as effectiveness and coverage, Table 4 and Figure 2).
A second imperative is to identify the target outcome to optimise, such as total mortality, morbidity, or NMB. This
is clearly demonstrated in the model results, where optimal prioritisation strategies vary between baseline and
worst-case scenarios, for total infections, hospitalisation, and NMB outcomes (Tables 4 and 5). For example, in
both worst-case scenarios with seniors reducing contacts, and school closures, age-based prioritisation performed
worst for iNB, but risk-based prioritisation performed worst for cumulative cases.
Vaccine scenarios achieve better outcomes than those without vaccine, and iNB gains are very large. While this
result indicates that the relative benefits of vaccine may justify a range of vaccine prices, our results do not
indicate that vaccine price should be set such that iNB is zero between vaccine and non-vaccine scenarios. Our
study applies a health system perspective and incurs immunisation costs, whereas a societal perspective is
expected to strictly increase NMB in immunisation scenarios relative to a health system perspective as broader
social, economic, and health impacts are considered. Furthermore, the model is run for a short time frame such
that the full benefits of immunisation are not fully realised, and important dynamics such as waning immunity are
unknown and not considered in this analysis.
Some results are unexpected warrant further explanation. The impact of a lower vaccine effectiveness in those
aged 55 and over was relatively lower than expected (Table 4). This is likely due to the expectation that physical
distancing is maintained at a higher level for seniors throughout the fall. In practice, it may be the case that public
health guidance or population attitudes shift such that seniors have increasing contact rates relative to the model
fitting period. This would impact the relative benefits of prioritizing older individuals, and therefore lower
effectiveness for older age groups would have a greater impact if guidance or behaviours changed.
A second unexpected result is the performance of the strategy where all age and risk groups are given equal
priority (Tables 4 and 5). While age and risk groups that are older or high risk have a greater risk of hospitalisation
or death from COVID, these outcomes interact dynamically with greater transmission, driven by higher rates of
contact, in young and low risk age groups. This dynamic may also be due to relatively high coverage in the
scenarios evaluated, but herd effects may not be as strong with coverage lower than 40%, as evaluated in our
baseline scenarios. Decision makers must consider the degree to which it is equitable for those at lowest risk of
poor outcomes to have equal access to a vaccine, which is a current point of debate [8-14]. Under these
considerations, equal prioritisation for all age and risk groups may represent a possible scenario which is not
acceptable to decision makers, and therefore should not be considered alongside the others. Finally, if the high-
19
efficacy preliminary trial results hold for future vaccines [33], the results of scenarios with 80% effectiveness, as in
the baseline scenario may be the most relevant.
Vaccine allocation is likely to be most contentious when the total number of doses is limited, and effectiveness is
low. The worst-case scenario simulations indicate that a vaccine prioritisation policy where all age and risk groups
are given equal initial access to vaccine maximises NMB (Table 5). When other public health policies and human
behaviours change, for example through school closures, reduced contact rates for seniors or high-risk individuals,
or in workplaces, NMB increases, but the optimal vaccine prioritisation policy does not shift. The degree to which
this finding holds will depend on the interventions in place at the time of vaccine allocation, so these scenarios
should be re-evaluated to capture policy and behaviour dynamics at the time of allocation.
Comparisons to Other Studies

While we are not aware of any other studies examining vaccine allocation within a health economic perspective, a
handful of preprint studies have used analytic methods to test outcomes under different vaccine allocation
scenarios [13, 15-17]. The results from the CRAM align with the other studies: the age-prioritised scenario
minimises hospitalisations, while the unrestricted scenario minimises total infections (Table 4). Like [17], who
project that coverage between 40-59% of the population will be required for suppression in Canada, the CRAM
demonstrates that infections are significantly reduced with coverage of 70%, when age and risk prioritisation is
implemented.
Several elements of CRAM compare favourably to the other studies. With the exception of Chen et al. [13], other
studies do not consider concurrent or future public health policies/ and or behavioural factors. This appears to be a
large oversight given that any allocation decision must not only consider the epidemiology at the time of
allocation, but also which other interventions might be used to amplify the impact of targeted vaccine strategies.
CRAM allows allocation to both age and risk groups, which are not evaluated together in many other studies [13,
15-17]. It is likely that governments would wish to consider allocation to high risk individuals rather than broad age
groups, and high-risk individuals are likely to take greater infection prevention measures, so these dynamics are
important to incorporate. Furthermore, most studies are calibrated to published R0 figures. Because the CRAM is
fit to surveillance data, model predictions for a broader range of outcomes (infections, hospitalisations, deaths) are
provided with greater accuracy. Finally, model inputs to the CRAM were informed wherever possible with
jurisdictional data, and expert opinion was used rather than using assumed values, as in the studies discussed.
There are also elements where other approaches are strong relative to CRAM. Two studies [12, 15] consider
vaccines which are both ‘leaky’ (reducing the probability a susceptible individual is infected) and ‘all-or-none’ (as
in the CRAM). Meehan et al. [17] evaluate vaccines which either prevent infection, or prevent disease. Bubar et al.
[15] compare scenarios where serological testing is included in vaccine scenarios, or not. While inclusion and
evaluation of these elements is beyond the scope of this paper, it is reasonable to assume that the relative
dynamics demonstrated by other authors (for example relative outcomes under leaky or all-or-none vaccine
assumptions) will hold for the results of the CRAM.
Strengths
Our approach has several strengths relative to other modelling studies. The direct interpretation of the NMB from
a health system payer perspective provides important insights. Generally, all measures that can reduce the spread
20
of COVID-19 have a large social value, due to the very high economic losses it causes. [35] A general expectation is
that economic and social outcomes will be relatively worse under more restrictive public health responses (such as
business and school closures), while direct health outcomes will improve. The inclusion of broader health
outcomes is undetermined because outcomes like excess mortality and morbidity, due to health system avoidance
or deferral of services will have dynamic impacts on NMB under various public health responses. Interventions
such as vaccine are expected to increase NMB from a societal perspective, limited only by the cost of vaccine
procurement and delivery. By estimating the INB from a payer perspective for each strategy, the minimum NMB of
economic and social impacts to indicate a different policy direction are made explicit.
The structure of CRAM was developed to align with available Alberta surveillance data, namely the number of
hospitalisations and cumulative infections on key dates indicated by seroprevalence studies. While case
classification in each healthcare setting provides unique challenges and remains subject to ascertainment and
reporting biases, these measurements can be estimated with the greatest amount of certainty, and therefore both
the fitting and compartmental structure of CRAM are developed around these data, mitigating these issues to the
greatest extent possible. Laboratory-confirmed cases and asymptomatic infections are not included within the
model structure due to data quality. Laboratory testing protocols have changed rapidly since March 2020 [36], and
therefore incorporating the number of laboratory-confirmed cases into the model structure would have been
inappropriate.
An important feature of CRAM structure is that it allows users to evaluate a broad and flexible range of concurrent
interventions. Our results demonstrate that optimal vaccine allocation changes depending on the public health
measures (like school closures) or behavioural changes (contact reductions for seniors) as well as vaccine program
features. By creating a structure that captures relative changes in behaviours and policies by age and risk group
and by location over time, our model is designed to capture dynamic population changes to the greatest degree
possible.
Limitations
A key challenge for all COVID-19 modelling and simulation studies has been data quality. The challenge is two-fold:
model inputs and parameters (such as the duration of infection and transmissibility by age group) are uncertain
and estimated from small observational studies. In addition, jurisdictional observational data are subject to bias.
The high degree of uncertainty in many of the model elements necessitates constant review and reappraisal of the
model structure and inputs. Constant evolution of the evidence also highlights the importance of the stochastic
approach used by this study. By sampling from the distributions of several parameters, the stochastic results
quantify the resulting uncertainty in model outputs.
Seasonal dynamics related to COVID-19 remain controversial. Several studies have found empirical and modelling
evidence to support changes in virus transmission consistent with seasonal behaviour, and these seasonal changes
in transmission can be driven by both behavioural and environmental factors. A review on the stability of
coronaviruses including COVID-19 under various environmental and climatic conditions [37] found early evidence
of virus sensitivity to temperature, light, and humidity in laboratory studies [38-42]. Many studies analysing the
natural history of the virus imply the presence of seasonal dynamics [43-50], while others found a limited effect
[51, 52]. Natural history studies are challenged with data quality issues, confounding, and a very short period of
observation [53]. There is overall agreement that the transmission of COVID-19 is likely to demonstrate some
seasonal behaviour, but that dynamics are complex due to the high susceptibility of the global population and that
effective pandemic control cannot rely on seasonal dynamics. As demonstrated by the findings of a rapid literature
review [54], the evidence around relative differences in susceptibility and transmission by age is highly uncertain,
21
in particular for children. None of the studies identified through the rapid review included parameters suitable for
inputting into CRAM. A modelling study estimating relative susceptibility by age group was instead used as a
source for these inputs [23]. Like any of the inputs, if evidence were to emerge that the dynamics of susceptibility
by age are different from those used to fit our model, there would be implications for the validity of these results.
Due to the widespread impacts of COVID-19 on population health, economy, and social outcomes, a societal
perspective is likely to be the most appropriate perspective for economic evaluations of COVID-19 response
policies. The estimates from a health system payer perspective generated in this study do not capture the full
health, social, and economic impacts of the COVID-19 pandemic and response. Accurate estimates of the
additional societal impacts are unlikely to become available in the near future. In this context, our analysis
provides an important baseline estimate of the cost-effectiveness of vaccines which future evaluation can build
upon. As additional research estimates the broader economic impact of different public health policies (for
example, business or school closures) these may be layered on to estimates from the CRAM, the flexible structure
of which allows the estimation NMB for a broad range of strategies.
Community infection comprises multiple disease states, which are not separated due to data quality. There is no
reliable estimate of the number of infected individuals in the community for Alberta, as current testing protocols
and changes to testing eligibility undermine the estimation of the number of infected individuals over time.
Similarly, evidence on the proportion of infected cases which are asymptomatic is rapidly evolving [55]. Due to the
absence of these inputs, community cases are not differentiated by the presence of symptoms. Related to this is
the exclusion of costs related to community infections, which could be included in future work. We did not
consider non-sterilizing vaccines, which would require use of an asymptomatic compartment. Finally, CRAM
structure does not allow for the isolation of close contacts, limiting the range of isolation and quarantine scenarios
that can be evaluated.
Policy Implications
The results of this paper have three key policy implications: (i) that optimal vaccine allocation will depend on the
public health policies, and human behaviours in place at the time of allocation; (ii) the outcomes of vaccine
allocation policies can be greatly supported with interventions targeting contact reduction in critical sub-
populations; and (iii) the identification of the optimal strategy depends on which outcomes are prioritised.
In allocating vaccine access, decision makers must not only consider the interventions and behaviours in place at
the time of allocation, but also those which could be in place in the future. Optimal allocation will depend on
public health policy and human behaviour, and therefore decision makers must understand that allocation
decisions are made within a dynamic environment. The value of vaccination may be improved by adjustments to
complementary public health policies.
Decision makers should define the range of public health policies that can be implemented alongside vaccine
allocation within an ethical framework. For example, in our model NMB will be maximised if seniors continue to
reduce their contacts through the fall, and vaccines were allocated without prioritizing older and high-risk
population groups. This would mean that the people with the greatest risk of disease and poor outcomes, would
also bear the highest burden of additional social isolation. Because of this, the suite of public health policy options
should be determined prior to dynamic policy optimisation, to ensure that the optimal strategy is achieved within
a framework reflective of social values and equity [8].
22
Decision makers must evaluate a broad range of outcome measures to prioritise and provide robust justification
for their choice. This paper uses NMB because it captures the relative disutility for epidemiologic outcomes (total
cases, hospitalisations, and deaths), as well as health system resource implications due to illness. In this way,
multiple factors such as age and outcome severity are built into the measure. NMB will also allow for simple and
direct inclusion of societal outcomes, if and when they become available.
Future Research
Ongoing research and observational data should continue to be incorporated into the model, through re-fitting
data, and adjusting model structure where evidence indicates important new relationships and dynamics. The
results of this paper indicate that outcomes are highly responsive to behavioural factors such as changes in contact
rates. This indicates that while the results of this paper provide important insights into the general allocation
problem, model reanalysis with the most up to date available information is critical to optimal strategy
identification.
While the results presented in this study pertain to a model calibrated to Alberta, Canada, the key findings are
relevant to many other jurisdictions. The approach demonstrates how the net benefit approach, implemented
through mathematical modelling, can support a principled and defensible vaccine allocation within a dynamic
policy environment.
23
Appendix 1 Model Details

Equations
The full set of equations used to describe CRAM is provided in this appendix. Each compartment is divided into 17
age groups and two risk groups. The 34 sub-compartments are indexed by a,r, where a is the age group and r is the
risk group. The population size of each group is given by , and a compartment sub-indexed by ,· represents
the sum of both risk groups for one age group.

,
,
,· ,· ,· ,· ,· , ,, , ,, ,
,·

,
, ,, , ,· ,· ,· ,· ,· , ,, ,
,·

,
, ,, , ,, ,
,· ,· ,· ,· ,· ,
,·

,

,
, !1 #$ ,
,· ,· ,· ,· ,·
,·

,
#
,
,· ,· ,· ,· ,· ! , , , $
,·

,
,· ,· ,· ,· ,· , # ,
,·

, # , ,

, %1 & , ' , ( !1 ) $ , ,

, %1 & , ' , ( !) $ , ,

, , ,

, ' , , ,

, & , , ,
*
, # , !1 +1 $ ! , , $ !1 +2 $ , !1
+3 $ ,
.
, +1 ! , , $ +2 +3 ,
24
Some parameters are functions, as in the following.

, 0
where is the age group and , are fitted parameters from an exponential regression.

,
where is the risk group index.

2

365 0.19# $1 %
&
Where is the numeric constant Pi, is the date, and
controls the amplitude of the seasonal effect, larger
numbers will increase the difference in transmission maximum and minimum values. The shift of 0.19 centres the
minimum on July 30 , the centre date between the hottest days of the year in Edmonton and Calgary.
th
The contact matrix is decomposed into four locations: home, workplace, school, and other, where other comprises
all remaining possible locations. Total contacts are given as the sum of contacts for each age to age (a,k) entry in
each of the matrices. Decomposing the total contact matrix by location allows for adjustments in contact rates at
locations to align with policy interventions. For example, school closure is reflected by adjusting the school contact
matrix ('( ,, ) by a scalar value, for the duration of dates during which the school is closed:
'( ,, ) '( , ,

where ) 0 while schools are closed, and ) 1 otherwise (that is, 0 * ) * 1).
Similarly, for interventions which would reduce contact rates in schools, ), may be applied to specific age
groups, adjusting the school contact matrix at various points in time according to age group.
Reductions ) , ) , and ) are applied to the remaining contact matrices, which are then summed
to generate the total matrix, '' , , at each time step. Reduction in contacts for seniors age 75+ are applied
directly to the total contact matrix, as a scalar multiplied through the rows and columns where a or k equals 16.
Reduction in contacts for high-risk individuals are applied to the total contact matrix, only for those individuals
whose r equals 2. For example:
,
+ ∑!
" )
,·
-./,· -,· -01,· -2,· -1,· / ,
where ) 1 for non-high-risk individuals.
Assumptions
The CRAM makes the following assumptions:
• Only susceptible individuals who have not been exposed to COVID-19 are immunised. In practice, this may
require serology testing prior to vaccine.
• The vaccine behaves as an all-or nothing vaccine, where the vaccine protects a proportion of those
immunised corresponding to vaccine effectiveness.
25
• Daily contact rates are based on age group in CRAM. This limits the types of contacts represented in the
model; for example, the model cannot represent cohorts of students (that is, students seeing the same
limited group of students/staff each day).
• There are no additional school contacts between individuals older than 20 years in our model (for
example, between school staff). Therefore, when contacts were increased with in-person schooling,
increases in contact rates only applied to contacts between students and other students, and contacts
between students and adults.
• All public health interventions were modelled by reducing the contacts of specific age or risk groups. This
necessitates assumptions about contact reductions associated with each public health intervention (for
example, how masks reduce contacts between school-aged children). Therefore, CRAM cannot test which
public health interventions are most likely to reduce contacts; rather, it estimates the impact that
reducing contacts through public health interventions may have on overall disease outcomes.
• Provincial level modelling is broadly informative at a local level. Where local case counts increase, this
may indicate that additional actions need to be taken, including possible changes to schooling
arrangements, but local factors relative to the provincial dynamics presented in these results should be
taken into context.
• All strategies in place at the start of the period, other than vaccine and other clearly stated changes will
remain in place for the duration of the model period. Therefore, the results reflect what might happen in
a scenario where no mitigating actions are taken. Any interventions implemented at a local or regional
level in response to an outbreak would result in a different number of expected infections and
hospitalisations.
26
Appendix 2 Expert Elicitation

To generate estimates for parameters for which there was no information source, a rapid expert elicitation
exercise was developed. The elicitation drew from the SHELF tool and adapted the questionnaire to be given
through a web-based survey. Expert opinions were pooled to derive the median and 90% CI estimates for each
value of interest.
An initial set of questions was disseminated to public health experts, identified in partnership with Alberta Health.
Experts were provided with a short online training module, and responses were restricted to a 11-point Likert scale
between 0 and 10. The median, and upper and lower tertiles were elicited, and beta distributions were fit to
estimate response values.
A second elicitation was disseminated to public health and education experts, using two return to school scenarios
defined by the Alberta Government. Experts were provided with a webinar training, and then the questionnaire
was disseminated online, with responses restricted to an 11-point Likert scale as before. The median, 5 and 95 th th
quantiles were elicited, and beta distributions were fit to estimate response values.
Note that following the expert elicitation, Alberta Health mandated mask use in schools. An additional reduction in
contact rates using a uniform distribution ranging from 20-80% relative to the elicited values was applied to the
school contact matrix. This value was based on a rapid review which found evidence of mask effectiveness at
reducing spread, but uncertainty in the degree of effectiveness [56].
TABLE E.1: Elicitation question and responses
Quantity Pooled response values, median Elicited values and linear pool
[90% CI]
Factor by which 0.586 [0.070, 0.995]
Albertans have
reduced their
contacts at home,
given current public
health orders and
guidance
27
[90% CI]
Factor by which 0.387 [0.137, 0.945]
Albertans aged ≥75
years have reduced
their contacts in all
locations, given
current public health
orders and guidance
Factor by which 0.326 [0.065, 0.922]
Albertans with a high
risk of severe
outcomes have
reduced their
contacts in all
locations, given
current public health
orders and guidance
Scenario 1
Contact reduction –
Age group: 5 9 years
Age
Contact reduction
group
5–9 0.692 [0.358, 0.869]
years
10–14 0.573 [0.314, 0.846]

years
15–19 0.522 [0.278, 0.861]

years
Age group: 10 –14 years
28
[90% CI]
Age group: 15 –19 years
Proportion returning –
Age
Proportion returning
group
5–9 0.760 [0.399, 0.933]
years
10–14 0.831 [0.399, 0.941]

years
15–19 0.860 [0.634, 0.994]

years
–
29
[90% CI]
–
Scenario 2
Contact reduction –
Age
Contact reduction
group
5–9 0.406 [0.174, 0.634]
years
10–14 0.358 [0.142, 0.664]

years
15–19 0.333 [0.199, 0.611]

years
–
30
[90% CI]
–
Proportion returning –
Age
Proportion returning
group
5–9 0.800 [0.590, 0.992]
years
10–14 0.812 [0.511, 0.987]

years
15–19 0.877 [0.691, 0.991]

years
–
31
[90% CI]
–
CI: confidence interval
32
References
1. WHO Coronavirus Disease (COVID-19) Dashboard.
2. COVID-19 Alberta statistics.
3. Who, The push for a COVID-19 vaccine. 2020.
4. Hassoun, N., How to Distribute a COVID-19 Vaccine Ethically, in Scientific American.
2020.
5. Kupferschmidt, K., WHO unveils global plan to fairly distribute COVID-19 vaccine, but
challenges await. 2020, Science.
6. Weintraub, R., P. Yadav, and S. Berkley, A Covid-19 Vaccine Will Need Equitable, Global
Distribution, in Harvard Business Review. 2020.
7. Gayle, H., et al., Framework for Equitable Allocation of COVID-19 Vaccine. 2020,
Washington (DC): National Academies Press (US).
8. Roope, L.S.J., et al., How Should a Safe and Effective COVID-19 Vaccine be Allocated?
Health Economists Need to be Ready to Take the Baton. Pharmacoecon Open, 2020.
9. Persad, G., M.E. Peek, and E.J. Emanuel, Fairly Prioritizing Groups for Access to COVID-19
Vaccines. JAMA, 2020.
10. Emanuel, E.J., et al., An ethical framework for global vaccine allocation. Science, 2020.
369(6509): p. 1309-1312.
11. Emanuel, E.J., et al., Fair Allocation of Scarce Medical Resources in the Time of Covid-19.
N. Engl. J. Med., 2020.
12. Buckner, J.H., G. Chowell, and M.R. Springborn, Optimal Dynamic Prioritization of Scarce
COVID-19 Vaccines. medRxiv, 2020.
13. Chen, X., et al., Allocation of COVID-19 Vaccines Under Limited Supply. 2020.
14. Liu, Y., S. Salwi, and B.C. Drolet, Multivalue ethical framework for fair global allocation of
a COVID-19 vaccine. J. Med. Ethics, 2020. 46(8): p. 499-501.
15. Bubar, K.M., et al., Model-informed COVID-19 vaccine prioritization strategies by age
and serostatus. Infectious Diseases (except HIV/AIDS), 2020.
16. Matrajt, L., et al., Vaccine optimization for COVID-19, who to vaccinate first? medRxiv,
2020.
17. Meehan, M.T., et al., Age-targeted dose allocation can halve COVID-19 vaccine
requirements. Infectious Diseases (except HIV/AIDS), 2020.
18. Stinnett, A.A. and J. Mullahy, Net health benefits: a new framework for the analysis of
uncertainty in cost-effectiveness analysis. Med. Decis. Making, 1998. 18(2 Suppl): p. S68-
80.
19. Prem, K., A.R. Cook, and M. Jit, Projecting social contact matrices in 152 countries using
contact surveys and demographic data. PLoS Comput. Biol., 2017. 13(9): p. e1005697.
20. Google, COVID-19 Community Mobility Reports.
21. O’Hagan, A. and J. Oakley, SHELF: The Sheffield Elicitation Framework. 2013.
22. Grassly, N.C. and C. Fraser, Seasonal infectious disease epidemiology. Proc. Biol. Sci.,
2006. 273(1600): p. 2541-2550.
23. Davies, N.G., et al., Age-dependent effects in the transmission and control of COVID-19
epidemics. Nat. Med., 2020.
33
24. Public Health Agency of, C., COVID-19: PHAC Modelling Group Weekly Report. 2020.
25. Ochalek, J., J. Lomas, and K. Claxton, Assessing health opportunity costs for the Canadian
health care systems.
26. Cadth, Guidelines for the Economic Evaluation of Health Technologies: Canada. 2017.
27. Fragaszy, E.B., et al., Effects of seasonal and pandemic influenza on health-related
quality of life, work and school absence in England: Results from the Flu Watch cohort
study. Influenza Other Respi. Viruses, 2018. 12(1): p. 171-182.
28. Mangen, M.-J.J., et al., The impact of community-acquired pneumonia on the health-
related quality-of-life in elderly. BMC Infect. Dis., 2017. 17(1): p. 208.
29. Briggs, A., et al., Estimating (quality-adjusted) life-year losses associated with deaths:
with application to COVID-19. Health Economics Letters - Under Review, 2020.
30. Alberta, H., Interactive Health Data Application - Health Costing Data.
31. Government of Canada, S.C., CANSIM Table 18-10-0005-01.
32. van Rijt, A.M., et al., Cost analysis of outbreaks with Methicillin-resistant Staphylococcus
aureus (MRSA) in Dutch long-term care facilities (LTCF). PLoS One, 2018. 13(11): p.
e0208092.
33. Callaway, E., Why Oxford's positive COVID vaccine results are puzzling scientists. Nature,
2020.
34. Dover, D.C., et al., Pandemic Risk Assessment Model (PRAM): a mathematical modeling
approach to pandemic influenza planning. Epidemiology and Infection, 2016. 144(16): p.
3400-3411.
35. Cutler, D.M. and L.H. Summers, The COVID-19 Pandemic and the $16 Trillion Virus.
JAMA, 2020.
36. Alberta Health, S., Laboratory Bulletins | Alberta Health Services.
37. Aboubakr, H.A., T.A. Sharafeldin, and S.M. Goyal, Stability of SARS-CoV-2 and other
coronaviruses in the environment and on common touch surfaces and the influence of
climatic conditions: a review. Transbound. Emerg. Dis., 2020.
38. Chin, A.W.H., et al., Stability of SARS-CoV-2 in different environmental conditions.
39. Kampf, G., et al., Persistence of coronaviruses on inanimate surfaces and their
inactivation with biocidal agents. J. Hosp. Infect., 2020. 104(3): p. 246-251.
40. van Doremalen, N., et al., Aerosol and Surface Stability of SARS-CoV-2 as Compared with
SARS-CoV-1. N. Engl. J. Med., 2020. 382(16): p. 1564-1567.
41. Setti, L., et al., SARS-Cov-2 RNA Found on Particulate Matter of Bergamo in Northern
Italy: First Preliminary Evidence. Infectious Diseases (except HIV/AIDS), 2020.
42. Wang, J., et al., High Temperature and High Humidity Reduce the Transmission of COVID-
19. 2020.
43. Sajadi, M.M., et al., Temperature, Humidity, and Latitude Analysis to Estimate Potential
Spread and Seasonality of Coronavirus Disease 2019 (COVID-19). JAMA Netw Open,
2020. 3(6): p. e2011834.
44. Vinoj, V., et al., The COVID-19 Spread in India and Its Dependence on Temperature and
Relative Humidity. other, 2020.
45. Mandal, M., et al., Study of the Dependence of Effective Reproduction Number of COVID-
19 on the Temperature and Humidity: A Case Study with the Indian States. Epidemiology,
2020.
34
46. Notari, A.,Temperature dependence of COVID-19 transmission. Epidemiology, 2020.

47. Ficetola, G.F. and D. Rubolini,Climate affects global patterns of COVID-19 early outbreak
dynamics. Infectious Diseases (except HIV/AIDS), 2020.
48. Merow, C. and M.C. Urban, Seasonality and uncertainty in COVID-19 growth rates.
Epidemiology, 2020.
49. COVID-19 transmission in Mainland China is associated with temperature
Qi, H., et al.,
and humidity: A time-series analysis. Sci. Total Environ., 2020. 728: p. 138778.
50. Susceptible supply limits the role of climate in the early SARS-CoV-2
Baker, R.E., et al.,
pandemic. Science, 2020.
51. Briz-Redón, Á. and Á. Serrano-Aroca, A spatio-temporal analysis for exploring the effect
of temperature on COVID-19 early evolution in Spain. Sci. Total Environ., 2020. 728: p.
138811.
52. Impact of climate and public health interventions on the COVID-19
Jüni, P., et al.,
pandemic: a prospective cohort study. CMAJ, 2020. 192(21): p. E566-E573.
53. Effective transmission across the globe: the role of climate in
O'Reilly, K.M., et al.,
COVID-19 mitigation strategies. Lancet Planet Health, 2020. 4(5): p. e172.
54. Covid 19 Transmission From and Infection of Children.
Warkentin, L., et al., 2020.
55. Byambasuren, O., et al., Estimating the extent of asymptomatic COVID-19 and its
potential for community transmission: Systematic review and meta-analysis. JAMMI,
2020.
56. The Effectiveness of Community Face Mask Use for the Prevention
Warkentin, L., et al.,
of Covid 19 Transmission and Infection. 2020.
35

A Net Benefit Approach For The Optimal Allocation of A COVID-19 Vaccine

Uploaded by

Copyright:

Available Formats

A Net Benefit Approach For The Optimal Allocation of A COVID-19 Vaccine

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

A Net Benefit Approach For The Optimal Allocation of A COVID-19 Vaccine

Uploaded by

Copyright:

Available Formats

medRxiv preprint doi: https://doi.org/10.1101/2020.11.30.20240986; this version posted December 2, 2020.

The copyright holder for this preprint

A Net Benefit Approach for the Optimal

Figure 1. CRAM Structure

Note: Figure intended for colour printing.

Table 1. Epidemiologic Parameters

Symbol Parameter Baseline value Description and

individuals on model day 1.

transmission modified, as given in Appendix C.

year in Edmonton and Calgary.

In this report,  = 0.3, which solves to 0.4 at its minimum,

meaning that transmission at the seasonal minimum is 40% of

transmission at the seasonal maximum.

 Susceptibility by 0.489 if aged <10 years, 1 otherwise. Estimates taken

age group from a modelling

schools baseline scenario. exercise.

return to school (median and 90% CI):

Age group Contact reduction Proportion

10–14 years 0.573 [0.314, 0.846] 0.831 [0.399, 0.941]

15–19 years 0.522 [0.278, 0.861] 0.860 [0.634, 0.994]

homes at time t exercise.

at time t contact rates. combined time

seniors 75 and above. exercise.

reduction for distancing for high-risk individuals after 15 March. elicitation

, , Proportion of User input day for scenario analysis. User/scenario

community seroprevalence studies conducted in Alberta. Note that the input.

infections Isolation compartment represents case isolation only.

  Latent period Mean 3.3 days. PHAC Modelling

maximum values identified in the PHAC report (21.0 and 33.0, as

 Days in hospital Mean 8.08 days. Alberta Health

  Days prior to 5 Assumed.

  Days prior to 14 Assumed.

 OR of hospital 6.304 (95% CI [4.973, 7.991]). Alberta Health

for high-risk data.

  OR of death for 42.384 (95% CI [22.449, 80.018]). Alberta Health

 Fitting term for Fitted.

risk by age converted to proportion by age and analysed in an exponential data.

regression for model input.

Fitted values: Max[ 6.346  0.3093" #, 0$

 Fitting term for Fitted.

*, Probability of Function of &) .

+- Probability of Function of ., '( , /2.

death from LTC

+1 Probability of Function of ., '( , /3.

 Fitting term for Fitted.

2- LTC proportion 116/474 Alberta Health

21 Hospital 93/460 Alberta Health

poor outcome outcome related to COVID-19.

High-risk is defined as individuals who have met an administrative

I25), respiratory diseases including asthma (ICD-10 codes: J45, J46)

and COPD (ICD-10 codes: J41–J44).

A subset of the high-risk population aged ≥70 years is shifted to a

isolation policies applied.

Table 2. Economic Parameters

Parameter Name Baseline Value Source and Description

Hospitalisation, ICU 0.58. QALY loss at admission multiplied by duration in the

admission, LTC case infectious compartment. SE or other measures can not

16-65 (00-10.83), 65+ (1.84-7.17).

0-9 68.96 (61.45-76.47) standard deviations. Values represent net present

In this report, = 0.3, which solves to 0.4 at its minimum,

Susceptibility by 0.489 if aged <10 years, 1 otherwise. Estimates taken

, , Proportion of User input day for scenario analysis. User/scenario

Latent period Mean 3.3 days. PHAC Modelling

Days in hospital Mean 8.08 days. Alberta Health

Days prior to 5 Assumed.

Days prior to 14 Assumed.

OR of hospital 6.304 (95% CI [4.973, 7.991]). Alberta Health

OR of death for 42.384 (95% CI [22.449, 80.018]). Alberta Health

Fitting term for Fitted.

Fitted values: Max[ 6.346 0.3093" #, 0$

Fitting term for Fitted.

*, Probability of Function of &) .

+- Probability of Function of ., '( , /2.

+1 Probability of Function of ., '( , /3.

Fitting term for Fitted.

where is the risk group index.

'( ,, ) '( , ,