Preprint Not Peer Reviewed: Modeling Vaccine Allocations in The COVID-19 Pandemic: A Case Study in Australia

Modeling Vaccine Allocations in the COVID-19 Pandemic: A Case Study in
Australia
ed
a,∗ a a b a
Babak Abbasi , Masih Fadaki , Olga Kokshagina , Naima Saeed , Prem Chhetri
a College of Business and Law, RMIT University, Melbourne, VIC 3000, Australia
iew
b School of Business and Law, University of Agder, Kristiansand, Norway
Abstract
ev
Timely allocation and distribution of COVID-19 vaccines on a large scale is a highly complex, dynamic, and
context-specific task. The focus of this research on optimizing vaccine allocation is on the downstream part
of a COVID-19 vaccine supply chain. Previous research on vaccine supply chains and pandemic supply chains
r
has not fully incorporated the multitude of factors and underlying constraints affecting a vaccine supply chain
which can be optimized to mitigate the risk of infection. An effective model is needed to conceptualize the
er
process of the downstream vaccine supply chain to ensure efficient coordination and timely distribution of vaccines
to the population. This paper develops a mathematical model to support vaccine allocation decisions based
pe
on exposure risk, susceptibility rate, and operational constraints including capacity of medical centers, vaccine
stocks, and transshipment capacity. Our conceptual model integrates a centralized booking system, risk profiling
and prioritization, and a vaccine distribution system, to develop an effective vaccine allocation model based on the
parameters of the total population susceptible to COVID-19 and the density-based exposure risk in the catchment
of each medical center. We have incorporated the possibility of transshipment between medical centers and a
ot
variety of different vaccine package sizes. Using the state of Victoria, Australia as a case study, we applied
the proposed model to test different scenarios of vaccine allocation and distribution. This research proposes
tn
specific guidelines for COVID-19 vaccine distribution and makes recommendations on how healthcare providers
and government entities should work together to establish more efficient logistical capabilities.
Keywords: COVID-19 Pandemic, Vaccine Supply Chain, Allocation Decisions, Decision-Making

rin
1. Introduction
“Getting a coronavirus vaccine in record time is hard. Distributing it to tens of millions may be equally daunting”
ep
- Washington Post, Aug, 4, 20201 . An innovative and robust pandemic vaccine supply chain needs to be designed
and developed to tackle the daunting task of mass vaccination under stringent operating constraints. Pandemics
∗ Correspondence to [email protected]
Pr
Email addresses: [email protected] (Babak Abbasi ), [email protected] (Masih Fadaki ),

[email protected] (Olga Kokshagina ), [email protected] (Naima Saeed ), [email protected] (Prem
Chhetri )
1 https://www.washingtonpost.com/health/2020/08/03/coronavirus-vaccine-distribution-confusion/
1
This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=3744520
such as the coronavirus disease (hereafter COVID-19) exert severe pressure on healthcare systems, which in turn
affects timely delivery and distribution of vaccines to healthcare centers. Most governments are responding to
ed
this distribution challenge by building or upgrading healthcare infrastructure to enhance geographic accessibility
of health services. Nevertheless, it is equally important to design an optimal vaccine supply chain that supports
an effective, agile, and responsive distribution network to maximize geographic coverage of populations at greater
iew
risk while keeping distribution lean. Evidence-based decision-making to help optimize and allocate vaccines in a
timely manner is critical to protect lives during the COVID-19 pandemic. Health organizations are calling for
novel approaches and methods to optimize immunization supply chains and meet the demands of an increasingly
large and costly portfolio of vaccines [1].
ev
COVID-19 is an infectious disease which was initially reported in December 2019 in Wuhan, Hubei Province,
China and is now transmitting globally at an alarming rate [2]. It is caused by severe acute respiratory syndrome
coronavirus 2 (SARS-CoV-2). The WHO declared COVID-19 a world pandemic on 11 March 2020 [3], with 213
countries currently impacted by the pandemic. COVID-19 is highly contagious compared to previous pandemics
r
[4] and it has dramatically hindered production in entire industries and exposed the fragility of global supply
er
chains. The economic impacts of the pandemic are severe and have caused contraction of GDP and widespread
job losses [5]. The long-term economic and societal impacts are projected to be significant, but are not yet properly
understood. In the initial stage of COVID-19 diffusion, Western Europe was severely impacted, and in particular
pe
the older population with a weak immune system suffered the most. Currently, USA, India, and Brazil are the
epicenters of the pandemic. With a global death toll approaching 1 million in September 2020, WHO warns of
another 1 million deaths in the coming months.
Many agencies are simultaneously working to develop and manufacture a vaccine to effectively cease or at least
ot
drastically decelerate the spread of the pandemic [2]. Most governments have imposed precautionary measures,
such as self-isolation, wearing of masks, and social distancing, to control the diffusion of COVID-19. However, in
the long run, it is necessary to produce a vaccine and, even more importantly, ensure that it is effectively distributed
tn
and delivered to the populace [6]. As the National Governors Association Center for Best Practices [6] puts it,
“Immunizing the U.S. population against COVID-19 will likely require the single largest vaccination campaign
ever undertaken and governors will play a key role in bringing together leaders from their state public health,
rin
immunization, and emergency management systems to design and execute the operation.” Given the complexity
of global vaccine supply chains and the constraints related to supply, demand, and capacity, various distribution
scenarios should be formulated to help optimize the system for acquiring, prioritizing, and distributing vaccines
to the populace [7, 8, 9, 10, 11, 12, 13, 14, 15, 16]. Despite the prevalence of epidemics and pandemics, only a few
ep
studies have designed and built an efficient distribution network to ensure that vaccines are effectively delivered to
the neediest people (see [8, 9, 10, 13, 14, 15, 16]). However, not even these studies fully considered the multitude of
factors and constraints affecting a vaccine supply chain which can be optimized to mitigate the risk of infection. A
Pr
robust model is therefore needed to conceptualize the process of the downstream vaccine supply chain and ensure
efficient coordination and distribution of vaccines to the population. Regardless of the type of deployment system,
it is critical to track supply, manage allocation, mitigate adverse events, and ensure transparency.
2
This research develops a mathematical model to support vaccine allocation decisions based on risk of exposure,
susceptibility, and the capacity constraints of medical centers. The focus of this research is on the downstream
ed
part of the COVID-19 vaccine supply chain, to help optimize vaccine allocation to inoculate people with different
priority levels. In this study we assume: that the vaccine is already manufactured and available in the market
for distribution; that limited supply and excessive demand necessitate optimizing the allocation of vaccine by
iew
prioritizing people with higher risk of infection and greater probability of social contact with others; and that the
vaccine is to be administered at state-run medical centers.
This study makes several conceptual and methodological contributions. The first is the conceptual model that
integrates all the key components of a vaccine supply chain, including a centralized booking system to help capture
ev
demand variability, a risk-driven prioritization of recipients, and a vaccine distribution system to optimally allocate
and effectively deliver vaccines to medical centers. Secondly, the model incorporates both susceptibility risk and
exposure risk in order to optimize vaccine allocation. Susceptibility refers to the quality or condition of being
susceptible to a pathogen or familial disease. Here it is the state of being predisposed or sensitive to COVID-19.
r
People are susceptible when they become a conduit for person-to-person transmission of infection. An exposure is
er
the likelihood of a contact between an individual and a potential hazard, i.e. an infected person or contaminated
surface. Risk of exposure is calculated from the ratio of the total population represented by a priority group
assigned to a medical center and the total population of the same priority group for all medical centers. Thirdly,
pe
the possibility of transshipment between medical centers is embedded in our model to improve the effectiveness
of vaccine delivery between over-supplied and under-supplied demand points. Fourth, the delivery of vaccines
in different package sizes is also considered, to optimize and improve distribution efficiency. We argue that the
package size is an important decision variable that has to be negotiated by governments and providers prior to
supplying vaccines.
ot
Using the state of Victoria as a case study, we apply the proposed model to test vaccine allocation and distribution.
This research develops specific guidelines for COVID-19 vaccine distribution and makes recommendations on how
tn
healthcare providers and government entities should work together to establish more efficient logistical capability.
In particular, our results indicate that including the possibility of transshipment and of different package sizes
can have a significant effect on the overall time needed for vaccine administration. We anticipate important
rin
implications for vaccine allocation and transshipment processes if pack sizes are allowed to be determined by
manufacturers alone. Furthermore, the capacity of medical centers will have a considerable impact on residual
risk and could also impact the vaccine administration period.
ep
The remainder of this paper is organized as follows. Firstly, we outline previous work on vaccine supply chains and
pandemic vaccine supply chains. Secondly, we introduce our conceptual model and the underlying mathematical
model. Thirdly, we present the results of the model and the key contributions and implications of this research.
Pr
3
2. Literature review
ed
Effective distribution of vaccines to the populace is paramount for risk mitigation of community transmission in
a pandemic. A typical vaccine supply chain involves manufacture, packaging, storage, and domestic and global
distribution, aiming to ensure cost-effective, timely, and uninterrupted supply of vaccines to the populace [17].
The difference between a typical vaccine supply chain and a pandemic vaccine chain is that in the former, vaccine
iew
providers (i.e. healthcare service providers) buy vaccines directly from manufacturers or via wholesalers and
distributors. In the case of a pandemic, government agencies are more likely to purchase all vaccines directly
from manufacturers to ensure quick and early vaccination. Once these vaccines are procured, government agencies
distribute them to public health centers to be administered. In a pandemic vaccine supply chain, healthcare service
ev
providers, including pharmacists, often collaborate with the government and register their interest under public
health programs instead of collaborating with supply chain vendors to procure pandemic vaccines [17]. Brown
et al. [18] conceptualized a typical pandemic vaccine supply chain as a four-level delivery system that integrates
r
the national depot, departmental stores and one regional store, commune stores, and hundreds of health posts.
However, they restructured the vaccine supply chain based on the following assumptions: (a) the commune level is
er
consolidated to a health zone level; (b) the commune level is completely removed, reducing the four-level delivery
system to three levels; and (c) the commune level is replaced by 12 more departmental stores.
The pandemic vaccine supply chain is a multi-tiered decision process where key allocation decisions hinge on
pe
infection transmission behavior and distribution-related factors. Brandeau et al. [19] demonstrated that optimal
resource allocation depends on many factors such as population size and susceptibility, local pandemic status,
time-horizon, prevention policies like mask wearing, likelihood of social contact, and the community transmission
rate. Furthermore, demand fluctuation and capacity to distribute and dispense vaccines to a population are vital
ot
parameters in the vaccine distribution model. These parameters represent the variegated demand for vaccines
via matching capacity[20], ensure that capacity is consistently available at peak demand levels, control capacity
to constantly maintain average benchmark levels based on planned schedules, and influence peak demand by
tn
diffusing it to lower demand periods[21, 22]. However, unpredictable and dynamic emergency situations pose
several challenges when implementing these strategies to solve highly complicated vaccine allocation decision
problems.[21] applied a cost-benefit-based model to optimize regional aid during public health emergencies. These
rin
are the three key results of their research: first, greater flexibility can be achieved by postponing deciding how much
to pre-allocate; second, small counties benefit most from mutual aid; and third, prioritizing groups in allocating
vaccines can result in significant savings.
ep
The development of an optimal COVID-19 vaccine supply chain is driven by constraints imposed by the specific
context of vaccination. It is anticipated that at least one or two doses will be required to achieve herd immunity
globally. Currently, 5.6 billion people need to be inoculated, which means mass production of vaccines needs to
be carried out in a short time. Hence, COVID-19 vaccines will not be supplied all at once, but in batches, to
Pr
maintain a sustainable and equitable system to satisfy burgeoning global demand. Vaccine supply is also impeded
by the capacity of the delivery mode and the dispensing capacity of health centers to effectively inoculate people
4
within the expected time threshold.
ed
In the following sub-sections, previous literature on vaccine supply chains and pandemic supply chains is presented
and explored. The extant literature on vaccine supply chains in a pandemic context is limited and fragmented.
Models were built using different process architecture, methods, and input parameters. However, an integrated
discussion is presented to differentiate vaccine supply chains from pandemic supply chains to ensure consideration
iew
of the key drivers, constraints, and mapping of vaccine allocation processes.
2.1. Vaccine supply chains
The current body of literature on vaccine supply chains is vast and well theorized. The majority of these studies
ev
were driven by cost-efficiency, as well as effectiveness of vaccine procurement, allocation, and distribution to
vaccinate a wider population. Lee et al. [23] explored the impact of a newly introduced vaccine on the existing
vaccine supply chain by combining a discrete event simulation model – Highly Extensible Resource for Modeling
Event-Driven Simulation (HERMES) – with a deterministic mathematical Equation-Based Model (EBM). Their
r
results showed that the distribution of the new vaccine required additional transport and storage capacity to be
able to effectively implement the vaccination program. Similarly, by taking into account scheduling preferences
er
of patients and scheduling inconvenience,Abrahams and Ragsdale [24] argued that the distribution of vaccines
presents a number of operations management challenges such as multi-dose vaccine packages, rapid spoilage upon
pe
opening, high cost of wastage, and unique vaccination needs of patients. Building a HERMES-generated simulation
model, Brown et al. (2014) explored how the introduction of the new vaccine affected its cost and availability in
the downstream supply chain. The model showed that when a new vaccine is introduced in an existing four-level
distribution system, the logistics cost per dose was relatively higher and vaccine availability was reduced. The
results indicate that the best option in terms of cost and efficiency was the one where the ‘commune’ level is
ot
consolidated into the ‘health zone’ level.
Another factor that influences vaccine supply chains is storage capacity. Shittu et al. [25] analyzed the impact of
tn
variance in supply and demand under various scenarios to enhance the supply chain’s competence to meet storage
requirements. By developing a simulation of vaccine storage capacity in Nigeria, their study found that a 55
percent increase in storage capacity was required to meet vaccination needs if no changes were implemented in the
rin
current supply chain. However, the establishment of three more vaccine delivery hubs could result in a decrease of
cold storage requirement from 55 percent to 33 percent. The redesign of old vaccine supply chains can therefore
be crucial for capacity utilization. Based on data from a vaccine supply chain in Mozambique, Lee et al. [26]
developed a simulation model based on HERMES to examine the efficiency of the existing multi-tiered vaccine
ep
supply chain compared to an alternative distribution system. In the existing system, manufacturers deliver the
vaccine to one national depot, the depot sends the vaccine to provisional stores, each provisional store provides
the vaccine to district stores, and finally health centers receive vaccines from the district stores. In their proposed
Pr
distribution system, the vaccine was directly delivered by stores in each province to health centers. The results
revealed that the alternative distribution design tended to increase the availability of vaccines and reduce logistics
costs per dose, as compared to the existing multi-tiered system.
5
The service providers in the downstream supply chain, such as hospitals or clinics, play a decisive role in ensuring
the effective distribution and dispensing of vaccines. For instance,Lin et al. [27] developed a mathematical model,
ed
firstly to analyze the distributor’s transportation decision to use a cold chain or non-cold chain to deliver vaccines
to retailers. Secondly, they extended the model to analyze the impacts of either a single-step or a two-step standard
inspection policy of retailers on the distributor’s decision whether to use the cold chain or not. The results revealed
iew
that the two-step policy, despite being stricter and more costly, was less effective in influencing the distributor to
select the cold chain option than the single-step policy was.
2.2. Pandemic vaccine supply chains
Vaccine supply chains during pandemics have their own specificities in terms of scale, exposure, time-space thresh-
ev
olds, and operating constraints. Uscher-Pines et al. [7] reviewed a sample of 45 national pandemic influenza
prioritization plans, including 19 developed and 26 developing countries. They found that 28 (14 developed and
14 developing countries) out of 45 nations have provision for prioritizing vulnerable groups of the population for
r
vaccination. In some countries, higher prioritization of high-risk individuals by healthcare workers and service
workers is embedded in the national pandemic vaccination plan (ibid). The study concluded by emphasising the
er
need to establish priority settings based on individualized modeling or impact estimates to enhance the effec-
tiveness of large-scale vaccination programs to mitigate the risk of community transmission during a pandemic
[7].
pe
Araz et al. [10] prioritized 15 counties of Arizona based on four distinct H1N1 pandemic vaccine distribution
strategies: pro rata distribution; sequential distribution by population size; sequential distribution by estimated
periods of pandemic peaks; and reverse sequential distribution by estimated order of pandemic peaks. Their study
identified the policies that would be effective to reduce the pandemic’s impact by optimizing waiting times for
ot
vaccines. The results revealed that the two most effective policies for controlling the epidemic and reducing unmet
demand are pro rata distribution and prioritization of communities expected to experience the latest outbreak.
tn
In fact, prior research such as [14, 16, 11, 9, 8] has extensively investigated risk-based pro rata distribution
and prioritization for vaccine allocation in order to minimize transmission. Medlock and Galvani [8] used a
parametrized model with survey-based contact and mortality data from influenza pandemics to determine optimal
rin
vaccine allocation minimizing five outcome measures: deaths, infections, years of life lost, contingent valuation,
and economic costs. They [8]found that optimal vaccination is feasible via prioritization of school children and
adults aged 30 to 39 years. Similarly, Lee et al. [9] recommended that the population group of ages 20-39 must be
given priority, followed by schoolchildren (6-12 years). The authors applied a non-linear dynamic mathematical
ep
model to determine optimal vaccination allocation policies during the Spring 2009 H1N1 pandemic in Mexico.
The main parameters used in the model were the age distribution of the population, age-specific vaccine efficacy,
hospitalization rates, mortality rates, and contact rates. The main reason for the overall transmission of the
Pr
pandemic was that schoolchildren (6-12 years) are often in high contact because of their activities [9]. Buccieri
and Gaetz [11] argued that in case of a pandemic outbreak, priority for vaccination should be given to population
groups at high risk and who would experience difficulty reaching health centres, such as homeless people and
6
other disadvantaged groups, to ensure equity and utility. During the H1N1 outbreak in Toronto, the city managed
to vaccinate 38 percent of homeless people via highly accessible community-based vaccine clinics. Davila-Payan
ed
et al. [12] applied multivariate linear regression to analyze the factors associated with the vaccination coverage
rates of two priority groups, children and high-risk adults, during the 2009-2010 H1N1 pandemic. The results
showed that the most critical determinants of vaccine coverage rates for children and high-risk adults are related
iew
to distributional and systemic decisions, such as vaccination venues. Huang et al. [14] considered five priority
groups – pregnant women, infants (0-3 years old), people 4-24, adults at high risk, and infant caregivers – to
optimally allocate different vaccine types in Texas, USA. They applied a pull-based strategy to distribute the
vaccine to healthcare providers and local health departments and a push-based strategy to deliver vaccines to
regional offices. Their results showed that a small cache of discretionary doses is sufficient to reach priority groups
ev
equitably across all counties in Texas. Taking New York City as a case study, Chen et al. [16] applied an age-
structured simulation model to explore the optimal allocation strategy for the COVID-19 vaccine. They divided
the population into seven compartments, and then each compartment was further divided into five age-groups.
r
They analyzed the impact of both static and dynamic policies. The results showed that, when the objective is to
minimize deaths, the optimal static approach is to vaccinate the oldest group first and then the younger groups.
er
But, when the target is to mitigate total confirmed cases, then the optimal static policy is to allocate vaccines to
younger people even if the supply is scarce. The best dynamic policies are ‘myopic’ and ‘two-day myopic.’ The
former policy distributes vaccines for the current day in such a way that tomorrow’s total new cases of all groups
pe
will be minimized. In contrast, the latter policy allocates vaccines for today and tomorrow to minimize the total
new cases of the day after tomorrow.
Optimal vaccine allocation and distribution depends on coordinated communication between vaccine providers and
Departments of Health. Marcello et al. [28] studied the effectiveness of communication in the delivery of H1N1
ot
pandemic vaccine to healthcare providers by the New York City Department of Health in 2009. They used data
from the Citywide Immunization Registry and a survey administered to vaccine providers to estimate the number
tn
of vaccine doses dispensed. The study concluded that nearly all (96 percent) the vaccine providers were satisfied
with the way information about the vaccine supply chain was communicated and coordinated by the Department
of Health. Fitzgerald et al. [29] studied the collaboration and coordination between vaccine providers, public
health departments, and pharmacies for pandemic vaccine planning in the USA. The results showed that about 89
rin
percent of jurisdictions involved pharmacies in the pandemic vaccine distribution plan. However, to improve the
pandemic vaccine supply chain’s efficiency and effectiveness, there is still a need to establish formal agreements
between public health departments and pharmacies. Turner [30] examined the procurement methods used by
ep
governments of developing countries for the pandemic influenza vaccine by considering procurement procedures
during the 2009 H1N1 pandemic as a case study. The results showed that traditional procurement tools are
ineffective and warned of the possibility of extended delays in vaccine distribution and dispensing. Biggerstaff
Pr
et al. [13] used a spreadsheet-based model to analyze several vaccination scenarios to prevent hospitalizations and
deaths in the United States. The variables used in the model were the number of people getting sick, by time,
the time difference between the launch of the vaccination program and emergence of the pandemic in the United
7
States, the number of doses administered per week, allocation by age-group, the clinical attack rate, hospitalization
rate, and case fatality ratios. They found that the time difference between the launch of the vaccination program
ed
and the start of the pandemic was a significant driver of vaccination.
Using a mixed-methods approach, Seib et al. [31] investigated the effect of vaccine-related standards of care
after H1N1, and found that only 40 percent of vaccine service providers had received training in emergency
iew
preparedness, actual disaster response, or surge capacity initiatives. The study recommended improving staffing
and surge capacity via preparedness training programs.
Li et al. [32] applied a detailed Susceptible-Exposed-Infected-Recovered (SEIR) model to explore the importance
of information availability on vaccine inventory status during an influenza pandemic. They studied two scenarios:
ev
firstly, where only details on population are available; and secondly, when both population and vaccine inventory
information are available. They found that the latter produced better service delivery. Their study recommended
improving information visibility about vaccine inventory levels across the supply chain by adopting systems and
r
practices to reduce infection risk. In a recent study, Govindan et al. [15] applied the Mamdani fuzzy inference
system (FIS) to develop a decision support system to manage demand in the healthcare system to help mitigate
er
the impact of the COVID-19 pandemic. They first divided the community into four groups based on their immune
system (very sensitive, sensitive, slightly sensitive, and normal) and then by two additional factors, their age and
pre-existing diseases. The proposed decision support system classified users based on the information collected
pe
by the system and then updated them about the kind of services they should receive from service providers. The
service recipients’ classifications provided the basis for managing demand in the healthcare supply chain. Finally,
the proposed decision support model was tested with the real data, and the simulation confirmed the validity of
the proposed decision support system.
ot
The studies discussed in this sub-section can be categorized into two groups of techniques, statistical (Table 1)
and mathematical (Table 2),based on their approaches and methods for modeling pandemic vaccine supply chains.
tn
Our literature review indicates that only a few studies have designed and built an efficient distribution network
which ensures that the vaccine will be effectively delivered to the neediest people (see[8, 9, 10, 13, 14, 15, 16]).
However, even these studies have not fully incorporated the multitude of factors and constraints, based on levels of
susceptibility and exposure, within which vaccine supply chains can be optimized to mitigate the risk of infection.
rin
None of the previous studies have considered transshipment between medical centers or interactions between
transshipment and vaccine package sizes. To deal with these gaps, this research develops a conceptual model of
the downstream vaccine supply chain to ensure efficient coordination and distribution of vaccines to the population,
ep
which will be discussed in the following section.

Pr
8
Table 1: A summary of literature on pandemic vaccine supply chains using statistical models. ’Relevance’ column indicates the relevance of the paper to a component of our
conceptual model.
Authors Objectives Priority group Methodology Variables Findings Relevance

Pr
Uscher- To review national pan- Healthcare workers, essen- Descriptive Vaccine and antiviral priority groups, group Countries gave top pri- Susceptibility
Pines demic influenza prioritiza- tial service providers, peo- statistics rankings, goals of pharmaceutical interven- oritization to high-risk
et al. tion plans. ple at high risk, children, el- tions, the inclusion of scenarios and popula- individuals, followed
ep
[7] derly, key decision-makers, tion size. by health care workers,
influenza cases, hospitalized and service workers.
cases, and the unvaccinated
Seib To investigate the effect of None Mixed meth- Type of vaccine providers, standard of care, More attention should None
et al. vaccine-related standard ods (linear willingness to respond, surge capacity, and be given to the size of
rin
[31] of care on willingness to regression and practice size (number of patients seen per the staff and surge ca-
respond among vaccine interviews) day). pacity in preparedness
providers after H1N1. training programs.
Buccieri To evaluate ethical pan- Homeless individuals in Mixed methods Gender, demographical factors, fear of infec- The findings showed Susceptibility
and demic planning policies. Toronto. (descriptive tion, lack of concern, access to community- that the immunization
tn
Gaetz statistics and based clinics, access to a regular doctor, pro- rate for the homeless
[11] interviews) motional campaigns. was higher than the ex-
pected average rate.
Davila- To explore the factors re- Children (6 months to 17 Linear regres- State campaign information, demographics, The most significant Susceptibility
Payan lated to a vaccination cover- years) and high-risk adults sion preventive or health seeking behavior, pre- factors of vaccine cov-
et al. age rates of priority groups. (25-64 years). paredness funding, providers, state character- erage rates are related
ot
[12] istics, and surveillance data. to the distributional
and systemic decisions.
9
Marcello To estimate the number of None Descriptive Facility type, number of doses administered, The reporting about Susceptibility
et al. vaccine doses administered. statistics number of doses in stock, number of doses the pandemic vaccine
[28] reported, and questions about the quality of was visible across all
pe
communications. vaccine providers.
FitzgeraldTo analyze the importance None Descriptive Data related to the public health program Formalized agreements None
et al. of cooperation between statistics plans for the following: recruiting, enrolling, between public health
[29] public health agencies and registering pharmacists as pandemic in- departments and phar-
and pandemic vaccine fluenza vaccine providers; vaccine allocation macies should be es-
er
providers. and distribution of pandemic vaccine to com- tablished.
munity pharmacies; weekly allocation of pan-
demic influenza vaccine by provider type;
r
and immunization policies, formalized agree-
ments, and memoranda of understanding be-
tween public health departments and commu-
nity pharmacies.
ev
Turner To evaluate the procure- None Literature Total number of doses received from all Traditional procure- None
[30] ment procedures for the review and sources, percentage of vaccine received from ment methods are
pandemic influenza vaccine quantitative WHO’s vaccine deployment initiative (VDI), not sufficient for
applied by the governments analysis. percentage of vaccine self-procured, percent- developing countries.
of developing countries. age of vaccine procured from other sources.
iew
ed
Table 2: A summary of work on pandemic vaccine supply chains using mathematical methods. ’Relevance’ column indicates its relevance to a component of our conceptual model.
Authors Objectives Priority group Methodology Variables Findings Relevance
Medlock To evaluate current vac- 17 age groups (ages 0, 1 to Age-structured Number of fatalities, contact rates, duration Optimal vaccination is Susceptibility
and cine allocation policies and 4; 5 to 9; 10 to 14; ...; 70 to simulation of the infectious period, years of life lost, achieved by prioritiza- and expo-
Gal- to determine the optimal 74; and 75 and older). model weighing deaths against the expected remain- tion of schoolchildren sure
Pr
vani strategy. ing years of life for different ages, contingent and adults aged 30 to
[8] valuation, costs associated with vaccination, 39 years.
costs associated with illness and death values.
Lee To determine optimal vac- 6 age groups (1 = 0–5 yrs, 2 Non-linear Age distribution of the population, age- The population group Susceptibility
ep
et al. cination allocation policies = 6–12 yrs, 3 = 13–19 yrs, 4 dynamic math- specific vaccine efficacy, hospitalization rates, aged 20-39 must be and expo-
[9] during the Spring 2009 = 20–39 yrs, 5= 40–59 yrs,6 ematical model mortality rates, and contact rates. given priority, followed sure
H1N1 pandemic in Mexico. = ¿60 yrs). by school children (6-
12 years).
Araz To identify the distribution Preschool age children (0–4 Mathematical Transmission probability per contact, age- Pro rata distribution Susceptibility
rin
et al. policies that would be ef- years), school age children model specific contact rate, force of infection, infec- and prioritization of and expo-
[10] fective to reduce the pan- (5–19 years), adults (20–64 tious period, incubation period, case fatality communities expected sure
demic’s impact. years), and older adults rate, vaccination rate, vaccine efficacy, vaccine to experience the latest
(65+ years). pre-protection period, and numbers of people outbreak are the two
in each county and age-group. most effective policies.
tn
BiggerstaffTo analyze different vacci- People presenting to receive A spreadsheet The number of people getting sick by time, the Strategies related to None
et al. nation scenario their second dose were pri- simulation time difference between the launch of the vac- improvements in time-
[13] oritized over those receiving model cination program and emergence of the pan- liness of vaccine pro-
their first dose. demic, the number of doses administered per duction are crucial for
week, and the allocation by age group, the future pandemic vacci-
clinical attack rate, hospitalization rate, vac- nation programs.
ot
cine effectiveness and case fatality ratios.
Huang To explore the optimal al- Pregnant women, infants Optimization The five priority groups and regions were taken A small cache of discre- Susceptibility
10
et al. location of several vaccine (0-3 years old); people be- model as input. tionary doses is enough
[14] types to certain priority tween age 4-24; and adults to achieve the vaccine’s
groups. at high risk and infant care- optimal distribution.
This preprint research paper has not been peer reviewed. Electronic
givers.
pe
Li To discuss the necessity of None Agent-based Every individual either: susceptible, exposed, It is recommended to Exposure
et al. having information on vac- simulation pre-symptomatic, asymptomatic, symp- make vaccine inven-
[32] cine inventory status during model tomatic, hospitalized, recovered, or dead. tory levels available
an influenza pandemic. Three categories of mixing: community (day throughout the sup-
and night), peer groups (day), and household ply chain to mitigate
er
(night). the number of infected
r cases.
Govindan To formulate a decision sup- Four categories based on Mamdani fuzzy Fever, tiredness, and dry cough as the input The results for differ- Susceptibility
et al. port system to manage the the risk level of their im- inference sys- variables. The output variable is the classifi- ent scenarios confirm
[15] demand and to control the mune system (very sensi- tem (FIS) cation of the community members. that the proposed deci-
outbreak of an epidemic. tive, sensitive, slightly sen- sion support system is
sitive, and normal). sound and reliable.
ev
Chen To determine the optimal Seven compartments Age-structured The number of individuals in each of the The results show that Susceptibility
et al. allocation policies for the (susceptible, exposed, simulation seven compartments, the population size, the among the static poli- and expo-
[16] COVID-19 vaccine. presymptomatic infectious, model transmission rate, the contact rate, the dis- cies, the optimal ap- sure
unascertained infectious, count factor of the transmission rate, the av- proach is to vaccinate
ascertained infectious, erage times (from exposed to infectious, from the oldest group first
isolated, and removed) presymptomatic infectious to symptomatic in- and then the younger
iew
and five age-groups (0-17, fectious, from symptomatic infectious to re- group. In the case of
18-44, 45-64, 65-74, and covered, from ascertained infectious to isola- dynamic policies, the
75+). tion, from isolation to recovered), the fraction results reveal that the
of ascertainment for each age group, the level best policies are my-
of permitted economic activities, the amount opic and two-day my-
of vaccine allocated to each age-group. opic.
ed
copy available at: https://ssrn.com/abstract=3744520

3. COVID-19 vaccine allocation conceptual model
ed
Building an optimization model based on a prioritization strategy for distributing vaccines is a complex and multi-
layered decision process which needs to be conceptualized and parameterized to reflect susceptibility, exposure
risk, and the operating environment. In this study, a new conceptual model for COVID-19 vaccine allocation is
developed. This model comprises three key stages: a centralized booking system; risk profiling and prioritization
iew
modeling; and the vaccine distribution system (see Figure 1).
1 - Centralized booking system (CBS): A central system to administer the booking and scheduling of appoint-
ments for vaccination needs to be designed and developed to enhance the effectiveness of antiviral immunization
for the vaccinated population, whilst mitigating the spread of the pandemic in the community. It is a self-service-
ev
oriented workflow system with a web-portal interface that enables potential recipients of vaccines to enter their
details, including their age, location, existing health conditions, type of work (e.g. healthcare workers, train
drivers), gender, the oldest person in their household who has not been vaccinated yet, household size, and daily
r
transport mode. The key functions of this web-portal are to allocate individuals to the appropriate vaccination
wave based on susceptibility and to identify the appropriate vaccination center based on their location.
er
2- Risk profiling and prioritization model: Based on the information provided through the CBS, each
individual is classified into a specific risk priority group based on susceptibility risk. We propose the segmentation
pe
system based on the one recommended by the US government and in published research on COVID-19 and influenza
vaccine distribution [19, 15, 6]. The segmentation system allocates individuals to five different groups based on
the type of occupation (e.g. essential workers), age cohort, co-morbidities, and pre-existing diseases. As shown in
Figure 1 below, these priority groups are based on combinations of the criteria and include: Extremely Susceptible,
Highly Susceptible, Moderately Susceptible, Partially Susceptible, and Least Susceptible. CBS enables the initial
ot
filtering and prioritizing of individuals by assigning them to predetermined risk clusters and then allocating them
to the nearest medical center. The capacity of medical centers remains a constraint on administering vaccines
tn
within the execution horizon.
3 - Vaccine distribution process: This defines the procedures for the allocation of COVID-19 vaccine by state
and local authorities (i.e. Departments of Health) to authorized healthcare facilities. Initially, we assume that the
rin
supply of COVID-19 vaccines will be limited and that therefore they will need to be batched using a centralized
distribution model. In a centralized distribution model, vaccines from suppliers are first transported to a central
location in full truckloads and then directly distributed to medical centers for capacity-led mass vaccination. In a
centralized system, the decision-making power is located in a single administrative authority which is responsible
ep
for performing key distribution functions for the state. The allocation decisions are based on minimizing the risk
to the population and reducing infection transmission. Supply of vaccines to different medical centers is based
on operational capacity and vaccine availability. We assume that the vaccine will be provided through a central
Pr
distribution center, national or regional, where stock will be replenished regularly. Both the size and number
of packages and different package sizes might influence allocation efficiency. Prioritized users are scheduled for
vaccination based on operational capacity, vaccine availability, and exposure risk in their location. The proposed
11
model allows for transshipment from the nearest medical centers or for centers to keep any excess vaccine, so that
centers have the option to procure vaccines from the nearest centers with excess supply.
ed
The COVID-19 Vaccine Allocation conceptual model, as shown in Figure 1, is designed as a multi-layered decision
process to improve the effectiveness of a vaccine supply chain. It is driven by: exposure risk and susceptibility
representing populations at risk; distribution of vaccines in different pack sizes; transshipment to improve opera-
iew
tional efficiency; and the capacity of medical centers to administer vaccines to assigned residents in the designated
catchment.
Figure 1: Vaccine allocation system
r ev
er
pe
ot
tn
rin
4. Mathematical model
The notations employed in the models of this study are as follows:

ep
Pr
12
Table 3: Notations and Definitions.
ed
Notation Definition
Sets and indices

R The set of priority groups, R = {1, 2, ..., R}.
iew
D The set of medical centers (demand points), D = {1, 2, ..., D}.
Parameters
L Lot size. The number of vaccine in each package.
nij The number of registered people for receiving the vaccine in demand point i ∈ D who are
in priority group j ∈ R.
ev
S The number of available vaccine packages per time period.
wj The weight of vaccination of a person in priority group j ∈ R, it is relatively defined for
PR
each group such that j=1 wj = 1 .
r
Ci The maximum number of vaccine that demand point i ∈ D can administer per time period.
M A big number.

γ
er
A very small number (e.g., 0.00001).
A coefficient defined as γ = min(wj−1 − wj ). It is the minimum deviation between the
weights of two priority groups.
pe
tij The travel time between demand points i ∈ D and j ∈ D \ {i}. ṫ = max(tij ).
α A parameter to cap number of transshipment per day in the network. It is the total travel
time that available vehicles can travel per day to do the transshipment. For example if
there are 2 vehicles α = 2 × 8 = 16.
ot
νij A parameter is the model which takes 0 or 1. It guarantees that each demand point is
linked for transshipment to at least G other demand points that have non-zero demand. If
P
only k ≤ G demand points meet m∈R njm > 0, all k demand points will be connected.
tn
Decision variables
φi The number of vaccine packages allocated to demand point (e.g. medical center) i ∈ D. It
is a decision variable in the model.
rin
xij The number of vaccine units administrated to individuals at demand point (e.g. medical
center) i ∈ D who are in priority group j ∈ R.
yij The number of vaccine units transshipped demand point i ∈ D to demand point j 6= i ∈ D.
ep
zij The binary decision variables to define if a transshipment from demand point i ∈ D to
demand point j 6= i ∈ D is occurred.
δ An exogenous decision variable to determine the number of units available at demand points
but not used due to no demand or capacity limitations.
Pr
τiin , τiout The binary decision variable to define transshipment in or transshipment out in node i.
13
DC(S)
ed
L
φi
×
×
φ2
L
iew
Di , Ci D2 , C2

L × φ1
ev
D1 , C1
î
y12 , t12
r
ω1 ω2 ω3 . . . ωj
P
n1j
g g g g g j
g g g g g
er
Figure 2: Schematic overview of vaccine distribution network.
pe
The objective of the mathematical model is to allocate the vaccine to medical centers based on the total risk
of individuals booked in there to receive the vaccine. A vaccine recipient prioritization and distribution system
should be encapsulated in the model whereby susceptibility rate and exposure risk are jointly addressed. The central
booking system uses the demographic data and medical history of applicants to assign them to the predetermined
ot
priority groups. In this study, the susceptibility rate of a person in a priority group is estimated via the weight of
the cluster that they belong to, e.g. age over 65. However, the exposure risk is calculated from the quotient of the
total number of people in each priority group in a specific medical center and the total number in the same priority
tn
group in all medical centers. Therefore, people in the same priority group will have higher priority to receive the
vaccine in a medical center where more people in the same priority group have registered for the vaccine.
Given that vaccines are delivered in packs, each demand point i receives φi number of packs or L × φi number
rin
of vaccines from the DC (as depicted in Figure 2). The complexity of designing an optimal distribution network
arises from the fact that some of the vaccines in a package delivered to one medical center could subsequently
be subsequently used in another medical center to further reduce total risk. Moreover, once the demand of a
ep
particular medical center is satisfied, the surplus (remainder of the last package) can be redistributed to other
medical centers to improve the overall efficiency of the distribution network. This challenge can be addressed using
one of the following two approaches: (i) using an inventory where each demand point keeps surplus vaccines for
the next period; or (ii) transshipping the vaccines to other demand nodes. For non-vital products and commodi-
Pr
ties, the first approach might deliver a simple but efficient solution. However, when people’s lives are at stake,
administering even one unit of vaccine, when it is available, should not be postponed. In this regard, the priority
of operational efficiency in designing a vaccine distribution and allocation network relies on effective embedding
14
of vaccine transshipment between demand nodes. Although embedding transshipment significantly improves the
operational efficiency of the allocation model, it imposes new challenges which should be properly addressed. For
ed
instance, each country/state has limited transportation capacity for transshipment network. Moreover, transship-
ping vaccines among medical centers should only be permitted with the remainder of vaccine units as a fraction
of a pack.
iew
Next, we present a Mixed Integer Linear Programming (MIP) model to address this problem:
XX nij X X
min (wj + γ P )(nij − xij ) + (ṫδ + tij yij ) (1)
nlj
ev
i∈D j∈R l∈D i∈D j∈D\{i}
S.t :
X X X
xij ≤ Lφi + yji − yij , ∀i ∈ D (2)
j∈R j∈D\{i} j∈D\{i}
r
XX X
xij = L φi − δ (3)
i∈D j∈R i∈D
X
i∈D
X
φi ≤ S
xij ≤ Ci , ∀i ∈ D
er (4)
(5)
pe
j∈R
X X
zij tij ≤ α (6)
i∈D j∈D\{i}
X
zij ≤ M τiout , ∀i ∈ D (7)
j∈D
ot
X
zji ≤ M τiin , ∀i ∈ D (8)
j∈D
τiout + τiin ≤ 1, ∀i ∈ D (9)

tn
zij ≤ νij , ∀i ∈ D, j ∈ D \ {i} (10)
xij 6 nij , ∀i ∈ D, j ∈ R (11)
yij ≤ M zij , ∀i ∈ D, j ∈ D \ {i} (12)

rin
φi , xij ∈ Z0+ , ∀i ∈ D, ∀j ∈ R (13)
yij ≥ 0, ∀i ∈ D, j ∈ D \ {i} (14)
δ≥0 (15)
ep
zij ∈ {0, 1}, ∀i ∈ D, j ∈ D \ {i} (16)
τiin , τiout ∈ {0, 1}, ∀i ∈ D (17)

Pr
The objective function has two main components. The first component minimizes the remaining risk among the
population registered to receive the vaccine. It considers a weight (wj ) for each priority group (j) which is a
15
measure of the susceptibility of each cluster and also prioritizes the most crowded demand points, defined by the
number of registered people linked to the density of a region. It would be reasonable to first vaccinate the most
ed
crowded regions in order to minimize the number of community transmissions. In other words, when the exposure
nij
risk ( P nlj ) of a region is high as a result of a high density of registered population for each priority group, the
l∈D
model gives that region higher priority in vaccine allocation. However, there is a need to consider a moderating
iew
coefficient to strike a balance between the exposure risk and the susceptibility rating. To address this, γ is
formulated as γ = min(wj−1 −wj ) to ensure that a person in a higher-risk group in terms of the susceptibility rating
j∈R
is given higher priority. In other words, the allocation mechanism is triggered by the susceptibility rating prior
P P
to considering the exposure risk. Given that i∈D j∈R (nij − xij ) represents the total number of unvaccinated
P P n
people who have been booked to receive the vaccine, we refer to i∈D j∈R (wj + γ P ij nlj )(nij − xij ) as the
ev
l∈D
Residual Risk of the allocation model.
The second component of the objective function has two elements. The first element (ṫδ) is developed to minimize
the number of over-supplied vaccines (δ) returning to the centralized distribution center (DC). The second element
r
ensures that the total transshipment time between all demand points which reflects transshipment quantity is
er
minimized. The common penalty coefficient for the transshipment and vaccine returns, , needs to be carefully
selected to ensure that the transshipment mechanism is triggered only when there is no further opportunity to
allocate vaccines directly from the DC. Similarly, ṫ where ṫ = max(tij ) ensures that the entire transshipment
pe
capacity of the network is used prior to returning δ units to the DC. The following conditions are defined into
to guarantee that the model meets the above requirements.
Condition 1 on . Sending a package to a node which only requires one unit of vaccine with no possibility of
transshipment to other nodes is better than keeping the package in the DC. It means the node needs to return
ot
min(wj)
L − 1 units to DC, given L > 1. Therefore, the following condition should be satisfied: ≤ ṫ(L−1)
. Note that
“≤” is used instead of “<” because a positive term is always added to wj in the objective function (refer to the
first term of the objective function).
tn
Properly 1. If L = 1, no transshipment is required. As the costs of delivery of a package from the DC to all
demand points are the same, direct delivery from the DC is less costly than using transshipment.
Constraints (2) limit the total number of vaccines received by each demand center, either via direct shipment
rin
from DC or via net transshipment from/to other medical centers. Constraint (3) ensures that if the total number
of vaccines which are allocated to all demand points is not divisible by the vaccine’s pack size, the over-supplied
units will be returned to the DC through the decision variable δ. Constraint (4) guarantees that total allocated
ep
vaccines are less than the total available vaccines in the DC. Constraints (5) reflect the capacity of each center
to administer vaccines. Constraint (6) defines the capacity for transshipment, which is related to the availability
of vehicles and their corresponding travel times. Constraints (7-9) prevent unnecessary transshipment between
Pr
demand nodes where a demand node could either send or receive to/from other medical centers.
Constraints (10) guarantee that each demand point can receive vaccines through the transshipment mechanism
from k nearest demand points that have existing demand themselves. Constraints (11) specify that the number
16
of vaccines allocated to a center should be less than its demand. Constraints (12) ensure that transshipment
of vaccines occurs only when the corresponding route is open for transshipment. Constraints (13-17) define the
ed
domain of variables where Z0+ is the set of non-negative integers.
4.1. Cuts to tighten oversupply
iew
Although the original model is fully capable of effectively allocating the vaccines, to improve solution time in
particular when large size problems are dealt with, a cut to the solution space is proposed:
î
δ≤ (L − 1) + M (1 − u) (18)
ev
min(k, î)
XX
αu ≥ α − tij yij − ṫ0 (19)
i∈D j∈D
r
There are three new terms which should be defined before delineating the impact of a designed cut. Although we
have developed a single-stage model, the allocation mechanism is used on a daily basis during the target horizon
er
and subsequent periods until the entire population is vaccinated. Note that the demand of each medical center
is updated based on the most recent booking numbers and the remaining unvaccinated people from previous
periods. Moreover, it is preferable to keep the unused vaccine units in the DC rather than in medical centers. The
pe
proposed cut has no impact on the objective function and is highly efficient in periods when the total supply is
greater than total demand. In each period, ṫ0 = max(tij ◦ νij ) where ’◦’ is the Hadamard product of matrices tij
and νij . ṫ0 captures the maximum time of available transshipment for all demand nodes with non-zero demand
at the current time. The second parameter (î) represents the remaining number of medical centers with non-zero
ot
P
demand (î = |{ j∈R nij > 0, ∀i ∈ D}|). Note that u is a binary decision variable in these constraints.
Given the defined parameters, u becomes 1 when any transshipment capacity is still available in the network. Note
tn
P P
that in the right hand side of the Constraint (19), α − i∈D j∈D tij yij represents the unused transshipment
capacity of the network. If this value is greater than the maximum capacity of any particular route in the network
(ṫ0 ), it implies that there is still transshipment capacity in the network, and thus u becomes 1 because α is greater
than any remaining transshipment capacity. When u = 1, Constraint (18) cuts the solution space by limiting the
rin
î
values of δ to be min(k,î)
(L − 1) units at most. To understand the rationale behind the upper limit of δ, note
that in each period there are î number of medical nodes with non-zero demand in the network. Each demand
node is able to transship at most (L − 1) number of vaccines to k number of other demand nodes, if î > k. It
ep
is obvious that the maximum number of transshipping units is (L − 1) as L is the pack size and transshipment
î
can be triggered for the remaining unassigned units in a pack. Therefore, min(k,î)
(L − 1) represents the maximum
number of vaccine units which can be transshipped among medical centers with non-zero demand. As explained
earlier, vaccines can be returned to the DC when there is no further opportunity in the network to transship them
Pr
î
among demand nodes, in order to decrease the residual risk. Therefore, the upper limit for δ is min(k,î)
(L − 1).
î
This cut is also valid for the final periods of vaccine distribution where î < k as min(k,î)
= 1 and the upper limit
for the number of units which can be returned to the DC (δ) is (L − 1).
17
5. Case study
ed
To assess the effectiveness of the proposed model, we apply it to allocating COVID-19 vaccines to the residents of
Victoria, a federated state in southeastern Australia. Victoria’s population is expected to be 6,909,840 in 2021 [33],
the majority of which is concentrated in the central south area surrounding Port Phillip Bay, and particularly in
the metropolitan area of Greater Melbourne, Victoria’s state capital and largest city and also Australia’s second-
iew
largest city, where over three-quarters of the Victorian population live. We assume that the Victorian government
has set a target horizon of 60 days to vaccinate the population of Victoria. The information regarding the medical
centers in Victoria is obtained from the Department of Health, Victoria, Australia [34]. Figure 3 shows the spatial
distribution of 325 medical centers (demand points) in the state of Victoria. In Figure 3a, the intensity of demand
ev
for each local government area (SA2 ) is depicted, and Figure 3b shows the intensity of demand for each medical
center.
r
er
pe
ot
(a) (b)
Figure 3: Spatial distribution of Victoria’s 325 medical centers

tn
The centralized booking system uses the demographic data, medical history, and other characteristics of applicants
to assign them to predetermined groups. To cluster the population of Victoria and quantify the susceptibility rate
of each priority group, various criteria (e.g., age, medical history, etc.) and methods (AHP, ANP, etc.) can be
rin
used. However, since clustering is a trivial process, without loss of generality we can use a simplified format to
maintain our focus on the allocation mechanism. We cluster the population of Victoria into five priority groups (R)
based on age-groups. As explained earlier, there is compelling evidence to show that one of the key contributing
factors in quantifying the susceptibility of people is ’age’. This finding can also be verified by comparison with the
ep
population of Victoria in 2021 [33], as illustrated in Figure 4a, and the number of COVID-19 deaths in Victoria
up to 20 August 2020 by age-group [35], as shown in Figure 4b.
To estimate wj , the priority of a risk cluster j (priority group) to receive the vaccine, we obtained the quotient
Pr
of the number of COVID-19 deaths and the population of Victoria for each age group. Then, the result was
normalized to 100%. The final values of wj were estimated as 0.00053243, 0.00728099, 0.02144653, 0.12265288,
0.84808718 for five age groups of 0-49, 50-59, 60-69, 70-79, >80, respectively.
18
5 400 386
ed
4
Population of Victoria (.106 )

300
No of COVID-19 Deaths
3
200
iew
98
100
1
25
4 10
0 0
0-49 50-59 60-69 70-79 >80 0-49 50-59 60-69 70-79 >80
Age Groups Age Groups
(a) Population of Victoria in 2021 per age groups (b) Number of COVID-19 deaths in Victoria by age groups
ev
Figure 4: Victoria’s population and number of COVID-19 deaths by age group
Next, we need to estimate nij , the number of people registered to receive the vaccine in medical centers i ∈ D
who are in priority group j ∈ R. Furthermore, values of tij as the travel time between medical centers should be
r
determined. As delineated in Algorithm 1, the geospatial data of medical centers [34] are employed as inputs to
estimate the travel times between medical centers (tij ) using the Geo-location API of the Google Cloud platform.
er
Next, to determine the values of nij , the Australian Statistical Geography Standard (ASGS) main structure is
applied [36]. This structure has seven hierarchical levels comprising in ascending order: Mesh Block (M B);
pe
Statistical Area Level 1 (SA1 ); Statistical Area Level 2 (SA2 ); Statistical Area Level 3 (SA3 ); Statistical Area
Level 4 (SA4 ); State and Territory (S/T ); and Australia (AU S). Each level directly aggregates into the level
above. Therefore, SA1 s are aggregates of Mesh Blocks and they aggregate into SA2 s.
After obtaining the geospatial data for the Mesh Blocks of Victoria [37], the latitudes and longitudes of SA2 s are
ot
computed. Then, the population of each SA2 by age group is obtained [38] and then age-groups are re-mapped
to the age-groups considered for the risk clusters. Next, given that A is defined as a set of SA2 s, distances
(dij ) between each pair of SA2 s (∀j, j ∈ A) and medical centers (∀i, i ∈ D) are computed. Moreover, the total
tn
population of each SA2 (∀j, j ∈ A), aggregated for all age groups, is represented by the variable pj .
Estimated value of dij and pj are incorporated into an assignment model (Model 2 of Algorithm 1) in which SA2 s
are assigned to medical centers in such a way that total travel distance is minimized. In this assignment model,
rin
qij is a binary decision variable which equals 1 if SA2 j is assigned to medical center i. Constraints (21) ensure
that only one medical center is assigned to each SA2 . Constraints (22) specify that at least one SA2 should be
assigned to a medical center. Constraints (23) prevent assignment of only one SA2 with zero population to a
ep
medical center. Constraints (24) define the domain of decision variables. The output from solving Model 2 is the
optimal values of qij which determine as whether SA2 j is assigned to medical center i. The results from solving
the assignment problem enable us to estimate the number of people (nij ) receiving the vaccine in each medical
center i ∈ D broken down by priority groups j ∈ R, by adding up the populations of all SA2 s for the same priority
Pr
groups which are assigned to medical center i.
Finally, to randomly distribute nij over the target horizon of 60 days, a vector of 60 random values was generated
19
from the Dirichlet distribution. Multiplying nij by a descending sorted vector of 60 Dirichlet random values
provides an estimation of the daily numbers of registered people receiving the vaccine at demand point i ∈ D who
ed
are in risk cluster j ∈ R. The rationale behind sorting the vector of 60 Dirichlet random values in ascending order
is that the probability of booking a higher number of people at the beginning of the target horizon is higher than
in later stages.
iew
Algorithm 1: Estimating nij
Input: Input parameters are Sets, parameters, and variables as listed in Table 3, A as a set of SA2 s in Victoria, Lat and Long
of M Bs in Victoria, Lat and Long of medical centers in Victoria and Population of each SA2 by age group.
Output: nij .
Model 1
ev
begin Model 1
Lat, Long of centroid of SA2 ← Mean of Lats, Longs of M Bs.
Population of SA2 s by age groups ← Mapping population of SA2 s to the risk cluster age groups.
dij ← Compute distance between the centroid of each SA2 and each medical centers.
pj ← Total population of each SA2 which is aggregated by age groups (∀j, j ∈ A).
qij ← Solve Model 2.
nij ← Sum of population of all SA2 s assigned to medical center i for each risk cluster j.
r
Return nij
Model 2
Input: dij : Distance between the centroid of each SA2 (∀j, j ∈ A) and each medical centers (∀i, i ∈ D), and pj : Total
Output: qij
begin Model 2
er
population of each SA2 which is aggregated by age groups (∀j, j ∈ A).
pe
XX
min dij qij (20)
i∈D j∈A
S. t :
X
qij = 1, ∀j ∈ A (21)
i∈D
X
qij ≥ 1, ∀i ∈ D (22)
ot
j∈A
X
qij pj ≥ 1, ∀i ∈ D (23)
j∈A
qij ∈ {0, 1}, ∀i ∈ D, ∀j ∈ A (24)

tn
Return qij
Regarding the daily capacity of medical centers Ci , the physical size or number of beds of medical centers is not an
rin
accurate measure to estimate their capacity to administer vaccines, because local governments will definitely make
the necessary arrangements to adjust the operational capacity of medical centers to administer vaccines based on
the total population of all SA2 s assigned to a particular medical center. Therefore, to obtain the most realistic
ep
value for medical center capacity, we generate Uniform random values for Ci with lower bound of 0.8
P
60 nij and
j∈R
upper bound of 1.2
P
60 nij , given that the plan is to vaccinate the whole population in 60 days. The values of tij
j∈R
are also computed by measuring the travel time between each pair of all the demand nodes, using the latitude and
Pr
longitude of medical centers.
After estimating the key sets and parameters of the model (D, R, wj , nij , tij , ṫ, γ, and Ci ) for the state of
Victoria, the remaining parameters of the vaccine allocation model are selected as: M = 1e5; = 6e − 8; k = 10;
20
target horizon=60; α = b 3i × 5 × 60c = 32, 499 minutes, assuming that there is one vehicle for every three medical
P P
nij
ed
i∈D j∈R
centers to facilitate the transshipment of vaccines and operate five hours a day; S = b target horizon×L c = 9, 597
packs of available vaccines per day in a DC. The vaccine allocation model is run for each day within the target
horizon and continued after the target horizon until the entire population of Victoria is vaccinated. Note that
if any portion of the daily demand is not satisfied due to any limitation in the capacity of medical centers, the
iew
available supply of vaccines, or the capacity for transshipment, this unmet demand is added to the demand of the
next day. The analysis is conducted on a computer with Intel Xeon CPU E5-2680 V3 @2.5GHz and 64GB RAM.
The model is implemented in Python 3.6, and the CPLEX 12.8 optimizer is used for the optimization process. To
instruct CPLEX to stop as soon as a feasible integer solution is identified, the relative MIP gap tolerance is set to
ev
1e-4. The results from executing the vaccine allocation model are shown in Figure 5.
3 4
Cumulative Unmet Demand (.106 )
2.5
3
r
Residual Risk (.103 )
2
1.5 2
0.5
0
er 1
0
pe
0 20 40 60 80 100 0 20 40 60 80 100
t (day) t (day)
(a) Cumulative unmet demand (b) Residual risk.
Figure 5: Cumulative unmet demand and residual risk of vaccine allocation mechanism.
ot
Figure 5a shows the cumulative number of people who have not yet received the vaccine for each day within the
P P t t−1
P P P τ
target horizon. In other words, it is an illustration of nij + (nij − xτij ). As per the settings of
i∈D j∈R τ =0 i∈D j∈R
P P t
this experiment, nij is a decreasing function in t, given that the number of people who book to receive the
tn
i∈D j∈R
vaccine is higher at the beginning of the target horizon than in later stages. However, adding the accumulated
t−1
P P P τ
unmet demand ( (nij − xτij )) from previous periods to the demand of the current period leads to an
τ =0 i∈D j∈R
increased total demand for the first 20 days and then a decreased total accumulated demand of vaccines, given the
rin
limitations on the availability of vaccines (S), capacity of medical centers (Ci ), and the transshipment capacity (α).
It can also be observed that the allocation of vaccines continues 30 days after the target horizon. This extension of
time for administering vaccines to the whole population provides a useful insight for vaccine supply chain planners
ep
as to how the actual time to complete the allocation process (the administering horizon) may deviate from the
target horizon.
0 t−1
nijt 0 0
)(nijt −xij )), where nijt = ntij + (nτij −
P P P P P P P
Figure 5b represents the residual risk ( (wj +γ P 0t
i∈D j∈D l∈D nlj i∈D j∈R τ =0 i∈D j∈R
Pr
xτij ). This is a crucial factor reflecting the weighted number of booked but unvaccinated people in each period. We
can see that the model is capable of effectively decreasing the value of the residual risk during the target horizon
and the subsequent 30 days until all Victorian residents are immunized.
21
12
10
ed
Total Transshipment (.103 )
8
iew
2
0 20 40 60 80 100
t (day)
Figure 6: Transshipment of vaccines among medical centers.
ev
To assess the effectiveness of the designed transshipment mechanism, Figure 6 presents the daily total number
of vaccines transshipped between medical centers. Transshipping up to 12,000 units of vaccines on a daily basis
clearly demonstrates the need to incorporate this process into the vaccine allocation model. In the next section,
r
we compare the performance of the proposed vaccine allocation model and the traditional risk-based allocation
policy.
5.1. Performance of the vaccine allocation mechanism er

In the absence of an optimal allocation mechanism, a Risk-Based Allocation (RBA) heuristic is a viable option.
pe
In this section, we compare the performance of the proposed allocation model with that of the RBA heuristic. In
nij
the RBA heuristic developed, vaccines are allocated based on the coefficient of the residual risk (wj + γ P nlj )
l∈D
of each medical center i for risk category j. To determine the daily allocated of vaccines (xij ), a two-stage process
is used. In the first stage, vaccines are pre-allocated based on the coefficient of the residual risk. At each step
ot
of the process, allocated stocks are examined to ensure that constraints on the number of available vaccines and
the capacity of medical centers are not violated. However, this allocation may not be fully feasible because the
tn
number of vaccines allocated to each medical center should be divisible by the pack size (L). Therefore, in the
second stage the pre-allocated stocks are adjusted to ensure that only whole packs of vaccines are distributed and
that the remaining pre-allocated vaccines are excluded from the next risk category in each medical center. Details
of the RBA heuristic developed to allocate vaccines are provided in Algorithm 2.
rin
ep
Pr
22
Algorithm 2: Risk-Based Allocation (RBA) Heuristic Algorithm
Input: Input parameters are Sets, parameters, and variables as listed in Table 3.
ed
Output: Outputs are rrt , ntsum .
Initialization
t ← 0 initialize the administering horizon.
ξ ← initialize a vector of 60 random values generated from the Dirichlet distribution.
rrt ← ∅ initialize an empty vector for daily residual risks over the administering horizon.
iew
nij
crr
ij ← ∅ initialize an empty array for coefficient of daily residual risks (wj + γ
P
nlj
).
l∈D
ntij ← ∅ initialize an empty array for daily nij over the administering horizon.
xtij ← ∅ initialize an empty array for daily allocated stock to medical center i for risk category j over the administering horizon.
0
xijt ← ∅ initialize an empty array for daily allocated stock to medical center i for risk category j over the administering horizon.
ntsum ← ∅ initialize an empty vector for daily total demand of vaccine units over the administering horizon.
s0 ← 0 initialize the remaining number of vaccines.
0
Ci ← ∅ initialize an empty vector for the remaining capacity of medical centers.
ev
φti ← ∅ initialize an empty array for number of allocated vaccine packs to medical centers over the administering horizon.
while max(ntij ) ≥ k do
if t=0 then
n0ij ← dnij × ξ0 e
r
else
if t ≤ 60 then
ntij ← dnij × ξt e + (nt−1 t−1
ij − xij )
0
else
ntij ← (nt−1 t−1
ij − xij )
xijt ← fx0 (ntij ), ∀i ∈ D, ∀j ∈ R

er
pe
ij
P 0
φti [i] ← xij [i][j], ∀i
j∈R
xtij ← fxij (φti ), ∀i ∈ D, ∀j ∈ R
nt
rrt ← (wj + γ P ijnt )(ntij − xtij )
P P
i∈D j∈D l∈D lj
ntsum ←
P P t
nij
i∈D j∈R
t=t+1
ot
return rrt , ntsum

function fx0 (nij ):
ij
s0 ← s × k
0
tn
Ci ← Ci
nij
crr
ij ← wj + γ
P
nlj
, ∀i, ∀j
l∈D
0
crr rr
i×j ← Asc Sort(F latten(cij ))
c←0
0
while c < crr
i×j do

rin
i, j ← indice(crr rr 0
ij = ci×j [c])
0 0
xij [i][j] ← min(s0 , Ci , nij [i][j])
0 0 0
Ci ← Ci − xij [i][j]
0 0 0
s ← s − xij [i][j]
c←c+1
ep
0
return xij
function fxij (φi ):

xφi ← φi ∗ k

for i ← 0 toD by 1 do

Pr
for j ← R to 1 by −1 do
xij [i][j] ← min(xφ
i [i], nij [i][j])
xφ φ
i [i] ← xi [i] − xij [i][j]
return xij
23
The result of comparison between the performance of the proposed optimal allocation model and that of the RBA
heuristic is provided in Figure 7. The optimal model clearly outperforms the RBA heuristic from the perspective
ed
of both cumulative unmet demand (Figure 7a) and residual risk (Figure 7b). The gap between the performance
of these models becomes more significant in the period between 10 and 50, when the system is in a severe supply
shortage position.
iew
3 4
Proposed Model Proposed Model
Cumulative Unmet Demand (.106 )
RBA Heuristic RBA Heuristic

2
ev
1
1
0 0
0 20 40 60 80 100 0 20 40 60 80 100
r
t (day) t (day)
(a) Cumulative unmet demand (b) Residual risk.
er
Figure 7: Comparing the cumulative unmet demand and residual risk between the proposed vaccine allocation model and RBA
heuristic.
pe
5.2. Impact of capacity of medical centers
It is crucial for planning bodies to consider how the capacity of medical centers may affect residual risk and the
vaccine administering period. To study the impact of Ci , the standard settings of the original experiment (e.g.
L = 12, one vehicle per three medical centers etc.) are used in the current experiment except the daily available
number of vaccine packs (S), which is selected as 20,000, to compare to the original experiment (S = 9, 597).
ot
The reason behind this selection is that relaxing the cap on S provides an opportunity to investigate the unbiased
impact of medical center capacities on the vaccine allocation model. The original capacities of all medical centers
tn
are increased by the following Enlargement Scale Factors (ESF ): {1, 1.2, 1.4, 1.6, 1.8, 2}. Results of these
experiments are presented in Figure 8.
3 7
1.0 Ci
Accumulated Residual Risk (.104 )
rin
1.2 Ci
1.4 Ci
1.6 Ci
1.8 Ci 5
2
2.0 Ci
3
ep
1
0 1 1.2 1.4 1.6 1.8 2
10 30 50 70 90
Enlargement scale factor for Ci
t (day)
Pr
(b) Accumulated Residual risk for various capacities of medical

(a) Residual risk for various capacities of medical centers. centers.
Figure 8: Residual and accumulative residual risk for various capacities of medical centers.
24
Figure 8a shows how the residual risk of allocating vaccines drops during the administering horizon. It can be
observed that when the capacity of medical centers is sufficiently large (ESF ∈ {1.6, 1.8, 2}), the entire population
ed
can be vaccinated within the target horizon of 60 days. However, for ESF s of 1, 1.2, and 1.4, the administering
periods are 64, 75, and 90 days, respectively, which all exceed the target horizon. It is obvious that the capacities
of medical centers to administer vaccines are not fixed, and that they are basically controllable inputs for a decision
iew
support system. This flexibility derives from the fact that medical centers can be instructed to allocate more/less
of their capacities to vaccinate the people booked. In case of severe restrictions on administering capacity caused
by any operational limitations, there is still an option for the decision body to increase the capacity of particular
medical centers using mobile units. The designed allocation model enables decision makers to strike a balance
between the cost of building an additional administering capacity and trying to meet the target horizon or using
ev
a less costly approach where a threshold of deviation from the target horizon is deemed acceptable. It can also be
seen that dedicating extra capacity over a certain threshold (given the current experiment’s settings, when ESF
> 1.6) has no additional positive impact on the administering horizon.
r
The second insight which can be gleaned from Figure 8a is that the residual risk is decreased by dedicating
er
additional capacity for administering vaccines to medical centers. To further scrutinize this trend, the accumulated
residual risk, represented by the area under the residual risk curve, is presented in Figure 8b for various values of
ESF . Although Figure 8a indicates that no further decrease of the administering period is derived from increasing
pe
the capacity of medical centers above 1.6, , Figure 8b shows that accumulated residual risk declines exponentially
for ESF s above 1.6. This extra buffer may help dampen the variability of allocation and administering mechanisms.
For example, variability resulting from registered vaccine recipients not attending their scheduled appointments
or postponing their appointments, or from other sources of variability in the distribution and/or administering
process can be addressed by considering additional capacity for medical centers. In the case of Victoria, we can see
ot
that doubling the capacity of medical centers decreases the weighted number of unvaccinated people approximately
seven fold.
tn
5.3. Impact of daily available vaccine packs (S)
In addition to the capacities of medical centers (Ci ), the daily number of available vaccine packs (i.e. base-stock
level, S) is the second most important restricting factor for timely fulfillment of medical centers’ demands. To
rin
study the impact of the base-stock level, while settings of the original experiment are employed (e.g. L = 12,
one vehicle per three medical centers etc.), relaxed values for Ci are considered, where ESF = 2. Figure 9
shows the daily and accumulated residual risk of each experiment for various values of the base-stock level,
ep
S ∈ {5000, 7500, 10000, 12500, 15000, 17500, 20000}.

Pr
25
10 3
S = 5000
Accumulated Residual Risk (.105 )

S = 7500
ed
8 S = 10000
S = 12500
S = 15000 2
S = 17500
6
S = 20000
4
1
iew
2
0
0 6 8 10 12 14 16 18 20
0 20 40 60 80 100 120
Daily Vaccine Capacity (S) (.103 )
t (day)
(b) Accumulated Residual risk for various daily available vaccine
(a) Residual risk for various daily available vaccine packs. packs.
ev
Figure 9: Residual and accumulated residual risk for various daily available vaccine packs.
Although this experiment is considered as a post-optimal analysis, it helps to validate the feasibility of the original
r
plan as to whether it is capable of completing the administering process within the target horizon. Intuition
suggests that increasing the base-stock level leads to lower levels of residual risk (Figure 9a). However, vaccinating
er
the entire population of Victoria within the target horizon relies on a minimum level of stock availability in the
DC. Although experiments, as depicted in Figure 9a, can be conducted at more granular levels, given the selected
levels of S and the relaxed restrictions on Ci values, it can be seen that a minimum of roughly 10,000 vaccine
pe
packs is required on a daily basis to complete the administering process within 60 days.
Figure 9b also demonstrates that the accumulated residual risk (i.e. the area under the residual risk curve)
exponentially decreases with increasing base-stock levels. This exponential decline confirms that other binding
parameters, such as capacity of medical centers (Ci ), confines the effectiveness of increasing S values to improve
ot
the residual risk of the vaccine administering process. Therefore, this experiment provides the insight for vac-
cine distribution designers that capacity building decisions should be made considering operational and logistical
restrictions, to ensure that additional endeavors in capacity building are worth the marginal benefits.
tn
5.4. Impact of pack size (L)
When a vaccine becomes available, there is always an opportunity for the decision-making authorities to negotiate
rin
a customized pack size (L) of vaccines with producers. Although the proposed vaccine allocation model is capable
of managing any pack size, to effectively manage the operational risk of the transshipment process it is worth
investigating how the number of transshipments is affected by changing the pack size. To conduct this experiment,
a relaxed range for capacities of medical centers is selected, where ESF = 2. Pack size of vaccines is selected as
ep
L ∈ {5, 12, 20, 30, 50}. To ensure

P P
the same number of vaccine units is available in all experiments, the base-stock
nij
i∈D j∈R
level is updated as S = b target horizon×L c for each selected value of L. Note that since the value of depends on
min(wj)
the pack size (as per Condition 1 on ), the value of is selected in a way that the condition ≤ ṫ(L−1)
is not
Pr
i
violated. The transshipment capacity (α) is also updated, given that α = b (1v/nM Cs) × 5 × 60c, where 1v/nMCs
denotes one vehicle per n number of medical centers (nMCs) in the transshipment network and this parameter is
selected to be 30, 40, and 50.
26
6
L = 5 25 1V/30MCs
Mean Daily Transshipped Vaccines (.104 )

L = 12
Daily Transshipped Vaccines (.103 )

1V/40MCs
5
ed
L = 20
1V/50MCs
L = 30
20
L = 50
4
15
3
2 10
iew
1
5
0 10 20 30 40 50 60 5 12 20 30 50
t (day) Pack size (L)
(a) Daily number of transshipped vaccines among medical cen- (b) Mean number of transshipped vaccines among medical cen-
ters for various pack sizes (L) where there is one vehicle available ters for various pack sizes (L) where there is one vehicle available
per thirty medicals centers. The dashed lines show the average per thirty/forthy/fifty medicals centers.
ev
of daily transshipment.
Figure 10: Number of transshipped vaccines among medical centers for various pack sizes (L) and available transshipment capacities.
r
Figure 10a portrays a daily number of transshipped vaccines for various values of pack size where one vehicle per
30 medical centers is available. The dotted lines show the average number of transshipped units within the target
er
horizon for each value of L. It can be seen that increasing the pack size leads to a higher number of vaccines being
transshipped between medical centers. It is obvious that when L = 1, there is no transshipment in the network.
Increasing the pack size increases the probability of a union of events which create a situation where x number of
pe
vaccine units are left in a medical center after fulfilling existing demand, where x ∈ {1, 2, 3, . . . , k − 1}. The unused
vaccines can then be transshipped to other medical centers to improve the number of vaccinated people. Please note
that there may be an unmet demand in a particular medical center which triggers the transshipment mechanism
to transfer vaccines to that medical center. However, fulfilling more demand is not the only mechanism triggering
ot
the transshipment mechanism, which is also triggered when there is an opportunity to provide the vaccine to a
person in a higher priority group at another medical center.
Intuition suggests that when higher capacity for transshipment (α) is available, the number of transshipped vaccines
tn
will grow as well. The result of analysis, as depicted in Figure 10b, shows that the expected trend is not always
realized when the model manages the allocation of vaccines within the target horizon. This can be justified by
considering that additional transshipment for a particular day definitely has a positive impact on minimizing the
rin
residual risk for that day, but decreasing the number of unvaccinated people on that day changes the total demand
of the next day. Changes in the next day’s demand may have a positive or negative impact on the optimality of
the allocation model for the next day, and the mean of transshipped units within the target horizon may thus be
ep
increased or decreased.
5.5. Computational time
The proposed allocation model effectively distributes vaccines to the Victorian population of 6.9M, based on the
Pr
susceptibility rate and exposure risk factors, within the administering horizon. We have assumed that vaccines are
administered by the 325 medical centers in Victoria. The computational time for each period (day) within the
administering horizon is less than a minute (48 seconds), on average, given that there are 325 demand nodes in
27
the network. Although this is a fairly reasonable computational time for the selected case, it may be negatively
impacted by two factors: (1) If the government decides to also employ General Practitioners (GPs) to admin-
ed
ister vaccines, the number of demand nodes will significantly increase; (2) If the proposed model is used in a
state/country with higher population, there is a chance that a higher number of demand nodes will be involved.
It is therefore prudent to investigate how the size of the model impacts the computational time needed to identify
iew
the optimal solution.
It is obvious that two parameters directly impact the size of the model. The first parameter is the number of
medical centers (i), which it exponentially increases the size of the model. The second parameter is î, which is
used in Constraint 18 and refers to the number of medical centers with non-zero demand which are located at the
ev
closest distance from a particular medical center. For example, if î = 10, it means each medical center is only
able to transship vaccines to/from the 10 nearest medical centers with non-zero demand. The binary decision
variables of yij and zij are set to zero for all other medical centers (∀j ∈ D) not amongst the 10 nearest medical
centers with non-zero demand. It is obvious that increasing î leads to an exponential growth in the problem size.
r
To measure the computational times for various sizes of the problem, the settings of the original experiment are
î ∈ {10, 20, 30, 40, 50}.
60
er
employed (e.g., L = 12, one vehicle per three medical centers, etc.) while i ∈ {325, 500, 700, 1000, 1500, 2000} and
î = 10
pe
î = 20
Computational Time (min)
î = 30
î = 40
40
î = 50
20
ot
0
tn
325 500 700 1,000 1,500 2,000

Model Size (number of medical centers)
Figure 11: Computational time for various sizes of a single-period vaccine allocation model.
rin
Figure 11 shows the computational times needed to identify the optimal solution for various sizes of the problem.
It can be seen that, even for the worst-case scenario with 2000 demand points (i = 2000) and a high value of î = 50,
the computational time is less than an hour (45:42 mm:ss). Therefore, the model is fully capable of managing
reasonably large problems and it does not seem necessary to develop a heuristic specifically to solve large instances
ep
of the problem.
6. Discussion and implications

Pr
Robust allocation mechanisms are crucial for successful emergency responses including efficient vaccine distribution
during a pandemic [6]. With respect to the COVID-19 pandemic, the need for an effective allocation model has
been recognized as one of the six major logistics challenges for enhancing community resilience to the COVID-19
28
pandemic [39]. Although every health crisis has its own unique characteristics, an effective resource allocation
model is a common need of all emergency responses. In fact, a well-established emergency response plan should
ed
predefine in advance any components it can, including vaccine allocation systems, with risk-based prioritization,
while other aspects must be tackled on an ad hoc basis. This centralized and proactive strategy allows governments
to best deploy their finite resources in critical, unexpected, and ad hoc decision-making situations.
iew
In contrast to the prevalent vaccine allocation models from previous epidemics and pandemics (see [8, 9, 10, 13,
14, 15, 16]), this study has fully incorporated the multitude of factors and constraints within which vaccine supply
chains can be optimized to mitigate the risk of transmission. Both scholars and practitioners have stressed the need
for a more robust model of the downstream vaccine supply chain to ensure efficient coordination and distribution
ev
of vaccines. The current study is motivated by the need to develop an effective allocation model for distributing
COVID-19 vaccines, as a vital component of the emergency decision support system. This research contributes
to the literature on pandemic vaccine supply chains by developing a system that integrates a centralized booking
system, risk profiling and prioritization, and a vaccine allocation system.
r
It is undeniable that designing an effective allocation mechanism for downstream vaccine supply chains is complex
er
and multi-factored, in particular for large states. This is particularly challenging given the fragmented nature of
local government, with 31 municipalities in the Melbourne metropolitan area. To address this level of complexity,
we have applied our model to different scenarios of vaccine distribution to the 6.9M residents of Victoria, Australia.
pe
First, an algorithm has been developed to assign the population of each local government area to medical centers.
Second, the distance between medical centers has been measured using the Google Cloud mapping service. Finally,
the risk for each priority group has been estimated using the total deaths from COVID-19 for each age-group.
Applying the model to administering the vaccine in Victoria confirms its effectiveness to vaccinate the Victorian
ot
population in 90 days, given limitations on vaccine availability and capacity of medical centers. Comparing the
performance of the proposed model and the RBA heuristic provides additional evidence that using the RBA
heuristic leads to operational efficiency deviating from the optimal position. It should be emphasized that, in the
tn
case of vaccine allocation, deviations causing a decline in the number of vaccinated people at any point during the
administering period will put people’s lives in danger.
The key theoretical contributions of this study to the field of vaccine allocation are manifold, in particular its
rin
covering the case where a vaccine is introduced to the market for the first time. First, typical allocation mechanisms,
such as proportional, linear, uniform, lexicographic, and two-stage models, have limited use in pandemic vaccine
supply chains because they are generally centered on a typical cost minimization paradigm. The salient feature of a
ep
vaccine allocation decision support system is minimizing the severity and reducing the mortality rate of a pandemic
by prioritizing the most vulnerable people in the community for treatment and vaccination. The novelty of this
model is its ability to encapsulate the susceptibility rating and exposure risk factors in the proposed allocation
model to ensure that the predefined risk categories determine priority for receiving vaccines.
Pr
Second, the integration of a transshipment mechanism into the optimization routine significantly improves the
operational efficiency of this allocation model. Previous models for pandemic vaccine distribution have largely
29
overlooked transshipment, except that of Arora et al. [21]. However, their study modeled vaccine transshipment
between regions, while this study has examined the local distribution of vaccines between medical centers within
ed
each region to optimize pandemic vaccine allocation in Victoria, Australia. As highlighted by Arora et al. [21],
without transshipment the medical centers serving regions less affected by the pandemic are likely to have sur-
plus vaccines, while others with more pandemic cases might experience shortages. Providing the opportunity to
iew
transship vaccines among medical centers delivers two key advantages for the model. First, risk is optimally mini-
mized because people in higher priority groups within a medical center can be vaccinated sooner by transshipping
vaccine packs from nearby centers. Second, given that vaccines are delivered in a predetermined pack size, the
transshipment mechanism also boosts the operational efficiency of the allocation model because the remainder of a
pack is used at other demand points rather than being stored in a center or being returned to the DC. It basically
ev
improves distribution efficiency while minimizing the population at increased risk. The novelty of this model thus
is the integration of transshipment and vaccine pack size into a vaccine allocation model, to eliminate storage in
medical centers and reduce the rate of vaccine returns to the DC.
r
Furthermore, the model is also capable of prioritizing allocation tasks, whereby as the first priority medical centers
er
receive vaccines from the distribution center based on their demand and capacity. Then, as the second priority, the
remainders of vaccine packs are redistributed via the transshipment mechanism. Finally, the very small portion
of vaccines which cannot contribute to either of these two priority tasks are returned to the distribution center.
pe
The findings of this study have significant managerial and policy implications for COVID-19 vaccine allocation
and distribution decision-making. Estimating the administering horizon, based on available resources, guides the
development of an operational plan to distribute and allocate vaccines. Our model can be used as a strategic tool by
epidemiologists to direct actions to mitigate potential risk of community spread in high-risk areas. Additionally,
ot
the model provides evidence that the impact of pack size and transshipment mechanism should be carefully
considered by the government, as their absence significantly decreases the effectiveness of the allocation model.
The post-optimality analysis also provides useful managerial insights for effectively managing the downstream
tn
vaccine allocation process. For instance, we observe that the capacity of medical centers has a considerable impact
on residual risk and the administering horizon. Therefore, to expedite the vaccination process, governments may
consider mobile units to provide extra support to medical centers with high demand. With respect to the number
rin
of daily available vaccine packs, we have found that there is a critical point for this number, where any value
less than that results in a prolongation of the administering horizon beyond the target horizon. However, if daily
available vaccine packs are more than this threshold, assuming everything else to be constant, residual risk can be
reduced faster. This is particularly helpful for managing unexpected logistics disruptions in the entire network.
ep
We have also identified a relationship between the capacity for transshipment and pack size. Increasing the pack
size inevitably leads to a growth in vaccine surpluses in medical centers, which in turn boosts the number of
transshipped vaccines. Therefore, higher capacity for transshipment is required across the network to manage this
Pr
process. This finding provides a guideline for decision-makers that if vaccine providers are limited to delivering
vaccines in a large pack size, a transshipment network with higher capacity should be designed.
30
From the planning perspective, it is crucial to generate the results from the allocation model within a reasonable
time. Otherwise, an alternative heuristic algorithm should be provided to support the model for large-size prob-
ed
lems. Assessing the computational time to solve the model shows that, even for large-size problems, the model
can be solved in less than an hour and thus no complementary heuristic algorithm is required.
iew
7. Conclusion
In excess of 250 candidates for a COVID-19 vaccine are currently in various stages of development, all hoping
to have a vaccine as soon as possible. As previous work indicates, governments should be proactive in planning
logistical activities prior to availability of the vaccine. One of the key logistics hurdles is the equitable allocation of
ev
vaccines to various priority groups in a community, subject to any limitations of resources. To plan this, a response
agency needs an effective allocation model for this critical activity. To address this need, we have developed a
tailored allocation mechanism for distributing the vaccine to a diversified range of priority groups and minimizing
r
the residual risk of having unvaccinated people in the community. The proposed model encapsulates the logistics
aspects of both vaccine pack size and transshipment of vaccines among medical centers.
er
Although we have tested our proposed model for vaccine distribution to the Victorian population, this study has a
number of limitations which could be addressed in future research. First, there is a chance of disruption in numbers
of daily available vaccine packs. Therefore, future research could study the impact of stochasticity in terms of
pe
vaccine availability. Second, pack sizes may be non-identical if vaccines are sourced from different suppliers.
Accordingly, another avenue for future research would be to incorporate various pack sizes into the allocation
model. Third, the response agency might consider mobile units to minimize the gap between the target horizon
and administering horizon. Hence, future studies could investigate how mobile units could be assigned across
ot
the distribution network to further minimize the residual risk of having unvaccinated people in the community.
Finally, there is a risk that people who book to receive the vaccine may not turn up on their scheduled date. Future
research could build a procedure to accommodate probabilistic demand arising from cancellation or postponement
tn
of appointments.
In conclusion, synthesizing lessons learned from logistics issues in the downstream distribution of the personal
protective equipment (PPE) supply chain during the first COVID-19 wave, and detailing the upcoming challenges
rin
for vaccine distribution, should be the thrust of governmental attempts to alleviate the negative consequences of
the current pandemic. This study has provided managerial guidelines and policy recommendations for response
agencies regarding incorporating salient logistics aspects into allocation mechanisms and making administering
ep
efforts more effective and productive in the overall vaccine allocation process.
References
Pr
[1] World Health Organization, Immunization, Vaccines and Biologicals, 2020. URL: https://www.who.int/immunization/
programmes_systems/supply_chain/en/.
[2] D. Calina, A. O. Docea, D. Petrakis, A. M. Egorov, A. A. Ishmukhametov, A. G. Gabibov, M. I. Shtilman, R. Kostoff, F. Carvalho,
31
M. Vinceti, et al., Towards effective covid-19 vaccines: Updates, perspectives and challenges, International Journal of Molecular
Medicine 46 (2020) 3–16.
ed
[3] M. M. Queiroz, D. Ivanov, A. Dolgui, S. F. Wamba, Impacts of epidemic outbreaks on supply chains: mapping a research agenda
amid the covid-19 pandemic through a structured literature review, Annals of Operations Research (2020) 1–38.
[4] D. D. Wu, D. L. Olson, Comparison with past pandemics, Pandemic Risk Management in Operations and Finance (2020) 7–17.
iew
[5] D. D. Wu, D. L. Olson, The effect of covid-19 on the banking sector, in: Pandemic Risk Management in Operations and Finance,
Springer, 2020, pp. 89–99.
[6] National Governors Association Center for Best Practices, Preparing for the COVID-19 Vaccine and
Considerations for Mass Distribution , 2020. URL: https://www.nga.org/wp-content/uploads/2020/08/
NGA-Memorandum-on-COVID-19-Vaccine-and-Mass-Distribution-Considerations.pdf.
ev
[7] L. Uscher-Pines, S. B. Omer, D. J. Barnett, T. A. Burke, R. D. Balicer, Priority setting for pandemic influenza: an analysis of
national preparedness plans, PLOS medicine 3 (2006) e436.
[8] J. Medlock, A. P. Galvani, Optimizing influenza vaccine distribution, Science 325 (2009) 1705–1708.
r
[9] S. Lee, M. Golinski, G. Chowell, Modeling optimal age-specific vaccination strategies against pandemic influenza, Bulletin of
mathematical biology 74 (2012) 958–980.
er
[10] O. M. Araz, A. Galvani, L. A. Meyers, Geographic prioritization of distributing pandemic influenza vaccines, Health Care
Management Science 15 (2012) 175–187.
pe
[11] K. Buccieri, S. Gaetz, Ethical vaccine distribution planning for pandemic influenza: Prioritizing homeless and hard-to-reach
populations, Public Health Ethics 6 (2013) 185–196.
[12] C. Davila-Payan, J. Swann, P. M. Wortley, System factors to explain 2009 pandemic h1n1 state vaccination rates for children
and high-risk adults in us emergency response to pandemic, Vaccine 32 (2014) 246–251.
ot
[13] M. Biggerstaff, C. Reed, D. L. Swerdlow, M. Gambhir, S. Graitcer, L. Finelli, R. H. Borse, S. A. Rasmussen, M. I. Meltzer, C. B.
Bridges, Estimating the potential effects of a vaccine program against an emerging influenza pandemic—united states, Clinical
infectious diseases 60 (2015) S20–S29.
tn
[14] H.-C. Huang, B. Singh, D. P. Morton, G. P. Johnson, B. Clements, L. A. Meyers, Equalizing access to pandemic influenza
vaccines through optimal allocation to public health distribution points, PloS one 12 (2017) e0182720.
[15] K. Govindan, H. Mina, B. Alavi, A decision support system for demand management in healthcare supply chains considering
the epidemic outbreaks: A case study of coronavirus disease 2019 (covid-19), Transportation Research Part E: Logistics and
rin
Transportation Review 138 (2020) 101967.
[16] X. Chen, M. Li, D. Simchi-Levi, T. Zhao, Allocation of covid-19 vaccines under limited supply, Available at SSRN 3678986
(2020).
ep
[17] U.S. Department of Health and Human Services , HHS Pandemic Influenza Plan, 2005. URL: https://www.cdc.gov/flu/pdf/
professionals/hhspandemicinfluenzaplan.pdf?fbclid=IwAR0KGbTVDQj2SovXHddSNa3k8kRj5_3IJD988kqDfQF5Rvxu1sFDTITtmPE.
[18] S. T. Brown, B. Schreiber, B. E. Cakouros, A. R. Wateska, H. M. Dicko, D. L. Connor, P. Jaillard, M. Mvundura, B. A. Norman,
C. Levin, et al., The benefits of redesigning benin’s vaccine supply chain, Vaccine 32 (2014) 4097–4103.
Pr
[19] M. L. Brandeau, G. S. Zaric, A. Richter, Resource allocation for control of infectious diseases in multiple independent populations:
beyond cost-effectiveness analysis, Journal of health economics 22 (2003) 575–598.
32
[20] K. J. Klassen, T. R. Rohleder, Combining operations and marketing to manage capacity and demand in services, Service
Industries Journal 21 (2001) 1–30.
ed
[21] H. Arora, T. Raghu, A. Vinze, Resource allocation for demand surge mitigation during disaster response, Decision Support
Systems 50 (2010) 304–315.
[22] R. E. Crandall, R. E. Markland, Demand management-today’s challenge for service industries, Production and Operations
iew
Management 5 (1996) 106–120.
[23] B. Y. Lee, T.-M. Assi, K. Rookkapan, A. R. Wateska, J. Rajgopal, V. Sornsrivichai, S.-I. Chen, S. T. Brown, J. Welling, B. A.
Norman, et al., Maintaining vaccine delivery following the introduction of the rotavirus and pneumococcal vaccines in thailand,
PloS one 6 (2011) e24673.
[24] A. S. Abrahams, C. T. Ragsdale, A decision support system for patient scheduling in travel vaccine administration, Decision
ev
Support Systems 54 (2012) 215–225.
[25] E. Shittu, M. Harnly, S. Whitaker, R. Miller, Reorganizing nigeria’s vaccine supply chain reduces need for additional storage
facilities, but more storage is required, Health Affairs 35 (2016) 293–300.
r
[26] B. Y. Lee, L. A. Haidari, W. Prosser, D. L. Connor, R. Bechtel, A. Dipuve, H. Kassim, B. Khanlawia, S. T. Brown, Re-designing
the mozambique vaccine supply chain to improve access to vaccines, Vaccine 34 (2016) 4998–5004.
Research 283 (2020) 182–195. er

[27] Q. Lin, Q. Zhao, B. Lev, Cold chain transportation decision in the vaccine supply chain, European Journal of Operational
[28] R. K. Marcello, V. Papadouka, M. Misener, E. Wake, R. Mandell, J. R. Zucker, Distribution of pandemic influenza vaccine and
pe
reporting of doses administered, new york, new york, usa, Emerging infectious diseases 20 (2014) 525.
[29] T. J. Fitzgerald, Y. Kang, C. B. Bridges, T. Talbert, S. J. Vagi, B. Lamont, S. B. Graitcer, Integrating pharmacies into public
health program planning for pandemic influenza vaccine response, Vaccine 34 (2016) 5643–5648.
[30] M. Turner, Vaccine procurement during an influenza pandemic and the role of advance purchase agreements: Lessons from
ot
2009-h1n1, Global public health 11 (2016) 322–335.
[31] K. Seib, D. J. Barnett, P. S. Weiss, S. B. Omer, Vaccine-related standard of care and willingness to respond to public health
emergencies: A cross-sectional survey of california vaccine providers, Vaccine 31 (2012) 196–201.
tn
[32] Z. Li, J. L. Swann, P. Keskinocak, Value of inventory information in allocating a limited supply of influenza vaccine during a
pandemic, PloS one 13 (2018) e0206293.
[33] Australian Bureau of Statistics, Population Projections, Australia, 2020. URL: https://www.abs.gov.au/statistics/people/
rin
population/population-projections-australia/latest-release.
[34] Department of Health and Human Services, Hospital Locations, 2020. URL: https://discover.data.vic.gov.au/dataset/
hospital-locations-spatial.
ep
[35] Department of Health Australia, COVID-19 deaths by age group and sex, 2020. URL: https://www.health.gov.au/resources/
covid-19-deaths-by-age-group-and-sex{#}:{~}:text=Themajorityofdeathshave,ofconfirmedCOVID-19deaths.
[36] Australian Bureau of Statistics, Australian Statistical Geography Standard, 2016. URL: https://www.abs.gov.au/ausstats/
[email protected]/Lookup/bySubject/1270.0.55.001{~}July2016{~}MainFeatures{~}Mainstructure{~}10002.
Pr
[37] S. Barbieri, L. Jorm, Travel times to hospitals in australia, Scientific data 6 (2019) 1–6.
[38] Victoria State Givernment Planning, Land use and population research, 2019. URL: https://www.planning.vic.gov.au/
land-use-and-population-research/victoria-in-future.
33
[39] DHL, DHL WHITE PAPER DELIVERING PANDEMIC RESILIENCE, Technical Report, 2020. URL: https://www.dhl.com/
global-en/home/insights-and-innovation/thought-leadership/white-papers/delivering-pandemic-resilience.html.
ed
iew
r ev
er
pe
ot
tn
rin
ep
Pr
34

Preprint Not Peer Reviewed: Modeling Vaccine Allocations in The COVID-19 Pandemic: A Case Study in Australia

Uploaded by

Document Information

Original Description:

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

Preprint Not Peer Reviewed: Modeling Vaccine Allocations in The COVID-19 Pandemic: A Case Study in Australia

Uploaded by

Copyright:

Available Formats

Modeling Vaccine Allocations in the COVID-19 Pandemic: A Case Study in

Keywords: COVID-19 Pandemic, Vaccine Supply Chain, Allocation Decisions, Decision-Making

Email addresses: [email protected] (Babak Abbasi ), [email protected] (Masih Fadaki ),

2.1. Vaccine supply chains

consolidated into the ‘health zone’ level.

2.2. Pandemic vaccine supply chains

which will be discussed in the following section.

Authors Objectives Priority group Methodology Variables Findings Relevance

Authors Objectives Priority group Methodology Variables Findings Relevance

copy available at: https://ssrn.com/abstract=3744520

within the execution horizon.

The notations employed in the models of this study are as follows:

Sets and indices

τiout + τiin ≤ 1, ∀i ∈ D (9)

zij ≤ νij , ∀i ∈ D, j ∈ D \ {i} (10)

xij 6 nij , ∀i ∈ D, j ∈ R (11)

yij ≤ M zij , ∀i ∈ D, j ∈ D \ {i} (12)

φi , xij ∈ Z0+ , ∀i ∈ D, ∀j ∈ R (13)

yij ≥ 0, ∀i ∈ D, j ∈ D \ {i} (14)

zij ∈ {0, 1}, ∀i ∈ D, j ∈ D \ {i} (16)

τiin , τiout ∈ {0, 1}, ∀i ∈ D (17)

Residual Risk of the allocation model.

4.1. Cuts to tighten oversupply

Figure 3: Spatial distribution of Victoria’s 325 medical centers

Population of Victoria (.106 )

groups which are assigned to medical center i.

qij ∈ {0, 1}, ∀i ∈ D, ∀j ∈ A (24)

longitude of medical centers.

(a) Cumulative unmet demand (b) Residual risk.

Figure 6: Transshipment of vaccines among medical centers.

5.1. Performance of the vaccine allocation mechanism er

xijt ← fx0 (ntij ), ∀i ∈ D, ∀j ∈ R

return rrt , ntsum

function fxij (φi ):

RBA Heuristic RBA Heuristic

Residual Risk (.103 )

(a) Cumulative unmet demand (b) Residual risk.

(b) Accumulated Residual risk for various capacities of medical

5.3. Impact of daily available vaccine packs (S)

S ∈ {5000, 7500, 10000, 12500, 15000, 17500, 20000}.

Accumulated Residual Risk (.105 )

5.4. Impact of pack size (L)

L ∈ {5, 12, 20, 30, 50}. To ensure

Mean Daily Transshipped Vaccines (.104 )

Daily Transshipped Vaccines (.103 )

5.5. Computational time

î ∈ {10, 20, 30, 40, 50}.

325 500 700 1,000 1,500 2,000

6. Discussion and implications

Transportation Review 138 (2020) 101967.

Research 283 (2020) 182–195. er

2009-h1n1, Global public health 11 (2016) 322–335.

You might also like