Urban-regional disparities in mental health signals in Australia during the COVID-19 pandemic: a study via Twitter data and machine learning models

Abstract

This study establishes a novel empirical framework using machine learning techniques to measure the urban-regional disparity of the public's mental health signals in Australia during the pandemic, and to examine the interrelationships amongst mental health, demographic and socioeconomic profiles of neighbourhoods, health risks and healthcare access. Our results show that the public's mental health signals in capital cities were better than those in regional areas. The negative mental health signals in capital cities are associated with a lower level of income, more crowded living space, a lower level of healthcare availability and more difficulties in healthcare access.

Full text

Cambridge Journal of Regions, Economy and Society
https://doi.org/10.1093/cjres/rsac025
This is an Open Access article distributed under the terms of the Creative Commons Attribution-NonCommercial License
(https://creativecommons.org/licenses/by-nc/4.0/), which permits non-commercial re-use, distribution, and reproduction in any
medium, provided the original work is properly cited. For commercial re-use, please contact journals.permissions@oup.com
© The Author(s) 2022. Published by Oxford University Press on behalf of the Cambridge Political Economy Society.
Urban-regional disparities in mental health signals in
Australia during the COVID-19 pandemic: a study via
Twitter data and machine learningmodels
SiqinWanga,, MengxiZhangb,, XiaoHuangc,, TaoHud,, ZhenlongLie,,
QianChaynSunf,, YanLiua,
aSchool of Earth and Environmental Sciences, University of Queensland, Room 537,
Chamberlain Building, St Lucia, Brisbane, QLD, 4076, Australia, S.W.: s.wang6@uq.edu.
au; Y.L.: yan.liu@uq.edu.au
bDepartment of Nutrition and Health Science, Ball State University, Indiana, USA,
mzhang2@bsu.edu
cDepartment of Geosciences, University of Arkansas, Arkansas, USA, xh010@uark.edu
dDepartment of Geography, Oklahoma State University, Oklahoma, 74078, USA, tao.hu@
okstate.edu
eGeoinformation and Big Data Research Laboratory, Department of Geography,
University of South Carolina, South Carolina, USA, zhenlong@sc.edu;
fSchool of Science, RMIT University, Melbourne, Victoria, Australia, chayn.sun@rmit.edu.au
Received on August 27, 2021; editorial decision on May 20, 2022; accepted on May 30, 2022
This study establishes a novel empirical framework using machine learning techniques
to measure the urban-regional disparity of the public’s mental health signals in Australia
during the pandemic, and to examine the interrelationships amongst mental health, demo-
graphic and socioeconomic proles of neighbourhoods, health risks and healthcare access.
Our results show that the public’s mental health signals in capital cities were better than
those in regional areas. The negative mental health signals in capital cities are associated
with a lower level of income, more crowded living space, a lower level of healthcare avail-
ability and more difculties in healthcare access.
Keywords: COVID-19, mental health, healthcare access, spatial disparity, Twitter, machine
learning models
JEL classications: C31, R50, R23, C23
Introduction
The COVID-19 pandemic and the policies im-
plemented to control the spread of the virus
have created stressful situations and chal-
lenges around the world, including the fear of
contracting the virus, nancial and employment
losses, and government-mandated restrictions on
movement and physical and social interactions
(Betsch, 2020). These challenges are particularly
detrimental to the mental health of individuals.
Downloaded from https://academic.oup.com/cjres/advance-article/doi/10.1093/cjres/rsac025/6610966 by guest on 18 June 2022

Page 2 of 20
Wang etal.
The effects of COVID-19 on people’s mental
health may vary across individuals with different
demographic and socioeconomic status, and dis-
tinguishable across the urban and regional/rural
space with distinct features in demographic and
socioeconomic proles, health risks, and access
to healthcare facilities (Summers-Gabr, 2020;
Sun and Lyu, 2020). Monitoring and measuring
the urban-regional disparity of people’s mental
health status during the pandemic, and exam-
ining the potential factors contributing to such
disparity, are crucial to provide evidence for
place-based health planning policy interven-
tion, as well as the provision of mental health
services towards the post-pandemicera.
The current studies on mental health during
the COVID-19 pandemic apply either survey-
based assessments or advanced modelling tech-
niques, such as machine learning algorithms
(Balcombe, 2020). Survey-based studies con-
ducted in different geographic contexts are
inevitably subject to issues associated with
the small data size, such as high cost, under-
representativeness, and limited spatial and tem-
poral coverage (e.g., Fisher etal., 2020; Newby
etal. 2020; Van Rheenen et al., 2020). Most of
the survey-based studies are limited to a cer-
tain period of the pandemic and focus on the
early stage of the pandemic, which also lacks
geographic information to reect the spatial
variation of mental health conditions. Another
stream of research uses social media data and
advanced modelling techniques to understand
public’s mental reactions to a range of COVID-
related issues such as home schooling (Ewing
and Vu, 2021), social restriction policies (Zhou
etal., 2021) and vaccination (Hu etal., 2021).
However, these studies are limited to cer-
tain states in Australia or relatively short time
periods and also lack the coverage to under-
stand the disparity between capital cities and
regional areas. Subsequently, there is a pressing
need to advance our understanding of people’s
mental health conditions across the urban and
regional areas over the full temporal spectrum
of the pandemic.
To full the knowledge decit, our study
aims to investigate the urban-regional dis-
parity of the public’s mental health during
the pandemic by addressing the following
research questions: 1) How do the public’s
mental health signals vary across capital cities
and regional areas? 2) How do the public’s
mental health signals change along the pan-
demic timeline? and 3) To what extent the
public’s mental health signals are associated
with the demographic and socioeconomic
prole of areas, their health risks and access
to healthcare services? Drawing on 244,406
geotagged tweets in Australia from 1 January,
2020 to 31 May, 2021, we employed machine
learning techniques to measure and classify
the disparity in the public’s mental health sig-
nals across the capital cities and regional areas
in Australia. We further examined the inter-
relationships among mental health, the demo-
graphic and socioeconomic proles of the
neighbourhoods they live, their health risks
and access to healthcare services, drawing on
the demographic and socioeconomic data,
health risk and healthcare access data re-
trieved from the Australian Urban Research
Infrastructure Network (AURIN), 2020. Using
the social media data with a large spatial and
temporal coverage, our study contributes, for
the rst time, a nationwide examination of the
public’s mental health signals in Australia and
the disparity between capital cities and their
regional counterparts. We also demonstrate
a novel empirical framework to systematic-
ally measure, classify, and map mental health
signals nationwide, through which the role
of public health policies and mental health
services can be assessed in the wake of the
global pandemic.
Background
Measuring mental health in the
COVID-19studies
Current studies on mental health relating to the
COVID-19 pandemic are largely survey-based,
Downloaded from https://academic.oup.com/cjres/advance-article/doi/10.1093/cjres/rsac025/6610966 by guest on 18 June 2022

Page 3 of 20
Urban-regional disparities in mental health signals in Australia
focusing on certain social groups or popula-
tions in particular regions (Balcombe, 2020).
Survey-based studies have been conducted
with different sizes of study populations in dif-
ferent geographic contexts; however, they are
inevitably subject to small-data issues such as
the limited sample sizes and temporal cover-
ages due to restricted nancial resources (e.g.,
Fisher etal., 2020; Van Rheenen etal., 2020).
These studies offer some valuable insights on
factors that may increase a person’s vulner-
ability to experiencing psychological distress
such as depression, nancial stress, health anx-
iety, contamination fears accompanied by the
increasing level of alcohol use and decreasing
level of physical activities (e.g., Newby et al.,
2020; Van Rheenen etal., 2020). However, most
existing studies focus on the early stage of the
pandemic, with survey data also lacking geo-
graphic information to reect the spatial vari-
ation of mental health signals.
Another set of research utilises
crowdsourcing data, combined with the rap-
idly evolving computational infrastructures
and intelligent algorithms (e.g., machine and/
or deep learning) that offer exciting possibil-
ities for monitoring both population-level and
individual-level mental health status (Cotfas
et al., 2021). In particular, social media data,
as a well-established data source that has
been applied in politics, businesses and dis-
aster management, has been increasingly used
in population health monitoring and other
mental health applications (Conway and
O’Connor, 2016). Qualitative or text-based
data in social media (e.g., third-person pro-
nouns and anger words) are the potential in-
dicators of social media users’ self-reported
mental health problems (Coppersmith et al.,
2014). The massive and insightful content por-
trayed and outlined by highly engaged social
media users provide unprecedented opportun-
ities for collective emotion and affective ana-
lysis (Liang etal., 2019). Based on the nature
of social media data, a series of analyses via ad-
vanced machine learning and natural language
processing techniques have been developed to
monitor and track the public’s mental health
signals towards vaccination (Hu etal., 2021),
social distancing policies, stay-at-home and
lockdown orders (Zhou etal., 2021) and work-
from-home requirements (Ewing and Vu,
2021) during the pandemic. There is great po-
tential for using machine learning techniques
within mental health studies, while there is
an emerging critique that the effective ap-
plication of machine learning is mediated by
research design and bound up with a wide
range of complex, interwoven challenges
(Thieme etal., 2020). These challenges include
generating large-scale, high-quality datasets
to represent the diversity of the population
(Bone et al., 2017), mitigating the obstacles
(e.g., errors, uncertainty and bias) for the de-
ployment of machine learning algorithms into
real-world systems (Thornicroft etal., 2007),
as well as considering far reaching personal,
societal and economic implications in mental
health contexts (Rudin etal., 2019). Despite of
these challenges, machine learning techniques
have offered new routes for learning pat-
terns of human behaviour, identifying mental
health symptoms and risk factors, and assisting
in the detection, diagnosis and treatment of
mental health problems. Considering that trad-
itional survey methods are time and labour-
consuming with limited spatial and temporal
coverages, investigating the public’s mental
health during the COVID-19 pandemic over a
longer timeline and with a larger spatial scale,
and exploring the potential factors impacting
on the disparity in mental health between
urban and regional areas using public sourced
social media data is needed.
Mental health and demographic
and socioeconomic proles of
neighbourhoods
The current studies examining the preva-
lence of mental health problems have been
largely inuenced by Diez Roux’s pathways
Downloaded from https://academic.oup.com/cjres/advance-article/doi/10.1093/cjres/rsac025/6610966 by guest on 18 June 2022

Page 4 of 20
Wang etal.
theory (Diez-Roux, 1998). The pathways
theory describes how the demographic and
socioeconomic prole at the individual level
and collectively at the aggregated population
level (e.g., neighbourhoods) might contribute
to the disparities of mental health across areas
via individual and contextual pathways. The
socioeconomic prole of neighbourhoods is
commonly measured in several ways, including
labour forces, income and nancial status (e.g.,
mortgage stress), which may further affect
travel and residential choices and dwelling
types (e.g., living space, difculty in moving
or access to transport) (Meyer et al., 2014).
These measures in socioeconomic proles of
neighbourhoods are particularly important in
investigating mental health status during a pan-
demic. The common ndings show that mental
health problems (e.g., depression) are more
likely to appear in low-socioeconomic neigh-
bourhoods (Meyer etal., 2014). These studies
also imply that the mental health problems that
are associated with the pandemic appear more
prevalent in low-socioeconomic areas with a
concentrated low-income and unemployed
population, or more essential workers who
need to work onsite, as well as in areas with
relatively crowded living space where the virus
transmission is high (Zhang et al., 2021). In
addition, the socioeconomic prole can be fur-
ther compounded by the demographic prole
of neighbourhoods, such as the age structure
and race/ethnicity. For instance, the ageing
neighbourhoods (e.g., retirement villages)
and minority-concentrated suburbs (e.g., the
Hispanic group in the U.S.) are subject to more
severe mental health problems (e.g., worry
and fear of infection), given the higher infec-
tion rate among the elderly and the Hispanic
group comprising a higher proportion of es-
sential workers compared to non-Hispanic
groups (Penner etal., 2021). Given these con-
siderations, we take onboard a set of indicators
measuring the demographic and socioeconomic
proles of neighbourhoods as potential factors
inuencing the public’s mental health.
Mental health and physicalhealth
Another set of factors potentially associated
with mental health are individuals’ physical
health status, such as chronic conditions (Talen
and Mann, 2009). People with chronic diseases
may experience negative emotions that further
increase the probability of developing mental
health issues. Such chronic diseases, including
hypertension (Sparrenberger etal., 2009) and
overweight and obesity (Talen and Mann, 2009),
have been widely discussed in the current lit-
erature and observed to be associated with the
onset of anxiety, stress and depression (Kretchy
etal. 2014; Pan etal., 2015). Youth and adoles-
cents who were overweight were more likely
to report self-stigma, depression, anxiety and
feelings of worthlessness (Chan et al., 2019),
which may have long-lasting consequences on
mental health. Furthermore, health risk behav-
iours such as alcohol consumption and smoking
are potentially associated with various mental
health issues, and poor mental health could be
an enduring risk factor for heavy alcohol con-
sumption (Shahab et al., 2014). The onset of
these chronic conditions and health risk be-
haviours at the individual level assembles the
prevalence of health risks at the aggregated
population level, that is, the level of health risks
in neighbourhoods. Subsequently, our study
takes into account health risks in neighbour-
hoods when examining people’s mental health
during the pandemic.
Mental health and the spatial disparity
of healthcareaccess
A study by Fisher etal. (2020) shows that around
25% of Australians had reported experiencing
mild to moderate symptoms of depression or
anxiety at the early stage of the COVID-19
pandemic. Lack of access to mental healthcare
services and shortage of mental health providers
may result in such mental health issues not
being resolved properly or in a timely fashion
(Lake and Turner, 2017). In Australia, the access
to mental health services in rural/regional areas
Downloaded from https://academic.oup.com/cjres/advance-article/doi/10.1093/cjres/rsac025/6610966 by guest on 18 June 2022

Page 5 of 20
Urban-regional disparities in mental health signals in Australia
is considerably lower than that in major cities
(Australian Institute of Health and Welfare
(AIHW), 2019). In particular, people living in
remote regional areas reported three times less
access to mental health services subsidised by
Medicare (i.e., Australia’s universal health in-
surance scheme), compared to those who live
in major cities (AIHW, 2019). The low rate of
access to mental health services might be at-
tributed to the limited number of mental health
professionals and healthcare facilities in rural
and remote Australia. Recent studies indicate
an unequal distribution of mental health pro-
fessionals (e.g., nurses and psychologists) across
the metropolitan areas and regional/rural areas
in Australia, with major cities having the most
adequate workforce resource compared to
other areas (Sutarsa et al., 2021). Apart from
the availability/provision of healthcare services,
access to mental health services is also inu-
enced by other barriers that may access, such
as the lack of connectivity to public transport
services or the lack of private vehicles to drive
to the healthcare facilities. Subsequently, people
tend to use healthcare services and facilities
that are within shorter travel distances or time
from their homes more often (Ghorbanzadeh
etal., 2020). Thus, it is critical to adjust for the
availability of healthcare services and the level
of difculty in healthcare access in the examin-
ation of mental health across capital cities and
regionalareas.
In summary, through an empirical study in
Australia, the key objective of our study is to
provide nuanced but plausible insights into
understanding the mental health of a nation
and its urban-regional disparities, which is ap-
plicable in other countries aiming for better
spatial justice and social harmony.
Study context, data and analysis
Studycontext
Australia is the largest developed country in the
Southern Hemisphere, with a total population
of nearly 26 million and a total area of around
7.61 million square kilometres (Australia
Bureau of Statistics (ABS), 2020). Australia is
highly urbanised, with over 80% of its popula-
tion living in cities. The nation’s capital city is
Canberra, also known as the Australian Capital
Territory (ACT), and the other states/terri-
tories are (the state capital cities are listed in
brackets): New South Wales (Sydney), Victoria
(Melbourne), Queensland (Brisbane), West
Australia (Perth), South Australia (Adelaide),
Tasmania (Hobart), and Northern Territory
(Darwin). According to the Greater Capital
City Statistical Area Structure (ABS, 2016),
each state is divided into a greater capital city
area and the remaining regional area. For in-
stance, the State of New South Wales (NSW)
is divided into the Greater Sydney Area and
Rest of NSW. In this paper, we simplied the
terminology using Sydney and Beyond Sydney
to represent Greater Sydney and the Rest of
NSW, respectively, in later analysis (Figure 1);
such simplication also applied to other states/
territories with the exception of ACT having
only the capital city for the whole territory.
Over 66% of the Australian population live in
the greater capital city areas of the eight states/
territories, with Sydney being the largest (with
around 4.9 million population in 2016), fol-
lowed by Melbourne (4.5 million) and Brisbane
(2 million) (ABS, 2020). Our analysis rst
looked at the comparison between the capital
cities aggregated as one unit and the regional
areas, also aggregated as one unit, and then fo-
cused on the comparison between each capital
city and its regional counterpart.
Data
We utilised Twitter academic full track ap-
plication programming interface (AFT-API)
to search and retrieve geotweets in Australia
(Twitter, 2020). Compared to the normal
Twitter API that returns 1% of the total tweets
due to privacy concerns, AFT-API enables us
to fully retrieve tweets with the pre-dened
Downloaded from https://academic.oup.com/cjres/advance-article/doi/10.1093/cjres/rsac025/6610966 by guest on 18 June 2022

Page 6 of 20
Wang etal.
queries, which improves the data coverage
and representativeness (Twitter, 2020). We de-
ned the searching terms as ‘pandemic, epi-
demic, virus, covid*, and vaccine*’; the search
timespan was dened as from 1 January, 2020 to
31 May, 2021 with the country dened as ‘AU’
(Australia). Consequently, 244,406 geotweets
were retrieved from the total of over 860 mil-
lion tweets in Australia. These geotweets con-
tain X, Y coordinates that were retrieved in two
ways: 1)an accurate pair of X, Y coordinates of
a place where a geotweet was posted if a user
activated the positioning function in Twitter,
and 2)an rough pair of X, Y coordinates con-
verted from the name of a place (including
cities and neighbourhood) if a user only indi-
cated the name of that place in Twitter; such
places were then assigned the X, Y coordinates
as the centroid of that place. The geotweets
were further aggregated to Statistical Area
level 2 (SA2) as the second smallest unit in the
Greater Capital City Statistical Area Structure
(ABS, 2016) given that SA2 areas with an
average population of about 10,000 persons
serve as an appropriate unit compared to the
full size of capital cities and the name of SA2
areas are identiable. The SA2 areas with the
number of tweets less than 17 were excluded,
given it was assumed that there should be at
least one tweet posted in each SA2 over the
entire research period of 17months (Jan 2020
to May 2021). We recognised that there were
a disproportionate number of tweets concen-
trated in the SA2 area where the centroid of
a capital city was located (not always the city
centre) if users roughly tag a city in a tweet but
they may reside in other SA2 areas. Thus, we
further excluded eight SA2 areas containing
the centroid of each capital city (e.g., the SA2
named ‘Hughesdale’ containing the centroid of
Melbourne) to reduce the data bias. The spatial
coverage of geotweets ranged from the highest
of 69.9% in Melbourne to the lowest of 28.24%
in Canberra, shown in the statistical summary
Figure 1. Capital cities and regional areas in Australia
Downloaded from https://academic.oup.com/cjres/advance-article/doi/10.1093/cjres/rsac025/6610966 by guest on 18 June 2022

Page 7 of 20
Urban-regional disparities in mental health signals in Australia
and distribution of tweets provided in the sup-
plementary material (Table S1 and Figure S1).
Demographic and socioeconomic data, and
health risk and healthcare access data were
retrieved from the AURIN online data portal
(2020) at the level of Population Health Areas
(PHA). PHA was developed by Public Health
Information Development Unit (2016) and
comprised a combination of whole SA2s and/
or multiple segmented SA2s. We then used
a correspondence le, as a cross-tabulated
table containing the proportions of each SA2
falling into each PHAs), to convert the data
in PHAs to SA2s in 2016, in order to join with
the spatial boundary of SA2s and to align
with the geotweets data aggregated at the
SA2 level (ABS, 2016). The demographic and
socioeconomic prole of one SA2 includes
the measures of age structures, income, labour
force, living space, needs for childcare, whether
to have vehicles and be able to move. Health
risk data include the proportion of people with
poor health status, alcohol use, overweight, and
high blood pressure in one SA2. Healthcare ac-
cess data include the measures of hospital ad-
missions and general practitioners (GP), and
difculty in healthcare access. More detailed
denitions of these measures are provided in
Table 1.
Analysis
Machine learning models to detect
sentiment andemotion
We commenced with a series of data pre-
processing to the geotweet records using
Python 3.9.6 (details provided in ‘Data pre-
processing’ section in the Supplementary
materials). We then employed the Valence
Aware Dictionary for sEntiment Reasoning
(VADER) model to analyse the sentiment of
each geotweet (VADER, 2021). The VADER
model is a lexicon and rule-based sentiment
analysis tool that has been specically attuned
to sentiments expressed in qualitative contexts
(e.g., social media posts) (Hutto et al., 2014).
VADER sentiment analysis relies on a machine
learning algorithm and an open-source dic-
tionary library that maps lexical features and
heuristic expressions to emotion intensities
known as sentiment scores (VADER, 2021). It
returns four sentiment scores, including posi-
tive, negative, neutral, and compound scores.
The rst three scores are measured as a ratio of
the number of words that fall in the respective
categories (positive, negative and neutral sen-
timent) to the total number of words in each
geotweet record, respectively. The compound
score is a weighted composite score that is fur-
ther generated based on the ratio of positive,
negative and neutral sentiment. The value of
a compound score ranges from ‐1 (most nega-
tive) to +1 (most positive). In this study, we
used the compound score as the indicator of
mental health given it is comparable across dif-
ferent geographic contexts (e.g., capital cities
and regional areas). More details of VADER
are provided in the work by Hutto etal. (2014).
To further interrogate the wide range of
emotions that may be not reected roughly by
the positive and/or negative sentiment gener-
ated by VADER, we further used NRCLex,
developed based on the National Research
Council Canada Affect Lexicon and the
Natural Language ToolKit (NLTK) library’s
WordNet synonym sets (Bird etal., 2009), con-
taining approximately 27,000 words, to detect
the emotional tendency of a given body of texts
in geotweets. NRCLex differentiates the types
of emotions via a word matching algorithm
based on a documented affection dictionary
and the association of the texts with four pairs
of primary bipolar emotions (i.e., eight basic
emotions): joy (feeling happy) versus sadness
(feeling sad); anger (feeling angry) versus fear
(feeling of being afraid); trust (stronger ad-
miration and weaker acceptance) versus dis-
gust (feeling something is wrong or nasty);
and surprise (feeling unprepared for some-
thing) versus anticipation (looking forward
positively to something). Among these eight
types of emotions, fear, sadness, anger and
Downloaded from https://academic.oup.com/cjres/advance-article/doi/10.1093/cjres/rsac025/6610966 by guest on 18 June 2022

Page 8 of 20
Wang etal.
Table 1. Name and denition of variables used in the regression modelling.
Dependent variable Denition
Binary variable representing a SA2
belonging to either the positive or
negative group in mental health signals
1 (positive group): SA2s with a sentiment score higher than the average of all SA2s in the study area ** (this is
the reference group)
0 (negative group):SA2s with a sentiment score lower than the average of all SA2s in the study area **
Independent variables
Demographic and socioeconomic proles
Age 65 and over % of people at and above 65 over the total population in one SA2
Aboriginal*% of aboriginal population over the total population in one SA2
Unemployment % of unemployed people over the total population in one SA2
Low-income % of low-income households over the total number of households in one SA2
Mortgage stress % of households under nancial stress from mortgage or rent over the total number of households in one SA2
No vehicle*% of households with no motor vehicle over the total number of households in one SA2
Crowded living*% of people living in crowded dwellings over the total population in one SA2
Childcare need % of households needed by childcare over the total number of households in one SA2
Difculty in moving % of people aged 18years and over who encountered barriers to access to places (e.g. lack of public transport)
over the total population in one SA2
Health risk
Poor health % of people aged 15years and over with fair or poor self-assessed health over the total population in one SA2
Alcohol use % of people aged 18years and over who consumed more than two standard alcoholic drinks per day on
average over the total population in one SA2
Overweight % of people aged 18years and over who were overweight over the total population in one SA2
High blood pressure % of people aged 18years and over who had high blood pressure over the total population in one SA2
Healthcare access
Hospital and general practitioners (GP) Total admissions of hospitals and GP numbers per 100 population in one SA2
Difculty to access healthcare service % of people aged 18years and over who experienced a barrier to accessing healthcare when needed it in the
last 12months over the total population in one SA2
*This variable was log transformed to ensure the data is normally distributed.
**The study area is all regional areas in Australia for Model 1, all capital cities in Australia for Model 2, Greater Sydney metropolitan area for Model 3,
Greater Melbourne metropolitan area for Model 4, and Greater Brisbane metropolitan area for Model 5.
Downloaded from https://academic.oup.com/cjres/advance-article/doi/10.1093/cjres/rsac025/6610966 by guest on 18 June 2022

Page 9 of 20
Urban-regional disparities in mental health signals in Australia
disgust correspond to negative mental sig-
nals, while joy, anticipation, trust and surprise
correspond to positive mental signals. The re-
sult of NRCLex was the number of words in
each type of emotion in each geotweet record,
which was further aggregated to a certain spa-
tial unit (e.g., a capital city) and calculated as
the proportion of words in each type of emo-
tion over the total (named as percentages of
emotion hereinafter).
Binary logistic regression
We then applied a binary logistic regression
(BLR) (Hosmer et al., 1997) to examine how
the mental health signals associate with the
three sets of indicators measuring the demo-
graphic and socioeconomic proles, health
risks and healthcare access. SA2s with the com-
pound sentiment score was transformed into
a binary variable by comparing the sentiment
score in each SA2 with the average in all SA2s
of the study area, with 1 indicating higher-
than-the-average SA2s (termed as the positive
group; this group is set as the reference group in
subsequent modelling) and 0 indicating lower-
than-the-average SA2s (termed as the nega-
tive group). This binary variable is used as the
dependent variable in the BLR, given that a
binary dependent variable is most common in
logistic regression which can produce a prob-
ability of a certain class, i.e., a SA2 being posi-
tive or negative in mental health signals. The
BLR model is expressedas:
Log
ÅP
(
Y
=
0
)
P(Y=1)
ã
=β0+β1x1+β2x2+β3x3···+βixi
(1)
where,
Y
is the dependent variable with two
possible values, i.e.,
Y=1
(the reference group)
when the sentiment score in a SA2 is higher
than the average sentiment score in all SA2s
and
Y=0
otherwise;
x1
to
xi
are the inde-
pendent variables dened in Table 1; and
β0
to
βi
are the regression coefcients for each vari-
able. The odds (
o
) indicating the likelihood of a
SA2 belonging to a positive or negative SA2 is
computedas:
o=10β
0
+β
1
x
1
+β
2
x
2
+β
3
x
3
···+β
n
xn
(2)
The corresponding probability (p) is
calculatedas:
p=
o
1+o
(3)
We ran a total of ve BLR models. Models
1 and 2 are for regional areas and capital cities,
respectively, and Models 3–5 are for Sydney,
Melbourne and Brisbane as the three largest
capital cities, respectively. Within each model,
there are three sub-models (e.g., Models 1–1,
1–2, and 1–3), with the rst sub-model involving
the demographic and socioeconomic prole of
neighbourhoods as the independent variables,
the second also including the level of health
risks, and the third further including healthcare
access factor to the model. The signicance
levels were set at 0.1, 0.05 and 0.01.
Results
Change of mental health signals along
the pandemic timeline
Figure 2 shows the temporal change of the
public’s mental health signals, illustrated by the
sentiment scores in capital cities and regional
areas. The sentiment score in capital cities is
higher than that in regional areas over the whole
timeline of the pandemic except for two short
periods, one in January 2020 and one in March
2021. From February to March 2020, the senti-
ment scores in both capital cities and regional
areas had an obvious increase and continued
to increase afterwards to May 2020 but slightly
decreased from May to July 2020. This cor-
responds to the second wave of the pandemic
Australians experienced during this period.
From September to November 2020, the senti-
ment scores in both capital cities and regional
Downloaded from https://academic.oup.com/cjres/advance-article/doi/10.1093/cjres/rsac025/6610966 by guest on 18 June 2022

Page 10 of 20
Wang etal.
areas increased to the peak in November 2020.
From December 2020 to March 2021, the sen-
timent score in capital cities stayed relatively
stable compared to that in regional areas ex-
periencing more uctuations. However, the sen-
timent scores in both capital cities and regional
areas had an obvious decrease from March to
April 2021. Overall, the public’s mental health
signals in capital cities were observed to be
higher than that of their counterparts in the
regional areas over the year-long period from
February 2020 to February2021.
The temporal change of the eight types of
emotions provides a more detailed map of the
variations of the public’s mental health sig-
nals between the capital cities and the regional
areas (Figure 3). At the early stage of the pan-
demic (January to March 2020), the feelings
of fear had an obvious decrease in both the
capital cities and regional areas while there
were increases in trust and joy. From March
2020 to May 2021, the overall trend of these
eight types of emotion in both capital cities
and regional areas are relatively stable, and the
feeling of fear accounts for the largest propor-
tion of emotion (reected by the highest pos-
ition of the dark blue solid line), followed by
trust, anticipation, sadness, joy, anger, surprise
and disgust. The comparison of the percentages
of emotion in capital cities (solid lines) and re-
gional areas (dash lines) shows that there are
no substantial differences in the percentages
of emotion between capital cities and regional
areas from March 2020 to May 2021, reected
by the small gap in a pair of solid and dash lines
in one colour; while before March 2020, in re-
gional areas there was a higher percentage in
fear and a lower percentage in trust compared
to capital cities. However, during the rest of the
pandemic, the variations of emotion in between
capital cities and regional areas is less clear,
with the observation that the percentage in a
certain type of emotion in capital cities is higher
than that in regional areas in some months (e.g.,
April to July 2020 for fear) but lower in other
months (e.g., April to May 2021 forfear).
Spatial disparity of mental healthsignals
We present the sentiment scores for each greater
city area and their respective regional counter-
parts (e.g., Sydney and Beyond Sydney as we
dened in Section 3.1), illustrating the subtle
variations of the public’s mental health signals
of the country (Figure 4). Overall, the mean
values of sentiment scores in all capital cities of
the eight states/territories are higher than that
of their regional counterparts, indicating that
the public’s mental health signals in cities tend
to be more positive than those in regional areas.
Moreover, the range of sentiment scores in one
capital city (e.g., Sydney, Melbourne, Brisbane,
and Perth) is wider than that in its counter-
part (e.g., beyond Sydney, beyond Melbourne,
Figure 2. Temporal change of the sentiment score in capital cities and regional areas
Downloaded from https://academic.oup.com/cjres/advance-article/doi/10.1093/cjres/rsac025/6610966 by guest on 18 June 2022

Page 11 of 20
Urban-regional disparities in mental health signals in Australia
beyond Brisbane, and beyond Perth, respect-
ively). In addition, the range of sentiment scores
in the three largest state capital cities (i.e.,
Sydney, Melbourne, and Brisbane) is wider than
that in the medium- and small-size capital cities
(i.e., Perth, Adelaide, Hobart and Darwin). This
Figure 3. Temporal change of emotion by type in capital cities and regional areas
Figure 4. Box-plots illustrating the variations of sentiment scores between capital cities and regional areas
Downloaded from https://academic.oup.com/cjres/advance-article/doi/10.1093/cjres/rsac025/6610966 by guest on 18 June 2022

Page 12 of 20
Wang etal.
indicates that although the mental health status
of people in capital cities tends to be more posi-
tive than those in the regional areas, such sig-
nals are also more diverse and varied in cities
than their regional counterparts.
Figure 5 reveals the variations of the emo-
tions by type in the eight capital cities and
their respective regional areas. In each of the
graphs in Figure 5, a bar in a positive direction
indicates a certain type of emotions (coloured
differently) has a larger percentage in the city
than that in its regional counterpart; the height
of the stacked bar in each month reects the
extent the sum of the percentages for different
types of emotion in the capital cities is higher
or lower than that its counterpart of regional
areas. In January and February 2020, all capital
cities except Brisbane have higher percentages
of negative emotion (including fear, anger, sad-
ness and disgust, as illustrated in a blue colour
theme in Figure 5) compared to the regional
areas, shown by the stacked bar in the positive
direction. In contrast, Brisbane has a higher
percentage of positive emotion (including joy,
trust, anticipation and surprise, as illustrated
in a red–orange colour theme in Figure 5)
compared to its counterpart of regional areas
in January 2020. From February 2020 to May
2021, there are relatively minor variations in
emotion between Melbourne and Brisbane
and their regional areas compared to other
capital cities. From January to April 2021, the
percentages of positive emotion in Brisbane,
Perth, Adelaide and Darwin are higher than
their counterparts of regional areas, while the
percentages of negative emotion in the re-
gional areas beyond Sydney, Melbourne and
Hobart are higher than the percentages in
their capital cities. It is quite surprising to see
that Melbourne, where the second wave of the
pandemic in Australia was centered had higher
percentages in positive emotion than its re-
gional areas from June to October 2020, but
shifted to higher percentages in negative emo-
tion than its regional areas from November
2020 to March 2021. It is possibly due to the
Figure 5. Variations of the emotion types between capital cities and regional areas
Downloaded from https://academic.oup.com/cjres/advance-article/doi/10.1093/cjres/rsac025/6610966 by guest on 18 June 2022

Page 13 of 20
Urban-regional disparities in mental health signals in Australia
long-term implementation of social restriction
policies (e.g., lockdown and home-dwelling or-
ders) after the second wave of the pandemic
from June to October2020.
Relationships between mental
health signals and the demographic
and socioeconomic proles of
neighbourhoods, health risks and
healthcareaccess
We further explored the extent to which
the spatial and temporal variations of the
mental health signals between capital cities
and regional areas are associated with the
demographic and socioeconomic prole
of neighbourhoods, their health risks and
healthcare access. In Model 1 (including
sub-Models 1–1, 1–2, and 1–3 in Table 2) ,
low-income has negative coefcients (at least
p < 0.1, and with odds ratios smaller than
1) while the areas with crowded living and
childcare need have positive coefcients (at
least p<0.1, correspondingly with odds ratios
larger than 1). It means that people living in
regional areas with lower income levels, more
crowded living spaces, and higher needs for
childcare are more likely to exhibit negative
mental signals. The proportion of the abori-
ginal population is observed to have negative
coefcients (at least p< 0.1, correspondingly
with odds ratios smaller than 1)only in Model
1–2 and 1–3. By adding the health risks and
healthcare access variables in Models 1–2 and
1–3, the Nagelkerke R2 increased slightly, from
0.12 in Model 1–1 to 0.13 in Model 1–2 and
0.14 in Model 1–3; however, none of the health
risks and healthcare access variables are sig-
nicant at p<0.1. This implies that the level of
health risks and healthcare access in regional
areas is not relevant to the mental health sig-
nals of people living in those regions.
For capital cities (Model 2, including Model
2–1, 2–2 and 2–3 in Table 3), the demographic
and socioeconomic proles of neighbour-
hoods that are signicantly (at least p < 0.1)
associated with mental health signals include
age at and above 65 (a positive coefcient of
0.086 in Model 2–1 correspondingly the odds
ratio larger than 1), aboriginal (the odds ratio
smaller than 1 in Model 2–3), unemployment
and low-income (all odds ratios smaller than
1), crowded living (odds ratios larger than 1 in
Model 2–1 and 2–3), and difculty in moving
(all odds ratios larger than 1). It means that
people living in areas with a smaller proportion
of aboriginal population, more crowded living
space, and more difculty in moving and trans-
port tend to exhibit negative mental signals.
The additional involvement of health risk vari-
ables in Model 2–2 increases the Nagelkerke
R2 from 0.21 in Model 2–1 to 0.33 in Model
2–2. All health risk and healthcare access vari-
ables are observed to be signicantly (at least
p<0.05) associated with mental health signals.
More specically, poor health, alcohol use and
high blood pressure have positive coefcients
(odds ratios larger than 1); while being over-
weight has negative coefcients (odds ratios
smaller than 1). It reveals that people living in
the areas with a higher proportion of the popu-
lation having issues in poor health, alcohol use
and high blood pressure tend to be more likely
to display negative mental signals. Further
adding in healthcare access variables in Model
2–3 increases the Nagelkerke R2 from 0.33
in Model 2–2 to 0.48 in Model 2–3 and both
healthcare access variables are signicantly
(p < 0.01) associated with mental health sig-
nals. More specically, the number of hospital
admissions and GPs has an odds ratio of 0.598
[0.391, 0.915] while difculty to healthcare ac-
cess has an odds ratio of 1.047 [1.032, 1.062]. It
means that people residing in areas with lower
availability of hospitals and GPs and more dif-
culties in healthcare access tend to be more
likely to have negative mental signals. This
nding in capital cities is distinct from that in
regional areas, reecting the spatial disparity
of health risks and healthcare access in capital
cities and regional areas that is further related
to the public’s mental healthstatus.
Downloaded from https://academic.oup.com/cjres/advance-article/doi/10.1093/cjres/rsac025/6610966 by guest on 18 June 2022

Page 14 of 20
Wang etal.
Table 2. Regression results for regional areas.
Regional areas (beyond capital cities) Model 1–1 Model 1–2 Model 1–3
Coefcient Odds ratio A
(95% CI)
Coefcient Odds ratio A
(95% CI)
Coefcient Odds ratio A
(95% CI)
Demographic and socioeconomic prole
Age 65 and over 0.059 1.061 (0.985, 1.143) 0.056 1.058 (0.979, 1.143) 0.049 1.051 (0.971, 1.137)
Aboriginal ‐0.062 0.939 (0.864, 1.021) ‐0.083* 0.921 (0.833,1.018) ‐0.082* 0.921 (0.833, 1.018)
Unemployment 0.101* 1.106 (0.974, 1.256) 0.100 1.106 (0.961, 1.272) 0.098 1.102 (0.959, 1.267)
Low-income ‐0.037** 0.964 (0.929, 1.000) ‐0.041** 0.960 (0.921, 1.001) ‐0.039* 0.962 (0.922, 1.003)
Mortgage stress ‐0.029 0.971 (0.928, 1.016) ‐0.029 0.972 (0.925, 1.020) ‐0.037 0.964 (0.914, 1.017)
No vehicles 0.016 1.016 (0.949, 1.088) 0.022 1.022 (0.950, 1.100) 0.028 1.028 (0.954, 1.108)
Crowded living 0.185** 1.204 (1.041, 1.392) 0.212** 1.236 (1.046, 1.460) 0.207** 1.230 (1.042, 1.453)
Childcare need 0.062* 1.064 (0.990, 1.143) 0.074* 1.077 (0.996, 1.165) 0.071* 1.074 (0.992, 1.162)
Difculty in moving ‐0.039 0.962 (0.690, 1.343) ‐0.105 0.90 0 (0.593, 1.368) ‐0.145 0.865 (0.563, 1.329)
Health risk
Poor health 0.032 1.033 (0.920, 1.160) 0.019 1.019 (0.903, 1.150)
Alcohol use 0.025 1.026 (0.949, 1.108) 0.024 1.024 (0.948, 1.106)
Overweight ‐0.0 10 0.990 (0.876, 1.119) 0.002 1.002 (0.883, 1.136)
High blood pressure ‐0.021 0.979 (0.782, 1.226) ‐0.029 0.972 (0.777, 1.216)
Healthcare access
Hospital and GP ‐0.004 0.996 (0.986, 1.007)
Difculty in healthcare access 0.177 1.193 (0.738, 1.930)
Nagelkerke R20.12 0.13 0.14
OCP 65.92 67.41 67.16
Number of SA2 660 660 660
Note: *p<0.1; **p<0.05; **p<0.01.
The reference group is the areas concentrated by people with relatively positive mental status.
CI, condence interval; OCP, an overall correct percentage of prediction.
ANumbers in the bracket under each odd radio is the range of condence interval of that odd radio at the level of 95%.
Downloaded from https://academic.oup.com/cjres/advance-article/doi/10.1093/cjres/rsac025/6610966 by guest on 18 June 2022

Page 15 of 20
Urban-regional disparities in mental health signals in Australia
Table 3. Regression results for capital cities.
All eight capital cities Model 2–1 Model 2–2 Model 2–3
Coefcient Odds ratio A
(95% CI)
Coefcient Odds ratio A
(95% CI)
Coefcient Odds ratio A
(95% CI)
Demographic and socioeconomic prole
Age 65 and over 0.086*** 1.09 (1.021, 1.164) 0.599 1.82 (1.334, 2.483) 0.036 1.037 (0.954, 1.126)
Aboriginal 0.011 1.011 (0.968, 1.056) 0.078 1.081 (1.005, 1.163) ‐0.039** 0.962 (0.904, 1.024)
Unemployment ‐0.210*** 0.811 (0.717, 0.917) ‐0.042* 0.959 (0.90 1, 1.02) ‐0.126* 0.882 (0.759, 1.025)
Low-income ‐0.026* 0.975 (0.941, 1.01) ‐0.100 0.905 (0.784, 1.043) ‐0.004** 0.996 (0.951, 1.044)
Mortgage stress 0.019 1.019 (0.978, 1.062) ‐0.019 0.981 (0.94, 1.023) 0.033 1.033 (0.981, 1.088)
No vehicles 0.022 1.022 (0.955, 1.093) 0.012 1.012 (0.968, 1.059) ‐0.014 0.986 (0.909, 1.071)
Crowded living 0.041** 1.042 (0.979, 1.109) ‐0.006 0.994 (0.921, 1.074) 0.092*** 1.097 (1.014, 1.186)
Childcare need 0.035 1.036 (0.963,1.114) 0.105*** 1.111 (1.026, 1.202) ‐0.040** 0.961 (0.882, 1.047)
Difculty in moving 0.599*** 1.82 (1.334, 2.483) 0.005** 1.005 (0.928, 1.089) 0.268*** 1.308 (0.832, 2.055)
Health risk
Poor health 0.219*** ‐0.803 (‐0.909, ‐0.71) 0.112** 1.118 (1.019, 1.226)
Alcohol use 0.078** 1.081 (1.001, 1.168) 0.072* 1.074 (0.991, 1.164)
Overweight ‐0.292*** 0.747 (0.652, 0.856) ‐0.198*** 0.820 (0.709, 0.95)
High blood pressure 0.488*** 1.629 (1.294, 2.05) 0.304** 1.355 (1.064, 1.724)
Healthcare access
Hospital and GP ‐0.513** 0.598 (0.391, 0.915)
Difculty in healthcare access 0.046*** 1.047 (1.032, 1.062)
Nagelkerke R20.21 0.33 0.48
OCP 66.40 71.70 74.5
Number of SA2 705 705 705
Note: *p<0.1; **p<0.05; **p<0.01.
The reference group is the areas concentrated by people with relatively positive mental status.
CI, condence interval; OCP, an overall correct percentage of prediction.
ANumbers in the bracket under each odd radio is the range of condence interval of that odd radio at the level of 95%.
Downloaded from https://academic.oup.com/cjres/advance-article/doi/10.1093/cjres/rsac025/6610966 by guest on 18 June 2022

Page 16 of 20
Wang etal.
We also examined the extent the public’s
mental health is associated with the demo-
graphic and socioeconomic prole of neigh-
bourhoods, their health risks and healthcare
access in the three largest capital cities, including
Sydney, Melbourne and Brisbane (Table S2–4).
The results show that the social and spatial
disparity in demographic and socioeconomic
prole across areas in capital cities are poten-
tially associated with the public’s mental health.
More specically, low-income (with odds ratios
larger than 1)and crowded living space (with
odds ratios smaller than 1)are observed to be
associated with mental health signals in Sydney
and Melbourne; while difculty in moving and
transport (with odds ratios larger than 1) is
associated with mental health signals only in
Sydney and mortgage stress (with odds ratios
larger than 1)is associated with mental health
signals only in Brisbane. However, health risks
and healthcare access variables are observed
to be not signicantly associated with mental
health signals in these three capital cities.
Discussion and conclusion
Drawing on social media data and machine
learning techniques, this study makes an initial
attempt to reveal the urban-regional disparity
of the public’s mental health status during a
global scale pandemic and examine the extent
to which mental health signals are associated
with the demographic and socioeconomic pro-
les of neighbourhoods, their health risks and
healthcare access. This has been under-explored
in current scholarship in the Australian context.
Urban and regional/rural areas are featured by
different demographic and socioeconomic char-
acteristics and congured by various healthcare
facilities and services, which can affect people’s
sentimental and emotional reactions to the
pandemic. It is especially the case for capital
cities where people with negative mental health
signals are observed to live in areas with limited
healthcare resources and more difculty in
healthcare access. Our nding align with the
observations in previous studies focusing on
the mental health of the general population in
different geographic contexts. More specically,
the minority-concentrated neighbourhoods
(e.g., with a higher proportion of aboriginal
population) have a higher likelihood of nega-
tive mental health signals; this nding echoes
the ndings in the US that Hispanic/Latinx
groups are subject to greater risks for worsened
mental health due to the pandemic (Penner
et al., 2021). Socioeconomic disparity across
neighbourhoods with different levels of income,
mortgage stress and living space are subject to
the discrepancy of mental health signals; such
an observation is also found in case studies in
China (Liu et al., 2021), US (McKnight-Eily
etal., 2021), Caribbean regions (Llibre-Guerra
etal., 2020), the United Kingdom (Pierce etal.,
2020) and European countries (Reznik etal.,
2020). Moreover, neighbourhoods with the
concentrations of the population at higher
health risks are subject to a higher likelihood
of negative mental signals, reecting that the
onset of health risks (e.g., poor health status,
high blood pressure and drinking habits) may
worsen mental health in the face of the pan-
demic. This nding is less observed in COVID-
19 related studies; however, more prevalent
in mental health studies before the pandemic
(e.g., Hardy et al., 2013; Kwan et al., 2016).
Furthermore, another disparity between capital
cities and regional areas observed in our study
is the availability of healthcare facilities and
services and the difculty in healthcare access
which is further associated with the prevalence
of negative mental health signals. Easy access
to mental health facilities could help persons
with mental illness to integrate into the patient
community and facilitate peer-led interventions
to improve medical self-management (Druss
et al., 2010). It has been observed that older
men living in rural Australia may have a similar
incidence of mental health problems compared
to older women, despite a lower rate of diag-
nosis (Fitzpatrick etal., 2021). The obstacles to
men obtaining mental health treatment in rural
Downloaded from https://academic.oup.com/cjres/advance-article/doi/10.1093/cjres/rsac025/6610966 by guest on 18 June 2022

Page 17 of 20
Urban-regional disparities in mental health signals in Australia
Australia include the reluctancy of emotional
expression, non-disclosure of distress and dif-
culties in seeking and getting help (Kennedy
etal., 2021). Thus, neighbourhoods congured
by better healthcare facilities serve as a con-
textual pathway to protect against the onset of
psychological problems, improving the percep-
tion of local residents in feeling safe, secure and
protected and thus reducing mental disorders.
Our ndings provide evidence for policy
making and implementation in public health.
In response to the urban-regional disparity
of the public’s mental health signals, regional
health authorities and government agencies
should distribute more healthcare services
in these vulnerable areas prior to any unpre-
cedented events; intervention programs and
mental health services should be introduced
or put in place in regional areas and urban
areas in capital cities which are currently less
covered by healthcare facilities. National level
assistance, such as digital mental health guide-
lines, are also recommended in regional areas
with a higher proportion of socioeconomically
disadvantaged populations. Mental health
strategies should be varied along pandemic
event timelines, with monitoring and more
effective decision-making processes to allow
prompt responses to the rapidly changing pan-
demic. Our models and research framework
may be useful for governments and policy-
makers at various levels to monitor public’s
mental health signals in future pandemics and
public healthcrises.
This study has several limitations that need
to be noted and can be improved in future
studies. First, the geotweet data used in our
study is non-random sample, since the number
of Twitter users is dependant on population
density and thus tend to concentrate in densely
populated areas (e.g., capital cities) whilst are
less prevalently in regional/remote areas, which
may bring bias to the analytical results (Zhou
etal., 2021). Second, Twitter users are typically
younger, avid users of social media apps and
the Internet (Huang et al., 2020), while older
individuals are less likely to use smartphones
and Aboriginal Australians have lower internet
accessibility and are relatively inactive in social
media (Povey et al., 2016). Thus, the current
geotweets may be skewed to represent the opin-
ions and perceptions of the subsection of the
population. In addition, the current geotweets
are in English, with non-English ones excluded
from sampling. Future studies using multilin-
gual tweets might better represent different
ethnicities and spoken languages. Third, our
analysis did not involve the COVID-19 meas-
ures (e.g., conrmed cases, and mortality and
morbidity rates) given such data is not avail-
able at the SA2 level. The spread and severity
of COVID-19 infection have been reported
to associated with the public’s mental status
(Kenerly et al., 2022), thus we encourage fu-
ture studies to take on board these COVID-19
measures to enhance the analytical robustness.
Fourth, the built environment in the capital
cities and regional areas can be very different
(e.g., the density, diversity, and design of neigh-
bourhoods); the measures of the built envir-
onment need to be included in the regression
modelling to better reveal the context pathway
through which urban/rural space affects the
public’s mental health during the pandemic.
In addition, it has been reported that the type
and stringency of policy implementation over
different phases of the pandemic also affects
people’s overall sentiment (Zhou et al., 2021).
Thus, the measures of policies can be included
in future modellingwork.
While our study provides empirical evidence
of using machine learning techniques in mental
health studies, there remain difculties and
challenges in translating the analytical nd-
ings to the real-world treatment of at-risk co-
horts. The empirical framework developed in
our study can be improved by integrating with
highly scalable and accurate machine learning
platforms (e.g., symbiotic marching learning
systems) and trials of local system dynamics
models which have been used in monitoring
the suicidal behaviours of Australian youth
Downloaded from https://academic.oup.com/cjres/advance-article/doi/10.1093/cjres/rsac025/6610966 by guest on 18 June 2022

Page 18 of 20
Wang etal.
(Iorno etal., 2021). Future work should also
give attentions to the usability challenges for
machine learning (Zytek et al., 2021). It re-
quires the researchers to have sufcient skills
and contextual knowledge in mental health and
medicine to develop human-centred research
procedure, reliable interpretation of model-
ling results, and cross validation of ndings
based on different approaches (e.g., traditional
survey-based assessment, and clinic protocol
and trials) (Thieme etal., 2020).
To conclude, our study contributes a novel
empirical framework using social media and
machine learning techniques to systematic-
ally classify and measure the urban-regional
disparity of mental health signals of a nation.
Our approach is designed in a manner that
can readily be augmented into an ongoing
monitoring capacity and extended to other
nations. Our ndings in the interrelationship
among mental health, the demographic and
socioeconomic proles of neighbourhoods,
their health risk and healthcare access provide
important evidence for the smart deployment
of nite mental health services and place-based
health planning towards the post-pandemic
period andbeyond.
Acknowledgments
This study is supported by the NCRIS-enabled
Australian Urban Research Infrastructure Network
(AURIN) with the project name—Integrated Heat
Vulnerability Assessment Toolkit for Australian
Cities, AURIN, High Impact Projects 2021.
Conict of interest statement
None.
References
Australia Bureau of Statistics. (20 16) 1270.0.55.00 1—
Australian Statistical Geography Standard
(ASGS): volume 1—main structure and greater
capital city statistical areas. Available at: https://
www.abs.gov.au/ausstats/abs@.nsf/mf/1270.
0.55.001 [accessed 18 Aug 2021].
Australia Bureau of Statistics. (2020) National, state
and territory population. Available at: https://
www.abs.gov.au/statistics/people/population/
national-state-and-territory-population/latest-
release [Accessed 18 Aug 2021].
Australian Institute of Health and Welfare. (2019)
Rural and remote health. Available at: https://
www.aihw.gov.au/reports/australias-health/rural-
and-remote-health [Accessed 18 Aug 2021].
Australian Urban Research Infrastructure Network.
(2020) AURIN data. Available at: https://aurin.
org.au/resources/data/ [Accessed 18 Aug 2021].
BalcombeL., and DeLeo D. (2020) An integrated
blueprint for digital mental health services amidst
COVID-19. JMIR Mental Health, 7: e21718.
BetschC. (2020) How behavioural science data helps
mitigate the COVID-19 crisis. Nature Human
Behaviour, 4: 438–438.
Bird, S., Loper, E., and Klein, E. (2009) Natural
Language Processing with Python. Available at:
https://www.nltk.org/ [Accessed 18 Aug 2021].
Bone D., Lee C. C., Chaspari T., Gibson J., and
NarayananS. (2017) Signal processing and machine
learning for mental health research and clinical ap-
plications [perspectives]. IEEE Signal Processing
Magazine, 34: 196–195.
Chan K. L., Lee C. S., and Cheng C. M., et al
(2019) Investigating the relationship between
weight-related self-stigma and mental health
for overweight/obese children in Hong Kong.
The Journal of Nervous and Mental Disease, 207:
637–641.
ConwayM., and O’ConnorD. (2016) Social media,
big data, and mental health: current advances
and ethical implications. Current Opinion in
Psychology, 9: 77–82.
Coppersmith,G., Harman,C., and Dredze,M. (2014)
Measuring post traumatic stress disorder in Twitter.
In Eighth International AAAI Conference on
Weblogs and Social Media (pp. 579–582).
Cotfas L. A., Delcea C., Roxin I., Ioanăş C.,
Gherai D. S., and Tajariol F. (2021) The longest
month: analyzing COVID-19 vaccination opin-
ions dynamics from tweets in the month following
the rst vaccine announcement. IEEE Access, 9:
33203–33223.
Diez-RouxA. V. (1998) Bringing context back into
epidemiology: variables and fallacies in multilevel
analysis. American Journal of Public Health, 88:
216–22.
DrussB.G., ZhaoL., and SilkeA. etal (2010) The
Health and Recovery Peer (HARP) Program:
a peer-led intervention to improve medical
self-management for persons with serious mental
illness. Schizophrenia Research, 118: 264–270.
Downloaded from https://academic.oup.com/cjres/advance-article/doi/10.1093/cjres/rsac025/6610966 by guest on 18 June 2022

Page 19 of 20
Urban-regional disparities in mental health signals in Australia
EwingL.A., and VuH.Q. (2021) Navigating ‘Home
Schooling’during COVID-19: Australian public re-
sponse on Twitter. Media International Australia,
178: 77–86.
FisherJ.R., Tran T. D., and Hammarberg K. et al
(2020) Mental health of people in Australia in the
rst month of COVID‐19 restrictions: a national
survey. Medical Journal of Australia, 213: 458–464.
FitzpatrickS.J., ReadD., BrewB.K., and PerkinsD.
(2021) A sociological autopsy lens on older adult
suicide in rural Australia: addressing health, psy-
chosocial factors and care practices at the intersec-
tion of policies and institutions. Social Science and
Medicine, 284: 114196.
Ghorbanzadeh M., Kim K., Ozguven E. E., and
Horner M. W. (2020) A comparative analysis
of transportation-based accessibility to mental
health services. Transportation Research Part D
Transport and Environment, 81: 102278.
Hardy S. A., Francis S. W., Zamboanga B. L.,
KimS.Y., AndersonS.G., and ForthunL.F. (2013)
The roles of identity formation and moral identity
in college student mental health, health‐risk be-
haviors, and psychological well‐being. Journal of
Clinical Psychology, 69: 364–382.
Hosmer D. W., Hosmer T., Le Cessie S., and
Lemeshow S. (1997) A comparison of good-
ness‐of‐t tests for the logistic regression model.
Statistics in Medicine, 16: 965–80.
HuT., WangS., and LuoW., etal (2021) Revealing
public opinion towards COVID-19 vaccines with
twitter data in the United States: spatiotemporal
perspective. Journal of Medical Internet Research,
23: e30854.
Huang X., Li Z., Jiang Y., Li X., and Porter D.
(2020) Twitter reveals human mobility dynamics
during the COVID-19 pandemic. PLoS One, 15:
e0241957.
Hutto C., and Gilbert E. (20 14) Vader: A parsi-
monious rule-based model for sentiment ana-
lysis of social media text. In Proceedings of the
International AAAI Conference on Web and
Social Media (Vol. 8, No. 1).
Iorno F., Occhipinti J. A., and Skinner A. et al
(2021) The impact of technology-enabled care co-
ordination in a complex mental health system: a
local system dynamics model. Journal of Medical
Internet Research, 23: e25331.
Kenerly M. J., Shah P., and Patel H. et al (2022)
Altered mental status is an independent predictor
of mortality in hospitalized COVID-19 patients.
Irish Journal of Medical Science, 191: 21–26.
Kennedy A., CosgraveC., MacdonaldJ., Gunn K.,
Dietrich T., and Brumby S. (2021) Translating
co-design from face-to-face to online: an
Australian primary producer project conducted
during COVID-19. International Journal of
Environmental Research and Public Health, 18:
4147.
KretchyI.A., Owusu-DaakuF.T., and DanquahS.A.
(2014) Mental health in hypertension: assessing
symptoms of anxiety, depression and stress on anti-
hypertensive medication adherence. International
Journal of Mental Health Systems, 8: 1–6.
KwanM. Y., Arbour-Nicitopoulos K. P., Duku E.,
and FaulknerG. (2016) Patterns of multiple health
risk–behaviours in university students and their
association with mental health: application of la-
tent class analysis. Health Promotion and Chronic
Disease Prevention in Canada rRsearch, Policy
and Practice, 36: 163–170.
Lake J., and Turner M. S. (2017) Urgent need for
improved mental health care and a more collab-
orative model of care. The Permanente Journal, 21:
17–024.
LiangY., ZhengX., and ZengD. D. (2019) A survey
on big data-driven digital phenotyping of mental
health. Information Fusion, 52: 290–307.
Liu L., Xue P., Li S. X., Zhang J., Zhou J., and
ZhangW. (2021) Urban-rural disparities in mental
health problems related to COVID-19 in China.
General Hospital Psychiatry, 69: 119–120.
Llibre-GuerraJ.J., Jiménez-VelázquezI.Z., Llibre-
RodriguezJ.J., and AcostaD. (2020) The impact
of COVID-19 on mental health in the Hispanic
Caribbean region. International Psychogeriatrics,
32: 1143–1146.
McKnight-EilyL.R., OkoroC.A., and StrineT.W.
et al (2021) Racial and ethnic disparities in the
prevalence of stress and worry, mental health con-
ditions, and increased substance use among adults
during the COVID-19 pandemic—United States,
April and May 2020. Morbidity and Mortality
Weekly Report, 70: 162–66.
Meyer O. L., Castro-Schilo L., and Aguilar-
GaxiolaS. (2014) Determinants of mental health
and self-rated health: a model of socioeconomic
status, neighbourhood safety, and physical activity.
American Journal of Public Health, 104: 1734–1741.
NewbyJ.M., O’MooreK., Tang S., ChristensenH.,
and Faasse K. (2020) Acute mental health re-
sponses during the COVID-19 pandemic in
Australia. PLoS One, 15: e0236562.
PanY., CaiW., ChengQ., DongW., AnT., and YanJ.
(2015) Association between anxiety and hyper-
tension: a systematic review and meta-analysis of
epidemiological studies. Neuropsychiatric Disease
and Treatment, 11: 1121–30.
PennerF., Ortiz J.H., and SharpC. (2021) Change
in youth mental health during the COVID-19
Downloaded from https://academic.oup.com/cjres/advance-article/doi/10.1093/cjres/rsac025/6610966 by guest on 18 June 2022

Page 20 of 20
Wang etal.
pandemic in a majority Hispanic/Latinx US
sample. Journal of the American Academy of Child
and Adolescent Psychiatry, 60: 513–523.
PierceM., HopeH., and FordT. etal (2020) Mental
health before and during the COVID-19 pandemic:
a longitudinal probability sample survey of the UK
population. The Lancet Psychiatry, 7: 883–892.
PoveyJ., MillsP. P. J.R., and DingwallK. M. etal
(2016) Acceptability of mental health apps for
Aboriginal and Torres Strait Islander Australians:
a qualitative study. Journal of Medical Internet
Research, 18: e5314.
Public Health Information Development Unit.
(2016) Population health areas: overview.
Available at: https://phidu.torrens.edu.au/about-
phidu [Accessed 18 Aug 2021].
Reznik A., Gritsenko V., Konstantinov V.,
KhamenkaN., and IsralowitzR. (2020) COVID-
19 fear in Eastern Europe: validation of the fear of
COVID-19 scale. International Journal of Mental
Health and Addiction, 19: 1903–1908. doi:10.1007/
s11469-020-00283-3.
RudinC. (2019) Stop explaining black box machine
learning models for high stakes decisions and use
interpretable models instead. Nature Machine
Intelligence, 1: 206–215.
ShahabL., AndrewS., and WestR. (2014) Changes
in prevalence of depression and anxiety fol-
lowing smoking cessation: results from an inter-
national cohort study (ATTEMPT). Psychological
Medicine, 44: 127–41.
SparrenbergerF., CicheleroF., and AscoliA., etal
(2009) Does psychosocial stress cause hyperten-
sion? Asystematic review of observational studies.
Journal of Human Hypertension, 23: 12–19.
Summers-Gabr N. M. (2020) Rural–urban mental
health disparities in the United States during
COVID-19. Psychological Trauma Theory,
Research, Practice, and Policy, 12: S222–24.
Sun J., and Lyu S. (2020) Social participation and
urban-rural disparity in mental health among older
adults in China. Journal of Affective Disorders, 274:
399–404.
Sutarsa N., Baneld M., Passioura J., Konings P.,
and MooreM. (2021) Spatial inequities of mental
health nurses in rural and remote Australia.
International Journal of Mental Health Nursing,
30: 167–176.
Talen M. R., and Mann M. M. (2009) Obesity and
mental health. Primary Care: Clinics in Ofce
Practice, 36: 287–305.
Thieme A., Belgrave D., and Doherty G. (2020)
Machine learning in mental health: a systematic
review of the HCI literature to support the devel-
opment of effective and implementable ML sys-
tems. ACM Transactions on Computer–Human
Interaction (TOCHI), 27: 1–53.
ThornicroftG., RoseD., KassamA., and SartoriusN.
(2007) Stigma: ignorance, prejudice or discrim-
ination?. The British Journal of Psychiatry, 190:
192–193.
Twitter. Twitter API—Academic Research product
track. (2020) Available at:https://developer.twitter.
com/en/products/twitter-api/academic-research
[Accessed 18 Aug 2021].
VADER. (2021) VADER library. Available at:
https://github.com/cjhutto/vaderSentiment
[Accessed 18 Aug 2021].
VanRheenenT.E., MeyerD., and NeillE., etal
(2020) Mental health status of individuals with a
mood-disorder during the COVID-19 pandemic
in Australia: initial results from the COLLATE
project. Journal of Affective Disorders, 275:
69–77.
ZhangM., WangS., and HuT. et al (2021) Human
mobility and COVID-19 transmission: a system-
atic review and future directions. Annals of GIS,
1–14. doi:10.1080/19475683.2022.2041725.
Zhou J., Yang S., Xiao C., and Chen F. (2021)
Examination of community sentiment dynamics
due to COVID-19 pandemic: a case study from
a state in Australia. SN Computer Science, 2:
1–11.
Zhou J., Zogan H., Yang S., Jameel S., Xu G.,
and Chen F. (2021) Detecting community de-
pression dynamics due to COVID-19 pan-
demic in Australia. IEEE Transactions on
Computational Social Systems, 8: 1–10. Available
online at: https://ieeexplore.ieee.org/stamp/stamp.
jsp?tp=&arnumber=9325873.
Zytek A., Liu D., Vaithianathan R., and
Veeramachaneni K. (2021) Sibyl: understanding
and addressing the usability challenges of machine
learning in high-stakes decision making. IEEE
Transactions on Visualization and Computer
Graphics, 28: 1161–71.
Downloaded from https://academic.oup.com/cjres/advance-article/doi/10.1093/cjres/rsac025/6610966 by guest on 18 June 2022