Content uploaded by Sebastien Bourdin
Author content
All content in this area was uploaded by Sebastien Bourdin on Feb 03, 2021
Content may be subject to copyright.
Article accepted in Regional Studies
1 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
Does lockdown work? A spatial analysis of
the spread and concentration of COVID-19
in Italy
Sebastien Bourdin, Ludovic Jeanne, Fabien Nadou,
Gabriel Noiret
EM Normandie Business School – Métis Lab
Department of regional economics and sustainable development
9, rue Claude Bloch, 14 000 Caen (France)
sbourdin@em-normandie.fr
Summary:
The spread of COVID-19 is a global concern, especially in the most developed countries where
the rapid spread of the virus has taken governments by surprise. Adopting a spatial approach to
this issue, we identify the spatial factors that help explain why some areas are hit harder than
others, based on the Italian example (with the Lombardy region as the epicentre in Europe).
Our analysis combines an autoregressive spatial model and a bivariate spatial autocorrelation
from a pool of data collected from the Italian provinces to propose a real-time analysis of the
spread and concentration of the virus. Our findings suggest that the most globally connected
areas are also the worst hit ones, and that the implementation of a lockdown at the beginning
of March was a crucial and effective approach to slowing the spread of the virus further.
Keywords: COVID-19, spatial analysis, geography, lockdown, proximities
Article accepted in Regional Studies
2 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
1. Introduction
In December 2019, a new virus from the SARS (Severe Acute Respiratory Syndrome) family, called
COVID-19 (WHO, 2020), began to spread across China to reach virtually every other country in the
world. The coronavirus 2019-nCoV can progress to pneumonia (Wang et al., 2020), and its mortality
rate has been estimated at a maximum of 3% (WHO, 2020). As with the transmission of SARS-CoV,
there is evidence that COVID-19 is transmitted person-to-person via airborne respiratory droplets, direct
contact with body fluids or secretions, or via contaminated objects (Xu et al., 2020). Initial findings
from an epidemiological study (Chen et al., 2020) suggest that the 2019-nCoV virus is 80% identical to
SARS, with 20% of unknown origin, and that it may be less pathogenic than MERS-CoV and SARS-
CoV. On the other hand, while the virus strain appears to be less deadly than SARS-CoV, it is more
contagious. Consequently, governments are on global alert due to the rapid spread of the infection.
As Munster et al. (2020) explained, the absence of severe outbreaks of the disease affected our ability
to contain its spread as people do not necessarily know if they are infected or not, and can pass it on
unwittingly. In addition, the potentially long incubation period - up to 14 days for COVID-19 versus
about two days for influenza - means that people can become ill and transmit the disease before any
symptoms have appeared (Wang et al., 2020). As a result, infected people are unaware of their potential
to the spread the virus and continue to work and travel, increasing the risk of passing it on to their
contacts, both at home and abroad.
Given the "alarming levels of spread and inaction" that have occurred, the WHO decided to officially
declare a pandemic on 11 March 2020. From 13 March 2020, Europe in turn became the epicentre of
the outbreak. In Italy, the population went into full lockdown on 9 March 2020. Spain and France were
also heavily affected and subsequently adopted lockdown measures. The problem is all the more acute
as the European population's demographic structure has a relatively high share of elderly people, on
average, and is thus more vulnerable compared to other regions of the world.
In this context, it is important to understand how the virus spreads in order to better contain it as there
is no epidemic without transmission. Mapping pandemics is an increasingly widespread scientific
Article accepted in Regional Studies
3 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
approach. As Koch and Koch (2005) argue, GIS mapping tools have greatly improved our capacity to
study and predict epidemics around the world. By using the example of Covid-19, Ahasan et al. (2020)
explain that GIS not only show the impact of the disease, but can also be a useful decision-making tool.
A geographical reading of epidemics provides a better understanding of how they spread spatially,
helping us understand the underlying social processes of diffusion. Whether the plague or other diseases,
epidemics have frequently followed major military and trade routes (Lai et al., 2008). Several studies
have highlighted the important explanatory role of geographical factors, such as air flows during the
SARS epidemic in 2003 (Bowen and Laroe, 2006), the porosity of regional borders during the Ebola
epidemic of 2014 (Kramer et al., 2016) and the impact of urbanisation in the case of Zika in 2015
(Lourenco, 2017).
Traditionally, epidemics spread according to topographical and isotropic factors. This means that they
are transmitted from one place to another over large areas – the notion of proximity being defined in
kilometres – and this occurs gradually and comprehensively. Two factors should be kept in mind here:
a place close to the epidemic’s starting point is almost always affected before a location that is very far
from it; and all locations equidistant from the start of the epidemic are expected to be affected at the
same time. Of course, the spread of an outbreak can sometimes give rise to a "tunnel effect" when the
pathogen is transported over long distances by a long-range transport network. In this way, the pathogen
may skip an entire area while continuing to spread on a much wider scale.
What is important here, however, is topographical diffusion, with the networks ultimately playing only
a triggering role that allows the pathogen to cover a much larger area. According to Cliff et al. (2004),
authors of the World Atlas of Epidemic Diseases, a reference book published in 2004, epidemic diseases
spread in three ways: hierarchical mode (between large cities), local mode (through neighbourhoods)
and "jump" mode (along transport routes). This was especially true with HIV/AIDS and SARS, whose
spread was largely topological, and where the notion of proximity is defined in terms of connectivity
between different points in a network. This changes the rules of diffusion completely, in other words, a
place that is strongly connected to the epidemic’s starting point will generally be affected before a place
that is poorly connected, regardless of the distance in kilometres. Places affected simultaneously from
Article accepted in Regional Studies
4 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
the same starting point of the epidemic have a similar degree of connection (they can thus be located
thousands of kilometres apart). Consequently, the topographical logic still plays a role, but in parallel
with the network (topological) logic. It is thus sometimes possible to distinguish two scales of diffusion
for the same epidemic: one local and territorial, the other global and reticular. For example, the 1889-
90 influenza in Switzerland spread first from the municipalities served by the railway (Hogbin, 1985)
and then via neighbourhoods to surrounding municipalities (Le Goff, 2011). With increased mobility
and the general development of transport, it is crucial to take flows involved in the rapid spread of
pathogens into account. The explosion of air flows tends to support a jumping mode of diffusion. For
this reason, airports have become strategic surveillance and control points when trying to contain the
spread of epidemics. In a recent metapopulation study, Chinazzi et al. (2020) modelled the transmission
of COVID-19 to project the impact of travel restrictions on the national and international spread of the
epidemic. They showed only a modest effect of global travel restrictions unless the measures were
combined with behavioural changes advocated by health services. In similar vein, Adiga et al. (2020)
quantified the risk of population exposure based on decisions to suspend flights, proposing a
measurement system to classify vulnerable countries at immediate risk of case emergence.
Hence, one of the key lessons to be learned from the geographical study of epidemics is that we need an
effective global health network to deal with the rapid spread of disease. This depends in part on our
ability to identify potential global health threats and to respond to them at local level, which in turn
involves understanding the mechanisms of how a pathogen spreads at global and regional levels. The
development of mapping methods has allowed us to create a more dynamic and accurate representation
of the transmission of COVID-19. Consequently, new questions can now be raised: which areas are
most affected by the virus? How can regional variations be explained? And last but not least, has
lockdown had the desired effect? Numerous studies have explored this topic with regard to other
epidemics. Several researchers have thus applied mapping and geostatistical methods to analyse disease
spread patterns during epidemics such as tuberculosis (Roth et al., 2016), cholera (Ali et al., 2002),
SARS-CoV (Wang et al., 2008; Meade, 2014), MERS-CoV (Cotten et al., 2014), H1N1 influenza
(Smallman-Raynor and Cliff, 2008; Souris et al., 2010), HIV (Wood et al., 2000), dengue (Atique et al.,
Article accepted in Regional Studies
5 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
2018; Zhu et al., 2019) and, more recently, COVID-19 (Desjardin et al., 2020; Paez et al., 2020; Orea
and Álvarez, 2020). Conducted on different scales and for different diseases, these studies highlight the
effects of spatial dependence between regions, partially explaining the spatial heterogeneity of the
spread of epidemics. Spatial dependence effects refer directly to the issue of spatial autocorrelation (Le
Gallo, 2014), i.e., the coincidence of similarity of values with similarity of location (Anselin, 1995;
Anselin, 1999). In their article on autonomous provinces (NUTS2), Paez et al. used a spatial model to
measure the effectiveness of lockdowns. Applying a spatial SUR model, they pointed to a decrease in
the contagion effect. We attempt to go further by (i) using global and local Moran statistics (LISA and
BiLISA) to measure the evolution of the spatial concentration of incidence and its persistence across
time, and (ii) analyse the incidence of COVID-19 on a finer scale (NUTS3) to better capture the spatial
effects.
While, in medicine, the most common hypothesis involves a process of diffusion across an area by way
of gradual transmission, we hypothesise that the spread of the virus and its spatial concentration (at the
beginning of the epidemic) is conditioned by the pre-existing spatial organisation (in socio-economic
and regional terms). It is within this framework that we attempt to analyse the dynamics of virus spread
by (i) highlighting the spatial concentration of COVID-19 cases and (ii) examining the links between
the spread of the disease and the variables likely to influence it, at the same time taking into
consideration the effects of spatial dependence (neighbourhood effects and spatial autocorrelation). To
this end, we adopted an exploratory method in spatial statistics not previously used to examine such
issues. Spatial relationships between geographical entities have been a key focus for geographers for
several decades, including in medical geography (Dobis, 2020). One research avenue in this area is the
analysis of spatial autocorrelation, which measures the association between objects of the same type
(Cliff and Ord, 1972). Existing measures include global (Moran, 1950; Geary, 1954) and local (Anselin,
1995) statistics. Another line of research focuses on spatial correlation or co-location, which measures
the spatial relationship between objects of different types (Leslie and Kronenfeld, 2011). We used the
method of local indicators of bivariate spatial association (Anselin et al., 2000; Anselin, 2010) to
parametrically identify the spatial concentrations of covariations between recorded cases of COVID-19
Article accepted in Regional Studies
6 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
and diverse explanatory variables (population density, level of economic development, air traffic, health
services). Moreover, in addition to a spatial exploratory approach, we developed a model to explain
cases of COVID-19. In this respect, OLS estimators are often biased and inefficient when considering
the existence of a spatial autocorrelation event, calling into question the statistical inference approach.
In order to correct this bias, a spatial lag model can be used to model the incidence of COVID-19 as
suggested by Paez et al. (2020), since it results from a dynamic contagion process. The model provides
us with estimators of the parameter ρ, a parameter that characterizes transmission effects. This is an
important factor and several researchers (Spielman and Yoo, 2008; Chen and Wen, 2010; Cromley et
al., 2012) have now called for the spatial dimension of neighbourhood effects to be taken into
consideration.
Our study attempts to analyse the mechanisms of the spread of the virus in the national context of Italy
that became the European epicentre of the outbreak during the first wave. We thus address the issue of
the virus diffusion model, examining in particular the spatial autocorrelation of the incidence of COVID-
19. We then analyse the links between regional factors and case density. Finally, the study reflects on
the strategies required to contain the epidemic by highlighting the central role of geography.
2. Materials and method
2.1. Data
Our study focuses on the data presented below in Table 1. Lee and Wong (2010) note that surveillance
data collected on a daily basis can be productively used to describe an epidemic and its geography,
supporting the introduction of more effective interventions at local level to contain the spread. In 2016,
the WHO published a study that identified all reported epidemics in Africa between 1970 and 2016. The
report showed the difficulty of collecting homogeneous fine-scale data and of making temporal
comparisons between the data at national level. For our example of Italy, we collected data from the
Ministry of Health in Italy at the level of provinces. Our dependant variable is the density of cases (i.e.,
the incidence of COVID-19) measured by the number of COVID-19 cases per 100,000 inhabitants.
Article accepted in Regional Studies
7 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
Table 1: Descriptive statistics
The WHO demonstrated (2004) that severe acute respiratory syndrome (SARS) spreads rapidly along
international air transport routes (the largest outbreaks were observed in air transport hubs or densely
populated areas), confirming other studies on the virus that underscored these factors (Wang et al., 2008;
Meade, 2014). Mayer (2000) also identified migration and mobility as a determining variable in a paper
that traced the key factors of epidemic emergence and spread. We therefore included the density of
annual airport passengers per km² (DensPassAero) as a variable to measure this factor in our study.
Several studies have shown the link between the spread of epidemics and the socio-economic
vulnerability of populations (Barnes, 2014; Stanturf et al., 2015; Bonifay et al., 2018). However, these
studies do not focus on the spatial dimensions of the epidemic’s transmission, nor do they account for
what is at stake with respect to the scale of the analysis. Our hypothesis seeks to address these limitations
by assuming that the most economically developed areas are more likely to concentrate the largest
number of cases since they are more closely connected to the rest of the world, especially other highly
affluent regions. This in turn can lead to greater exposure of the most vulnerable socio-economic groups
at local level. However, it can also keep areas with populations that are extremely socio-economically
vulnerable more distant from the virus transmission process or, at the very least, reduce their exposure
and postpone the moment they might be affected. We therefore added the variable GDP, which
represents GDP/capita in PPP.
Variable Min Max Mean Std. dev. Date Source
Incid08 0 370.55 13.93 44.39 8 March. 2020 Ministry of Health, Italy
Incid18 1.18 627.72 60.23 106.01 18 March. 2020 Ministry of Health, Italy
Incid28 5.52 1004.30 153.13 181.72 28 March. 2020 Ministry of Health, Italy
GDP 15200 55100 26259 7551.18 2018 Eurostat
DensAero 0 25285.51 900.90 3026.31 2019
Ministry of Infrastructures
and Transport, Italy
Density 29 823.7 193.83 150.63 2018 Eurostat
Share65+ 17.512 29.12 23.80 2.39 2018 Eurostat
Article accepted in Regional Studies
8 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
Incorporating various examples, Reyes et al. (2013) explained how the urban environment can be an
aggravating factor in the fight against infectious diseases. In their study on swine flu (H1N1) in Hong
Kong, Lee and Wong (2010) showed that densely populated areas are more affected overall. Given the
effects of densely built-up areas, we argue that since the spatial proximity of individuals is greater in
densely populated areas, this exacerbates more rapid circulation of the virus. We therefore added a
variable relative to population density (Density), allowing us to identify the potential effects of
thresholds, with the relation between the speed of virus transmission on the one hand and population
density on the other, which are not necessarily in direct proportion. Some threshold effects may result
from significant variations between specific areas and societies with regard to housing characteristics,
the composition of family groups living under the same roof, private and public sociability norms,
hygiene practices and proxemic structures (Hall, 1966 & 1968). However, it is not easy to identify such
threshold effects within a country as behavioural norms may not be sufficiently marked.
Finally, several recent studies have shown that older people have a higher likelihood of being affected
by the virus (Xu et al., 2020; Chen et al., 2020). Our model therefore includes the percentage of the
population aged over 65 (Share65+).
2.2 Methodology
The methodology adopted is designed to define a geography of the spatial spread and concentration of
the virus, in similar vein to several previous studies on virus diffusion (Vaneckova et al., 2010;
Robertson and Nelson, 2014; Weimann et al., 2016). It is based on the hypothesis from the literature of
spatial effects in linkages between the spread of COVID-19 and its explanatory factors. We propose a
geography of the outbreak by highlighting the spatial concentration and transmission effects (spatial
autocorrelation and co-location effects) associated with a model that includes spatial interaction effects
to identify the factors that help to explain the observed event.
2.2.1. Spatial autocorrelation and co-location effects
A co-location analysis technique was first developed by Anselin et al. (2002) who designed a bivariate
local spatial association indicator (BiLISA) to study the spatial correlation patterns between two
Article accepted in Regional Studies
9 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
georeferenced variables (i.e., number of alcohol outlets relative to the number of criminal incidents in
the areas in question). BiLISA is used to assess the link between a variable in one area and a second
variable in neighbouring areas. Bivariate analysis allows the relationship between the dependent variable
and the explanatory variables to be isolated one at a time. It is also possible to study spatiotemporal
autocorrelation, in other words, the correlation of a variable with reference to the spatial location of a
single variable within a given time interval. The correlation of a variable with itself in space and time is
then analysed (Anselin et al., 2002). The idea here is to conduct a parametric examination (beyond visual
inspection) of the way spatial patterns are correlated between several variables in order to draw
conclusions about the spatial dynamics that link a set of localized data. In more technical terms,
examining the similarity of spatial processes between variables can be seen as a way to test the
robustness or persistence of a given spatial pattern/scheme over time, such as by comparing patterns of
local spatial association between COVID-19 and population density (robustness) or between density of
COVID-19 cases at two reasonably spaced periods in time (persistence), for example. Thus, the method
applied allows the spatial covariations of the explanatory factors behind the spatial concentration
phenomena of the virus to be captured. We can then identify whether, for each spatial area, the virus is
significantly spatially correlated with another variable.
In order to identify and assess the magnitude of spatial relationships, we used Moran's I to measure
spatial autocorrelation and identify spatial clusters in the data. Four types of spatial associations can be
derived from this statistic for our study: i.e., high-high (HH - spatial concentration of high values of
incidence and high values of the independent variable from neighbouring regions) and low-low (LL -
spatial concentration of low values of incidence and low values of the independent variable from
neighbouring regions) types for spatial clustering of similar values, and high-low (HL - spatial
concentration of high values of incidence and low values of the independent variable from neighbouring
regions) and low-high (LH - spatial concentration of low values of incidence and high values of the
independent variable from neighbouring regions) types for spatial clustering of dissimilar values.
The measure of local bivariate spatial autocorrelation using Moran's I statistic, derived from Anselin's
(1995) formula, is:
Article accepted in Regional Studies
10 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
where represents the incidence of COVID-19 and the independent variable for the province and
the neighbouring province , while are the standardized z-scores of both the variables and
and is the spatial weight matrix. The choice to model spatial proximity and interdependence between
provinces is explained in Part 2.2.2.
A permutation approach (here 999) was used to assess the statistical significance of Moran's I and
BiLISA results. Randomisation assumes that the location of the values and their spatial arrangement are
unimportant. Based on randomisation, various theoretical standard deviations for Moran's I were
obtained, each giving a different p-value as a pseudo-significance. The threshold value of p=0.05 defined
its significance. The p-values follow an asymptotically standard-normal distribution, allowing us to
evaluate their level of significance by comparing them to a reference distribution (Anselin 1995), and
thus defining the thresholds beyond which the HH, LL, LH, HL relationships were no longer significant.
2.2.2. Empirical model and specification strategy
Our model used an equation for each time period (8th, 18th, 28th March), estimated for a cross section of
spatial units. Use of an appropriate econometric model is recommended when dealing with the spatial
autocorrelation of data to avoid any bias in the estimates. To this end, in the case of spatial interactions
(ρ≠0 or λ≠0), Elhorst (2010) proposed a "mixed" approach (between the classic specific-to-general
strategy or any of the general-to-specific ones), which consists of starting with the bottom-up approach
instead of directly choosing a SAR (spatial autoregressive) or SEM (spatial autocorrelation of errors)
model, and then studying the Durbin spatial model. This makes it possible to confirm the relevance of
the chosen model by means of several tests (Lagrange multiplier, likelihood ratio). It also allows
exogenous interactions to be added to the analysis.
In practical terms, Anselin and Florax (2012) propose using the Lagrange multiplier (LM) (lag and error)
and their robust versions to opt between two spatial econometric models (SEM model versus SAR
Article accepted in Regional Studies
11 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
model). Whatever the spatial weight matrix, LM tests reject the null hypothesis, confirming the existence
of spatial autocorrelation in the model’s error term (Table 2).
The decision-making rule suggested by the same authors is used to determine the most appropriate
specification. If LMLAG is more significant than LMERR, and RLMLAG is significant while
RLMERR is not, then the appropriate model is the auto-regressive spatial model. The test results are
presented in Table 2. In all specifications, the standard and robust tests on spatial lag dependence are
more significant and, consequently, the SAR model is more appropriate than the SEM model.
The SAR model is written as follows:
If is the incidence of COVID-19 in province i at date t, represents the matrix of exogenous
variables, with the weighting matrix noted , where ρ is the parameter of the spatially lagged
dependent variable that captures the spatial interaction effect, indicating the degree to which the
incidence of COVID-19 in one province is determined by the incidence of COVID-19 in its
neighbouring provinces.
Table 2: Results for LM tests
To determine the appropriate specification, some additional tests can be conducted (LeSage and Pace,
2009; Elhorst, 2010). More specifically, if LM tests are significant, it is preferable to estimate a Spatial
Durbin Model (SDM) since the SEM and SAR models are specific forms of the latter. As the SDM
Spatial weight matrix
Spatial lag Spatial error Spatial lag Spatial error
1st order contiguity 76.92*** 34.56*** 56.11*** 1.74
2nd order continguity 121.76*** 65.53*** 72.51*** 2.25
3rd order contiguity 71.19*** 45.12*** 38.96*** 0.36
5 nearest neighbours 120.73*** 57.36*** 76.98*** 0.24
10 nearest neighbours 118.45*** 41.87*** 89.48*** 0.01
15 nearest neighbours 156.08*** 67.18*** 101.25*** 0.31
Robust LM tests
LM tests
Article accepted in Regional Studies
12 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
contains the SAR, we use the likelihood ratio (LR) derived from the Maximum Likelihood (ML)
estimation to identify which estimation strategy should be adopted. Indeed, the SAR model corresponds
to the specific case where spatial autocorrelation is only due to the influence of neighbouring provinces
(when 0). If the hypothesis holds, the SDM could be simplified to the SAR model (LeSage and
Pace, 2009). As shown in Table 3, the 0 hypothesis cannot be rejected, indicating that the SDM can
be simplified as a SAR model. Consequently, since this test also points to a SAR model, we can use the
latter to describe the data in the best way possible. Following Millo and Piras (2012) and Elhorst (2014),
we used a standard method in the spatial model literature, the ML procedure for fixed effects, to capture
unobservable heterogeneity (Baltagi, 2008).
Table 3: Tests for choosing the spatial weight matrix and model comparisons
Before running our model, we need to select the best spatial weight matrix for our analysis. To this end,
we align with other researchers (LeSage and Pace, 2009; Stakhovych and Bijmolt, 2009; Elhorst, 2010;
Vega and Elhorst, 2013) who propose goodness-of-fit measures (as in the case of the log-likelihood
function value (LIK)) to discriminate between different spatial matrix specifications when there are no
clear theoretical reasons for any specific form. We estimate the SAR model by LM using different matrix
specifications. According to the results (Table 3), we select the “first order contiguity matrix” as it
exhibits the highest LIK. Note that the “5 nearest neighbours” could also have been a good alternative
since the LIK is quite similar.
Finally, we follow LeSage and Pace (2009) who distinguish between direct, indirect and total effects.
The mean direct effect corresponds to the impact of a change in an explanatory variable on the incidence
of COVID-19 in province i. The average indirect effect corresponds to the impact of a change in an
explanatory variable in all the provinces other than province i on province i. Symmetrically, it
Spatial weight matrix LIK LR test: θ=0
1st order contiguity 59,457 49,542
2nd order continguity 37,048 28,494
3rd order contiguity 25,20 21,231
5 nearest neighbours 55,134 41,769
10 nearest neighbours 48,318 39,91
15 nearest neighbours 32,854 27,607
Article accepted in Regional Studies
13 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
corresponds to the impact of a change in an explanatory variable in all provinces on province i. This is
the spillover or spatial diffusion effect. The total average effect is, of course, the sum of the average
direct and indirect effects. Interpretation of the results of a SAR model with spatial effects is not
immediate due to the presence of a lag in the dependent variable, as these results do not correspond to
marginal effects. Thus, we estimate both direct and indirect effects. For the estimation, we applied a log
transformation of our variables since they are ratios. In line with Paez et al. (2020), these effects are
interpreted as a percentage change in the incidence of COVID-19 as a consequence of a 1% change in
the variable.
3. Results
Analysis of the Italian case is crucial since it was the first country to be confronted with out-of-control
transmission of the COVID-19 virus, and therefore provides a good basis for assessing lockdown
strategies. In the case of China, suspicion surrounding the accuracy of the public data released prevents
us from fully understanding the kinetics of the epidemic contained through a lockdown strategy. On the
other hand, the numerous social, political, economic and institutional similarities between European
countries compared to the rest of the world mean that it is possible to draw useful and relevant lessons
for other European countries, in particular Spain and France which have also been strongly affected.
Our first goal was therefore to assess the geography of the effects of the spatial concentration of COVID-
19 incidence in Italy. Calculations of Moran’s Index on March 8 (first day of the application of the
general lockdown decision made by the Italian government), March 18 and March 28 revealed a strong
and significant global spatial autocorrelation (I 8th March = 0,421**; I 18th March = 0,385***; I 28th
March = 0,372**). It appears that the clustering of incidence remains high across time.
In order to assess the effects of the spatial dependency of COVID-19 incidence, we applied the spatial
statistics of the Moran Index at local level, first testing the persistency of the spatial concentration of the
number of observed cases over time. To this end, we calculated the BiLISA of the incidence of COVID-
19 between March 8 and March 28 (Moran’s I=0.465) and then mapped the results (Figure 1). The
provinces where the spatial concentration of a small incidence can be observed between the two dates
Article accepted in Regional Studies
14 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
in the Mezzogiorno region are in dark blue, while the spatial concentration of high incidence between
the two dates is in red. The spatial concentration of a high incidence in the province of Rome by 8
March, surrounded by areas with a low incidence by 28 March is in pink, and the spatial concentration
of provinces with a low incidence surrounded by areas with a high incidence by 28 March is in light
blue. We can thus conclude that there was persistency in the spatial concentration of incidence in
Lombardy (high – areas in red) and in the south of Italy (low – areas in dark blue). This persistency can
be interpreted as lockdown effectiveness, since the spatial concentration of high (respective low)
incidence between 8 and 28 March, 2020 remains high and localised in space.
Figure 1: Co-localisation of spatial concentration between 8 March and 28 March
We can observe a strong concentration of the epidemic in Northern Italy, especially in the central
provinces of Lombardy, with a contained spread in the adjacent provinces. On the other hand, the
situation in the south of Italy presents a much weaker spatial concentration of the virus and incidence of
cases. Only the province of Rome, which has strong international connections, experienced an
intermediary situation with a relatively high concentration of incidence and moderate spread through
the provinces in the vicinity of Lazio (High-Low – in Figure 1). Given the current evidence, it is
generally understood that the virus entered Italy through Lombardy and, more specifically, the Milan
Article accepted in Regional Studies
15 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
region, notably Codogno to the south-east of the city. Its acceleration occurred in Bergamo at the time
of a European football match with Valencia (Spain) football team.
A traditional approach would consist of speculating that it would gradually spread to all areas of the
country from the place where the highly contagious virus was first introduced, with gradual spatial
homogenisation in incidence. However, this did not occur. Further analysis of the data shows that
between March 8 and March 28, there was very clear persistence in the relative incidence of COVID-
19 (Figure 1). Since the lockdown in Italy began on 8 March, the co-localisation map of the spatial
concentration of COVID-19 incidence appears to confirm the anticipated effects of quarantine,
especially since eleven towns were locked down from 22 February 2020, six of them in Lombardy.
When a lockdown is introduced early enough, it certainly seems to contain the situation, in other words
it contains the spread in areas with a very high density of cases. This result is all the more remarkable
since SARS-CoV-2 is not only highly contagious but also presents a large number of asymptomatic
cases. The combination of these two factors inevitably hampers detection of the virus’s circulation and
reduces the chances of identifying the best moment for a lockdown. In the absence of other means to
introduce different strategies (systematic and reliable testing, targeted quarantine), lockdown appears to
be an effective solution in the face of an infectious agent for which we have neither a vaccine nor
efficient treatment to counter the acute cases that carry a high risk of death.
Figure 2: Co-localisation of the spatial concentration between the incidence on 8 March (start date of lockdown) and
GDP/cap.
Article accepted in Regional Studies
16 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
Moreover, in addition to examining the efficiency of the lockdown, we also calculated the BiLISA of
incidence by 28 March (Moran’s I= 0.345) with the GDP/cap, and then mapped the results (Figure 2).
We identified a (i) a spatial concentration of high values of incidence and high values of GDP from
neighbouring regions and (ii) a spatial concentration of low values of incidence and low values of GDP
from neighbouring regions. There is also a spatial concentration of low values of incidence and high
values of GDP from neighbouring regions presented in light blue.
The explanatory economic factors for such discrepancies appear to match the reality. Italy’s north/south
economic divide appears to be accentuated by the epidemic’s geography. The highest incidence of
COVID-19 is found in the most economically globalised areas, which explains the localisation of the
place where SARS-CoV-2 first broke out in Italy and the localisation of the first infection cluster. Figure
2 is particularly enlightening as it not only shows the localisation of the place where the virus emerged,
but also an area of propagation in the country’s most economically developed region and a few adjacent
provinces.
Table 4: SAR model of case density in Italian provinces
***, **, and *significance at the 1%, 5% and 10% levels, respectively
Estimation methods OLS SAR 1 SAR 1 SAR 1
(8th March) (18th March) (28th March)
Direct
logGDP 0.364*** logGDP 0.345*** 0.313*** 0.293***
logDensAero 0.019* logDensAero 0.014* 0.007* 0.009*
logDensity 0.058** logDensity 0.042** 0.033** 0.034**
logShare65+ -0.207 logShare65+ -0.182 -0.175 -0.178
Indirect
logGDP 0.158*** 0.093** 0.081***
logDensAero 0.006* 0.005* 0.007*
logDensity 0.014*** 0.011*** 0.009**
logShare65+ -0.334 -0.311 -0.259
Total
logGDP 0.503*** 0.406*** 0.374**
logDensAero 0.02* 0.012* 0.016*
logDensity 0.056** 0.044** 0.043**
logShare65+ -0.516 -0.486 -0.437
ρ0.565*** 0.522*** 0.306***
R² 0.457
Pseudo R² 0.536 0.545 0.533
Multicollinearity 6.5441
Jarque-Bera test 5.5689*
Breush-Pagan test 22.534**
Article accepted in Regional Studies
17 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
After identifying these spatial concentrations, we drew up a spatial model that allowed us to test the
influence of different variables on the density of cases (Table 4). There are no multicollinearity issues
(since the conditional multicollinearity number is 6.5441) or heteroskedasticity (Breusch-Pagan test –
BP=22.534, rejecting the null hypothesis of homoscedasticity at 5%), and the Jarque-Bera non-
normality statistic on the residuals shows a non-significant value (JB=5.5689) at 5%.
We found that ρ is positive and significant, confirming the spatial dependency effect. Digging further,
its minimal value declined between March 8 and March 28, indicating a transmission effect, but one
that decreases over time. We can thus assume that the lockdown implemented on March 8 did indeed
help to slow the spread of COVID-19. This affirmation is confirmed by lower estimates of indirect
effects (i.e., contagion) as we move forward in time.
The findings show that the level of wealth influences the incidence of COVID-19 (strong and highly
significant coefficient). More specifically, the estimated effect of GDP/c indicates an increase in the
density of COVID-19 cases by 0.345% if we look at the direct effects. We can also note that the number
of passengers is positive but less significant. For Italy, we can therefore say that the higher the number
of visitors at an airport in a province, the higher the number of cases detected per 100,000 inhabitants.
From this perspective, Lombardy is an interesting case as it includes the city of Milan and its
international airport and thereby concentrates international connections and flows of people in Europe
and internationally. Given the intense economic and commercial exchanges between Italy and China,
Lombardy is an area that is highly vulnerable to the emergence of the virus in Italy, especially via Milan,
as are the adjacent regions and the Emilia-Romagna and Veneto regions whose economic activities
(industry, tourism) are closely tied to Lombardy. Strong connectivity with the rest of the world is a great
asset for regions in terms of economic attractiveness, but our findings show that it rapidly becomes a
threat during a disease outbreak as such areas are liable to be the first and most severely affected.
Population density also appears to play a role in the case density observed. Both the direct and indirect
effects are positive and significant, in line with our expectations, although the direct effect is the most
significant. An increase of 1% in provincial density leads to an increase of 0.056% in the incidence of
COVID-19 with respect to total effect, and 0.014% in indirect effect. Thus, high population density in
Article accepted in Regional Studies
18 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
a province and its neighbouring provinces appears to increase the incidence. With regard to density, the
role of spatial proximity in everyday situations is shown to be a key factor in the spread of viruses such
as SARS-CoV-2. This confirms what has already been observed on a global scale, where a higher
incidence has been detected in large cities.
On the other hand, the percentage of elderly people (65+ yo) in the overall population does not explain
the incidence of infection (not significant). It thus appears that the spatial conditions of social life are
more likely to explain the spread of an epidemic than the presence of at-risk populations. This can be
explained by the fact that age and personal social attributes have little or no impact on a person’s
contagiousness, and that they only have an incidence (notably age) on determining populations liable to
suffer from acute cases and death. The kinetics of spatial dissemination of SARS-CoV-2 is therefore
highly dependent on the spatial organisation of societies from the moment one or several cases is
introduced without being immediately detected. This dependency is interesting as, to a certain extent, it
can guide the construction of alternative scenarios of spatial spread of SARS-CoV-2 in a given area,
keeping in mind the scientific knowledge available in terms of its social and economic geography,
spatial practices and population(s) (including its mobility patterns), and housing characteristics in
relation to the composition of households.
4. Conclusion and discussion regarding lockdown strategies against the
spread of COVID-19
The COVID-19 outbreak is unprecedented in recent human history. The reasons for its singularity are
geographic and spatial (outbreaks concomitant with areas of high population density and economic
development).
In tune with several other researchers (Paez et al., 2020; Hellewell et al., 2020), we demonstrated the
impact of lockdown measures in controlling the spread of COVID-19. We used two techniques to detect
the positive effect of lockdown (local Moran statistics and SAR model) and showed the persistence of
spatial autocorrelation across time, signifying that the spread of the virus was controlled by the
Article accepted in Regional Studies
19 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
lockdown. Our study illustrates why some geographical areas are less affected than others. In the case
of Italy, the Mezzogiorno region clearly reported a far lower number of cases than the rest of the country.
Thus, areas that are generally associated with socio-economic difficulties end up being protected. In
addition, the rapid lockdown strategy introduced by the Italian government on 8 March 2020 was almost
certainly a key factor in slowing the spread of the virus to the south of the country. Furthermore, we
demonstrated a decline in the spatial dependence of the incidence of cases, which could indicate a
deceleration in its transmission linked to the positive effects of the containment measure. We also
showed that wealthy, high-density regions that are well connected by air are affected to a greater degree.
Our knowledge of the second wave to date, however, indicates that the geographical pattern of COVID-
19 has changed, with more regions now being affected.
While globalisation can be an accelerator of epidemics or pandemics, we are now seeing that it can also
be a factor in the expansion of ways to prevent and cure such diseases. The largest research centres have
undertaken to publish all their work on the virus in open access as quickly as possible. At the same time,
we could also ask whether the globalisation of information is partially responsible for fuelling panic and
alarm, especially as viruses have always been vectors of fear (Smith, 2006; Kott and Limaye, 2016).
The resilience of the COVID-19 pandemic and the exponential growth in casualties has given rise to
numerous questions about the strategies put in place to control it. In our paper, the international
community refers to all nation states (or groups of nations, such as the European Union) and their
respective public opinion leaders, multilateral organisations and NGOs for health protection and
monitoring, and world-renowned health research institutes mobilised to combat the virus and its spread.
At the start of the outbreak, however, its international diffusion was less evident, so much so that more
than a month after its first public appearance, the new virus had been met with relatively widespread
polite indifference. The epidemic’s expansion was indirectly favoured by, on the one hand, an
underestimation by the major powers, convinced that they were safe from contamination and, on the
other hand, procrastination by much of the population who did not take the virus and its contagiousness
seriously. In addition, we should not forget the deep suspicion regarding the official data shared by the
Article accepted in Regional Studies
20 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
Chinese authorities concerning the reality of the epidemic in China, especially in Wuhan, the city at the
centre of the outbreak.
The analysis presented above, and based on the Italian case, offers some strategic guidelines in the event
of a new COVID-19-type epidemic breaking out in one of the major centres of economic globalisation.
The introduction of a general and complete lockdown as early as possible certainly appears an effective
response as it can contain the virus spatially and reduce the internal diffusion kinetics within each area,
as demonstrated by Orea and Álvarez (2020). The results are interesting and encouraging for countries
such as France and Spain that could potentially have adopted a different approach, but which opted for
this strategy due to the lack of alternative resources (mass testing, mass mask wearing, etc.). It is
especially useful in terms of public health policy to compare with pandemic situations that can evolve
in countries where the government introduces neither a targeted strategy nor a lockdown (US, Brazil).
During a first ‘preventive’ quarantine, the authorities can prepare for a partitioned and targeted lifting
of the lockdown, identifying the most concentrated areas of COVID-19 cases where epidemic kinetics
are highest, and then pursuing systematic testing and a targeted rather than generalised lockdown of
clusters of infected areas and people. The first areas where the lockdown is lifted, in accordance with
the average incubation period, should be those furthest from the initial epidemic centre or from areas
that have the most cases and the highest kinetics. The areas “kept distant from the epidemic" should be
those that require the lowest amount of testing. In the meanwhile, the government can concentrate
resources in areas where the most infections and the most acute cases are concentrated. However, given
the apparent benefits of the initial lockdown indicated by the Italian case, in the course of a partitioned
and targeted lifting of the lockdown, all inter-territorial mobility will need to be banned or reduced to a
minimum. Incidentally, this should also be a fundamental principal from the very outset of the initial
lockdown (which was not the case in France that saw numerous movements from Paris to the French
coastal regions between 14 and 17 March). The absence of inter-territorial travel would help to preserve
the disassociation between areas observed in the epidemic kinetics thanks to the lockdown.
With regard to the limitations of our study, we do not as yet have sufficient micro-data to map human
mobility. Indeed, as Sorichetta et al. (2016) argued in the case of malaria, very fine-scale data on
Article accepted in Regional Studies
21 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
migration makes it possible to trace the connectivity between regions and population movements, and
thus to identify the best local and national strategies to eradicate epidemics. In future studies, it would
therefore be interesting to analyse such intra-territorial mobility. In the case of northern Italy and in
relation to the density of the fabric of economic activity (notably with the presence of industrial districts
comprising thousands of SMEs), we can assume that inter and intra-regional migrations of people
(between Lombardy, the Veneto region, Emila Romagna and Piedmont) are relatively high. It would
thus be interesting to take this study further by exploring relational and proximity effects using indicators
of people’s mobility such as the commute between home and work, for instance. Following Paez et al.
(2020), climate/environmental explanatory variables could be introduced in models in future studies to
help refine them. Finally, another line of research concerns the modelling strategy adopted to evaluate
the spread of Covid-19 or other infections. The dependent variable could be transformed into a
probability ratio, i.e. a propensity to be infected or an incidence ratio. Thus, an alternative model would
consist of using spatial probit model for grouped-data, similar to the one proposed by Chasco et al
(2019). This would involve analysing the likelihood of being infected by Covid-19 in a formal context
of spatio-temporal data.
Article accepted in Regional Studies
22 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
Bibliography
Adiga, A., Venkatramanan, S., Schlitt, J., Peddireddy, A., Dickerman, A., Bura, A., ... & Machi, D.
(2020). Evaluating the impact of international airline suspensions on the early global spread of COVID-
19. medRxiv.
Ahasan R, Hossain MM. (2020). Leveraging GIS and spatial analysis for informed decision-making in
COVID-19 pandemic. Health Policy Technol. doi:10.1016/j.hlpt.2020.11.009
Ali, M., Emch, M., Donnay, J. P., Yunus, M., & Sack, R. B. (2002). The spatial epidemiology of cholera
in an endemic area of Bangladesh. Social science & medicine, 55(6), 1015-1024.
Anselin, L. (1995). Local indicators of spatial association—LISA. Geographical analysis, 27(2), 93-
115.
Anselin, L. (1999). The future of spatial analysis in the social sciences. Geographic information
sciences, 5(2), 67-76.
Anselin, L. (2010). Thirty years of spatial econometrics. Papers in regional science, 89(1), 3-25.
Anselin, L., Syabri, I., & Kho, Y. (2010). GeoDa: an introduction to spatial data analysis. In Handbook
of applied spatial analysis (pp. 73-89). Springer, Berlin, Heidelberg.
Anselin, L., Syabri, I., & Smirnov, O. (2002). Visualizing multivariate spatial correlation with
dynamically linked windows. In Proceedings, CSISS Workshop on New Tools for Spatial Data
Analysis, Santa Barbara, CA.
Atique, S., Chan, T. C., Chen, C. C., Hsu, C. Y., Iqtidar, S., Louis, V. R., ... & Chuang, T. W. (2018).
Investigating spatio-temporal distribution and diffusion patterns of the dengue outbreak in Swat,
Pakistan. Journal of infection and public health, 11(4), 550-557.
Baltagi, B. (2008). Econometric analysis of panel data. John Wiley & Sons.
Bloom, B.R. (2003). Editorial: “Lessons from SARS”, Science, vol. 300, p. 701.
Article accepted in Regional Studies
23 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
Bowen Jr, J. T., & Laroe, C. (2006). Airline networks and the international diffusion of severe acute
respiratory syndrome (SARS). Geographical Journal, 172(2), 130-144.
Chasco C, Aroca P, Anselin L (2019) Probit Models for Grouped-data Migration Flows: A Theoretical
Note. Economía 42(84):1-8.
Chen, D. R., & Wen, T. H. (2010). Socio-spatial patterns of neighborhood effects on adult obesity in
Taiwan: a multi-level model. Social Science & Medicine, 70(6), 823-833.
Chen, N., Zhou, M., Dong, X., Qu, J., Gong, F., Han, Y., ... & Yu, T. (2020). Epidemiological and
clinical characteristics of 99 cases of 2019 novel coronavirus pneumonia in Wuhan, China: a descriptive
study. Chinese Journal of Epidemiology, 395(10223), 507-513
Chinazzi, M., Davis, J. T., Ajelli, M., Gioannini, C., Litvinova, M., Merler, S., ... & Viboud, C. (2020).
The effect of travel restrictions on the spread of the 2019 novel coronavirus (COVID-19) outbreak.
Science.
Cliff, A. D., Haggett, P., & Ord, J. K. (1986). Spatial aspects of influenza epidemics. Routledge.
Cliff, A. D., Haggett, P., & Smallman-Raynor, M. (2004). World atlas of epidemic diseases. Oxford
University Press
Cliff, A., & Ord, K. (1972). Testing for spatial autocorrelation among regression residuals. Geographical
analysis, 4(3), 267-284.
Cotten, M., Watson, S. J., Zumla, A. I., Makhdoom, H. Q., Palser, A. L., Ong, S. H., ... & Albarrak, A.
(2014). Spread, circulation, and evolution of the Middle East respiratory syndrome coronavirus. MBio,
5(1), e01062-13.
Cromley, E. K., Wilson-Genderson, M., & Pruchno, R. A. (2012). Neighborhood characteristics and
depressive symptoms of older people: Local spatial analyses. Social science & medicine, 75(12), 2307-
2316.
Article accepted in Regional Studies
24 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
Desjardins, M. R., Hohl, A., & Delmelle, E. M. (2020). Rapid surveillance of COVID-19 in the United
States using a prospective space-time scan statistic: Detecting and evaluating emerging clusters. Applied
Geography, 102202.
Dobis, E. A., Stephens, H. M., Skidmore, M., & Goetz, S. J. (2020). Explaining the spatial variation in
American life expectancy. Social Science & Medicine, 246, 112759.
Elhorst, J. P. (2010). Applied spatial econometrics: raising the bar. Spatial economic analysis, 5(1), 9-
28.
Elhorst, J. P. (2014). Spatial econometrics: from cross-sectional data to spatial panels (Vol. 479, p. 480).
Heidelberg: Springer.
Fleuret, S., & Charreire, H. (2019). Practicing health geography in 2017 IMGS 2017 Editorial. Social
Science & Medicine, 112123.
Geary, R. C. (1954). The contiguity ratio and statistical mapping. The incorporated statistician, 5(3),
115-146.
Hellewell, J., Abbott, S., Gimma, A., Bosse, N. I., Jarvis, C. I., Russell, T. W., ... & Eggo, R. M. (2020).
Feasibility of controlling 2019-nCoV outbreaks by isolation of cases and contacts. medRxiv.
Hogbin, V. (1985). Railways, disease and health in South Africa. Social Science & Medicine, 20(9),
933-938.
Holshue, M. L., DeBolt, C., Lindquist, S., Lofy, K. H., Wiesman, J., Bruce, H., ... & Diaz, G. (2020).
First case of 2019 novel coronavirus in the United States. New England Journal of Medicine.
Kearns, R., & Moon, G. (2002). From medical to health geography: novelty, place and theory after a
decade of change. Progress in Human Geography, 26(5), 605-625.
Koch, T., & Koch, T. (2005). Cartographies of disease: maps, mapping, and medicine (p. 840).
Redlands, CA: Esri Press.
Article accepted in Regional Studies
25 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
Kott, A., & Limaye, R. J. (2016). Delivering risk information in a dynamic information environment:
Framing and authoritative voice in Centers for Disease Control (CDC) and primetime broadcast news
media communications during the 2014 Ebola outbreak. Social Science & Medicine, 169, 42-49.
Kramer, A. M., Pulliam, J. T., Alexander, L. W., Park, A. W., Rohani, P., & Drake, J. M. (2016). Spatial
spread of the West Africa Ebola epidemic. Royal Society open science, 3(8), 160294.
Lai, P. C., So, F. M., & Chan, K. W. (2008). Spatial epidemiological approaches in disease mapping
and analysis. CRC press.
Le Gallo, J. (2014). Cross-Section Spatial Regression Models. In Fischer, M., and Nijkamp, P.,
Handbook of regional science, 1511.
Le Goff, J. M. (2011). Diffusion of influenza during the winter of 1889-1890 in Switzerland. Genus,
67(2).
Lee, S. S., & Wong, N. S. (2010). Characterizing the initial diffusion pattern of pandemic (H1N1) 2009
using surveillance data. PLoS currents, 2.
LeSage, J. P., & Pace, R. K. (2009). Introduction to Spatial Econometrics, CRC Press/Taylor & Francis
Group.
Leslie, T. F., & Kronenfeld, B. J. (2011). The Colocation Quotient: A New Measure of Spatial
Association Between Categorical Subsets of Points. Geographical Analysis, 43(3), 306-326.
Lourenco, J., de Lima, M. M., Faria, N. R., Walker, A., Kraemer, M. U., Villabona-Arenas, C. J., ... &
Recker, M. (2017). Epidemiological and ecological determinants of Zika virus transmission in an urban
setting. Elife, 6, e29820.
Mayer, J. D. (2000). Geography, ecology and emerging infectious diseases. Social science & medicine,
50(7-8), 937-952.
Meade, M. S. (2014). Medical geography. The Wiley Blackwell Encyclopedia of Health, Illness,
Behavior, and Society, 1375-1381.
Article accepted in Regional Studies
26 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
Millo, G., & Piras, G. (2012). splm: Spatial panel data models in R. Journal of Statistical Software,
47(1), 1-38.
Moran, P. A. (1950). Notes on continuous stochastic phenomena. Biometrika, 37(1/2), 17-23.
Munster, V. J., Koopmans, M., van Doremalen, N., van Riel, D., & de Wit, E. (2020). A novel
coronavirus emerging in China—key questions for impact assessment. New England Journal of
Medicine, 382(8), 692-694.
Orea L., and Álvarez, I. (2020) How effective has the Spanish lockdown been to battle COVID-1 9? A
spatial analysis of the coronavirus propagation across provinces. Documentos de trabajo FEDEA, Nº. 3,
pp. 1-33
Reyes, R., Ahn, R., Thurber, K., & Burke, T. F. (2013). Urbanization and infectious diseases: general
principles, historical perspectives, and contemporary challenges. In Challenges in Infectious Diseases
(pp. 123-146). Springer, New York, NY.
Robertson, C., & Nelson, T. A. (2014). An overview of spatial analysis of emerging infectious diseases.
The Professional Geographer, 66(4), 579-588.
Roth, D., Otterstatter, M., Wong, J., Cook, V., Johnston, J., & Mak, S. (2016). Identification of spatial
and cohort clustering of tuberculosis using surveillance data from British Columbia, Canada, 1990–
2013. Social Science & Medicine, 168, 214-222.
Smallman-Raynor, M., & Cliff, A. D. (2008). The geographical spread of avian influenza A (H5N1):
panzootic transmission (December 2003–May 2006), pandemic potential, and implications. Annals of
the Association of American Geographers, 98(3), 553-582.
Smith, R. D. (2006). Responding to global infectious disease outbreaks: lessons from SARS on the role
of risk perception, communication and management. Social science & medicine, 63(12), 3113-3123.
Sorichetta, A., Bird, T. J., Ruktanonchai, N. W., zu Erbach-Schoenberg, E., Pezzulo, C., Tejedor, N., ...
& Tatem, A. J. (2016). Mapping internal connectivity through human migration in malaria endemic
countries. Nature. Scientific data, 3(1), 1-16.
Article accepted in Regional Studies
27 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
Souris, M., Gonzalez, J. P., Shanmugasundaram, J., Corvest, V., & Kittayapong, P. (2010).
Retrospective space-time analysis of H5N1 Avian Influenza emergence in Thailand. International
journal of health geographics, 9(1), 3.
Spielman, S. E., & Yoo, E. H. (2009). The spatial dimensions of neighborhood effects. Social science
& medicine, 68(6), 1098-1105.
Sridharan, S., Tunstall, H., Lawder, R., & Mitchell, R. (2007). An exploratory spatial data analysis
approach to understanding the relationship between deprivation and mortality in Scotland. Social
Science & Medicine, 65(9), 1942-1952.
Stakhovych, S., & Bijmolt, T. H. (2009). Specification of spatial models: A simulation study on weights
matrices. Papers in Regional Science, 88(2), 389-408.
Vaneckova, P., Beggs, P. J., & Jacobson, C. R. (2010). Spatial analysis of heat-related mortality among
the elderly between 1993 and 2004 in Sydney, Australia. Social science & medicine, 70(2), 293-304.
Vega, S. H., & Elhorst, J. P. (2013, August). On spatial econometric models, spillover effects, and W.
Working Paper.
Wang, D., Hu, B., Hu, C., Zhu, F., Liu, X., Zhang, J., ... & Zhao, Y. (2020). Clinical characteristics of
138 hospitalized patients with 2019 novel coronavirus–infected pneumonia in Wuhan, China. Jama.
Wang, J. F., Christakos, G., Han, W. G., & Meng, B. (2008). Data-driven exploration of ‘spatial pattern-
time process-driving forces’ associations of SARS epidemic in Beijing, China. Journal of Public Health,
30(3), 234-244.
Weimann, A., Dai, D., & Oni, T. (2016). A cross-sectional and spatial analysis of the prevalence of
multimorbidity and its association with socioeconomic disadvantage in South Africa: A comparison
between 2008 and 2012. Social Science & Medicine, 163, 144-156.
Wood, E., Chan, K., Montaner, J. S., Schechter, M. T., Tyndall, M., O’Shaughnessy, M. V., & Hogg,
R. S. (2000). The end of the line: has rapid transit contributed to the spatial diffusion of HIV in one of
Canada’s largest metropolitan areas?. Social science & medicine, 51(5), 741-748.
Article accepted in Regional Studies
28 contact : Sebastien Bourdin, sbourdin@em-normandie.fr www.sebastienbourdin.com
Wood, E., Chan, K., Montaner, J. S., Schechter, M. T., Tyndall, M., O’Shaughnessy, M. V., & Hogg,
R. S. (2000). The end of the line: has rapid transit contributed to the spatial diffusion of HIV in one of
Canada’s largest metropolitan areas?. Social science & medicine, 51(5), 741-748.
World Health Organization (2004) Guidelines for the global surveillance of severe acute respiratory
syndrome (SARS). WHO Publication
World Health Organization (2016) Mapping the risk and distribution of epidemics in the WHO African
Region, Technical report
World Health Organization (2020) Naming the coronavirus disease (COVID-19) and the virus that
causes it, Technical guidance
Xu, X. W., Wu, X. X., Jiang, X. G., Xu, K. J., Ying, L. J., Ma, C. L., ... & Sheng, J. F. (2020). Clinical
findings in a group of patients infected with the 2019 novel coronavirus (SARS-Cov-2) outside of
Wuhan, China: retrospective case series. Bmj, 368.
Zhu, G., Xiao, J., Liu, T., Zhang, B., Hao, Y., & Ma, W. (2019). Spatiotemporal analysis of the dengue
outbreak in Guangdong Province, China. BMC infectious diseases, 19(1), 493.
