Resilience to health shocks and the spatial extent of local labour markets: evidence from the Covid-19 outbreak in Italy

ABSTRACT In addition to the general issue that fewer interpersonal contacts reduce the speed of contagion, less attention has been paid to the spatial configuration of such contacts. In Italy, Covid-19 severely affected the most industrialized area of the country, where the network of commuting flows is particularly dense. We investigate the relationship between workers’ mobility and the diffusion of the disease by computing, for each municipality, the intensive and extensive margins of commuting flows and by measuring excess mortality over the period January–May 2020. We find that if commuting patterns were 90% of those observed in the data, Italy would have suffered approximately 2300 fewer fatalities during the first pandemic cycle.


INTRODUCTION
The daily mobility of individuals for motives of labour is one of the main features of developed societies, so the spatial extent of local labour markets (LLMs) is in fact defined on the basis of the geography of commuting flows.The openness of such areas generates costs and benefits for the governance of the local economy, particularly in the case of a pandemic, as it is one of the conditions that determine their ability to contain such an intense health disaster (Gong et al., 2020).In this article we investigate how the openness of LLMs, as defined by the structure of the commuting network, influences the resilience of cities to health shocks, such as the Covid-19 outbreak.In particular, we explore the aforementioned dynamics in Italy, the first Western country to be deeply affected by the disease.
There are several reasons why we believe that such an empirical analysis is needed for a thorough comprehension of the phenomenon.First, the initial burst of the virus spread across Italy first and foremost in the most industrialized area of the country, suggesting a correlation between the structural features of local economies and the epidemic.
Second, places characterized by a high density of economic activities also exhibit dense networks of spatial interactions (especially in the form of commuting flows), and these flows in their turn place such areas at more severe epidemiological risk, as shown by past (Charu et al., 2017;Zhou et al., 2019) and present (Fang et al., 2020;Glaeser et al., 2020) epidemics.In fact, Bergamo, the city among the provincial capitals with the largest share of incoming and outgoing workers compared with the population overall, also experienced the greatest increase in fatalities recorded in March 2020 (+429%) when compared with the 2015-19 average.Unsurprisingly, the openness of Bergamo's labour market is remarkable because it is the epicentre of the largest industrial district in the whole nation. 1  In response to the diffusion of Covid-19 during the first months of 2020, several national governments imposed unprecedented lockdown restrictions to slow the infection rate and save lives.Indeed, in the absence of medical treatments, such as vaccines or pharmaceuticals, the limitation of interpersonal contacts was the key policy for containing viral infections (Haushofer & Metcalf, 2020;Van Bavel et al., 2020).As a result, the travel behaviours of people have been drastically altered (De Vos, 2020), with dramatic economic and social consequences (Bonaccorsi et al., 2020).
Although the literature exploring the main drivers of the geographical diffusion of Covid-19 is sizeable, understanding how local economies are related to the perturbation caused by the pandemic is still an open research question.In this article we contribute to this ongoing debate by investigating the role played by the spatial extent of LLMs in filtering the initial spread of the disease through workers' mobility and, therefore, in influencing the resilience of cities to the health shock.
To this end, we analyse the pre-existing characteristics of commuting patterns at the municipality level 2 using data from the latest official countrywide assessment of mobility for Italy.Similarly, we measure the local depth of the pandemic shock by considering excess mortality over the period January-May 2020, comprising several weeks both before and after the most critical part of the first pandemic cycle.We also consider a broad set of additional municipality characteristics to control for other specific dynamics.After assembling the novel dataset, our empirical strategy exploits within-municipality variation in excess mortality over time by estimating a two-way fixed effects model in which all our explanatory variables are interacted with month dummies.
This article provides some relevant novelties in two directions: first, we examine the structure of the commuting network by computing both the intensive and the extensive margins of commuting flows; and second, we exploit more granular and heterogeneous data by performing the analysis at the municipality level, while most of the previous studies focused on main cities, provinces or regions.
More precisely, we compute two synthetic indices that describe commuting flows under different perspectives: the intensity of external mobility and the centrality of each municipality.The first indexthe intensive marginis defined as the total number of workers moving from and to a municipality over its population, similar to what is proposed by Murgante et al. (2020).In other words, it is a proxy for the share of the population exposed to the possibility of the virus being imported from elsewhere.The second indexthe extensive marginis based on the topological concept of relative degree centrality of a node within a network, measuring the importance and openness of a municipality, as defined by Patuelli et al. (2009Patuelli et al. ( , 2010)).In other words, the aim is to measure the number of other different places (each of which may have a different infection rate) to which the municipality is connected.
In addition, we further investigate the commuting dynamics by exploring the spatial heterogeneity of lockdown intensities induced by different government policies, such as the anticipation of mobility restrictions (imposed using containment areas) and the reduction of active workers (imposed using the closure of non-essential economic activities).Shedding light on whether such measures played a role in flattening the mortality curve is therefore important in the design of future policies aimed at containing new outbreaks.
Our findings suggest that the spatial extent of LLMs was crucial in influencing the resilience of cities to the Covid-19 health shock.In particular, a 1 percentage point increase in the intensive margin is associated, on average, with 1.43 and 0.91 percentage point increases in excess mortality in March and April 2020, respectively, while the same increase in the extensive margin is associated with a 3.44 increase in our outcome of interest in April.As a result, more isolated and less central places are found to be more resilient than others.Moreover, we report suggestive evidence on the role of containment areas and businesses closure in reducing Covid-19-related fatalitiesand therefore in increasing the resilience of local economiesby cutting down mobility among municipalities.
Within the massive empirical literature on Covid-19, our research is mainly related to two separate lines of work that have proposed commuting flows and the characteristics of local economies as major explanations for the observed unequal spread of Covid-19 across subnational areas.By focusing on those contributions dealing with the Italian context, 3 a first strand of research has shown that human mobility played a crucial role in the propagation of the disease during the first wave of the pandemic, as highlighted by the striking relationship between mobility flows and both the net reproduction number (R t ) of the virus (Cintia et al., 2020) and the resulting excess mortality (Ascani et al., 2021b;Iacus et al., 2020).Linked to this line of work, a group of studies has analysed the mobility patterns of people during the emergency and the consequent change in the structure of commuting flows, including Beria and Lunkar (2021) and Pepe et al. (2020).A second strand of research has emphasized how the most industrialized area of the country was more severely affected by the earliest phase of the pandemic due to its socioeconomic characteristics, such as greater domestic and international connectivity (Bourdin et al., 2021), and a greater degree of interaction between workers employed in locations endowed with a high density of industries (Ascani et al., 2021a).Hence, more productive and interconnected areas were found to be more exposed to the spread of infectious diseases than others (Bloise & Tancioni, 2021).
Within the broad literature on regional resilience, our research is also related to a rich amount of studies that have analysed how the social, demographic and economic characteristics of territories determine their heterogeneous resilience to a particular shock, including (among others) Bristow and Healy (2014), Diodato and Weterings (2015), Kitsos and Bishop (2018) and Martin et al. (2016).
In this article we connect these three strands of research by investigating the interplay between labour market dynamics, the initial spread of Covid-19 and the resilience to the shock.
The remainder of the article is organized as follows.Next we provide a conceptual framework to the notion of resilience in times of a pandemic and briefly summarize the timeline of the Covid-19 crisis in Italy.We then describe the data used in the analysis, followed by a discussion of the empirical strategy, the main results and the robustness checks.Finally, we explore the spatial heterogeneity of lockdown intensities, and conclude.

THE ANATOMY OF RESILIENCE SHAPED BY COVID-19
Resilience in a time of pandemic In the current Covid-19 crisis, the notion of resilience has increasingly taken up within not only academic research but also institutional and policy debates.Such a multidisciplinary concept does not have a unique definition because its meaning varies according to the purposes for which it is used and depends on the scale, nature and duration of the shock to which it refers (Martin, 2018).Broadly speaking, it defines the ability of a system to absorb an external shock, bounce back and reorganize itself afterwards.Such a concept has been progressively introduced into the regional economic research when a proliferation of studies started to tackle the question of why some local economies where more resilient than others in facing an increasing number of shocks and disruptions, such as natural hazards (e.g., Hong et al., 2021;Zhou et al., 2010) and financial crisis (e.g., Capello et al., 2015;Davies, 2011).
Following the conceptual framework developed by Martin and Sunley (2015) and Martin et al. (2016), the resilience of a local system is usually defined as a complex process involving several phases, such as (1) the vulnerability to a shock (defined as the propensity to be hit by that shock), (2) the resistance to it (measured as the impact of the shock on a specific outcome), (3) the reorientation after the shock (defined as the ability to adjust and adapt to the shock), and (4) the recoverability from the shock (measured by the speed of return to a previous equilibrium).According to the same framework, these phases are influenced by many factors, such as local economic characteristics and any supportive policy aimed at softening the impact of the shock (for a detailed explanation of each phase of the resilience process, see Figure 1).
Compared with the other shocks to which the literature on regional resilience normally refers, the specific features of the disruption caused by Covid-19 are of a different nature, so that their analysis has given rise to new research questions (Gong et al., 2020).For instance, understanding why some places were more severely affected than others during the first wave of the pandemic is still a relatively unexplored issue.Given that infectious diseases tend to spread through human interaction, the risk for a local economy of being hit by a health shock strongly depends on its openness, which is shaped by the daily mobility of individuals for motives of labour.Therefore, the analysis of the geography of commuting flows is crucial when it comes to understanding the vulnerability phase (Massaro et al., 2018).Furthermore, these flows affect a type of resistance called 'human' resistance (Ascani et al., 2021b), which can be measured as excess mortality compared with a previous average.In its turn, such an unfortunate outcome directly affects the 'economic' resistance of a local system because it determines the introduction of lockdown restrictions (aimed at slowing down the infection rate) on which the reorientation and recoverability phases crucially depend.
Since our period of analysis is one in which the health crisis has always been far from over, this article focuses on the first two phases of the resilience process (i.e., those included in the area bounded by the dotted line in Figure 1).On this basis, we investigate the extent to which the pre-existing commuting networkshaped by the structural features of local economiesaffected excess mortality during the first pandemic cycle in Italy.Although partial, this analysis may represent a first assessment of the resilience of cities to the Covid-19 health shock.

Covid-19 in Italy
Our empirical analysis focuses on Italy, the first Western country that was forced to shut down its economy to 'flatten the curve' and contain the diffusion of Covid-19.Therefore, Italy represents the ideal scenario for investigating the relationship between commuting flows and the initial diffusion of the virus because government and citizens were unprepared to face the pandemic, while both policymakers and populations of other European countries have been influenced by the Italian case.Such an unfortunate situation limits the number of confounding factors because there were no countermeasures or policy responses during the first weeks of the outbreak.
The timeline of the main events that occurred during the first wave of the pandemic (summarized in Figure 2) is the following.The first two Covid-19 cases in Italy were officially detected on 30 January, after a Chinese couple travelled from Wuhan to Milan, Verona, Parma and Florence.The first cases of secondary transmission were identified near Codogno and Vo' (two municipalities in the Lombardy and Veneto regions, respectively) on 21 February, and two days later the Italian government enforced mobility restrictions into and from these areas (DPCM1, 2020).On 4 March, all schools and universities were closed (DPCM2, 2020).On 8 March, the lockdown was imposed for the first relevant 'red zone' of the country (DPCM3, 2020), that is, the whole Lombardy region and 14 additional provinces within the Emilia-Romagna, Marche, Piedmont and Veneto regions (for a detailed map, see Figure 7). 4On 11 March, the lockdown was extended to the whole nation (DPCM4, 2020), and many business activities open to the public were forced to close.Between 22 and 25 March, the 'economic' lockdown was tightened further by shutting down all nonessential economic activities and prohibiting any movement of people on Italian soil with few exceptions, such as for work or health reasons (DPCM5, 2020;DPCM6, 2020).This step marked the so called 'phase 1' of the epidemic, which gradually ended between 4 and 18 May.

DATA
To study the spatial diffusion of the recent Covid-19 pandemic, we rely on two main data sources: the Italian National Institute of Statistics (ISTAT) and the Italian Institute for Environmental Protection and Research (ISPRA).In the following section we describe the variables used in the empirical analysis.

Measuring resilience through excess mortality
For 7357 Italian municipalities out of 7904 (covering approximately 95% of the total population), we obtain data released by ISTAT on 9 July 2020, that is, the monthly number of fatalities occurring during the first five months of 2020 and the average monthly number of fatalities occurring during the same period in 2015-19.For simplicity, we refer to the latter data as the 'baseline' throughout the article.Our outcome of interest is then mortality growth, defined as the increase in fatalities recorded in January-May 2020 compared with the same period in the 'baseline': where i and t denote the municipality and month, respectively.This measure of the incidence of Covid-19 is directly related to the notion of local resilience (Boschma, 2015) since it computes the burden of the disease as a deviation from a pre-existing trend.We consider excess mortality our main outcome of interest over the official number of Covid-19 cases because it allows us to overcome, at least partially, major measurement errors and endogeneity issues related to the number of reported cases, such as non-random differences in screening procedures and testing capacity among areas.Indeed, it allows us to observe any Covid-19-related fatalities, even before 21 February, when the first Italian Covid-19 hotspots were identified. 6 Similarly, we prefer total fatalities over official Covid-19 fatalities because the latter are no longer considered a reliable measure due to differences in classification among hospitals (Buonanno et al., 2020).Moreover, it is plausible to expect that the official numbers are underestimating the true increase in mortality since a substantial number of people died without being tested (Bartoszek et al., 2020;Ciminelli & Garcia-Mandicó, 2020).Indeed, during the first quarter of 2020, Italy experienced 46,909 more deaths with respect to the average number of fatalities occurring in the same period during 2015-19, while the official Covid-19 fatalities declared by the Department of Civil Protection numbered 27,938 (INPS, 2020).Hence, it is likely that the majority of the remaining 18,971 fatalities were also caused by the pandemic. 7In addition, the use of such measures also allows us to consider the indirect effects of the pandemic, such as the Note: The timeline of the main events that occurred in Italy during the first wave of the pandemic is shown; hence, dates refer to 2020.Source: Authors' own elaboration.

REGIONAL STUDIES
possible increase in fatalities caused by other diseases that were not treated as usual due to hospital congestion.

Measuring the spatial extent of LLMs
The aim of this article is to investigate the role played by the spatial extent of LLMs in influencing their resilience to the spread of Covid-19.To this end, we use data on the network of commuting flows reported in the 2011 census in the form of a nationwide origin-destination matrix.
We measure the intensity of external mobility of each municipality by considering both the outflows, indicating the total number of workers w ij moving from their residential municipality i to any other municipality j ¼ 1, …, n (excluding j ¼ i), and the inflows, indicating the total number of workers w ji moving to municipality i from any other municipality j.We compute, for each municipality, the intensive margin of commuting, defined as the sum of the incoming and outgoing flows over the 2011 population of the area: We also consider a topological index.We first compute the total number of direct outward and inward connections of each municipality (degree centrality), that is, the set of origin-destination routes used by at least one worker to commute.We then define the extensive margin of commuting as the ratio between the observed and the maximum possible number of connections (n − 1) of a municipality:

Control variables
To separate the effect of commuting flows from other confounding factors, we consider another important dimension linked to the movement of people, such as internal mobility.To this endand by relying on the same 2011 censuswe compute an internal mobility index as the ratio of self-flows, indicating the total number of workers w ii moving within their residential municipality i to reach the workplace, to the 2011 population of the area.
We then further control for other variables potentially correlated with both excess mortality and commuting patterns.In particular, we add all those predictors that are essential in standard epidemiological models to explain the spatial diffusion of a disease (e.g., Bisin & Moro, 2020).Given that living in urban areas and in close proximity is likely to increase the probability of infection (Armillei et al., 2021;Desmet & Wacziarg, 2021), we first capture relevant geographical and demographic characteristics by including two dummy variables that take the value of 1 if a municipality is located near the sea (coastal) or at medium-high altitude (mountainous), and 0 otherwise, the log of the population density (ln density), and a proxy of physical proximity, defined as the log of the average number of m 2 per inhabitant in occupied dwellings (ln house m 2 pc).
Second, given that the fatality rates for males are two to three times higher than for females (Porcheddu et al., 2020), that the fatality rate is positively correlated with a larger presence of elderly people (Knittel & Ozaltun, 2020), that nursing homes and hospitals were the locations of the first outbreaks of the pandemic (Barnett & Grabowski, 2020), and that pollution can be an important codeterminant of Covid-19-related fatalities in northern Italy (Coker et al., 2020;Conticini et al., 2020;Dettori et al., 2021), 8 we also control for five measures of vulnerability to the pandemic: the share of the male population at the municipality level (share males); the share of the population older than 75 years at the municipality level (share_over75); the share of individuals older than 65 years cohabiting with younger individuals at the municipality level (share cohab over65); the number of hospital beds per inhabitant at the province level (hospital beds pc); and particulate matter PM 10 , defined as the average concentration (µg/m 3 ) at the province level (pm10).
Third, we account for differences in economic structure between areas by including a dummy variable that takes the value of 1 if a municipality is located within an industrial district (district), and 0 otherwise.Indeed, recent literature shows how thicker LLMs (characterized by a high density of industries) may foster higher levels of business and social interactions (e.g., Ascani et al., 2021b).Finally, we have seen how the pandemic has induced many workers to perform their duties from home, preventing them from travelling.Thus, it might be that municipalities with larger numbers of 'remote' workers experienced fewer Covid-19-related fatalities with respect to others.To capture this possible dynamic, we compute a working remotely index (remote working) by weighting the set of working remotely indices provided by Barbieri et al. (2021) by the labour force composition of each municipality (as defined by the one-digit ATECO sections 9 ).
All the data are publicly available. 10Table A1 in the supplemental data online reports standard descriptive statistics of the variables used in the empirical analysis; Table A2 online summarizes their definition (as well as their reference year, unit of observation and data source); while Figure B2 online reports a correlation matrix among covariates.

Descriptive evidence
In this section we briefly describe the spatial patterns of our main variables of interest.Figure 3 plots the spatial evolution of mortality growth in March 2020, that is, when Italy was severely affected by the pandemic (for other months, see Figure B3 in the supplemental data online).Clearly, we can note how Covid-19-related fatalities appear to be spatially clustered in the northern part of Italy, particularly in the Lombardy region and across the Po Valley area. 11Overall, the virus spread first and foremost in the most industrialized area of the country, suggesting a possible correlation between the structural

Econometric model
To examine the relationship between the characteristics of commuting flows and excess mortality, we estimate the following equation: where mortality growth it measures the increase in fatalities occurring in municipality i in month t, compared with the same period at 'baseline'.On the right-hand side, intensive margin i and extensive margin i are our municipality commuting indices interacted with a vector of monthly specific fixed effects, d t , accounting for the nationwide common evolution of excess mortality in a given month, such as the seasonal trend.By excluding January as the pre-outbreak period, the vectors of coefficients b t and g t capture the impact of the structural characteristics of commuting flows on excess mortality over the various months of the pandemic cycle.Z i × d t indicates the internal   mobility, geographical, demographic, vulnerability and economic controls also interacted with month dummies.
Then, a i is a full set of municipality-level fixed effects intended to absorb any difference in excess mortality due to time-invariant characteristics.Hence, by controlling for all these observed and unobserved characteristics, our identifying assumption is that no other factor correlated with workers' commuting systematically affects excess mortality.Finally, given that the geography of commuting flows analysed in this article essentially describes the spatial extent of LLMs (Kropp & Schwengler, 2016), 1 it are heteroskedasticity-and autocorrelation-consistent standard errors, respectively clustered at the LLM level.

Estimation results
Tables 1 and 2 report regression results for equation ( 4).
The rationale for the structure of Tables 1 and 2 (which can be read sequentially) is to progressively include fixed effects and control variables to test the strength of our estimates.
In Table 1 the first two columns report the estimated coefficients for the specifications in which the intensive and extensive margins, interacted with month dummies, are included one at a time.Accordingly, the main effects of the interactions are included as well.In column (3), the two margins are simultaneously estimated, while in column (4), the specification adds a full set of region fixed effects because the Italian national health system is managed at the regional level.Finally, column (5) substitutes the region fixed effects with a full set of municipality fixed effects to better control for time-invariant characteristics of each observation potentially correlated with both excess mortality and commuting flows. 12Overall, almost all the estimated coefficients of the two margins preserve their signs and significance throughout the columns, their magnitudes decreasing as the specifications become less parsimonious.
Table 2 reports estimates of regressions in which we have extended the set of controls.Interestingly, the estimated coefficients of the two margins remain consistent because, moving from the most parsimonious specification in column (1) to the most extended in column (4), their magnitude decreases without leading to a substantial increase in the standard error.Thus, our estimates suggest an important role played by the spatial extent of LLMs in influencing the resilience of municipalities during the Covid-19 outbreak.Indeed, the intensity of external mobilitythe intensive marginand the topological centrality of a municipalitythe extensive marginare positively correlated with excess mortality during the most critical part of the pandemic.This empirical evidence suggests how greater connectivity renders places less resilient to epidemic health shocks.
For simplicity, we discuss further only the estimates in column (4) because they are obtained with the most complete specification in relation to our data.Given that January is the reference period, regression results are close to zero and not statistically significant in February, that is, when the Covid-19 virus had just begun to spread.As expected, intensive margin shows its strongest correlation with excess mortality in March, when Italy was suddenly and severely affected by the pandemic.The coefficient indicates that, holding constant the other variables, a 1 percentage point increase in the share of population moving from and to a municipality is associated, on average, with a 1.43 percentage point increase in excess mortality.Following the introduction of all the containment measures previously described, this positive correlation then remains significant in April, but with a smaller magnitude (0.91), while it loses significance and approaches zero in May, hinting at how the lockdown was crucial in reducing excess mortality by cutting down workers' mobility among municipalities.The extensive margin, instead, shows its statistically significant correlation only in April, likely because the most central nodes of the commuting network played a pivotal role in spreading the disease later.The coefficient indicates that a 1 percentage point increase in the ratio between the observed and the maximum possible number of connections of a municipality is associated, on average, with a 3.44 percentage point increase in our outcome of interest, all else being equal. 13 That said, we provide some back-of-the-envelope calculations by considering three scenarios in which the

REGIONAL STUDIES
intensive margins among Italian municipalities would be equal to 90%, 80% and 70% of those actually observed in our data.In other words, we are interested in understanding what the reduction in mortality growth would have been had commuting flows been lower.For each scenario, Figure 6 shows these median reductions for the months in which our intensive margin coefficients are strongly significant.By focusing on the mildest scenario, 14 where our commuting index is cut by 10.0%, 4.8% and 5.3%, median reductions in mortality growth on March and April would translates to 1346 and 997 lives saved 15 across Italy, respectively.

Robustness checks
In the following section we briefly describe a set of robustness checks aimed at corroborating our empirical findings.First, intensive margin, which is defined as the sum of incoming and outgoing flows over the population of the area, could have some 'extreme' values.Indeed, as shown in Figure 4, for 58 of 7345 municipalities, the value of this index is > 1, implying that the number of workers moving from and to the municipality is greater than the number of residents.To check that these possible outliers are not affecting the results, we winsorize intensive margin by setting all the data > 99th percentile to the 99th percentile and all the data < 1st percentile to the 1st percentile.By so doing, we obtain an index that takes values between 0 and 1. Accordingly, we also winsorize in the same way extensive margin.We then estimate the most complete specification of equation ( 4) with these new variables.As shown in column (1) of Table 3, the regression results are very consistent with the main ones provided in Table 2, indicating that these possible outliers are not driving our estimates.Second, the first wave of Covid-19 has spread dramatically in some regions and not in others.The reasons for this phenomenon are difficult to assess since they most likely depend on many factors that favour the spread of the disease through different channels.For instance, in Italy, the virus severely affected the most industrialized regions, such as Lombardy, Emilia-Romagna, Piedmont and Veneto, which differ from the rest of the country in several characteristics.Thus, it might be relevant to verify that our previous findings are not affected by such differences among areas.To this end, we estimate the most complete specification of equation ( 4) with a more 'balanced' sample by considering only the municipalities located within these four regions.The regression results provided in column (2) of Table 3 suggest that the intensity of external mobility remains a determining factor in spread of the disease in the northern regions, while there is no evidence that the topology of the network contributed as well.Our interpretation is that, within the most infected areas of the country, what matters most is the total number of workers moving between municipalities, rather than the number of different connections.
Third, despite nine-year lagged explanatory variables perhaps solving some endogeneity issues, a reasonable concern is whether the 2011 commuting flows remain informative about the current ones.As proved by Gatto et al. (2020), they are since the spatial patterns of workrelated mobility seem to be remarkably preserved over such a long time interval.Moreover, we further test the consistency over time of the commuting network by comparing the 2011 share of outgoing flows, indicating the total number of workers moving from a municipality over its population, with the 2019 ones. 16As shown by the regression line depicted in Figure B6 in the supplemental data online, we find an almost one-to-one association between the two shares (R 2 ¼ 0.95).Finally, we computed the intensive and extensive margins using the 2001 and 1991 official countrywide assessments of mobility for Italy.If our 2011 mobility patterns are truly 'structural', we should expect similar estimates by relying on the 2001 and 1991 data.Once again, we estimate the most complete specification of equation ( 4), and the related regression results are provided in columns ( 3) and ( 4) of Table 3.The estimated coefficients involving the intensive margin are consistent in sign, significance and magnitude with the main ones provided by Table 2 (lending additional reliability to our empirical findings), while the estimated coefficient involving the extensive margin in April is barely not statistically significant using the 2001 data.

SPATIAL HETEROGENEITY IMPLIED BY LOCKDOWN INTENSITIES
In this section we provide some further evidence for the relationship between the spatial extent of LLMs and the initial diffusion of the virus by exploring the spatial heterogeneity of lockdown intensities induced by two policy interventions.The first source of geographical heterogeneity is based on some municipalities being located within the first relevant 'red zone' of the country, which was enforced on 8 March (DPCM3, 2020).In this area, mobility restrictions were anticipated compared with the rest of Italy; hence, it is plausible to expect that this early reduction in workers' commuting played a role in flattening the mortality curve more rapidly inside the 'red zone' than outside. 17The second source of geographical heterogeneity is based on the 'economic' lockdown imposed between 22 and 25 March (DPCM5, 2020;DPCM6, 2020), which forced the closure of non-essential economic activities, as well as those with high indices of physical proximity (Barbieri et al., 2021), indicating that the different sectoral composition of economic activities among municipalities leads to different shares of inactive workers, which consequently translate into different reductions in commuting flows between areas.

The introduction of the 'red zone'
We start our analysis by first considering the heterogeneity imposed by the introduction of a containment area, such as the 'red zone'.To this end, we first set a dummy variable (red zone) equal to 1 if a municipality is located within the locked area, the boundaries of which are drawn in Figure 7.
We then estimate the following augmented version of equation ( 4): where we add the triple interactions among: (1) the intensive and extensive margins; (2) the red zone dummy; and (3) the set of month dummies.Accordingly, the area main effects are also included.Table 4 reports regression results for equation ( 5).Similar to Table 2, all the specifications include month and municipality fixed effects, while columns (1-4) progressively add our sets of control variables.We focus on the coefficients estimated by the most complete specification in column (4).Given that the 'red zone' was enforced on 8 March, that the incubation time of the disease can be approximated in approximately five days (Lauer et al., 2020), and that reported Covid-19 fatalities tend to occur around 18-21 days after infection (Yang et al., 2020), we should observe the impact of the anticipated mobility restrictions in the area in reducing mortality from April onwards.As expected, the coefficients associated with the triple interactions involving the intensive margin in April and May are negative, but only the latter is statistically significant.This outcome suggests how an early reduction in the intensity of commuting flows, induced by an anticipated lockdown, could fosterafter some weeksa faster reduction in excess mortality related to external mobility, compared with areas without restrictions.Thus, containment areas could be useful to increase the resilience of local economies.Here, the triple interactions involving the extensive margin are not at all significant, likely because of collinearity with the red zone dummy, which captures most of the variability (Figure 5). 18Finally, we clearly find a positive and consequently decreasing correlation between being located within the 'red zone' and excess mortality.This finding confirms that the boundaries of the containment area were based on the high infection rate of the municipalities within it.

Introduction of the 'economic' lockdown
We now turn to exploiting the variation in the share of inactive workers due to the closure of non-essential economic activities.To this end, we rely on the most recent official data provided by ISTAT 19 to compute the number of active and inactive workers for each municipality, which are based on the list of ATECO sectors not suspended by the Italian government during the first part of 2020 (for a detailed list of sectors that were allowed to operate, see Table A4 in the supplemental data online).We then compute our share of interest by simply dividing the number of inactive workers by the total number of workers in the area: At this point, we are interested in understanding how much this 'economic' lockdown has tightened commuting flows among municipalities, given that many workers no longer had to reach their workplaces.To do so, we compute the share of inactive commuters for each municipality in the following way: where the total number of workers moving from and to a municipality (as explained in equation 2) has been first multiplied by the share of inactive workers in the municipality of destination and then weighted by the total incoming and outgoing flows.Next, we define municipalities with the largest share of inactive commuters by setting a dummy variable (high inactive) that equals 1 if the value computed through equation ( 7) is greater than the 66th percentile.These municipalitiesvisually correlated with the most industrialized regions of the countryare plotted in Figure 8.To test whether the closure of non-essential economic activities played a role in reducing Covid-19-related fatalities by tightening commuting flows further, we estimate the following augmented version of equation ( 4): where we add the triple interaction among: (1) the intensive margin; (2) the high inactive dummy; and (3) the set of month dummies.Accordingly, the area main effects are included.Note that we do not add the triple interaction involving the extensive margin because the closure of nonessential economic activities affected the intensity of

REGIONAL STUDIES
commuting flows, rather than the number of connections between municipalities.
Table 5 reports regression results for equation (8).Different from previous tables, column (1) directly reports the estimated coefficients for the most complete specification.Here, the negative and significant coefficient associated with the triple interaction in April suggests that municipalities with the largest share of inactive commuters would benefit from a faster reduction in excess mortality.Interestingly, this finding is in line with the recent empirical evidence provided by Borri et al. (2020) andDi Porto et al. (2020).In columns (2) and (3), we further examined this point by splitting the sample between municipalities located inside and outside the 'red zone'.The rationale for the sample split is testing whether this second policy made an additional contribution to reducing Covid-19-related fatalitiesthrough a further restriction of workers' commutingeven within an area that had already been affected by the first policy.As it is plausible to expect, the effectiveness of the 'economic' lockdown in reducing commuting flows further (and therefore in better controlling virus transmissions) lessened within the 'red zone'.In fact, the coefficients associated with the triple interactions are nowhere significant in column (2).Conversely, the coefficients retain their magnitudes and significance in column (3), indicating that the national dynamics also hold outside the 'red zone'.Accordingly, the same explanation applies.

CONCLUSIONS
The diffusion of Covid-19 is imposing tremendous challenges on our society, and it seems that now, more than in the past few decades, geography is considered a crucial feature for resilience to such a shock.With reference to the Italian case, the virus spread first and foremost in the most industrialized area of the country, where the high density of economic activities also exhibits dense networks of commuting flows.To the best of our knowledge, this article is among the very few exploring the role played by the openness of LLMs, as defined by the structure of the commuting network, in filtering the initial spread of the virus and, therefore, in influencing the resilience of Italian municipalities to the health shock.To this end, we computed the intensive and extensive margins of commuting flows, and we measured the spread of Covid-19 by considering excess mortality over the first five months of 2020, with clear implications in terms of measurement of resilience.
Using a rich and novel dataset, we have found that, during the most critical part of the first pandemic cycle (i.e., March-April 2020), municipalities with larger shares of population commuting from and to their borders for motives of labour tended to have higher Covid-19-related fatalities.Moreover, our findings also indicate that it is not only the intensity of external mobility that can influence the speed of diffusion of the virus and the depth of the shock but also the centrality of each municipality within a network of commuting flows.Indeed, municipalities strongly connected to many other different places experienced higher excess mortality in April as well.A back-of-the-envelope calculation suggests that if structural commuting patterns were 90% of those observed in the data, Italy would have suffered 1346 and 997 fewer fatalities in March and April 2020, respectively.Finally, we explored the spatial heterogeneity of lockdown intensities induced by different government 0.12 0.12 0.12 0.13  Resilience to health shocks and the spatial extent of local labour markets: evidence from the Covid-19 outbreak in Italy 2515 REGIONAL STUDIES policies, such as the introduction of the first relevant 'red zone' of the country and the closure of non-essential economic activities.We report suggestive evidence on the role of these policies in favouring a faster reduction in excess mortality and, therefore, in increasing the resilience of local economies.The overall conclusion arising from our analysis is that places more isolated and less central are found to be more resilient than others, all else being equal.This finding, in turn, suggests policy actions to strengthen the resistance to the shockand overcome the epidemicnot only considering the intensity of commuting flows but also addressing specific hotspots, central in the network of commuting flows.

DISCLOSURE STATEMENT
No potential conflict of interest was reported by the authors.

NOTES
1. Industrial districts are 'self-contained' labour markets mainly consisting of small and medium-sized enterprises specializing in the same economic activity.According to the latest industry and services national census, the industrial district of Bergamo is the largest in terms of population (802,731) and embedded municipalities (123).2. To avoid misunderstandings, we understand the openness of LLMs by means of commuting flows at the municipality level.This choice is driven by the need to exploit excess mortality data that is as granular as possible.Therefore, our empirical analysis does not use the official definition of LLM (as defined by the Italian National Institute of Statistics -ISTAT) as the unit of observation.3.For the sake of conciseness, this unavoidably incomplete literature review focuses only on those contributions that have used the Italian scenario as their case study.4. The 14 additional provinces that completed the containment areas are: Modena, Parma, Piacenza, Reggio nell'Emilia, Rimini, Pesaro e Urbino, Alessandria, Asti, Novara, Verbano-Cusio-Ossola, Vercelli, Padova, Treviso and Venezia. 5.The evolution of excess mortality in Italy during the period of analysis is plotted in Figure B1 in the supplemental data online.We acknowledge an anonymous referee for pointing out the existence of alternative estimates of excess mortality for Italy, such as that provided by Cerqua et al. (2021) using machine-learning techniques.Unfortunately, we are not able to fit our model specification (which is based on monthly data) with estimates computed over different time intervals.6.By analysing the first three complete genomes of SARS-CoV-2, Zehender et al. (2020) showed that the virus was present in Italy weeks before the first reported case.7. Between 25 May and 15 July 2020, the Italian Ministry of Health and ISTAT conducted an epidemiological investigation to estimate the percentage of the population that likely contracted the infection by sampling 150,000 individuals throughout the whole Italian territory.The results (based on 64,660 serological tests) show that the number of people who contracted the virus is equal to 2.5% of the population and, therefore, six times more than the official Covid-19 cases detected over the pandemic cycle (ISTAT, 2020).8. Several studies in the medical literature have shown that individuals living in highly polluted areas have a reduced capacity to react to respiratory diseases and pneumonias (Pope & Dockery, 2006).9.The ATECO 2007 classification is the Italian equivalent of the European NACE Rev. 2 classification.10.For mortality_growth data, see https://www.istat.it/it/archivio/240401; for intensive margin, extensive margin and internal mobility data, see https://www.istat.it/it/archivio/157423; for coastal, mountainous and ln density data, see https://www.istat.it/it/archivio/156224;for ln house m 2 pc, share over75 and share cohab over65 data, see http://ottomilacensus.istat.it/;for share males and hospital beds pc data, see http://dati.istat.it/;for pm10 data, see https://www.isprambiente.gov.it/it/pubblicazioni/stato-dellambiente; for district data, see https://www.istat.it/it/archivio/150320; and for remote working data, see http://dati-censimentoindustriaeservizi.istat.it/Index.aspxand Barbieri et al. (2021).11.Nevertheless, by relying on a spatial weights matrix constructed through Euclidean distances without neighbourless municipalities, the Moran's I index for spatial autocorrelation of our outcome of interest is relatively low (0.13).12.Given that intensive margin and extensive margin are time-invariant variables, they are omitted from column (5) because of collinearity with municipality fixed effects.13.For the sake of completeness, estimates of all the control variables (supported by a brief discussion) are reported in Table A3 in the supplemental data online, while Figure B5 online plots the coefficients of the most complete specification of Table 2 with their 95% and 99% confidence intervals.14.By way of example, this scenario would correspond to the situation in which the city of Bergamo, the provincial capital with both the highest intensive margin and mortality growth in March (as discussed in the introduction), would have commuting flows comparable with the provincial capital of Monza.15.With reference to our 7357 municipalities, the average numbers of fatalities occurring in Italy at the 'baseline' were 55,065 in March and 49,144 in April, while the total number of fatalities occurring during the same months in 2020 were 82,867 (+50.5%) and 67,805 (+38.0%),respectively.According to the reductions in mortality growth for these months computed by our back-of-the-envelope calculations, the mildest scenario would have led to mortality growth of 48.0% (50.5% -(50.5%*4.8%)) in March and 35.9% (38.0%-(38.0%*5.3%)) in April.Hence, the 'counterfactual' number of fatalities during the most critical part of the pandemic cycle would have been 81,521 in March and 66,808 in April. 16.Although a more up-to-date origin-destination matrix would undoubtedly be preferred, it is not available.At the municipality level, the most recent data on commuting (see http://dati-censimentipermanenti.istat.it/Index.aspx)provide information only on the aggregate 2019 outflows.17.Caselli et al. (2020) provide empirical evidence that this containment area significantly lowered individual mobility.18.For the same reason, the estimated coefficients of the extensive margin interacted with month dummies are not directly interpretable.19.For data, see https://www.istat.it/it/archivio/241341.For the sake of clarity, ISTAT data (which are based on the 2017 Frame Territoriale register) focus on workers in the industrial and service sectors.Workers employed in other economic activities, such as agriculture and public administration, are excluded from the registry because these sectors are outside the scope of business statistics.

Figure 2 .
Figure2.Timeline of the main events.Note: The timeline of the main events that occurred in Italy during the first wave of the pandemic is shown; hence, dates refer to 2020.Source: Authors' own elaboration.
Resilience to health shocks and the spatial extent of local labour markets: evidence from the Covid-19 outbreak in Italy 2507 REGIONAL STUDIES features of local economies, such as the spatial interactions of workers, and the epidemic.As we can see in Figures4 and 5, this area also shows high density of commuting flows in both the intensive and extensive components (for additional maps of other control variables, see FigureB4 online).The visual correlation, especially between excess mortality and the intensity of external mobility, is striking and suggests a specific role of commuting flows in placing more connected places at more severe epidemiological risks.

Figure 6 .
Figure 6.Reduction in mortality_growth, by month and scenario.Note: Estimates are based on back-of-the-envelope calculations for three scenarios in which the intensive margins would be 90%, 80% and 70% of those really observed in our data, respectively.
Note: All the specifications present ordinary least squares (OLS) estimates and include month and municipality fixed effects.Standard errors clustered at the local labour market (LLM) level are shown in parentheses.Significance: ***p,0.01,**p,0.05,*p,0.10.

Figure 8 .
Figure 8. share_inactive_commuters, by municipality.Note: Readers of the print article can view the figures in colour online at https://doi.org/10.1080/00343404.2022.2035708Source: Authors' own elaboration.

Table 1 .
Commuting indices and mortality growth (part 1).Resilience to health shocks and the spatial extent of local labour markets: evidence from the Covid-19 outbreak in Italy 2509 Note: All the specifications present ordinary least squares (OLS) estimates and include month, region and municipality fixed effects as indicated.Standard errors clustered at the local labour market (LLM) level are shown in parentheses.Significance: ***p,0.01,**p,0.05,*p,0.10.
Note: All the specifications present ordinary least squares (OLS) estimates and include month and municipality fixed effects.Standard errors clustered at the local labour market (LLM) level are shown in parentheses.Significance: ***p,0.01,**p,0.05,*p,0.10.
Note: All the specifications present ordinary least squares (OLS) estimates and include month and municipality fixed effects.Standard errors clustered at the local labour market (LLM) level are shown in parentheses.Significance: ***p,0.01,**p,0.05,*p,0.10.2512MattiaBorsati et al.
Note: All the specifications present ordinary least squares (OLS) estimates and include month and municipality fixed effects.Standard errors clustered at the local labour market (LLM) level are shown in parentheses.Significance: ***p,0.01,**p,0.05,*p,0.10.2516MattiaBorsati et al.