Local inequalities of the COVID-19 crisis

Abstract

This paper assesses the pandemic's impact on Italian local economies with the newly developed machine learning control method for counterfactual building. Our results document that the economic effects of the COVID-19 shock vary dramatically across the Italian territory and are spatially uncorrelated with the epidemiological pattern of the first wave. The largest employment losses occurred in areas characterized by high exposure to social aggregation risks and pre-existing labor market fragilities. Lastly, we show that the hotspots of the COVID-19 crisis do not overlap with those of the Great Recession. These findings call for a place-based policy response to address the uneven economic geography of the pandemic.

1. Introduction

With over 132,200 deaths and more than 4,780,000 cases (as of November 4, 2021), Italy ranks among the worst-hit countries by COVID-19.1 The Italian government was the first in Europe to declare, on March 9, 2020, an unprecedented national lockdown that paralyzed the country. From March 25, productive activities were shut down, except for those deemed ‘essential’ for the functioning of the country's economic system. On May 4, lockdown rules started to be lifted, and, from June 15, almost all economic activities were finally allowed to re-open, albeit under strict safety protocols. The suspension of restrictive measures continued throughout the summer until the impressive resurgence of the contagion in the fall of 2020 forced the authorities to issue new social distancing policies, including the reintroduction of restrictive measures targeting economic activities.

The Italian government tried to attenuate the impacts of such disruptive events via the adoption of several emergency measures and fiscal packages.2 In order to increase workers’ protection, the government also issued an ad hoc Decree-Law on March 17, 2020, which introduced two exceptional labor market policies: a special COVID-19 short-time work retroactive compensation scheme and a freezing of layoffs, which have later been repeatedly extended and are still partially in place at the time of writing.

Despite the implementation of a wide range of policy interventions, Italy's GDP contracted by 8.9% in 2020.3 Besides, the Bank of Italy reports, for 2020, a reduction of 11% in the number of hours worked and a decrease of 2.1% in the number of persons employed.

Remarkably, ex-post evaluations of the spatial distribution of the economic effects of the COVID-19 emergency are still missing. Such a vacuum was hardly surprising initially, as real-time microdata are scarce, but as the epidemiological impacts finally attenuate, there is a strong case for a comprehensive analysis of the economic geography of the pandemic crisis. On top of data scarcity, rigorous impact evaluation is challenging for econometric issues: the COVID-19 shock virtually left no part of the Italian territory unaffected, and the national lockdown involved the entire country. In econometric jargon, it is hard to find a control group because the treatment affected all units simultaneously or with short lags. Therefore, while in countries and areas where no total lockdowns were implemented one can exploit staggered or heterogenous policy responses to generate a counterfactual scenario,4 standard evaluation techniques, such as difference-in-difference or the synthetic control method (SCM), are not applicable to the Italian context. This is probably the reason why, although the literature on the pandemic is flourishing (Adams-Prassl et al., 2020; Baker et al., 2020; Bartik et al., 2020; Benedetti et al., 2020; Bick and Blandin, 2020; Bloom et al., 2021; Blundell et al., 2020; Cajner et al., 2020; Carvalho et al., 2020; Chetty et al., 2020; Chudik et al., 2021; Forsythe et al., 2020; Gourinchas et al., 2020; Von Gaudecker et al., 2020), the Italian case still remains relatively unexplored.

Among the few exceptions, Ascani et al. (2021) provide evidence of a close relationship between COVID-19 disease patterns and local economies' characteristics. Through a linear probability model, Casarico and Lattanzio (2020) find that workers already in disadvantaged conditions before the shock (young, low-skilled, and seasonal workers) have substantially higher risks of losing their jobs. Carta and De Philippis (2021) simulate the pandemic's effects on the labor income distribution of Italian households in the first two quarters of 2020, and assess the effects of government policies to reduce labor income losses. Their estimates suggest that the social insurance policies were effective in preventing a significant increase in income inequality, and that the pandemic increased labor market income inequality. These studies underline important local and sectoral components of the impacts of the shock in Italy. Indeed, in Europe as elsewhere, the current crisis is undoubtedly a regional one, so regional perspectives are essential to understand the unequal impacts of the pandemic (Bailey et al., 2020).

We quantify and map the heterogeneous impacts of the COVID-19 crisis in 2020 on labor and firm outcomes for all 610 Italian local labor markets (LLMs),5 investigate the main territorial features of such unevenness, and compare the magnitude and spatial distribution of these impacts with those of the 2008–2009 Great Recession. To this end, we leverage quarterly LLMs data, collected from the Business Register kept by the Union of the Italian Chambers of Commerce, combined with a counterfactual application of machine learning (ML) techniques, namely the newly developed machine learning approach for counterfactual building (Varian, 2016; Burlig et al., 2020; Cerqua et al., 2021) which we call the Machine Learning Control Method (MLCM). The MLCM draws on the predictive ability of ML algorithms to generate a no-COVID counterfactual scenario (i.e. a ‘business-as-usual’ scenario) in such a peculiar econometric setting. The use of the MLCM is made possible by constructing a comprehensive time-series cross-sectional database on LLMs.

Thanks to this counterfactual approach, we document several key findings. First, by the end of 2020, the shock has caused a steep decrease in firm entries and abnormal drops in employment and firm exits at the national level. Second, the effects have been markedly heterogeneous across the Italian territory, but not in the way one would expect. Italy is known for its historical and persistent North-South divide, with the South being poorer and characterized by high unemployment rates and vulnerable labor markets. Besides, the micro-level evidence mentioned above shows that those who have been more affected by the pandemic recession are temporary and less-educated workers. Such workers are notably more employed in Southern Italian regions. Given these aspects, one could a priori conclude that the South endured the most severe effects of the crisis. We do not find that this is the case: the largest employment impacts are more concentrated in the Centre-North, albeit they did not occur in the areas that experienced the highest death toll due to COVID-19. Third, the level of within-region heterogeneity across local economies is even more surprising than these macro-area patterns. We thus use a regression tree to identify the factors that matter the most in explaining the local heterogeneity of employment changes, and find that the features more significantly associated with severe employment losses are a high share of workers in sectors more exposed to social aggregation risks and pre-existing labor market fragilities.

From a methodological perspective, we also provide a number of contributions. We show that causal ML tools can be employed to credibly estimate treatment effects, even in settings without available control groups. The complex algorithm we employ, the random forest, outperforms a set of other common approaches in the predictive ‘race’ to build a plausible counterfactual scenario. In comparison, a traditional regression approach fares far worse, as it is sensitive to extreme values and prone to overfitting. These issues do not affect ML techniques thanks to cross-validation and their inner ability to detect recurrent patterns as well as complex interactions and non-linearities in the data. At the same time, we also find that the simplest possible approach, a before-after analysis using the last pre-shock period's figures as a counterfactual, delivers more accurate predictions than data-requiring predictive models such as Ordinary Least Squares (OLS) and a regression tree. This lends credibility to several works published at the onset of the pandemic based on pre-/post-comparative analyses of economic outcomes in the short-run (see, among many, Baker et al. (2020); Chen et al. (2021); Sheridan et al. (2020)), and suggests that this straightforward methodology can be an acceptable empirical solution during ‘emergency’ situations in which issues such as time-consuming data collection or econometric challenges prevent the immediate use of more sophisticated techniques. Note, however, that such an intuitive methodology is likely to be reasonably reliable only for short-term estimates. Finally, the association analysis we carry out to uncover the most relevant territorial predictors of impacts reveals that, even in this case, ML methods are more fit for the purpose compared to traditional approaches. While a plain regression model can detect the importance of the predominant predictors, it is unable to capture the relevant interactions and non-linearities among these features that are instead detected by a fully non-parametric regression tree. For instance, according to OLS, all the areas with fragile labor markets (proxied by a high share of temporary contracts) experienced negative effects. In contrast, our tree suggests that conditional on threshold values of other key features, this variable can be either negatively or positively associated with the employment losses, depending on the way in which it combines with the other features. In sum, the replacement of traditional methods with data-driven techniques leads to more realistic counterfactuals and qualitatively different conclusions and policy implications.

Lastly, we quantify and map the employment losses of the Great Recession and show that the impact of the pandemic is larger in absolute terms and that the spatial distribution of the two crises does not overlap, suggesting that the territorial patterns of the COVID-19 recession are different from those of the previous economic downturns. Our sectorial analysis on COVID-19 adds that the pandemic is also dissimilar from the Great Recession in that it damaged the services sector more than the manufacturing one. In the context of the comparative literature on economic crises, our results support the conclusions of Cajner et al. (2020) and Coibion et al. (2020), who find a more pronounced employment decline during the coronavirus crisis than the Great Recession in the US. Furthermore, Cajner et al. (2020) argue that, while the manufacturing sector contracted the most during previous recessions, the COVID-19 shock is hitting different industries, including leisure, hospitality, and tourism, which is also confirmed by our analysis. Finally, our results are in line with those of Chetty et al. (2020), who document sharp differences in the geographic patterns of employment changes between the COVID crisis and Great Recession in the US, and point out that job losses during the pandemic tend to be more concentrated in affluent areas. In conclusion, we provide empirical support in favor of the view that the pandemic crisis is different from a typical recession.

2. Data

Our primary dependent variable is the log of overall employment in the private non-financial sector. In addition, we also split employment between manufacturing and services, and investigate the impact of COVID-19 on the number of new business registrations (births) and cessations of trading (deaths).6 All these variables are generated using restricted-access data collected from the Business Register kept by the Union of the Italian Chambers of Commerce (Unioncamere). The Business Register is based on administrative data on the Italian companies gathered by the provincial Chambers of Commerce. It contains information on the registration data of the universe of Italian private non-financial sector firms. The Business Register quarterly data on local employment have been made available by the Italian Social Security Institute (INPS) since the third trimester of 2014.

To estimate the impact of COVID-19 on each LLM, we build a comprehensive, balanced panel of all 610 Italian LLMs from 2016 Q3 to 2020 Q4 and employ the random forest algorithm described in Section 3.7 The counterfactual is estimated by controlling for the industrial structure of each LLM. To this end, we exploit the classification by the Italian National Institute of Statistics (Istat), which splits the Italian LLMs into four classes: without specialization, non-manufacturing, made in Italy,8 and other manufacturing. Furthermore, in light of the expected plunge in tourism-related employment, we split the non-manufacturing class into touristic and non-touristic. We then control for LLM size, geographical dummies (North-East, North-West, Centre and South), log of population, population density, unemployment rate, activity rate, yearly and quarterly fixed effects, share of foreign population and trends in employment, business births, and business deaths. For each of the latter three variables, we control for two lags of the same quarter, the lags of the four preceding quarters, all the available yearly lags, the last pre-treatment lag of the outcome variable, and the same lags of the other outcome variables.

In the second phase of the empirical analysis, the association analysis uses the estimated COVID-19 impact on employment for all LLMs as the outcome of interest to uncover its primary predictors. For this analysis, we collected variables plausibly correlated with the employment change due to COVID-19. We use the dependency ratio to control for the population structure and its implications for the productive part of the population. As a measure of the spread of COVID-19, we use the excess mortality estimates provided by Cerqua et al. (2021), updated to September 30, 2020.9 We also employ two variables which capture the criticality of the tasks performed by employees, the possibility of exposure to the virus and physical proximity to the workplace, all highlighted as relevant factors in the literature (see Barbieri et al., 2021): the share of jobs having a high risk of social aggregation and the share of jobs having a high ‘integrated’ risk. These variables proxy for the demand-side changes due to peoples' immediate response to the pandemic and are generated on the basis of the work conducted by an ad hoc task force,10 which linked a level of social aggregation to each economic sector (2-digit NACE Rev.2 classification) and integrated risks from low to high. Activities at high integrated risk are those associated with the risk of coming into contact with sources of contagion at work, especially those connected to work processes (e.g. human health services, sewerage, public administration and defense), while activities at high risk of social aggregation are those that involve contact with other subjects in addition to the company's workers (e.g. catering, entertainment, hospitality).

As the geography of industries highly exposed to the ‘COVID-19 shock’ is heterogeneous (Krueger et al., 2020), we create a variable that incorporates the predicted supply-side sectoral shocks to each LLM. Specifically, we generate the share of jobs in suspended economic activities from March to May 2020.11 In addition, we build the share of temporary contracts as a metric for temporary jobs' local relative importance.12

Other economic variables included in this phase of the analysis are per capita income, unemployment rate, the share of innovative start-ups as a proxy for local innovation, and a measure of economic fragility, i.e. the share of firms having employees in Cassa Integrazione Guadagni Straordinaria (CIGS), namely the most utilized Italian short-time work program providing subsidies for temporary reductions in the number of hours worked.13 We also add two variables that consider the densities of health care personnel and hospitals: i) the number of hospital beds per 1,000 inhabitants, and ii) the share of workers employed in the NACE 2-digit sectors ‘human health activities’ and ‘residential care activities’.

Lastly, as mobility is one of the critical aspects linked to the epidemiological spread of COVID-19, we take this into account by using three variables:

• -the number of road accidents per 10,000 inhabitants;
• -the share of population living in peripheral areas;
• -the index of relational intensity (IIRFL) within the local labor market. The higher the IIRFL, the greater the inter-municipal turbulence in terms of flows.

In the Appendix, Table A1 includes a more detailed description of all the variables, while Table A2 provides descriptive statistics. The availability of these indicators will allow us to identify the LLM characteristics that matter the most in explaining the treatment effects’ heterogeneity.

3. Methods

Our empirical exercise consists of three tasks – a counterfactual analysis, an association analysis, and a comparative analysis. For all three steps, we harness ML's statistical power: the random forest algorithm for the counterfactual and comparative analyses, and a regression tree for the association analysis. Below, we separately discuss the three empirical analyses.

3.1. Counterfactual analysis: the machine learning control method

To tackle the econometric challenges related to the pandemic shock's pervasive nature and establish causality, we draw on the newly developed MLCM to generate a counterfactual scenario in which the COVID-19 crisis never hit Italy. In other words, we employ the MLCM to address the fundamental problem of causal inference, i.e. the impossibility of observing the potential outcome in the no-treatment scenario, a curse that affects all LLMs.

Although ML algorithms primarily deal with out-of-sample predictions or ‘prediction policy problems’ (see Kleinberg et al., 2015), more recently, they have been combined with causal inference approaches (Athey and Imbens, 2016; Athey et al., 2021; Athey et al., 2019; Belloni et al., 2017; Hofman et al., 2021; Varian, 2014, 2016; Wager and Athey, 2018). Varian (2014, 2016) was among the first to note that counterfactual building is essentially a predictive task, which is exactly the task at which ML excels. In a panel or time-series setting, he noted that one could exploit pre-treatment observations to generate an artificial control group that acts as a counterfactual in the no-treatment, ‘business-as-usual’ scenario. This way, one could readily retrieve treatment effects as the difference between the observed outcome and the ML-generated potential outcome. Varian called this straightforward counterfactual method the ‘train-test-treat-compare’ process. This process is similar to the SCM developed by Abadie et al. (2010), with the key difference that it does not require the availability of untreated units, as it draws on pre-treatment information to generate a credible estimate of the ‘outcome for the treated if not treated’.

Early empirical applications of this intuitive methodology for counterfactual building have recently appeared (Abrell et al., 2019; Benatia, 2020; Benatia and de Villemeur, 2019; Bijnens et al., 2019; Burlig et al., 2020; Cerqua et al., 2021; Souza, 2019). Except Burlig et al. (2020) and Souza (2019), all the other studies cannot rely on an original control group in their research design because they only observe treated units in settings with simultaneous treatment, just as in our case.

Benatia (2020) and Cerqua et al. (2021) are the most closely related to this study because they both investigate the causal effects of the COVID-19 crisis. Benatia (2020) applies a neural network model to study the impact of containment measures on the demand reduction in New York's electricity markets; Cerqua et al. (2021) employ three different ML routines (LASSO, random forest, and stochastic gradient boosting) to derive municipality-level excess mortality estimates during the first wave of the COVID-19 pandemic in Italy.

In the spirit of this nascent evaluation approach, we apply the MLCM to pursue our causal inference analysis of COVID-19 local economic impacts in Italy. Our artificial control group comes from an ML predictive model developed to forecast a post-treatment counterfactual for each LLM. In this way, under the crucial assumption of stable trends in the absence of the shock, we can assess the LLM-specific causal impact of the exogenous shock by comparing the observed post-shock trajectory with the most credible trajectory the LLM unit would have followed in a no-shock scenario. A critical requirement for the validity of this approach is that the predictive ML model must not include predictors that may be affected by the treatment (Varian, 2016). We avert this issue by employing only pre-2020 features in our counterfactual building. Finally, the use of the MLCM is made possible from the construction of a comprehensive time-series cross-sectional database on LLMs (see Section 2).

We apply a powerful and popular ML method: the random forest, which has been defined the most successful general-purpose algorithm in modern times (Howard and Bowles, 2012; Varian, 2014). The random forest is a fully non-linear technique based on the aggregation of many decision trees. In particular, random forest builds many trees (1,000, in our case) based on bootstrapped training samples and, at each split of a tree, uses only a random subset of the predictors as split candidates, thus introducing a double layer of decorrelation of the trees from one another (Hastie et al., 2009).

Drawing from the routine established by Cerqua et al. (2021), our counterfactual analysis is based, for each outcome variable, on the following 7-step methodological sequence:

• 1)We randomly split the pre-2019 quarterly dataset into a training sample, made up of 80% of the LLMs, and a disjoint test set, consisting of the remaining 20%14 ;
• 2)We train our random forest algorithm on the training set and perform a 10-fold cross-validation to select the best-performing tuning hyperparameter15 ;
• 3)We test the out-of-sample predictive performance on the corresponding pre-2019 testing sample16 ;
• 4)We test model accuracy on the entire 2019 sample and compare its predictive performance with those of a battery of other alternative approaches, namely a before-after analysis, which has become a common and intuitive metric to gauge the magnitude of the pandemic's impact, a traditional OLS model, and a simpler ML algorithm, the regression tree;
• 5)We repeat the same routine on the entire pre-2020 dataset (which, having four more quarters of observations, enhances random forest's predictive power compared to the one on which we develop the model) and finally predict, for the 2020 sample, employment levels, business births, and business deaths in a no-COVID (‘business-as-usual’) scenario;
• 6)We derive individual treatment effects for all LLMs as the difference between the observed 2020 outcomes and the ML-generated potential outcomes;
• 7)We map the individual treatment effects of the LLM-level economic impacts of COVID-19.

The critical assumption behind this MLCM routine is that the difference between our observed and counterfactual economic outcomes is the causal impact of the COVID-19 shock. We deem it plausible given the massive disruption to the economy brought about by the sudden unexpected arrival of the pandemic. Lastly, please note that, by ‘COVID-19 shock’, we mean the economic shock, i.e. we refer not only to the epidemiological spread of the virus per se, but also to the related behavioral changes and, above all, to the national lockdown and the other non-pharmaceutical interventions that were adopted to contain the health emergency. This implies that, via our counterfactual approach, we capture the total impact of the pandemic on each LLM.

3.2. Association analysis: the employment change regression tree

To estimate the relationship between the estimated employment outcomes and potentially relevant covariates linked to economic, mobility, and pandemic-related LLM features, we harness the efficacy and power of another well-known ML algorithm: the regression tree.

First and foremost, bear in mind that here we abandon the causal inference setting to go back to the original ML habitat, i.e. the realm of pure prediction. What we want to do in this analysis is to get an idea of the factors which matter most in predicting the heterogeneous local economic impact of the pandemic.

Regression trees are an ideal tool to fulfill this purpose for two reasons: i) differently from complex, black-box ML methods such as random forest, regression trees allow an intuitive understanding of the mechanism through which the outcome variable of interest is linked to its most relevant predictors, thus producing an easy-to-interpret output which can be particularly valuable when the model must be shared to support public decision-making (Andini et al., 2018; Lantz, 2019); ii) unlike traditional approaches, regression trees are extremely flexible methods that can easily capture, in the sequence of splits, the entire range of potential non-linearities and interactions between the features, without imposing any parametric functional form to the underlying data-generating process.

From a technical point of view, this ML algorithm divides the data into progressively smaller subsets to identify significant patterns that are then used to predict the continuous output. Compared to standard regression tree analyses, two necessary clarifications are in order. First, we do not divide our sample into a training and testing set. The reason is straightforward: instead of testing for the out-of-sample accuracy of our regression tree model, we want to investigate the main predictors of our outcome variable, i.e. the estimated treatment effect for employment change in 2020 Q4, on the full sample of Italy's LLMs. In other words, we are not interested in out-of-sample prediction, but only in the associations between the features and the outcome. Second, and related, we do not apply cross-validation to select the hyperparameter of the regression tree method (named ‘complexity parameter’, cp) and adopt the commonly adopted default value of 0.01.

Therefore, we run a basic regression tree model of the employment effects to uncover the most relevant predictors of treatment effect unevenness at the local level. Notably, the associations emerging from the regression tree should not be interpreted in a causal sense, but rather as a way to uncover significant correlations between the most important features and the outcome variable of interest.

Lastly, for the sake of comparison with a traditional regression approach, we also run an OLS-based association analysis using as outcome variable the employment treatment effects estimated using the standard before-after predictive method. This additional check, which draws on micro literature works using a linear model to retrieve associations with the estimated changes (Casarico and Lattanzio, 2020), allows us to show that more advanced and fully nonparametric statistical techniques do not only lead to more accurate impact estimates, but also to a more comprehensive assessment of territorial-level associations.

3.3. Comparative analysis: the COVID-19 crisis vs the Great Recession

Is the pandemic crisis different from previous recessions? To answer this question, we compare the estimated employment losses due to the COVID-19 shock with those of the 2008–2009 Great Recession. Specifically, we collect annual employment data from Istat for the period 2006–2009,17 as well as feature data as similar as possible to those employed in the main counterfactual analysis, and apply the same machine learning approach to generate local-level employment losses during the period 2008–2009. Due to data availability constraints, we use 2006–2007 data to generate a counterfactual prediction of employment levels in each LLM at the end of 2009. The procedure is exactly the same as that outlined in subsection 3.1. The most substantial difference, other than the use of annual in place of quarterly LLM data and the more limited amount of information available, is that here we estimate the impacts of the Great Recession across two years, 2008 and 2009, whereas impacts for the pandemic refer to a single year (2020). In fact, the repercussions of the global crisis on employment in Italy started to appear during the last months of 2008, meaning that data from this year cannot be employed to train the ML model. For this reason, all the feature data used for the 2009 employment prediction are lagged of two years, i.e. they refer to 2007, so that no post-treatment information is provided to the algorithm.18 Table A1, Table A2 in the Appendix provide further details on the data used for this analysis. Finally, after estimating and mapping employment losses during the Great Recession, we compare their spatial distribution with that of the employment change registered during the COVID-19 crisis.

4. Counterfactual analysis

We begin by reporting in Table 1 the random forest technique's predictive performance compared to a battery of alternative predictive methods. First, we compare it with the intuitive before-after method often adopted to gauge the magnitude of the COVID-19 shock. The before-after analysis estimates the impact of COVID-19 as the difference between the trend of a given outcome (in this case, employment) in 2020 (after the pandemic's arrival) and the pre-pandemic average figures of the past year(s). The underlying assumption is that, without the pandemic, the number of employees would have remained constant (zero growth). Examples of this intuitive approach to gauge the magnitude of the COVID-19 impact on employment and firm outcomes in the Italian context can be found in Casarico and Lattanzio (2020), Giacomelli et al. (2021), and Viviano (2020). Next, we also assess the predictive power of a more traditional fixed-effects OLS model (fitted on the full 2016–2018 sample) and a simpler machine learning routine, the regression tree (here used following the same routine implemented for the random forest). In this way, we can get a full picture of the predictive ability of our methodology compared to other approaches.

Table 1.

Predictive performances for 2019 (log) overall employment levels.

Panel A – Performance on all LLMs
Predictive method	MSE	MEDSE
Corresponding quarter – last year (2018)	0.00101	0.00040
Corresponding quarter – 3-year average (2016–2018)	0.00320	0.00214
OLS	0.08510	0.04799
Regression tree	0.05669	0.02907
Random forest	0.00082	0.00026
Panel B – Performance by population size
≤ 50,000 inhabitants
Corresponding quarter – last year (2018)	0.00120	0.00042
Corresponding quarter – 3-year average (2016–2018)	0.00325	0.00189
OLS	0.08067	0.03839
Regression tree	0.04337	0.02366
Random forest	0.00107	0.00034
Between 50,000 and 200,000 inhabitants
Corresponding quarter – last year (2018)	0.00075	0.00037
Corresponding quarter – 3-year average (2016–2018)	0.00312	0.00224
OLS	0.06415	0.05289
Regression tree	0.04382	0.02943
Random forest	0.00052	0.00021
≥ 200,000 inhabitants
Corresponding quarter – last year (2018)	0.00090	0.00042
Corresponding quarter – 3-year average (2016–2018)	0.00324	0.00274
OLS	0.20479	0.18258
Regression tree	0.19768	0.08120
Random forest	0.00056	0.00020

Notes: Estimates on the 2019 sample. MSE stands for Mean Squared Error; MEDSE for Median Squared Error.

To assess the predictive performance of the various methods, we use two different measures of the typical prediction error, i.e. the Mean Squared Error (MSE) and Median Squared Error (MEDSE). The figures reported in Table 1 reveal that random forest predictions substantially outperform all the other methodologies in the out-of-sample predictive test on the 2019 sample. Let us first focus on the mean predictive performances across the entire 2019 sample (Panel A). Using MSE as the reference metric, the predictive gain of the random forest performance is of about 19% compared to last year's before-after figures, and of almost 75% compared to the three-year (2016–2018) average of the outcome variable. MEDSE performances are even more dramatically unbalanced in favor of the random forest. The random forest also strongly outperforms an OLS model run on the full 2016–2018 sample, with a reduction of the 2019 MSE larger than 90% compared to this traditional approach, which also performs far worse than both the before-after approaches and seems to be very sensitive to extreme values, in line with the results of Burlig et al. (2020). Finally, the regression tree predictive performance is in between, with MSE and MEDSE values lower than OLS but definitely larger than the before-after metrics and the random forest. This suggests that one needs to apply more powerful ensemble methods such as the random forest in order to maximize out-of-sample performance and unleash the full predictive potential of ML for counterfactual building. In addition, the relatively good performance of the before-after approach using the last pre-shock corresponding quarter's average as counterfactual suggests that this straightforward methodology can be an acceptable empirical solution for short-term assessments during ‘emergency’ situations.

Overall, the random forest outperforms all the other methods. This key insight is not just valid on average, but also for each of the three population subsamples in which we split our full sample (Panel B). The subsample comparison reveals that the random forest predictive gain becomes especially sizable when moving from less-populated LLMs (in which employment annual growth rates can vary greatly due to small employment levels) to bigger local economies, where the gap between this complex algorithm and all the other approaches becomes large. In sum, this test demonstrates that complex data-driven methodologies can lead to consistently more accurate predictions of potential outcomes in a given, ‘ordinary’ year.19

Having established that the random forest exploits past information to predict future outcomes much better than standard methods, we take a quick look at the aggregate treatment effects of the coronavirus crisis (weighted by LLM population) for the employment outcome. By the end of 2020, the pandemic has entailed a 3% decrease in overall employment in Italy, compared to what employment levels would have been had the pandemic never reached the country. This national-level estimate is larger than the 2.1% annual reduction in the number of persons employed in 2020 estimated by the Bank of Italy.20 The reason for the discrepancy is that this institution measures average employment change with respect to 2019, which does not account for the potential growth in employment that could have been registered in 2020 in the absence of the shock.

As we mainly focus on the local heterogeneous impact of COVID-19, in the following sections, we first map LLM-specific treatment effects and then gauge the heterogeneity in COVID-19 impacts across local economies.

4.1. Employment

Fig. 1 shows the map of employment change in 2020 Q4 ​at the LLM level. The degree of treatment effect heterogeneity is striking. At first glance, it is evident that the crisis hit more severely local economies located in the Centre-North (−3.7% vs −1.8% in the South). Nevertheless, many LLMs in Southern regions and in the islands have also been sharply affected. We also remark that, in the so-called Mezzogiorno, the weight of the informal economy is substantial, and we cannot capture those losses with official data. The North-South gap in employment losses might thus be smaller than it is apparent from this figure. Despite these caveats, the map looks different from what one would expect based on micro-level evidence (Casarico and Lattanzio, 2020; Carta and De Philippis, 2021), because temporary and less-educated workers with unstable and less protected jobs are more commonly employed in Southern regions, less developed than the richer North. Our finding that losses are more substantial in the affluent areas of the country is in line with the evidence provided by Chetty et al. (2020) for the US.

Fig. 1.

Employment change in 2020.

More generally, some local economies have been hit much harder than others, with impacts ranging from drops larger than 2.5% in most LLMs of the entire Centre-North, Abruzzi, Basilicata and the costs of Sardinia, to small decreases or even mildly positive effects in parts of Lazio, Campania, Calabria, Sicily, and central Sardinia. What is even more relevant than the macro-area patterns is, in our view, the within-region degree of heterogeneity, which shows how, in virtually all Italian regions, some LLMs fared much better than others despite being geographically close and often contiguous. Such local disparities would be averaged away if using regional data. Besides, it is worth noting that the documented employment impacts are net of the Italian government's protective measures. This means that without these protective measures (the layoff freeze and CIGS extensions in particular), local impacts would have been more sizeable.

Where does such a striking heterogeneity come from? We first inspect the geographic distribution of the employment and epidemiological outcomes engendered by COVID-19. Fig. 2 presents a visual comparison between the economic vs epidemiological effects of COVID-19 in Italy. Looking at the maps, the geographic distribution of impacts does not mirror the COVID-19 epidemiological spread during the first wave, which is proxied by excess mortality estimates from February 21, 2020, to September 30, 2020. To estimate the spatial correlation between these outcomes, we measure their overall spatial relationship across all LLMs using the bivariate Moran's I. This index ranges from −1 (perfect negative spatial correlation) to 1 (perfect positive spatial correlation), and we obtained a Moran's I coefficient close to 0 (−0.092), which suggests a lack of significant spatial correlation between employment and epidemiological impacts.

Fig. 2.

Economic versus epidemiological impacts of the COVID-19 pandemic across Italy. Notes: Excess mortality data are from the municipality-level estimates of Cerqua et al. (2021), aggregated at the LLM level.

4.2. Employment by sector

If LLMs’ COVID-19 death toll is not a primary driver, where does the heterogeneous impact on overall employment originate? Sectoral specialization of LLMs is part of the answer. As shown in the maps of employment change in manufacturing and services, depicted in Fig. 3 , the tertiary sector was more severely affected than the manufacturing one.21

Fig. 3.

Employment change in 2020 by sector.

This is not unexpected, as workplace closures primarily affected economic activities in the tertiary sector. At the same time, a large share of manufacturing firms could avert the shutdown thanks to being comprised in the list of ‘essential activities’ that the government decided to keep open to guarantee the basic functioning of Italy's economic system. The tertiary sector is also notably the one with the highest prevalence of temporary jobs and seasonal workers, which could only marginally benefit from the layoff freeze measure. Given these facets, it comes as no surprise that employment losses primarily affected LLMs specialized in services, and tourism in particular. For instance, the case of Tuscany, home to many cities renowned for their art and culture, is particularly emblematic of the losses suffered in the tourism sector.

4.3. Business demography

We then look at how COVID-19 affected business demography outcomes. At the national level, by the end of 2020, the crisis determined a 19.5% decrease in business births and a 7.5% decrease in business deaths. Fig. 4 disaggregates these country-level estimates and maps the cumulative impact of COVID-19 for business births change (i.e. firm entries) and business deaths change (firm exits) in 2020.

Fig. 4.

Business births and deaths changes in 2020.

The impact on business births is particularly acute and, with almost no exception, involves the entire national territory. This anomalous plunge happened despite the so-called Decreto Rilancio (May 14, 2020), which included a set of protective measures intended to support investments in start-ups. By contrast, the impact on firm exits is more polarized and geographically dispersed, with several LLMs experiencing substantial reductions in cessations of trading, whereas others saw a significant increase in firm exits.

The generalized drop in the number of newly-born firms across the country is particularly troublesome because start-ups and young firms are usually the most innovative ones, thus pointing to dire forecasts about the potentially long-lasting effects of the fall in business births in terms of aggregate productivity growth. Moreover, this lost generation of firms creates a persistent dent in overall employment as subsequent years will be characterized by a lower number of firms (Sedláček, 2020). This is all the more worrying in Italy, a country whose economic dynamism – its ability and willingness to allocate resources efficiently – has been steadily declining in the last quarter of a century (Rossi and Mingardi, 2020). The results on firm closures, instead, should be interpreted with caution, as many firm exits could have been temporarily frozen by the supportive measures adopted by the government.

5. Association analysis

The counterfactual analysis revealed a substantial heterogeneity of the pandemic's economic effects. Such heterogeneity is partly driven by the territorial sectoral specialization. Nevertheless, we want to go further than this and understand the factors that matter the most in generating such a fragmented landscape. Therefore, we use a regression tree to examine the main predictors of employment losses.

Fig. 5 illustrates the regression tree of the LLM-specific overall employment treatment effects. The tree reveals interesting patterns. First, only four variables generate the tree: the share of jobs having a high aggregation risk, the share of temporary contracts, the unemployment rate, and per capita income. Second, the most severely affected LLMs are those in which there is a substantial share of jobs at a high risk of social aggregation and a high share of temporary contracts. For instance, the tree predicts that LLMs with a share of jobs having a risk of aggregation equal to or higher than 43% and a share of temporary contracts equal to or higher than 29%, will experience a 21% drop in employment.22 Third, excess deaths are not picked up by the tree, confirming that there is no discernible association between employment outcomes and the spread of COVID-19.

Fig. 5.

Regression tree on 2020 employment change.

Exposure to high aggregation and proximity risk seems to be a primary discriminant of impacts across LLMs with different shares or ‘workers at risk’ (Barbieri et al., 2021). In turn, the relevance of the labor market attributes in generating the regression tree provides empirical support for the above discussion on the unequal exposure of different workers' categories and types of contracts in the face of the emergency, in line with the heterogeneous findings of Casarico and Lattanzio (2020) and Carta and De Philippis (2021) for Italy and Blundell et al. (2020) for the UK. This analysis also suggests that emergency measures were by design effective only for specific categories of workers and types of contracts. More fragile categories (think of seasonal workers and occasional jobs) proved to be more vulnerable to the crisis's labor market consequences. We remark, however, that only LLMs characterized by economic sectors having both high social aggregation risks and fragile labor markets endured the sharpest drops in overall employment levels. It is their combination that matters. In fact, the only slightly positive impacts predicted by the tree occur in poorer LLMs with low risks of social aggregation, high unemployment rates, and high shares of temporary contracts. These territories, mainly corresponding to remote, inner, and rural areas of Calabria and Sicily, might have benefitted from the protective measures adopted by the government without paying a high price due to their structural characteristics that made them less exposed to the crisis.23 Finally, local economies specialized in sectors with a high aggregation risk, lower share of temporary contracts, but higher per capita income also experienced more severe losses. Such insights add nuance to the micro-level evidence cited above and complement the key insight from Fig. 1 that the largest employment changes have been observed in the Centre-North.

As a sensitivity check, we replace our variables on the share of jobs having a high risk of social aggregation, the share of jobs having a high ‘integrated’ risk, and the share of jobs in suspended economic activities with alternative measures of the expected sectoral shocks: the observed demand- and supply-side changes. These two variables weigh the expected supply and demand sectoral shocks reported in del Rio-Chanona et al. (2020) by each LLM's sectoral composition (see Table A1, Table A2 in the Appendix for definitions and descriptive statistics), and are highly correlated with the three replaced features.24 The corresponding regression tree is presented in Fig. A1. The tree confirms that the largest employment losses originate from severe demand shocks experienced in areas characterized by high risks of social aggregation and more vulnerable labor markets.

Lastly, Table A3 in the Appendix reports the results of the OLS-based association analysis run on the 1-year before-after treatment effects. As one can see, while a traditional model can detect the importance of the predominant predictors, as the coefficients of the key variables (social aggregation risks, the share of temporary contracts, per capita income, and the unemployment rate) are all strongly significant and negative (except for the unemployment rate, whose sign is positive), it is intrinsically unable to capture the relevant interactions among these features that are instead promptly detected by the tree. Consider, for instance, the role played by the share of temporary contracts. According to OLS, it is negatively associated with the treatment effects. But the regression tree suggests that, conditional on threshold values of other key features (risks of social aggregation and unemployment rate), this variable can be either negatively or positively associated with the employment losses. The switch depends on the way in which it combines with the other features. Based on the OLS association analysis, instead, one would conclude that all areas with a high share of temporary contracts experienced negative effects. Instead, the areas with the highest values of this variable, which are more isolated and with higher unemployment rates, saw slight increases in employment levels. The reason for this contradiction is that, unlike the regression tree, which is a fully nonparametric and flexible statistical learning method designed to automatically capture the interactions and non-linearities among features with relevant predictive power, OLS cannot capture the important interactions and non-linearities at play due to its parametric functional form restrictions. Lastly, OLS suggests that other variables, such as the share of jobs in suspended economic activities and the dependency ratio, are significantly associated with treatment effects, but these features do not appear at all in the tree. Such differences suggest that simple linear techniques provide only an incomplete assessment of the more significant territorial-level associations. In sum, the replacement of traditional methods with more advanced data-driven techniques leads not only to more accurate estimates of treatment effects, but also to qualitatively different conclusions regarding their most relevant territorial factors and, ultimately, to different policy prescriptions.

6. Comparative analysis

The foregoing has provided evidence about the magnitude, heterogeneity, and characteristics of the COVID-9 shock in Italy. But is this recession different from previous ones? Fig. 6 reports an LLM-level comparison between two maps, both estimated using a counterfactual generated through the random forest: the employment losses of the COVID-19 crisis, i.e. the same map of Fig. 1, and those of the Great Recession.

Fig. 6.

Employment change in different recessions.

Three key insights stand out: i) in absolute terms, the magnitude of the national-level impact is stronger in the pandemic recession: −3% in 2020 vs −1.8% in 2008–2009, in line with the findings of Cajner et al. (2020) and Coibion et al. (2020) for the US; ii) the two crises hit different territories: the spatial correlation between the two sets of LLM-level estimates is −0.104, confirming the differences in the geographic patterns of employment losses during the two recessions documented by Chetty et al. (2020) for the US; iii) the macro-patterns are also clearly different: while Southern regions were the most severely affected during the Great Recession (−3.8%), the COVID-19 crisis has a more homogeneous impact across macro-areas and the largest employment losses are predominantly concentrated in the Centre-North.

Even though, due to the lack of granular data by sector, we cannot replicate the subsample analysis for manufacturing and services that we presented in Fig. 3, official data by the Bank of Italy report that the manufacturing sector was substantially more affected than the services sector during the Great Recession.25 In sum, while the Great Recession mainly hit the South and manufacturing activities, the pandemic recession triggered employment drops across the entire country and damaged the services sector more. The COVID-19 crisis is thus less similar to previous economic downturns than one would perhaps expect, not just in absolute terms but also concerning the most affected territories and sectors.

7. Conclusion

We have documented the striking local inequalities of the coronavirus crisis across the Italian territory. The largest employment losses are more concentrated in the Centre-North and associated with a combination of LLM-specific features such as sectoral specialization, exposure of economic activities to high social aggregation risks, and pre-existing labor market vulnerabilities. By contrast, there is no discernible spatial correlation between the economic and epidemiological patterns of the pandemic. Lastly, the territorial hotspots of the COVID-19 crisis do not overlap with those of the Great Recession.

These insights are provisional. First, the pandemic did not end in 2020. Second, as the exceptional protective labor market policies implemented by the government will be gradually lifted, the magnitude and spatial distribution of local impacts may change. More research is thus needed to monitor the evolution of local economies and provide rigorous, granular, and constantly updated empirical evidence. As it stands, we deem these local and spatial dimensions of the crisis to be policy-relevant, especially in light of the forthcoming resources from the NextGenerationEU initiative. There is growing evidence that the pandemic is increasing inequality at all levels, including territorial and regional divides (Stantcheva, 2021). Our work corroborates these findings. In this respect, the Italian case is emblematic, and the results we provide can be thoughtfully extended to similar dynamics that might be taking place in other countries too.

From a prescriptive viewpoint, our analysis calls for a place-based approach in the policy response to the crisis. As national policies and top-down plans will be insufficient to lead the recovery (Bailey et al., 2020), policymakers should not neglect the local evolution of this unprecedented shock. Therefore, such diverging trajectories emphasize the need for ad hoc policy interventions based on the territorial profile and sectoral specialization of local economic systems (Ascani et al., 2021). The place-based policy perspective we advocate is, in our view, the best possible approach to target the local hotspots of the COVID-19 crisis and prevent the pandemic from further exacerbating pre-existing territorial vulnerabilities.

Credit author statement

Augusto Cerqua: Conceptualization, Data curation, Validation, Visualization, Investigation, Writing- Reviewing and Editing, Funding acquisition, Marco Letta: Methodology, Software, Formal analysis, Writing – original draft preparation, Project administration, Supervision.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A.

Table A1.

Definition of the variables included in the empirical analysis.

Variable name	Definition	Time period	Source
Counterfactual analysis
Employment	Overall employment of private non-financial sector firms	2014 Q3–2020 Q4	Business Register
Employment in manufacturing	Overall manufacturing employment	2014 Q3–2020 Q4	Business Register
Employment in services	Overall services employment	2014 Q3–2020 Q4	Business Register
Business births	Companies that have registered in the period under review	2014 Q1–2020 Q4	Business Register
Business deaths	Companies that went out of business in the period under review	2014 Q1–2020 Q4	Business Register
Economic classification dummies	Without specialization, non-manufacturing (touristic), non-manufacturing (non-touristic), made in Italy, other manufacturing		Istat
Geographical dummies	North-East, North-West, Centre, South		Istat
Log of population	Log of the resident population	2014–2019	Istat
Population density	Resident population per unit area	2014–2019	Istat
Unemployment rate	Resident population aged 15+ not in employment but currently available for work	2014–2019	Istat
Activity rate	The number of people employed and those unemployed as a % of the total population	2014–2019	Istat
Share of foreign population	Foreigners/population	2014–2019	Istat
Association analysis
Employment change in 2020	Treatment effect of the COVID-19 crisis on overall employment levels	2020 Q4	Estimated via the MLCM
Unemployment rate	Resident population aged 15+ not in employment but currently available for work	2019	Istat
Excess mortality estimates	Municipality-level excess mortality estimated by applying ML techniques to all-cause deaths data, aggregated at the LLM level	From Feb 21, 2020 to Sep 30, 2020	Cerqua et al. (2021),
Share of jobs having a high risk of social aggregation	Number of employees exposed to a medium-high or high risk of social aggregation divided by the number of employees	2019	Own calculations using Business Register data
Share of jobs having a high integrated risk	Number of employees exposed to a medium-high or high integrated risk divided by the number of employees	2019	Own calculations using Business Register data
Share of temporary contracts	Number of employees with temporary contracts in October divided by the number of employees in October	2015	Istat
Share of jobs in suspended economic activities	Share of jobs in activities suspended in March 2020 by the Italian Government due to the spread of the pandemic	2017	Istat
Per capita income	The amount of money earned per person	2019	Ministry of Economy and Finance
Share of innovative start-ups	The ratio between innovative start-ups and the universe of firms registered in the Business Register	Average (2016–2019)	Business Register
Share of firms having employees in CIGS	The number of firms with employees in CIGS divided by the universe of firms registered in the Business Register	Average (2015–2018)	Ministry of Labor and Social Policies
Number of road accidents per 10,000 inhabitants	The number of road accidents with injuries to persons divided by resident population ∗ 10,000	2019	Istat
Dependency ratio	The ratio of those typically not in the labor force (the dependent part, ages 0 to 14 and 65+) and those typically in the labor force (the productive part, ages 15 to 64)	Jan 1, 2020	Istat
Share of population living in peripheral areas	Share of population living in areas defined by Istat as peripheral or ultra-peripheral	Jan 1, 2020	Istat
Index of relational intensity (IIRFL)	The percentage of flows within an LLM that connect different municipalities on the total of flows within the LLM. This indicator ranges from values close to 0 to 100 (case in which all the workers of the municipalities of the LLM go to work in another municipality). The higher the indicator, the greater the inter-municipal turbulence in terms of flows	2011	Istat
Number of hospital beds per 1000 inhabitants	Number of hospital beds divided by resident population ∗ 1000	2018	Ministry of Health
Share of workers employed in health care occupations	Share of jobs in the NACE 2-digit sectors ‘human health activities’ and ‘residential care activities’	2019	Own calculations using Business Register data
Supply-side changes	Supply-side changes due to the closure of non-essential industries and workers not being able to perform their activities at home	2019	Own calculations using forecasts by del Rio-Chanona et al. (2020)
Demand-side changes	Demand-side changes due to people's immediate response to the pandemic, such as reduced demand for goods or services that are likely to place people at risk of infection	2019	Own calculations using forecasts by del Rio-Chanona et al. (2020)
Great Recession analysis
Employment	Overall employment, including also all public employees	2006–2009	Istat
Economic classification dummies	Without specialization, non-manufacturing (touristic), non-manufacturing (non-touristic), made in Italy, other manufacturing		Istat
Geographical dummies	North-East, North-West, Centre, South		Istat
Business births	Companies that have registered in the period under review	2005–2007	Business Register
Business deaths	Companies that went out of business in the period under review	2005–2007	Business Register
Population density	Resident population per unit area	2005–2007	Istat
Per capita income	The amount of money earned per person	2005–2007	Ministry of Economy and Finance

Notes: Outcome variables in bold.

Table A2.

Descriptive statistics. Association analysis.

Variable name	Mean	SD	Min	Max
Counterfactual analysis
Employment (log)	9.30	1.25	5.95	14.41
Employment in manufacturing (log)	7.52	1.61	3.50	12.65
Employment in services (log)	8.88	1.28	5.51	14.22
Business births	57.96	243.92	1	5173
Business deaths	46.72	204.10	1	9685
Share of LLMs without specialization	0.19	0.39	0	1
Share of touristic LLMs	0.14	0.34	0	1
Share of non-manufacturing (non-touristic) LLMs	0.23	0.42	0	1
Share of made in Italy LLMs	0.31	0.46	0	1
Share of manufacturing LLMs	0.14	0.35	0	1
<=10,000 inhabitants	0.08	0.28	0	1
(10,000; 50,000]	0.46	0.50	0	1
(50,000; 100,000]	0.25	0.43	0	1
(100,000; 500,000]	0.18	0.39	0	1
>500,000 inhabitants	0.03	0.16	0	1
Activity rate (%)	48.10	6.67	29.79	63.91
Unemployment rate (%)	12.14	6.19	1.19	39.08
Population (log)	10.71	1.13	8.05	15.18
Population density	0.21	0.30	0.01	3.17
Share of foreign population	0.07	0.04	0.00	0.18
Number of LLM-quarters	10,980
Association analysis
Employment change in 2020 (%)	−3.47	4.20	−26.90	9.71
Unemployment rate (%)	10.99	5.91	1.19	36.19
Excess mortality estimates (%)	7.99	19.72	−34.30	148.07
Share of jobs in suspended economic activities	0.47	0.08	0.25	0.79
Per capita income (€)	16559	4109	8050	27664
Share of firms having employees in CIGS	0.0008	0.0007	0	0.0046
Share of population living in peripheral areas	0.29	0.40	0	1
Share of temporary contracts	0.19	0.08	0.10	0.56
Number of road accidents per 10,000 inhabitants	2.18	1.20	0	6.94
Index of relational intensity (IIRFL)	25.70	14.48	0.2	66.1
Dependency ratio	0.58	0.05	0.43	0.78
Share of innovative start-ups	0.003	0.003	0	0.017
Share of jobs having a high risk of social aggregation	0.23	0.11	0.06	0.76
Share of jobs having a high integrated risk	0.06	0.03	0.01	0.37
Number of hospital beds per 1000 inhabitants	2.43	3.16	0	24.27
Share of workers employed in health care occupations	0.0253	0.0265	0	0.3530
Supply-side changes (used in the sensitivity check)	−0.27	0.06	−0.51	−0.10
Demand-side changes (used in the sensitivity check)	−0.21	0.08	−0.08	−0.61
Number of LLM	610
Great Recession analysis
Employment (log) – Istat data	9.71	1.17	7.21	14.31
Share of LLMs without specialization	0.19	0.39	0	1
Share of touristic LLMs	0.14	0.34	0	1
Share of non-manufacturing (non-touristic) LLMs	0.23	0.42	0	1
Share of made in Italy LLMs	0.31	0.46	0	1
Share of manufacturing LLMs	0.14	0.35	0	1
<=10,000 inhabitants	0.08	0.28	0	1
(10,000; 50,000]	0.46	0.50	0	1
(50,000; 100,000]	0.25	0.43	0	1
(100,000; 500,000]	0.18	0.39	0	1
>500,000 inhabitants	0.03	0.16	0	1
Population density	0.20	0.29	0.01	3.12
Per capita income (€)	15408	3136	8757	25172
Number of LLM-years	1830

Table A3.

OLS-based association analysis

Feature	Coefficient	Standard error	p-value
Unemployment rate	12.525	3.498	0.000370
Excess mortality estimates	0.00901	0.00647	0.165
Share of jobs in suspended economic activities	−6.352	1.577	0.0000639
Per capita income	−0.000250	0.0000623	0.0000684
Share of firms having employees in CIGS	−107.763	208.675	0.606
Share of population living in peripheral areas	0.464	0.367	0.206
Share of temporary contracts	−13.783	2.160	0.000
Number of road accidents per 10,000 inhabitants	0.0859	0.131	0.511
Index of relational intensity	−0.00919	0.0106	0.385
Dependency ratio	−5.990	2.542	0.0188
Share of innovative start-ups	63.971	45.158	0.157
Number of hospital beds per 1,000 inhabitants	−0.0108	0.0437	0.805
Share of workers employed in health care occupations	14.931	8.177	0.0684
Share of jobs having a high risk of social aggregation	−19.656	1.236	0.000
Share of jobs having a high integrated risk	−4.517	7.213	0.531

Notes: The outcome variable is 2020 employment change estimated with the 1-year before-after methodology. Intercept not reported.

Fig. A1.

– Regression tree on 2020 employment change using demand- and supply-side changes reported indel Rio-Chanona et al. (2020) .

Appendix A. Supplementary data

The following are the Supplementary data to this article: