The impact of COVID‐19 on international tourism flows to Italy: Evidence from mobile phone data

Abstract

This paper analyses the response to the COVID‐19 pandemic of inbound tourism to Italy looking at variation across countries and provinces. To this end, it uses weekly data on the number of foreign visitors in Italy from January 2019 until February 2021, as provided by a primary mobile telephony operator. We document a very robust negative relation at the province level between the local epidemic situation and the inflow of foreign travellers. Moreover, provinces with a historically higher share in art‐tourism, and those that used to be ‘hotel intensive’ were hit the most during the pandemic, while provinces with a more prevalent orientation to business tourism proved to be more resilient. Entry restrictions with varying degrees of strictness played a key role in explaining cross‐country patterns. After controlling for these restrictions, we observed that the number of travellers that could arrive by private means of transportation decreased proportionally less. Overall, this evidence emphasises that contagion risk considerations played a significant role in shaping international tourism patterns during the pandemic.

1. INTRODUCTION

The outbreak of the COVID‐19 pandemic in the early months of 2020 caused unprecedented disruption to tourism flows. 1 According to the World Tourism Organization (UNWTO), in 2020 international arrivals worldwide dropped by 74% (1 billion arrivals less than the previous year). Italy, a country for which the tourism industry is very important, 2 was among the first EU countries to be hit by the pandemic: between February and April 2020, positive cases rapidly rose from a few hundreds to over a hundred thousand, with a surge in the number of patients needing intensive care and in the number of deaths. 3

Fear of contagion and containment measures (including travel bans) resulted in tourism flows dropping to near‐zero levels since the end of March 2020. During the second quarter of 2020 conditions improved, allowing for the lifting of travel restrictions at the EU level in the summer. Italy, among other southern European countries (Spain, Portugal and Greece), benefited from the recovery of cross‐border tourism, although flows remained at around a half of pre‐pandemic levels. The second wave of the pandemic that hit Italy after the end of the summer halted again tourism flows. Overall, in 2020 foreign travellers' expenditure in Italy fell by about three fifths compared with previous year (from €44 to 17 billion), and the travel surplus of the balance of payments was halved to 0.5 per cent of GDP (from 1.0 per cent in 2019).

In this context, the adequate design and evaluation of policy responses clearly requires a thorough understanding of how inbound tourism is affected by contagion risk and to containment measures of different intensity. In particular, two main questions deserve closer investigation to inform policy decisions. The first is to what extent the fall in foreign arrivals reflects not only regulatory restrictions and containment measures (travel bans, quarantines, etc.) but also fears of contagion that spontaneously lead travellers to stay away from destinations with a locally higher epidemiological risk. Answering this question is highly relevant from a policy perspective: lifting restrictions while the epidemic is still not under control might not be sufficient to revamp tourism flows if travellers' behaviour actively responds to the risk of contagion. In fact, it may affect tourists decision in the future period.

The second related question is how travel preferences changed in reaction to the pandemic, looking at characteristics that were indirectly related to contagion risk, such as transport means, type of accommodation and amenities at the destination. A proper understanding of these factors is needed to formulate reasonable predictions about which destinations are going to record a larger drop in tourism inflows, so that adequate policy responses can be prepared. Understanding how tourists react to travel restrictions of varying intensity (from quarantine requirements to screening tests) would also be useful for the same purpose.

This paper uses a unique combination of weekly mobile phone data and survey data for Italy to provide answers to the above questions, through an overarching analysis of international tourism flows during the pandemic. The high frequency of mobile phone data on the number of foreign visitors by nationality and province allows us to identify precisely the impact of changing patterns in the epidemics and of the adopted policy measures. We estimate reduced‐form models (consistent with a gravity framework) where the number of foreign travellers in a given location is related to the risk of contagion in the province of stay as well as in the source country, controlling for an extensive set of fixed effects. We also look at how structural characteristics of destinations shaped the dynamics of tourism flows in interaction with the contagion dynamics. We provide compelling evidence that travellers paid a lot of attention to contagion risk during the second wave of contagion—when travel restrictions were looser—avoiding local Italian destinations with a higher number of COVID‐19 cases. Furthermore, destinations that were perceived as ‘less risky’ by tourists (for instance because they were reachable by private means of transport or had a larger share of private accommodations), were hit less, all other things being equal.

This paper is at the intersection of two strands of literature. The first and larger strand is about the adverse effects that infectious diseases cast on the economy, and on tourism in particular. It received an important boost in the 2000s, after the outbreak of the SARS and the ‘aviary flu’ in Asia (Chou et al., 2004; Hanna & Huang, 2004; McKercher & Chon, 2004), followed by studies on MERS (Joo et al., 2019) and the H1N1 influenza (Rassy & Smith 2013). All of these studies show that the tourism sector was hit the hardest, finding a negative relationship between contagion dynamics and foreign arrivals. In particular, Hanna and Huang (2004) find that the impact was higher in regions characterised by higher population density, higher mobility of people, and where public health infrastructure was less developed. Chou et al. (2004) conclude that a failure in disclosing the actual number of SARS cases can deliver additional GDP loss in the longer run, pointing to the fact that not only international travellers but also foreign investors need accurate information on the dynamic of the epidemic. More recently, Cevik (2020) compares the impact of different kind of diseases on bilateral tourism flows, showing that the impact on tourism is due more to the contagiousness of the disease than to its severity, and that negative effects are stronger for developing countries.

With the outbreak of COVID‐19, the first truly global pandemic after the 1918–1919 influenza (so‐called ‘Spanish flu’), a large and growing bulk of papers was added to this workstream. Given the pervasiveness of the shock and the strictness of countermeasures that were adopted worldwide, studies have analysed the impact not only on tourism but also on trade of goods (Bas et al., 2022; Berthou & Stumpner, 2022; Liu et al., 2021) and services in general (Ando & Hayakawa, 2022; Minondo, 2021). 4 The present crisis is in fact characterised by quick and wider developments, impacting all countries across the globe. As regards impact of COVID‐19 on tourism, existing studies are largely descriptive (MacDonald et al., 2020; Metaxas & Folinas, 2020; Uğur & Akbıyık, 2020; see Sigala, 2020 for a preliminary survey) or focus on specific segments of the tourism industry, such as short‐term rental: Hu and Lee (2020) quantify the impact of lockdown on global AirBnB bookings. Focusing on the European short‐term rental market, Guglielminetti et al. (2021) find that the epidemic reduced markedly both the supply of apartments available for rents and the consumers' demand. Our paper contributes to this literature with a rich econometric analysis of the effects of COVID‐19 on foreign arrivals in Italy. We believe that Italy is an ideal setting for this analysis, for three reasons. First, it is one of the largest exporters of tourism services (Italian tourism exports rank sixth in the world, according to UNWTO), so it is a very relevant case study. Second, it is endowed with a well‐diversified range of destinations associated with different travel purposes (business trips, art visits, beach or mountain holidays, etc.), and it attracts visitors from a very diverse set of departure countries, which allows to study the interaction between characteristics of both local destinations and countries of departure. Third, the significant heterogeneity in the spread of contagion across the country allows a quite neat identification of the response of tourism to the differential level of the epidemic among local destinations, while controlling for developments at the country level. This allows us to draw several conclusions on the response of international tourism to the pandemic which are potentially useful for policymaking purposes.

The second strand of literature this study is related to is the growing number of research papers using location data derived from mobile phone networks for the analysis of mobility and consumer behaviour (Hu et al., 2009; Tucker & Yu, 2020). Mobile phone data have been used in behavioural studies for almost two decades (Spinney, 2003) and the use of this data for tourism analysis is not entirely new. 5 The availability of such data accelerated when smartphones massively replaced first‐generation mobile phones. As this paper confirms, this type of data has become a very valuable complement to more conventional data sources (e.g. survey data), especially for tourism analysis.

The paper is structured as follows: Section 2 provides descriptive evidence on the changes that occurred in incoming tourism flows after the pandemic along various dimensions, paving the way for the subsequent econometric analysis. Section 3 presents the database and the empirical model adopted to measure the impact of the pandemic on the incoming tourism flows and its interaction with variables at the province and the country of departure level. In Section 4, we present and discuss estimation results, robustness evaluations and economic interpretation of regression coefficients. Finally, Section 5 summarises our findings and draws concluding remarks.

2. AGGREGATE PATTERNS OF FOREIGN TOURISM FLOWS IN ITALY

This section of the paper presents the main aggregate patterns in foreign tourism to Italy in 2020, highlighting the heterogeneous impact of the pandemic. This evidence guides us in the selection of relevant variables for the empirical model presented in Section 3.

The COVID‐19 disease started to spread in Italy in the second half of February 2020. The lockdown was applied initially in selected Northern provinces and, since March 9, in the entire country. It included a stay‐at‐home order, the shutdown of all non‐essential economic activities and restrictions to both internal and international mobility. In this phase, the outbreak remained concentrated in Northern Italy. These restrictions were lifted during the month of May 2020. The strong containment measures proved to be effective in halting the spread of the disease, and Italy benefited of near‐zero rate of new COVID‐19 cases throughout the summer. In early June travel restrictions between EU member countries, Schengen Area countries and United Kingdom were lifted, and inbound tourism gradually resumed. New case rates started picking up again at the end of August, and in the fall a second wave of contagion hit Italy throughout the country, with virtually no province spared from a rise in infections.

According to official statistics, in 2020 foreign visitors in Italy (i.e. including those who did not stay in Italy overnight) were 39 million overall, about 60% less than the previous year. 6 The drop in inbound tourism was sharp from all countries of origin, but particularly severe from farther countries (Table 1 and Figure 1): the number of arrivals from Europe (both EU and non‐EU) decreased by 56.2% with respect to 2019; those from the Americas and Asia fell by 87% and 81% respectively.

TABLE 1.

Changes in the number of foreign travellers in Italy.

Area of departure	% Change in arrivals (YoY)
European Union	−54.1
Non‐EU Europe	−61.7
Americas	−86.7
Asia	−81.1
Rest of the World	−75.5

Source: BISIT data. Changes refer to 2020 with respect to 2019.

FIGURE 1.

Changes in the number of foreign arrivals by area of origin and new COVID‐19 cases. Lines represent monthly foreign arrivals in 2020 vis‐à‐vis the corresponding months in 2019 in percentage terms (scale on the left‐hand axis). Histograms (scale on the right‐hand side axis) represent the number of new COVID‐19 cases occurred in Italy in each month.

These patterns were likely affected by the travel bans adopted in many countries throughout the world (including Italy), but they may also reflect a preference by foreign tourists for destinations closer to home that can be reached by private means of transport. Indeed, the drop of arrivals in regions closer to Italian borders (such as Veneto and Lombardy) was relatively smaller than in the other regions.

The pandemic also induced changes along the dimension of the travel's motive, as suggested by the correlation between the ex ante shares of various travel purposes in each Italian province (which capture their ‘touristic specialisation’) and the change in arrivals between 2019 and 2020 (Figure 2). 7 Arrivals dropped systematically more in provinces specialised in cultural tourism purposes, while this correlation is weaker for ‘sea and nature’ holidays. The correlation is instead positive in the case of business tourism, meaning that the provinces that used to have a relatively higher share of tourism related to business reasons suffered much less in terms of decline in foreign arrivals.

FIGURE 2.

Correlation between change in arrivals and travel purpose shares at province level. Each dot represents an Italian province. In all graphs, vertical axis reports the drop in arrivals between 2019 and 2020 in % terms, while horizontal axis reports the share of travellers that used to visit the province before 2020 for the specified travel purpose.

Finally, another relevant change was observed along a third dimension of interest: the type of accommodation chosen by visitors during their sojourn in Italy. As shown in Table 2, comparing 2020 data with the pre‐COVID‐19 three‐year period (2017–2019), shares of ‘traditional’ accommodations (hotel, B&B, tourist resort) decreased significantly (for over 14 percentage points), mainly to the advantage of independent non‐shared accommodations (rented houses or own properties) or other less common accommodations (campers, tents, caravans, etc.). The share of visitors who stayed at home with relatives or friends during their sojourn also grew significantly.

TABLE 2.

Accommodation choices pre‐ and post‐COVID‐19.

Accommodation type	2017–2019	2020
Hotel, resort, and B&B	57.6	43.5
Hosted by friends or relatives	15.6	20.6
Rented house or own house	10.1	13.1
Other accommodations n.i.e.	16.7	22.9
Total	100	100

Note: ‘Other accommodations’ includes also camping, caravans and farmhouses.

Source: BISIT data. All values are shares. Values for 2017–2019 are averages.

3. THE HETEROGENEOUS IMPACT OF COVID‐19 ON TOURISM: DATA AND EMPIRICAL MODEL

3.1. Data sources and variables definition

We combine various sources of information about tourism, epidemiological patterns and policy measures, to build a comprehensive and detailed dataset for our empirical exercise. The dataset covers the period from January 2019 to February 2021.

Two main sources are used for tourism data, to quantify the number of foreign tourists and to gather information on tourism characteristics. The first source of data comes from a primary Italian mobile phone operator. It provides the total number of foreign phone SIM cards on the Italian territory, by province and by issuer country. We use the former as information about the province of destination and the latter as a proxy for the country of origin of the traveller. Mobile phone data are available at a daily frequency (we aggregate them into weekly data). This source provides several important advantages. First, the data cover also the months in which the Bank of Italy Survey on International Travel was discontinued because of the restrictions against the spread of the pandemic. Second, the higher (weekly) data frequency allows to assess the impact that the contagion dynamics and the policy responses had on tourism patterns in a much more precise way than what could be done with monthly data: for instance, we can match the increase in cases occurred in a given week with the tourism flows observed in subsequent weeks, while controlling for the travel restriction in place in that specific week of the year. Finally, the extensive coverage provided by mobile phone data allows to look at combinations of ‘country of origin – province – time’ that in BISIT data may be subject to significant measurement error (e.g. for smaller countries and provinces). One limitation however is that the number of foreign tourists derived from mobile phone data may be distorted by the presence of communities of foreign residents in Italy. To avoid this potential bias, in our analysis we considered the first 40 countries, in terms of the number of tourists in 2017–2019, excluding those having large communities of residents in Italy. The selected countries account for about 94 per cent of the total inbound tourism flows to Italy (over the period 2017–2019); half of them belong to the European Union. 8

The second source of tourism data is the Bank of Italy Survey on International Tourism (BISIT). The survey questionnaire asks the interviewed traveller to provide information about the kind of transportation used to reach the destination, the purpose of the trip and the type of accommodation used during the trip (if any). We use data for the period 2017–2019 to construct indicators before the pandemic outbreak: for each province and origin, we quantify the shares of travellers by travel purpose, accommodation type and means of transport.

The epidemiological data regarding the spread of the contagion in Italy are sourced from the Italian Civil Protection Department. 9 At province level, the only available information is the cumulative number of positive COVID‐19 cases, at a daily frequency. From this, we compute the number of new cases of COVID‐19 (gross of recovered patients) over a period of 14 days, per 1000 inhabitants. The resident population in the province at the end of 2019 is retrieved from ISTAT, the Italian national statistical institute.

The corresponding information on the evolution of COVID‐19 in the foreign countries of origin was obtained from the European Center for Disease Prevention and Control (ECDC), which provides harmonised and comparable data on the rate of contagion in all European countries and in all other non‐European countries considered in our analysis.

As for the containment measures adopted by foreign countries, we used the Oxford Stringency Index (Hale et al., 2021), which reflects restrictions to different aspects of economic and social life, such as mandatory closure of schools and offices to remote functioning, shops and restaurants closures, restrictions on public transportation and international travel bans. To control for the different intensity of the restrictions by Italian regions enforced since November 2020, we relied on the index developed for Italy by Conteduca (2021). 10

We also constructed a set of dummies related to the intensity of bilateral travel restrictions enforced by the Italian Government. This information was collected from the legislation acts adopted throughout the period, also relying on the website ‘reopen.europa.eu’, and on the website of Italy's Foreign affairs Ministry ‘www.viaggiaresicuri.it’.

Finally, variables on bilateral distance were retrieved from the CEPII data warehouse (Mayer & Zignago, 2011).

3.2. The empirical strategy

Our empirical exercise aims at explaining the heterogeneous impact of COVID‐19 on international tourism to Italy disentangling the contribution of various factors at the province and the country‐of‐origin level. In practice, the empirical strategy relies on two mirror‐like reduced‐form models for inbound tourism to Italy that are in line with a gravity framework. We estimate those models using the Poisson pseudo maximum likelihood estimator on weekly data from January 2019 to February 2021. 11

Our first model estimates the effects of contagion at the province level and of province's characteristics, while controlling for time‐varying characteristics of tourists' countries of origin with fixed effects (Equation 1).

(1)

The dependent variable, ${\mathit{\text{Tourists}}}_{\mathit{opt}}$, is the total number of days spent by tourists from country $o$ in province $p$ at time $t$, where temporal unit $t$ denotes a combination of year‐week. Our identification strategy exploits the granularity of the data set, and it includes an extensive set of fixed effects to control for unobservable factors. Country–province–week factors (${\alpha }_{\mathit{opw}}$) control for the preference of travellers from a specific country for a specific province $p$ in a week $w$. 12 Such preferences may be motivated by the availability of convenient flight connections, by business links and of course by the characteristics of the touristic offer of the destination compared to the domestic market (for instance, German tourists may favour beach destinations in Italy in summer weeks relatively more than French tourists, because France also offers attractive seaside destinations to domestic tourists). We also include time‐varying factors related to the country of departure ${\alpha }_{\mathit{ot}}$, which control for all developments that occurred at time $t$ in the country of departure, in Italy, or third countries, that could affect the number of arrivals (for instance in terms of the epidemic or in containment measures). 13

Our main explanatory variable in Equation (1) is cases _pt − 1, which is the number of new COVID‐19 cases on 1000 inhabitants that were recorded in the province during the previous 2 weeks, a commonly used metric to measure epidemic developments. This variable allows us to verify whether tourists were concerned about the level of contagion risk not only at the country level (which is captured by the fixed effects) but also at the local level. Indeed, information on local developments of the COVID‐19 epidemic is widely and easily available on the web. Therefore such information may be consulted by travellers before travelling to a given country, in order to avoid destinations where the epidemic is spreading faster.

To elicit the effect of the pandemic outbreak on tourists' choices, we interact variables Purpose _op , Accommodation _op and Transport _op with a dichotomic variable that marks the COVID‐19 period, taking value one from the last week of February 2020 onward. These variables are vectors of shares extracted from BISIT data for the years 2017–2019, as explained in Section 3.1. Purpose _op reports the shares of various purposes of the trips, as declared by foreign travellers from country o when they visited province p before the pandemic: ‘art and culture holiday’, ‘sea and nature holiday’, ‘other purposes trip’ and ‘business reasons’ (the latter being the base category). These shares are computed for each season to take into account possible seasonality in the purpose of travel for some destination.

In the same fashion, Accommodation _op reports the shares of various accommodation choices made by travellers: ‘hotels and hostels’, ‘camping, farmhouses, and caravans’, ‘day‐trip (and others)’, which is associated with no accommodation at all or with alternative types of accommodation, and the base category ‘own house, or hosted by relatives/friends, or at a rented house/flat’. Finally, Transport _op indicates the shares of transports typologies chosen by travellers from country $o$ to reach their destination $p$ before the pandemic. We classified them into two categories: (i) collective and/or mass transports (planes, ships and trains) and (ii) individual/private transports (cars, caravans, bikes and motorcycles), our base category.

As mentioned, we estimate the model by Pseudo Poisson Maximum Likelihood regression—PPML, in line with the literature on gravity models of trade (Santos Silva & Tenreyro, 2006). 14 An advantage of PPML is that it allows the inclusion of null observations, namely provinces that tourists from country $o$ visited in week $w$ in 2019 but they did not visit in 2020. In our case, these are potentially meaningful observations as they refer to flows that were hit the hardest by the pandemic. Moreover, PPML is a consistent estimator in presence of heteroskedasticity (even if the dependent variable does not follow a Poisson distribution) and lends itself well to model count variables, as it is our dependent variable. In our inference, we assume double‐clustering by country of departure–time and by province–time.

In a second step, we drop the country–time fixed effects ${\alpha }_{\mathit{ot}}$ from the model and introduce variables related to the evolution of the epidemic, the containment measures, the bilateral entry restrictions imposed by Italy, and distance, to explain the cross‐country variation in international tourism inflows (Equation 2):

(2)

Here, we include fixed effects ${\alpha }_{\mathit{pt}}$ to control for any factor at play at time $t$ in province $p$ (including COVID‐19) that can have an impact on inbound tourism in that province from any destination. This specification is thus designed to estimate the effects of variables indexed by ot (country‐of‐origin and time), exploiting variation across countries at time t, while controlling for time‐varying province‐specific pt factors.

We consider the following additional explanatory variables: ${\mathit{\text{cases}}}_{\mathit{ot}-1}$ is the number of new COVID‐19 cases over 1000 inhabitants over a period of 14 days ending in week t − 1 in the country of departure o (Section 3.1). ${\mathit{\text{Entry restrictions}}}_{\mathit{ot}}$ is a set of dummies indicating the bilateral travel restrictions (if any) imposed by Italy vis‐à‐vis other countries. We distinguished between (i) the travel restrictions that allow entry from a country only for urgent and/or essential reasons, like health motives or repatriations (${\mathit{\text{Necessity only}}}_{\mathit{ot},\mathit{IT}}$), (ii) restrictions that allow entry only for work reasons and/or upon a quarantine period (${\mathit{\text{Quarantine}}}_{\mathit{ot},\mathit{IT}}$), (iii) restrictions that allow entry upon a negative result of swab test (either at arrival or before departure) (${\mathit{\text{Swab}}}_{\mathit{ot},\mathit{IT}}$). ${\mathit{\text{Stringency}}}_{\mathit{ot}}$ is the Oxford Stringency index (which takes values in the 0–100 interval, depending on the intensity of containment measures adopted by the country o at time t). 15

We further interact the indicator variable for the COVID‐19 period with two variables measuring distance, to check whether foreign tourists from closer countries reduced their presence in Italy relatively less than tourists from more distant countries, in addition to what is already captured by the variable ${\mathit{\text{Transport}}}_{\mathit{op}}$, which varies by the province of destination p and country of departure o. These two variables are the logarithm of the bilateral population‐weighted distance between Italy and country o, and an indicator variable which is equal to one if the country has a common border with Italy. 16

4. RESULTS AND DISCUSSION

4.1. Analysis by local destination

Table 3 reports results from the estimation of the model in Equation (1). Column (1) includes only the ‘local contagion’ variable (new positive cases in the province) and the full set of fixed effects: the coefficient of the contagion variable is negative and statistically significant. Given our specification of fixed effects, it means that if a province records 100 new positive cases per 100,000 inhabitants more than other provinces over 2 weeks, that province will experience on average a reduction in the number of foreign tourists about 6 percentage points larger than other provinces in the subsequent week, ceteris paribus. The contagion variable remains highly significant, with a slightly larger coefficient (column 2), when we add controls for the interaction between province–country structural characteristics and a dichotomous variable signalling the start of the pandemic. 17 The remaining columns of Table 3 report the results for three different phases of the epidemic in Italy. The first phase goes from February 25th to June 2nd 2020, and it covers the lockdown period (column 3). The second phase includes the summer period until September 15th, a period characterised by a gradual recovery of inbound tourism and by negligible rates of new COVID‐19 cases (column 4). The third phase covers the second wave of contagion, and it extends from mid‐September 2020 to February 2021 (column 5). Estimation results show that the negative relation between the number of foreign arrivals and new COVID‐19 cases materialised only during the latter phase, which includes the second wave of COVID‐19. This period is in our view the most appropriate setting to study the impact of new cases on inbound tourism because it was characterised by milder travel restrictions, by a larger degree of awareness about the health situation and by more information accessible to tourists about the local evolution of the epidemic. 18 On the contrary, we do not include the cases variable for contagion when estimating the model for the summer period (column 4), given the extremely low number of new cases in most provinces during summertime, as otherwise the estimate would be driven by a few observations only. For similar reasons, when we include the cases variable and estimate the model for the first wave, we are aware that the subsequent results should be taken with caution, since the travel restrictions in place during that period effectively blocked all tourists, except those travelling for reasons of need or work. Indeed, the sign of the estimated coefficient over this period is found to be, counter‐intuitively, positive. This result could be however rationalised considering that during the first wave, contagion occurred overwhelmingly in northern regions of Italy, which were also the regions more frequented by foreigners travelling for business reasons. For instance, cross‐border workers, which typically work in northern Italy, could enter and exit the country even during the first wave. When we exclude from ‘the first wave’ estimation the countries bordering Italy (Table A3 in Appendix 1), the significance of the coefficient on the ‘new cases’ vanishes for this period. 19

TABLE 3.

Analysis by province.

Dep. var: ${\mathit{\text{Tourists}}}_{\mathit{opt}}$
	All sample		$1\mathrm{st}$ wave	Summer	$2\mathrm{nd}$ wave
	(1)	(2)	(3)	(4)	(5)	(6)	(7)
${\mathit{\text{cases}}}_{\mathit{pt}-1}$	−0.0553***	−0.0695***	0.0427**		−0.0664***	−0.0606***	−0.0396***
	(0.00900)	(0.00871)	(0.0217)		(0.00754)	(0.00688)	(0.00349)
${\mathit{d}}_{\mathit{Cov}19}$ × ${\mathit{\text{purpose}}}_{\mathit{op}}$
Nature and beach		−0.513***	−1.139***	−0.435***	−0.414***	−0.412***	−0.492***
		(0.0340)	(0.119)	(0.0445)	(0.0460)	(0.0452)	(0.0415)
Art and culture		−1.066***	−1.495***	−0.945***	−1.079***	−1.090***	−1.127***
		(0.0354)	(0.103)	(0.0500)	(0.0515)	(0.0515)	(0.0422)
Other pers. reasons		−0.361***	−0.190***	−0.305***	−0.443***	−0.444***	−0.310***
		(0.0259)	(0.0626)	(0.0362)	(0.0339)	(0.0339)	(0.0257)
$d_{\mathit{Cov}19}$ × ${\mathit{\text{accommodation}}}_{\mathit{op}}$
Hotels/hostels		−0.670***	−1.006***	−0.401***	−0.967***	−0.973***	−0.499***
		(0.0307)	(0.0802)	(0.0385)	(0.0368)	(0.0365)	(0.0237)
Camping/farmhouse		−0.284***	−0.739***	−0.488***	0.406***	0.400***	0.224***
		(0.0456)	(0.127)	(0.0522)	(0.0688)	(0.0680)	(0.0476)
Others		−0.453***	−0.579***	−0.288***	−0.713***	−0.721***	−0.491***
$d_{\mathit{Cov}19}$ × ${\mathit{\text{airplane}}}_{\mathit{op}}$		−0.347***	−0.411***	−0.401***	−0.263***	−0.263***	−0.513***
		(0.0257)	(0.0642)	(0.0380)	(0.0362)	(0.0363)	(0.0224)
RR index						−0.00955***
						(0.00345)
FE country#prov#week	Yes	Yes	Yes	Yes	Yes	Yes	Yes
FE country#time	Yes	Yes	Yes	Yes	Yes	Yes	Yes
FE reg.#time	No	No	No	No	No	No	Yes
Pseudo $R^2$	.99	.99	.99	.99	.98	.98	.99
Observations	340,496	311,082	75,956	84,300	127,660	127,660	127,660

Note: The table presents results of the model 1 estimated over different periods. Columns (1) and (2) look at the whole sample (January 2019–February 2021). The 1st wave period (column 3) includes the weeks from 25 February to 2 June 2020. The Summer period (column 4) goes from 3 June to 15 September 2020. The 2nd wave (columns 5–7) goes from 16 September 2020 onward. The reference categories for the variable interacted with a dummy for the COVID‐19 period are, respectively, ‘business‐reasons’ for the purpose of travel, and ‘rented house or private house’ for the accommodation. The reference category for ‘airplane’ (short for public means of transport) is private means of transport (like own car). Standard errors, in parenthesis, are clustered by province–time and country of departure–time. Stars (***, ** and *) indicate statistical significance at 1, 5 and 10 per cent, respectively. Fixed effects by country of departure–province–week (${\alpha }_{\mathit{opw}}$) and country of departure–time (${\alpha }_{\mathit{ot}}$) are always included.

Since early November 2020, new restrictive measures were introduced in Italy, based on an assessment of epidemiological risk at the regional level. After this policy change, epidemiological risk per se cannot be considered anymore the main explanatory variable for the decrease in inbound tourism, as internal mobility restrictions may also contribute to it, reducing the attractiveness of a province. We thus include in column (6) a one‐week lag of the regional restriction index (RR‐Index) constructed by Conteduca (2021). 20 As expected, we find the coefficient of the RR‐Index to be negative and statistically significant, meaning that tourists avoided provinces where more stringent restrictions were in place. Nevertheless, the coefficient of our contagion variable remains significant and almost unaffected in size, meaning that even after controlling for internal mobility restrictions, foreign tourists decreased more in provinces where contagion risk was higher, all other things being equal.

As a further robustness check, we replace the restriction index with region–time fixed effects (column 7). This structure of fixed effects is able to control for the new system of region‐based restrictions while also capturing the correlation of the epidemic within provinces of the same region. Yet, the contagion variable remains negative and significant, and only marginally lower, further corroborating the robustness of our results about the adverse effect of contagion on foreign arrivals. 21 Overall, this result suggests that tourists paid attention not only to the national dynamics of the epidemic (which in our case is captured by the country–time fixed effects ${\alpha }_{\mathit{ot}}$) but also to local developments of the epidemic, with noteworthy policy implications: even at the local level, there is a trade‐off implied by loosening restrictions: on the one hand, it may attract more tourists in the short term; on the other hand, if more arrivals are associated with an increase in the number of cases, it may discourage inbound tourism later in time.

Results from the interaction between province–country structural characteristics and the COVID‐19 period also indicate that travellers took contagion risk into account in their decisions. The coefficients of these variables can be interpreted as the average differential impact of the outbreak of the pandemic across our observational units (country–province–week). Column (2) shows that provinces that were more ‘specialised’ in art and culture tourism were hit the hardest: a coefficient of −1.1 for art tourism means that an increase in $10\%$ in the proportion of tourists that used to visit the province for that purpose is associated to a $7\%$ larger drop in inbound tourism. The drop would be only about $4\%$ for provinces visited for beach or nature holidays, with tourism for personal reasons purposes (like leisure tourism) hit generally harder than business tourism (our base level). A possible driving factor underlying this result is related to the fact that trips motivated by work reasons were generally exempted by travel restrictions, hence visitors travelling for work reasons could come to Italy even when tourists that would visit for holiday reasons could not (for instance, this was the case during the first lockdown for visitors arriving from countries outside Europe). This may have favoured provinces receiving historically higher shares of business travellers, even in a period when conferences and big events were moved on virtual platforms or cancelled.

Results also show that provinces in which tourists used to stay in ‘hotel‐like’ accommodations were hit harder than provinces characterised by a larger share of private housing and/or rental houses (our base level). The latter type of accommodation may indeed be perceived as relatively safer by tourists, as it implies less social interaction with other people. Provinces with a higher share of tourists staying in ‘green’ accommodations, like camping and farmhouses, also appear to have been more resilient on average.

Finally, the third feature of interest under consideration is the means of transport used; in line with our expectations, provinces that used to have a larger share of visitors arriving by plane (or other shared means of transport, like train or ship) were hit harder, reflecting the perception of a higher risk of infection compared to private non‐mass transport means, like cars or caravans. Using an extreme case as an illustrative example, the number of visitors in a province from a country in which all tourists come by collective means of transport recorded a $30\%$ larger drop than a province in which tourists from the same country arrive by car.

The behaviour of these variables in the different sub‐samples is overall consistent to what described for the whole sample. In the summer, interestingly, the relative loss by hotel‐intensive provinces appears to be only a half than what estimated for the overall period, suggesting that during this period tourists may have been less concerned with contagion risk, consistently with the near‐zero cases in most provinces.

In Table 4, we report several robustness tests of the result on the variable measuring contagion at the province level using different metrics and specifications, finding robust and statistically significant coefficients with a comparable size. First, we consider a longer temporal lag (4 weeks, rather than 1 week) to compute the number of new cases, to account for the fact that tourists may make their travel plans sufficiently in advance. We obtain a coefficient almost identical (column 2). 22 We further check against the effects of few big outliers by winsorising the variable ${\mathit{\text{cases}}}_{p,t-1}$ at the 1st and 99th percentiles. Doing so delivers an even higher coefficient (column 3). In column (4) we include a quadratic term, which we find to be significant, suggesting a non‐linearity in the impact of this variable on arrivals: in other words, tourists seem to refrain more from travelling to Italian locations when the notification rate of new positive cases becomes high. We then include the cumulative number of positive cases at the province level (column 5). This metric takes into account the hypothesis that tourists may be sensitive to the past dynamics of positive cases in the destination province, rather than only to the current situation (although the two variables are to some extent correlated). We find that the notification rate of new positive cases remains highly significant and of similar magnitude. 23 As a robustness check, we estimate the baseline model 1 in log‐linear formulation by OLS (column 6). The coefficient on our contagion variable again remains negative and statistically significant, and only marginally lower. Finally, we consider two sub‐samples: first, we limit the analysis to the first 40 provinces in terms of inbound tourism in previous years (column 7), obtaining similar results. Second, we exclude the first 2 months of 2021 from our sample, to rule out the possibility that our results are distorted by a change occurred in the way new positive cases were recorded before and after 15 January 2021 (before the date, new positive cases were counted based only on the results of PCR molecular tests, while after that date, positive cases detected through rapid antigenic tests were included in the counter for the number of cases). Our results remain substantially unchanged. 24

TABLE 4.

Analysis by province: Different measures of COVID‐19 spread.

Dep. variable: ${\mathit{\text{Tour}}}_{\mathit{opt}}$						$\log \left({\mathit{Tor}}_{\mathit{opt}}\right)$	First 40 prov.	Up to December 2020
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
${\mathit{\text{cases}}}_{\mathit{pt}-1}$	−0.0645***			−0.0290**	−0.0398***	−0.0420***	−0.0605***	−0.0422***
	(0.00850)			(0.0148)	(0.00648)	(0.00576)	(0.00942)	(0.00460)
${\mathit{\text{cases}}}_{\mathit{pt}-4}$		−0.0524***
		(0.00858)
Wins. ${\mathit{\text{cases}}}_{\mathit{pt}-1}$			−0.0749***
			(0.0108)
${\mathit{\text{cases}}}_{\mathit{pt}-1}^2$				−0.00247**
				(0.00114)
Cumulative ${\mathit{\text{cases}}}_{\mathit{pt}-1}$					−0.0131***
					(0.00123)
$d_{\mathit{\text{Covid}}19}$ × ${\mathit{\text{purpose}}}_{\mathit{op}}$
Nature and beach	−0.756***	−0.758***	−0.762***	−0.754***	−0.799***	−1.064***	−0.829***	−0.532***
	(0.0525)	(0.0553)	(0.0536)	(0.0523)	(0.0510)	(0.0584)	(0.0639)	(0.0471)
Art and culture	−1.050***	−1.043***	−1.051***	−1.045***	−1.194***	−1.207***	−1.076***	−0.963***
	(0.0530)	(0.0544)	(0.0538)	(0.0525)	(0.0558)	(0.0474)	(0.0647)	(0.0562)
Other personal reasons	−0.724***	−0.726***	−0.739***	−0.715***	−0.788***	−0.593***	−0.685***	−0.698***
	(0.0356)	(0.0358)	(0.0358)	(0.0359)	(0.0361)	(0.0353)	(0.0797)	(0.0371)
$d_{\mathit{\text{Covid}}19}$ × ${\mathit{\text{accommodation}}}_{\mathit{op}}$
Hotels/hostels	−1.287***	−1.305***	−1.290***	−1.296***	−1.074***	−1.432***	−1.626***	−1.221***
	(0.0749)	(0.0770)	(0.0756)	(0.0746)	(0.0763)	(0.0613)	(0.105)	(0.0713)
Camping/farmhouse	−0.124	−0.106	−0.115	−0.125	−0.128	0.764***	−0.855***	−0.480***
	(0.130)	(0.136)	(0.132)	(0.131)	(0.126)	(0.142)	(0.163)	(0.123)
Others	−0.571***	−0.576***	−0.568***	−0.580***	−0.439***	−0.667***	−0.904***	−0.509***
	(0.0420)	(0.0427)	(0.0423)	(0.0422)	(0.0425)	(0.0378)	(0.0608)	(0.0408)
$d_{\mathit{Cov}19}$ × ${\mathit{\text{airplane}}}_{\mathit{op}}$	−0.279***	−0.266***	−0.278***	−0.278***	−0.309***	−0.144***	−0.378***	−0.329***
	(0.0239)	(0.0255)	(0.0241)	(0.0240)	(0.0226)	(0.0144)	(0.0276)	(0.0232)
FE country#prov#week	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
FE country#time	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Pseudo $R^2$/adj.$R^2$	.99	.99	.99	.99	.99	.90	.99	.99
Observations	323,747	323,747	323,747	323,747	323,747	227,702	153,337	300,244

Note: The table reports estimates of the model 1 over the period January 2019–February 2021, for different specifications of the variable measuring contagion at local level (columns 1–5). Column (6) report estimates of the model rewritten in log form and estimated with OLS. Column (7) restricts the sample to the first 40 provinces. Column (8) excludes the first months of 2021. The reference categories for the variable interacted with a dummy for the COVID‐19 period are, respectively, ‘business‐reasons’ for the purpose of travel, and ‘rented house or private house’ for the accommodation. The reference category for ‘airplane’ (short for public means of transport) is private means of transport (like own car). Standard errors, in parenthesis, are clustered by province–time and country of departure–time. Stars (***, ** and *) indicate statistical significance at 1, 5 and 10 per cent level, respectively. Fixed effects by country of departure–province–week (${\alpha }_{opw}$) and country of departure–time (${\alpha }_{ot}$) are always included.

Table A2 shows the estimates of our model in Equation (1) in which only EU countries, Schengen members and the United Kingdom are included in the analysis. Travellers from these countries were allowed to enter Italy for tourism after June 3rd without quarantine requirements (unlike other countries), and they accounted for most of inbound tourism to Italy in our sample period. We obtain almost identical results. Finally, we estimate our model excluding travellers from countries sharing a common border with Italy to remove the impact of cross‐border workers. Again, we obtain similar results overall and an even larger coefficient on the contagion variable (Table A3).

4.2. Analysis by country of departure

In this section we shift our focus to the variation of incoming tourism flows by country of departure of the tourists. To do so, as explained in Section 3.2, we drop our Country–Time fixed effects and we augment our model with the variables described in Section 3 (Equation 2). We estimate the model over two sets of countries: the entire sample of 40 countries (Table 5) and the sub‐sample of ‘passport‐free’ countries (EU and Schengen Area member countries, and the United Kingdom, whose citizens from 3 June onward were allowed to enter Italy for touristic reasons without almost any quarantine requirements). This sub‐sample includes the countries that account for most of the inbound tourism in our period of analysis and that faced very similar restrictions, which makes them more comparable. 25

TABLE 5.

Analysis by country of departure: All countries.

	All	All	$1\mathrm{st}$ wave	Summer	$2\mathrm{nd}$ wave
	(1)	(2)	(3)	(4)	(5)
${\mathit{\text{Cases}}}_{\mathit{ot}-1}$	0.0169*	0.00979	0.368***	0.432***	0.00726
	(0.00907)	(0.00866)	(0.0805)	(0.0487)	(0.00833)
$d_{\mathit{Cov}19}$ × $\ln {\left(\text{dist}\right)}_o$	−0.0625	−0.0610	0.247***	−0.531***	−0.129**
	(0.0479)	(0.0454)	(0.0486)	(0.0724)	(0.0527)
$d_{\mathit{Cov}19}$ × $d_{\mathit{\text{border}}o}$	−0.00724	−0.00521	−0.263***	−0.00598	0.0451
	(0.0286)	(0.0293)	(0.0672)	(0.0389)	(0.0427)
${\mathit{\text{Quarantine}}}_{\mathit{ot},\mathit{IT}}$	−0.893***	−0.904***	−0.669**	−1.206***	−0.502***
	(0.0892)	(0.0898)	(0.272)	(0.135)	(0.0886)
${\mathit{\text{Necessity only}}}_{\mathit{ot},\mathit{IT}}$	−0.967***	−1.079***		−1.964***	−0.616***
	(0.196)	(0.205)		(0.189)	(0.202)
${\mathit{\text{Swab}}}_{\mathit{ot},\mathit{IT}}$	−0.278***	−0.325***		−0.925***	−0.170***
	(0.0474)	(0.0463)		(0.0958)	(0.0492)
${\mathit{\text{Stringency}}}_{\mathit{ot}}$	0.00829***	0.00398***	0.00993***	0.0152***	−0.000165
	(0.00121)	(0.00150)	(0.00304)	(0.00264)	(0.00203)
$d_{\mathit{\text{stayathome}}}$		0.187***	−0.0982	−0.124*	0.283***
		(0.0384)	(0.0829)	(0.0748)	(0.0510)
$d_{\mathit{\text{noreg}}.\mathit{\text{movement}}}$		0.146***	0.192**	0.218***	0.0981**
		(0.0325)	(0.0787)	(0.0495)	(0.0394)
Travel restr. in $o$ (1)		−0.244***
		(0.0442)
Travel restr. in $o$ (2)		−0.240***
		(0.0694)
Travel restr. in $o$ (3)		−0.278***
		(0.0623)
Travel restr. in $o$ (4)		−0.681***
		(0.0805)
FE country#prov#week	Yes	Yes	Yes	Yes	Yes
FE prov#time	Yes	Yes	Yes	Yes	Yes
Pseudo $R^2$	.98	.98	.98	.99	.98
Observations	325,880	325,880	79,492	88,506	134,324

Note: The table presents estimates of the model 2 over different periods for the first 40 countries in terms of tourism receipts to Italy. The $1\mathrm{st}$ wave period (column 1) includes the weeks from 25 February to 2 June 2020. The Summer period (column 2) goes from 3 June to 15 September 2020. The $2\mathrm{nd}$ wave, column (5), goes from 16 September 2020 onward. Variable ${\mathit{\text{cases}}}_{\mathit{ot}-1}$ was winsorised at the 1%–99% per cent level to mitigate possible measurement errors and outliers (there are cases in the original dataset where the number of new cases is negative). The model includes the variables $X_{\mathit{op}}$ (coefficients not shown), namely ${d_{\mathit{\text{COVID}}\mathit{19}}}^{*}\left({\beta }_1^{\prime }{\mathit{\text{Purpose}}}_{\mathit{op}}+{\beta }_2^{\prime }{\mathit{\text{Accommodation}}}_{\mathit{op}}+{\beta }_3^{\prime }{\mathit{\text{Transport}}}_{\mathit{op}}\right)$. Standard errors, in parenthesis, are clustered by province–time and country of departure–time. ***, ** and * indicate statistical significance at 1, 5 and 10 per cent, respectively. Fixed effects by country of departure–province–week (${\alpha }_{\mathit{opw}}$) and province–time (${\alpha }_{\mathit{ot}}$) are always included.

Column (1) in Table 5 indicates that, unsurprisingly, the most important variables in explaining cross‐country variation in the presence of foreign tourists are related to the strictness of the bilateral travel restrictions imposed by Italy. The coefficient of the dummy variable Quarantine _ot,IT (which takes value 1 if there is either a mandatory quarantine period for tourists coming from that country, or if entry for leisure tourism is forbidden), implies a reduction in tourist presence by about 60 per cent larger than what are recorded by countries not subject to this requirement. The relative drop in international tourism is even more dramatic when entry was allowed only for urgent/essential reasons. On the contrary, screening measures at entry (e.g. swab tests) cause a substantially milder reduction in entry flows: the coefficient of the dummy for swab test requirement indicates a 20 per cent decrease in tourism flows. In fact, the coefficient of the swab test requirement is not statistically different from zero when we limit the analysis to EU and Schengen countries (and United Kingdom; Table 6), suggesting that this type of screening could limit the international spread of contagion without significantly hampering inbound tourism flows.

TABLE 6.

Analysis by country of departure: EU, Schengen members and UK.

	All	All	$1\mathrm{st}$ wave	Summer	$2\mathrm{nd}$ wave
	(1)	(2)	(3)	(4)	(5)
${\mathit{\text{Cases}}}_{\mathit{ot}-1}$	−0.0115	−0.0426***	0.0949	0.275***	−0.0379***
	(0.00790)	(0.0116)	(0.0861)	(0.0486)	(0.0105)
$d_{\mathit{Cov}19}$ × $\ln {\left(\text{dist}\right)}_o$	−0.176***	−0.278***	0.620***	−0.787***	−0.0352
	(0.0490)	(0.0533)	(0.107)	(0.0789)	(0.0706)
$d_{\mathit{Cov}19}$ × $d_{\mathit{\text{border}}\mathit{o}}$	−0.00435	0.00419	−0.174**	−0.0734*	0.0866*
	(0.0288)	(0.0305)	(0.0725)	(0.0384)	(0.0447)
${\mathit{\text{Quarantine}}}_{\mathit{ot},\mathit{IT}}$
${\mathit{\text{Necessity only}}}_{\mathit{ot},\mathit{IT}}$
${\mathit{\text{Swab}}}_{\mathit{ot},\mathit{IT}}$	−0.0526	−0.0436		−0.763***	0.130***
	(0.0359)	(0.0446)		(0.0754)	(0.0438)
${\mathit{\text{Stringency}}}_{\mathit{ot}}$	0.00992***	0.0139***	0.0160***	0.0184***	0.00652***
	(0.00131)	(0.00173)	(0.00349)	(0.00264)	(0.00251)
$d_{\mathit{\text{stayathome}}}$		0.0723*	−0.118	−0.108*	0.154***
		(0.0394)	(0.0816)	(0.0632)	(0.0574)
$d_{\mathit{\text{noreg}}.\mathit{\text{movement}}}$		0.0530	0.181**	0.154***	−0.0597
		(0.0332)	(0.0778)	(0.0475)	(0.0431)
Travel restr. in $o$ (1)		−0.311***
		(0.0519)
Travel restr. in $o$ (2)		−0.715***
		(0.0668)
Travel restr. in $o$ (3)		−0.770***
		(0.0641)
Travel restr. in $o$ (4)		−0.548***
		(0.0894)
FE country#prov#week	Yes	Yes	Yes	Yes	Yes
FE prov#time	Yes	Yes	Yes	Yes	Yes
Pseudo $R^2$	.98	.99	.98	.99	.98
Observations	237,421	219,356	58,016	62,394	66,670

Note: The table presents estimates of the model 2 over different periods for EU countries, Schengen countries plus United Kingdom. These were the only countries for which after the first wave visits for holiday tourism were allowed without the need to quarantine and accounted for about two‐thirds of total tourism receipts in 2020. The $1\mathrm{st}$ wave period (column 1) includes the weeks from 25 February to 2 June 2020. The Summer period (column 2) goes from 3 June to 15 September 2020. The $2\mathrm{nd}$ wave, column (5), goes from 16 September 2020 onward. Variable ${\mathit{\text{cases}}}_{\mathit{ot}-1}$ was winsorised at the 1%–99% per cent level to mitigate possible measurement errors (there are cases in the original database where the number of new cases is negative). The model includes the variables $X_{\mathit{op}}$ (coefficients not shown), namely ${d_{\mathit{\text{COVID}}19}}^{*}\left({\beta }_1^{\prime }{\mathit{\text{Purpose}}}_{\mathit{op}}+{\beta }_2^{\prime }{\mathit{\text{Accommodation}}}_{\mathit{op}}+{\beta }_3^{\prime }{\mathit{\text{Transport}}}_{\mathit{op}}\right)$. Standard errors, in parenthesis, are clustered by province–time and country of departure–time. ***, **, and *Statistical significance at 1, 5 and 10 per cent, respectively. Fixed effects by country of departure–province–week (${\alpha }_{\mathit{opw}}$) and province–time (${\alpha }_{\mathit{ot}}$) are always included.

A second result is related to the impact of distance. Our model includes the interaction between a dummy for the COVID‐19 period and the share of visitors that used to arrive at the local destination by plane or other public means of transport, which displayed a negative and statistically significant coefficient, pointing to the renewed importance of distance during the pandemic. We further add a variable measuring bilateral distance between Italy and the country of departure and a dummy for bordering countries. We find that distance also had an additional negative effect on the number of foreign travellers when we consider European countries, in particular during the summer months, suggesting that tourists preferred closer destinations, ceteris paribus. This remains true, with the only exception of the first wave, if we exclude countries bordering with Italy (Table A4).

While we had clear priors about the coefficients of the above‐mentioned variables, we had ambiguous expectations about the effect of contagion and stringency measures in the country of departure. On one hand, an increase of COVID‐19 cases in the home country of the tourists may induce them to raise caution and curb their plans to travel abroad, given the uncertainty of the health situation at home. By the same token, a tightening in containment measures in the home country may produce a similar effect, also in consideration that future stronger containment policies may hinder the travel on the way back home or make it more costly (e.g. because of reduced number of flights). On the other hand, a surge of positive cases at home may push the tourist to travel abroad (if the destination is perceived as ‘safer’) to minimise contagion risk during holidays and/or avoid domestic restrictions (substitution effect).

As regards the new COVID‐19 cases variable, our results are inconclusive: the coefficient is not consistently different from zero in the whole sample (columns 1 and 2) in Table 5, while it turns out to be negative on the sub‐sample of European countries (Table 6, columns 2 and 5). Moreover, the coefficient is positive during the summer months while negative or not statistically different from zero afterward. 26 The coefficient of the stringency index is instead more stable, as we find consistent positive estimates over the whole sample (column 1 in Tables 5 and 6). The sign of the stringency index coefficient remains positive even if we separately introduce dummies that control for mobility restrictions at home (column 2). 27

A possible relevant source of cross‐country variation that we are not controlling for in column (1) stems from travel restrictions to outbound tourism in the countries of origin. Unfortunately, we do not have information on these restrictions. As a proxy remedy to this concern, we include a categorical variable from the Oxford database (Hale et al., 2021) that measures the strictness of travel restrictions to inbound tourism in the tourist's home, as we assume that the restrictions to outbound tourism are generally symmetric with restrictions on inbound tourism, as suggested by anecdotal evidence observed for the Italian case. Our assumption seems validated, as we find that the introduction of these measures is negatively associated with a reduction in the number of arrivals, but their inclusion does not alter our results (column 2).

5. CONCLUDING REMARKS

In this paper we analysed inbound tourism to Italy during the COVID‐19 pandemic, looking at variation across Italian provinces of destination as well as across travellers' countries of origin. To this end, we relied on unique mobile phone data about the weekly number of foreign visitors in Italy, broken down by Italian province of stay and by visitors' nationality, for a period going from January 2019 to February 2021.

Our first result is that there is a negative and statistically significant relationship between the flow of foreign travellers in a given province and the local epidemiological situation, even after controlling for restrictive measures at the national and regional level. In other words, tourists appear to have paid a lot of attention to the risk of contagion not only at the national level (as somewhat expected), but also at the local destination level, and they make their travel plans accordingly. The resulting policy implication is that revamping international tourism flows during an epidemic is not simply a matter of lifting restrictions, but it also requires a substantial reduction of contagion risk, at least until the immunisation of the population reduces the health risks associated with getting ill with COVID‐19. With this regard, we can expect the negative elasticity of tourism flows to contagion to be sensibly reduced by progress in the vaccination campaigns.

Our second related result is that, since the start of the pandemic, provinces specialised in art tourism were hit the most, while provinces with a more prevalent orientation to business tourism proved to be significantly more resilient. Furthermore, provinces that used to be more ‘hotel intensive’ in terms of accommodation choices made by visitors were hit harder than provinces characterised by a larger use of private housing and/or rental houses. Finally, we also found that arrivals to local destinations more easily reachable by private means of transport (such as cars) decreased significantly less. This evidence is overall consistent with the hypothesis that contagion risk significantly affects not only tourists' decisions to travel but also how to travel and where to stay, thus implying heterogeneous effects across local destinations. Some local destinations appear to have suffered a larger fall in international arrivals, because they were perceived as ‘riskier’, given their local characteristics. Therefore, well‐diversified accommodation facilities and travel infrastructures enhance the resilience of a touristic destination to this type of adverse shocks.

Thirdly, we found that the different degrees in intensity and extension of entry restrictions across countries were key factors in explaining cross‐country patterns in international arrivals. However, screening requirements for incoming visitors (such as swab tests) do not seem to significantly discourage arrivals. Screening upon entry may thus be considered by policymakers an effective tool to reconcile the need to contain the expansion of the epidemic with the need to mitigate its impact on tourism flows. We also observed that arrivals from more remote European countries decreased comparatively more, even after controlling for entry restrictions and excluding neighbouring countries, pointing to the increased importance of distance in affecting tourists' choices during a pandemic.

ACKNOWLEDGEMENTS

The authors wish to thank Silvia Fabiani, Stefano Federico, Fadi Hassan, Alfonso Rosolia, Simonetta Zappa, Alessandro Borin, the editor and two anonymous referees for useful comments and suggestions on a previous version of this paper, while retaining full responsibility for all remaining errors and omissions. The views expressed in this study are those of the authors and do not involve the responsibility of the Bank of Italy.

APPENDIX 1.

DATA ANNEX

In this section, we provide further details on the data we used.

Mobile phone data

The total number of foreign SIM cards in Italy was calculated by the mobile operator based on roaming data. The cell network coverage of our provider is very large, so the number of foreign SIM cards detected is also large. In practice, however, not all foreign SIM cards are captured by this network, because there are also other Italian mobile phone operators offering roaming services to foreign SIM cards. In order to overcome this issue and estimate the total population of foreign SIM cards in Italy, our provider added also an estimate of the number of SIM cards roaming on other competing Italian networks, based on proprietary commercial data and market shares calibration. While we do not have access to their methodology, we could verify that their final data are consistent with BISIT data for the period common to the two data sources, and the two time series show very similar dynamics (Figure A1). This suggests that mobile phone data provides a good tracker for inbound tourism flows, supporting the use of this source for the analysis.

FIGURE A1.

Number of inbound travellers (indices: August 2019 = 100).

A SIM card (Subscriber Identity Module) contains an integrated circuit that encodes the subscriber's identity and the nationality of the operating company that has issued the card. We take this information as a proxy for the residency of the card owner (i.e. country of departure). This approximation is good as far as phone users resort to resident mobile companies. This may not be the case for migrants, as mentioned in Section 3.1, since they may prefer SIM cards issued in their home country instead of cards issued in their host country, to call their relatives at home at cheaper prices. For this reason, we excluded foreign SIM cards issued by countries associated with large immigrant communities in Italy.

As for the location, foreign SIM cards were attributed to Italian provinces based on the ‘cells’ (i.e. mobile phone antenna towers) they were connected with. If a SIM is detected in more than one province on the same day, it is assigned to the province where it was detected for a longer time. The Italian data protection legislation does not allow the diffusion of information derived from mobile phone data referring to less than 15 individual users. Therefore, if the three dimensions day, country of origin and province of destination are populated by 15 or less observations, the phone operator set the province of destination equal to ‘non specified’. The impact of this censoring on the data used in the paper is however quite low: the share of SIMs in the ‘undisclosed’ provinces was about 1.5 per cent in 2019 (2.5 in 2020). Moreover, those SIMs are prevalently associated with relatively ‘small’ countries, that were already excluded from the analysis for the reasons specified in the sub‐section List of countries included.

BISIT data

The Bank of Italy Survey on International Tourism (BISIT) is based on two pillars: (i) counting the number of travellers that enter/leave the country at a selected number of border crossing points, and (ii) conducting interviews with a sample of international travellers, both residents and not residents, crossing the Italian borders. The counting process aims at estimating the reference universe (i.e. the total number of inbound and outbound travellers), broken down by country of residence or destination, while the survey collects information about tourists' expenditure and their personal characteristics.

The BISIT survey asks the surveyed traveller to specify the reason for her trip to Italy choosing one among the possible answers: (A) personal reasons (it includes: A1 holidays and leisure; A2 Studying; A3 Pilgrimage or other religious reasons; A4 health or thermal tourism; A5 honeymoon; A6 visiting relatives and/or friends; A7 shopping; A8 other personal reasons). (B) Business reasons. (C) Transit only. If the respondent chooses A1, she is invited to further specify if it was holidays A1.1 at the beach; A1.2 on the mountains; A1.3 at the lake; A1.4 in a città d'arte (city of art); A1.5 green holidays; A1.6 sport and fitness holidays; A1.7 wine & food holidays. The complete questionnaire form can be downloaded from the Bank of Italy website section on international tourism statistics.

List of countries included

For readers' information, we list here (according to the alphabetical order of their ISO code) the 40 countries of origin included in our sample: Argentina, Austria, Australia, Bosnia and Herzegovina, Belgium, Brazil, Belarus, Canada, Switzerland, Chile, Czech Republic, Germany, Denmark, Spain, Finland, France, Great Britain, Greece, Croatia, Hungary, Ireland, Israel, Japan, Lithuania, Luxembourg, Latvia, Malta, Mexico, Netherlands, Norway, New Zealand, Poland, Portugal, Russia, Saudi Arabia, Sweden, Slovenia, Slovakia, Turkey and the United States. From this selection we excluded the Principality of Monaco (as it was not identifiable using mobile phone data) and the countries with large foreign resident communities, namely: Roumania, Bulgaria, Colombia, China, Serbia, Ukraine, Albania, India, Macedonia, and Moldova.

In guiding our choice, we adopted a simple quantitative criterion based on the ratio between the number of foreign residents living in Italy by country of origin (from the National Statistical Agency; ISTAT) and the number of travellers from the same country in 2019 (from the BISIT). We excluded the countries for which this ratio exceeds 10 per cent, because for such countries data may be distorted by the travels and the foreign SIMS owned by foreign residents living in Italy. Since the choice of the 10 per cent threshold is somewhat discretionary, we checked that lowering the threshold to 5 per cent leaves the results of the analysis substantially unchanged.

Table A1 reports some statistics on their weight on total inbound tourism, both in terms of night spent and in terms of total travellers, and a comparison between BISIT data and mobile phone data.

TABLE A1.

Weight of included countries in terms of inbound tourism to Italy.

	BISIT travellers		BISIT nights		Daily SIMs
	2017–2019		2017–2019		2019
	Number (thous.)	Avg. share	Number (thous.)	Avg. share	Number (thous.)	Avg. share
First 40 countries (net of excluded countries)	88,093	94.1	353,159	91.5	409,067	86.3
Excluded countries a	3945	4.2	20,306	5.3	46,576	9.8
Other countries	1591	1.7	12,612	3.3	18,326	3.9
Total	93,629	100.0	386,077	100.0	473,969	100.0

Source: BISIT and mobile phone data.

^a Countries with a large community resident in Italy: Poland, Romania, Bulgaria, Bosnia and Herzegovina, China, Serbia, Ukraine, Albania and Moldavia.

ROBUSTNESS ANALYSIS

TABLE A2.

Analysis by province: EU, Schengen members and UK.

Dep. variable: ${\mathit{\text{Tourists}}}_{\mathit{opt}}$
	All	All	$1\mathrm{st}$ wave	Summer	$2\mathrm{nd}$ wave
	(1)	(2)	(3)	(4)	(5)	(6)	(7)
${\mathit{\text{cases}}}_{\mathit{pt}-1}$	−0.0578***	−0.0721***	0.0410*		−0.0635***	−0.0594***	−0.0421***
	(0.00906)	(0.00880)	(0.0225)		(0.00712)	(0.00678)	(0.00344)
$d_{\mathit{Cov}19}$ × ${\mathit{\text{purpose}}}_{\mathit{op}}$
Nature and beach		−0.486***	−1.122***	−0.392***	−0.465***	−0.469***	−0.524***
		(0.0342)	(0.124)	(0.0439)	(0.0452)	(0.0452)	(0.0410)
Art and culture		−1.021***	−1.472***	−0.868***	−1.108***	−1.118***	−1.137***
		(0.0358)	(0.109)	(0.0480)	(0.0514)	(0.0513)	(0.0416)
Other pers. reasons		−0.356***	−0.192***	−0.268***	−0.473***	−0.477***	−0.341***
		(0.0264)	(0.0625)	(0.0352)	(0.0364)	(0.0362)	(0.0270)
$d_{\mathit{Cov}19}$ × ${\mathit{\text{accommodation}}}_{\mathit{op}}$
Hotels/hostels		−0.592***	−0.755***	−0.345***	−0.973***	−0.977***	−0.437***
		(0.0303)	(0.0678)	(0.0388)	(0.0398)	(0.0395)	(0.0235)
Camping/farmhouse		−0.198***	−0.326***	−0.462***	0.364***	0.359***	0.222***
		(0.0461)	(0.107)	(0.0506)	(0.0718)	(0.0714)	(0.0456)
Others		−0.373***	−0.342***	−0.212***	−0.599***	−0.604***	−0.358***
$d_{\mathit{Cov}19}$ × ${\mathit{\text{airplane}}}_{\mathit{op}}$		−0.327***	−0.315***	−0.384***	−0.215***	−0.214***	−0.453***
		(0.0256)	(0.0659)	(0.0369)	(0.0364)	(0.0365)	(0.0213)
RR index						−0.00620**
						(0.00285)
FE country#prov#week	Yes	Yes	Yes	Yes	Yes	Yes	Yes
FE country#time	Yes	Yes	Yes	Yes	Yes	Yes	Yes
FE reg.#time	No	No	No	No	No	No	Yes
Pseudo $R^2$	.99	.99	.99	.99	.98	.98	.99
Observations	255,793	237,421	58,016	62,394	114,642	114,642	114,642

Note: The table reports estimates of the model 1 over different periods. Column (1) and (2) look at the whole sample (January 2019–February 2021). The 1st wave period (column 3) includes the weeks from 25 February to 2 June 2020. The Summer period (column 4) goes from 3 June to 15 September 2020. The 2nd wave (columns 5–7) goes from 16 September 2020 onward. The reference categories for the variable interacted with a dummy for the COVID‐19 period are, respectively, ‘business‐reasons’ for the purpose of travel, and ‘rented house or private house’ for the accommodation. The reference category for ‘airplane’ (short for public means of transport) is private means of transport (like own car). Standard errors, in parenthesis, are clustered by province–time and country of departure–time. ***, ** and *Statistical significance at 1, 5, and 10 per cent, respectively. Fixed effects by country of departure – province – week (${\alpha }_{\mathit{opw}}$) and country of departure – time (${\alpha }_{\mathit{ot}}$) are always included.

TABLE A3.

Analysis by province: Exclusion of countries with a border in common with Italy.

Dep. variable: ${\mathit{\text{Tourists}}}_{\mathit{opt}}$
	All	All	$1\mathrm{st}$ wave	Summer	$2\mathrm{nd}$ wave
	(1)	(2)	(3)	(4)	(5)	(6)	(7)
${\mathit{\text{cases}}}_{\mathit{pt}-1}$	−0.0664***	−0.0814***	0.0266		−0.0760***	−0.0713***	−0.0433***
	(0.0109)	(0.0106)	(0.0248)		(0.00918)	(0.00841)	(0.00414)
$d_{\mathit{Cov}19}$ × ${\mathit{\text{purpose}}}_{\mathit{op}}$
Nature and beach		−0.639***	−1.117***	−0.552***	−0.567***	−0.567***	−0.589***
		(0.0376)	(0.115)	(0.0502)	(0.0530)	(0.0523)	(0.0443)
Art and culture		−1.191***	−1.506***	−1.063***	−1.177***	−1.188***	−1.119***
		(0.0388)	(0.105)	(0.0569)	(0.0555)	(0.0557)	(0.0448)
Other pers. reasons		−0.475***	−0.326***	−0.364***	−0.586***	−0.589***	−0.324***
		(0.0282)	(0.0599)	(0.0348)	(0.0453)	(0.0454)	(0.0268)
$d_{\mathit{Cov}19}$ × ${\mathit{\text{accommodation}}}_{\mathit{op}}$
Hotels/hostels		−0.822***	−1.188***	−0.517***	−1.131***	−1.137***	−0.579***
		(0.0348)	(0.0859)	(0.0432)	(0.0466)	(0.0460)	(0.0249)
Camping/farmhouse		−0.370***	−0.868***	−0.688***	0.356***	0.349***	0.257***
		(0.0580)	(0.157)	(0.0624)	(0.0791)	(0.0779)	(0.0534)
Others		−0.621***	−0.704***	−0.549***	−0.667***	−0.675***	−0.429***
$d_{\mathit{Cov}19}$ × ${\mathit{\text{airplane}}}_{\mathit{op}}$		−0.403***	−0.403***	−0.496***	−0.244***	−0.242***	−0.538***
		(0.0292)	(0.0749)	(0.0359)	(0.0527)	(0.0530)	(0.0296)
RR index						−0.00771*
						(0.00424)
FE country#prov#week	Yes	Yes	Yes	Yes	Yes	Yes	Yes
FE country#time	Yes	Yes	Yes	Yes	Yes	Yes	Yes
FE reg.#time	No	No	No	No	No	No	Yes
Pseudo $R^2$	.98	.99	.99	.99	.98	.98	.99
Observations	305,042	281,004	68,572	76,762	114,486	114,486	114,486

Note: The table reports estimates of the model 1 over different periods but excluding countries of origin that share a common border with Italy, namely Austria, France, Switzerland and Slovenia. Columns (1) and (2) look at the whole sample (January 2019–February 2021). The 1st wave period (column 3) includes the weeks from 25 February to 2 June 2020. The Summer period (column 4) goes from 3 June to 15 September 2020. The 2nd wave (columns 5–7) goes from 16 September 2020 onward. Standard errors, in parenthesis, are clustered by province–time and country of departure–time. The reference categories for the variable interacted with a dummy for the COVID‐19 period are, respectively, ‘business‐reasons’ for the purpose of travel, and ‘rented house or private house’ for the accommodation. The reference category for ‘airplane’ (short for public means of transport) is private means of transport (like own car). ***, ** and *Statistical significance at 1, 5 and 10 per cent, respectively. Fixed effects by country of departure – province – week (${\alpha }_{\mathit{opw}}$) and country of departure – time (${\alpha }_{\mathit{ot}}$) are always included.

TABLE A4.

Analysis by countries of origin: Exclusion of countries with a border in common with Italy.

	All	All	$1\mathrm{st}$ wave	Summer	$2\mathrm{nd}$ wave
	(1)	(2)	(3)	(4)	(5)
${\mathit{\text{Cases}}}_{\mathit{ot}-1}$	0.0522***	0.0459***	0.368***	0.433***	0.0456***
	(0.00752)	(0.00755)	(0.0870)	(0.0501)	(0.00754)
$d_{\mathit{Cov}19}$ × $\ln {\left(\text{dist}\right)}_o$	−0.0981**	−0.0968**	0.233***	−0.460***	−0.217***
	(0.0449)	(0.0403)	(0.0486)	(0.0774)	(0.0504)
${\mathit{\text{Quarantine}}}_{\mathit{ot},\mathit{IT}}$	−0.935***	−0.826***	−0.606**	−1.260***	−0.531***
	(0.0799)	(0.0784)	(0.270)	(0.138)	(0.0799)
${\mathit{\text{Necessity only}}}_{\mathit{ot},\mathit{IT}}$	−0.994***	−0.902***		−2.032***	−0.546***
	(0.181)	(0.189)		(0.195)	(0.174)
${\mathit{\text{Swab}}}_{\mathit{ot},\mathit{IT}}$	−0.533***	−0.537***		−0.942***	−0.466***
	(0.0413)	(0.0396)		(0.0934)	(0.0524)
${\mathit{\text{Stringency}}}_{\mathit{ot}}$	0.0109***	0.00872***	0.00416	0.0169***	−0.00151
	(0.00134)	(0.00172)	(0.00326)	(0.00287)	(0.00227)
$d_{\mathit{\text{stayathome}}}$		−0.0497	−0.0604	−0.160**	−0.0658
		(0.0403)	(0.0881)	(0.0733)	(0.0493)
$d_{\mathit{\text{noreg}}.\mathit{\text{movement}}}$		0.139***	0.0515	0.179***	0.295***
		(0.0364)	(0.0806)	(0.0503)	(0.0509)
Travel restr. in $o$ (1)		−0.110*
		(0.0623)
Travel restr. in $o$ (2)		−0.0419
		(0.0699)
Travel restr. in $o$ (3)		−0.159***
		(0.0576)
Travel restr. in $o$ (4)		−0.753***
		(0.0750)
FE country#prov#week	Yes	Yes	Yes	Yes	Yes
FE prov#time	Yes	Yes	Yes	Yes	Yes
Pseudo $R^2$	.98	.98	.98	.99	.98
Observations	283,627	283,627	69,020	77,798	116,000

Note: The table reports estimates of the model 2 over different periods but excluding countries of origin that share a common border with Italy, namely Austria, France, Switzerland and Slovenia. Columns (1) and (2) look at the whole sample (January 2019–February 2021). The 1st wave period (column 3) includes the weeks from 25 February to 2 June 2020. The Summer period (column 4) goes from 3 June to 15 September 2020. The 2nd wave (columns 5–7) goes from 16 September 2020 onward. Variable ${\mathit{\text{cases}}}_{\mathit{ot}-1}$ was winsorised at the 1%–99% per cent level to mitigate possible measurement errors (there are cases in the original database where the number of new cases is negative). The model includes the variables $X_{\mathit{op}}$ (coefficients not shown), namely ${d_{\mathit{\text{COVID}}19}}^{*}\left({\beta }_1^{\prime }{\mathit{\text{Purpose}}}_{\mathit{op}}+{\beta }_2^{\prime }{\mathit{\text{Accommodation}}}_{\mathit{op}}+{\beta }_3^{\prime }{\mathit{\text{Transport}}}_{\mathit{op}}\right)$. Standard errors, in parenthesis, are clustered by province–time and country of departure–time. ***, ** and *Statistical significance at 1, 5, and 10 per cent, respectively. Fixed effects by country of departure – province – week (${\alpha }_{\mathit{opw}}$) and province – time (${\alpha }_{\mathit{pt}}$) are always included.

VARIANCE DECOMPOSITION

As discussed in Section 3.2, our empirical approach relies on fixed effects to achieve a clean identification of the variation explained by our set of independent variables. In this respect, this section shows a comparison exercise on the amount of variance that our models can explain, with a view of assessing the relative importance of variables by country of origin and by province. We do this exercise by incrementally adding variables and fixed effects to a model and looking at the square of the correlation between our dependent variable and its fitted values. This is conceptually equivalent to looking at the R2 in the case of a linear model. Results are reported in Table A5.

TABLE A5.

A variance decomposition.

	(1)	(2)	(3)	(4)	(5)
Explained variance (%)	83.8	95.4	97.3	99	98.7
Residual variance explained (%)	–	71.7	83.4	94	92.1
FE orig#prov#week	Yes	Yes	Yes	Yes	Yes
FE time	No	Yes	Yes	No	No
$X_{\mathit{pt}}$	No	No	Yes	Yes	No
$X_{\mathit{ot}}$	No	No	Yes	No	Yes
$d_{\mathit{\text{covid}}19}\#X_{\mathit{op}}$	No	No	Yes	Yes	Yes
FE orig # time	No	No	No	Yes	No
FE prov # time	No	No	No	No	Yes

Note: The table shows the variance explained by several models with different sets of fixed effects and variables. The explained variance is the square of the correlation between fitted values and observed values. The residual variance is computed as the share of variance in addition to model (1), taken as a reference term. $X_{\mathit{pt}}$ and $d_{\mathit{\text{COVID}}19}\#X_{\mathit{op}}$ are the variables included in Equation (1); $X_{\mathit{ot}}$ are the variables added in Equation (2).

As a first comparison term, we compute this statistic for a model in which we only include the fixed effects ${\alpha }_{\mathit{opw}}$ (column 1). These fixed effects, as discussed in Section 3.2, control for all factors that render a province more attractive for tourists from a specific country, as well as for possible seasonal patterns in these relationships. This simple model alone can explain about 84% of the total variation of the data, leaving 16% residual variance, exhaustively capturing the gravity structure in our tourism data. We then add to the model the time fixed effects ${\alpha }_t$ (column 2). They capture the effect of time‐varying shocks that affect all Italian destinations and flows from all countries of origin in the same way. As expected, this model explains a large share of the residual variance (about 70%), clearly reflecting the nature of COVID‐19 as a common shock that hit international tourism flows. The residual 30% is the variation during the pandemic that was country or destination specific and which is the focus of this paper. In column (3) we thus report the same statistic as we add to the model all our explanatory variables. Overall, augmenting the model with our variables lead to a significant improvement in terms of explained variance (by about 12%).

We then look at the explanatory power of our variables along a specific dimension (province versus country of origin), controlling for the other with fixed effects. In particular, we first include country‐of‐origin fixed effects, leaving the province‐time variation explained by our variables (column 4). This model explains 94% of residual variance. Adding province–time fixed effects to the model in column (3) leads to a similar accounting, as it raises the explained variance to 92% (column 5). To sum up, this evidence suggests that province characteristics and country factors played a comparable role in explaining heterogeneous patterns at the country–province level during the pandemic.

Della Corte, V. , Doria, C. , & Oddo, G. (2023). The impact of COVID‐19 on international tourism flows to Italy: Evidence from mobile phone data. The World Economy, 00, 1–24. 10.1111/twec.13380

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available upon request from the corresponding author, with the only exception of mobile phone data, as they were purchased from a private mobile phone company and cannot be made publicly available due to ownership restrictions. Data from the survey on international tourism in Italy (see Section Data annex in Appendix 1 for details) can be freely downloaded from the official website of the Bank of Italy in the following section: Home/Statistics/External transactions and positions/International tourism/Distribution of microdata.