Social capital and patients’ mobility in Italy

ABSTRACT Using a 10-year (2006–15) regional dataset on hospital discharges, we estimate the determinants of Italian regional outflow rates while also including three proxies for social capital: the quality of friendships, the involvement in social activities and the ratio of blood donors to the population. We find a significant push effect from the lack of social capital intended as generalized expectation of cooperative behaviour (proxied by blood donors); this does not hold true for the other two proxies of social capital. The lack of a cultural context where norms of reciprocity matter may reduce regions’ ability to retain their patients.


INTRODUCTION
In healthcare systems that allow citizens to choose where to receive treatment (e.g., in Denmark, the UK, the Netherlands, Italy and Sweden), extra-regional patient mobility is often observed. The economic literature has widely examined to what extent this phenomenon is associated with patient, provider and area characteristics (for a systematic review, see Aggarwal et al., 2017). However, despite this interest, no one, to the best of our knowledge, has studied the effect of the regional endowment of social capital on patient mobility. Since the healthcare sector is characterized by a high level of asymmetric information and by principal-agent relations in which trust is crucial (Arrow, 1963), we believe the mechanism through which social capital possibly influences patients' decision to move from their regions of origin to seek hospital care is of great interest. Trust and social capital are closely related. Some authors consider social capital as pertaining essentially to individuals (Glaeser et al., 2002), while others view it as a feature of the community. Guiso et al. (2008) define social capital as a 'good' culture, that is, a set of individual beliefs and values that enables cooperation among the members of a community. Inasmuch as social capital reduces information asymmetries and the probability of moral hazard in the healthcare sector (Folland, 2018;Rocco et al., 2014), patients living in regions characterized by higher social capital in terms of attitude towards 'civicness' and pro-social behaviour might be more likely to prefer local hospital care systems. This study aims at investigating whether regions' ability to retain their patients also depends on the local endowment of social capital.
We do not build a gravity model on bilateral patient flows such as described in the existing literature (i.e., Balia et al., 2014Balia et al., , 2018Fabbri & Robone, 2010;Fattore et al., 2014). Instead, we propose a model that estimates the determinants of regional outflow rates including health structure variables, demographic and economic variables and proxies for social capital. We focus on the concept of social capital put forth by Putnam et al. (1994) in their seminal work, which considers three dimensions of social capital: 'interpersonal trust', 'active participation in public affairs' and 'generalized expectations of cooperative behaviour'. Since each dimension may affect information asymmetries between patients and hospitals in a different way, we investigate each one separately. In this paper we use the following proxies to measure the three dimensions of social capital: the quality of friendship relations (accounting for interpersonal trust), the involvement in social activities (for active participation) and the number of blood donors (for generalized expectation of cooperative behaviour).
Using a 10-year (2006-15) regional dataset on hospital discharges, we examine the drivers of outflow rates in Italian regions. Italy represents an interesting case study. First, in 2001, a reform introduced 21 separate and autonomous regional health services, allowing patients to choose freely where to be treated. Since then, Italy registers a relevant degree of interregional patient mobility. Second, since official hospital quality measures specifically for 'common citizens' are not yet easily available to the public in Italy and quality measures published by the Health Ministry are too technical, it is very likely that information on the quality of Italian hospitals still spreads mainly through informal networks and social interactions. Finally, Italy represents an exemplary case study in the literature on social capital (e.g., Banfield, 1958;Guiso et al., 2004;Putnam et al., 1994). Putnam et al.'s (1994) seminal work on social capital analyses the effect of social trust on the efficiency of local governments in Italy because the country was, and is still, characterized by relevant differences among regions in both the variables. This also makes Italy an interesting case study to analyse the effect of social capital on patient choice to travel to other regions to receive hospital care.
The extant literature on hospital choice has examined in great depth the key roles of both distance and quality. Balia et al. (2020) find that, in Italy, patients' choices are sensitive to variations in hospital quality when hospitals are located several hundred kilometres away; however, this does not hold true when local hospitals are considered. In light of their findings, we examine separately the determinants of total regional outflow rates from those of outflow rates directed toward bordering regions. Moreover, we analyse whether the occurrence of patient migration from an area is affected by one or more features in neighbouring regions. We apply a simple spatial model, the SLX model (Le Sage & Pace, 2009), that adds spatially lagged variables to the non-spatial model.
This study offers a valuable contribution to the ongoing debate on the effectiveness of patient empowerment through free mobility as a tool for enhancing competition between regional healthcare systems. According to the literature, competition should stimulate quality levelling and, therefore, in the long run it should reduce interregional mobility (Brekke et al., 2012;Brekke et al., 2014). However, for Italy, data show that since the 2001 reform, interregional patient mobility remains high. Berta et al. (2016) find that in Italy the choice of a specific hospital depends significantly on the number of people who lived in the same area who had previously chosen the hospital. They suggest that asymmetric information may act as a barrier for competition to work effectively. As stated above, social capital may reduce information asymmetries and the probability of moral hazard in the healthcare sector. Our results confirm that while high mobility across Italian regions may be explained by relevant differences in the quality of regional healthcare services, there is also a significant push effect from the lack of social capital, intended as generalized expectation of cooperative behaviour.
The paper is organized as follows: The next section gives a review of the literature related to social capital and to the main determinants of patient mobility. The third and fourth sections describe the empirical framework and the data, respectively. The fifth section provides results for the analyses. Finally, the sixth section concludes.

THE RELEVANT LITERATURE
2.1. Social capital Several aspects of social capital, such as trust between community members and local officials, knowledge and information sharing, and/or greater participation in voluntary organizations, are shown to contribute in improving health governance and service delivery (e.g., Harphman et al., 2002;James et al., 2001). In the presence of asymmetric information, the choice of moving for hospital care is likely to depend on social capital. In particular, referring to the three dimensions of social capital identified in the introduction, we may argue that: . 'Interpersonal trust', resulting from intense social interaction, may provide easier access to relevant information concerning health (Berkman & Glass, 2000). In turn, this may reduce ex-ante asymmetric information, that is, the probability of buying a service that is below the expected quality. The presence of friends signalling the best hospitals in a region on the basis of their experience may therefore reduce regional outflows. We must highlight that they could also become a 'push factor', whenever the quality of local care is low, by pointing out the relatively low quality of local hospitals. 1 Moreover, interpersonal trust could reduce outflow rates when it is likely to provide informal healthcare and psychological support (Murgai et al., 2002). Market or public healthcare systems do not usually provide such services, so mechanisms of informal assistance are agreed upon between neighbours or friends. This support tends to arise only in a context of reciprocal trust, as there is no enforceable contract guaranteeing obligations. Interpersonal trust may be proxied by the quality of friendships. . 'Active participation in public affairs' may make people's lobbying efforts and coordination to obtain health-enhancing goods and services more successful (Kawachi et al., 1997;Mellor & Milyo, 2005). Communities characterized by higher civic engagement could therefore be more likely to have high-quality healthcare providers. In the presence of ex-ante asymmetric information, patients may prefer to remain in those communities that are relatively more endowed with this kind of social capital because they believe that they are likely to benefit from high-quality health providers. Active participation in public affairs is proxied by participation in non-profit associations. . 'Generalized expectations of cooperative behaviour' and norms of reciprocity may be important because they reduce the probability of moral hazard. In particular, in the presence of ex-post asymmetric information, when it is impossible to monitor the behaviour of the agent, principals prefer communities where agents are likely to behave consistently with the prevailing 'good' culture. Most of the measures used in the literature are outcome-based, and the majority result from good law enforcement rather than from a high level of social capital. The number of blood donors has the advantage of being an outcome-based measure of social capital that does not undergo this critique given that in Italy, there is no incentive (neither legal nor economic) encouraging blood donation; blood donation remains completely free as national regulations do not allow any form of compensation (Guiso et al., 2004). In addition, data provided by AVIS (the Italian National Association of Blood Donors) show that blood donation centres are equally accessible and evenly distributed across provinces and regions in the country (as documented by Crescenzi et al., 2013).

The determinants of patient mobility
In countries such as Italy, where healthcare prices are regulated by central and local governments and patients are free to choose where to be treated, hospitals can compete only through quality. In the literature, results highlight that a higher degree of competition is likely to produce better quality (Faber et al., 2009;Marshall et al., 2004). As documented by Aggarwal et al. (2017), patient mobility should represent an effective driver for improving provider performance. According to Berwick et al. (2003) and Le Grand (2009), even movements concerning only 5-10% of patients may provide the necessary incentive to improve quality. Nevertheless, a methodological challenge faced when using patient mobility as a proxy for patient choice is to separate the impact based on differences in hospital quality, from that of other factors. In fact, patient outflows may be influenced by physicians' preferences (Ringard, 2010) or by insufficient local supply. Furthermore, patient choice cannot foster quality levelling in a scenario of asymmetric information. Hospital rankings are available in some countries, such as in the US or UK, but not yet in Italy. The so-called Programma Nazionale Esiti (PNE), which was first made available in 2012, provides outcome quality indicators for the Italian healthcare sector but the PNE's institutional objective, as declared, is not to criticize hospitals' performance or to create hospital rankings. Rather, it aims at providing experts with a scientific tool of evaluation and monitoring. In other words, PNE data are not for the primary use of 'common citizens'. It follows that, as mentioned in the Introduction, in Italy information on hospital quality still spreads mainly through informal networks (word of mouth, the experience reported by friends or neighbours, etc.).
PNE indicators may be divided between 'straight' outcome indicators, measuring the result of a healthcare process in terms of its clinical outcome (mortality, morbidity, etc.), and process indicators, measuring the adherence of a care process to reference standards of the best evidencebased clinical practice. Among process indicators published by PNE, one of proven clinical effectiveness is the surgical treatment of femoral neck fractures in elderly patients carried out within the first 48 h of the trauma. A femoral neck fracture is one of the most common traumatic injuries in elderly patients and is associated with high rates of mortality and functional loss. Timely intervention significantly reduces both mortality and post-surgical complications. Therefore, the percentage of patients over 65 with a femur neck fracture receiving surgical treatment within the first two days may be an effective index to verify the quality of a hospital or of a healthcare organization. Outcome indicators such as surgical treatment of femoral neck fractures in elderly patients can be easily deduced by listening to other people's experiences. In the present framework, it may represent a good indicator to measure quality.
We are interested in quality at a regional level, which can be described not only by 'output/outcome' indicators, but also by 'input' levels that refer to hospitals' structural variables, such as the number of beds (a proxy for hospital capacity in a region) and/or the technological endowment. Studying the Italian case, Levaggi and Zanola (2004) and Balia et al. (2018) find that a higher public hospital capacity discourages outflows, and Fabbri and Robone (2010) show that technology endowment in healthcare does not affect patient outflow. Still at a regional level, related to quality, we find in the Ministero della Salute database the comparative index of performance (CIP) and the case mix index (CMI) as proxies for the efficiency of the healthcare sector. Higher values of the CIP, indicating some inefficiency in managing the length of stays and longer waiting times, may be associated with higher outflows (in this case outflows may be the result of the insufficient local supply and not of patient choice). We might have a negative relationship between CIP and outflows if patients perceive longer stays as an insurance against bad health at home. At a regional level, the CMI can be viewed as an index for specialization in cases with higher resource intensity. Patients could prefer to remain in their regional healthcare system because it is more specialized in highly complex cases (signalled by a higher CMI); however, specialization could also contribute to increasing outflows whenever patients are put off by long waiting lists and look for less complex care in other regions (again, outflows may result from insufficient local supply and not from patient choice). Balia et al. (2018) find that higher CIP and CMI levels in Italy are associated with higher outflows.
Among the other variables that can in some way influence the decision to migrate for health reasons are economic and demographic variables, such as per capita income, regional agglomeration and the ratio of the aged population. Regarding income, higher GDP per capita in the region of origin could lead to more mobility since people can afford travel expenses. On the other hand, considering that the Italian healthcare system is mainly regionally funded, higher income at the origin may also mean higher quality hospitals and better services, discouraging patients from moving. Results from Italian data show that patients move from the poorer regions to the richer ones (Balia et al., 2014;Fabbri & Robone, 2010;Fattore et al., 2014). Another important variable could be the population density, since people living in less urbanized areas may move to areas providing a wider range of services. Frailer population groups, such as those over 75 years of age, have a higher demand for health services and hence could migrate more. However, patients belonging to older age groups may also be less likely to move outside their region because they may be less prone to travel (Levaggi & Zanola, 2004).

EMPIRICAL FRAMEWORK
In order to evaluate the determinants of patients' outflow rates, we perform an econometric analysis using a panel dataset consisting of 20 Italian regions (currently 21 NUTS-II territorial units where data relative to Trentino-Alto Adige are taken separately for the two provinces of Trento and Bolzano) during the period of 2006-15. The dependent variable is the outflow rate (i.e., the passive mobility ratio), while explanatory variables are divided among proxies measuring the main aspects of social capital, variables representing the quality of the regional healthcare system, and economic and demographic factors.
First, in order to gain an initial insight into the factors affecting patients moving from one region to another, we estimate a pooled ordinary least squares (OLS) regression with clustered standard errors. We are well aware of the limitations of this approach compared with panel data models: intercepts are forced to be the same for all regions and the pooled OLS estimator is not consistent when the relationship between the two variables under investigation is governed by a third omitted variable. Second, we address the panel structure of the data by estimating a fixed effect (FE) model. Third, due to the presence of two time-invariant explanatory variables, 2 we move to a random effect (RE) model (Wooldridge, 2011).
Finally, in order to account for unobserved heterogeneity, we estimate a correlated random effects (CRE) model in which group means of variables, which vary within groups, are added to regressors (the so-called Mundlak correction). We introduce this correction in order to combine the advantages of both fixed and REs. This technique was put forth by Mundlak (1978) as a way to relax the assumption of the RE estimator according to which the observed variables are uncorrelated to the unobserved variables (Bell & Jones, 2015;Mavromaras et al., 2013). The introduction of means should capture the correlation between unobserved heterogeneity and the covariates that could make the RE model inconsistent. In the context we are studying, the efficiency of public administration for instance, is a possible latent factor of heterogeneity across regions. Indeed, organizational efficiency in the public sector is quite difficult to measure: the choice of the most reliable empirical measurement method is still an object of controversy (Smith & Street, 2005). Moreover, the efficiency of public administration is a latent factor most likely correlated to some of the independent variables in our specification, especially those proxying social capital (Andrews, 2012).
The CRE final specification is as follows: with: where x it defines the time-varying explanatory variables for region i at time t; z i represents the two constantover-time explanatory variables for region i; x i denotes the regional-level mean of the time-varying covariates; and u it is the error term. Furthermore, in order to infer the possible presence of spatial interaction effects, we adopt a simple spatially lagged regression model, the SLX model (Le Sage & Pace, 2009), by adding spatial lags of the exogenous variables in the original RE and CRE specifications. As described by Elhorst and Vega (2015) and pointed out by Gibbons and Overman (2012), the SLX model is the simplest among the spatial models used to consider local spatial spillover effects. Furthermore, the SLX supersedes some identification issues that can be typical of an alternative model, such as the spatial Durbin model (SDM).
A crucial element of any model involving spatial lags is the building of the spatial weight matrix. We rely on a distance-based matrix assuming that the intensity of interactions depends on the distance between regions. A distance matrix can be defined on the basis of various indicators, depending both on the definition of distance (great circle distance, driving distance, etc.) and on the functional form chosen (the inverse of the distance, the inverse of the squared distance, etc.). Finally, a distance cut-off beyond which spatial interactions are negligible must be identified. Following a common practice in the literature (e.g., Dall'erba & Le Gallo, 2008), we use the great circle distance between regional centroids. In particular, each element of the spatial weight matrix is defined as follows: where w ij is an element of the row standardized weight matrix W (row-standardized spatially lagged variables represent an average across bordering regions), d ij is the great circle distance between centroids of regions i and j, k defines the functional form and D is the cut-off parameter above which spatial interactions are assumed to be negligible. In our specification we take the inverse of the squared distance (k ¼ 2) and we choose the median distance as a cut-off. 3 The complete SLX CRE is as follows: where h it , x it , z i , x i and u it are the same as in equation (1) and W is the spatial weight matrix. In order to have a complete overview, we then estimate an FE model including  (1) and (4). Table 1 reports definitions, descriptive statistics and sources for all the variables included in the estimations.

DATA
In the first specification, the dependent variable is the regional patients outflow rate (outflow), given by the percentage of residents hospitalized in other regions over the total number of residents in a given region admitted to hospital during the year (coming from both the same and other regions).
The SDO database (Archivio Nazionale Schede di Dimissione Ospedaliera) published online by the Health Ministry does not distinguish between patient mobility and emergency medical care. Therefore, we cannot eliminate any non-elective care due to accidents that occurred while patients were in another region for vacation or work. Still, we believe that this does not reduce the reliability of our results in a significant way. 4 We obtained disaggregated SDO data only relative to 2015. They show that the standard deviation of the percentage of residents hospitalized in a different region for emergencies is very small (0.03) and much lower than the standard deviation of the percentage of hospitalizations for emergencies in Italian regions (0.06). Its mean is only 18% compared with a mean of 27% of the regional percentage of hospitalizations for emergencies. We must outline that the latter ranges from a minimum of 18% in the virtuous region of Lombardy, to a maximum of 39% in Campania, one of the less attractive regional healthcare systems in Italy. These data suggest that in some regions more than in others, Italian patients are likely to use an emergency to bypass administrative red tape and delays to be hospitalized. Therefore, non-elective care might include many patients who are intentionally 'choosing' to be cured in a specific region'. Balia et al. (2020) also include in their analysis of Italian patient mobility non-elective admissions to verify whether their main findings, obtained considering only elective care, display a more general validity. They obtain similar results in both analyses.
In a second specification, the dependent variable is the outflow rate towards bordering regions (bordoutflow). The rate is given by the percentage of residents hospitalized in bordering regions over the total number of residents in a given region who are admitted to hospital during the year. 5 As specified in the previous section, the explanatory variables in our panel dataset are social capital variables, healthcare system variables and demographic and economic variables: . Social capital variables: we consider 'interpersonal trust' (friend), proxied by the percentage of householders aged over 14 stating that they are very satisfied with their friendships; 'active participation' (socpart), proxied by the percentage of householders aged over 14 stating that they were involved in a social activity during the last year; 'generalized cooperative behaviour' (blood), proxied by the percentage of blood donors out of the population. The data source for the variables friend and socpart is the 'Aspects of daily life Survey' of ISTAT, a large annual sample survey that covers the resident population in private households, by interviewing a sample of 20,000 households and 50,000 people. It provides population estimates for the main topics of daily life and behaviours. The data source for the variable blood is the 'National and regional register of blood and plasma' filled out by every Transfusion Service published by the ISTISAN (Istituto Superiore di Sanità), the Italian National Institute of Health. . Healthcare system variables: We consider the average percentage of patients over 65 with a fracture of the femur neck operated on within 48 h of admission to hospital during 2010-15 (PNE) as an index of the quality of hospital healthcare. 6 Among the structural variables, we include the number of beds (beds), the variables CIP and CMI and the ratio of Magnetic Resonances with resolution higher than 0.5 teslas over total Magnetic Resonances in 2016 (MRquality). 7 Finally, given that during the period under consideration, regions with persistent deficits in the area of healthcare were obliged to reduce them (the so-called Piani di Rientro), we include a dummy variable (Plan) equal to 1 for every year of the recovery plan, under the assumption that the implied expenditure cuts may push local patients to move to other regions. MRquality and PNE are the two time-invariant variables in our panel dataset (z i in equations 1 and 4). . Demographic and economic variables: We consider per capita income (income), population/km 2 (popdens) and the ratio of total population aged over 75 (over75). Table 1 shows that the average outflow rate in Italy is much higher in the South than in the Centre or in the North of the country. The average regional outflow rate rose between 2006 and 2015, from 10.32% to 10.95%, but the situation is not the same in all Italian regions. Going from Northern regions (on the left) to Southern ones (on the right), Figure 1 shows decreasing or stable outflow rates (with the exception of Liguria, which experienced a significant increase, and Veneto) changing to rising ones (all the regions to the right of Umbria show growing outflow rates, with the exception of Basilicata, which is in any case characterized by a very high rate).
Tables A1 and A2 in the supplemental data online show that the flow of patients mainly occurs from the South to the North. The percentages of patients of each region hospitalized in non-bordering regions (indicated with a star in Tables A1 and A2 and mainly directed to Lombardia, Emilia-Romagna, Toscana and Lazio) are much higher in the South than in the rest of the country.
Data reported in Table 1 show that social capital is lower in the South than in the Centre and North for all the three considered dimensions (friend, socpart and blood). Figures A1 and A2 in the supplemental data online show that, while all Southern regions (with the exception of Sardinia) present lower values of friend and socpart than those of the Northern and Central ones, this is not true for the variable blood. Table A3 in the supplemental data online reports a relatively low correlation rate of the variable blood to variables friend and socpart (0.49 and 0.43, respectively) and a much higher correlation between the variables friend and socpart (0.82). Furthermore, the variable blood is not strongly correlated to hospital quality variables, ruling out the hypothesis that in regions with good health services, there is also a larger fraction of people choosing to become blood donors.
The healthcare system variables show that, whereas the average number of beds (beds) is much lower in the South than in the rest of the country (indicating that Southern regional healthcare systems are characterized, on average, by weaker infrastructures), the means of CIP and CMI in Southern regions are approximately 1, that is, they are 'in line' with the national mean. Concerning technology, the Magnetic Resonance quality (MRquality) seems even higher in the South than in the North and Centre of Italy. 8 The mean of the variable PNE, in contrast, is much lower in the South than in the rest of the country (see also Figure A2 in the supplemental data online). Southern regions are significantly poorer, younger, and less densely populated than Northern and Central ones. Table 2 reports the results of pooled OLS, FE, RE and CRE model estimates for variables outflow and bordoutflow. The most remarkable result emerging from the estimates of both dependent variables is that the coefficient of the variable blood is negative and significant in all models, while those of the variables friend and socpart are significant with a negative sign only in OLS. This result lends support to the hypothesis that the 'generalized expectation of cooperative behaviour', by reducing the probability of moral hazard, matters for reducing patient outflows. On the opposite, neither 'interpersonal trust' resulting from more intense social interactions nor 'active participation in public affairs', which were both expected to reduce exante asymmetric information on hospital quality, seem to affect patient migration choice significantly.

RESULTS
With respect to the healthcare system variables, FE, RE and CRE model estimates show a significant negative effect of the variable beds and a positive significant influence of CIP on the variable outflow. A possible interpretation of this is that a larger number of hospital beds and a lower CIP shorten waiting lists, thereby discouraging outflows. Similar results are found in Levaggi and Zanola (2004) and Balia et al. (2018). Moreover, all models show a positive and significant impact of CMI on the variable outflow, consistent with the results obtained by Balia et al. (2018). The quality of technology (MRquality) has a negative and significant impact on the variable outflow only in pooled OLS. PNE is significantly and negatively related to variable outflow in all models. Column 8 shows a significant positive effect of both the variables CPI and CMI on the variable bordoutflow. These results suggest that longer waiting lists push patients toward bordering regions to seek hospital care. In contrast, we do not find a significant effect of variables beds and PNE on bordoutflow: patients living in regions not endowed with large hospitals and with low PNE seem to be pushed to migrate toward distant regions.
With respect to economic and demographic variables, Table 2 shows that the older the population in the region, the higher the total outflow rate (this is supported by the FE, RE and CRE models). Furthermore, looking at the  results of the CRE model reported in Column 4, population density and income negatively affect outflow but are no longer significant when considering the determinants of bordoutflow. This result suggests that patients living in relatively poor and less populated regions are pushed to migrate to non-bordering regions. The fractional nature of the dependent variable could lead to some distortion in the results. In this case, as underlined by Wooldridge (1996, 2008) and by Wooldridge (2011), a linear model could present the same advantages and drawbacks of a linear probability model for a dichotomous dependent variable, that is, senseless predictions outside the unit interval. To handle this problem, and to confirm the robustness of our results, we apply both a logit and a log-odds transformation of the response variable outflow, with boundary values strictly between zero and one, obtaining results similar to those presented in Table 2. The same transformation is not possible for the dependent variable bordoutflow because of the presence of zero values. Therefore, we follow the CRE approach of Papke and Wooldridge (2008) to estimate fractional response models for panel data. Again, we obtain similar results. 9 The results of the spatial CRE model reported in Table 3 confirm that the older and less populated the region of origin, the higher the patient outflow rate. All the quality variables included in the estimates are significant and have the expected sign. The variable blood affects the outflow rate negatively and significantly. By looking at the spatially lagged variables, we find a negative and significant coefficient for spatially weighted CMI (for FE and CRE models). This result suggests that patients are more likely to migrate to less specialized nearby healthcare systems, consistent with the interpretation of the results presented in Table 2, showing that high CMI pushes patients to migrate toward bordering regions.
Furthermore, when we look at the CRE model results, we find that patients living in regions surrounded by poor and non-densely populated regions, lacking large hospitals, and not endowed with high-quality technological equipment tend to migrate more. These results require careful interpretation. A possible explanation is that patients residing in the South, especially in peripheral areas, tend to migrate more and to move to more distant regions endowed by significantly larger hospitals. In other words, possibly, in bordering regions they do not find healthcare systems that are better than the ones in their home region. This interpretation is consistent with the evidence reported in Table 2 showing that patients living in regions that are poor and without large hospitals do not migrate to bordering ones. We also find a significant push effect associated with the variable blood; being in a region surrounded by neighbours with a low endowment of this kind of social capital is a stimulus for migration. Taken together, our results suggest that patients living in regions characterized by a low expectation of cooperative behaviour migrate more to both close and distant regions ( Table 2) and that patients living in regions close to other ones characterized by a low expectation of cooperative behaviour tend to migrate more (Table 3), possibly to distant regions.
As previously discussed, the matrix W assigns a weight to the spatial effects between regions. The latter is based on geographical distance due to its unambiguous exogeneity (Anselin & Bera, 1998;Anselin et al., 1996). In the case of Italy, this raises the question of the two big islands, Sicilia and Sardinia. For islands, indeed, simple contiguity matrices are not applied to rule out the case of a weight matrix W presenting rows and columns with only zeros. We take this issue into account by estimating the SLX model (FE, RE and CRE) without Sicilia and Sardinia and we obtain robust results. 10

CONCLUSIONS
In the last two decades several European countries have implemented free patient choice in their healthcare systems on the basis of the assumption that hospitals competing for patients would be pushed to improve quality. Several studies have investigated the determinants of interregional patient mobility with the aim of evaluating the impact of healthcare quality on patient migration. This paper highlights the importance of investigating how patient choice to move to other regions depends on local social capital in a contest characterized by a relevant degree of asymmetric information. Our results show, both in the non-spatial and spatial analysis, that regional outflow rates fall together with the ratio of blood donors to the population. Per capita blood donation is widely recognized as an accurate latent proxy of social capital (e.g., Guiso et al., 2004Guiso et al., , 2008Guiso et al., , 2011Nannicini et al., 2013), specifically to measure the 'generalized expectation of cooperative behaviour'.
Surprisingly, the other two dimensions of social capital, that is, 'interpersonal trust', proxied by the quality of friendships, and 'active participation', proxied by the involvement in social activity, are not relevant. It seems that what is crucial in determining the decision of a patient to remain in his/her regional healthcare system is the respect of norms of reciprocity, an element that reduces the probability of moral hazard. The reduction of exante information asymmetries on hospital quality, coming from the signals of good friends or from the belief that communities characterized by higher civic engagement are more likely to provide high-quality healthcare systems, does not seem to affect the choice of patients to migrate. This result is interesting, since in Italy information on the quality of healthcare is still channelled mainly through informal networks. The treatment of a femoral neck fracture, a traumatic injury that is very common among elderly patients, is the kind of information that patients are likely to collect through informal networks. We find that the PNE outcome indicator relative to the timely surgical treatment of femoral neck fractures negatively affects the outflow of patients. Furthermore, patients tend to remain in their own region whenever it is endowed with large hospitals and it is efficient in managing hospital length of stays. A deeper investigation of Italian patient migration shows that patients living in regions that are poor and not endowed with large hospitals are more likely to migrate to distant regions. This may raise some concerns since Balia et al. (2020) find that only young and highly educated individuals are likely to react to lower quality by taking long distance journeys, while the frailer population might remain captive to the local services.
In conclusion, our results show that the choice of a patient to trust and remain in his/her regional healthcare system does not depend only on the latter's structural aspects, but also on the local community's cultural features. The effects of providers' policies addressed at influencing where patients decide to receive treatment may be attenuated because people fear moral hazard. Further research is required to understand the role of trust as a driver of patient mobility; a big shock such as that caused by the COVID-19 pandemic in 2020 is likely to affect Italian regions' social capital as well as the trust in regional hospital care providers. It will be very interesting to investigate if and how the 2020 and 2021 shocks will affect patient regional mobility and its determinants in the future.

DISCLOSURE STATEMENT
No potential conflict of interest was reported by the authors. Note: Number of regional units ¼ 21; total number of observations ¼ 210.Standard errors are shown in parentheses. *p < 0.10, **p < 0.05, ***p < 0.01.
Social capital and patients' mobility in Italy 917