TY - JOUR
AU1 - Mastrobuoni,, Giovanni
AB - Abstract This study uses declassified data on US mafia members of the 1950s and 1960s to estimate the criminal network effect on their economic status. I measure economic status exploiting detailed information about their place of residence. Housing values are reconstructed using current deflated transactions data. I deal with non‐random sampling of mobsters modelling investigations on connections as Markov chains. Reverse causality between economic status and the gangster’s position in the network is solved exploiting exogenous exposure to pre‐immigration connections. A standard deviation increase in closeness centrality increases economic status by between one‐quarter and three‐quarters of a standard deviation. In January 2011, exactly 50 years after Robert F. Kennedy’s first concentrated attack on the American mafia as the newly appointed attorney general of the US, nearly 125 people were arrested on federal charges, leading to what federal officials called the ‘largest mob roundup in FBI history’.1 Over the last 50 years the mafia has continued following the same rules and is still active in many countries, including the US.2 Despite this, the illicit nature of organised crime activities has precluded empirical analysis and the literature has overwhelmingly been anecdotal or theoretical (Reuter, 1994; Williams, 2001). This study uses declassified data on 800 mafia members, who were active just before the 1961 crackdown, to study the importance of criminal connections inside such a secret society (linking the network position of mobsters to an economic measure of their success). The records are based on an exact facsimile of the Federal Bureau of Narcotics (FBN) secret files on American mafia members in 1960 (MAF, 2007).3 The data contain information collected from FBN agents on the gangsters’ closest criminal associates, which I use to reconstruct the criminal network.4 Connections are believed to be the building blocks of secret societies and of organised crime groups, including the mafia. Francisco Costiglia, alias Frank Costello, a mafia boss who according to the data was connected to 34 gangsters, would say ‘he is connected’ to describe someone’s affiliation to the mafia (Wolf and DiMona, 1974). Indeed, the first rule in mafia’s decalogue states that ‘No one can present himself directly to another of our friends. There must be a third person to do it’, who knows both affiliates (Maas, 1968).5 As a consequence, gangsters who are on average closer to all the other gangsters need fewer interconnecting associates to expand their network. While gangsters who bridge connections across separate clusters of the network can maintain such monopoly power. There are measures of importance of members inside networks that are based on the average closeness and on the bridging capacity, called closeness centrality and betweenness centrality.6 But even just the number of connections, known as degree centrality, might be important to reach leadership positions, as in the mafia these are not simply inherited. Soldiers elect their bosses using secret ballots (Falcone and Padovani, 1991, p. 101).7 Three main empirical challenges emerge when estimating how a gangster’s network centrality influences his economic prospects, called the network effect: (i) the measurement of economic prospects in the absence of information about illegal proceeds; (ii) the non‐random; and (iii) endogenous nature of the network. Regarding the first issue, since illicit transactions and criminal proceeds inside the mafia are unobservable, I use the value of the house or the apartment where such criminals presumably resided (or nearby housing) to measure their economic success. Such value is reconstructed based on the deflated value of the current selling price of their housing based on the internet site Zillow.com. Prices are deflated using the metropolitan statistical areas’ (MSAs) average housing values from Gyourko et al. (2013). Given that most mobsters who were active in 1960 were born into very poor families (Lupo, 2009), the value of the house where they resided, whether it was owned or rented, is arguably a reasonable measure of their illegal proceeds,8 though reconstructing the original value is certainly prone to error.9 Regarding the second issue, in the 1960s, the total estimated number of mafia members was around 5,000 (Maas, 1968). Since almost all high‐ranking members have a record, the 800 criminal profiles are clearly a potentially non‐representative sample of mafia members. To deal with the incompleteness and non‐randomness of the network, in Section 1 I model law enforcement’s surveillance and detection of mafia network nodes (mobsters) as a Markov chain (Mastrobuoni and Patacchini, 2012). The final issue about the potential endogeneity of the network requires a longer discussion. Sparrow (1991) and Coles (2001) propose the use of network analysis to study criminal networks, however, apart from some event studies based on a handful of connections, empirical evidence on criminal networks is scarce and never addresses the potential endogeneity of the network.10 In non‐experimental settings the variation that identifies the effect of networks may be partly driven by homophily (the tendency of individuals to be linked to others with similar characteristics), or unobserved characteristics which determine someone’s position in the network as well as his or her outcomes. Mobsters might, for example, use their (unobserved) wealth to build connections and buy more expensive housing. Since real networks can hardly be generated entirely through an intervention, there are three ways to estimate (causal) network effects:11 (i) modelling sequentially the network formation and the network effects (Chandrasekhar and Lewis, 2011); (ii) experimenting with networks;12 and (iii) instrumenting the position in the network.13 ,14 Since connections, their number, as well as their quality, are potentially even more important in a world without enforceable contracts, a world where secrecy, reputation and violence prevail, such bonds are even more likely to be endogenous. Several factors might influence the decision to connect and do business with another gangster. When gangsters expand their network they are trading off the increased risk of whistleblowing with increased criminal proceeds (Bonanno, 1983). Criminal hierarchies, kinships, complementarity and substitutability in criminal as well as non‐criminal activities are just some of the factors that are likely to influence the gangster’s decision to expand his network and, thus, his network centrality. Instead of modelling the entire network formation mechanism, I rely on an instrumental variable that influences mobsters’ centrality in the mafia network. Such instrument is based on information collected from the pre‐immigration communities (Munshi, 2003), that the mobsters left several decades before I observe their network and their housing wealth in 1960. The identification strategy is based on advantages in building connections that originate from the gangsters’ or from his ancestors’ place of origin (Italy). Conditional on the region of birth, a detailed description of their legal and illegal activities and other individual characteristics of the mobsters these innate connections should not influence mobsters’ housing wealth in the US (other than through such connections).15 While this exclusion restriction remains untestable, I address potential pitfalls; mainly endogenous migration driven by successful mobsters and direct effects of potential innate connections on housing values. Several factors might have helped building more connections when the gangster’s ancestors were concentrated in more traditional mafia territories: (i) increased trust that spills over from known and reputable families;16 (ii) easier punishments through left‐behind kinships; and (iii) knowledge about the rules and traditions of the secret society (e.g. omertá, which is a vow of silence). Without detailed information on the communities of origin for all the members born in the US, I use the informative content of surnames on the place of origin, more commonly known as isonomy. Such a measure of potential innate interactions predicts the gangster’s individual number and quality of connections. Exploiting differences in isonomy between Southern and Northern Italy, I perform falsification tests of the first stage regressions which validate the instrument’s capability to proxy for the place of origin. The instrument is weak, though I show that restricting the analysis to gangsters who have ever been arrested (making up 80% of the sample), for which surnames are arguably measured with greater precision, strengthens the instrument without changing the results. When using the instrument, a one standard deviation increase in network closeness centrality (the inverse of the average network distance from all other gangsters) increases housing value by three‐quarters of a standard deviation, with the p‐values on the endogeneity tests being usually close to 10%. The results are similar for eigenvector centrality (which is a function of the prestige of the connected members), while the results for degree (the simple count of connections) and betweenness centrality (the bridging capacity across different clusters of the network) tend to be weaker but for different reasons. On one hand, degree appears to be a crude measure of someone’s importance (the value of connections is increasing in the rank of the gangster). On the other hand, gangsters with high betweenness were more likely to be part of the commissione, the governing body of the mafia. These members of the mafia often kept a lower profile by living in more humble housing, which unresolvably biases the corresponding results downward. 1. Sampling and Estimating Network effects A quick look at record number one, Joe Bonanno (see the online Appendix Figure B2), reveals the kind of information that will be used to link a mobster’s network centrality to his economic success. According to the FBN he was born on 18 January 1905 in Castellamare (Sicily), and resided in 1847 East Elm Street in Tucson (Arizona). He had interests in three legal businesses: Grande Cheese Co., Fond du Lac (Wisconsin), Alliance Realty & Insurance (Tucson, Arizona) and Brunswick Laundry Service (Brooklyn, New York)e.tc.Finally, his closest criminal associates were Lucky Luciano, Francisco Costiglia (Frank Costello), Giuseppe Profaci, Anthony Corallo, Thomas Lucchese, and Carmine Galante. I use: (i) the value of the house where mobsters reside to measure economic success y (subsection 2.1.); (ii) information on their associates to reconstruct the network G(subsection 2.2.);17 (iii) the informational content of surnames to build the instrumental variable (subsection 2.4). But before analysing networks, particularly when such networks are hidden, it is important to take into account that the observed network Grepresents a subset (subgraph) of the entire network Gand not necessarily a random one. More formally, the goal is to model an economic outcome y as a linear function of network centrality c (abstracting for simplicity from other covariates), y=α+c(G)β+ϵ,(1) when the observed network Gis a subset of G. In general the measurement error that biases the estimate β^ will not be classical. Chandrasekhar and Lewis (2011) show that when c measures degree centrality, the network is exogenous, and the sampling is random (with sampling rate ψ), plimβ^=β·ψ−1· attenuation bias. The reason for the scaling factor ψ−1 is that under random sampling the expectation of c(G) is approximately equal to ψc(G). However, the observed mafia records are unlikely to represent a random sample of mobsters. The 800 criminal files come from an exact facsimile of a Federal Bureau of Narcotics report of which 50 copies were circulated within the Bureau starting in the 1950s. They come from more than 20 years of investigations and several successful infiltrations by undercover agents (McWilliams, 1990). The FBN data represent a snapshot of what the authorities knew in 1960 and thus do not contain exact information about the hierarchies within the organisation. Such information was revealed only a few years later, when Joe Valachi, a mafia associate, became the first FBN and later FBI informant.18 Joe Valachi’s testimony confirmed FBN’s view that the mafia had a pyramidal structure with connections leading towards every single member.19 Indeed, the observed network is connected (or ergodic), meaning that from each node (gangster) one can reach any other node. Moreover, it is a ‘small‐world’ network, as the average path length between gangsters is just 3.7 steps. Given the hierarchical structure of the mafia and the estimated 5,000 associates who were active during those years, the 800 gangsters are likely to be a non‐random sample of Cosa Nostra members. More active, more important, and more connected mobsters were certainly more likely to be noticed and tracked. Indeed, all known mafia bosses who were alive in 1960 are listed in the records. This means that the observed 800 gangsters are likely to be more connected than the average one and that part of the network is unobserved. There are no written records about how the FBN followed mobsters and constructed the network. Through surveillance posts and undercover agents, the agency was likely to have been discovering previously unknown mobsters by following known ones. Two surveillance photographs of Italian mobsters taken in 1980 and in 1988 show evidence of these patterns (see the online Appendix Figure B1). As a consequence more connected gangsters were more likely to be ‘sampled’ by the FBN. In order to produce a representative sample of mobsters, ‘sampling’ weights should underweigh highly connected mobsters and vice versa. This kind of sampling resembles a procedure that is used to sample hidden populations, called snowball sampling (Goodman, 1961; Granovetter, 1976; Frank, 1979; Snijders, 1992; Rothenberg, 1995). Let us assume the FBN starts observing one or more mobsters out of N, and that such known mobsters are indicated with the number one in the 1 × N vector of zeros and ones p0 ⁠, called the seed. Following the initial mobsters’ links the FBN will observe more and more mobsters. Starting with the N × N symmetric matrix A=[aij] with elements equal to 1 when mobsters i and j are connected and zero otherwise (called the adjacency matrix), one can obtain the transition matrix Tnormalising the columns to sum up to one. The element tij of Tmeasures the likelihood of discovering mobster j when mobster i is under surveillance. After k steps the likelihood of discovering mobster j is equal to the j‐th of the vector pk=p0Tk ⁠. The corresponding stationary distribution p, defined as a vector that does not change under application of the transition matrix, is independent of the seed, where p= pT.20 Element pi of the probability vector pcan be interpreted as the likelihood of observing gangster i if one randomly picked and followed an edge of the network. The resampling weights are thus equal to the inverse of such probability wi=1/pi ⁠, with 0<pi<1 ⁠. Since pi is almost proportional to the number of connections, such weights are quite intuitive. Gangsters with fewer connections, who are less likely to be spotted by the FBN and as a result are under‐represented, receive larger weights. The weighting corrects for the selection bias when describing the individual characteristics of the mobsters.21 Moreover, using a Monte Carlo simulation, Lee et al. (2006) show that under snowball sampling the scale bias is less than ψ−1 as, for the most central nodes, c(G) and c(G) coincide.22 Since exact re‐scaling would only be applicable to a degree, later I use logarithmic transformations of c to address the scaling bias ψ−1 and an instrumental variable approach to address the attenuation bias (as well as to address the potential endogeneity of the network). Expressing the variables of (1) in logarithms has several advantages. First, we will see that the densities of log (y) and log (c) are approximately normal. Second, it deals with the scaling bias. The (weighted) regression model in logarithms becomes: log(y)=α~+log[c(G)]β~+ϵ~,(2) where α~ includes − log (ψ). The summary statistics (Table 1) describe the individual characteristics as well as the network characteristics of the gangsters, with and without correcting for the non‐random sampling design. The average number of identical surnames (possibly ‘extended family members’) is about 1.5. Most gangsters migrated between the turn of the century and the 1930s (see Appendix A), though the records contain no information on when the gangsters, or their families, migrated to the US. About 70% of mobsters who were active in 1960 were born in the US, while the rest was split between Sicily (about 20%) and the rest of Italy (about 10%). In 1960 the average age of the associates is about 48, and 76% of them are married (10% are divorced). About 15% of the times the wife’s maiden name is shared with other gangsters, leading to a presumably ‘connected wife’. The average number of known children is about 1, while the average number of known siblings is 2. The FBN records contain also some information about the mobsters’ legal and illegal businesses. The average number of businesses is about 1. A little more than half of them are involved in drug dealing, but in line with long criminal careers (the average age at first arrest is 23) the average number of crime types listed in their records is close to 3. Only 14% of gangsters have no arrest record. For these gangsters some variables (e.g. the place of residence, or the exact surname) might be measured with more noise. Table 1 Summary Statistics . Unweighted . Weighted . . . . Mean . SD . Mean . SD . Min . Max . Individual characteristics Housing value 1960 397,678 805,186 378,196 770,101 5,286 7,866,203 Potential innate interaction 36.18 76.79 26.10 51.59 0 654.02 Housing built after 1960 0.10 0.30 0.14 0.35 0 1 Year built unknown 0.12 0.33 0.12 0.33 0 1 Extended family members 1.58 1.01 1.30 0.73 1 6 Born outside Italy (mainly US) 0.67 0.47 0.71 0.46 0 1 Born in Sicily 0.24 0.43 0.20 0.40 0 1 Age 48.63 7.59 48.20 7.57 24 72 Year of birth unknown 0.21 0.41 0.19 0.39 0 1 Height in feet 5.61 0.20 5.60 0.19 5 6.17 Weight in pounds 177.66 26.87 176.63 29.12 95 365 Age at first arrest 23.03 7.32 23.09 7.11 8 57 Never arrested 0.14 0.34 0.14 0.35 0 1 Connected wife 0.19 0.39 0.15 0.36 0 1 Married 0.80 0.40 0.76 0.43 0 1 Divorced 0.07 0.26 0.10 0.30 0 1 Number of children 1.07 1.47 1.06 1.44 0 8 Siblings 2.10 2.14 2.02 2.08 0 11 Types of crime committed 2.64 1.68 2.71 1.75 0 9 Number of businesses 1.08 1.01 1.00 0.94 0 5 Involved in drug dealing 0.56 0.50 0.59 0.49 0 1 Network characteristics Degree 11.10 9.28 6.21 5.52 1 71 Eigenvector (std.) 12.77 14.12 6.59 8.31 0.05 100 Betweenness (std.) 4.94 9.33 1.90 4.21 0 100 Closeness (std.) 75.12 8.76 69.78 8.27 50.51 100 . Unweighted . Weighted . . . . Mean . SD . Mean . SD . Min . Max . Individual characteristics Housing value 1960 397,678 805,186 378,196 770,101 5,286 7,866,203 Potential innate interaction 36.18 76.79 26.10 51.59 0 654.02 Housing built after 1960 0.10 0.30 0.14 0.35 0 1 Year built unknown 0.12 0.33 0.12 0.33 0 1 Extended family members 1.58 1.01 1.30 0.73 1 6 Born outside Italy (mainly US) 0.67 0.47 0.71 0.46 0 1 Born in Sicily 0.24 0.43 0.20 0.40 0 1 Age 48.63 7.59 48.20 7.57 24 72 Year of birth unknown 0.21 0.41 0.19 0.39 0 1 Height in feet 5.61 0.20 5.60 0.19 5 6.17 Weight in pounds 177.66 26.87 176.63 29.12 95 365 Age at first arrest 23.03 7.32 23.09 7.11 8 57 Never arrested 0.14 0.34 0.14 0.35 0 1 Connected wife 0.19 0.39 0.15 0.36 0 1 Married 0.80 0.40 0.76 0.43 0 1 Divorced 0.07 0.26 0.10 0.30 0 1 Number of children 1.07 1.47 1.06 1.44 0 8 Siblings 2.10 2.14 2.02 2.08 0 11 Types of crime committed 2.64 1.68 2.71 1.75 0 9 Number of businesses 1.08 1.01 1.00 0.94 0 5 Involved in drug dealing 0.56 0.50 0.59 0.49 0 1 Network characteristics Degree 11.10 9.28 6.21 5.52 1 71 Eigenvector (std.) 12.77 14.12 6.59 8.31 0.05 100 Betweenness (std.) 4.94 9.33 1.90 4.21 0 100 Closeness (std.) 75.12 8.76 69.78 8.27 50.51 100 Open in new tab Table 1 Summary Statistics . Unweighted . Weighted . . . . Mean . SD . Mean . SD . Min . Max . Individual characteristics Housing value 1960 397,678 805,186 378,196 770,101 5,286 7,866,203 Potential innate interaction 36.18 76.79 26.10 51.59 0 654.02 Housing built after 1960 0.10 0.30 0.14 0.35 0 1 Year built unknown 0.12 0.33 0.12 0.33 0 1 Extended family members 1.58 1.01 1.30 0.73 1 6 Born outside Italy (mainly US) 0.67 0.47 0.71 0.46 0 1 Born in Sicily 0.24 0.43 0.20 0.40 0 1 Age 48.63 7.59 48.20 7.57 24 72 Year of birth unknown 0.21 0.41 0.19 0.39 0 1 Height in feet 5.61 0.20 5.60 0.19 5 6.17 Weight in pounds 177.66 26.87 176.63 29.12 95 365 Age at first arrest 23.03 7.32 23.09 7.11 8 57 Never arrested 0.14 0.34 0.14 0.35 0 1 Connected wife 0.19 0.39 0.15 0.36 0 1 Married 0.80 0.40 0.76 0.43 0 1 Divorced 0.07 0.26 0.10 0.30 0 1 Number of children 1.07 1.47 1.06 1.44 0 8 Siblings 2.10 2.14 2.02 2.08 0 11 Types of crime committed 2.64 1.68 2.71 1.75 0 9 Number of businesses 1.08 1.01 1.00 0.94 0 5 Involved in drug dealing 0.56 0.50 0.59 0.49 0 1 Network characteristics Degree 11.10 9.28 6.21 5.52 1 71 Eigenvector (std.) 12.77 14.12 6.59 8.31 0.05 100 Betweenness (std.) 4.94 9.33 1.90 4.21 0 100 Closeness (std.) 75.12 8.76 69.78 8.27 50.51 100 . Unweighted . Weighted . . . . Mean . SD . Mean . SD . Min . Max . Individual characteristics Housing value 1960 397,678 805,186 378,196 770,101 5,286 7,866,203 Potential innate interaction 36.18 76.79 26.10 51.59 0 654.02 Housing built after 1960 0.10 0.30 0.14 0.35 0 1 Year built unknown 0.12 0.33 0.12 0.33 0 1 Extended family members 1.58 1.01 1.30 0.73 1 6 Born outside Italy (mainly US) 0.67 0.47 0.71 0.46 0 1 Born in Sicily 0.24 0.43 0.20 0.40 0 1 Age 48.63 7.59 48.20 7.57 24 72 Year of birth unknown 0.21 0.41 0.19 0.39 0 1 Height in feet 5.61 0.20 5.60 0.19 5 6.17 Weight in pounds 177.66 26.87 176.63 29.12 95 365 Age at first arrest 23.03 7.32 23.09 7.11 8 57 Never arrested 0.14 0.34 0.14 0.35 0 1 Connected wife 0.19 0.39 0.15 0.36 0 1 Married 0.80 0.40 0.76 0.43 0 1 Divorced 0.07 0.26 0.10 0.30 0 1 Number of children 1.07 1.47 1.06 1.44 0 8 Siblings 2.10 2.14 2.02 2.08 0 11 Types of crime committed 2.64 1.68 2.71 1.75 0 9 Number of businesses 1.08 1.01 1.00 0.94 0 5 Involved in drug dealing 0.56 0.50 0.59 0.49 0 1 Network characteristics Degree 11.10 9.28 6.21 5.52 1 71 Eigenvector (std.) 12.77 14.12 6.59 8.31 0.05 100 Betweenness (std.) 4.94 9.33 1.90 4.21 0 100 Closeness (std.) 75.12 8.76 69.78 8.27 50.51 100 Open in new tab Weighting introduces larger changes in average centrality. For instance, the average degree drops from about 11 to about six when weighting. The remaining centrality indices are divided by their maximum value (multiplied by 100). The housing characteristics and the instrumental variable ‘potential innate interactions’ are discussed later in subsections 2.1 and 2.4. 2. Descriptive Evidence 2.1. Housing Values and Number of Legal Businesses There is no database on individual housing values of 1960 properties,23 but feeding the exact residence address into Zillow.com produces 641 current real estate values and, for 561 homes (about 90% of the sample), there is also information on the year the house was built.24 The remaining 159 mobsters were not residing in the US anymore (such as Lucky Luciano, who had already been expelled from the country), or never lived in the US (most of these mobsters were living in Italy).25 Zillow’s estimated price is calculated from public data and from data submitted by users (e.g. real estate agents or appraisers physically inspect the home and take special features, location, and market conditions into account). Comparing historic estimates based on Zillow with the actual transaction prices of homes that sold, about 80–90% of the time the measurement error is within 20% of the sale price (more than half of the time it is within 10%).26 In order to deflate contemporary housing values to 1960 ones, I use the average housing value in 1960 and in 2000 taken from Gyourko et al. (2013) for the 608 homes that are in an MSA, while I use state‐level census data for the remaining 33 homes.27 The upper panel of Figure 1 shows the relationship (truncated at the 90th percentile) between the current and the 1960 values. Housing prices have approximately doubled over the last 50 years, though they have increased almost five times in San Francisco, while they stayed almost constant in Binghampton, Utica/Rome, or Buffalo. In the New York MSA, where almost 300 gangsters reside, prices doubled. Fig. 1. Open in new tabDownload slide Housing Value Deflators and Distribution Notes. Current housing prices (2010) are taken from Zillow.com. Deflators use MSA or state‐level housing values from the census (Gyourko et al., 2013). Fig. 1. Open in new tabDownload slide Housing Value Deflators and Distribution Notes. Current housing prices (2010) are taken from Zillow.com. Deflators use MSA or state‐level housing values from the census (Gyourko et al., 2013). The 5th, 10th, 25th, 50th, 75th, and 90th percentile of the housing value in 1960 were 39, 50, 95, 190, 325, and 662,000 dollars. The 95th and 99th percentiles were 1.5 and 4.7 million dollars. The lower panel of Figure 1 shows the housing value density (truncated at the 90th percentile). The mean housing value in 1960 is $400,000 when not weighting the sample, and is smaller ($379,000) when weighting.28 As long as the remaining within‐city variation is stable between 1960 and today, the deflated housing prices represent a good approximation of the gangsters’ economic status. Unfortunately there is little empirical literature about the persistence in housing prices within cities over decades. A recent study (Villarreal, 2013) shows that natural drainage conditions influenced the desirability of the local environment during the settlement of New York City in the nineteenth century and that, despite private residential redevelopment, those historical conditions explain contemporary variation in housing prices and household incomes.29 Zillow.com publishes more than 10,000 zip code‐level median housing prices going back 16 years. Using such data to compute the correlation between current (T) and lagged (T − t) prices, one can show that as t goes to 16 years the correlation declines from 100% to 92%.30 Moreover, deflating prices using city‐level inflation rates the decline drops from 8 to just 2 percentage points. That said, the correlation might decrease (and the measurement error increase) when moving from median prices to prices of individual housing units, and when reducing t for the remaining 38 years. The construction of housing value is also subject to additional problems. Wealthy mobsters might own several houses and their place of residence might not perfectly represent their wealth whenever it is in the hands of figureheads.31 All these measurement problems render the results more conservative. Later in the network effects regressions I also control for the legitimate earnings opportunities, measured by the number of legal business that gangsters own. 2.2. Network‐based Measures of Importance Each criminal record contains a list of criminal associates. Online Appendix Figure B2 indicates, for example, that Joe Bonanno was associated with Luciano, Costello, Profaci, Corallo, Lucchese, and Galante. There is no evidence about how the FBN established such associations but each record tends to list the most important (and connected) associates. Indirected connections are clearly more numerous, as mobsters can be listed as associates in several records. As in Mastrobuoni and Patacchini (2012), I define two mobsters to be connected whenever the FBN lists one of the two mobsters as an associate in the other mobster’s file. Outdegree (the number of associates listed in one file) is bounded by the available space on a record (the maximum is 13), while indegree (the number of times someone is listed in other records) is not. For this reason degree (the number of undirected connections) is only weakly correlated with outdegree (37%), and mainly depends on indegree.32 The number of connections is clearly the simplest but crudest way to measure the importance of members. In recent years social network theorists proposed different centrality measures to account for the importance of someone’s connections (Wasserman and Faust, 1994; Borgatti, 2004).33 Unlike degree, which weights every contact equally, the eigenvector index weighs contacts according to their centralities.34 The index takes the whole network into account (direct and indirect connections).35 Given the first mafia rule that guarantees secrecy, gangsters who are closer to other gangsters need fewer interlinking associates to reach a randomly chosen gangster, while gangsters with important bridging capacities across clusters of the network have more monopoly power in establishing such new links. While degree centrality might in principle be inferior to the other centrality measures, I will estimate network effects based on all four measures of centrality. All these measures are positively skewed (Figure 2), especially betweenness centrality, indicating that a few mobsters represent the bridges between subsets of the network. Given the skewness and the scaling bias that I discussed prior to (2) in order to estimate the network effects all measures of network centrality are taken in logarithms.36 Ductor et al. (2014), also use a specification in logarithms. The corresponding (more symmetric) densities are plotted in the online Appendix Figure B5. Another online Appendix Figure B6 shows that all centrality measures are positively related to each other. Plotting log eigenvector against log closeness generates a thick line (ρ = 96%), which shows that once one penalises the larger outliers the two centrality measures are quite similar. Fig. 2. Open in new tabDownload slide Densities of Centrality Indices Note. The density shown excludes the top 1% of the distribution. Fig. 2. Open in new tabDownload slide Densities of Centrality Indices Note. The density shown excludes the top 1% of the distribution. But such large correlation masks a very different variability. The ratio between the standard deviation of the log eigenvector index and the log closeness index is about 10 to 1. This has to be taken into account when interpreting standardised variations. The correlations are lower with respect to the other two measures, especially in the lower tail. For betweenness the low correlation is driven by the fact that several mobsters, despite having many connections, have extremely low levels of betweenness. Hierarchies are the likely reason. For about 400 mobsters I managed to reconstruct their position within the mafia (not always referred to their status in 1960). Underbosses and captains who head several soldiers within one mafia ‘family’, tend to have large degrees but low betweenness. Counsellors and bosses have the largest median betweenness measure (about 1/2), while those of captains and underbosses are half that large.37 Differences in closeness and eigenvector centrality across ranks are less striking. With respect to degree the low correlation with other indices is driven by small values. At larger degrees most indices tend to be highly correlated with each other. The next subsection describes the relationship between housing prices and network centrality. 2.3. Economic Status, Network Centrality, and Potential Biases Figure 3 shows that the unconditional weighted local polynomial regression of log‐housing value is increasing in all log‐centrality measures and is not far from being linear. The correlation is stronger when using the eigenvector index and the closeness index, than when using the simple degree or the betweenness index (approximately, 20% versus 10%). Fig. 3. Open in new tabDownload slide Housing Values and Network Centrality (in Logs) Notes. The size of the circles is proportional to the sampling weight. The solid line corresponds to the local polynomial smoothing regression line (using the Epanechnikov kernel). Fig. 3. Open in new tabDownload slide Housing Values and Network Centrality (in Logs) Notes. The size of the circles is proportional to the sampling weight. The solid line corresponds to the local polynomial smoothing regression line (using the Epanechnikov kernel). On the one hand, degree is likely to be a poor measure of centrality when the connections are scarce but valuable. On the other hand, mobsters with larger betweenness indices, represent bridges between clusters of the network, most likely different mafia families. Since the detection of a such high ranking bosses would have put the whole organisation in peril (Baccara and Bar‐Isaac, 2008), there is anecdotal evidence that some bosses kept a low profile. For example, Joseph Bonanno tells the story about when he decided not to join Lucky Luciano’s very lucrative garment industry in New York to avoid being in the spotlight (Bonanno, 1983). Such members’ true economic outcomes might also be hidden whenever their wealth is in the hands of figureheads. There is indeed evidence that some mafia leaders preferred to ‘officially’ reside in unpretentious housing. Using the FBN’s description of the associates, one can show that those described as ‘leaders’ or ‘bosses’ tend to be more central in the network (Table 2). They tend to have considerably larger betweenness centrality than lower‐ranked gangsters, about 40% larger, while other centrality measures differ less between bosses and lower‐ranked gangsters. Despite this, such leaders tend to live in less expensive housing.38 Table 2 Housing Values and Centralities by Leadership Status . Mafia leader (weighted) . Mafia leader (unweighted) . . No (91%) . Yes (9%) . No (87%) . Yes (13%) . Housing value 380,453 362,479 399,654 387,587 Involved in drug dealing 0.62 0.39 0.58 0.45 Betweenness 1.76 2.86 4.26 8.46 Closeness 69.63 70.85 74.51 78.23 Eigenvector 6.21 9.26 11.54 19.03 Degree 5.94 8.10 10.25 15.46 . Mafia leader (weighted) . Mafia leader (unweighted) . . No (91%) . Yes (9%) . No (87%) . Yes (13%) . Housing value 380,453 362,479 399,654 387,587 Involved in drug dealing 0.62 0.39 0.58 0.45 Betweenness 1.76 2.86 4.26 8.46 Closeness 69.63 70.85 74.51 78.23 Eigenvector 6.21 9.26 11.54 19.03 Degree 5.94 8.10 10.25 15.46 Note. About 13% of gangsters are described as either ‘leader’ or ‘boss’ in the FBN records. Open in new tab Table 2 Housing Values and Centralities by Leadership Status . Mafia leader (weighted) . Mafia leader (unweighted) . . No (91%) . Yes (9%) . No (87%) . Yes (13%) . Housing value 380,453 362,479 399,654 387,587 Involved in drug dealing 0.62 0.39 0.58 0.45 Betweenness 1.76 2.86 4.26 8.46 Closeness 69.63 70.85 74.51 78.23 Eigenvector 6.21 9.26 11.54 19.03 Degree 5.94 8.10 10.25 15.46 . Mafia leader (weighted) . Mafia leader (unweighted) . . No (91%) . Yes (9%) . No (87%) . Yes (13%) . Housing value 380,453 362,479 399,654 387,587 Involved in drug dealing 0.62 0.39 0.58 0.45 Betweenness 1.76 2.86 4.26 8.46 Closeness 69.63 70.85 74.51 78.23 Eigenvector 6.21 9.26 11.54 19.03 Degree 5.94 8.10 10.25 15.46 Note. About 13% of gangsters are described as either ‘leader’ or ‘boss’ in the FBN records. Open in new tab Moreover, the Sicilian origin seems to influence the decision to keep a lower profile.39 Figure 4 shows that despite the fact that Sicilians are usually more centrally located in the mafia network, they tend to live in considerably cheaper housing; especially at the top of the distribution of housing values. Later in Section 3 I test the robustness of the network effect when controlling for the place of origin. Fig. 4. Open in new tabDownload slide Distribution of Housing Values by Region of Birth Fig. 4. Open in new tabDownload slide Distribution of Housing Values by Region of Birth Other potential omitted variables when regressing housing values on network centrality are the family composition and the gangsters’ initial wealth. Larger families are instrumental to the bosses’ success but need also larger and more expensive housing. Wealth might be used to buy both, power inside the mafia and more expensive housing. While the data have no good proxy for initial wealth, in the next subsection I devise a presumably exogenous instrument that is based on the innate potential connection back in Italy. In particular, I exploit the joint spatial distribution of the gangsters’ surnames in Italy to construct a measure of inherited (potential) connections, which is moderately correlated with network centrality. 2.4. Birthplace and Potential Innate Interactions Several authors have highlighted the importance of familial, interpersonal and communal relationships in determining criminals’ success inside organised crime groups (Ianni and Reuss‐Ianni, 1972; Falcone and Padovani, 1991; Coles, 2001). Most of these relationships are also likely to influence housing decisions and would lead to implausible exclusion restrictions. For example, marrying a gangster’s daughter is likely to boost someone’s power inside the mafia, but might also change someone’s housing budget directly. Ideally, one would use mobsters’ innate characteristics which influence his future chances of building connections but are unrelated to his housing choice (other than through the derived centrality in the network). Proximity to other mobsters represents a natural choice. But such proximity should not be related to inherited wealth, as wealth might be used to acquire centrality. Moreover, geographic proximity based on the place of US residence is likely to be endogenous with respect to network centrality, as more powerful mobsters might decide to live in the middle of their sphere of influence. For these reasons I use a measure of proximity that is based on Italian and not US residencies. Mafia associates born in Italy40 and also those born in the US from Italian immigrants often kept strong links with the Italian communities of origin.41 At the beginning of this study, I listed a number of reasons why such innate connections might influence the gangsters’ connections back in the US (reputation, trust, punishability and knowledge of the rules). The main untestable identification assumption is that such interactions at the origin do not shape the gangsters’ housing preferences, at least not conditional on other covariates (including the region of birth). The records do not contain pre‐immigration information on the exact place of origin in Italy but one can approximate such information using the informative content of surnames. For at least 30 years researchers in human biology have been exploiting the analogy between patrilineal surname transmission and the characterisation of families and communities (Lasker, 1977). For geographic, historical, as well as social reasons, surnames tend to be highly geographically clustered, particularly in countries with low internal mobility such as Italy (Zei et al., 1993; Barrai et al., 1999; Allesina, 2011).42 The geographic distribution of surnames, called isonomy, contains a strong signal about someone’s origin.43 The surname of record number 1, ‘Bonanno’, is more widespread across the whole country, though, again, most Bonanno families live in Sicily, and a non‐negligible fraction lives in Castellamare del Golfo, which is where Bonanno’s family was coming from.44 Given that: (i) 30% of gangsters were born in Italy (including Sicily) and later moved to the US and even those who were born in the US were likely to keep links with Italy; and (ii) surnames tend to be geographically clustered, the way the current distribution of a given surname overlaps with the distribution of all the other surnames represents one way, possibly the only way, to measure the connections stemming from the gangsters’ origin country (thus unrelated to US housing prices). Since poor living conditions in Italy were the main driver for migrating for the US, such innate connections are hardly measuring innate wealth. Conditional on the region of birth and on all the other individual characteristics of the mobsters they are also unlikely to be related to individual preferences for housing. Since some of the mobsters might have migrated as a response to increased wealth levels by the potential innate connections (violating the exclusion restriction between potential innate connections and housing wealth), later in the robustness regressions I control for the place of origin of the gangsters’ direct associates (direct links). Figure 5 visualises how I construct the index. Starting from the current postal code level distribution of the members’ surnames,45 I compute the probability that each members’ surname shares a randomly chosen postal code located in Italy (as a robustness check I also limit the attention to the South) with other surnames from the list.46 To be more precise, the index for member i is equal to 1,000,000 times the sum across postal codes j of the fraction of surnames of member i present in postal codes j times the fraction of surnames of the other members (−i) in the same postal code: Potential innate interactionsindexi=109∑j#surnamei,j∑j#surnamei,j#surname−i,j∑j#surname−i,j.(3) Fig. 5. Open in new tabDownload slide Geographical Distribution of Mafia Surnames Notes. Each circle represents a postal code. The size of the circles is proportional to the number of US mafia members’ surnames found in today’s Italian phone directory. Fig. 5. Open in new tabDownload slide Geographical Distribution of Mafia Surnames Notes. Each circle represents a postal code. The size of the circles is proportional to the number of US mafia members’ surnames found in today’s Italian phone directory. There are 4,748 postal codes for about 60 million Italians, thus each postal code covers a little more than 12,000 Italians, and an area of about 23 square miles, a reasonable area within which most relationships are likely to get established. In Figure 5 each circle is proportional to the number of surnames present within each postal code. Not surprisingly many surnames show up in Sicily, in Naples, and in Calabria. Many of these surnames appear also in large cities that were subject to immigratory flows from the south, like Milan, Rome, and Turin. Such migration patterns introduce some noise in the instrument, which is why later I also compute a Potential Innate Interactions measure that is just based on Southern regions (Campania, Molise, Calabria, Basilicata, Sicilia, Sardegna, Puglia), with Northern regions acting as imperfect falsification samples (as migration might depend on ethnic networks). The Potential innate interactions index tends to be small when the fraction of surnames i overlaps little with the fraction of all other surnames −i. Dividing by the total number of surnames (i and −i) takes into account that some surnames are more frequent than others. The average index is equal to 36 per 1 million, though taking the sampling into account it drops to 26, already indicating that more connected mobsters have more interactions (see Table 1). About 10% of the times the index is zero, either because the postal codes do not overlap or because the surname is not in the phone directory.47 Figure 6 shows the distribution of the interaction index. Mobsters born in Italy tend to have more innate potential interactions than those born outside of Italy (Figure 7). Most mobsters with very large interactions were born in the Western part of Sicily (the 10 cities of birth corresponding to the largest interaction indices are in major US cities (Chicago, New York, St. Louis), but also Palermo (Sicily), Cerda (Sicily), Trapani (Sicily), Amantea (Calabria). Most of these are well‐known mafia enclaves. Fig. 6. Open in new tabDownload slide Density of Potential Innate Interactions Fig. 6. Open in new tabDownload slide Density of Potential Innate Interactions Fig. 7. Open in new tabDownload slide Distribution of the Potential Innate Interactions Index by Region of Birth Fig. 7. Open in new tabDownload slide Distribution of the Potential Innate Interactions Index by Region of Birth When measuring the interactions focusing on postal codes in Southern Italy the shape of the bars is very similar, while when focusing on postal codes in Central and Northern Italy the interaction probabilities are considerably lower and are lowest among Sicilians, which is consistent with the index capturing interactions in the place of origin. Since Northern interactions are likely to be the product of recent migrations from the South to the North of Italy rather than true potential connections in 1960 they represent a useful placebo case.48 Potential interactions influence the centrality measures (see Figure 8 and Table 3). Correlations are between 15% (eigenvector, closeness and degree) and 30% (betweenness), indicating that such initial interactions are important, though neither necessary, nor sufficient, to reach the top positions in the organisation (those positions that generate bridges across families). Moreover, the correlation is larger when restricting the interactions to Southern Italy and considerably weaker when restricting the interactions to Northern Italy. For this reason, in most two‐stage regressions I will use the interactions based on Southern regions as instruments. Fig. 8. Open in new tabDownload slide Log‐network Centrality and Potential Innate Interactions Index Notes. The size of the circles is proportional to the sampling weight. The solid line corresponds to the local polynomial smoothing regression line (using the Epanechnikov kernel). Fig. 8. Open in new tabDownload slide Log‐network Centrality and Potential Innate Interactions Index Notes. The size of the circles is proportional to the sampling weight. The solid line corresponds to the local polynomial smoothing regression line (using the Epanechnikov kernel). Table 3 Correlation Table . 0log Clo. . log Bet. . log Eig. . log Deg. . Pot. Int. . Pot. Int. South . log Betweenness 0.533 log Eigenvector 0.968 0.460 log Degree 0.801 0.646 0.806 Potential innate interactions 0.159 0.309 0.121 0.165 Pot. Int. in the South 0.185 0.345 0.148 0.191 0.959 Pot. Int. in the North 0.060 0.150 0.034 0.066 0.811 0.610 . 0log Clo. . log Bet. . log Eig. . log Deg. . Pot. Int. . Pot. Int. South . log Betweenness 0.533 log Eigenvector 0.968 0.460 log Degree 0.801 0.646 0.806 Potential innate interactions 0.159 0.309 0.121 0.165 Pot. Int. in the South 0.185 0.345 0.148 0.191 0.959 Pot. Int. in the North 0.060 0.150 0.034 0.066 0.811 0.610 Notes. Clustered (by surname) standard errors in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1. Open in new tab Table 3 Correlation Table . 0log Clo. . log Bet. . log Eig. . log Deg. . Pot. Int. . Pot. Int. South . log Betweenness 0.533 log Eigenvector 0.968 0.460 log Degree 0.801 0.646 0.806 Potential innate interactions 0.159 0.309 0.121 0.165 Pot. Int. in the South 0.185 0.345 0.148 0.191 0.959 Pot. Int. in the North 0.060 0.150 0.034 0.066 0.811 0.610 . 0log Clo. . log Bet. . log Eig. . log Deg. . Pot. Int. . Pot. Int. South . log Betweenness 0.533 log Eigenvector 0.968 0.460 log Degree 0.801 0.646 0.806 Potential innate interactions 0.159 0.309 0.121 0.165 Pot. Int. in the South 0.185 0.345 0.148 0.191 0.959 Pot. Int. in the North 0.060 0.150 0.034 0.066 0.811 0.610 Notes. Clustered (by surname) standard errors in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1. Open in new tab The next Section assigns confidence intervals to the unconditional network effects shown in Figure 3 and adds further controls and an IV strategy to pinpoint such effects. 3. Regression Results 3.1. Evidence Based on Ordinary Least Squares Regressions When estimating (2) using closeness centrality by ordinary least squared, I find that doubling the centrality measure increases housing values by about 200% (Table 4).49 Such large elasticities are driven by a very compressed distribution of closeness centrality. In terms of standard deviations (SDs), an SD increase in log closeness (0.12) increases housing values by one quarter. Table 4 Housing Value Regressions and Closeness Centrality . (1) . (2) . (3) . (4) . (5) . . log Housing value . log Closeness 2.069*** 2.524*** 1.418*** (0.406) (0.452) (0.497) Closeness 0.036*** 0.020*** (0.007) (0.007) Housing built after 1960 0.100 0.054 0.048 0.051 0.045 (0.156) (0.159) (0.157) (0.160) (0.158) Year built unknown −0.346*** −0.277** −0.298** −0.276** −0.298** (0.121) (0.122) (0.140) (0.122) (0.140) Extended family members −0.216*** −0.067 −0.224*** −0.069 (0.073) (0.056) (0.074) (0.056) Born outside Italy (mainly US) −0.089 −0.038 −0.085 −0.035 (0.173) (0.160) (0.172) (0.160) Born in Sicily −0.273 −0.173 −0.269 −0.169 (0.188) (0.161) (0.188) (0.161) Age −0.065 −0.076 −0.063 −0.075 (0.074) (0.075) (0.074) (0.075) Age 2/100 0.067 0.082 0.065 0.081 (0.081) (0.080) (0.081) (0.080) Year of birth unknown −1.750 −1.935 −1.729 −1.920 (1.602) (1.663) (1.610) (1.664) Height in feet 0.530* 0.224 0.536* 0.224 (0.297) (0.279) (0.299) (0.280) Weight in pounds −0.001 −0.000 −0.001 −0.000 (0.002) (0.002) (0.002) (0.002) Age at first arrest −0.013* −0.016*** −0.013* −0.016*** (0.007) (0.006) (0.007) (0.006) Never arrested −0.127 −0.181 −0.141 −0.188 (0.204) (0.183) (0.205) (0.183) Married −0.080 −0.000 −0.083 −0.001 (0.151) (0.138) (0.151) (0.139) Divorced −0.042 −0.005 −0.052 −0.010 (0.222) (0.184) (0.222) (0.184) Number of children −0.030 −0.062** −0.031 −0.064** (0.030) (0.028) (0.030) (0.028) Siblings 0.016 −0.010 0.016 −0.010 (0.025) (0.025) (0.026) (0.025) Types of crime committed 0.021 −0.002 0.019 −0.003 (0.038) (0.034) (0.038) (0.034) Number of businesses 0.043 0.092** 0.043 0.092** (0.051) (0.043) (0.052) (0.043) Involved in drug dealing 0.346*** 0.003 0.344*** −0.000 (0.099) (0.091) (0.099) (0.091) State of residence fixed effects X X Observations 641 641 637 641 637 R‐squared 0.066 0.144 0.349 0.141 0.348 . (1) . (2) . (3) . (4) . (5) . . log Housing value . log Closeness 2.069*** 2.524*** 1.418*** (0.406) (0.452) (0.497) Closeness 0.036*** 0.020*** (0.007) (0.007) Housing built after 1960 0.100 0.054 0.048 0.051 0.045 (0.156) (0.159) (0.157) (0.160) (0.158) Year built unknown −0.346*** −0.277** −0.298** −0.276** −0.298** (0.121) (0.122) (0.140) (0.122) (0.140) Extended family members −0.216*** −0.067 −0.224*** −0.069 (0.073) (0.056) (0.074) (0.056) Born outside Italy (mainly US) −0.089 −0.038 −0.085 −0.035 (0.173) (0.160) (0.172) (0.160) Born in Sicily −0.273 −0.173 −0.269 −0.169 (0.188) (0.161) (0.188) (0.161) Age −0.065 −0.076 −0.063 −0.075 (0.074) (0.075) (0.074) (0.075) Age 2/100 0.067 0.082 0.065 0.081 (0.081) (0.080) (0.081) (0.080) Year of birth unknown −1.750 −1.935 −1.729 −1.920 (1.602) (1.663) (1.610) (1.664) Height in feet 0.530* 0.224 0.536* 0.224 (0.297) (0.279) (0.299) (0.280) Weight in pounds −0.001 −0.000 −0.001 −0.000 (0.002) (0.002) (0.002) (0.002) Age at first arrest −0.013* −0.016*** −0.013* −0.016*** (0.007) (0.006) (0.007) (0.006) Never arrested −0.127 −0.181 −0.141 −0.188 (0.204) (0.183) (0.205) (0.183) Married −0.080 −0.000 −0.083 −0.001 (0.151) (0.138) (0.151) (0.139) Divorced −0.042 −0.005 −0.052 −0.010 (0.222) (0.184) (0.222) (0.184) Number of children −0.030 −0.062** −0.031 −0.064** (0.030) (0.028) (0.030) (0.028) Siblings 0.016 −0.010 0.016 −0.010 (0.025) (0.025) (0.026) (0.025) Types of crime committed 0.021 −0.002 0.019 −0.003 (0.038) (0.034) (0.038) (0.034) Number of businesses 0.043 0.092** 0.043 0.092** (0.051) (0.043) (0.052) (0.043) Involved in drug dealing 0.346*** 0.003 0.344*** −0.000 (0.099) (0.091) (0.099) (0.091) State of residence fixed effects X X Observations 641 641 637 641 637 R‐squared 0.066 0.144 0.349 0.141 0.348 Notes. Clustered (by 530 surnames) standard errors in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1. Open in new tab Table 4 Housing Value Regressions and Closeness Centrality . (1) . (2) . (3) . (4) . (5) . . log Housing value . log Closeness 2.069*** 2.524*** 1.418*** (0.406) (0.452) (0.497) Closeness 0.036*** 0.020*** (0.007) (0.007) Housing built after 1960 0.100 0.054 0.048 0.051 0.045 (0.156) (0.159) (0.157) (0.160) (0.158) Year built unknown −0.346*** −0.277** −0.298** −0.276** −0.298** (0.121) (0.122) (0.140) (0.122) (0.140) Extended family members −0.216*** −0.067 −0.224*** −0.069 (0.073) (0.056) (0.074) (0.056) Born outside Italy (mainly US) −0.089 −0.038 −0.085 −0.035 (0.173) (0.160) (0.172) (0.160) Born in Sicily −0.273 −0.173 −0.269 −0.169 (0.188) (0.161) (0.188) (0.161) Age −0.065 −0.076 −0.063 −0.075 (0.074) (0.075) (0.074) (0.075) Age 2/100 0.067 0.082 0.065 0.081 (0.081) (0.080) (0.081) (0.080) Year of birth unknown −1.750 −1.935 −1.729 −1.920 (1.602) (1.663) (1.610) (1.664) Height in feet 0.530* 0.224 0.536* 0.224 (0.297) (0.279) (0.299) (0.280) Weight in pounds −0.001 −0.000 −0.001 −0.000 (0.002) (0.002) (0.002) (0.002) Age at first arrest −0.013* −0.016*** −0.013* −0.016*** (0.007) (0.006) (0.007) (0.006) Never arrested −0.127 −0.181 −0.141 −0.188 (0.204) (0.183) (0.205) (0.183) Married −0.080 −0.000 −0.083 −0.001 (0.151) (0.138) (0.151) (0.139) Divorced −0.042 −0.005 −0.052 −0.010 (0.222) (0.184) (0.222) (0.184) Number of children −0.030 −0.062** −0.031 −0.064** (0.030) (0.028) (0.030) (0.028) Siblings 0.016 −0.010 0.016 −0.010 (0.025) (0.025) (0.026) (0.025) Types of crime committed 0.021 −0.002 0.019 −0.003 (0.038) (0.034) (0.038) (0.034) Number of businesses 0.043 0.092** 0.043 0.092** (0.051) (0.043) (0.052) (0.043) Involved in drug dealing 0.346*** 0.003 0.344*** −0.000 (0.099) (0.091) (0.099) (0.091) State of residence fixed effects X X Observations 641 641 637 641 637 R‐squared 0.066 0.144 0.349 0.141 0.348 . (1) . (2) . (3) . (4) . (5) . . log Housing value . log Closeness 2.069*** 2.524*** 1.418*** (0.406) (0.452) (0.497) Closeness 0.036*** 0.020*** (0.007) (0.007) Housing built after 1960 0.100 0.054 0.048 0.051 0.045 (0.156) (0.159) (0.157) (0.160) (0.158) Year built unknown −0.346*** −0.277** −0.298** −0.276** −0.298** (0.121) (0.122) (0.140) (0.122) (0.140) Extended family members −0.216*** −0.067 −0.224*** −0.069 (0.073) (0.056) (0.074) (0.056) Born outside Italy (mainly US) −0.089 −0.038 −0.085 −0.035 (0.173) (0.160) (0.172) (0.160) Born in Sicily −0.273 −0.173 −0.269 −0.169 (0.188) (0.161) (0.188) (0.161) Age −0.065 −0.076 −0.063 −0.075 (0.074) (0.075) (0.074) (0.075) Age 2/100 0.067 0.082 0.065 0.081 (0.081) (0.080) (0.081) (0.080) Year of birth unknown −1.750 −1.935 −1.729 −1.920 (1.602) (1.663) (1.610) (1.664) Height in feet 0.530* 0.224 0.536* 0.224 (0.297) (0.279) (0.299) (0.280) Weight in pounds −0.001 −0.000 −0.001 −0.000 (0.002) (0.002) (0.002) (0.002) Age at first arrest −0.013* −0.016*** −0.013* −0.016*** (0.007) (0.006) (0.007) (0.006) Never arrested −0.127 −0.181 −0.141 −0.188 (0.204) (0.183) (0.205) (0.183) Married −0.080 −0.000 −0.083 −0.001 (0.151) (0.138) (0.151) (0.139) Divorced −0.042 −0.005 −0.052 −0.010 (0.222) (0.184) (0.222) (0.184) Number of children −0.030 −0.062** −0.031 −0.064** (0.030) (0.028) (0.030) (0.028) Siblings 0.016 −0.010 0.016 −0.010 (0.025) (0.025) (0.026) (0.025) Types of crime committed 0.021 −0.002 0.019 −0.003 (0.038) (0.034) (0.038) (0.034) Number of businesses 0.043 0.092** 0.043 0.092** (0.051) (0.043) (0.052) (0.043) Involved in drug dealing 0.346*** 0.003 0.344*** −0.000 (0.099) (0.091) (0.099) (0.091) State of residence fixed effects X X Observations 641 641 637 641 637 R‐squared 0.066 0.144 0.349 0.141 0.348 Notes. Clustered (by 530 surnames) standard errors in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1. Open in new tab The first column controls only for variables collected from Zillow.com, in particular whether the housing unit has been built after 1960 (the year of the FBN records) and whether such information is available. The negative coefficient on this variable is capturing that for more expensive housing Zillow.com is more likely to collect information on the year the house was built. Controlling for additional variables (column 2) increases somehow the effect, while adding US state of residence fixed effects reduces them (column 3); partly because the more influential mobsters resided in New York and New Jersey (where housing prices tend to be high).50 Given that the state of residence measures part of the centrality of mobsters in the remainder of the study I will not control for it. The instrument used in subsection 3.2. is by construction unrelated to the state of residence (other than through connections). I also show that the results change little when controlling for the log state‐level average housing price (thus de facto using the log of relative housing prices). The last two columns show that similar results are found when using closeness in levels rather than in logarithms, though in such case the predictive power of centrality drops slightly (in line with the scaling bias). As for the logarithm, a standard deviation increase in closeness centrality (equal to 8.27) increases housing values by about one quarter. Before showing the results for other centrality measures, let me briefly discuss the coefficients on the other regressors. Gangsters born in Italy, in particular those born in Sicily tend to live in cheaper housing, which might either be due to housing preferences or to a more pronounced avoidance to attract attention (though the effects are not significantly different from 0).51 Age at first arrest, which might represent a (negative) measure of career experience within the organisation tends to be negatively related to housing values, meaning that earlier arrests tend to be related to increased housing values. Not just experience, but also being involved with drug dealing seems to be related to higher housing values, though the coefficient stops being significant once state of residence effects are added to the regression (indicating that the drug dealing business used to be geographically clustered). The number of legal businesses tend to be positively associated with housing values, which is not surprising. Going back to the centrality measures, in Table 5 I substitute closeness centrality with the other measures, with and without controlling for additional regressors. A quick look at the R‐squared reveals that closeness centrality is the strongest predictor of housing value. This is coherent with the first rule of the mafia, which states that only interlinking affiliates, for example B in a network A–B–C, have the power to introduce a direct link between A and C. Mobsters who are on average closer to all other mobsters will thus need fewer interlinking associates to establish new connections. Eigenvector centrality has similar predictive power, which is not surprising given how strongly correlated the two measures are.52 Table 5 Housing Value Regressions with Other Centrality Measures . (1) . (2) . (3) . (4) . (5) . (6) . . log Housing Value . . . . . . log Eigenvector 0.184*** 0.218*** (0.040) (0.043) log Degree 0.148** 0.220*** (0.068) (0.083) log Betweenness −0.085 0.036 (0.115) (0.158) Other Xs X X X Observations 641 641 641 641 641 641 R‐squared 0.062 0.136 0.025 0.098 0.014 0.076 . (1) . (2) . (3) . (4) . (5) . (6) . . log Housing Value . . . . . . log Eigenvector 0.184*** 0.218*** (0.040) (0.043) log Degree 0.148** 0.220*** (0.068) (0.083) log Betweenness −0.085 0.036 (0.115) (0.158) Other Xs X X X Observations 641 641 641 641 641 641 R‐squared 0.062 0.136 0.025 0.098 0.014 0.076 Notes. All regressions control for the housing variables. The additional regressors (‘Other Xs’) are the same as in Table 4. Clustered (by 530 surnames) standard errors in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1. Open in new tab Table 5 Housing Value Regressions with Other Centrality Measures . (1) . (2) . (3) . (4) . (5) . (6) . . log Housing Value . . . . . . log Eigenvector 0.184*** 0.218*** (0.040) (0.043) log Degree 0.148** 0.220*** (0.068) (0.083) log Betweenness −0.085 0.036 (0.115) (0.158) Other Xs X X X Observations 641 641 641 641 641 641 R‐squared 0.062 0.136 0.025 0.098 0.014 0.076 . (1) . (2) . (3) . (4) . (5) . (6) . . log Housing Value . . . . . . log Eigenvector 0.184*** 0.218*** (0.040) (0.043) log Degree 0.148** 0.220*** (0.068) (0.083) log Betweenness −0.085 0.036 (0.115) (0.158) Other Xs X X X Observations 641 641 641 641 641 641 R‐squared 0.062 0.136 0.025 0.098 0.014 0.076 Notes. All regressions control for the housing variables. The additional regressors (‘Other Xs’) are the same as in Table 4. Clustered (by 530 surnames) standard errors in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1. Open in new tab The coefficient on degree, instead, has relatively larger standard errors, and a standard deviation increase in log degree increases housing value by about 11%; in line with what emerged in Figure 3, the flattest relationship is the one between log housing value and log betweenness centrality. Statistically speaking, the slope is 0.53 This is likely driven by the bosses’ preference for keeping a low profile and, as argued before, such bias cannot be eliminated. 3.2. Evidence Based on Instrumental Variable Regressions As previously discussed, for a number of reasons (including measurement error) OLS coefficients on centrality measures might be inconsistent. Table 6 presents the first stage (sketched in Figure 8) and the main regressions in a compact form, focusing on the instruments and on the endogenous variable.54 Table 6 Two Stage Least Squares Regressions . (1) . (2) . (3) . (4) . Panel (a): First stage regression log Closeness log Eigenvector log Degree log Betweennes Potential innate interactions (in %) 0.033** 0.268* 0.116 0.138*** (0.015) (0.158) (0.094) (0.053) Zero potential innate interactions −0.019 −0.308 −0.236 0.011 (0.020) (0.234) (0.153) (0.032) R‐squared 0.160 0.146 0.205 0.326 Kleibergen‐Paap rk Wald F statistic 3.479 2.663 2.343 12.60 F‐stat on potential innate interactions 5.208 2.887 1.523 6.726 Panel (b): Main regression log Housing value log Closeness 6.350** (3.196) log Eigenvector 0.627* (0.329) log Degree 0.977* (0.537) log Betweenness 1.201 (1.054) Observations 641 641 641 641 Endogeneity p‐value 0.305 0.214 0.146 0.244 Kleibergen‐Paap rk Wald F statistic 3.479 2.663 2.343 3.399 Anderson and Rubin p‐value 0.212 . (1) . (2) . (3) . (4) . Panel (a): First stage regression log Closeness log Eigenvector log Degree log Betweennes Potential innate interactions (in %) 0.033** 0.268* 0.116 0.138*** (0.015) (0.158) (0.094) (0.053) Zero potential innate interactions −0.019 −0.308 −0.236 0.011 (0.020) (0.234) (0.153) (0.032) R‐squared 0.160 0.146 0.205 0.326 Kleibergen‐Paap rk Wald F statistic 3.479 2.663 2.343 12.60 F‐stat on potential innate interactions 5.208 2.887 1.523 6.726 Panel (b): Main regression log Housing value log Closeness 6.350** (3.196) log Eigenvector 0.627* (0.329) log Degree 0.977* (0.537) log Betweenness 1.201 (1.054) Observations 641 641 641 641 Endogeneity p‐value 0.305 0.214 0.146 0.244 Kleibergen‐Paap rk Wald F statistic 3.479 2.663 2.343 3.399 Anderson and Rubin p‐value 0.212 Notes. Panel (a) shows the first stage regressions, Panel (b) the main log Housing value 2SLS regressions. Additional housing variables and the additional regressors (‘Other Xs’) are the same as in Table 4. The Potential Innate Interactions are based on Southern regions. The Stock‐Yogo weak ID test critical values for F tests on single endogenous regressor are: 16.38 (10% maximal IV size), 8.96 (15%), 6.66 (20%) and 5.53 (25%). Clustered (by surname) standard errors in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1. Open in new tab Table 6 Two Stage Least Squares Regressions . (1) . (2) . (3) . (4) . Panel (a): First stage regression log Closeness log Eigenvector log Degree log Betweennes Potential innate interactions (in %) 0.033** 0.268* 0.116 0.138*** (0.015) (0.158) (0.094) (0.053) Zero potential innate interactions −0.019 −0.308 −0.236 0.011 (0.020) (0.234) (0.153) (0.032) R‐squared 0.160 0.146 0.205 0.326 Kleibergen‐Paap rk Wald F statistic 3.479 2.663 2.343 12.60 F‐stat on potential innate interactions 5.208 2.887 1.523 6.726 Panel (b): Main regression log Housing value log Closeness 6.350** (3.196) log Eigenvector 0.627* (0.329) log Degree 0.977* (0.537) log Betweenness 1.201 (1.054) Observations 641 641 641 641 Endogeneity p‐value 0.305 0.214 0.146 0.244 Kleibergen‐Paap rk Wald F statistic 3.479 2.663 2.343 3.399 Anderson and Rubin p‐value 0.212 . (1) . (2) . (3) . (4) . Panel (a): First stage regression log Closeness log Eigenvector log Degree log Betweennes Potential innate interactions (in %) 0.033** 0.268* 0.116 0.138*** (0.015) (0.158) (0.094) (0.053) Zero potential innate interactions −0.019 −0.308 −0.236 0.011 (0.020) (0.234) (0.153) (0.032) R‐squared 0.160 0.146 0.205 0.326 Kleibergen‐Paap rk Wald F statistic 3.479 2.663 2.343 12.60 F‐stat on potential innate interactions 5.208 2.887 1.523 6.726 Panel (b): Main regression log Housing value log Closeness 6.350** (3.196) log Eigenvector 0.627* (0.329) log Degree 0.977* (0.537) log Betweenness 1.201 (1.054) Observations 641 641 641 641 Endogeneity p‐value 0.305 0.214 0.146 0.244 Kleibergen‐Paap rk Wald F statistic 3.479 2.663 2.343 3.399 Anderson and Rubin p‐value 0.212 Notes. Panel (a) shows the first stage regressions, Panel (b) the main log Housing value 2SLS regressions. Additional housing variables and the additional regressors (‘Other Xs’) are the same as in Table 4. The Potential Innate Interactions are based on Southern regions. The Stock‐Yogo weak ID test critical values for F tests on single endogenous regressor are: 16.38 (10% maximal IV size), 8.96 (15%), 6.66 (20%) and 5.53 (25%). Clustered (by surname) standard errors in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1. Open in new tab Having no potential innate connections is allowed to have its own influence on centrality. All centrality measures are positively correlated with the potential connections, but the coefficient is more significant for closeness and betweenness centralities, though the instrument is not very strong. In particular, for closeness centrality the F‐statistic for the excluded instruments is around 3.5 when considering the joint significance of interactions and zero interactions (otherwise it is close to 5). The F‐statistic is largest and above 10 for the betweenness index. While there are no Montecarlo simulations to determine the bias for clustered standard errors, we should take into account that the estimates are likely to be biased towards OLS. Later I increase the strength of the instrument by focusing on the 86% of gangsters who have been arrested at least once, and for whom all variables are more likely to be measured with precision. But test statistics related to the reduced form regressions, which do not suffer from weak‐instrument bias (the Anderson and Rubin p‐value as well as the p‐value on potential innate interactions alone) tend to be below 10%, indicating that the instrument does indeed influence housing values. Instrumenting the centrality measures the coefficients on closeness, eigenvector, as well as degree centrality tend to be three times as large as the OLS equivalents, meaning that a standard deviation increase in either closeness or eigenvector centrality almost doubles the housing value. Such differences are likely to be partly explained by measurement error bias, while there is little evidence that the weakness of the instrument is driving these differences. As in the OLS case, betweenness does not appear to be significantly different from 0, confirming that the gangsters with large bridging capacity tend to keep a lower profile, no matter whether the centrality has been reached exploiting potential innate connections or not. Overall the relative precision of the estimates tends to be smaller than for the OLS regressions, which is in part due to the instrument’s weakness. Indeed, the standard errors are large, and endogeneity test p‐values fluctuate around 10%. Table 7 improves upon the previous results by focusing on samples where variables are measured with greater care. Surnames, which represent the core information for the instrumental variable strategy, are not easily measured inside the mafia. For secrecy reasons, gangsters are typically known by their nickname. This is why the FBN was also collecting the gangsters’ aliases.55 Knowing the exact name is clearly important to reconstruct its geographical distribution in Italy, and for arrested gangsters such information the FBN could be double‐checked. Column 1 shows the baseline results. In column 2 I restrict the analysis to gangsters who have at least one arrest record. These represent about 85% of the total, and are potentially a more selected sample. The Kleibergen‐Paap Wald F‐statistic does indeed improve, and the IV coefficients increase slightly. Table 7 Two‐stage Regressions with Stronger First Stages . (1) . (2) . (3) . (4) . Arrested gangsters only ✓ US born gangsters only ✓ Panel (a): First stage regression log Closeness Potential innate interactions (in %) 0.033** 0.036** 0.027 (0.015) (0.014) (0.018) Zero potential innate interactions −0.019 −0.028 −0.024 (0.020) (0.022) (0.020) PII using whole Italy 0.023** (0.011) Zero PII in whole Italy 0.000 (0.000) R‐squared 0.160 0.163 0.155 0.177 Kleibergen‐Paap rk Wald F statistic 3.479 4.328 2.670 2.234 F‐stat on potential innate interactions 5.208 6.540 4.899 2.394 Panel (b): Main regression log Housing value log Closeness 6.350** 7.103** 8.343** 8.784** (3.196) (3.223) (4.144) (4.344) Observations 641 554 641 429 Endogeneity p‐value 0.212 0.0838 0.168 0.115 p value on PII 0.257 0.0657 0.111 0.0908 Anderson and Rubin p‐value 0.212 0.0838 0.168 0.115 . (1) . (2) . (3) . (4) . Arrested gangsters only ✓ US born gangsters only ✓ Panel (a): First stage regression log Closeness Potential innate interactions (in %) 0.033** 0.036** 0.027 (0.015) (0.014) (0.018) Zero potential innate interactions −0.019 −0.028 −0.024 (0.020) (0.022) (0.020) PII using whole Italy 0.023** (0.011) Zero PII in whole Italy 0.000 (0.000) R‐squared 0.160 0.163 0.155 0.177 Kleibergen‐Paap rk Wald F statistic 3.479 4.328 2.670 2.234 F‐stat on potential innate interactions 5.208 6.540 4.899 2.394 Panel (b): Main regression log Housing value log Closeness 6.350** 7.103** 8.343** 8.784** (3.196) (3.223) (4.144) (4.344) Observations 641 554 641 429 Endogeneity p‐value 0.212 0.0838 0.168 0.115 p value on PII 0.257 0.0657 0.111 0.0908 Anderson and Rubin p‐value 0.212 0.0838 0.168 0.115 Notes. All regressions control for the additional housing variables and the additional regressors as in Table 4. The Southern regions are Puglia, Campania, Calabria, Molise, and Sicilia. The Potential Innate Interactions are based on Southern regions. Clustered (by surname) standard errors in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1. Open in new tab Table 7 Two‐stage Regressions with Stronger First Stages . (1) . (2) . (3) . (4) . Arrested gangsters only ✓ US born gangsters only ✓ Panel (a): First stage regression log Closeness Potential innate interactions (in %) 0.033** 0.036** 0.027 (0.015) (0.014) (0.018) Zero potential innate interactions −0.019 −0.028 −0.024 (0.020) (0.022) (0.020) PII using whole Italy 0.023** (0.011) Zero PII in whole Italy 0.000 (0.000) R‐squared 0.160 0.163 0.155 0.177 Kleibergen‐Paap rk Wald F statistic 3.479 4.328 2.670 2.234 F‐stat on potential innate interactions 5.208 6.540 4.899 2.394 Panel (b): Main regression log Housing value log Closeness 6.350** 7.103** 8.343** 8.784** (3.196) (3.223) (4.144) (4.344) Observations 641 554 641 429 Endogeneity p‐value 0.212 0.0838 0.168 0.115 p value on PII 0.257 0.0657 0.111 0.0908 Anderson and Rubin p‐value 0.212 0.0838 0.168 0.115 . (1) . (2) . (3) . (4) . Arrested gangsters only ✓ US born gangsters only ✓ Panel (a): First stage regression log Closeness Potential innate interactions (in %) 0.033** 0.036** 0.027 (0.015) (0.014) (0.018) Zero potential innate interactions −0.019 −0.028 −0.024 (0.020) (0.022) (0.020) PII using whole Italy 0.023** (0.011) Zero PII in whole Italy 0.000 (0.000) R‐squared 0.160 0.163 0.155 0.177 Kleibergen‐Paap rk Wald F statistic 3.479 4.328 2.670 2.234 F‐stat on potential innate interactions 5.208 6.540 4.899 2.394 Panel (b): Main regression log Housing value log Closeness 6.350** 7.103** 8.343** 8.784** (3.196) (3.223) (4.144) (4.344) Observations 641 554 641 429 Endogeneity p‐value 0.212 0.0838 0.168 0.115 p value on PII 0.257 0.0657 0.111 0.0908 Anderson and Rubin p‐value 0.212 0.0838 0.168 0.115 Notes. All regressions control for the additional housing variables and the additional regressors as in Table 4. The Southern regions are Puglia, Campania, Calabria, Molise, and Sicilia. The Potential Innate Interactions are based on Southern regions. Clustered (by surname) standard errors in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1. Open in new tab Column 3 shows that the results do not change much when computing the potential interactions using all Italian regions, though the power of the instrument decreases considerably. Another sample for which the information might be more precise are the mobsters born in the US. Focusing on this group of criminals addresses also the concern that innate connections might still have independent (direct) influence on current wealth, as this should be less likely to happen for US born ones. The smaller sample reduces the power of the instrument but the instrumental variable coefficient does not change much. In Table 8 I perform a whole series of additional robustness checks, always keeping the entire sample. The first column adds the average log housing value at the MSA level to control for the possibility that the results are simply due to sorting of the most influential mobsters in cities that have the highest housing values. The network effect becomes just a little smaller when controlling for such relative differences in housing prices (indicating that centrality is positively correlated with average housing values). Table 8 Robustness Regressions . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . . 2SLS . 2SLS . 2SLS . 2SLS . 2SLS . 2SLS . OLS . 2SLS . Panel (a): First stage regression log Closeness Potential innate interactions (in %) 0.032*** 0.033** 0.036** 0.032*** 0.027** 0.040 0.034** (0.012) (0.015) (0.015) (0.012) (0.013) (0.039) (0.015) Zero potential innate interactions −0.019 −0.016 −0.012 −0.011 −0.020 −0.017 (0.019) (0.020) (0.019) (0.019) (0.020) (0.020) PII squared −0.007 (0.036) PII cube 0.001 (0.005) R‐squared 0.224 0.178 0.238 0.228 0.306 0.155 0.165 F‐stat on potential innate interactions 7.531 4.950 6.084 6.693 4.165 1.009 5.413 Panel (b): Main regression log Housing value log Closeness 6.370** 6.175* 6.004* 7.294** 6.893* 5.085* 2.486*** 6.274** (3.231) (3.326) (3.108) (3.695) (3.701) (3.010) (0.453) (3.174) Fraction of Sicilian associates −0.634 (0.393) Fraction of Italian associates −0.049 (0.443) Connected wife 0.139 −0.022 (0.143) (0.228) Average log housing value (MSA) ✓ Without controlling for legal businesses ✓ 13 job dummies ✓ 5 modus operandi dummies ✓ 12 crime convictions dummies ✓ Kleibergen‐Paap rk Wald F statistic 4.685 3.118 3.664 3.910 2.829 8.331 3.527 Endogeneity p‐value 0.205 0.355 0.405 0.255 0.309 0.196 0.302 Anderson and Rubin p‐value 0.149 0.257 0.196 0.151 0.227 0.0156 0.212 p‐value on additional dummies 0.229 0.0322 0.590 . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . . 2SLS . 2SLS . 2SLS . 2SLS . 2SLS . 2SLS . OLS . 2SLS . Panel (a): First stage regression log Closeness Potential innate interactions (in %) 0.032*** 0.033** 0.036** 0.032*** 0.027** 0.040 0.034** (0.012) (0.015) (0.015) (0.012) (0.013) (0.039) (0.015) Zero potential innate interactions −0.019 −0.016 −0.012 −0.011 −0.020 −0.017 (0.019) (0.020) (0.019) (0.019) (0.020) (0.020) PII squared −0.007 (0.036) PII cube 0.001 (0.005) R‐squared 0.224 0.178 0.238 0.228 0.306 0.155 0.165 F‐stat on potential innate interactions 7.531 4.950 6.084 6.693 4.165 1.009 5.413 Panel (b): Main regression log Housing value log Closeness 6.370** 6.175* 6.004* 7.294** 6.893* 5.085* 2.486*** 6.274** (3.231) (3.326) (3.108) (3.695) (3.701) (3.010) (0.453) (3.174) Fraction of Sicilian associates −0.634 (0.393) Fraction of Italian associates −0.049 (0.443) Connected wife 0.139 −0.022 (0.143) (0.228) Average log housing value (MSA) ✓ Without controlling for legal businesses ✓ 13 job dummies ✓ 5 modus operandi dummies ✓ 12 crime convictions dummies ✓ Kleibergen‐Paap rk Wald F statistic 4.685 3.118 3.664 3.910 2.829 8.331 3.527 Endogeneity p‐value 0.205 0.355 0.405 0.255 0.309 0.196 0.302 Anderson and Rubin p‐value 0.149 0.257 0.196 0.151 0.227 0.0156 0.212 p‐value on additional dummies 0.229 0.0322 0.590 Notes. The additional controls are the same as in the last columns of Table 4. Column 4 is based on the subset of mobsters born in the US, all the other regressions use the entire sample. The Potential Innate Interactions are based on Southern regions. Clustered (by 530 surnames) standard errors in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1. Open in new tab Table 8 Robustness Regressions . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . . 2SLS . 2SLS . 2SLS . 2SLS . 2SLS . 2SLS . OLS . 2SLS . Panel (a): First stage regression log Closeness Potential innate interactions (in %) 0.032*** 0.033** 0.036** 0.032*** 0.027** 0.040 0.034** (0.012) (0.015) (0.015) (0.012) (0.013) (0.039) (0.015) Zero potential innate interactions −0.019 −0.016 −0.012 −0.011 −0.020 −0.017 (0.019) (0.020) (0.019) (0.019) (0.020) (0.020) PII squared −0.007 (0.036) PII cube 0.001 (0.005) R‐squared 0.224 0.178 0.238 0.228 0.306 0.155 0.165 F‐stat on potential innate interactions 7.531 4.950 6.084 6.693 4.165 1.009 5.413 Panel (b): Main regression log Housing value log Closeness 6.370** 6.175* 6.004* 7.294** 6.893* 5.085* 2.486*** 6.274** (3.231) (3.326) (3.108) (3.695) (3.701) (3.010) (0.453) (3.174) Fraction of Sicilian associates −0.634 (0.393) Fraction of Italian associates −0.049 (0.443) Connected wife 0.139 −0.022 (0.143) (0.228) Average log housing value (MSA) ✓ Without controlling for legal businesses ✓ 13 job dummies ✓ 5 modus operandi dummies ✓ 12 crime convictions dummies ✓ Kleibergen‐Paap rk Wald F statistic 4.685 3.118 3.664 3.910 2.829 8.331 3.527 Endogeneity p‐value 0.205 0.355 0.405 0.255 0.309 0.196 0.302 Anderson and Rubin p‐value 0.149 0.257 0.196 0.151 0.227 0.0156 0.212 p‐value on additional dummies 0.229 0.0322 0.590 . (1) . (2) . (3) . (4) . (5) . (6) . (7) . (8) . . 2SLS . 2SLS . 2SLS . 2SLS . 2SLS . 2SLS . OLS . 2SLS . Panel (a): First stage regression log Closeness Potential innate interactions (in %) 0.032*** 0.033** 0.036** 0.032*** 0.027** 0.040 0.034** (0.012) (0.015) (0.015) (0.012) (0.013) (0.039) (0.015) Zero potential innate interactions −0.019 −0.016 −0.012 −0.011 −0.020 −0.017 (0.019) (0.020) (0.019) (0.019) (0.020) (0.020) PII squared −0.007 (0.036) PII cube 0.001 (0.005) R‐squared 0.224 0.178 0.238 0.228 0.306 0.155 0.165 F‐stat on potential innate interactions 7.531 4.950 6.084 6.693 4.165 1.009 5.413 Panel (b): Main regression log Housing value log Closeness 6.370** 6.175* 6.004* 7.294** 6.893* 5.085* 2.486*** 6.274** (3.231) (3.326) (3.108) (3.695) (3.701) (3.010) (0.453) (3.174) Fraction of Sicilian associates −0.634 (0.393) Fraction of Italian associates −0.049 (0.443) Connected wife 0.139 −0.022 (0.143) (0.228) Average log housing value (MSA) ✓ Without controlling for legal businesses ✓ 13 job dummies ✓ 5 modus operandi dummies ✓ 12 crime convictions dummies ✓ Kleibergen‐Paap rk Wald F statistic 4.685 3.118 3.664 3.910 2.829 8.331 3.527 Endogeneity p‐value 0.205 0.355 0.405 0.255 0.309 0.196 0.302 Anderson and Rubin p‐value 0.149 0.257 0.196 0.151 0.227 0.0156 0.212 p‐value on additional dummies 0.229 0.0322 0.590 Notes. The additional controls are the same as in the last columns of Table 4. Column 4 is based on the subset of mobsters born in the US, all the other regressions use the entire sample. The Potential Innate Interactions are based on Southern regions. Clustered (by 530 surnames) standard errors in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1. Open in new tab Column 2 does not control for the number of legal businesses owned by the mobster, which reduces slightly the coefficient. Column 3 deals again with the potential violations of the exclusion restriction. Since some of the mobsters, with their surnames, might have migrated as a response to increased wealth levels, I add the fraction of direct links (associates) that were born in Sicily and in peninsular Italy to the regression. Doing so leaves the results almost unchanged. One could argue that even controlling for the sum of the types of crimes for which the criminals have been arrested and for the sum of legal businesses that the gangster owns, the exclusion restrictions might not hold whenever the potential innate interactions not only influence the gangsters’ centrality but also additional remunerative abilities that are not channelled through connections. While there is no definite test to exclude such a possibility, in columns 4 and 5 I control for a series of dummies that describe the gangster’s crimes, modus operandi and legal jobs, and potentially control for such unobserved remunerative ability. The potential downside is that adding these controls might capture part of the network effect. There are 13 crime categories with at least 3% of arrested gangsters (murder, robbery, burglary, larceny, drug offence, prohibition offence, illegal detention of weapons, menace, theft, assault, fraud, disorderly behaviour and other crimes), six modus operandi that come from FBN descriptions of the gangster’s illegal activities (racketeering, drug dealing, gambling, bookmaking, murder and other), and finally 14 job descriptions with at least 3% of mobsters involved (restaurant, drugstore, food industry, real estate, transportation, hotel, import‐export, automotive, clothing, manual labour, raw materials, amusement, casino and others). While some of these dummy variables are correlated with both the centrality of mobsters and the log housing value, the network effect changes little when adding these controls and, if anything, becomes even larger. Column 6 tests the robustness of the results when getting rid of the zero interaction dummy (thus assuming continuity and linearity at 0) and allowing the instrument to have non‐linear effects on log closeness. The network effect because smaller but is still significant at the 10% level. Finally, in the last two columns I add a variable that is clearly endogenous, whether the mobster is married to a ‘connected’ wife, meaning a wife whose maiden name is also the surname of another mobster in the data. These marriages tend to be arranged for strategic reasons and allow gangsters to gain additional power (connections) as well as wealth. Adding this variable increases the OLS estimates while keeping the 2SLS estimate almost unchanged, suggesting that omitted variables (which the instrument takes care of) are indeed biasing the OLS estimate towards zero.56 Overall, no matter the specification an SD increase in closeness centrality – for instance, such difference would mean moving from a median closeness to the 90th percentile – increases housing values by about 75%. 4. Concluding Remarks This article estimates how, in 1960, network centrality inside the Italian‐American mafia influenced gangsters’ economic prosperity, measured by the value of their homes. In the overground world, the whole geometry of connections has been shown to be related to a variety of economic outcomes. In the underground world such connections are presumably even more important and yet evidence on this has mainly been based on ethnographic studies. Moreover, even in the overground world researchers have rarely gone beyond just documenting correlations, as networks tend to emerge endogenously out of complicated bilateral and multilateral decision processes, and researchers often observe only subsets of the entire network. I deal with the non‐random sampling design modelling the law‐enforcement data collection process within the network as a Markov process. With respect to the endogenous nature of networks, social scientists have been able to exploit the geometry of the network to develop identification strategies for direct connections (peer effects) but not yet for measures of how central agents are inside networks. Instruments for networks with reasonable exclusion restrictions are in short supply. Any characteristic that determines someone’s position inside his/her network is also likely to directly influence a multitude of other outcomes. In the mafia, for example, family relationships, wealth, place of birth etc. might help in securing a centrality in the network but could also be related to the demand for housing. For migrants with strong ethnic identities, instruments naturally evolve from shocks that happen in the country of origin and are thus less likely to influence economic outcomes in the country of destination. Munshi (2003) uses rainfall in Mexico to instrument for the network size of Mexican immigrants to study how such size influences labour market outcomes. Similarly, this article instruments network centrality using the potential exposure to connections in the gangster’s place of origin (30–50 years earlier). In the absence of pre‐immigration data, I use the informational content of surnames (isonomy) to measure such place. Individual home values in 1960 are reconstructed starting from today’s values. Notwithstanding the distant past and the approximation lead to weak instruments and larger standard errors, the evidence suggests that centrality measures have a sizeable influence on housing values. Mobsters who are on average closer to their peers tend to live in more expensive housing. This is consistent with the importance of interconnection capacity in a secret society where two unlinked mobsters need a common associate to generate a new (direct) link. But then, mobsters who act as bridges across clusters of the larger network57 tend to keep a lower profile and officially reside in less expensive housing (either because they hide their true residence through the use of figureheads or because they really prefer so). The evidence suggests that these tend to be the more important bosses, those who most likely form the governing body of the mafia (la Commissione). In line with Bonnano’s autobiography such bosses were less likely to be directly involved in the narcotics businesses, which might be part of the same ‘attention avoidance strategy’ (Bonanno, 1983). The geometry of the network is likely to influence not just the economic prosperity of the ‘made’ men but also their criminal activity (and vice versa); and therefore the overall crime rates.58 Understanding how network effects shape the portfolio of crimes that gangsters commit represents a natural extension of this work. Finally, while data restrictions prevent researchers from performing similar analyses based on more recent organised crime networks, this might hopefully change in the near future. Understanding how central figures grow up inside criminal networks is fundamental to the design of targeted law enforcement strategies. Appendix A: A Short History of the Mafia and The American Mafia This Section is a brief history of the so‐called ‘made’ men who came to the US and a description of how the mafia operated in the 1960s when the FBN was filing the 800 records. There is evidence of mafia style organisations (Camorra in Campania, ’Ndrangheta in Calabria, and mafia in Sicily) operating in Italian post‐unification (1861) territory (see Buonanno et al., 2015). The collapse of Southern feudal institutions when the Northern troops conquered their territories led to a sudden lack of law enforcement and, with a rising demand for protection, local bandits started offering such services (Gambetta, 1996). The additional ingredient that might have contributed to the emergence of the mafia is what Banfield (1967) described as a self‐interested, family centric society which sacrificed the public good, or social capital. Lack of social capital has probably prevented an early eradication of the mafia in Southern Italy.59 This means that the land that Southern Italian migrants were leaving behind some 50 years later was already in the hands of the mafia. Historians define two major waves of immigration from Sicily, before and after World War I (WWI). Before WWI immigrants were mainly driven by economic needs. Several mafia bosses, like Lucky Luciano, Tommaso Lucchese, Vito Genovese, Frank Costello etc, were children of these early immigrants. Even though between 1901 and 1913 almost a quarter of Sicily’s population departed for America, many of the early immigrant families were not from Sicily. In that period around 2 million Italians, mainly from the South emigrated to the US (Critchley, 2009). These baby immigrants later became street gang members in the slums; they spoke little Italian and worked side by side with criminals from other ethnicities, mainly Jewish and Irish (Lupo, 2009). Lured by the criminal successes of the first wave of immigrants and (paradoxically) facilitated by prohibition, the second wave of immigrants that went on to become mafia bosses were already criminals by the time they entered the US. Charles Gambino, Joe Profaci, Joe Bonanno, and others were in their 20s and 30s when they first entered the US (the average age in 1960 of Italian born gangsters is indeed six years higher (54 years) than for American born ones) and they all came from Sicily. Another reason for this selection of immigrants was the fascist crack‐down of the mafia, which forced some of these criminals to leave Sicily. After the second wave of immigration the mafia became more closely linked to the Sicilian mafia and started adopting its code of honour and tradition (see Gosch and Hammer, 1975). In 1930 and 1931 these new arrivals led to a mafia war, called the Castellamare war, named after a small city in Sicily where many of the new mafia bosses came from. The war lasted until Maranzano, who was trying to become the ‘Boss of the Bosses’, was killed, probably by Lucky Luciano who had joined the Masseria Family.60 This war put Lucky Luciano at the top of the mafia organisation but also led to a reaction by the media and the prosecutors.61 Between 1950 and 1951, the Kefauver Committee, officially the Senate Special Committee to Investigate Crime in Interstate Commerce, had a profound impact on the American public. It was the first committee set up to gain a better understanding of how to fight organised crime, and the main source of information was a list of 800 suspected criminals submitted by FBN’s Commissioner Anslinger, most likely an early version of the records used in this article(McWilliams, 1990, p. 141).62 Throughout the 1950s the FBN continued to investigate the mafia and, in 1957, an unexpected raid of an American mafia summit, the ‘Apalachin meeting’, captured considerable media attention. Police detained over 60 underworld bosses from the raid. After that meeting everyone had to agree with the FBN’s view that there was one large and well organised mafia society. Robert Kennedy, attorney general of the US, and J. Edgar Hoover, head of the Federal Bureau of Investigations, joined Harry J. Anslinger, the US Commissioner of Narcotics, in his war against the mob. The same year a permanent Senate Select Committee was formed – the McClellan commission. Anslinger’s FBN conducted the investigative work and coordinated nationwide arrests of Apalachin defendants. Lucky Luciano died of a heart attack at the airport of Naples in 1962. Footnotes 1 " See The New York Times, 21 January 2011, page A21 of the New York edition. 2 " According to the FBI, in 2005 there were 651 pending investigations related to the Italian‐American mafia; almost 1,500 mobsters were arrested and 824 were convicted; of the roughly 1,000 ‘made’ members of Italian organised crime groups estimated to be active in the US, 200 were in jail (see http://www.fbi.gov). 3 " The distribution of the year of the first arrest of mobsters has almost full support within the range 1908–60, so one can infer that the data refer to what the authorities knew in 1960. 4 " In the 1930s and up to the 1950s the FBN, which later merged with the Bureau of Drug Abuse Control to form the Bureau of Narcotics and Dangerous Drugs, was the main authority in the fight against the mafia (Critchley, 2009). For example, in New York, the Federal Bureau of Investigation had just four agents, mainly working in the office, assigned to the mafia, while in the same office more than 400 agents were fighting domestic communists (Maas, 1968). 5 " These rules were listed both, in a 1963 testimony by the first FBN informant, Joe Valachi, and in a piece of paper that belonged to the Italian mafia boss Salvatore Lo Piccolo during his 2007 arrest. 6 " The closeness centrality index measures the average distance between a node (a member) and all the other nodes, and its inverse, is a good measure for how isolated members are. The betweenness centrality index measures the number of times a node is on the shortest path between two randomly chosen nodes. Ductor et al. (2014) use both to predict research output of individual researchers. Kinnan and Townsend (2012) look at how network distance to a bank affects consumption smoothing. 7 " Degree is often used to measure network importance. For example, Kremer and Miguel (2007) use it to study the diffusion of a deworming pill take‐up, while Hochberg et al. (2007) show that venture capitalists’ investment performance depends on how connected they are. 8 " A large literature has shown the link between housing demand and income (See Goodman, 1988). The online Appendix Figure B3 shows a correlation of 60% between zip code‐level median housing prices and median household income, with an elasticity that is very close to 1. 9 " Any classical measurement error would inflate the standard errors, making the inference more conservative. 10 " Morselli (2003) analyses connections within a single New York based family (the Gambino family), Krebs (2002) analyses connections among the September 2001 hijackers’ terrorist cells, Natarajan (2000, 2006) analyses wiretap conversations among drug dealers and McGloin (2005) analyses the connections among gang members in Newark (NJ). There is considerable more theoretical work. Most studies have focused on a market structure view of organised crime, where the mafia generates monopoly power in legal (for a fee) and illegal markets. Among others, such a view is present in the collection of papers in Fiorentini and Peltzman (1997), and in Reuter (1983), Abadinsky (1990), Gambetta (1996) and Kumar and Skaperdas (2009). Only two theoretical papers have focused on the internal organisation of organised crime groups. Garoupa (2007) looks at the optimal size of these organisations, while Baccara and Bar‐Isaac (2008) look at the optimal internal structure (cells versus hierarchies). 11 " Alternatively, one can avoid making any causal claims. Ductor et al. (2014) focus on predictions, and show that researchers’ network centralities help to predict future research output. 12 " Experimental studies on networks usually take the network as given and randomly assign information or other treatments to single network nodes (Kremer and Miguel, 2007; Alatas et al., 2012; Fafchamps et al., 2013). While such experimental variation does not solve the endogeneity issue, Banerjee et al. (2013) develop and test a model of information diffusion, validating such model based on unexploited variation in the data (Blume et al., 2012). 13 " For example, Munshi (2003) uses rainfall in the origin‐community as an instrument for the size of the network at the destination (the US). Mexican immigrants with larger networks in the US face better labour conditions. 14 " The more commonly studied peer effects can be seen as specific network effects, where only direct links are assumed to matter. Researchers have used this restriction to estimate how criminals’ behaviour depends on the behaviour of their peers (Sarnecki, 1990,2001; Baker and Faulkner, 1993; Haynie, 2001; Sirakaya, 2006; Patacchini and Zenou, 2008; Bayer et al., 2009; Drago and Galbiati, 2012). An old and extensive literature in labour economics documents the importance of friends and relatives in providing job referrals (Montgomery, 1991; Glaeser et al., 1996; Bayer et al., 2008). In recent years the interest has shifted towards understanding not just peer influence, or the influence of direct links, but how the whole architecture of a network, thus including indirect links, influences behaviour and outcomes (Ballester et al., 2006; Goyal, 2007; Vega‐Redondo, 2007; Jackson, 2008). Empirical evidence on these ‘network effects’ is scarce but growing (see, among others Hochberg et al., 2007; Angelucci et al., 2010; Alatas et al., 2012; Kinnan and Townsend, 2012; Banerjee et al., 2013), with the main burden being the endogeneity of the network (Blume et al., 2012). 15 " See subsection 2.4 for a thorough discussion about the instrument. Such instrument is also related to the growing literature on trust and family values. Guiso et al. (2006) present an introduction to the importance of culture, defined as ‘customary beliefs and values that ethnic, religious, and social groups transmit fairly unchanged from generation to generation’, on economic behaviour. The same applies to criminal behaviour. 16 " Karlan et al. (2009) build and estimate a model of trust in social networks between a lender and a borrower based on the collateral that is provided by their weakest links. 17 " I construct the undirected network, meaning that i and j with at least one of the two records lists the other surname among the associates. 18 " Jacobs and Gouldin (1999) provide a relatively short overview about law enforcement’s unprecedented attack on Italian organised crime families following Valachi’s hearings. 19 " Valachi revealed that the Cosa Nostra was made of approximately 25 Families. Cosa Nostra was governed by a commissione of 7–12 bosses, which also acted as the final arbiter on disputes between families. The remaining 10–15 families were smaller and not part of Cosa Nostra’s governing body. Each family was structured in hierarchies with a boss (capo famiglia) at the top, a second in command, called underboss (sottocapo), a counsellor (consigliere) and several captains (caporegime) who head a group of soldiers (regime) (Maas, 1968). 20 " The Perron–Frobenius theorem ensures that when the Markov chain is ergodic such a vector exists and is unique. I approximate pby p40 ⁠, and compute such distribution multiplying a constant vector of size N (number of nodes) that sums up to one by the 40th power of T. This result is related to the convergence in snowball sampling, called respondent‐driven‐sampling (Heckathorn, 1997). 21 " See Golub and Jackson (2010) for a discussion about selection bias in networks. 22 " Such bias would be even less severe under respondent‐driven sampling. 23 " The only historical data on individual housing units that goes so far back in time are rental prices and sale prices from newspaper articles (Margo, 1996), but it would be close to impossible to find those exact same housing units occupied by the mobsters. 24 " One third of the times the exact address did not produce an estimated value, so the nearest house with such information was selected. Since housing values tend to be highly geographically clustered such proxy is likely to reduce the precision of our estimates by a small amount. 25 " While these mobsters do not contribute directly to the analysis, they are part of the network and are used to construct measures of network centrality. 26 " See http://www.zillow.com/howto/DataCoverageZestimateAccuracy.htm. 27 " See http://www.census.gov/hhes/www/housing/census/historic/values.html. 28 " Table 1 shows that 10% of the houses found on Zillow.com were built after 1960. To control for the fact that these houses might have a different valuation I am going to control for a dummy variable equal to one when the houses were built after 1960. 29 " Bleakley and Lin (2012) show persistence across cities driven by now obsolete portage locations. 30 " See the online Appendix Figure B4. 31 " Joe Bonnano is a good example. While being an NYC boss his official residence was for strategic reasons in Tucson, Arizona. 32 " In other words, I construct a symmetric adjacency matrix of indirected connections between mobsters’ last names. Dealing with changing first names would have been a complex task. 33 " See also Sparrow (1991) for a discussion on centrality indices in criminal networks. 34 " It equals the eigenvector of the largest positive eigenvalue of the adjacency matrix, the N × N matrix of zeros and ones (indicating whether gangster i and j are connected). 35 " As first noted by Granovetter (1973), weak ties (i.e. friends of friends) are an important source of information. While Ballester et al. (2006, 2010) show that in non‐cooperative games the activity of individuals is proportional to the eigenvector centrality (the ‘key‐player’ having the largest eigenvector centrality), several assumptions of that model would not hold for the mafia: conditional on operating inside the same area and being part of the same mafia ‘Family’, the hierarchy (introducing cooperativeness) as well the implied risk such connections are likely to influence the activity of mobsters. 36 " For the betweenness centrality index, since 4.5% of observations have such index equal to zero, I take the inverse hyperbolic sine transformation log[y+(y2+1)1/2] ⁠. 37 " Soldiers have a median betweenness index of about 0.18. 38 " Bosses also show an aversion to drug dealing, a very lucrative business. Later we will see that gangsters involved in drug dealing live is houses that are about 30% more expensive. 39 " Recently arrested bosses who were heading the entire Sicilian mafia, Totó Riina and Bernardo Provenzano, were living in very poor houses. Such cautious behaviour seems less present in other organised crime groups. A recently arrested boss from the Neapolitan Camorra, Francesco Schiavone, was living in a mansion, built in the style of the house in the Hollywood movie Scarface. This same pattern between the Italian region of birth and housing values is evident in the data. 40 " About a quarter of mobsters were born in Sicily, 2/3 were born outside of Italy (mostly in the US) and the rest in other regions of Italy. Properly weighting the data, these fractions are 2/10, 7/10 and 1/10, indicating that Sicilians tend to have more connections. 41 " Table 1 shows that the average age is 48 years, which means that the average year of birth is 1912, right in the middle of the Italian migration wave (see Appendix A). Most mobsters are either first or second generation immigrants. All but a handful of mobsters were of Italian origin, as this was a prerequisite to become a member. The few non‐Italian gangsters in the data were either French gangsters from Marseille or Corsica, or part of the, so‐called, Jewish mafia. 42 " See Colantonio et al. (2003) for an overview on recent developments on the use of surnames in human population biology. 43 " One can try out surnames of Italian economists on the following web sites: http://www.gens.info/ or http://www.paginebianche.it/. 44 " Guglielmino et al. (1991) show that in Sicily genetic and cultural transmission are revealed by surnames. 45 " Only four mobsters were neither born in Italy nor in the US. For two of these mobsters (Lansky and Genese) the Potential Innate Interactions index is zero. 46 " I use Italy’s phone directory http://www.paginebianche.it/ in 2010. Ideally one would use the distribution of surnames in 1960, though previous research has shown how persistent such distribution is (Colantonio et al., 2003). 47 " In the regressions I allow the zeros to have an independent effect. Moreover, ideally one would us the mother’s surname as well, though such information is not always available. 48 " Pull factors of migration that are driven by social networks would preserve some of the original predictive power. 49 " Given that there is one large network, the residuals might be correlated across mobsters. Assuming that such correlation depends on the shortest distance between pairs of gangsters the variance of the residual of gangster i is going to be a function of the sum of all such distances dij over j. In particular, σi=∑jσij(dij) ⁠. Approximating such variance with either a linear function or an exponential function in average distances (the inverse of closeness) one always rejects that distance influences the squared residuals and thus the correlations across mobsters (see online Appendix Table C1). Nevertheless, in all regressions the standard errors are clustered at the surname level, which is the level at which the centrality measures are calculated. 50 " Measurement error might also be co‐responsible for such drop. 51 " Online Appendix Table C2 shows that the place of birth, while influencing the housing value, does not introduce heterogeneity in the effect of centrality. 52 " The elasticity is larger for closeness centrality only because its variation is about 10 times smaller than the one of eigenvector centrality. One standard deviation increase in log‐closeness centrality has about the same impact on housing value as a one standard deviation increase in log eigenvector centrality. 53 " All the regressions use the weighting strategy developed earlier on, but the results based on unweighted regressions can be seen in the online Appendix Table C3. 54 " Online Appendix Tables C4 and C5 show all the coefficients. 55 " Some of the aliases for gangsters mentioned before were: Don Vitone, The Old Man (Vito Genovese); Francisco Castiglia, Frank Costello, Frank Saverio, Saveria (Francisco Costiglia); Joe Bananas, Joe Bononno, Joe Bonnano, Joe Bouventre (Joseph Bonanno), Joe Proface (Giuseppe Profaci); Carlo Gambrino, Carlo Gambrieno, Don Carlo (Carlo Gambino). 56 " Given the potential endogeneity of such marriage one cannot draw more definite conclusions. 57 " Given what is known about the mafia these are mafia families, clans, or mandamenti. 58 " Mastrobuoni and Patacchini (2012) show that cities with larger mafia densities (measured by the average eigenvector index) exhibit larger violent crime rates. 59 " Putnam et al. (1994) documents large differences in social capital between Southern and Northern Italy and Pinotti (2015) shows not only that the mafia is expanding to other Southern regions, but also that the cost of such expansions are large in terms of lost economic growth. 60 " Before this event, in order to end the power‐struggle between Masseria and Maranzano, Lucky Luciano had offered to eliminate Joe ‘the Boss’ Masseria, which he did at an Italian restaurant by poisoning Masseria’s food with lead. 61 " In 1936 Thomas E. Dewey, appointed as New York City special prosecutor to crack down on the rackets, managed to obtain Luciano’s conviction with charges on multiple counts of forced prostitution. Luciano served only 10 of the 30–50 years sentenced. In 1946 thanks to an alleged involvement in the Allied troops’ landing in Sicily he was deported to Italy, from where he tried to keep organising ‘the organisation’. 62 " The Committee could not prove the existence of a mafia and after Luciano’s expatriation several other families headed the organisation: Costello, Profaci, Bonanno, and Gambino. Family ties were of utmost importance. According to Bonanno’s autobiography (Bonanno, 1983), he became the Boss of the Bosses in part by organising the marriage between his son Bill and the daughter of Profaci, Rosalia in 1956. In 1957 Gambino took over the leadership. References Abadinsky , H . ( 1990 ). Organized Crime , Belmont, CA . Wadsworth Publishing . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Alatas , V ., Banerjee , A., Arun G. Chandrasekhar , R.H. and Olken , B.A. ( 2012 ). ‘ Network structure and the aggregation of information: theory and evidence from Indonesia ’, NBER working paper series, Number 18351. Allesina , S . ( 2011 ). ‘ Measuring nepotism through shared last names: the case of Italian academia ’, PLoS One , vol. 6 ( 8 ), p. e21160 . Google Scholar Crossref Search ADS PubMed WorldCat Angelucci , M ., De Giorgi , G., Rangel , M.A. and Rasul , I. ( 2010 ). ‘ Family networks and school enrolment: evidence from a randomized social experiment ’, Journal of Public Economics , vol. 94 ( 3 ), pp. 197 – 221 . Google Scholar Crossref Search ADS WorldCat Baccara , M . and Bar‐Isaac , H. ( 2008 ). ‘ How to organize crime ’, Review of Economic Studies , vol. 75 ( 4 ), pp. 1039 – 67 . Google Scholar Crossref Search ADS WorldCat Baker , W.E . and Faulkner , R.R. ( 1993 ). ‘ The social organisation of conspiracy: illegal networks in the heavy electrical equipment industry ’, American Sociological Review , vol. 58 ( 6 ), pp. 837 – 60 . Google Scholar Crossref Search ADS WorldCat Ballester , C ., Zenou , Y. and Calvó‐Armengol , A. ( 2006 ). ‘ Who’s who in networks. Wanted: the key player ’, Econometrica , vol. 74 ( 5 ), pp. 1403 – 17 . Google Scholar Crossref Search ADS WorldCat Ballester , C ., Zenou , Y. and Calvó‐Armengol , A. ( 2010 ). ‘ Delinquent networks ’, Journal of the European Economic Association , vol. 8 ( 1 ), pp. 34 – 61 . Google Scholar Crossref Search ADS WorldCat Banerjee , A ., Banerjee , A., Chandrasekhar , A.G., Duflo , E. and Jackson , M.O. ( 2013 ). ‘ The diffusion of microfinance ’, Science , vol. 341 ( 1236498 ), pp. 363 – 70 . OpenURL Placeholder Text WorldCat Banfield , E.C . ( 1967 ). The Moral Basis of a Backward Society , New York, NY : Free Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Barrai , I ., Rodriguez‐Larralde , A., Mamolini , E. and Scapoli , C. ( 1999 ). ‘ Isonymy and isolation by distance in italy ’, Human Biology , vol. 71 ( 6 ), pp. 947 – 61 . Google Scholar PubMed OpenURL Placeholder Text WorldCat Bayer , P. , Hjalmarsson , R. and Pozen , D. ( 2009 ). ‘ Building criminal capital behind bars: peer effects in juvenile corrections ’, Quarterly Journal of Economics , vol. 124 (1), pp. 105 – 47 . Google Scholar Crossref Search ADS WorldCat Bayer , P. , Ross , S. and Topa , G. ( 2008 ). ‘ Place of work and place of residence: informal hiring networks and labor market outcomes ’, Journal of Political Economy , vol. 116 ( 6 ), pp. 1150 – 96 . Google Scholar Crossref Search ADS WorldCat Bleakley , H . and Lin , J. ( 2012 ). ‘ Portage and path dependence ’, The Quarterly Journal of Economics , vol. 127 (2), pp. 587 – 644 . Google Scholar Crossref Search ADS PubMed WorldCat Blume , L.E ., Brock , W.A., Durlauf , S.N. and Jayaraman , R. ( 2012 ). ‘ Linear Social Network Models ’, NBER Working Paper Series, Number. 19212. Bonanno , J . ( 1983 ). A Man of Honor: The Autobiography of Joseph Bonanno , New York : St. Martin’s . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Borgatti , S.P . ( 2004 ). ‘The key player proble111’, in ( R. Breiger, K.M. Carley and P. Pattison, eds.), Dynamic Social Network Modeling and Analysis: 2002 Workshop Summary and Papers , pp. 241 – 52 , Washington, DC : National Academies Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Buonanno , P ., Durante , R., Prarolo , G. and Vanin , P. ( 2015 ). ‘ Poor institutions, rich mines: resource curse and the origins of the sicilian mafia ’, Economic Journal . OpenURL Placeholder Text WorldCat Chandrasekhar , A.G . and Lewis , R. ( 2011 ). ‘ Econometrics of sampled networks ’, mimeo, Stanford University, available at http://stanford.edu/~arungc/CL.pdf (last accessed 8 February 2015). Colantonio , S ., Lasker , G.W., Kaplan , B.A. and Fuster , V. ( 2003 ). ‘ Use of surname models in human population biology: a review of recent developments ’, Human Biology , vol. 75 ( 6 ), pp. 785 – 807 . Google Scholar Crossref Search ADS PubMed WorldCat Coles , N . ( 2001 ). ‘ It’s not what you know it’s who you know that counts. Analysing serious crime groups as social networks ’, British Journal of Criminology , vol. 41 ( 4 ), pp. 580 . Google Scholar Crossref Search ADS WorldCat Critchley , D . ( 2009 ). The Origin of Organized Crime in America: The New York City Mafia, 1891–1931 , New York, NY : Routledge . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Drago , F . and Galbiati , R. ( 2012 ). ‘ Indirect effects of a policy altering criminal behavior: evidence from the italian prison experiment ’, American Economic Journal: Applied Economics , vol. 4 ( 2 ), pp. 199 – 218 . Google Scholar Crossref Search ADS WorldCat Ductor , L ., Fafchamps , M., Goyal , S. and van der Leij , M.J. ( 2014 ). ‘ Social networks and research output ’, Review of Economics and Statistics , vol. 96 ( 5 ), pp. 936 – 48 . Google Scholar Crossref Search ADS WorldCat Fafchamps , M ., Vaz , A. and Vicente , P.C. ( 2013 ). ‘ Voting and peer effects: experimental evidence from Mozambique ’, NOVAFRICA Working Paper No. 1303. Falcone , G . and Padovani , M. ( 1991 ). Cose di Cosa Nostra , Milan : Rizzoli . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Fiorentini , G . and Peltzman , S. ( 1997 ). The Economics of Organised Crime , Cambridge : Cambridge University Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Frank , O . ( 1979 ). ‘ Sampling and estimation in large social networks ’, Social Networks , vol. 1 ( 1 ), pp. 91 – 101 . Google Scholar Crossref Search ADS WorldCat Gambetta , D . ( 1996 ). The Sicilian Mafia: The Business of Private Protection , Cambridge, MA : Harvard University Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Garoupa , N . ( 2007 ). ‘ Optimal law enforcement and criminal organization ’, Journal of Economic Behavior & Organization , vol. 63 ( 3 ), pp. 461 – 74 . Google Scholar Crossref Search ADS WorldCat Glaeser , E.L ., Sacerdote , B. and Scheinkman , J.A. ( 1996 ). ‘ Crime and social interactions ’, Quarterly Journal of Economics , vol. 111 ( 2 ), pp. 507 – 48 . Google Scholar Crossref Search ADS WorldCat Golub , B . and Jackson , M.O. ( 2010 ). ‘ Using selection bias to explain the observed structure of internet diffusions ’, Proceedings of the National Academy of Sciences , vol. 107 ( 24 ), pp. 10833 – 6 . Google Scholar Crossref Search ADS WorldCat Goodman , A.C . ( 1988 ). ‘ An econometric model of housing price, permanent income, tenure choice, and housing demand ’, Journal of Urban Economics , vol. 23 (3), pp. 327 – 53 . Google Scholar Crossref Search ADS WorldCat Goodman , L.A . ( 1961 ). ‘ Snowball sampling ’, Annals of Mathematical Statistics , vol. 32 ( 1 ), pp. 148 – 70 . Google Scholar Crossref Search ADS WorldCat Gosch , M.A . and Hammer , R. ( 1975 ). The Last Testament of Lucky Luciano , Boston, MA : Little, Brown . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Goyal , S . ( 2007 ). Connections: An Introduction to the Economics of Networks , Princeton, NJ : Princeton University Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Granovetter , M.S . ( 1973 ). ‘ The strength of weak ties ’, American Journal of Sociology , vol. 78 ( 6 ), pp. 1360 – 80 . Google Scholar Crossref Search ADS WorldCat Granovetter , M . ( 1976 ). ‘ Network sampling: some first steps ’, American Journal of Sociology , vol. 81 ( 6 ), pp. 1287 – 303 . Google Scholar Crossref Search ADS WorldCat Guglielmino , C ., Zei , G. and Cavalli‐Sforza , L. ( 1991 ). ‘ Genetic and cultural transmission in sicily as revealed by names and surnames ’, Human Biology , vol. 63 ( 5 ), pp. 607 – 27 . Google Scholar PubMed OpenURL Placeholder Text WorldCat Guiso , L ., Sapienza , P. and Zingales , L. ( 2006 ). ‘ Does culture affect economic outcomes? ’, Journal of Economic Perspectives , vol. 20 ( 2 ), pp. 23 – 48 . Google Scholar Crossref Search ADS WorldCat Gyourko , J ., Mayer , C. and Sinai , T. ( 2013 ). ‘ Superstar cities ’, American Economic Journal: Economic Policy , vol. 5 ( 4 ), pp. 167 – 99 . Google Scholar Crossref Search ADS WorldCat Haynie , D.L . ( 2001 ). ‘ Delinquent peers revisited: does network structure matter? ’, American Journal of Sociology , vol. 106 ( 4 ), pp. 1013 – 57 . Google Scholar Crossref Search ADS WorldCat Heckathorn , D.D . ( 1997 ). ‘ Respondent‐driven sampling: a new approach to the study of hidden populations ’, Social Problems , vol. 44 ( 2 ), pp. 174 – 99 . Google Scholar Crossref Search ADS WorldCat Hochberg , Y.V ., Ljungqvist , A. and Lu , Y. ( 2007 ). ‘ Whom you know matters: venture capital networks and investment performance ’, Journal of Finance , vol. 62 ( 1 ), pp. 251 – 301 . Google Scholar Crossref Search ADS WorldCat Ianni , F.A . and Reuss‐Ianni , E. ( 1972 ). A Family Business: Kinship and Social Control in Organized Crime , New York, NY : Russell Sage Foundation . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Jackson , M.O . ( 2008 ). Social and Economic Networks , Princeton, NJ : Princeton University Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Jacobs , J.B . and Gouldin , L.P. ( 1999 ). ‘ Cosa nostra: the final chapter? ’, Crime and Justice , vol. 25 , pp. 129 – 89 . Google Scholar Crossref Search ADS WorldCat Karlan , D ., Mobius , M., Rosenblat , T. and Szeidl , A. ( 2009 ). ‘ Trust and social collateral ’, Quarterly Journal of Economics , vol. 124 ( 3 ), pp. 1307 – 61 . Google Scholar Crossref Search ADS WorldCat Kinnan , C . and Townsend , R. ( 2012 ). ‘ Kinship and financial networks, formal financial access, and risk reduction ’, American Economic Review , vol. 102 ( 3 ), pp. 289 – 93 . Google Scholar Crossref Search ADS WorldCat Krebs , V.E . ( 2002 ). ‘ Mapping networks of terrorist cells ’, Connections , vol. 24 ( 3 ), pp. 43 – 52 . OpenURL Placeholder Text WorldCat Kremer , M . and Miguel , E. ( 2007 ). ‘ The illusion of sustainability ’, Quarterly Journal of Economics , vol. 122 ( 3 ), pp. 1007 – 65 . Google Scholar Crossref Search ADS WorldCat Kumar , V . and Skaperdas , S. ( 2009 ). ‘Organized Crime’, in (N. Garoupa, ed.) Criminal Law and Economics , pp. 257 – 70 , Cheltenham : Edward Elgar . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Lasker , G.W . ( 1977 ). ‘ A coefficient of relationship by isonymy: a method for estimating the genetic relationship between populations ’, Human Biology , vol. 49 , pp. 489 – 93 . Google Scholar PubMed OpenURL Placeholder Text WorldCat Lee , S.H ., Kim , P.J. and Jeong , H. ( 2006 ). ‘ Statistical properties of sampled networks ’, Physical Review E , vol. 73 ( 1 ), pp. 016102 . Google Scholar Crossref Search ADS WorldCat Lupo , S . ( 2009 ). Quando la Mafia Trov‘o l’America , Turin : Einaudi . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Maas , P . ( 1968 ). The Valachi Papers , New York : Putnam . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC MAF . ( 2007 ). MAFIA: The Government’s Secret File on Organized Crime. By the US Treasury Department, Bureau of Narcotics , New York, NY : HarperCollins Publishers . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Margo , R.A . ( 1996 ). ‘ The rental price of housing in New York City, 1830–1860 ’, Journal of Economic History , vol. 56 ( 3 ), pp. 605 – 25 . Google Scholar Crossref Search ADS WorldCat Mastrobuoni , G . and Patacchini , E. ( 2012 ). ‘ Organized crime networks: an application of network analysis techniques to the American mafia ’, Review of Network Economics , vol. 3 ( 3 ), pp. 1 – 43 . OpenURL Placeholder Text WorldCat McGloin , J.M . ( 2005 ). ‘ Policy and intervention considerations of a network analysis of street gangs ’, Criminology & Public Policy , vol. 4 ( 3 ), pp. 607 – 35 . Google Scholar Crossref Search ADS WorldCat McWilliams , J.C . ( 1990 ). The Protectors: Harry J. Anslinger and the Federal Bureau of Narcotics, 1930–1962 , Newark, DE : University of Delaware Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Montgomery , J.D . ( 1991 ). ‘ Social networks and labor‐market outcomes: toward an economic analysis ’, American Economic Review , vol. 81 ( 5 ), pp. 1408 – 18 . OpenURL Placeholder Text WorldCat Morselli , C . ( 2003 ). ‘ Career opportunities and network‐based privileges in the Cosa Nostra ’, Crime, Law and Social Change , vol. 39 ( 4 ), pp. 383 – 418 . Google Scholar Crossref Search ADS WorldCat Munshi , K . ( 2003 ). ‘ Networks in the modern economy: Mexican migrants in the US labor market ’, Quarterly Journal of Economics , vol. 118 ( 2 ), pp. 549 – 99 . Google Scholar Crossref Search ADS WorldCat Natarajan , M . ( 2000 ) ‘Understanding the structure of a drug trafficking organisation: a conversational analysis’, in ( Natarajan M. and M. Hough, eds.), Illegal drug markets: from research to prevention policy , pp. 273 – 98 , Monsey, NY : Criminal Justice Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Natarajan , M . ( 2006 ). ‘ Understanding the structure of a large heroin distribution network: a quantitative analysis of qualitative data ’, Journal of Quantitative Criminology , vol. 22 ( 2 ), pp. 171 – 92 . Google Scholar Crossref Search ADS WorldCat Patacchini , E . and Zenou , Y. ( 2008 ). ‘ The strength of weak ties in crime ’, European Economic Review , vol. 52 ( 2 ), pp. 209 – 36 . Google Scholar Crossref Search ADS WorldCat Pinotti , P . ( 2015 ). ‘ The economic costs of organized crime: evidence from southern italy ’, Economic Journal . OpenURL Placeholder Text WorldCat Putnam , R.D ., Leonardi , R. and Nanetti , R.Y. ( 1994 ). Making Democracy Work: Civic Traditions in Modern Italy , Princeton, NJ : Princeton University Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Reuter , P . ( 1983 ). Disorganized Crime: The Economics of the Visible Hand , Cambridge, MA : MIT press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Reuter , P . ( 1994 ). Handbook of Organized Crime in the United States , Westport, CT : Greenwood Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Rothenberg , R.B . ( 1995 ). ‘ Commentary: sampling in social networks ’, Connections , vol. 18 ( 1 ), pp. 104 – 10 . OpenURL Placeholder Text WorldCat Sarnecki , J . ( 1990 ). ‘ Delinquent networks in Sweden ’, Journal of Quantitative Criminology , vol. 6 ( 1 ), pp. 31 – 50 . Google Scholar Crossref Search ADS WorldCat Sarnecki , J . ( 2001 ). Delinquent Networks: Youth Co‐offending in Stockholm , Cambridge : Cambridge University Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Sirakaya , S . ( 2006 ). ‘ Recidivism and social interactions ’, Journal of the American Statistical Association , vol. 101 ( 475 ), pp. 863 – 77 . Google Scholar Crossref Search ADS WorldCat Snijders , T.A . ( 1992 ). ‘ Estimation on the basis of snowball samples: how to weight? ’, Bulletin de méthodologie sociologique , vol. 36 ( 1 ), pp. 59 – 70 . Google Scholar Crossref Search ADS WorldCat Sparrow , M.K . ( 1991 ). ‘ The application of network analysis to criminal intelligence: an assessment of the prospects ’, Social Networks , vol. 13 ( 3 ), pp. 251 – 74 . Google Scholar Crossref Search ADS WorldCat Vega‐Redondo , F . ( 2007 ). Complex Social Networks , Econometric society monograph series, Cambridge : Cambridge University Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Villarreal , C . ( 2013 ). ‘ Where the other half lives: evidence on the origin and persistence of poor neighborhoods from New York city 1830–2012 ’, mimeo, University of Illinois at Chicago. Wasserman , S . and Faust , K. ( 1994 ). Social Network Analysis. Methods and Applications , Cambridge : Cambridge University Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Williams , P . ( 2001 ). ‘ Transnational criminal networks ’, Networks and Netwars: the Future of Terror, Crime, and Militancy, in pp. 61 – 97 , Rand Corporation, available at http://www.rand.org/content/dam/rand/pubs/monograph_reports/MR1382/MR1382.ch3.pdf (last accessed 8 February 2015). Wolf , G . and DiMona , J. ( 1974 ). Frank Costello: Prime Minister of the Underworld , New York: Morrow. Zei , G ., Barbujani , G., Lisa , A., Fiorani , O., Menozzi , P., Siri , E. and Cavalli‐Sforza , L.L. ( 1993 ). ‘ Barriers to gene flow estimated by surname distribution in Italy ’, Annals of Human Genetics , vol. 57 ( 2 ), pp. 123 – 40 . Google Scholar Crossref Search ADS PubMed WorldCat Author notes " I thank the editor Steven Machin, as well as an anonymous referee, Nadia Campaniello, Edoardo Gallo, Giacomo de Giorgi, Gaia Narciso, Michele Pellizzari, Santiago Oliveros, David Reinstein, Rodrigo Soares, Fabian Waldinger, and seminar participants at Trinity College Dublin, the University of Warwick, the University of Essex and the Brucchi Luchino conference for their comments. Martino Bernardi and Dominic Smith have provided excellent research assistance. © 2015 Royal Economic Society
TI - The Value of Connections: Evidence from the Italian‐American Mafia
JF - The Economic Journal
DO - 10.1111/ecoj.12234
DA - 2015-08-01
UR - https://www.deepdyve.com/lp/oxford-university-press/the-value-of-connections-evidence-from-the-italian-american-mafia-XwtRhbI7in
SP - F256
VL - 125
IS - 586
DP - DeepDyve
ER -