banking crises and financial integration: insights from networks science
TRANSCRIPT
Accepted Manuscript
Title: Banking Crises and Financial Integration, InsightsfromNetworks Science
Author: Julian Caballero
PII: S1042-4431(14)00130-9DOI: http://dx.doi.org/doi:10.1016/j.intfin.2014.11.005Reference: INTFIN 748
To appear in: Int. Fin. Markets, Inst. and Money
Received date: 23-9-2014Accepted date: 2-11-2014
Please cite this article as: Julian Caballero, Banking Crises and FinancialIntegration, Insights fromNetworks Science, Journal of International FinancialMarkets, Institutions & Money (2014), http://dx.doi.org/10.1016/j.intfin.2014.11.005
This is a PDF file of an unedited manuscript that has been accepted for publication.As a service to our customers we are providing this early version of the manuscript.The manuscript will undergo copyediting, typesetting, and review of the resulting proofbefore it is published in its final form. Please note that during the production processerrors may be discovered which could affect the content, and all legal disclaimers thatapply to the journal pertain.
Page 1 of 49
Accep
ted
Man
uscr
ipt
Banking Crises and Financial IntegrationInsights from Networks Science
Julian Caballero∗
Abstract
This paper exploreswhether the level of de factofinancial integration of banks in a country increasesthe incidence of systemic banking crises. The paper computes a measure of financial integrationbased on network statistics of banks participating in the global market of inter-bank syndicatedloans. The network statistics used are indegree, outdegree, betweenness, clustering coefficients,authority, and hub centrality. The paper fits a count data model in the cross-section for the pe-riod 1980-2007, and finds that the level of financial integration of the average bank in a countryis a robust determinant of the incidence of banking crises. While borrowing (weighted indegree)is positively associated with a higher incidence of crises, betweenness is associated with a lowerincidence. That is, the more important is the average bank of a country to the global bank network,as captured by betweenness, the smaller the number of crises the country experiences.
Keywords: Banking crises; Financial crises; Capital flows; International Banking; FinancialNetworksJEL classification: E44, F34, F36, G01, G21
∗E-mail: [email protected]. Tel: +1(202)623.3556. 1300 New York Ave. NW, Washington DC, 20577. I would like tothank the Federal Reserve Bank of San Francisco for providing a stimulating environment while I was visiting as a PhDintern in the summer of 2009 and prepared the first draft of this paper. I would especially like to thank Galina Hale, notleast for sharing her data. Michael Hutchison, and Andrew Powell made useful comments. All errors, of course, aremine. The findings, interpretations, and conclusions expressed in this article are entirely those of the author. They donot necessarily represent the views of the Inter-American Development Bank, its Executive Directors, or the countries itrepresents.
Draft. Comments welcome Current version as of September 6, 2014
1) Title Page (WITH Author Details)
Page 2 of 49
Accep
ted
Man
uscr
ipt
1. Introduction
Following the Global Financial Crisis of 2007/2008, there has been a revival of the literature on
financial crises and the factors exacerbating risks in the financial sector. The discussion emphasizes
the role of international financial integration as a factor exacerbating the vulnerability of the bank-
ing system. One strand of the literature focuses on the association between financial crises and de
facto financial integration, proxied by the stock of foreign liabilities, and finds non-conclusive re-
sults.1 On the other hand, a new breed of papers is increasingly successful in linking rapid inflow
growth with an increased probability of crises. This literature also finds that the type of inflows
does matter, debt inflows being particularly problematic.2
Despite the prolific of the literature, the issue of the association of criseswith increased financial
globalization through the banking system, or the banking flows themselves, has been less explored.
This paper aims to shed some light on this issue, employing a novel approach and modeling the
financial globalization of banks using tools from networks science.
An advantage of a network approach is that it can capture different dimensions of how con-
nected each node (bank or country) is to a network. It allows us to think not only in terms of the
size of the flows going into a node (borrowing or inflows), but also on how connected or important
is a node to the network (e.g., this can be captured by network statistics such as betweenness).
The paper studies the association of de facto financial integration of banks on the incidence
of systemic banking crises using network statistics for the average bank in a country to proxy for
the country’s de facto financial integration. These network statistics are computed from lending and
borrowing flows among banks in the inter-bank syndicated loans market. I then use these network
statistics averaged out at the country level to perform non-parametric and regression analyses of
the relationship between the de facto financial integration of the average bank and the incidence of
banking crises in 116 countries in the cross-section of the period 1980-2007.
The different degrees of connectedness or the position relative to the networkmay be important
for understanding howdifferent shocks affect each node in a network. Intuitively, we can think that
1For example, Bonfiglioli (2008) reports a positive link in developed countries between banking crises and the ag-gregate stock of foreign liabilities, but she finds no association in developing countries. Joyce (2010) and Ahrend andGoujard (2012) find a robust association between the likelihood of banking crises and the stock of foreign debt liabilitiesin emerging economies—with the latter study also reporting a greater probability of crises the larger the share of debtin foreign liabilities. However, Gourinchas and Obstfeld (2012) fail to find any association between the share of debtin total external liabilities and the probability of banking crises in emerging markets—although they do find a robustassociation in high-income countries.
2See e.g., Reinhart and Reinhart (2009), Furceri, Guichard and Rusticelli (2012), and Caballero (2012).
1
*2) Blinded Manuscript (WITHOUT Author Details)
Page 3 of 49
Accep
ted
Man
uscr
ipt
amore connected nodewould bemore affected by a systemic shock. However, as shown byAlbert,
Jeong and Barabási (2000) and, more recently emphasized by Haldane (2009), the structure of a
network can make it simultaneously fragile and robust. For example, being well connected to the
global banking networkmay allowa country easier access to the international capitalmarketswhen
it needs it themost and hence enable that country towithstand different shocks that otherwisemay
trigger a financial crisis.3
Furthermore, there is growing evidence showing that a higher level of connectedness may be
associated with an increased ability to dissipate economic shocks. For example, Kali and Reyes
(2010) find that countries that were well integrated into the global trade network were able to
cushion the impact of financial shocks, such as the Mexican and Asian crises, while Caballero,
Candelaria and Hale (2009) show that countries where banks were more connected to the global
network of syndicated loans prior to the 2008/2009 crisis were less affected by it. Chinazzi, Fagiolo,
Reyes and Schiavo (2013) found similar results for connectedness measured in a global network of
security holdings.
The paper uses the network statistics betweenness, clustering coefficient, hub centrality, authority,
indegree and outdegree to capture the connectedness of the average bank of a country to the global
bank network. Betweenness and clustering coefficient are measures based on the link structure of
the network and capture the importance of a bank as intermediary in the network (intermediary
in the sense of being a potential intermediary of bank relationships). Indegree and outdegree
are measures based on the number of incoming and outgoing links, measuring the number of
borrowing and lending relationships of a bank. Hub centrality and authority are hybrid measures,
based on both the link structure of the network and the number of links of the nodes. I use both
unweighted and weighted measures (weighting the bank-level statistics by the size of borrowing
and lending of each bank). Throughout the text, I refer to these network statistics as the de facto
financial integration of the average bank of a country.
The core of the analysis is based on a regression model for the number of systemic banking
crises in the period 1980-2007. The paper fits a count datamodel using a Generalized LinearModel
3One argument from relational banking is that bank lending, as well as much of economic activity, crucially dependson available information, trust, and relationships. One can expect that the higher the connectedness of a country in theglobal banking system, the stronger the relationships this country has, and, hence, the easier its access to internationalcapital markets. One can also think in terms of the literature on financial crises (e.g., Chang and Velasco, 2001) andsudden stops (e.g. Calvo, 1998), emphasizing the inability to obtain short-term debt or that being starved of financialinflows can trigger a banking crisis or a sudden stop event that ends in a crisis.
2
Page 4 of 49
Accep
ted
Man
uscr
ipt
technique, which is a novel approach to study financial crises.4
The main finding is that the level of financial integration of the average bank in a country is
a robust determinant of the incidence of banking crises. I find that increased de facto integration
of banks as measured by total borrowing (sum of weighted indegree for all banks in a country) is
positively associated with the incidence of banking crises. Furthermore, using the proxy for de jure
financial integration of Chinn and Ito (2008), I find that a higher level of de jure integration is also
associated with a higher incidence of crises.
The results also indicate that other factors are at work, and potentially can have a much bigger
role as determinants of banking crises. In particular, prudential banking regulation (supervision)
seems to play a crucial role in reducing the incidence of crises.
The results also indicate that the level of integration of banks into international markets, as
measured by betweenness of the average bank, is negatively associated with the incidence of bank-
ing crises. That is, the more important the average bank of a country is to the global bank network,
the smaller the number of crises the country experiences, even after controlling for borrowing, the
degree of de jure capital account openness, and the quality of banking supervision.
1.1. Related Literature
Network analysis is popular in many social and natural sciences, and it is now gaining popu-
larity in economics and finance; particularly in applications to banking. The banking system lends
itself to being represented by a network in which each bank is a node (or vertex) and the links (or
edges) between banks are the inter-bank borrowing and lending flows.
The bulk of the recent literature on financial networks focuses either on (i) understanding the
phenomena of contagion and systemic risk from a theoretical perspective,5 or on (ii) characterizing
the network structure of banks in a country.6 The focus of the empirical literature on domestic
4After an extensive search, I was not able to find other papers that fit a count data model for the number of financialcrises in a country. The most similar paper I found is Eichengreen (2002). However, this paper fits a count data modelon the number of crises in the world during a year, not for the number of crises in a country in a cross-section of time.
5For a review of the theoretical literature on financial networks see Allen and Babus (2009). Theoretical models ofcontagion in financial markets are developed, e.g., by Zawadowski (2013), Battiston, Delli Gatti, Gallegati, Greenwaldand Stiglitz (2012), and Georg (2013). Related models of cascading effects are studied by Acemoglu, Carvalho, Ozdaglarand Tahbaz-Salehi (2012), Acemoglu, Ozdaglar and Tahbaz-Salehi (2010), and Blume, Easley, Kleinberg, Kleinberg andTardos (2011). Schweitzer, Fagiolo, Sornette, Vega-Redondo and White (2009) survey the more general literature oneconomic networks.
6Boss, Elsinger, Summer and Thurner (2004) and Inaoka, Ninomiya, Taniguchi, Shimizu and Takayasu (2004) areearly studies on the network topology of inter-bank markets (for Austria and Japan, respectively). Other studies havelooked at networks of inter-bank payments or inter-bankmarkets in theUSA, theUK, Brazil, Italy, Denmark, TheNether-lands, Colombia, and Mexico. An example of a recent study of this kind is the paper on the network topology of theinter-bank money market in the USA (federal funds market) by Bech and Atalay (2010).
3
Page 5 of 49
Accep
ted
Man
uscr
ipt
markets is in part due to the lack of international bank-level data on lending and borrowing, which
only until recently has been an obstacle in this line of research. In fact, three of only a handful of
existing papers using a network approach to characterize a global bank network rely on country-
level data from the banking statistics of the Bank of International Settlements (Hattori and Suda,
2007; Minoiu and Reyes, 2013; von Peter, 2007).7
Hale (2012), Hale, Candelaria, Caballero and Borisov (2011), and Hale, Candelaria, Caballero
and Borisov (2013) take a different approach, using bank-level micro data to construct a global
bank network based on the global market of inter-bank syndicated loans. Hale (2012) uses data
from eight thousand banks that participate in the market and studies the evolution of the result-
ing network of inter-bank borrowing and lending. She compares some basic characteristics of the
network and the formation of new bank relationships after recessions and financial crises, finding
that crises have long-lasting negative effects on the formation of new bank relationships. Network
statistics at the country level based on the inter-bank syndicated loans network have been used by
Hale et al. (2011) to show that the location of a country in the network is associated with capital
flows into and out of a country. Similarly, Hale et al. (2013) compute a network distance between
bank pairs in this network and show that their measure of new connections between banks in a
given country pair is associated with increased bilateral trade flows.
The remaining of the paper is organized as follows. The next section introduces the network
statistics used as proxies for de facto financial integration of banks, focusing on the definition and
intuition of how these statistics reflect different aspects of the international integration of banks.
Section 3 describes the data used in the paper. Section 4 takes a first look at the basic characteristics
of the network of inter-bank syndicated loans, plotting its degree distribution and visualizing the
web of lending and borrowing connections among banks. Sections 5 and 6 contain the core of the
analysis of the association between the network statistics capturing the de facto integration of banks
and the incidence of banking crises. Section 7 closes with some concluding remarks.
7Other country-level studies using a network approach have looked at economic flows other than banking. Forexample, Kubelec and Sá (2012) construct a dataset on cross-border assets and liabilities in four asset classes: FDI,equity, debt, and foreign exchange reserves. Chinazzi et al. (2013) use holdings portfolio and equity securities. Bilateraltrade flows have been used by Kali and Reyes (2007) and Fagiolo, Reyes and Schiavo (2008) to construct network statisticsto measure international economic integration.
4
Page 6 of 49
Accep
ted
Man
uscr
ipt
2. Network Approach to International Financial Integration
Before discussing the network statistics used to proxy for the de facto financial integration of
banks, we need to introduce some definitions and notation. In this network each bank is a node
(or vertex) and is indexed by i = 1, . . . , n. The link or edge between banks i and j is denoted by aij .
We are characterizing a directed network, in which the direction of the link matters (borrowing or
lending), thus, aij 6= aji. Not all pairs of banks (nodes) are connected, and hence many of these
links do not exist. In a network, however, two disjointed nodes k and i are indirectly connected
through the direct connection between i and j. The link between banks i and j can also be charac-
terized by the size or intensity of the connection, and we denote this weighted link as wij (in this
directed network wij 6= wji).
The length of the path between nodes i and j is the number of links that must be covered
in the trajectory from i to j. A geodesic path is a path between two nodes that has the shortest
possible length (for two given nodes there may exist more than one geodesic path). The length
of the geodesic path from bank i to bank j is denoted as gij . The number of geodesic paths from
nodes i to j is denoted as pij . The number of geodesic paths that go from bank i to bank j passing
through bank k is denoted as pikj .
Network statistics that capture the connectedness of a given node are referred in the networks
science literature as centrality measures. The literature has studied different centrality measures,
with different proxies based on different aspects of the connections present in the network. The
most basic measures are called degree measures and are computed based on the number of links
going into or departing from a given node (e.g., indegree, outdegree, degree). There are measures
of centrality based on the length of geodesics between nodes (e.g., closeness). There are measures
that take into account the link structure of the whole network to identify central nodes (e.g., eigen-
vector centrality, PageRank, hub centrality, authority centrality). Alternatively, there are measures
that gauge how central a node is to the network by taking into account how important is its role as
intermediary in the network (e.g., betweenness, clustering coefficients).
To describe the de facto financial integration of a bank into the network, I will use the network
statistics indegree, outdegree, degree, betweenness, clustering coefficients, hub centrality, and au-
thority. These measures capture different dimensions of how interconnected a node is into the
network, allowing us to fully exploit the richness of a network approach.8
8Some of the existing network statistics are computed based solely on the information from the binary links between
5
Page 7 of 49
Accep
ted
Man
uscr
ipt
Degree measures (indegree, outdegree, degree) capture how connected is each node to the
network by simply counting the number of links going into or out of a node. This gives an idea
of the number of lending or borrowing relationships that a bank has in the network. We also can
construct the equivalent measure taking into account the intensity of the links, which in the global
bank network is just the size of the borrowing and lending of a node. Thus, the network statistic
indegree is simply the number of incoming links or borrowing relationships (kini =∑
j aji), while
outdegree is the number of outgoing links or lending relationships (kouti =∑
j aij). Degree is
just the sum of these two quantities. Each of these measures can be computed after weighting
by the size of the links. In the case of indegree, this would be the sum of all borrowing done
by a bank. I will use both unweighted and weighted indegree measures, and I will refer to the
former as indegree and to the latter as borrowing (computed as borrowingi =∑
j wji). I will also
use the equivalent measure for outdegree, and the corresponding weighted statistic (referred to as
lending).
Betweenness captures how important and central a node is to the network. It is not affected by
the size of the network (the number of nodes or of geodesic paths). It is constructed as the ratio of
the paths in which node i is an intermediary between nodes k and l to the total number of paths
between k and l. Formally, it is computed as betweennessi =∑
l
∑k(pkilpkl
). Note that this definition
takes into account the directed nature of the network.9
As betweenness, the clustering coefficient captures how central as an intermediary to the net-
work a node is. A difference is that betweenness takes into account how a given node plays as inter-
mediary between any two nodes in the whole network, while the clustering coefficient is based on
the role of intermediary in local triplets (i.e., two immediate neighbors of node i that are connected
to each other, so that the three nodes form a closed triangle). As in the case of betweenness, the
usual algorithms to compute clustering coefficients disregard the weights of the links between the
nodes and are based solely on the link structure of the network. Newman (2010, p. 202) remarks
nodes, disregarding the information from the direction and the intensity of the connection (this is because most of theinitial literature focused on unweighted and undirected networks). However, I use measures that take into account thedirected and weighted nature of the global bank network. In addition, I do not use measures based on the length of thegeodesic paths, such as closeness, because a largemeasure for a given node can be the result of a relatively large network(geodesic paths of large length for the average node) or that the node is relatively isolated (relatively large geodesic pathsfor a particular node). I also do not consider measures that have been proven of limited meaning in directed networks,such as eigenvector centrality, PageRank, and Katz centrality. For a detailed discussion of different issues with centralitymeasures see Newman (2010, Ch. 7).
9The implementation in this paper uses the algorithm of Brandes (2001) and normalizes betweenness by n2, so thatit is strictly in the [0,1] interval. The computation is done in the software developed by Batagelj and Mrvar (1998).
6
Page 8 of 49
Accep
ted
Man
uscr
ipt
that betweenness and the clustering coefficient are quite correlated in practice. Nonetheless, I will
use the clustering coefficient to complement the analysis based on betweenness.10
An alternative measure that considers the intensity of the connection is the weighted cluster-
ing coefficient of Barrat, Barthelemy, Pastor-Satorras and Vespignani (2004). This statistic weights
the node-level statistic by the intensity of the flows in the triplet. Thus, it considers not just the
number of closed triplets in the neighborhood of a node, but also their total relative weight with
respect to the intensity of links of the node. I use this measure to complement the analysis based
on betweenness and the clustering coefficient, which are unweighted statistics.11
I also use the measures of authority and hub centrality proposed by Kleinberg (1999). These
statistics are computed using information on the link structure of the whole network. These statis-
tics are based on the premise of the existence of a mutually reinforcing relationship between two
different types of nodes in a network: (i) authorities, which are nodes pointed to by important hubs,
and hence tend to exhibit a large indegree; and (ii) hubs, which point to important authorities, and
as a result tend to exhibit a large outdegree.12
One shortcoming of betweenness, clustering coefficient, hub centrality, and authority is that
they do not take into account the intensity of the flows. Thus, a node that serves as hub for banking
flows in some region can have a very large measure, even though it is not a world heavyweight in
terms of the size of the flows. To allay this issue, I additionally compute these statistics using as
weight the size of the share of the total flows in an out of bank i on the total global flows. Thus,
weighted statistics are computed aftermultiplying each network statistic byωi =∑i wij+
∑j wji∑
j(∑i wij+
∑j wji)
.
The idea is to use these network statistics to proxy for the de facto financial integration of the
banking system of a given country. The network statistics are computed at the bank level and for
all banks that have participated as lead arrangers in the inter-bank syndicated loans market. To
obtain a country level measure, averages over all lead banks in a country are taken. Thus, in effect
I am proposing that the de facto financial integration of a country be proxied by the network statistic
10The clustering coefficient is computed as ci = ∆i/∆ii, where ∆i is the number of closed triangles connected to node
i and ∆ii is the number of triples centered on node i (which is just the number of pairs of neighbors of i). Thus, the
clustering coefficient of node i captures the probability that i’s neighbors are also connected. Note that the role of ias local intermediary measured by the clustering coefficient decreases as the coefficient increases. The computationsof clustering coefficients are implemented in the software developed by Batagelj and Mrvar (1998), which is based ontriplets of the undirected network. Because some nodes have degree of less than 2, for some nodes the denominator iszero and ci is not defined. In those cases, I follow the convention and assign ci = 0.
11The weighted clustering coefficient is implemented using the software of Bastian, Heymann and Jacomy (2009).12The computations of the HITS algorithm of Kleinberg (1999) are implemented using Batagelj and Mrvar (1998).
7
Page 9 of 49
Accep
ted
Man
uscr
ipt
of the average lead bank in a country.13
Because a priori it is unclearwhether unweighted statistics are preferable overweighted ones, or
vice versa, I will use both. An advantage of weighting the bank-level statistics by the bank’s share
of total global flows (ωi) is that it allows to take into account the concentration of banking activity
in the syndicated loans market. This weighting approach assigns higher levels of betweenness and
other centrality measures to banks with larger shares of the market. Note that the country-level
averages of the weighted statistics carry this information into the country-level statistics.
Throughout the paper I use country-level statistics normalized so that the maximum is 1. Fi-
nally, in order to control for all borrowing and lending by banks in a country, borrowing and lend-
ing will be used as the sum of the flows of all banks in the country.
3. Data
The analysis in this paper centers in the cross-section of the period 1980-2007. I compiled a
database on financial crises with a set of macroeconomic and institutional variables, and, as de-
scribed before, I usemeasures of de facto financial integration based on network statistics computed
from inter-bank syndicated loans.
3.1. Loans Data
I use loans data from Hale (2012). These data comprises all syndicated loans between private
banks reported by Dealogic Loan Analytics (this database is also known as Loanware).14 Given
the composition of the syndicate, it is possible to create bank-to-bank links (lending or borrowing
relationships) between participating banks. To construct the network, each bank-to-bank link in a
deal is replicated as many times as there are lenders in the syndicate, and equal loan amounts are
assigned to each participating lender in the deal. The links between nodes in the network are then
13The paper uses country-level simple averages for both unweighted and weighted network statistics. Country-levelstatistics can also be computed as weighted averages. For example, it would be ideal to compute country-level averagesby weighting bank-level statistics by bank’s share of the country’s financial assets. However, this is not possible with thedata at hand.
14A syndicated loan is one jointly funded by a group of banks. The syndicated loan market follows an “originate-to-distribute model," whereby the originating bank, dubbed lead arranger or lead manager, establishes a lending re-lationship with the borrower and negotiates the terms of the loan agreement. The lead arranger, for reasons such asexposure and risk management, looks for participant banks which fund a share of the loan; in return, the participantbanks receive fees. The lead arrangers choose the participant lenders and administer the loan/syndicate, whereas par-ticipant lenders essentially just fund the loan. Large loans are typically structured in multiple facilities. All facilitiesare covered by the same loan agreement; however, they may have different maturity or draw-down terms. Syndicatedloans are a significant source of international financing, accounting for a third of all international financing, includingbond, commercial paper and equity. For an explanation on syndicated loans and their rationale and determinants see,for example, Pichler and Wilhelm (2002) and Gatev and Strahan (2009).
8
Page 10 of 49
Accep
ted
Man
uscr
ipt
computed aggregating over all the deals of a given pair of banks.
The original dataset has the universe of inter-bank loans reported in the period from January
1980 to December 2009, with a total of 15,324 deals and a total of 8,525 participating banks from
124 countries. I use the data to build a network of lending and borrowing relationships in the
deals between banks that have participated as lead arrangers. The total number of banks that
participated as lead arrangers is 4,806, and these banks were involved in 3,657 deals that include
at least one lead arranger. The network statistics described in Section 2 were computed based on
these data. Table A-1 in the Appendix presents the sample of countries for which network statistics
were computed. The nationality of banks is based on a locational basis.15
The focus is on banks that have participated as lead arrangers because the values of the network
statistics used in the analysis are obtained as an average of the statistics of all participating banks
from a country. Even though most banks participate as lead arrangers and usually alternate roles
in subsequent loans (Cai, 2010), there are banks that never participate as lead arrangers. Hence, if
all banks are included in the computation, the resulting statistics would be skewed downwards by
banks that rarely participate in the market. Thus, the paper uses network statistics computed from
the network of banks that have participated in at least one deal as lead arrangers.16
Besides allowing to construct a bank-level network of bank relationships, one advantage of us-
ing the syndicated loans market is that it is a good alternative for studying the international bank
market as a whole. As shown by Gadanecz and von Kleist (2002, p. 71), the lending flows of syn-
dicated private loans are a good proxy for the flows implied by the consolidated banking statistics
of the Bank of International Settlements (BIS), which are the authoritative source of bilateral bank
flows. According to these authors the proportional changes in both data sets seem to be closely
linked, with the change factor significantly different from zero and virtually identical to one. Hale
(2012) further documents that the total annual loan originations in the inter-bank market for syn-
dicated loans on average have amounted to 20 percent of inter-bank claims reported to the BIS
between 1980 and 2009—although during the late 1990s this ratio was as high as 34 percent.
The paper focuses on the cross-section of the period 1980-2007, that is, before the unwinding of
15In other words, the nationality of a bank is assigned depending on where the operation of the bank takes place. Inthis case, the country where the bank participating in the syndicate is located. For example, for Citibank Honk Kong Iassign the nationality of Hong Kong, not of the USA, which is the nationality of the parent. This approach is similar tothe one used in the balance of payments statistics.
16Focusing on lead banks is usual in the literature using data on syndicated loans, as lead banks control the lion’sshare of the market. For example, Cai, Saunders and Steffen (2012) document that in the United States the top 100 leadbanks controlled an aggregated share of between 99.7 percent and 100 percent of the market during 1988-2010.
9
Page 11 of 49
Accep
ted
Man
uscr
ipt
the subprime crisis in the USA and the subsequent global financial crisis of 2008/2009. Specifically,
the network statistics used in this paper as proxies of de facto financial integration were computed
using data for loans originated in the period spanning January 1980 to June 2007. The analysis
focuses on this period because the global financial crisis is in a league of its own, having a deep
impact on banking flows all around the globe (Hale, 2012; Minoiu and Reyes, 2013) and unique
characteristics—especially regarding the way it spread out to many developed countries.17
3.2. Banking Crises Data
To identify banking crises, I use the database of Laeven and Valencia (2013). In this database
a banking crisis is defined as a systemic banking crisis when two conditions are met: (i) significant
signs of financial distress in the banking system (as indicated by bank runs, losses in the banking
system, and/or bank liquidations); and (ii) significant banking policy intervention measures were
undertaken in response to losses in the banking system.
The starting year for a systemic banking crisis is identified by the two conditions mentioned
above, along with the fulfillment of at least three out of the following six policy interventions
Laeven and Valencia (2013): extensive liquidity support; large bank restructuring costs; significant
bank nationalizations; significant asset purchases; significant guarantees put in place; or deposit
freezes and bank holidays.18 Because the quantitative thresholds used in this definition of systemic
banking crises are ad hoc, events that almost meet these thresholds are classified as “borderline.”
The definition does not include isolated banks in distress. With the methodology just described,
Laeven and Valencia (2013) identify 147 crises in 115 countries for the period 1973–2009. Table A-1
in the Appendix makes explicit the sample of crises used in the paper.
3.3. Other Data
The set of macroeconomic variables used as controls in the regression analysis was obtained
from the World Development Indicators database of the World Bank, including GDP per capita,
openness in trade, bank credit to private sector, current account balance, and inflation. To account
for de jure capital account openness I use the database by Chinn and Ito (2008). The indexes of
17The dating of the start of the subprime crisis is imprecise. June 30, 2007 is chosen as a cut-off for the loans databecause it is just before the first clear signals of the start of the liquidity crunch in the United States. The events thatprecipitated the crash include the liquidation of two hedge funds by Bear Stearns in July 31 and the decision of BNPParibas of suspending three hedge funds focused on US mortgages in August 9.
18When a country has faced financial distress but less than three of these measures have been used, the event isnevertheless classified as a crisis if either the country’s banking system exhibits significant losses due to nonperformingloans or bank closures, or if fiscal restructuring costs of the banking sector exceed five percent of GDP.
10
Page 12 of 49
Accep
ted
Man
uscr
ipt
banking supervision is taken from Abiad, Detragiache and Tressel (2010). To proxy for quality of
political institutions, I use the index of political risk from ICRG. Details on each variable used in
the paper can be found in Table A-2 in the Appendix.
4. A First Look at the Global Network of Inter-bank Syndicated Loans
As described in Subsection 3.1, I use data from Hale (2012) to build a network of lending and
borrowing relationships between banks that have participated as lead arrangers in the global mar-
ket for inter-bank syndicated loans during 1980-2007. The resulting network comprisesN = 4, 806
nodes and E = 16, 063 directed links. This section takes a brief look at some properties of the
network. However, a full-blown network analysis is out of the scope of this paper and is left for
future research.19
4.1. Basic Network Characteristics
Table 1 presents basic characteristics of the network in the period 1980-2007, which is the base-
line used throughout the paper. The average degree of the network is 3.34.20 The degree measures
exhibit large dispersion and asymmetry: while some nodes exhibit zero indegree or outdegree
measures, maximal indegree is 128 and maximal outdegree is 229. Thus, in this network the most
connected of borrowers has almost half of the relationships than the most connected of lenders.
This indicates an asymmetry in the inter-bank syndicated loans market, with many more banks
that lend to a few institutions but do not borrow, than there are banks that borrow from a few in-
stitutions but do not lend. This asymmetry between lenders and borrowers in an inter-bankmarket
has been also documented by Bech and Atalay (2010) in the USA.
The average shortest path, measured as the average number of edges separating any two nodes
in the network, is ` = 4.45. This is quite small relative to the size of the network and suggests that
the network may exhibit the small-world property. The small-world property is the characteristic
of a network of having most nodes connected by a short path, much shorter than what it would be
19Local markets for syndicated loans have been studied from a network perspective, but, to the best of my knowledge,there is no study on the network topology of the globalmarket for syndicated loans. There are twopapers looking at localmarkets and both are focused on studying some aspects of the network of participating banks, more with an emphasison exploring how the network structure affects the terms of the loans or the structure of the syndicates, rather than afull network analysis. Godlewski, Sanditov and Burger-Helmchen (2012) study the network of lenders in France, whileCai et al. (2012) study the top lead banks in the USA (although, the latter paper does not build a network of borrowersand lenders, but rather constructs a measure of connectivity of banks based on the similarity of their loan portfolios byindustry).
20Because in a network it must be true that∑i,j aij =
∑i,j aji, by definition in a directed network average degree is
equal to both average indegree and average outdegree.
11
Page 13 of 49
Accep
ted
Man
uscr
ipt
implied by its size and random connections between nodes. This characteristic has obvious impli-
cations for the dynamics of processes taking place on the network. For example, when considering
the spread of a disease, the small-world property implies that the disease will spread very fast
throughout the population (for a discussion see Albert and Barabási, 2002 and Newman, 2003).
For comparison, Table 1 also presents the same characteristics of the network for two subsam-
ples: 1995-2007 and 2001-2007. The network in these later subsamples is smaller, but exhibits sim-
ilar characteristics as the full-sample network, with a lot of dispersion in the degree measures, and
with a maximal outdegree almost as twice as large as maximal indegree.
Table 1: Key Network Characteristics
1980-2007 Network 1995-2007 Network 2001-2007 NetworkNodes 4,806 2,861 1,486Edges 16,063 9,849 5,526Total geodesic paths 208,455 62,199 18,690Avg. geodesic path 4.45 3.89 2.57Diameter 13 14 11Density 0.001 0.001 0.003Avg. Degree 3.34 3.43 3.72Min Indegree 0 0 0Max Indegree 128 108 91Min Outdegree 0 0 0Max Outdegree 229 208 174
This table presents key network characteristics of three different samples of the data: the fullsample 1980-2007, and two subsamples for 1995-2007 and 2001-2007. The computation of de-gree measures, number of geodesic paths, diameter, and density is implemented in the soft-ware developed by Bastian et al. (2009). In all cases the network is treated as a weighted directedgraph.
Table 1 also shows two statistics of the whole network: density and diameter.21 Comparing the
density and diameters of the full sample network and the subsamples, it is clear that the network
has become denser and more connected in more recent years, as the process of financial globaliza-
tion has deepened. Hale (2012) studies in detail the time evolution of the number of banks, edges,
density, diameter, and geodesics of the global network of inter-bank syndicated loans, finding that
the network has become more connected over time.
21Density (ρ) is the number of links actually present in the network as a share of the possible number of links. Itdescribes the connectedness of the nodeswithin the network. Formally, ρ =
∑i
∑j(aij+aji)/(n(n−1)); with ρ ∈ [0, 1].
Diameter is the length of the longest geodesic path present in the network. It measures the span of the network.
12
Page 14 of 49
Accep
ted
Man
uscr
ipt
4.2. Degree Distribution
The topology of the network can further be described by its degree distribution. Figure A-1
in the Appendix shows histograms with the distributions of indegree and outdegree. There are
long tails in both measures, with a large number of banks with a very reduced number of borrow-
ing/lending connections and a very small number of banks with a large number of connections.
Histogramswith such long tails are difficult to read, however, and convey little information beyond
the presence of a long tail.
The preferred way of representing the degree distributions in the networks science literature
is by means of the cumulative distribution function. These plots show in a log-log scale the cumu-
lative fraction of nodes with degree greater than or equal to degree k as a function of the degree.
This is useful because the literature has found that many networks have degree distributions that
in their tails behave as Power Laws, with a decay parameter α between -2 and -3 (see surveys by
Albert and Barabási, 2002 and Newman, 2003). In a cumulative distribution plotted in a log-log
scale this translates into a, more or less, straight line with slope α. Networks that follow in their
tail a Power Law are also referred to as scale-free networks (Barabási and Albert, 1999). Impor-
tantly, scale-free networks have been shown to be very resilient to the random removal of nodes,
but highly vulnerable to deliberate removal of high-degree nodes (Albert et al., 2000).
Figure 1 shows the degree distributions for the full sample network. As a reference, the fig-
ure also shows the distributions of equivalent networks following a logarithmic function and two
different Power Laws (α = −2,−3). We see that both indegree and outdegree distributions have
heavy tails, somewhat similar to the ones observed in Power Laws, although not identical.22
Finally, since in the subsequent analysis in this paper we will aggregate the bank-level network
statistics into country-level measures, it is informative to explore if some of the characteristics of
the network are conserved after this aggregation. The second panel in Figure 1 provides this com-
parison, plotting the degree distributions of the actual country-level data used later in the paper. It
is clear from the plots that the indegree and outdegree distributions of the country-level aggregates
is similar to that of the bank-level statistics, with a large number of countries with a low degree
measure, and a small number of countries with a large degree statistic, and with the distributions
of indegree and outdegree scaling in similar ways.
22These plots are shown for illustrative purposes, to describe the network. They are not supposed to provide ultimateproof of the functional form of the degree distributions of the network, which is out of the scope of this paper.
13
Page 15 of 49
Accep
ted
Man
uscr
ipt
Figure 1: Cumulative Distribution Functions of Degree Measures
Panel A. Bank-Level Network. Sample 1980-2007
01
Cum
ulat
ive
Dis
trib
utio
n
Indegree
01
Cum
ulat
ive
Dis
trib
utio
n
Outdegree
01
Cum
ulat
ive
Dis
trib
utio
n
Degree
0.5
1C
umul
ativ
e D
istr
ibut
ion
Power Law a=-2
Power Law a=-3
Logarithmic
Reference Distributions
Panel B. Country-Level Averages. Sample 1980-2007
01
Cum
ulat
ive
Dis
trib
utio
n
Indegree
01
Cum
ulat
ive
Dis
trib
utio
n
Outdegree
01
Cum
ulat
ive
Dis
trib
utio
n
Degree
0.5
1C
umul
ativ
e D
istr
ibut
ion
Power Law a=-2
Power Law a=-3
Logarithmic
Reference Distributions
This figure presents the cumulative distribution functions of indegree, outdegree, and degree measures forthe full sample network (1980-2007). The plots present in a log-log scale the cumulative fraction of nodes withdegree greater than or equal to degree k as a function of the degree. As a reference, the figure also showsthe distributions of equivalent networks following a logarithmic function and two different Power Laws withdecaying parameters α = −2 and α = −3. These reference distributions are build using the same numberof nodes as the original networks. Panel A shows the cumulative distribution functions for the bank-levelnetwork, composed of n = 4, 806 nodes. Panel B shows the corresponding cumulative distribution functionsfor the country-level averages (n = 116).
14
Page 16 of 49
Accep
ted
Man
uscr
ipt
4.3. Network Visualization
A popular way of analyzing a network is through a visualization of the resulting graph from
the web of nodes and edges that compose the network. Figure 2 presents a visualization of the full
sample network. Each node in the graph represents a bank and the links between banks represent
lending and borrowing relationships. Convex edges represent lending from node i to node j. This
network is moderately large (N = 4, 806;E = 16, 063), and hence its structure is difficult to analyze
visually. Nonetheless, the visualization is useful to better understand the data.
The graph uses color keys to identify banks geographically with four distinct geographic re-
gions (the Americas, Western and Eastern Europe, Asia and the Pacific, and Africa and theMiddle
East). The nodes have the color according to their nationality, and the edges are colored according
to the color of the target. For example, a convex blue edge from a green-colored node repents a
loan from a bank located in Europe to a bank located in the Americas.
The graph is somewhat difficult to read due to its size, but several patterns emerge. On the
one hand, there are a myriad of dyads of banks that participate in deals that only include them.
These are banks with little integration into the market, as measured by the number of lending and
borrowing relationships, and are plotted in the periphery of the graph (a further analysis indicates
that a significant portion of these binary links are between banks located in the same country).
On the other hand, the more connected banks, towards the center of the graph, tend to be
located in Europe and the Americas, with Asian banks a little behind. Further analysis of the data
verifies this pattern, with most of the high-degree nodes from the UK, USA, Germany, France, and
Japan. Finally, the more connected banks tend to have lending and borrowing relationships with
many banks, not only domestically, but also internationally. Further analysis of the data indicates
that European banks are the ones with more cross-border relationships.
To better understand the data, Figure 3 presents a visualization of a subsample of the network.
This visualization shows the network composed of the top 100 lenders and borrowers and the links
between them that are worth US$500 million or more (N = 118; E = 212). As in the full sample
graph, the edges are colored according to the target country and the color of the nodes identify the
country where the bank is located—this time with a less coarse classification.
15
Page 17 of 49
Accep
ted
Man
uscr
ipt
Figu
re2:
Visua
lizationof
theNetwork.
Sample19
80-200
7
Amer
icas
Euro
peAs
ia &
Pacif
icAf
rica &
Mid
dle E
ast
A B
Link
from
B to
A
A B
Link
from
A to
B
Thisfig
urepresen
tsavisu
alizationof
thene
twork,
sample19
80-200
7.Ea
chno
dein
thegrap
hrepresen
tsaba
nkthat
have
participated
aslead
arrang
erin
theglob
almarke
tfor
synd
icated
loan
sin
thepe
riod19
80-200
7,an
dthelin
ksbe
tweenba
nksrepresen
tlen
ding
andbo
rrow
ingrelatio
nships
.Th
evisu
alizationpresen
tsthe
linka
gesb
etweenatotalo
fN=
4,8
06ba
nks(no
des);a
totalo
fE=
16,0
63lin
ks(edg
es).Th
istw
o-dimen
sion
alvisu
alizationisba
sedon
theforcedirected
algo
rithm
ofFruc
hterman
andRe
ingo
ld(199
1),w
hich
places
node
ssothat
allthe
edge
sare
ofmoreor
less
equa
llen
gthan
dthereareas
few
crossing
edge
sasp
ossible.
Link
saredirected
,with
conv
exed
gesrepresen
tinglend
ingfrom
node
ito
node
j.Th
esizesof
theno
desan
dof
thelin
ksareun
iform
.Th
ecoloro
fthe
node
siden
tifies
bank
sgeo
grap
hically
with
four
distinct
geog
raph
icregion
s,as
describ
edin
thelege
nd.T
helin
ksarecoloredaccordingto
thecoloro
fthe
target.
16
Page 18 of 49
Accep
ted
Man
uscr
ipt
In the visualization of Figure 3, however, the size of the nodes is proportional to the degree
of the bank (the sum of all lending and borrowing relationships) and the size of the links is pro-
portional to the size of the lending or borrowing deals between the banks. To better represent the
connections between banks, the graph also presents the names of the banks’ parents, followed by
a three letter acronym for the country where the bank is located.23
The graph is useful to visualize the interconnections between banks across the globe and do-
mestically. The graphmakes it evident how global banks increase the financial exposure of a coun-
try to the rest of theworld throughout their cross-border linkageswith other banks. The graph also
shows that the interconnections between banks across borders are not only due to global banks
expanding into new international markets (with cross-border flows likely driven by centralized
decisions on how to allocate capital internally within the organization), but also due to exposures
between global banks headquartered in different countries. The graph makes it clear that many of
these exposures have their origin in the operations of subsidiaries of global banks abroad, but also
from direct cross-border operations originated in the headquarters.
23Tomake explicit the connections between global banks, the graph presents the names of the bank’s parent when thebank is a subsidiary of a global bank (the nationality of the bank, reflected in the three-letter country code, still refersto where it is located). Because the data used in this paper is up to 2007, mergers and acquisitions after the 2008 wherenot considered in the process of assigning parent names. Since banks in the top 100 spots by lending and borrowinginclude several subsidiaries of global banks, the network is constructed after aggregating the links between the resultingdyads after assigning parent names. For example, in the top list of banks we have Lehman Brothers Holdings Inc.and Lehman Brothers North America, both located in the USA. I collapse the information of these two banks into oneconsolidated bank for Lehman Brothers in the USA. This process yielded a network of 118 nodes. For clarity, the graphonly visualizes the 212 edges between banks that are worth US$500 million or more; the total number of links betweenthe 118 consolidated banks is 513. The visualization of the complete graph is somewhat convoluted and is presented inthe Appendix as Figure A-2.
17
Page 19 of 49
Accep
ted
Man
uscr
ipt
Figu
re3:
Visua
lizationof
theNetwork.
Top10
0Bo
rrow
ersan
dLe
nders(link
s>
USD
$500
million).S
ample19
80-200
7
DA
IWA
EU
RO
PE L
TD
NIB
C B
ank-
NLD
Land
esba
nk-G
BR
SNS
Rea
al -N
LD
MB
NA
Cor
p-U
SA
DN
B B
ank-
NO
R
Deu
tsch
e A
usgl
eich
-DEU
SEB
-SW
EJP
Mor
gan-
IRL
HSB
C-N
LD
FCE
Ban
k-G
BR
RB
C-C
AN
US
Ban
corp
-USA
Rab
oban
k-IR
L
Wes
thyp
Fin
-DEU
FHL
Ban
ks-U
SA
Cha
se-U
SA
Blo
ck F
G-U
SA
Deu
tsch
e B
ank-
DEU
Cre
dit S
uiss
e-C
HE
Deu
tsch
e B
ank-
CH
E
Deu
tsch
e B
ank-
GB
R
Sven
ska
Han
dels
-SW
E
VTB
Cap
ital-R
US
Wes
tLB
-DEU
3CIF
-FR
A
UB
S-G
BR
Nat
ionw
ide
BS-
GB
R
Man
uf. H
anov
er-G
BR
Com
pagn
ie B
anca
ire-F
RA
HSB
C-G
BR
Euro
Cle
ar-B
EL
Gol
dman
Sac
hs-D
EU
DZ
Ban
k-D
EU
Mor
gan
Stan
ley-
USA
Com
mer
zban
k -IR
L
Sum
itom
o M
itsui
-GB
R
Sum
itom
o M
itsui
-JPN Citi
-GB
R
Inte
sa S
anpa
olo-
ITA
Bay
ernL
B-D
EU
FIH
-DN
K
Deu
tsch
e B
ank-
LUX
Gol
dman
Sac
hs-G
BR
Ban
k O
ne-U
SA
Bar
clay
s-G
BR
CIT
Gro
up -U
SA
AN
Z-H
KG
Lehm
an B
roth
-GB
R
Cre
dit A
gric
ole
-FR
A
Stan
dard
Cha
rter
ed B
ank-
GB
R
SG-A
US
Ban
k of
Am
eric
a -G
BR
Ced
el B
ank-
LUX
Kid
der P
eabo
dy-U
SA
HVB
-Uni
cred
it-D
EU
Hyp
o Ve
rein
sfin
-DEU
BN
P Pa
ribas
-FR
A
Dep
fa-D
EU
Nat
ixis
-FR
A
Firs
t Uni
on B
ank-
USA
BN
Y M
ello
n-U
SA
Rab
oban
k-A
US
Bra
dfor
d-B
ingl
ey-G
BR
Nor
ther
n R
ock-
GB
R
Scot
land
Inte
r Fin
-GB
R
Wac
hovi
a-U
SA
Isba
nk-T
UR
Mer
rill L
ynch
-DEU
AN
Z-A
US
Ban
k of
Am
eric
a -U
SA
Bea
r Ste
arns
-USA
Mits
ubis
hi U
FJ -J
PN
Ban
k of
Chi
na -C
HN
Rai
ffeis
en-A
UT
Turk
iye
Gar
anti-
TUR
LBB
W-D
EU
Han
a B
ank-
KO
R
JPM
orga
n-U
SA
AB
N A
MR
O-N
LD
Cre
dit S
uiss
e-U
SA
Mer
rill L
ynch
-USA
HSB
C-H
KG
NA
B-A
US
Ban
k of
Am
eric
a -H
KG
Sant
ande
r-G
BR
Deu
tsch
e B
ank-
HK
G
Wes
tpac
-AU
S
ING
-NLD
Nom
ura-
DEU
Cre
dit S
uiss
e-G
BR
CB
A-A
US
JPM
orga
n-G
BR
BA
-Uni
Cre
dit-A
UT
JP M
orga
n-D
EU
Hel
aba
Fina
nce-
DEU
UB
S-D
EU
UB
S-C
HE
Nom
ura-
GB
R
Rab
oban
k-N
LD
Mer
rill L
ynch
-GB
R
Ban
k of
Irel
and-
IRL
Miz
uho-
GB
R
Miz
uho-
JPN
Lehm
an B
roth
-USA
Cap
ital O
ne-U
SA
RB
S-G
BR
Ban
que
PSA
-FR
A
Mor
gan
Stan
ley-
DEU
Com
mer
zban
k -D
EU
Firs
t Nat
Ban
k-U
SA
HFC
Ban
k-G
BR
Mor
gan
Stan
ley-
LUX
Ban
k of
Sco
tland
-GB
R
Lloy
ds -G
BR
Citi
-USA SG
-FR
A
Mor
gan
Stan
ley-
GB
R
USA
& Ca
nada
UK
W
este
rn E
urop
e Ja
pan
East
Asia
& P
acifi
c Ea
ster
n E
urop
e & C
entra
l Asia
A
B Li
nk fr
om B
to A
A B
Link
from
A to
B
Thisfig
urepresen
tsavisu
alizationof
thene
tworkcompo
sedof
thetop10
0lend
ersa
ndbo
rrow
ersa
ndthelin
ksbe
tweenthem
that
areworth
US$
500millionor
more
inthesample19
80-200
7.Th
evisu
alizationpresen
tsthelin
kage
sbetweenatotalo
fN=
118ba
nks(
node
s);a
totalo
fE=
212lin
ks(edg
es).Link
sare
directed
,with
conv
exed
gesrepresen
tinglend
ingfrom
node
ito
node
j.Th
esize
oftheno
desisprop
ortio
naltothede
gree
ofeach
node
andthesize
ofthelin
ksisprop
ortio
nal
tothesize
ofthelend
ingor
borrow
ing.
Thecoloro
fthe
node
siden
tifies
bank
sge
ograph
ically
with
sixdistinct
geog
raph
icregion
s,as
describ
edin
thelege
nd.T
helin
ksarecoloredaccordingto
thecoloro
fthe
target.T
hegrap
hpresen
tsthena
mes
oftheba
nks’pa
rent,followed
byathreelette
racron
ymforthe
coun
trywhe
rethe
bank
islocated.
Fora
nexplan
ationof
theprocesso
fassigning
names
seefootno
te23
.
18
Page 20 of 49
Accep
ted
Man
uscr
ipt
5. Financial Integration and Banking Crises: Non-Parametric Analysis
To gauge the association between the de facto financial integration of banks with the incidence
of banking crises, this section performs a non-parametric analysis based on Chi-Squared indepen-
dence tests. Three independence tests are performed: Pearson Chi-squared, Likelihood-ratio, and
Fisher’s exact test.24
The data for this analysis use a dummy that takes the value 1 for countries that have suffered a
systemic banking crisis in the period 1980-2007, and zero otherwise. There are a total of 61 coun-
tries with at least one crisis in the period, out of a total of 116 in the sample with network statistics.
The sample of countries and number of crises per country is reported in Table A-1.
I compute the two-way tabulations and independence tests using quartiles of the network statis-
tics proxing for de facto financial integration of banks described in Section 2. I use a total of 22 dif-
ferent network statistics: sixteen statistics for average bank betweenness, weighted clustering co-
efficient, clustering coefficient, authority, hub centrality, degree, indegree, and outdegree, and the
corresponding unweighted and weighted measures. And six additional network statistics for the
sum of all banks in a country, including degree, indegree, and outdegree, and the corresponding
weighted statistics (recall that throughout the text weighted indegree is referred to as borrowing,
and weighted outdegree is referred to as lending).
As described in Section 2, the battery of network statistics captures different aspects of the de
facto financial integration of the country. Average betweenness, the clustering coefficients, author-
ity, and hub centrality capture how important and central the average bank of a country is to the
global bank network as an intermediary. The degree measures give us an idea of the financial in-
tegration of the average bank in terms of borrowing or lending relationships, or their size in the
case of borrowing and lending. The sum of all banks borrowing is simply the same as the usual
measure of flows into a country. It captures the size of borrowing by banks.
Table 2 reports summarized results of the independence tests for the network statistics and a
set of institutional variables of interest.25 The p-values of the independence tests indicate that an
24These tests are performed using two-way tabulations in which banking crises are on the rows and the networkstatistics proxing the de facto financial integration of banks are on the columns. The null hypothesis in these tests hasthe general form: Ho : Pij = Pi+ ∗ Pj+, that is, the probability that an observation selected at random will be classifiedin the ith row and the jth column is equal to the marginal probability that the observation is classified in the ith rowtimes the marginal probability of being classified in the jth column. Thus the null hypothesis implies that the rows arestatistically independent from the columns.
25The underlying two-way tabulations of the results summarized in Table 2 are available upon request.
19
Page 21 of 49
Accep
ted
Man
uscr
ipt
increased de facto financial integration of the average bank is statistically correlated with banking
crises. The independence tests tell us that unweighted betweenness of the average bank is statis-
tically associated with the incidence of banking crises over the period. An statistical significant
association with the occurrence of at least one banking crisis in the period is found for most other
measures that capture how important as intermediary is the average bank in a country, including
the measures of hub centrality and weighted clustering coefficients.
Outdegree and lending measures are also associated with the incidence of crises, indicating
that the larger the number of international borrowers or international lenders of the average bank
in a country, the more likely is the country to endure a banking crisis. The same is found for hub
centrality and its weighted version (no surprise here, since outdegree, lending, and hub centrality
are highly correlated by construction). The indegree and borrowingmeasures for the average bank
seem not to be associated with crises, as it is also the case for authority. As expected from most
empirical evidence, total bank lending is strongly associated with the incidence of crises.
To gauge the association between banking crises and financial globalization, I use a proxy for de
jure capital account openness. This index goes from 0 to 3 and is increasing in the level of openness
of the capital account transactions. The index is taken from Abiad et al. (2010).26 The tests point
that there seems to be no statistical relationship between the incidence of banking crises and de jure
capital account openness. One reason, as argued by Edwards (2007), who reports similar results,
may be that countries often find ways to circumvent regulations and de jure capital controls. The
p-values of the independence tests, however, do not suggest strong evidence of independence.
Employing the indexes elaborated by Abiad et al. (2010), I also perform this non-parametric
analysis for some other variables of interest. In particular, I perform independence tests for the
incidence of banking crises and an index of financial reform and for an index of banking super-
vision. The results indicate that there is a statistical association between the incidence of banking
crises and both the index of financial reform and the level of regulation of the banking system.
In summary, the non-parametric analysis via independence tests suggests that the incidence
of banking crises is associated with the level of de facto financial integration, as measured by be-
tweenness, clustering coefficients, hub centrality, outdegree, and lending of the average bank. Total
lending by banks in a country is also associated with the incidence of crises. The analysis also in-
26There are several indexes in the literature, and I will use Chinn and Ito (2008)’s KAOPEN index in the regressionanalysis later on. However, for this non-parametric analysis based on two-way tabulations I use the discrete index byAbiad et al. (2010). This index has a correlation of 0.73 with the continuous, and more nuanced, KAOPEN index.
20
Page 22 of 49
Accep
ted
Man
uscr
ipt
Table 2: Results of Independence Tests and Bivariate Probit Regressions of Indicator for aBanking Crisis in 1980-2007
Obs Pearson LR Fishers Stat Assoc Sign p-valueat 5% probit probit
Betweenness 116 0.071 0.069 0.074 No† - 0.002Betweenness (weighted) 116 0.142 0.14 0.152 No - 0.160Hub Centrality 116 0.015 0.013 0.015 Yes - 0.192Hub (weighted) 116 0.036 0.034 0.038 Yes - 0.488Weight. Clust. Coeff. 116 0.214 0.203 0.217 No + 0.925Weight. Clust. Coeff. (weighted) 116 0.006 0.005 0.005 Yes - 0.090Clust. Coeff. 116 0.214 0.203 0.217 No + 0.769Clust. Coeff. (weighted) 116 0.006 0.005 0.005 Yes - 0.104Authority 116 0.063 0.057 0.065 No† - 0.140Authority (weighted) 116 0.509 0.508 0.523 No - 0.074Degree 116 0.131 0.125 0.138 No - 0.267Borrowing+Lending 116 0.116 0.112 0.124 No - 0.001Indegree 116 0.331 0.329 0.359 No + 0.228Borrowing 116 0.147 0.144 0.16 No - 0.005Outdegree 116 0.001 0.001 0.001 Yes - 0.009Lending 116 0.006 0.005 0.005 Yes - 0.066Sum Degree 116 0.071 0.068 0.074 No† - 0.345Sum Borrowing+Lending 116 0.016 0.013 0.016 Yes - 0.333Sum Indegree 116 0.46 0.456 0.471 No + 0.505Sum Borrowing 116 0.081 0.074 0.082 No† - 0.555Sum Outdegree 116 0.02 0.018 0.019 Yes - 0.140Sum Lending 116 0.039 0.037 0.04 Yes - 0.250Cap. Acc. Liber. Index 76 0.177 0.171 0.198 No - 0.056Financial Reform Index 76 0.075 0.073 0.083 No† - 0.025Bank supervision 76 0.000 0.000 0.000 Yes - 0.001
This table presents a summary of the results of the non-parametric analysis. The first columns present theresults of three independence tests (Pearson Chi-squared, Likelihood-ratio, and Fisher’s exact test) testing thenull hypothesis of statistical independence between the occurrence of a systemic banking crisis in the period1970-2007 and the 25 variables listed in the rows of the table. The independence tests are performed on anindicator for a systemic banking crisis in the period against quartiles of each of the 25 listed variables. Thecolumns present the p-value of each test, and a summary result (Yes or No) of the association based on thesignificance of the three tests. † means significance at 10% level. The last two columns present the results ofbivariate Probit regressions of an indicator dummy for a banking crisis in the period 1980-2007 on each of the 25listed variables. All regressions include a constant and are estimated using heteroskedasticity-robust standarderrors. The table presents the sign of the estimated coefficient and its corresponding p-value.
dicates that banking regulation is statistically associated with the occurrence of crises.
The analysis so far cannot answer questions about the direction of the relationship, let alone
causality. The independence tests offer a simple but systematic way to study the statistical associ-
ation or independence between the incidence of two categorical variables. However, one should
not read more than that into these tests: the tests tell us if there is an association between the two
variables, but not what the association is.
21
Page 23 of 49
Accep
ted
Man
uscr
ipt
I complement this non-parametric analysis by running simple bivariate probit models, with
the dummy for banking crises as dependent variable and the different network statistics and other
variables of interest as explanatory variables. This approach allows us to gauge the sign and the
significance of the relationship parametrically. The last two columns of Table 2 show the sign of the
estimated coefficients and corresponding p-values of these bivariate probit regressions. Most of the
results from the non-parametric analysis are corroborated by the regressions, with (unweighted)
betweenness exhibiting a significant coefficient. Measures of borrowing, outdegree, and lending
of the average bank also show up significant.
One limitation of the approach employed so far is that it cannot capture the interactions of the
two variables once we control for other plausible determinants of the incidence of crises. More
importantly, the structure of the data only exploits the occurrence of one crisis in the cross-section.
Given that there are countries that experience two and four crisis in the period, this approach
disregards valuable information. To fully exploit the information in the data, in the next section I
perform a multivariate regression analysis using a count data model.
6. Financial Integration and Banking Crises: Regression Analysis
The regression analysis in this section explores the hypothesis whether or not higher financial
integration of banks increases the incidence of banking crises. The strategy is to fit an empirical
count model in the cross-section of the data. In this model the dependent variable is the number
of banking crises over the period in a given country, and the explanatory variables are a vector of
averages ofmacroeconomic and institutional variables the literature has found relevant to influence
the occurrence of banking crises, plus the proxies of de facto financial integration based on network
statistics. I estimate variations of the model:
bankcrisesi = α+ βFinInti + ψXi + ξi (1)
where bankcrisesi is the sum of banking crises over the period 1980-2007 in country i, α is a con-
stant, FinInti is a vector of de facto financial integration measures for country i based on network
statistics, and Xi is a vector of macroeconomic and institutional variables, including a dummy
for income,27 openness in trade, an index of institutional quality (political risk), banking credit to
private sector (as percentage of GDP), the current account balance (as percentage of GDP), and
27This dummy takes the value 1 if the country is a high-income OECD member and the value 0 otherwise.
22
Page 24 of 49
Accep
ted
Man
uscr
ipt
inflation. Controls for de jure financial integration are also included, as is an index of quality of
banking supervision. Details on each variable used in the paper can be found in Table A-2.28 Table
A-3 presents summary statistics of the variables.
This cross-sectional analysis should be taken with some caution. Since the end and start of the
period are arbitrary, the averages ofmacro variables and institutional indexes should be considered
carefully. On the other hand, since we are using averages, there are some relevant macroeconomic
variables that must be omitted because they do not carry relevant information in a cross-section
set up or may present some endogeneity with banking crises.29 The cross-section setting also pre-
cludes a discussion about causality. Nonetheless, this approach offers a baseline from which we
can evaluate the association between de facto financial integration and banking crises.
As explained in Cameron and Trivedi (1998), the model selection for count data depends on
the dispersion characteristics of the data. The Poisson model assumes equidispersion; that is, it
assumes that the conditional variance of the dependent variable is equal to its conditional mean.30
If the conditional mean is larger (smaller) than the variance, the data present overdispersion (un-
derdispersion). In the case of equidispersion, the Poisson model yields consistent and efficient
estimates. However, if the data are overdispersed, the Poisson regression loses efficiency.31
Figure A-3 shows a histogram of the distribution of banking crises. The figure shows that most
of the countries did not suffer a crisis in the period 1980-2007, and only one country (Argentina)
faced more than two crises, with a tally of four. The unconditional mean and variance reported in
Table A-3 do not suggest large differences in mean and variance, but we cannot be sure if the data
are under or overdispersed when including all covariates in the analysis (i.e., we do not know how
the conditional means will turn out).
28The literature on banking crises has found as determinants of banking system distress a set of macroeconomicand institutional variables which includes above-trend credit growth, financial (banking) system regulation, levels ofcorruption, exchange rate regime, and trade openness, among others. A rigorous study of the determinants of bankingcrises is beyond the scope of this paper. See Demirgüc-Kunt and Detragiache (2005) for a survey. After experimentingwith a large set of macroeconomic and institutional variables, I settled for the most parsimonious model that capturesthe main variables suggested in the literature and that are suited for a cross-sectional analysis.
29For example, interest rates have been found to be related to banking crises; however, once the crisis takes placethe interest rate becomes an endogenous variable. The same can be said about ratio of reserves to broad money andmoney growth. Thus, I selected as controls variables that would be meaningful after taking averages over a quarter-century period. The variable of credit as percentage of GDP is obviously affected by the occurrence of a crisis, sincelending collapses by definition. However, the average over a long period still carries information about the financialdevelopment of the country, and hence it is included in the analysis.
30Formally, the Poisson model assumes that the number of occurrences of the event y over a fixed exposure periodhas the probability mass function Pr(Y = y) = e−µµγ/γ!, y = 0, 1, 2, . . .; and assumes E(Y ) = V ar(Y ) = µ.
31For this reason the Negative Binomial model is preferred in the case of overdispersion. However, a limitation of thismodel is that its assumptions imply overdispersion in the data, making it unsuitable for underdispersed data.
23
Page 25 of 49
Accep
ted
Man
uscr
ipt
An alternative when the assumption of equidispersion is violated in the data is to estimate the
model using a Generalized Linear Model (GLM) technique. The GLM approach can take into ac-
count the over or underdispersion of the data and still yield consistent coefficients estimates. To
achieve consistency, the GLM is estimated assuming a Poisson distribution and using a pseudo-
maximum likelihood estimator (PMLE) and robust standard errors. The variance can be consis-
tently estimated using the “robust sandwich” estimator (Cameron and Trivedi, 1998, p.65). This
estimator will produce consistent estimates provided that the conditional mean is well specified.
This requires that the assumed density belongs to the Linear Exponential Family (Cameron and
Trivedi, 1998, p.31). In the case of mis-specification of the conditional mean, however, the GLM
PMLE loses efficiency.
Given the consistency in the coefficients under the GLMPMLE, even undermis-specification of
the conditional mean, this regression framework is enough to evaluate our hypothesis on whether
financial integration of the banking system is associated with an increase in the incidence of bank-
ing crisis. Furthermore, in the case of over or underdispersion the direction of the bias is known
for the Poisson model Winkelmann (2008, p.91). If the data exhibit underdispersion the variance
is overestimated, and the model tends to yield t-values too small and, as a result, the model will
produce more than necessary stringent p-values. In the case of overdispersion, the variance is
underestimated, resulting in excessively large t-values and possibly spurious inference.
6.1. Results
Table 3 presents results of estimating equation 1 assuming a Poisson distribution for the con-
ditional mean, and using the GLM PMLE and robust standard errors. The table shows the results
of the model including the network statistics that proxy for de facto financial integration of banks
entering as explanatory variables one by one in a baseline specification of the model, and the base-
line set of controls. The table reports coefficient estimates and some statistics of the fit of the model
including estimated deviance, the Pearson deviance, the dispersion of these twomeasures (i.e., the
variables scaled by the degrees of freedom),32 and the Bayesian Information Criterion.
32The deviance has an approximate chi-square distribution with N −K degrees of freedom, where N is the numberof observations and K is the number of predictor variables including the intercept. Since the expected value of a chi-square random variable is equal to the degrees of freedom, then if the model fits the data well the ratio of the devianceto degrees of freedom, or its dispersion, should be about one. Thus, the estimated dispersion can be used as a gaugeof the assumption of equidispersion. Values greater than 1 indicate overdispersion, and values smaller than 1 indicateunderdispersion. Section 6.1.2 discusses the fit of the model.
24
Page 26 of 49
Accep
ted
Man
uscr
ipt
Table3:
Results
ofGM
LM
ultiv
ariate
Reg
ressions
.Num
berof
Bank
ingCriseson
each
NetworkStatistic
,and
Baselin
eSe
tofC
ovariates.
Sample19
80-200
7
βs.e
.N
Dev
ianc
eDev
ianc
e_df
Pearson
Pearson_
dfBIC
Betw
eenn
ess
-2.425
***
[0.719]
7426.397
0.412
22.930
0.358
-249.063
Betw
eenn
ess(
weigh
ted)
-0.766
[2.091]
7428.591
0.447
26.434
0.413
-246.869
Hub
Cen
trality
-2.378
[2.534]
7428.230
0.441
25.436
0.397
-247.230
Hub
(weigh
ted)
-5.918
[10.809]
7428.262
0.442
25.171
0.393
-247.199
Weigh
t.Clust.C
oeff.
-0.749
[2.589]
7428.691
0.448
25.797
0.403
-246.769
Weigh
t.Clust.C
oeff.
(weigh
ted)
-8.949
*[4.927]
7427.778
0.434
23.963
0.374
-247.682
Clust.C
oeff.
-1.128
[2.476]
7428.653
0.448
25.667
0.401
-246.807
Clust.C
oeff.
(weigh
ted)
-6.684
*[3.435]
7427.582
0.431
23.493
0.367
-247.878
Autho
rity
0.367
[0.940]
7428.633
0.447
26.342
0.412
-246.827
Autho
rity(w
eigh
ted)
0.798
[1.236]
7428.446
0.444
26.091
0.408
-247.014
Deg
ree
-0.009
[0.030]
7428.649
0.448
25.701
0.402
-246.811
Borrow
ing+
Lend
ing
-0.386
[0.383]
7428.288
0.442
24.899
0.389
-247.172
Inde
gree
0.004
[0.028]
7428.709
0.449
25.989
0.406
-246.751
Borrow
ing
-0.127
[0.440]
7428.690
0.448
25.822
0.403
-246.770
Outde
gree
-0.155
*[0.093]
7427.262
0.426
23.206
0.363
-248.199
Lend
ing
-1.229
[0.879]
7427.791
0.434
24.354
0.381
-247.669
Sum
Deg
ree
0.020
[0.271]
7428.713
0.449
25.854
0.404
-246.747
Sum
Borrow
ing+
Lend
ing
0.528
[1.286]
7428.542
0.446
25.396
0.397
-246.918
Sum
Inde
gree
0.044
[0.407]
7428.708
0.449
25.812
0.403
-246.752
Sum
Borrow
ing
2.530
[2.797]
7427.969
0.437
24.960
0.390
-247.491
Sum
Outde
gree
0.005
[0.484]
7428.718
0.449
25.963
0.406
-246.742
Sum
Lend
ing
0.188
[2.305]
7428.710
0.449
25.888
0.404
-246.750
This
tablepresen
tsresu
ltsof
multiv
ariate
regression
sof
thecoun
tdatamod
elof
equa
tion1.
Inallreg
ressions
thede
pend
entv
ariable
isthenu
mbe
rof
system
icba
nkingcrises
inthepe
riod19
80-200
7.Ea
chrow
show
stheresu
ltsof
aregression
inwhich
thecova
riates
includ
eon
eprox
yford
efactofin
ancial
integrationba
sedon
netw
orkstatistic
sand
aseto
fbaselinecontrols(not
show
n).C
ontrolsinc
lude
trad
eop
enne
ss,p
oliticalrisk,
domestic
cred
itas
percen
tage
ofGDP,cu
rren
taccou
ntba
lanc
e,infla
tion,
bank
ingsu
pervision,
capitala
c-coun
tope
nness,an
dan
indicatorfor
high
-incomeOEC
Dcoun
try.
Theconstruc
tionof
thene
tworkstatistic
sise
xplained
inSe
ction2;
all
othe
rcon
trolsa
recoun
tryaverag
esdu
ringthepe
riod19
80-200
7.Th
eregression
sare
estim
ated
usingaGen
eralized
Line
arMod
el(G
LM)
approa
chin
which
theestim
ationassu
mes
aPo
issondistrib
utionfort
hecond
ition
almeanof
thede
pend
entv
ariable.
Theestim
ationis
implem
entedus
ingaps
eudo
-max
imum
likelihoo
destim
ator
(PMLE
)and
robu
ststan
dard
errors.B
esides
repo
rtingtheestim
ated
coeffi
-cien
tfor
theva
riableof
interestin
each
regression
,the
tablesh
owss
tatis
ticso
fthe
fitof
themod
el,inc
luding
theestim
ated
devian
ce,the
Pearsonde
vian
ce,the
disp
ersion
ofthesetw
omeasu
res(
i.e.,theva
riables
scaled
bythede
greeso
ffreed
om),an
dtheBa
yesian
Inform
a-tio
nCrit
erion.∗indicatess
ignific
ance
at10
percen
tlevel,∗∗indicatess
ignific
ance
at5pe
rcen
tlevel,a
nd∗∗∗indicatess
ignific
ance
at1
percen
tlevel.
25
Page 27 of 49
Accep
ted
Man
uscr
ipt
The results indicate that the incidence of crises is negatively correlated with measures that cap-
ture how important is the average bank as intermediary in the global bank network (intermediary
in the sense of being a potential intermediary of bank relationships). The coefficient for between-
ness shows up with a negative sign and significant at the 1% level. In all cases, the coefficients of
hub centrality are negative, although not statistically significant. The negative association of crises
with these measures is found in all cases, when computing the bank-level statistics taking into ac-
count only the link structure of the network (unweighted) and when weighting the statistics by the
bank’s share of global flows (ωi).
These results suggest that the more important is the average bank as intermediary in the global
bank network, the smaller the number of banking crises the country experiences. One way to
rationalize this result is to see betweenness as a measure capturing how easy is for a country to
tap international capital markets: a country that has relatively easy access to debt flows is able to
avoid a liquidity crisis.
One puzzling result is the negative coefficients for the clustering coefficients. As the centrality
of a node decreases the larger is the clustering coefficient, these results are opposite of what is
suggested by the other centrality measures (betweenness and hub centrality). However, it is hard
to know what to make of this result. Betweenness and hub centrality are measures based on the
link structure of the whole network, capturing how important is a node as an intermediary to any
pair of nodes. In contrast, the clustering coefficients are local in nature, capturing how important is
the role of the node as intermediary of two of its neighbors. Since the idea is to use these network
statistics to capture the de facto integration of banks into the global bank network, betweenness is
my preferred centrality measure, as it takes into account the link structure of the whole network.
On the other hand, the results also indicate a negative and significant correlation of crises with
the number of lending relationships that the average bank in a country has (outdegree). The statis-
tics based on the number of borrowing relationships (indegree and authority) and the statistics
measuring the actual size of the flows (borrowing, lending, and total flows) show up with positive
coefficients, suggesting that countries with many lending relationships and larger international
banking flows are more prone to crises (although, none of these coefficients is statistically signifi-
cant in these specifications where the regressions control for only one network statistic at a time).
Table 4 reports the results of estimating the model including at once all the network statistics
that exhibited a significant coefficient in the one-by-one regressions. The specification in column
26
Page 28 of 49
Accep
ted
Man
uscr
ipt
1 includes betweenness, the weighted clustering coefficient, and outdegree.33 The results of this
richer specification indicate that it is betweenness the only variable that survives with a statistically
significant coefficient.
Columns 2 adds hub centrality and authority to the specification in column 1. These two vari-
ables showed up as significant in the nonparametric analysis. The results of the last two specifi-
cation tell us that betweenness is the only centrality statistic that is strongly correlated with the
incidence of banking crises—even after controlling for all other significant centrality measures.
Onewould like to add to this specification lending, the remaining network statistic that showed
up with a significant coefficient in the nonparametric analysis. But to do so, we have to drop from
the model the non-significant variables of hub centrality and authority, as they have very high
correlations with lending. For the same reason we also have to drop outdegree. Specification
3, then, estimates the model including the baseline covariates, plus betweenness, the weighted
clustering coefficient, and lending. This specification actually shows a marginal better fit that the
specification with outdegree, reason why we will keep lending in our benchmark model.
Given the literature on banking crises and capital inflows discussed in the introduction, it is
desirable to add controls for total borrowing in the country (a variable that showed up significant
in the nonparametric analysis). Column 4 adds total borrowing to the benchmark of specification 3.
The coefficient of betweenness is always negative and highly statistically significant. The coefficient
for total borrowing is positive and highly significant, which is a result in line with the theoretical
and empirical literature linking debt flows and banking crises.
Regarding the covariates of the baseline model, in all cases the results of the model indicate
that measures of de jure financial integration (i.e., capital account openness) are positively corre-
lated with the incidence of crises. The other macroeconomic variables and institutional indexes
show up with the expected sign, and in most cases are significant. The coefficient estimates sug-
gest that the incidence of banking crises is decreasing in the degree of trade openness of a country,
as also found by Rose and Spiegel (2009). As expected, the larger the political risk in a country, the
33There is some degree of correlation between a few of the network statistics. In particular, the unweighted statisticsare highly correlated with their weighted versions. On the other hand, the weighted clustering coefficient is highlycorrelated with the normal clustering coefficient. As expected, there are high correlations of hub centrality with au-thority, and of these two statistics with outdegree/lending and indegree/borrowing, respectively. Betweenness is alsohighly correlated with outdegree. Unsurprisingly, there are high correlations of borrowing and lending with many ofthe weighted statistics. Thus, from an econometric perspective, we cannot include all variables in one specification. Ichoose to include the variables that showed up a significant coefficient in Table 3 andmake sense from an economic per-spective. I preferred the weighted clustering coefficient over the normal clustering coefficient, because it is computedtaking into account the size of flows of the neighbors. See correlations of country-level network statistics in Figure A-5.
27
Page 29 of 49
Accep
ted
Man
uscr
ipt
larger the number of banking crises (a coefficientwith a strong statistical significance). Not surpris-
ingly, with data up to 2007, the indicator for high income countries is negatively correlated with
the incidence of banking crises. The coefficient for the index of banking supervision is always sig-
nificant and with a negative sign, suggesting that the existence of prudential banking supervision
is an important factor in reducing the incidence of crises.
Finally, column 5 of Table 4 estimates the model with the variables that were found to be sig-
nificant in the richest specification. This is the preferred model, and we will use it to gauge how
economically significant are the effects of de facto financial integration in the incidence of crises.
Again, similar results for betweenness of the average bank and total borrowing are obtained.
Taken altogether, the results suggest that the higher is the international borrowing of banks
in a country, the higher is the incidence of banking crises. However, the occurrence of crises is
lower in countries where the average bank plays an important role as intermediary in the global
network of syndicated loans (proxied by the network statistic betweenness). I conjecture that a
large betweenness captures the ability of a country to roll over or obtain short-term debt, as banks
that are perceived as central to the network by other banks may be in capacity to tap international
markets more easily.
These two different results for borrowing and betweenness are not contradictory. One can think
in terms of the model proposed by Chang and Velasco (2001), in which large debt inflows make
more likely the occurrence of a crisis, while at the same time the inability of obtaining either on-
going or short-term debt increases financial system fragility.
28
Page 30 of 49
Accep
ted
Man
uscr
ipt
Table 4: Baseline Restuls of GLM Multivariate Regressions. Model Includ-ingCentrality andDegreeNetworkStatistics forAverageBank. Sample 1980-2007
(1) (2) (3) (4) (5)Betweenness -1.893** -1.922** -2.117*** -2.928*** -2.873***
[0.776] [0.786] [0.639] [0.682] [0.601]Weighted Clust. Coeff. (weighted) -6.773 -6.897 -5.463 -6.051
[4.533] [4.756] [4.046] [4.263]Outdegree -0.107 -0.086
[0.082] [0.108]Hub Centrality -0.458
[2.381]Authority 0.399
[0.914]Lending -0.763 -2.070
[0.757] [1.367]Sum Borrowing 5.570*** 3.753**
[1.741] [1.638]Trade openness -0.871*** -0.900*** -0.908*** -0.730*** -0.763***
[0.251] [0.250] [0.244] [0.235] [0.209]Political risk 0.048*** 0.047*** 0.049*** 0.051*** 0.050***
[0.012] [0.013] [0.012] [0.012] [0.012]Domestic credit 0.158 0.173 0.039 -0.016
[0.295] [0.314] [0.250] [0.306]Current account balance 0.017 0.014 0.016 0.024
[0.025] [0.026] [0.025] [0.025]Inflation 0.076 0.075 0.077 0.074 0.080
[0.052] [0.052] [0.050] [0.055] [0.050]High income-OECD country -1.717*** -1.911** -1.813*** -1.779*** -2.344***
[0.481] [0.814] [0.460] [0.515] [0.451]Bank supervision -0.583*** -0.539** -0.563*** -0.708*** -0.747***
[0.206] [0.210] [0.217] [0.187] [0.193]Capital account openness 0.123* 0.124* 0.124* 0.119* 0.120*
[0.069] [0.070] [0.067] [0.070] [0.070]Constant -1.783** -1.711** -1.761** -1.857*** -1.861***
[0.712] [0.736] [0.699] [0.702] [0.704]Deviance 25.212 25.112 25.560 22.449 24.782Deviance_df 0.407 0.419 0.412 0.368 0.381Pearson 20.466 20.810 21.799 16.897 20.800Pearson_df 0.330 0.347 0.352 0.277 0.320BIC -241.640 -233.132 -241.292 -240.099 -254.982Obs 74 74 74 74 74
This table presents results of multivariate regressions of the count data model of equation1. In all regressions the dependent variable is the number of systemic banking crises in theperiod 1980-2007. Each column shows the results of a regression in which the covariatesinclude a set of proxies for de facto financial integration based on network statistics and theset of baseline controls shown in the table. The construction of the network statistics is ex-plained in Section 2; all other controls are country averages during the period 1980-2007. Theregressions are estimated using a Generalized Linear Model (GLM) approach in which theestimation assumes a Poisson distribution for the conditional mean of the dependent vari-able. The estimation is implemented using a pseudo-maximum likelihood estimator (PMLE)and robust standard errors. The table shows at the bottom statistics of the fit of the model, in-cluding the estimated deviance, the Pearson deviance, the dispersion of these two measures(i.e., the variables scaled by the degrees of freedom), and the Bayesian Information Crite-rion. ∗ indicates significance at 10 percent level, ∗∗ significance at 5 percent level, and ∗ ∗ ∗significance at 1 percent level.
29
Page 31 of 49
Accep
ted
Man
uscr
ipt
6.1.1. Marginal Effects
To gauge the economic significance of the estimated coefficients, I estimate both Incidence Rate
Ratios (IRR) and Average Marginal Effects (AME) for the preferred specification in column 5 of
Table 4. IRRs in the Poisson model are calculated as the rate of change in the outcome (incidence)
of the dependent variable as a response of a one-unit change in an independent variable. In the
Poissonmodel the IRR are simply exponentiated coefficients (Hardin andHilbe, 2007, p. 197). With
two covariates and a constant, the calculation of the incidence rate ratio for variable 1 reduces to:
IRR1 = exp[β0+(x1+1)β1+x2β2]exp[β0+x1β1+x2β2]
= exp[β1]. Thus, the IRR is positive and constant (it does not depend
on the particular values of the regressors). However, a limitation of IRRs is that the scale, the units
in which the regressors are measured, affects the estimation.
Marginal effects in the Poissonmodel are calculated as elasticities and semi-elasticities. Cameron
and Trivedi (1998) suggest that the best way to gauge the average response across observations of a
change in regressor j is to compute AMEj = 1N
∑i∂E[yi|xi]∂xi
= 1N βjexp[x
′iβ]. In the Poisson model
with intercept this simplifies to AMEj = βj y. This measure gives the change in y, given a unit
change in x. Like IRRs, this measure will be affected by the scale. Thus, it is preferable to compute
average elasticity: eyex = AMEj × xy , which gives the average proportionate change in y associated
with a proportionate change in x. Alternatively, we can estimate the semi-elasticity dyex = AMEj× 1
y ,
which gives the average change in y associated with a proportionate change in x. These elasticities
and semi-elasticities, along with IRRs, are reported in Table 5 for specification 5 of Table 4.
The results suggest that the negative effects in the incidence of banking crises coming from de
facto and de jure financial integration are not large when compared with other covariates. Other
factors seem to play a more crucial role as determinants of banking crises, in particular prudential
regulation of the banking sector, political risk, and trade openness. Interestingly, the results sug-
gest that the beneficial effects of de facto financial integration coming from a higher connectedness,
rationalized here as ease of access to international lending, are an order of magnitude larger than
the negative effects.
As Table 5 reports, the estimated elasticity eyex suggests that a 10 percent increase in the borrow-
ing of all banks in a country increases the occurrence of crises by 1.1 percent. Also, a 10 percent
increase in the de jure openness of the current account would have an increase in the occurrence
of crises of 0.5 percent. These effects are relatively small when compared with the negative effects
from increased political risk, reduced banking supervision or reduced trade openness. For exam-
30
Page 32 of 49
Accep
ted
Man
uscr
ipt
Table 5: Incidence Rate Ratios and AverageMarginal Effects
IRR eyex
dyex
Betweenness 0.0565 -0.2943 -0.1636Sum Borrowing 42.6397 0.1144 0.0636Trade openness 0.4664 -0.5707 -0.3172Political risk 1.0509 3.3664 1.8710Inflation 1.0828 0.0454 0.0252Banking supervision 0.4739 -0.7756 -0.4310KA open 1.1272 0.0534 0.0297
This table presents the estimated marginal effects of themodel in specification 5 of Table 4. The table reportsaverage marginal effects expressed as incidence rate ra-tios (IRR) and semi-elasticities. ey
exis the average pro-
portionate change in y associated with a proportionatechange in x; while dy
exis the average change in y associ-
ated with a proportionate change in x.
ple, the elasticity eyex calculates that a 10 percent increase in the index of prudential bank regulation
reduces the occurrence of banking crises by 7.7 percent. Also, the estimated elasticity suggests that
a 10 percent increase in the measured betweenness of the average bank will reduce the occurrence
of banking crises by 2.9 percent.34
IRR estimates are also reported. However, as stated before, a limitation of this measure is that
it is not scale-free. This is particularly problematic for interpreting the coefficient of borrowing
by all banks, as it is expressed in billions in the data used in the estimation. The results suggest
that an increase in 1 unit in the current account openness index would increase the occurrence of
banking crises by 13 percent. On the other hand, a one-unit increase in the banking supervision
index would decrease the occurrence of banking crises by (1-0.47)×100 = 53 percent.
6.1.2. Goodness of Fit of the Model
Tables 3 and 4 report goodness of fit statistics of the model. The deviance and the Pearson
deviance (P ) are to be compared with N − K, where N is the number of observations and K
is the number of regressors. Obtaining P < N − K implies underdispersion of the data if the
conditional mean has been well specified. Scaled deviance statistics of less than one also indicate
34Alternatively, the semi-elasticity dyex
indicates that a 10 percent improvement in the banking supervision index willbe associated with 0.04 fewer banking crises. In comparison, an increase of 10 percent in the current account opennessindex will be associated with 0.003 more banking crises, while the same proportionate increase in total borrowing ofbanks will be associated with 0.006 more banking crises. On the other hand, a 10 percent increase in the measuredbetweenness of the average bank will be associated with 0.016 fewer banking crises.
31
Page 33 of 49
Accep
ted
Man
uscr
ipt
underdispersion. It is evident from the computed statistics that the data exhibit some degree of
underdispersion. This implies that the Poisson model is a better fit than the Negative Binomial,
and it is the reason why I followed this strategy.
However, we must bear in mind that P 6= N −K can also indicate that the conditional mean
has been mis-specified. This does not imply, however, that the results from the model are not
consistent. As discussed before, the consistency of the Poisson estimator is obtained by the use of
the GLM PMLE estimator (although it loses efficiency if data are underdispersed). Furthermore,
with underdispersion the direction of the bias is known for the Poisson model. Since the variance
is overestimated in this case and we obtain greater than necessary stringent p-values, we can be
confident about the statistical inference from the results reported above.
An overall test of adequacy of the model is to evaluate how close the residuals are to normality
(Cameron andTrivedi, 1998, p. 145). Table 6 presents the results of normality tests for the estimated
residuals of the model in specification 5 of Table 4. Figure A-4 in the Appendix plots the kernel
density of the residuals against a normal distribution. These tests indicate that the baseline model
provides a fair representation of the data.
Table 6: Results of Normality Tests of Residuals
Test W or W’ V or V’ Z Prob>zShapiro-Wilk W 0.98543 0.938 -0.139 0.55546Shapiro-Francia W’ 0.98419 1.119 0.225 0.41113
Pr(Skewness) Pr(Kurtosis) Adj chi2(2) Prob>chi2Skewness/Kurtosis 0.3279 0.4982 1.46 0.4817
This table presents the results of normality tests for residuals of the model of specifica-tion 5 of Table 4. The table shows the results of three normality tests, the Shapiro-Wilk,Shapiro-Francia, and the Skewness and Kurtosis test for normality. The null hypothe-sis in these tests is that the residuals behave as a normal distribution. The table reportsthe estimated values of the tests and corresponding p-values.
Table 4 also reports the deviance and the Pearson deviance, which are measures of aggregate
goodness of fit. The deviance is the counterpart of the sum of residuals in the linear regression
framework. The results also report Bayesian Information Criterion (BIC) statistics. These mea-
sures suggests that the model including the weighted clustering coefficient and lending (column 4
of Table 4) is the one that best fits the data. However, this specification performs poorly when we
evaluate the goodness of fit of the residuals of the model. We cannot reject the hypothesis of nor-
mality of the residuals, but at somewhat low confidence levels. Thus, the preferred specification is
column 5 of Table 4, which drops the variables that were not found to be significant, the weighted
32
Page 34 of 49
Accep
ted
Man
uscr
ipt
clustering coefficient, and lending.
6.1.3. Robustness Checks
The baseline results were obtained using all countries with at least one bank that have partici-
pated as lead arranger in the market for syndicated loans. The first robustness test is to constraint
the sample of analysis to include only countries with five or more lead banks, so that wemake sure
the results are not being driven by countries with very few banks (which may systematically show
lower levels of betweenness and other network statistics). Table 7 reports the results of estimating
themodel in this sample and including the full set of network statistics (replication of Table 4). The
sample size reduces to 67 countries, but the point estimates are basically the same as before.35
The results are also robust to the inclusion of different macroeconomic variables, including the
exchange rate regime, the coefficient of variation of the nominal and real exchange rates, and to the
use of a different proxy for quality of political institutions (Polity2 variable of the Polity IV project).
To save space, these results are available from the author upon request.
A concern with this cross-section analysis is that the network statistics may be unstable over
time and, hence, may not be suitable for a cross-section approach. Even though Hale (2012) shows
that there is little year-on-year variation in the network statistics based on the syndicated loans
market, the cross-section used in this paper is over a quarter century, and it is possible to have
important changes in the network statistics if they were computed in a different cross-section.
Figure A-6 in the Appendix documents the correlation between network statistics computed
in three different samples. The full sample used in the paper, spanning 1980-2007, one subsample
spanning the period from 1995 to 2007, and a second subsample from 2001 to 2007. The figure
reports the correlations of the country-level network statistics, which is the data used in the re-
gression models. Figure A-7 presents the correlations of the network statistics across the different
samples for the network statistics at the bank level. The figures document high correlations across
subsamples for most network statistics, indicating a high degree of inertia embedded in the net-
work statistics.36 This is actually reassuring for the cross-section analysis performed in this paper.
As a robustness check, nonetheless, I estimate the model using network statistics computed on
the two different subsamples. Table A-4 in the Appendix presents summarized results of repli-
cating the last three specifications of Table 4 with network statistics computed in the two different
35Similar results are also obtained for the non-parametric analysis. Results available upon request.36Section 4 also presented a few characteristics of these networks, finding that the two subsample networks exhibit
similar characteristics as the full-sample network.
33
Page 35 of 49
Accep
ted
Man
uscr
ipt
Table 7: Robustness Check Excluding Countries with Fewer than Five Banks.Model Including Centrality andDegreeNetwork Statistics for Average Bank. Sam-ple 1980-2007
(1) (2) (3) (4) (5)Betweenness -1.933** -1.962** -2.162*** -3.004*** -2.991***
[0.793] [0.794] [0.661] [0.699] [0.606]Weighted Clust. Coeff. (weighted) -7.436* -7.656 -6.132 -6.719
[4.514] [4.742] [4.026] [4.223]Outdegree -0.105 -0.082
[0.082] [0.108]Hub Centrality -0.408
[2.405]Authority 0.444
[0.930]Lending -0.701 -2.030
[0.752] [1.389]Sum Borrowing 5.591*** 3.727**
[1.871] [1.621]Baseline Controls Yes Yes Yes Yes YesDeviance 20.354 20.240 20.729 17.609 19.964Deviance_df 0.370 0.382 0.377 0.326 0.344Pearson 17.083 17.452 18.394 13.620 17.517Pearson_df 0.311 0.329 0.334 0.252 0.302BIC -210.904 -202.609 -210.529 -209.444 -223.908Obs 67 67 67 67 67
This table presents summarized results of multivariate regressions of the count data model of equa-tion 1, but excluding countries with less than five active lead banks in the period 1980-2007. In allregressions the dependent variable is the number of systemic banking crises in the period 1980-2007.Each column shows the results of a regression in which the covariates include a set of proxies forde facto financial integration based on network statistics and a set of baseline controls (not shown).Controls include trade openness, political risk, domestic credit as percentage ofGDP, current accountbalance, inflation, banking supervision, capital account openness, and an indicator for high-incomeOECD country. The construction of the network statistics is explained in Section 2; all other controlsare country averages during the period 1980-2007. The regressions are estimated using aGeneralizedLinear Model (GLM) approach in which the estimation assumes a Poisson distribution for the con-ditional mean of the dependent variable. The estimation is implemented using a pseudo-maximumlikelihood estimator (PMLE) and robust standard errors. The table shows at the bottom statistics ofthe fit of the model, including the estimated deviance, the Pearson deviance, the dispersion of thesetwo measures (i.e., the variables scaled by the degrees of freedom), and the Bayesian InformationCriterion. ∗ indicates significance at 10 percent level, ∗∗ significance at 5 percent level, and ∗ ∗ ∗significance at 1 percent level.
subsamples. The results are similar to the baseline obtained with the full sample.
7. Concluding Remarks
This paper explores the hypothesis ofwhether increased financial integration is associatedwith
an increased incidence of systemic banking crises. The paper uses a measure of de facto financial
integration based on the connectedness of the average bank in the global network of inter-bank
34
Page 36 of 49
Accep
ted
Man
uscr
ipt
syndicated loans. The results of both non-parametric tests and a regression analysis support the
hypothesis that financial integration is positively associated with the incidence of banking crises.
The larger the borrowing by banks in a country, the more prone the country is to financial distress.
Interestingly, the results indicate that betweenness of the average bank, which can be viewed as a
proxy for how readily a country can access international capital markets, plays an important role
in reducing the incidence of banking crises, even after controlling for the size of borrowing and the
capital account openness. The results also suggest that other factors are at work, and potentially
can have a much bigger role as determinants of banking crises. In particular, prudential banking
supervision seems to play a crucial role in reducing the occurrence of banking crises.
Even though a full-blown network analysis was out of the scope of this paper, the results
showed here yielded interesting insights on the topology of the network of inter-bank syndicated
loans. The first approximation to the degree distribution of this network offered here shows that it
is similar to other social networks, with a long tail and exhibiting a scaling similar to a Power Law.
Further exploration of the topology of this network is important, as the networks science literature
has shown that the characteristics documented in this paper have important implications for the
resilience and vulnerability of the network to different shocks.
ReferencesAbiad, A.,Detragiache, E. and Tressel, T. (2010). ANewDatabase of Financial Reforms. IMF Staff Papers, 57 (2), 281–302.
Acemoglu, D., Carvalho, V.,Ozdaglar, A. and Tahbaz-Salehi, A. (2012). The network origins of aggregate fluctuations.Econometrica, 80 (5), 1977–2016.
—,Ozdaglar, A. and Tahbaz-Salehi, A. (2010). Cascades in networks and aggregate volatility. Working Paper 16516, NBER.
Ahrend, R. andGoujard, A. (2012). International Capital Mobility and Financial Fragility - Part 1. Drivers of Systemic BankingCrises: The Role of Bank-Balance-Sheet Contagion and Financial Account Structure. EconomicsDepartmentWorking PapersNo. 902, OECD.
Albert, R. and Barabási, A.-L. (2002). Statistical mechanics of complex networks. Reviews of modern physics, 74 (1), 47–97.
—, Jeong, H. and Barabási, A.-L. (2000). Error and attack tolerance of complex networks. Nature, 406 (6794), 378–382.
Allen, F. and Babus, A. (2009). Networks in Finance. In P. Kleindorfer, Y. Wind and R. Gunther (eds.), The NetworkChallenge: Strategy, Profit, and Risk in an Interlinked World, Wharton School Publishing, pp. 367–382.
Barabási, A.-L. and Albert, R. (1999). Emergence of scaling in random networks. Science, 286 (5439), 509–512.
Barrat, A., Barthelemy, M., Pastor-Satorras, R. and Vespignani, A. (2004). The architecture of complex weighted net-works. Proceedings of the National Academy of Sciences of the United States of America, 101 (11), 3747–3752.
Bastian, M., Heymann, S. and Jacomy, M. (2009). Gephi: An Open Source Software for Exploring and Manipulat-ing Networks. International AAAI Conference on Weblogs and Social Media. Version 0.8.2, January 2013. https://gephi.org/.
Batagelj, V. and Mrvar, A. (1998). Pajek–Program for Large Network Analysis. Version 3.13, September 2013. http://pajek.imfm.si.
35
Page 37 of 49
Accep
ted
Man
uscr
ipt
Battiston, S., Delli Gatti, D., Gallegati, M., Greenwald, B. and Stiglitz, J. E. (2012). Liaisons dangereuses: Increasingconnectivity, risk sharing, and systemic risk. Journal of Economic Dynamics and Control, 36 (8), 1121–1141.
Bech, M. and Atalay, E. (2010). The Topology of the Federal Funds Market. Physica A: Statistical Mechanics and its Appli-cations, 389 (22), 5223–5246.
Blume, L., Easley, D., Kleinberg, J., Kleinberg, R. and Tardos, É. (2011). Which Networks are Least Susceptible to Cas-cading Failures? In 2011 IEEE 52nd Annual Symposium on Foundations of Computer Science, IEEE, pp. 393–402.
Bonfiglioli, A. (2008). Financial Integration, Productivity and Capital Accumulation. Journal of International Economics,76 (2), 337–355.
Boss, M., Elsinger, H., Summer, M. and Thurner, S. (2004). Network Topology of the Interbank Market. QuantitativeFinance, 4 (6), 677–684.
Brandes, U. (2001). A faster algorithm for betweenness centrality. Journal of Mathematical Sociology, 25 (2), 163–177.
Caballero, J. (2012). Do Surges in International Capital Flows Influence the Likelihood of Banking Crises. IDB Working PaperNo. 305, Inter-American Development Bank, Washington DC.
—, Candelaria, C. and Hale, G. (2009). Bank Relationships and the Depth of the Current Economic Crisis. FRBSFEconomic Letter, 2009, 38.
Cai, J. (2010). Competition or Collaboration? The Reciprocity Effect in Loan Syndication. Tech. rep., FRB of Cleveland PolicyDiscussion Paper No. 09-09R.
—, Saunders, A. and Steffen, S. (2012). Syndication, Interconnectedness, and Systemic Risk. Manuscript. http://ssrn.com/abstract=1508642.
Calvo, G. (1998). Capital Flows and Capital-Market Crises: The Simple Economics of Sudden Stops. Journal of AppliedEconomics, 1 (1), 35–54.
Cameron, A. and Trivedi, P. (1998). Regression Analysis of Count Data. Cambridge Books.
Chang, R. and Velasco, A. (2001). A Model of Financial Crises in Emerging Markets. The Quarterly Journal of Economics,116 (2), 489–517.
Chinazzi, M., Fagiolo, G., Reyes, J. A. and Schiavo, S. (2013). Post-mortem examination of the international financialnetwork. Journal of Economic Dynamics and Control, 37 (8), 1692–1713.
Chinn, M. and Ito, H. (2008). A NewMeasure of Financial Openness. Journal of Comparative Policy Analysis: Research andPractice, 10 (3), 309–322.
Demirgüc-Kunt, A. and Detragiache, E. (2005). Cross-Country Empirical Studies of Systemic Bank Distress: A Survey.National Institute Economic Review, 192 (1), 68.
Edwards, S. (2007). Capital Controls, Sudden Stops, and Current Account Reversals.
Eichengreen, B. (2002). International Financial Crises: Is the Problem Growing? Jahrbuch für Wirtschaftsgeschichte, 1,89–104.
Fagiolo, G.,Reyes, J. and Schiavo, S. (2008). On the Topological Properties of theWorld TradeWeb: AWeightedNetworkAnalysis. Physica A: Statistical Mechanics and its Applications, 387 (15), 3868–3873.
Fruchterman, T. M. and Reingold, E. M. (1991). Graph drawing by force-directed placement. Software: Practice andexperience, 21 (11), 1129–1164.
Furceri, D., Guichard, S. and Rusticelli, E. (2012). Episodes of Large Capital Inflows, Banking and Currency Crises,and Sudden Stops. International Finance, 15 (1), 1–35.
Gadanecz, B. and vonKleist, K. (2002). Do Syndicated Credits Anticipate BIS Consolidated Banking Data? BIS QuarterlyReview, pp. 65–74.
36
Page 38 of 49
Accep
ted
Man
uscr
ipt
Gatev, E. and Strahan, P. (2009). Liquidity Risk and Syndicate Structure. Journal of Financial Economics, 93 (3), 490–504.
Georg, C.-P. (2013). The effect of the interbank network structure on contagion and common shocks. Journal of Banking& Finance.
Godlewski, C., Sanditov, B. and Burger-Helmchen, T. (2012). Bank Lending Networks, Experience, Reputation, andBorrowing Costs: Empirical Evidence from the French Syndicated Lending Market. Journal of Business Finance &Accounting, 39 (1-2), 113–140.
Gourinchas, P. andObstfeld,M. (2012). Stories of the Twentieth Century for the Twenty-First.American Economic Journal:Macroeconomics, 4 (1), 226–65.
Haldane, A. (2009). Rethinking the Financial Network. Speech delivered at the Financial Student Association, Ams-terdam, April. http://http://www.bankofengland.co.uk/publications/Documents/speeches/2009/speech386.pdf.
Hale, G. (2012). Bank Relationships, Business Cycles, and Financial Crises. Journal of International Economics, (In press).
—, Candelaria, C., Caballero, J. and Borisov, S. (2011). Global Banking Network and Cross-Border Capital Flows.Mimeo. Federal Reserve Bank of San Francisco, and University of California, Santa Cruz.
—,—,— and— (2013). Bank Linkages and International Trade. IDB Working Paper No. 445, Inter-American DevelopmentBank, Washington DC.
Hardin, J. and Hilbe, J. (2007). Generalized Linear Models and Extensions. Stata Corp.
Hattori, M. and Suda, Y. (2007). Developments in a Cross-Border Bank Exposure Network. CGFS Papers Chapters, Bank forInternational Settlements.
Inaoka, H., Ninomiya, T., Taniguchi, K., Shimizu, T. and Takayasu, H. (2004). Fractal Network Derived from BankingTransaction–An Analysis of Network Structures Formed by Financial Institutions. Working Papers, No. 4, Bank of Japan.
Joyce, J. (2010). Financial Globalization and Banking Crises In Emerging Markets. Open Economies Review, pp. 1–21.
Kali, R. and Reyes, J. (2007). The Architecture of Globalization: A Network Approach to International Economic Inte-gration. Journal of International Business Studies, 38 (4), 595–620.
— and — (2010). Financial Contagion on the International Trade Network. Economic Inquiry, 48 (4), 1072–1101.
Kleinberg, J. M. (1999). Authoritative sources in a hyperlinked environment. Journal of the ACM (JACM), 46 (5), 604–632.
Kubelec, C. and Sá, F. (2012). The Geographical Composition of National External Balance Sheets: 1980-2005. Interna-tional Journal of Central Banking, 8 (2), 143–189.
Laeven, L. and Valencia, F. (2013). Systemic Banking Crises Database. IMF Econ Rev, 61 (2), 225–270.
Minoiu, C. and Reyes, J. A. (2013). A network analysis of global banking: 1978–2010. Journal of Financial Stability.
Newman, M. (2003). The structure and function of complex networks. SIAM review, 45 (2), 167–256.
— (2010). Networks: An Introduction. Oxford University Press.
Pichler, P. and Wilhelm, W. (2002). A Theory of the Syndicate: Form Follows Function. The Journal of Finance, 56 (6),2237–2264.
Reinhart, C. and Reinhart, V. (2009). Capital Flow Bonanzas: An Encompassing View of the Past and Present. In NBERMacroeconomics Annual, University of Chicago Press.
Rose, A. and Spiegel,M. (2009). International Financial Remoteness andMacroeconomic Volatility. Journal of DevelopmentEconomics, 89 (2), 250–257.
Schweitzer, F., Fagiolo, G., Sornette, D., Vega-Redondo, F. and White, D. R. (2009). Economic networks: What do weknow and what do we need to know? Advances in Complex Systems, 12 (04n05), 407–422.
37
Page 39 of 49
Accep
ted
Man
uscr
ipt
von Peter, G. (2007). International Banking Centres: a Network Perspective. BIS Quarterly Review, pp. 33–45.
Winkelmann, R. (2008). Econometric Analysis of Count Data. Springer Verlag.
Zawadowski, A. (2013). Entangled financial systems. Review of Financial Studies, 26 (5), 1291–1323.
38
Page 40 of 49
Accep
ted
Man
uscr
ipt
8. Appendix (Not to Print)
Figure A-1: Histograms of Degree Distributions. Sample 1980-2007
020
0040
00F
requ
ency
0 50 100 150
Indegree0
2000
4000
Fre
quen
cy
0 100 200
Outdegree
This figure presents the distributions of indegree and outdegreemeasures for the full samplenetwork (1980-2007)
39
Page 41 of 49
Accep
ted
Man
uscr
ipt
Figu
reA-2:V
isua
lizationof
theNetwork.
Top10
0Bo
rrow
ersan
dLe
nders.
Sample19
80-200
7
DA
IWA
EU
RO
PE L
TD
NIB
C B
ank-
NLD
Land
esba
nk-G
BR
SNS
Rea
al -N
LD
MB
NA
Cor
p-U
SA
DN
B B
ank-
NO
R
Deu
tsch
e A
usgl
eich
-DEU
SEB
-SW
EJP
Mor
gan-
IRL
HSB
C-N
LD
FCE
Ban
k-G
BR
RB
C-C
AN
US
Ban
corp
-USA
Rab
oban
k-IR
L
Wes
thyp
Fin
-DEU
FHL
Ban
ks-U
SA
Cha
se-U
SA
Blo
ck F
G-U
SA
Deu
tsch
e B
ank-
DEU
Cre
dit S
uiss
e-C
HE
Deu
tsch
e B
ank-
CH
E
Deu
tsch
e B
ank-
GB
R
Sven
ska
Han
dels
-SW
E
VTB
Cap
ital-R
US
Wes
tLB
-DEU
3CIF
-FR
A
UB
S-G
BR
Nat
ionw
ide
BS-
GB
R
Man
uf. H
anov
er-G
BR
Com
pagn
ie B
anca
ire-F
RA
HSB
C-G
BR
Euro
Cle
ar-B
EL
Gol
dman
Sac
hs-D
EU
DZ
Ban
k-D
EU
Mor
gan
Stan
ley-
USA
Com
mer
zban
k -IR
L
Sum
itom
o M
itsui
-GB
R
Sum
itom
o M
itsui
-JPN Citi
-GB
R
Inte
sa S
anpa
olo-
ITA
Bay
ernL
B-D
EU
FIH
-DN
K
Deu
tsch
e B
ank-
LUX
Gol
dman
Sac
hs-G
BR
Ban
k O
ne-U
SA
Bar
clay
s-G
BR
CIT
Gro
up -U
SA
AN
Z-H
KG
Lehm
an B
roth
-GB
R
Cre
dit A
gric
ole
-FR
A
Stan
dard
Cha
rter
ed B
ank-
GB
R
SG-A
US
Ban
k of
Am
eric
a -G
BR
Ced
el B
ank-
LUX
Kid
der P
eabo
dy-U
SA
HVB
-Uni
cred
it-D
EU
Hyp
o Ve
rein
sfin
-DEU
BN
P Pa
ribas
-FR
A
Dep
fa-D
EU
Nat
ixis
-FR
A
Firs
t Uni
on B
ank-
USA
BN
Y M
ello
n-U
SA
Rab
oban
k-A
US
Bra
dfor
d-B
ingl
ey-G
BR
Nor
ther
n R
ock-
GB
R
Scot
land
Inte
r Fin
-GB
R
Wac
hovi
a-U
SA
Isba
nk-T
UR
Mer
rill L
ynch
-DEU
AN
Z-A
US
Ban
k of
Am
eric
a -U
SA
Bea
r Ste
arns
-USA
Mits
ubis
hi U
FJ -J
PN
Ban
k of
Chi
na -C
HN
Rai
ffeis
en-A
UT
Turk
iye
Gar
anti-
TUR
LBB
W-D
EU
Han
a B
ank-
KO
R
JPM
orga
n-U
SA
AB
N A
MR
O-N
LD
Cre
dit S
uiss
e-U
SA
Mer
rill L
ynch
-USA
HSB
C-H
KG
NA
B-A
US
Ban
k of
Am
eric
a -H
KG
Sant
ande
r-G
BR
Deu
tsch
e B
ank-
HK
G
Wes
tpac
-AU
S
ING
-NLD
Nom
ura-
DEU
Cre
dit S
uiss
e-G
BR
CB
A-A
US
JPM
orga
n-G
BR
BA
-Uni
Cre
dit-A
UT
JP M
orga
n-D
EU
Hel
aba
Fina
nce-
DEU
UB
S-D
EU
UB
S-C
HE
Nom
ura-
GB
R
Rab
oban
k-N
LD
Mer
rill L
ynch
-GB
R
Ban
k of
Irel
and-
IRL
Miz
uho-
GB
R
Miz
uho-
JPN
Lehm
an B
roth
-USA
Cap
ital O
ne-U
SA
RB
S-G
BR
Ban
que
PSA
-FR
A
Mor
gan
Stan
ley-
DEU
Com
mer
zban
k -D
EU
Firs
t Nat
Ban
k-U
SA
HFC
Ban
k-G
BR
Mor
gan
Stan
ley-
LUX
Ban
k of
Sco
tland
-GB
R
Lloy
ds -G
BR
Citi
-USA SG
-FR
A
Mor
gan
Stan
ley-
GB
R
USA
& Ca
nada
UK
W
este
rn E
urop
e Ja
pan
East
Asia
& P
acifi
c Ea
ster
n E
urop
e & C
entra
l Asia
A
B Li
nk fr
om B
to A
A B
Link
from
A to
B
This
figurepresen
tsavisu
alizationof
thene
tworkcompo
sedof
thetop10
0lend
ersan
dbo
rrow
ersan
dallthe
links
betw
eenthem
inthesample19
80-200
7.Th
evisu
alizationpresen
tsthelin
kage
sbe
tweenatotalo
fN=
118ba
nks(nod
es);atotalo
fE=
513lin
ks(edg
es).
Link
saredirected
,with
conv
exed
gesrepresen
ting
lend
ingfrom
node
ito
node
j.Th
esize
oftheno
desisprop
ortio
naltothede
gree
ofeach
node
andthesize
ofthelin
ksisprop
ortio
naltothesize
ofthelend
ingor
borrow
ing.
Thecoloro
fthe
node
siden
tifies
bank
sge
ograph
ically
with
sixdistinct
geog
raph
icregion
s,as
describ
edin
thelege
nd.T
helin
ksarecoloredaccording
tothecoloro
fthe
target.T
hegrap
hpresen
tsthena
mes
oftheba
nks’pa
rents,follo
wed
byathreelette
racron
ymfort
hecoun
trywhe
retheba
nkislocated.
Fora
nexplan
ationof
theprocesso
fassigning
names
seefootno
te23
.
40
Page 42 of 49
Accep
ted
Man
uscr
ipt
Figure A-3: Histogram: Number of Banking Crises
0.1
.2.3
.4.5
Den
sity
0 1 2 3 4Number of banking crises
This figure shows a histogram with the distribution of systemic bankingcrises in the period 1980-2007 for the list of countries in Table A-1.
Figure A-4: Estimated Residuals (Deviance)
0.2
.4.6
.8D
ensi
ty
−2 −1 0 1 2Deviance residual Kernel
Kernel density estimate
Normal density
This figure compares the estimated residuals, or deviance, of the modelin specification 5 of Table 4 with a normal distribution. The plot uses theEpanechnikov kernel for the estimation of the kernel density with a band-width parameter of 0.2092.
41
Page 43 of 49
Accep
ted
Man
uscr
ipt
Figu
reA-5:C
orrelatio
nof
NetworkStatistic
sat
theCou
ntry
Leve
l.NetworkStatistic
sCom
putedin
FullSa
mple19
80-200
7
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
(11)
(12)
(13)
(14)
(15)
(16)
(1)
Betw
eenn
ess
1(2
)Hu
b Ce
ntra
lity
0.22
1(3
)W
eigh
ted
Clus
t. Co
eff.
-0.0
1-0
.01
1(4
)Cl
ust.
Coef
f.-0
.01
0.00
0.99
1(5
)Au
thor
ity0.
330.
44-0
.03
-0.0
21
(6)
Inde
gree
0.05
-0.1
20.
020.
02-0
.10
1(7
)O
utde
gree
0.53
0.49
0.03
0.02
0.37
-0.1
11
(8)
Degr
ee0.
380.
210.
040.
030.
150.
780.
541
(9)
Betw
eenn
ess (
wei
ghte
d)0.
620.
37-0
.01
-0.0
10.
44-0
.06
0.36
0.18
1(1
0)Hu
b Ce
ntra
lity
(wei
ghte
d)0.
040.
86-0
.01
-0.0
10.
21-0
.07
0.26
0.10
0.11
1(1
1)W
eigh
ted
Clus
t. Co
eff.
(wei
ghte
d)0.
090.
150.
320.
270.
17-0
.07
0.14
0.03
0.14
0.08
1(1
2)Cl
ust.
Coef
f. (w
eigh
ted)
0.12
0.21
0.38
0.35
0.25
-0.0
70.
170.
050.
200.
120.
991
(13)
Auth
ority
(wei
ghte
d)0.
580.
48-0
.04
-0.0
30.
78-0
.13
0.42
0.15
0.68
0.23
0.15
0.21
1(1
4)Bo
rrow
ing
0.44
0.36
0.06
0.04
0.65
0.14
0.46
0.41
0.43
0.15
0.59
0.63
0.69
1(1
5)Le
ndin
g0.
370.
890.
000.
000.
51-0
.16
0.72
0.32
0.49
0.64
0.17
0.23
0.55
0.47
1(1
6)Bo
rrow
ing+
Lend
ing
0.47
0.71
0.04
0.03
0.68
0.00
0.68
0.43
0.53
0.45
0.46
0.52
0.73
0.87
0.84
1
Thisfig
uresh
owspa
irwisecorrelations
ofthecoun
try-leve
lnetworkstatistic
sus
edas
prox
iesford
efacto
finan
cial
integrationin
theregression
analysisof
Section6.
Thene
tworkstatistic
sarecompu
tedfrom
bank
-leve
ldataan
dthen
aggreg
ated
toob
tain
coun
try-leve
lvariables
(see
explan
ationin
Section2on
how
coun
tryaverag
esareob
tained
).Asa
visu
alaid,
thefig
uresh
owsh
ighe
rcorrelatio
nva
lues
inada
rker,m
oreintens
ecolor.
42
Page 44 of 49
Accep
ted
Man
uscr
ipt
Figu
reA-6:C
orrelatio
nof
NetworkStatistic
sat
theCou
ntry
Leve
lAcros
sDiff
eren
tSam
ples
(1)
(2)
(3)
(4)
(5)
(6)
(1)
(2)
(3)
(4)
(5)
(6)
(1)
Net
w. S
tat.
80-0
71
(1)
Net
w. S
tat.
80-0
71
(2)
Net
w. S
tat.
95_0
70.
941
(2)
Net
w. S
tat.
95_0
70.
631
(3)
Net
w. S
tat.
01_0
70.
770.
891
(3)
Net
w. S
tat.
01_0
70.
621.
001
(4)
W. N
etw
. Sta
t. 80
_07
1(4
)W
. Net
w. S
tat.
80_0
71
(5)
W. N
etw
. Sta
t. 95
_07
0.92
1(5
)W
. Net
w. S
tat.
95_0
70.
171
(6)
W. N
etw
. Sta
t. 01
_07
0.76
0.87
1(6
)W
. Net
w. S
tat.
01_0
70.
161.
001
(1)
(2)
(3)
(4)
(5)
(6)
(1)
(2)
(3)
(4)
(5)
(6)
(1)
Net
w. S
tat.
80-0
71
(1)
Net
w. S
tat.
80-0
71
(2)
Net
w. S
tat.
95_0
70.
851
(2)
Net
w. S
tat.
95_0
7-0
.25
1(3
)N
etw
. Sta
t. 01
_07
0.43
0.52
1(3
)N
etw
. Sta
t. 01
_07
-0.2
80.
791
(4)
W. N
etw
. Sta
t. 80
_07
1(4
)W
. Net
w. S
tat.
80_0
71
(5)
W. N
etw
. Sta
t. 95
_07
0.52
1(5
)W
. Net
w. S
tat.
95_0
70.
621
(6)
W. N
etw
. Sta
t. 01
_07
0.44
0.74
1(6
)W
. Net
w. S
tat.
01_0
70.
070.
101
(1)
(2)
(3)
(4)
(5)
(6)
(1)
(2)
(3)
(4)
(5)
(6)
(1)
Net
w. S
tat.
80-0
71
(1)
Net
w. S
tat.
80-0
71
(2)
Net
w. S
tat.
95_0
70.
201
(2)
Net
w. S
tat.
95_0
70.
921
(3)
Net
w. S
tat.
01_0
70.
120.
991
(3)
Net
w. S
tat.
01_0
70.
830.
951
(4)
W. N
etw
. Sta
t. 80
_07
1(4
)W
. Net
w. S
tat.
80_0
71
(5)
W. N
etw
. Sta
t. 95
_07
0.53
1(5
)W
. Net
w. S
tat.
95_0
70.
911
(6)
W. N
etw
. Sta
t. 01
_07
0.51
1.00
1(6
)W
. Net
w. S
tat.
01_0
70.
600.
821
(1)
(2)
(3)
(4)
(5)
(6)
(1)
(2)
(3)
(4)
(5)
(6)
(1)
Net
w. S
tat.
80-0
71
(1)
Net
w. S
tat.
80-0
71
(2)
Net
w. S
tat.
95_0
70.
911
(2)
Net
w. S
tat.
95_0
70.
971
(3)
Net
w. S
tat.
01_0
70.
830.
941
(3)
Net
w. S
tat.
01_0
70.
900.
971
(4)
W. N
etw
. Sta
t. 80
_07
1(4
)W
. Net
w. S
tat.
80_0
71
(5)
W. N
etw
. Sta
t. 95
_07
0.83
1(5
)W
. Net
w. S
tat.
95_0
70.
951
(6)
W. N
etw
. Sta
t. 01
_07
0.48
0.81
1(6
)W
. Net
w. S
tat.
01_0
70.
830.
911
Inde
gree
Out
degr
ee
Betw
eenn
ess
Hub
Cent
ralit
y
Auth
ority
Wei
ghte
d Cl
uste
ring
Coef
ficie
ntCl
uste
ring
Coef
ficie
nt
Degr
ee
This
figuresh
owscorrelations
across
diffe
rent
subsam
ples
ofthecoun
try-leve
lnetworkstatistic
sus
edas
prox
iesforde
factofin
ancial
integrationin
the
regression
analysis
ofSe
ction6.
Thefig
urecontains
eigh
thpa
nels,e
achsh
owingthecorrelationof
ane
tworkstatistic
(unw
eigh
tedan
dweigh
ted)
across
threediffe
rent
samples:19
80-200
7,19
95-200
7,an
d20
01-200
7.Th
ene
tworkstatistic
sarecompu
tedforeach
subsam
plefrom
bank
-leve
ldataan
dthen
aggreg
ated
toob
tain
coun
try-leve
lvariables
(see
explan
ationin
Section2on
how
coun
tryaverag
esareob
tained
).Asavisu
alaid,
thefig
uresh
owshigh
ercorrelationva
lues
inada
rker,m
oreintens
ecolor.
43
Page 45 of 49
Accep
ted
Man
uscr
ipt
Figu
reA-7:C
orrelatio
nof
NetworkStatistic
sat
theBa
nkLe
velA
cros
sDiff
eren
tSam
ples
(1)
(2)
(3)
(4)
(5)
(6)
(1)
(2)
(3)
(4)
(5)
(6)
(1)
Net
w. S
tat.
80-0
71
(1)
Net
w. S
tat.
80-0
71
(2)
Net
w. S
tat.
95_0
70.
801
(2)
Net
w. S
tat.
95_0
70.
511
(3)
Net
w. S
tat.
01_0
70.
660.
901
(3)
Net
w. S
tat.
01_0
70.
511.
001
(4)
W. N
etw
. Sta
t. 80
_07
1(4
)W
. Net
w. S
tat.
80_0
71
(5)
W. N
etw
. Sta
t. 95
_07
0.86
1(5
)W
. Net
w. S
tat.
95_0
70.
821
(6)
W. N
etw
. Sta
t. 01
_07
0.34
0.53
1(6
)W
. Net
w. S
tat.
01_0
70.
821.
001
(1)
(2)
(3)
(4)
(5)
(6)
(1)
(2)
(3)
(4)
(5)
(6)
(1)
Net
w. S
tat.
80-0
71
(1)
Net
w. S
tat.
80-0
71
(2)
Net
w. S
tat.
95_0
70.
911
(2)
Net
w. S
tat.
95_0
7-0
.21
1(3
)N
etw
. Sta
t. 01
_07
0.79
0.87
1(3
)N
etw
. Sta
t. 01
_07
-0.2
00.
831
(4)
W. N
etw
. Sta
t. 80
_07
1(4
)W
. Net
w. S
tat.
80_0
71
(5)
W. N
etw
. Sta
t. 95
_07
0.51
1(5
)W
. Net
w. S
tat.
95_0
7-0
.01
1(6
)W
. Net
w. S
tat.
01_0
70.
370.
691
(6)
W. N
etw
. Sta
t. 01
_07
0.00
0.75
1
(1)
(2)
(3)
(4)
(5)
(6)
(1)
(2)
(3)
(4)
(5)
(6)
(1)
Net
w. S
tat.
80-0
71
(1)
Net
w. S
tat.
80-0
71
(2)
Net
w. S
tat.
95_0
70.
081
(2)
Net
w. S
tat.
95_0
70.
971
(3)
Net
w. S
tat.
01_0
70.
070.
991
(3)
Net
w. S
tat.
01_0
70.
870.
951
(4)
W. N
etw
. Sta
t. 80
_07
1(4
)W
. Net
w. S
tat.
80_0
71
(5)
W. N
etw
. Sta
t. 95
_07
0.07
1(5
)W
. Net
w. S
tat.
95_0
70.
921
(6)
W. N
etw
. Sta
t. 01
_07
0.06
1.00
1(6
)W
. Net
w. S
tat.
01_0
70.
650.
811
(1)
(2)
(3)
(4)
(5)
(6)
(1)
(2)
(3)
(4)
(5)
(6)
(1)
Net
w. S
tat.
80-0
71
(1)
Net
w. S
tat.
80-0
71
(2)
Net
w. S
tat.
95_0
70.
971
(2)
Net
w. S
tat.
95_0
70.
971
(3)
Net
w. S
tat.
01_0
70.
890.
951
(3)
Net
w. S
tat.
01_0
70.
880.
951
(4)
W. N
etw
. Sta
t. 80
_07
1(4
)W
. Net
w. S
tat.
80_0
71
(5)
W. N
etw
. Sta
t. 95
_07
0.89
1(5
)W
. Net
w. S
tat.
95_0
70.
921
(6)
W. N
etw
. Sta
t. 01
_07
0.73
0.86
1(6
)W
. Net
w. S
tat.
01_0
70.
610.
781
Inde
gree
Out
degr
ee
Betw
eenn
ess
Hub
Cent
ralit
y
Auth
ority
Wei
ghte
d Cl
uste
ring
Coef
ficie
ntCl
uste
ring
Coef
ficie
nt
Degr
ee
Thisfig
ures
howsc
orrelatio
nsacross
diffe
rent
subsam
ples
oftheb
ank-leve
lnetworkstatistic
scom
putedfrom
thes
yndicatedloan
s.Th
efigu
recontains
eigh
tpa
nels,eachsh
owingthecorrelationof
ane
tworkstatistic
(unw
eigh
tedan
dweigh
ted)
across
threediffe
rent
samples:1
980-20
07,1
995-20
07,a
nd20
01-200
7.Th
ene
tworkstatistic
sare
compu
tedfore
achsu
bsam
plefrom
bank
-leve
ldata.
Asa
visu
alaid,
thefig
uresh
owsh
ighe
rcorrelatio
nva
lues
inada
rker,m
ore
intens
ecolor.
44
Page 46 of 49
Accep
ted
Man
uscr
ipt
Table A-1: Sample. Number of Systemic Banking Crises during 1980-2007 in Parentheses
High income countriesAustralia (0) Finland (1) Korea (1) Qatar (0)Austria (0) France (0) Kuwait (1) Singapore (0)Bahamas (0) Germany (0) Luxembourg (0) Slovenia (1)Bahrain (0) Greece (0) Macao (0) Spain (0)Belgium (0) Hong Kong (0) Malta (0) Sweden (1)Bermuda (0) Iceland (0) Netherlands (0) Switzerland (0)Brunei Darussalam (0) Ireland (0) Netherlands Antilles (0) United Arab Emirates (0)Canada (0) Israel (0) New Zealand (0) United Kingdom (0)Cyprus (0) Italy (0) Norway (1) United States (1)Denmark (0) Japan (1) Portugal (0)
Developing countriesAlgeria (1) Egypt (1) Lithuania (1) Serbia (0)Angola (0) El Salvador (1) Macedonia (1) Slovakia (1)Argentina (4) Estonia (1) Malaysia (1) South Africa (0)Armenia (1) Ethiopia (0) Mauritius (0) Sri Lanka (1)Azerbaijan (1) Georgia (1) Mexico (2) Tajikistan (0)Belarus (1) Ghana (1) Moldova (0) Tanzania (1)Bolivia (2) Guatemala (0) Mongolia (0) Thailand (2)Bosnia and Herzegovina (1) Honduras (0) Morocco (1) Trinidad and Tobago (0)Brazil (2) Hungary (1) Namibia (0) Tunisia (1)Bulgaria (1) India (1) Nigeria (1) Turkey (2)Cayman Islands (0) Indonesia (1) Oman (0) Ukraine (1)Chile (1) Iran (0) Pakistan (0) Uruguay (2)China (1) Iraq (0) Panama (1) Uzbekistan (0)Colombia (2) Jamaica (1) Peru (1) Venezuela (1)Costa Rica (2) Jordan (1) Philippines (2) Yemen (1)Croatia (0) Kazakhstan (0) Poland (1) Zambia (1)Cuba (0) Kenya (2) Romania (1) Zimbabwe (1)Czech Republic (1) Latvia (1) Russia (1)Dominican Republic (1) Lebanon (1) Rwanda (0)Ecuador (2) Libya (0) Saudi Arabia (0)
This table shows the sample of countries and banking crises used in the paper. The source is ?. The total number of crisesfor each country in the period 1980-2007 is shown in parenthesis. A banking crisis is defined as a systemic banking crisiswhentwo conditions are met: (i) significant signs of financial distress in the banking system (as indicated by bank runs, losses inthe banking system, and/or bank liquidations); and (ii) significant banking policy intervention measures were undertakenin response to losses in the banking system. The year in which a systemic banking crisis starts is identified by these twoconditions and when at least three out of the following five policy interventions have been used:
• Extensive liquidity support (ratio of central bank claims on the financial sector to deposits and foreign liabilities exceeds5% and more than doubles relative to its pre-crisis level);• Large bank restructuring costs (at least 3% of GDP, excluding asset purchases and direct liquidity assistance from thetreasury);• Bank nationalizations (treasury or central bank asset purchases exceeding 5% of GDP);• Significant asset purchases;• Significant guarantees put in place (excluding increases in the level of deposit insurance coverage); or• Deposit freezes and bank holidays.
When a country has faced financial distress but fewer than three of these measures have been used, the event is classified as acrisis if one of the following two conditions has been met: (a) Country’s banking system exhibits significant losses resulting ina share of nonperforming loans above 20% or bank closures of at least 20% of banking system assets; or (b) Fiscal restructuringcosts of the banking sector exceed 5% of GDP.
45
Page 47 of 49
Accep
ted
Man
uscr
ipt
Table A-2: Data Description
Variable Definition Source
Banking crises Discrete variable equal to the total number of systemic banking crises in1980-2007. Detailed definition of crises in Section 3.2 and TableA-1. ?
Trade Open-ness
Total trade (sum of exports and imports of goods and services) as a percent-age of GDP. Variable NE.TRD.GNFS.ZS in WDI.
WDI database,World Bank
Political RiskPolitical risk index designed by the Political Risk Group, and known as In-ternational Country Risk Guide (ICRG). The index goes from 0 to100 and isdecreasing in the level of risk.
Political Risk Group.International Coun-try Risk Guide(ICRG)
Polity2Combined polity score (index) of strength of democratic institutions de-signed by Polity IV Project. The index is discrete and ranges from -10 to+10 and is increasing in the strength/quality of democratic institutions.
Polity IV Project
Domesticcredit to pri-vate sector
Domestic credit provided by the banking sector as percentage of GDP.Includes all credit to various sectors on a gross basis, with the ex-ception of credit to the central government, which is net. VariableFS.AST.DOMS.GD.ZS in WDI database.
WDI database,World Bank.
Current Ac-count Balance
Current account balance as percentage of GDP. VariableBN.CAB.XOKA.GD.ZS in WDI database.
WDI database,World Bank
Inflation Annual percentage change in consumer price index. VariableFP.CPI.TOTL.ZG in WDI database.
WDI database,World Bank
Capital Ac-count Open-ness (KA open)
Index that measures the extent of openness in capital account transactions(it tries to capture the extent and intensity of capital controls). It is builtbased on the binary dummy variables that codify the tabulation of restric-tions on cross-border financial transactions reported in the IMF’s AnnualReport on Exchange Arrangements and Exchange Restrictions (AREAER).The index is continuous and increasing in the openness of the capital ac-count transactions. For the available sample it ranges in the interval [-1.8,2.5].
Chinn and Ito (2008)
Banking super-vision Index
Banking supervision index. It is increasing in the level of regulation of thebanking system. The index is built using four dimensions: (i) adoption ofBasel standards on capital adequacy, (ii) independence of banking supervi-sory agency from executive’s influence, (iii) existence and effectiveness ofon-site and off-site examinations by the supervisory agency, and (iv) spec-trum of financial institutions covered by the supervisory agency. Index goesfrom 0 to 6 and is increasing in the level of regulation (however, the highestindex awarded in the database is 3).
Abiad et al. (2010)
Financialreform Index
Index is increasing in the level of financial reform achieved. The index isbuilt using seven dimensions: (i) credit controls and excessively high re-serve requirements, (ii) interest rate controls, (iii) entry barriers, (iv) stateownership of the banking sector, (v) capital account restrictions, (vi) bank-ing supervision, and (vii) security markets policy. Index goes from 0 to 21.To use the index in the non-parametric analysis the scores were adjusted tofit 3 categories as follows: value 1 for index values from 0 to 6; value 2 forindex values from 7 to 14; and value 3 for index values from 15 to 21.
Abiad et al. (2010)
Capital ac-count transac-tions index
Index is increasing in the liberalization of the CA transactions. The indexis built using three dimensions: (i) existence of a unified exchange rate sys-tem, (ii) extent of restrictions on capital inflows, (iii) extent of restrictions oncapital inflows. Index goes from 0 to 3.
Abiad et al. (2010)
IncomeDummy
Dummy variable that takes value 1 if country is high income country. In-come group is that ofWorld Bank. High income countries include all OECDcountries, plus Hong Kong, Israel, Kuwait and Slovenia. However, someOECD members are classified as developing countries: Chile, Czech Re-public, Hungary, Korea, Mexico, Poland, Slovak Republic, and Turkey.
World Bank, OECD
GNI per capita GNI per capita, Atlas method (current US$)Variable NY.GNP.PCAP.CD inWDI.
WDI database,World Bank
46
Page 48 of 49
Accep
ted
Man
uscr
ipt
Table A-3: Summary Statistics. Country-Level Variables. Sample 1980-2007
mean sd min max sum countNumber of banking crises 0.647 0.725 0.000 4.000 75.000 116Betweenness 0.083 0.200 0.000 1.000 9.610 116Betweenness (weighthed) 0.033 0.129 0.000 1.000 3.814 116Weighted Clust. Coeff. 0.040 0.102 0.000 1.000 4.633 116Weighted Clust. Coeff. (weighted) 0.028 0.099 0.000 1.000 3.264 116Clust. Coeff. 0.039 0.100 0.000 1.000 4.490 116Clust. Coeff. (weighted) 0.037 0.106 0.000 1.000 4.246 116Hub 0.023 0.117 0.000 1.000 2.633 116Hub (weighted) 0.010 0.094 0.000 1.000 1.214 116Authority 0.078 0.196 0.000 1.000 9.095 116Authority (weighted) 0.057 0.176 0.000 1.000 6.612 116Degree 4.842 3.729 1.000 19.318 561.668 116Borrowing+Lending 0.360 0.558 0.002 3.343 41.817 116Indegree 3.559 3.166 0.571 18.909 412.819 116Borrowing 0.260 0.341 0.002 1.755 30.159 116Outdegree 1.283 2.347 0.000 11.669 148.848 116Lending 0.101 0.309 0.000 2.001 11.658 116Sum Degree 0.275 0.654 0.001 4.120 31.926 116Sum Borrowing+Lending 0.041 0.167 0.000 1.374 4.715 116Sum Indegree 0.137 0.270 0.001 1.351 15.915 116Sum Borrowing 0.020 0.067 0.000 0.571 2.355 116Sum Outdegree 0.138 0.463 0.000 3.250 16.011 116Sum Lending 0.020 0.109 0.000 1.040 2.360 116Trade openness 0.838 0.521 0.198 4.117 92.990 111Political risk 66.719 12.588 32.904 92.617 6938.801 104Domestic credit 0.649 0.415 0.081 2.667 71.390 110Current account balance -1.127 7.219 -17.018 36.235 -121.736 108Inflation 0.549 1.244 0.006 6.794 59.346 108
47
Page 49 of 49
Accep
ted
Man
uscr
iptTableA-4: RobustnessCheckSub-Samples 1995-2007 and 2001-2007. Model Including
Centrality and Degree Network Statistics for Average Bank
(3) (4) (5) (3) (4) (5)1995-2007 Network 2001-2007 Network
Betweenness -2.603** -3.959** -3.967*** -1.295 -1.715 -2.122*[1.175] [1.596] [1.382] [0.939] [1.171] [1.120]
Weighted Clust. Coeff. (weighted) -0.339 -0.428 -0.384 -0.633[0.495] [0.534] [0.489] [0.633]
Lending -0.038 -0.774 0.050 -1.662[0.784] [0.889] [1.367] [2.208]
Sum Borrowing 8.580*** 7.137*** 24.496*** 18.645***[1.886] [1.942] [9.148] [4.830]
Baseline Controls Yes Yes Yes Yes Yes YesDeviance 26.936 24.129 24.783 25.969 23.224 23.987Deviance_df 0.442 0.402 0.387 0.499 0.455 0.436Pearson 23.901 18.710 20.536 23.147 18.819 21.034Pearson_df 0.392 0.312 0.321 0.445 0.369 0.382BIC -234.782 -233.298 -249.806 -190.293 -188.880 -204.752Obs 73 73 73 64 64 64
This table presents summarized results of multivariate regressions of the count data model of equa-tion 1 using proxies for de facto financial integration computed using data restricted to a sample period.The first three columns shows the results of regressions in which the proxies for de facto financial inte-gration based on network statistics are computed using data for the period 1995-2007. The last threecolumns show results for network statistics computed with data restricted to 2001-2007. In all regres-sions the dependent variable is the number of systemic banking crises in the period 1980-2007. Theconstruction of the network statistics is explained in Section 2. All regressions include a set of baselinecontrols (not shown), including trade openness, political risk, domestic credit as percentage of GDP,current account balance, inflation, banking supervision, capital account openness, and an indicator forhigh-income OECD country. All these controls are country averages during the period 1980-2007. Theregressions are estimated using a Generalized Linear Model (GLM) approach in which the estimationassumes a Poisson distribution for the conditional mean of the dependent variable. The estimation isimplemented using a pseudo-maximum likelihood estimator (PMLE) and robust standard errors. Thetable shows at the bottom statistics of the fit of the model, including the estimated deviance, the Pearsondeviance, the dispersion of these two measures (i.e., the variables scaled by the degrees of freedom),and the Bayesian Information Criterion. ∗ indicates significance at 10 percent level, ∗∗ significance at 5percent level, and ∗ ∗ ∗ significance at 1 percent level.
48