Sharing Private Information with
Customers: Strategic Default and
Lender Learning
Gerardo Pérez Cavazos
Working Paper 16-033
Sharing Private Information with
Customers: Strategic Default and Lender
Learning
Gerardo Pérez Cavazos
Harvard Business School
Working Paper 16-033
Copyright © 2015 by Gerardo Pérez Cavazos
Working papers are in draft form. This working paper is distributed for purposes of comment and discussion only. It may
not be reproduced without permission of the copyright holder. Copies of working papers are available from the author.
Sharing private information with customers:
Strategic default and lender learning
Gerardo Pérez Cavazos∗
The University of Chicago Booth School of Business
April 24, 2015
Abstract
I use a unique data set of loans to small business owners to examine whether lenders
face negative externalities when they share private information with borrowers. When
lenders grant debt forgiveness to borrowers, borrowers communicate that information
to other borrowers, who are then more likely to strategically default on their own
obligations. This strategic default contagion is economically large. When the lender
doubles debt forgiveness, contagion causes the default rate to increase by 10.9% on
average. Using an exogenous shock to the lender’s forgiveness policy, I further show
that as the lender learns about the extent of borrower communication it tightens its
debt forgiveness and origination policies to reduce information spillovers and mitigate
the default contagion. Collectively, these results provide new evidence on the strategic
interactions between a firm and its customers in a dynamic information environment.
JEL No.: D10, D83, G21, M41
Keywords: Information transmission, communication, strategic default contagion,
learning
∗
I am grateful to my dissertation committee: Douglas Skinner (Chair), Phil Berger, Christian Leuz,
and Haresh Sapra. I also thank Ray Ball, Andreas Bodmeier, Alejandro Cavazos, Hans Christensen, Merle
Erickson, John Gallemore, Joseph Gerakos, Mark Maffett, Mike Minnis, Adair Morse, Valeri Nikolaev, Antonio Picca, Andreya Silva, and Anastasia Zakolyukina for helpful comments. I received valuable feedback from
seminar participants at Duke Universiy, Harvard Business School, London Business School, Massachusetts
Intitute of Technology, New York University, Stanford University, University of California at Los Angeles,
University of Chicago, University of Illinois, University of Pennsylvania, and Washington University, as well
as conference participants at the 2014 AAA/Deloitte/J. Michael Cook Doctoral Consortium. Finally, I am
extremely grateful to Financiera Ayudamos for providing their data.
1
Introduction
A large accounting literature discusses the role of proprietary costs in firms’ disclosure decisions. Because observing proprietary costs is difficult, however, empirical evidence on the
magnitude of those costs is limited. I examine how information externalities in small business
lending affect the actions of borrowers and lenders, and in particular, how private information about lenders’ debt forgiveness policies is transmitted to other borrowers, a form of
information contagion that is costly to lenders. The debt forgiveness negotiation process
creates a dynamic information environment in which lenders and borrowers both share private information. Although the information exchange can help both parties achieve a more
efficient outcome, there is a risk of sharing too much information (Crawford and Sobel, 1982).
Theoretical work shows that economic agents can strategically use a firm’s proprietary information to their advantage (Verrecchia, 1983; Dye, 1986; Wagenhofer, 1990). My research
provides evidence on the costs that firms incur when they share private information with
customers.
I present two key findings. First, I show that lenders face strategic default contagion when
granting debt forgiveness to borrowers. This unintended consequence results from borrowers
communicating the private terms of their debt forgiveness agreements to other borrowers,
who are then more likely to strategically default on their own obligations. Second, I show
that lenders learn about the extent of borrower communication and alter their operating
policies to reduce information spillovers and mitigate the default contagion.
I use a detailed data set from Financiera Ayudamos (FA), a Mexican credit institution
that grants small business and consumer loans. FA selectively grants debt forgiveness in
cases of default in an effort to reduce losses on delinquent loans by incentivizing defaulters
to continue repaying their loans.1 I use these debt forgiveness events to analyze whether
1
For example, consider a borrower who is four weeks late on his payments and is only able to pay half
of his deficiency. This borrower may not make any payments, because he knows he will remain in default.
By contrast, if the lender offers to forgive two payments, the borrower can commit to paying the remaining
two installments and become current on his loan.
1
FA incurs costs due to private information transmission between borrowers regarding its
forgiveness policy. The idea is that a borrower who is granted debt forgiveness learns private
information regarding the extent of the lender’s willingness to give a discount on the loan
repayment and can subsequently share this information with other borrowers. Financially
able borrowers can strategically use this information to their advantage to get a reduction
on their loan balance. Any information asymmetry regarding a borrower’s ability to pay
exacerbates this issue. If the lender is unable to clearly distinguish between borrowers who
can and cannot repay their loans, it is more likely that he will unknowingly offer debt
forgiveness to strategic defaulters.
A distinct feature of this setting is that lenders only share private information through
private contracts with select borrowers. Although these borrowers can communicate with
others, the extent of communication is unknown ex-ante. It is therefore unclear whether
the information externalities are substantial enough for the lender to alter its operating
decisions to reduce the amount of private information shared with borrowers. Alternatively,
borrowers may not share the details of this forgiveness event with others, because unlike a
publicly disclosed event such as foreclosure, a default and the subsequent renegotiation with
the lender are private.
To identify costs due to information externalities it is necessary to disentangle strategic
defaults from defaults that are the result of economic adversity. My empirical approach
is based on the simple premise that borrowers’ primary economic shocks occur at work,
whereas communication among borrowers occurs in the home neighborhood. I am able to
draw this important distinction because of the particularities of the Mexican setting. First,
many individuals commute over an hour to work. This allows me to separate the default
rate in the work area from that in the home neighborhood. A high default rate in the work
area indicates that borrowers have suffered shocks to income, whereas a low default rate
indicates stable economic conditions. If an individual defaults but does not suffer a shock to
his income, I then attribute the default to events within his home neighborhood.
2
Defaults that originate in the home neighborhood have two possible explanations: (i) an
event that causes a shock to expenses or (ii) strategic default due to communication between
borrowers. Communication is high within Mexican neighborhoods because they comprise
tight-knit groups of families and friends who have lived there most of their lives and interact
through weekly community activities (Keefe, 1984).2 Therefore, to determine whether the
default stems from communication about the lender’s forgiveness policy, I examine the forgiveness rate within the home neighborhood. A high forgiveness rate within a neighborhood
indicates more borrowers have experienced the debt forgiveness process. As these informed
borrowers communicate about the lender’s willingness to forgive, neighborhoods with a high
forgiveness rate will have many defaults. I then control for the contemporaneous default
rate, inflation, and natural disasters to rule out any shocks to the neighborhood’s expenses.
My first set of results measures the extent of strategic default in a neighborhood after the
lender grants debt forgiveness to a borrower. I find evidence of strategic default contagion
resulting from communication among borrowers regarding the lender’s forgiveness policy. If
FA doubles its forgiveness rate in a neighborhood, the default rate increases by 1.7% in the
following month. This default contagion is equivalent to a 10.9% increase in FA’s monthly
default rate. These results are robust and valid for a wide range of econometric tests. In
addition to controlling for economic shocks, I include various fixed effects and complement
the regression analysis with an instrumental variables (IV) approach.
My findings have three critical implications. First, firms not only release information
through disclosure, but also through their operating actions. In my setting these operational
decisions are private internal policies; however, firms also make public operational decisions,
such as expanding their operations or releasing new products, which likely reveal proprietary
information to outsiders. Second, firms incur significant informational costs when negotiating
with customers. Although theoretical models define customers as potential opponents of
2
Keefe (1984) finds that Mexican Americans families “have relatively large kin networks with high rates
of visiting and exchange. Even immigrant Mexicans, who have experienced disruption of their kin group due
to migration, surpass Anglos in the number of relatives living nearby and their frequency of visiting kin.”
3
a firm, the extant empirical literature has focused almost exclusively on competitors as
users of firms’ proprietary information. Third, the evidence I provide raises questions about
whether increasing renegotiations is an optimal strategy to attenuate the high foreclosure
rate. This issue is important because of the recent foreclosure crisis. In 2008 the U.S.
government instituted the Home Affordable Modification Program (HAMP), which, in an
effort to solve the foreclosure crisis, provides lenders with incentives to renegotiate more
loans. Consistent with the conjecture of Posner and Zingales (2008), my results illustrate
that there are negative externalities associated with increasing loan renegotiations.
To provide additional evidence on the causal link between debt forgiveness and strategic
default, I exploit a plausibly exogenous event that caused a change in FA’s debt forgiveness
policy. In late August 2012, FA’s director of risk management unexpectedly resigned to
accept a position at a larger financial institution. While FA searched for a replacement,
responsibility for approving debt forgiveness was transferred to executives in the marketing
department who then granted debt forgiveness more freely. In September 2012, a borrower
in default was twice as likely to receive debt forgiveness as in the prior month. As a result,
within the following six months, the percentage of loans in default increased from 41.1% to
47.3%. Consistent with strategic default, the increase was concentrated in the group of loans
that were one to 29 days delinquent. By contrast, adverse economic conditions would affect
default across all delinquency groups.
Next, to further investigate the communication mechanism driving strategic default I
analyze borrower interconnectedness within neighborhoods. I expect that highly interconnected borrowers communicate more and thus exhibit higher strategic default contagion
after forgiveness has been granted within their home neighborhood. To test this conjecture,
I construct multiple measures of borrower interconnectedness: (i) the fraction of borrower
referrals within a neighborhood, (ii) the geographic concentration of loans within zip codes
surrounding each branch, (iii) a size-adjusted geographic concentration measure, and (iv)
the average time that the borrowers of each branch have spent living in their current home.
4
I find that after the change in forgiveness policy, branches holding a more concentrated loan
portfolio or a portfolio with a higher referral rate suffered a higher incidence of strategic
default. Taken together, my results provide evidence that communication among borrowers
results in information contagion that is costly to lenders.
Finally, I examine whether the lender learns from strategic default contagion. The motivation behind this analysis is to understand whether firms recognize that privately disclosed
information spreads between their counterparties and the extent to which they adjust their
policies to mitigate the negative externalities. Studying firm learning is challenging because
equilibrium outcomes are typically observed. However, the surge in defaults after FA’s exogenous event provides an ideal setting to examine lender learning. I test whether the lender
adjusts its origination and forgiveness policies to limit the spread of its private information.
I find that FA’s branches learn from strategic default contagion and react by increasing
the strictness of their forgiveness and origination policies within nine months of the event.
The policy changes are stronger for those branches that experienced higher strategic default
contagion.
These results illustrate that lenders recognize the informational trade-off they face when
making operational decisions that reveal proprietary information. As such, they alter their
policies to mitigate the negative externalities resulting from communication about their
willingness to modify loans. These findings also shed light on the observation that a low
incidence of loan renegotiations occurred during the financial crisis. Prior studies by Piskorski
et al. (2010) and Adelino et al. (2013) investigate whether institutional frictions arising from
securitizations explain the small number of loan renegotiations. I show that although the expost observation that lenders did not increasingly renegotiate during the financial crisis might
seem counterintuitive, lenders limit the number of renegotiations because they recognize the
informational costs ex-ante.
My study makes several contributions. To my knowledge, this study is the first to measure
the information externalities of sharing a lender’s private information with customers. I use
5
a unique approach to show that although the lender shares information privately, borrowers
communicate what they learn to other borrowers. This result provides empirical evidence
that supports the theory that customers can play the role of a firm’s strategic opponent.
In addition, although I study the lender-borrower relationship, my findings are applicable
to other settings where two parties transact with one another privately, such as supplier
negotiations, government-firm negotiations, or a central bank offering capital infusions to a
particular bank. Lastly, I am able to isolate the effect that a lender’s operating policies have
on borrowers’ decisions to strategically default. Disentangling this effect is a step toward
finding appropriate solutions for default and foreclosure contagion.
Section 2 reviews the relevant literature. Section 3 describes the data. Section 4 presents
results on the impact of forgiveness on strategic default contagion; and Section 5 further
explores communication as the mechanism driving strategic default. Section 6 examines
creditor learning. Section 7 concludes.
2
Related Literature
Customers are traditionally regarded as economic agents who contribute positively to the
firm’s bottom line; however, they can also play the role of a strategic opponent if they use
the firm’s private information to their advantage (Verrecchia, 1983; Dye, 1986; Wagenhofer,
1990). In particular, firms constantly share information with customers in negotiations
(Crawford and Sobel, 1982), and the spread of this information can impact the firms’ bargaining power and the outcome of other negotiations. For example, customers may demand
a discount if they learn that a price concession was granted to another customer. Therefore,
understanding whether information shared during private negotiations spreads and measuring the resulting costs for the firm is important.
Although offering debt forgiveness is not a disclosure choice in the traditional sense,
it is an operating decision that reveals private information to outsiders. In this respect,
6
my study is related to the literature on proprietary information, which analyzes whether
sharing information with outsiders is beneficial for the firm. In particular, the majority of
proprietary information studies extensively analyze the effect of competition on different
disclosure decisions, such as product development information (Guo et al., 2004; Jones,
2007), material contract filings (Verrecchia and Weber, 2006), sales and cost information
(Dedman and Lennox, 2009), and customer information (Ellis et al., 2012). These studies
find that despite market pressures, firms do not fully disclose all private information because
of competitors. My study is similar in that firms face a trade-off when deciding whether to
release more or less information in negotiations with customers.
A unique feature of my study is that I examine information sharing through a private
contract. Therefore, the firm cannot perfectly anticipate whether and to what degree information will spread, making it difficult to set optimal policies ex-ante. I exploit this feature to
analyze whether the lender learns about the extent of proprietary costs resulting from information transmission among its borrowers and how the lender alters its policies in response.
A recent study by Ali et al. (2014) shows that firms with high proprietary costs prefer
financing options that allow them to circumvent disclosure requirements. My study complements their findings by providing evidence that a financial institution modifies its debt
forgiveness and origination policies because of the high cost of sharing private information.
My work also relates to two recent streams within the mortgage literature. Guiso,
Sapienza, and Zingales (2013) use survey data to examine the moral and social determinants of homeowners’ attitudes toward strategic default. They find that homeowners are
more likely to default if they know of other strategic defaulters, in part because they perceive the bank is less likely to pursue them. Further work examines how foreclosure spreads
within neighborhoods, performing an analysis within Maryland (Towe and Lawley, 2013), an
analysis in Illinois (Munroe and Wilse-Samson, 2013), and a national analysis that controls
for neighborhood and zip-code (Goodstein et al., 2013). These studies find that foreclosure
is contagious and that neighbors’ behaviors influence default.
7
My study differs from these prior studies in three respects. First, my study is concerned
exclusively with how the spread of the lender’s private information affects strategic default.
I am able to identify this because I observe defaults and loan modifications that are private.
By contrast, a foreclosure is a publicly disclosed event making it difficult to disentangle
communication about the bank’s proprietary information from any learning that occurs
regarding the economic trajectory of the neighborhood or foreclosure procedures (Towe and
Lawley, 2013; Munro and Wilse-Samson, 2013).
Second, my study focuses on the period that spans a borrower’s initial default on a
loan through all renegotiations that occur until the lender sends the loan to a collection
agency. Thus, my setting closely parallels the renegotiation period prior to a foreclosure. As
such, I am able to investigate whether increasing loan modifications helps reduce foreclosure
contagion, as suggested by Munro and Wilse-Samson (2013) yet critiqued by Posner and
Zingales (2009). The paper most closely related to mine is Mayer et al. (2013), who use the
Countrywide lawsuit settlement to analyze how a publicly announced mortgage modification
policy affects strategic default. My approach differs from theirs because I focus on private
loan renegotiations and therefore communication among customers as the channel by which
learning about the lender’s policies occurs. Thus, my study more closely parallels the negotiation of a private contract, because banks and companies do not typically announce they will
grant concessions to all customers who need assistance meeting their financial obligations. In
doing so, I contribute new evidence showing that loan renegotiations cause strategic default
contagion even when renegotiations are conducted privately with only select borrowers in
default.
Lastly, I examine whether the lender identifies the informational costs associated with
renegotiating loans. Despite the importance of loan renegotiations as a tool to minimize bank
losses, there is little work that examines whether the low incidence of renegotiations occurs
because banks anticipate that loan modifications can generate information externalities.
This paucity of evidence is likely due to the difficulty of finding settings where the bank’s
8
renegotiation decisions vary exogenously. I circumvent this challenge by exploiting a plausible
exogenous change to the lender’s forgiveness policy that caused a surge in defaults.
3
Data
I use loan data from Financiera Ayudamos (FA), a subsidiary of Grupo Financiero BBVA
Bancomer, the largest commercial bank in Mexico. FA is a Sociedad Financiera de Objeto
Multiple (SOFOM), a credit institution subject to less government regulation than deposit
banks.3 SOFOMs primarily grant consumer credit, yet also provide financing for small and
medium enterprises, distributors, and intermediaries. SOFOMs are not allowed to accept
deposits; therefore, FA is fully funded by its parent company.
FA began its operations in May 2007 and currently operates 54 branches concentrated
in central Mexico. It mainly serves individuals with limited access to traditional credit.4
FA’s loans range from $1,500 to $50,000 MXN ($115 to $3,845 USD), have a 12 to 24 month
maturity, and have the same 66% interest rate across borrowers. Loans also incur a 16%
value-added tax on interest payments and a 9% origination fee or 5% renewal fee.
My main sample comprises 14,649 FA loans to small business owners, granted from January 2011 through March 2014. Panel A of Table 1 reports borrower and loan characteristics
at origination. On average, the loans have a principal of $6,940 MXN ($515 USD), an annual
payment rate (APR) of 92%, a maturity of 18 months, and a weekly payment frequency.
Borrowers are 57% female and, on average, 39 years old. Nearly two-thirds of borrowers have
a credit score and, of those, the mean score is 692 of 850. According to FA’s risk management department, a credit score of 692 corresponds to a medium-risk borrower. To assess
the credit-worthiness of all borrowers, FA constructs an internal credit score that aggregates
borrower characteristics. Lastly, 38% of borrowers have previously obtained a loan with FA.
3
SOFOMs do not require approval from the National Banking Regulator and are exempt from capital
requirements unless they have an economic relationship with a banking institution (Pena, 2008).
4
FA facilitates the borrowing process by accepting non-traditional documents as proof of income. For
example, FA accepts receipts from inventory purchases as proof of income, whereas most commercial banks
strictly require payroll receipts or bank statements.
9
My identification relies on borrowers living in a tight-knit neighborhood and commuting
to work. Consistent with this, on average, FA’s borrowers have spent 67% of their life at
their current home and travel 3.8 linear miles to work. Given the mountainous geography of
Mexico, the linear distance likely understates the commute to work. For example, in Mexico
City, a 3.8 linear mile can require travel of over six miles on the road, corresponding to a
commute in excess of one hour.
Panel B of Table 1 presents the monthly loan performance. FA monitors loan performance
on a weekly basis. In a given month, an average of 16% of borrowers miss a payment and
are considered to be in default. Each branch is required to pursue collection efforts, which
begin with loan officers placing telephone calls and making home visits. The branch can
also offer debt forgiveness should the first two methods become ineffective. Debt forgiveness
is contingent on the borrower (i) paying the remainder of his deficiency such that the loan
becomes fully current and (ii) committing to make the rest of his payments on time. Prior to
September 2012, FA granted forgiveness to an average of 10% of defaulters, on approximately
4% of the loan principal.5
Once delinquency surpasses 60 days, FA’s centralized collection division takes over the
collection process. This division is, on average, less than 15% successful at getting loans
back into repayment. Lastly, once the loan is more than 90 days delinquent, it is transferred
to an external collection agency that charges a percentage of any recovered portion of the
loan.
5
The average forgiveness rate over the whole sample period is 20% due to the period after August 2012,
as discussed in Section 5.
10
4
Borrower Communication and Strategic Default
Contagion
4.1
Identification Strategy
My first hypothesis is that lenders incur information externalities when they grant debt
forgiveness, because borrowers communicate with other borrowers who are then more likely
to default. To test this hypothesis, I must differentiate defaults that are due to information
spillovers from those that are purely due to adverse economic shocks. To do so, I use a novel
approach that exploits the particularities of the Mexican setting.
First, I account for economic shocks to a borrower’s income that could make him unable
to repay his loan. I exploit the fact that Mexicans typically commute to work, and separately
analyze the default rate of a borrower’s work area and home neighborhood.6 A high default
rate in a borrower’s work area indicates that local economic conditions that impact takehome pay are deteriorating. In particular, the income of small business owners, who sell
food, clothes, personal care items, repair cars, repair shoes, etc., is highly dependent on
common local factors such as a large store opening in the vicinity, a company closing down
or downsizing, nearby protests during the week, construction in the area, etc.
Next, if there are no economic shocks at the work area I consider shocks to a borrower’s
home expenses that might cause default. First, I consider a rise in housing costs. According
to the Mexican National Institute of Statistics and Geography (INEGI), individuals within
the corresponding income deciles of my sample of borrowers only spend 3.2% of their income
on rent. Consistent with this statistic, less than 1% of my sample of borrowers rent their
home. Therefore, in my setting, rent increases are unlikely to contribute to a rise in the
neighborhood’s default rate. Second, I consider a rise in the cost of consumption goods.
INEGI reports that, on average, the two greatest expenses for individuals within the corre6
A lack of urban planning, the lack of jobs in low-income communities, and government housing in the
outskirts of major cities contribute to the high number of commuters.
11
sponding income decile of my sample of borrowers are food (41%) and transportation (17%).
However, my sample period had modest annual inflation rates under 4.5%. Thus, the rise in
the cost of these goods is unlikely sufficient to cause an increase in neighborhood default.
Lastly, if a neighborhood does not experience a widespread shock that causes expenses
to increase, a borrower’s default is more likely driven by communication among neighbors
regarding the lender’s forgiveness policies. In particular, the majority of individuals in
Mexico live in the same neighborhood from infancy to adulthood, forming strong ties to their
community of large extended families and friends (Rodriguez et al., 2007). Weekly activities
within these neighborhoods also contribute to the high level of information sharing. As such,
I examine whether a borrower is more likely to default after forgiveness is granted to other
borrowers within his neighborhood.
Considering FA’s forgiveness policy, a strategic default due to communication would
occur as follows: (i) Borrower A defaults in month t, (ii) Borrower A receives forgiveness in
month t + 1, and (iii) Borrower B defaults in month t + 2. Because Borrower B’s decision
to default is separated from Borrower A’s default by two months, it is much less likely that
Borrower B’s default is due to a pervasive economic shock that impacts the entire home
neighborhood at one point in time. In sum, my identification strategy is that economic
shocks that impact the borrower’s ability to repay his loan are correlated among borrowers
that work in the same area, whereas defaults that are correlated among borrowers that live
in the same neighborhood are primarily due to information transmission.
Figure 1 illustrates the identification strategy. Borrowers A and B live in the same tightknit neighborhood and commute to their respective work areas to run small independent
businesses. Borrowers C and D live in a different neighborhood and commute to their
small businesses. I identify strategic default as follows: suppose Borrower A and Borrower
C default in t on their small consumer loans because of an economic shock that affects the
whole work area. The lender offers Borrower A debt forgiveness in t +1 as a means of getting
him out of default and back into repayment. Borrower A subsequently communicates with
12
his neighbor, Borrower B, regarding the lender’s forgiveness policy. Borrower B, who has not
suffered an economic shock, then chooses to strategically default in t + 2 with the objective
of getting a portion of his loan forgiven. Borrower D does not default, because he neither
experiences an economic shock nor has any interaction with Borrowers A or B. Moreover,
suppose communication occurs among local business owners in the work area. The result
would be an overstated coefficient on economic conditions and an understated coefficient on
forgiveness. Therefore, it will be more difficult to detect contagion.
4.2
Empirical Analysis
To implement the identification strategy I estimate the following panel regression:
Def aulti,t = α0 + α1 F orgiveness Ratehome,t−1 + α2 Def ault Ratework,t +
(1)
Controls + F ixed Ef f ects.
The dependent variable, Def aulti,t , is binary and takes the value of one if a borrower’s
loan i is in default at month t, but not at month t − 1, and zero otherwise. The main
explanatory variable, F orgiveness Ratehome,t−1 , is the forgiveness rate in borrower i’s home
zip code in the prior month. Using lagged forgiveness rate overcomes the reflection problem
that arises in social interaction models (Manski, 2000; Brock and Durlauf, 2001).7
To control for other determinants of default, I include measures of economic conditions,
borrower characteristics, and loan characteristics. I construct a variable, Def ault Ratework,t ,
as a proxy for local economic conditions that impact borrowers’ ability to repay their loans.8
This variable measures the proportion of loans that are delinquent in the zip code where
borrower i works (excluding borrower i). The relatively low correlation of 0.35 between
7
The reflection problem arises when determining whether group behavior affects the behavior of individuals (Manski, 1993). The issue is that group behavior is merely the aggregate of individual behaviors.
8
I construct this variable because a measure of local economic conditions is unavailable at the zip code
level. INEGI’s most granular economic measures are reported at the state level. In addition, INEGI’s state
data are reported quarterly and thus do not allow for a monthly analysis.
13
Def ault Ratework,t and Def ault Ratehome,t indicates that the economic conditions at the
work and home zip codes are distinct from each other (Appendix Table A.2). I also include
industry and time fixed effects to control for macroeconomic cycles and seasonality.
I control for demographic characteristics that affect a borrower’s ex-ante likelihood of
default: age, gender, education, credit score, number of years living in the current house as
a fraction of age, and number of years owning the business as a fraction of age. I expect that
a borrower who has owned his business for many years is more likely to generate a stable
income than a borrower who recently started his business.
I control for loan characteristics: payment frequency, size of the loan relative to income,
and the portion of the loan that is repaid. I expect that borrowers are not likely to default
immediately after obtaining a loan or when a few payments are left. I also anticipate that
default is more likely to occur on loans with weekly payments, because FA grants biweekly or
monthly payments only to borrowers that demonstrate a relatively stable source of income.
Lastly, I cluster standard errors at the branch level to account for branches operating in
distinct areas and therefore the possibility that they implement collection efforts differently.
4.3
4.3.1
Empirical Results
Linear regression analysis
Panel A of Table 2 presents results of the strategic default regressions using a linear probability model.9 A 1% increase in the debt forgiveness rate within a neighborhood increases the
likelihood that an additional borrower will default by 0.087% (column 1). This contagion is
large in terms of economic magnitude. If FA were to double the debt forgiveness rate within
a neighborhood, the likelihood of a borrower defaulting in a given month increases by 1.7%,
equivalent to approximately a 10.9% increase in FA’s monthly default rate. Further, the
joint effect of P ct P aid and P ct P aid2 indicates that the probability of default is a concave
9
I use a linear probability model for ease of interpreting coefficients and because it allows for a larger set
of fixed effects. However, in untabulated results I run logistic specifications and find similar results.
14
function that reaches its maximum when 66% of the loan is repaid. Therefore, as predicted,
defaults do not typically occur at the beginning or near the maturity of the loan.
In column 2, I use FA’s Internal Credit Score as a control, in lieu of borrower characteristics. The internal credit score variable aggregates borrower characteristics and can capture
nonlinearities that FA has already identified and incorporated into its scoring formula. It
also has the added benefit of increasing the sample size, because approximately 36% of FA’s
borrowers lack a credit score. I find that the coefficient of F orgiveness Ratehome,t−1 remains
positive and significant.
Next, I examine repeat defaulters to test whether the likelihood that a borrower defaults
increases after he receives debt forgiveness on his loan. An individual repeatedly defaults
for two reasons: (i) he is a bad borrower or (ii) he is taking advantage of the information
that he previously learned about FA’s willingness to offer forgiveness. Accordingly, I include
two additional explanatory variables (column 3). The first variable, N ot F orgiven, is binary
and takes the value of one for borrowers who defaulted in the past but did not receive
forgiveness. The second variable, P ast F orgiveness, is also binary and takes the value of
one for borrowers who received forgiveness in the past. The coefficient of N ot F orgiven is
positive, indicating that bad borrowers are more likely to default again. However, consistent
with borrower learning, the coefficient of P ast F orgiveness is greater than the coefficient of
N ot F orgiven. The difference in the two coefficients shows that conditional on being a bad
borrower, learning about forgiveness increases the likelihood of default by 14.3%.
In column 4, I include repeat borrowers. I am interested in this subset of borrowers
because their past borrowing history with the lender creates less information asymmetry
regarding their wealth. These borrowers are also likely interested in renewing their existing
loans or obtaining new loans, which increases the cost of taking actions that negatively affect
their credit score. In particular, when making renewal decisions, FA penalizes borrowers for
having received forgiveness on a loan. Generally, FA does not renew a loan if the borrower
has defaulted within six months or defaulted multiple times in the past. Approximately
15
5% of repeat borrowers received forgiveness on a previous loan. In line with my prediction, the coefficient of Repeat Borrower is negative and significant, indicating that repeat
borrowers default less on average. The interaction coefficient of F orgiveness Ratehome,t−1 ×
Repeat Borrower indicates that repeat borrowers do not default more when the forgiveness
rate increases in their neighborhood. This is consistent with the lender being more informed
about the repeat borrowers’ wealth, making it more difficult for this subset of borrowers to
strategically default and obtain forgiveness.
In column 4, I substitute Def ault Ratework with the interaction of work area and month
fixed effects. The purpose of the work-month fixed effects is to eliminate the effect of economic shocks that are common to all businesses in a given work area each month. Lastly, I
include home neighborhood fixed effects to control for time-invariant characteristics of the
home neighborhood that can influence the likelihood of default (column 5). The coefficients
are estimated from within-home-neighborhood variation, requiring that any potential confounding event be correlated with time-varying home neighborhood characteristics. The
coefficient of F orgiveness Ratehome,t−1 is positive and significant in all specifications.
Next, I further investigate the communication mechanism that drives strategic default. I
test whether borrowers with stronger ties to their neighborhood are more likely to strategically default after a neighbor receives forgiveness. In Panel B, I partition the main regression
analysis by high and low Lif e at Home, which is determined by the median percentage of
life spent in the current home. I find that the coefficients of F orgiveness Ratehome,t−1 are
positive and significant for borrowers in the high group. For the borrowers in the low group,
the coefficients are positive, but not significant. These results demonstrate that stronger
neighborhood ties lead to more pervasive communication, and ultimately higher levels of
strategic default.
16
4.3.2
Instrumental variables
I use instrumental variables (IV) as an alternative empirical method to strengthen the causal
link between prior forgiveness and default (Table 3). The objective of the IV approach
is to exploit shocks to the home neighborhoods’ forgiveness rate that are exogenous to
each borrower. This is different from the regression analysis, which attempts to control
for all economic conditions that affect borrowers. As such, IV allows me to circumvent
potential concerns about endogeneity in FA’s decisions to grant debt forgiveness. If FA
can detect different levels of communication across neighborhoods, its best response is to
lower debt forgiveness in areas with high communication. From a statistical perspective,
this rational behavior by FA would result in a downward bias in the ordinary least squares
(OLS) estimates.
For each borrower i, I instrument the home neighborhood’s forgiveness rate by using the
fraction of neighboring borrowers that work in a different area and became delinquent in
t − 2. The first-stage regressions show that this instrument, N ew Def ault Ratework(−i),t−2 ,
satisfies the relevance restriction (Table 3, Panel B). The coefficient of the instrument is
positive and significant and the Angrist and Pischke (2008) partial F -statistics are above
the critical value for a weak instrument as defined by Stock et al. (2002). This instrument
likely satisfies the exclusion restriction. First, it is constructed using the default rate of
borrowers that work in a different zip code and are therefore exposed to different local
economic conditions. Second, the instrument is lagged by two periods, further separating
economic shocks suffered by the defaulters from shocks that could affect borrower i. Given
that these borrowers have minimal or no savings, it is unlikely that they will be able to
withstand economic shocks for a prolonged period of time without defaulting. As such, it is
reasonable to assume that the only channel by which N ew Def ault Ratework(−i),t−2 affects
borrower i’s decision to default is through the level of forgiveness that i observes in his home
neighborhood.
In line with my previous results, the coefficient of F orgiveness Ratehome,t−1 is positive
17
and significant across all specifications. These coefficients are also larger than those of the
OLS regressions, which is consistent with a downward bias in the OLS estimates. In terms of
economic magnitude, a one standard deviation increase in the forgiveness rate in the home
neighborhood increases the default rate by 2.6%.
Overall, the results in this Section indicate that the lender’s day-to-day operations reveal
proprietary information to customers who then communicate the terms of their agreement to
other borrowers, causing negative externalities in the form of strategic default contagion. The
lender’s private information spreads even when it is exclusively disclosed in a private contract.
Considering that all types of firms engage in a multitude of private contracts this result is
likely to apply to other settings where two companies transact with one another privately.
For example, any firm that decides to grant better delivery terms, additional warranty, or
additional services to select customers can incur costs if doing so induces additional customers
to demand the same treatment. In addition, although my setting focuses on communication
among individual borrowers, prior research has found evidence of information transmission
among decision makers at different firms. For instance, manufacturing executives gather
and transmit information about retailers’ payment policies through business associations
(Doner and Schneider, 2000), investment managers communicate with other local investment
managers about portfolio allocation decisions (Hong et al., 2005), and executives within the
same social networks discuss and influence each other’s managerial decisions (Cohen et al.,
2008; Shue, 2013). Therefore, the mechanism of communication among individual customers
regarding a firm’s private information is also generalizable to communication among firms’
decision makers.
Lastly, my results shed light on the information externalities associated with loan renegotiations. In particular, I provide evidence consistent with the theoretical predictions of
Posner and Zingales (2009), which state that loan renegotiations, although efficient ex-post,
can create future costs for the bank. This result is important from a policy perspective
because renegotiations have received attention as a potential solution to reduce mortgage
18
foreclosures. For example, the Obama administration enacted the Home Affordable Modification Program in 2008 to provide incentives for lenders to increase loan renegotiations.10
Beyond demonstrating that renegotiations have a costly consequence for lenders, my results
indicate that the foreclosure contagion problem cannot be solved by exclusively prescribing
renegotiations.
4.4
Robustness and Alternative Explanations
I perform a number of additional tests to evaluate the robustness of my main results. First,
I rule out potential shocks to expenses in the home neighborhood. Then I use a placebo
test to examine whether default contagion is due to online communication regarding FA’s
forgiveness policy.
4.4.1
Shocks to expenses at the home
I use a series of alternative specifications to test for possible alternate explanations for strategic default contagion (Table 4). First, I include the contemporaneous default rate in the home
neighborhood as a control for potential correlated economic shocks in the home neighborhood
that drive borrowers’ default. A drawback of including this variable, Def ault Ratehome,t , is
that it is affected by the lagged forgiveness rate. As such, it will absorb a portion of the effect
associated with F orgiveness Ratehome,t−1 and thus, attenuate the coefficient. Although the
coefficient of F orgiveness Ratehome,t−1 is downwardly biased in this specification, it remains
positive and significant (column 1). Therefore, increasing debt forgiveness increases strategic
default contagion after controlling for economic conditions.
Next, I include inflation subcomponents to control for price increases in different types
of goods consumed by the borrowers. For example, according to INEGI, the average borrower in my sample spends 17% of his income on transportation. Therefore, I expect that
borrowers will be more likely to default when there is a high inflation rate with respect
10
For more details on the consequences of the HAMP see Agarwal et al. (2012).
19
to transportation goods and services (column 3). Including these controls does not affect
the coefficient of F orgiveness Ratehome,t−1 . The coefficients of inflation in transportation,
clothing, and housing are positive and significant. By contrast, the coefficient of inflation in
food prices is not statistically significant, likely due to the large portion of borrowers in my
sample that own food sales businesses.
I also analyze natural disasters because they are economic shocks that can cause a correlation in defaults among borrowers living in the same neighborhood. I include a binary
variable, N atural Disaster, that takes the value of one for loans in states that suffered
a natural disaster during the year, and the interaction term F orgiveness Ratehome,t−1 ×
N atural Disaster. The coefficients of these two variables are not statistically different from
zero. In this specification, I also include a binary variable, M ultiple Accounts, that takes
the value of one when a borrower has multiple credit accounts at origination, and the interaction term F orgiveness Ratehome,t−1 × M ultiple Accounts. A borrower that has multiple
outstanding loans is possibly learning from the actions of various lenders. By controlling for
borrowers with multiple credit accounts I disentangle whether the borrower is learning from
FA’s actions. I find that these borrowers are less likely to default but equally likely to learn
from neighboring forgiveness and to strategically default.
In column 4, I test whether the spread of a disease could explain default contagion.
According to the Centers for Disease Control and Prevention (CDC), flu activity peaks
between December and February. Therefore, I include a binary variable, W inter, which takes
the value of one for months between November and March. I also include the interaction term
F orgiveness Ratehome,t−1 × W inter. The coefficient of F orgiveness Ratehome,t−1 remains
largely unaffected, indicating that the spread of contagious diseases such as the flu do not
drive the main results.
Lastly, I use a data set of 62 Walmart store openings within my sample period to test
for local economic shocks to borrowers who work within 2 miles of the area. The idea is
that a Walmart store opening could cause an increase in competition to local businesses.
20
Conversely, I use the store openings as a positive shock to borrowers who live within 2
miles of the area, because it could allow them to purchase goods at lower prices. I also
include the interaction term of each of these two variables with F orgiveness Ratehome,t−1 .
The coefficients of F orgiveness Ratehome,t−1 remain largely unchanged, but the coefficients
of Walmart openings are not statistically significant. I attribute these results to low power
in this test, because the Walmart store openings did not occur within close proximity of the
majority of FA borrowers.11
4.4.2
Placebo test
Under my contagion hypothesis, I expect that granting forgiveness in a home neighborhood
reveals information that creates incentives for other borrowers in the same neighborhood to
strategically default. A potential concern is that borrowers could learn about FA’s forgiveness
policies through online communication as opposed to traditional word-of-mouth communication. To rule out this concern, I test whether randomized forgiveness rates explain future
defaults. Specifically, I conduct placebo tests that assign a random home neighborhood
to each borrower (column 6). I find that the average coefficient of forgiveness rate in the
placebo tests is not statistically different from zero. In addition, this coefficient is positive
and significant in only 3% of my simulations.
5
Communication Channel
5.1
The 2012 Event
As of 2007, FA’s debt forgiveness policy required that each branch file a formal request and
obtain approval from FA’s central risk management department before granting debt forgiveness to a borrower. The risk management department reviewed all formal requests to ensure
each loan was at least six months old and that each borrower signed a risk awareness form
11
Less than 0.4% of my sample of loan-months fall within this category.
21
that outlined the negative credit consequences of receiving debt forgiveness.12 On average,
only 10% of defaulters received debt forgiveness. Of these, few were repeat forgiveness cases
as FA typically declines requests for repeat forgiveness.
In August 2012, the director of the risk management department unexpectedly left FA for
a position in a much larger credit institution. Consequently, the department lacked a manager with the authority to approve debt forgiveness, and within two weeks, a large backlog of
requests accumulated. To alleviate the issue, the CEO/president of the company temporarily transferred the approval responsibilities to executives within the marketing department.
Given the marketing department’s lack of experience in assessing risk and their focus on
business growth, the executives rapidly granted most forgiveness requests as a means of getting clients to repay. The executives also eliminated the need for borrowers’ signatures on
the risk awareness form. As a result, FA’s total dollar amount forgiven increased 188% from
the previous month (Figure 2).
In light of my prior results, I anticipate that the loan performance worsened after this
event, because of a surge in strategic defaults. To examine this prediction, I compare monthly
transition matrices of delinquencies in FA’s loan portfolio during the six-month period before and after September 2012. Loans that were one to 29 days delinquent increased from
29.9% to 36.8%, whereas all delinquent loans over 30 days slightly decreased from 11.2%
to 10.5% (Table 5, Panels A and B). These results have two implications. First, because
the delinquency rate did not increase across all groups, adverse economic conditions do not
account for the high incidence of default. Second, the increase in debt forgiveness was not
an effective policy to reduce the overall number of loans in default. An effective policy would
have caused a larger decrease in delinquent loans over 30 days and little to no impact on the
rate of loans that were one to 29 days delinquent. In addition, the likelihood that, in a given
month, non-delinquent borrowers became delinquent increased by 2.24% and the likelihood
12
FA reports borrowers who received debt forgiveness to the Credit Bureau. The Credit Bureau then
factors the forgiveness into the borrowers credit score, making it more difficult for him to obtain credit in
the future.
22
that loans with delinquencies over 60 days re-entered repayment status increased by 0.51%
(0.39%+0.12% in Panel C).
In sum, the new forgiveness policy intended to help delinquent borrowers get back into repayment, yet it had the unintended consequence of incentivizing strategic default among nondelinquent borrowers. Moreover, an additional unintended consequence is a cross-subsidy between sophisticated and unsophisticated borrowers as described by Campbell (2006). Strategic defaulters are likely the higher credit quality borrowers among the group of defaulters,
and will most likely absorb the majority of the forgiveness offered by FA.13 Consistent with
this conjecture, the credit quality of borrowers that received forgiveness improved. The average internal credit score increased from 585 points to 597 points after the forgiveness policy
change.
5.2
Neighborhood Interconnectedness
To further analyze the mechanism driving my prior results, I test whether neighborhood
interconnectedness explains strategic default contagion in the cross-section of branches. I
expect that tight-knit neighborhoods exhibit a high level of inter-borrower communication.
Therefore, branches granting loans to borrowers residing within tight-knit neighborhoods
will experience a higher level of strategic default contagion as compared to branches that
grant loans to less inter-connected neighborhoods. A potential concern is the endogeneity of
branches’ decision to grant forgiveness, because, in equilibrium, branches that observe more
communication may grant less forgiveness. To address this concern, I exploit the change
to FA’s forgiveness policy in September 2012. I construct four measures of the level of
neighborhood interconnectedness one month prior to the policy change, and test whether
they explain the level of strategic default observed during the three months after the change.
The first variable, Ref erral Ratej , captures the fraction of borrowers referred to FA by
13
Campbell (2006) explains that some financial products create a cross subsidy between sophisticated
and unsophisticated households. For example, a cross subsidy arises when unsophisticated borrowers do
not optimally refinance, thus allowing the sophisticated borrowers to obtain more attractive terms from the
financial institution.
23
a family member or a friend. The second variable, Concentrationj , measures the geographic
concentration of the loans within each branch’s portfolio through a Herfindahl measure:
Concentrationj =
X
h2i,j ,
(2)
i
where hi.j is equal to the number of loans originated by branch j in zip code i as a fraction of
the number of loans in the portfolio of branch j. The third variable, Adj. Concentrationj ,
adjusts the Concentrationj measure to give less weight to large zip codes:
Adj. Concentrationj =
X
i
h2i,j ×
1
,
log(areai )
where areai is the area of a circle with radius equal to half the distance between the center
of zip code i and the center of the closest zip code. The last variable, F raction of Lif ej , is
the average number of years that borrowers of branch j have spent in their current home as
a fraction of their age.
I use the following linear specification to analyze the relation between strategic default
and neighborhood interconnectedness:
Strategic Def aultj = γ0 + γ1 Ref erral Ratej,Aug12 + γ2 Concentrationj,Aug12
(3)
+γ3 Adj. Concentrationj,Aug12 + γ4 F rac. of Lif ej,Aug12 .
In this regression, Strategic Def aultj is equal to αd
1,j × F orgiveness Ratej , the product
of the average marginal effect of the forgiveness rate and the average forgiveness rate. I
estimate αd
1,j from the following specification:
Def aulti,j,t =
X
(α1,j (Ij × F orgiveness Ratehome,t−1 )) + Controls +
j
Ij + M onth F.E.,
24
(4)
where Ij is an indicator variable that takes the value of one if loan i was originated in
branch j, and zero otherwise. The controls are the same as in equation 1. I use Monte
Carlo simulations to adjust the standard errors to account for the presence of the generated
14
regressor αd
1,j .
The coefficients of Ref erral Rate, Concentration, Adj. Concentration, and F raction of Lif e
are positive; however, only Ref erral Rate is statistically significant at the 10% level (Table
6). This result suggests a positive relation between the fraction of referred customers and
future contagion. The coefficient of Concentration is not statistically significant, which may
be due to the variable lacking more detailed geographic and demographic data. The coefficient of Adj. Concentration is larger, but remains statistically insignificant. This finding
suggests that this variable can be further improved by including more precise data about zip
code population density.15
In sum, neighborhood interconnectedness helps explain the level of strategic default observed after the lender modified its forgiveness policy. This result is consistent with borrower communication as the channel by which the lender’s private information spreads. As
additional borrowers communicate about the lender’s willingness to offer debt forgiveness,
strategic default contagion increases.
6
Lender Learning
In this Section, I examine whether the lender internalizes the costs resulting from borrowers communicating the terms of their renegotiation agreements. Specifically, I investigate
whether the lender alters its operating policies as it learns about strategic default contagion.
Although it is well understood that companies learn by doing (Arrow, 1962) and in particular
14
I use the coefficients α
d
1,j and their joint variance-covariance matrix obtained from regression 4 to
simulate 10,000 random draws of each coefficient αd
1,j,s from a multivariate normal distribution. I then
estimate regression 4 using each set of simulated coefficients α
d
1,j . I report the average coefficient and
p-value obtained after repeating the estimation of regression 4 10,000 times.
15
The INEGI does not report information about the geographic size and total population at the zip code
level, thereby limiting the set of variables available to construct more sophisticated measures.
25
by observing financial data (Pastor and Veronesi, 2009), typically, only equilibrium outcomes
are observable. Therefore, identifying whether, how, and to what degree lenders learn from
strategic default contagion is challenging. I overcome this obstacle by further exploiting
FA’s exogenous event in 2012. I examine FA’s actions after the event and test whether
branches that experienced a high level of contagion subsequently increased the strictness of
their forgiveness and origination policies.
6.1
Forgiveness Policy
FA’s branch managers receive a bonus as a function of delinquencies in their respective loan
portfolios. As such, they have incentives to minimize the number and extent of nonperforming loans. One method to reduce defaults is to modify the forgiveness policy. Given that
FA grants individual branches the autonomy to implement stricter forgiveness policies, I
expect that branches that experience high strategic default will learn about the information
externality and thus tighten their forgiveness policy. It is ultimately an empirical question
as to how long a branch would take to identify contagion; therefore, I conduct the analysis
using learning periods of three, six, and nine months (Figure 3).
To test my conjecture, I first use the regression in equation 4 to estimate the coefficient
of F orgiveness Rate for each branch j throughout each learning period. I then use the
estimates αd
1,j as a measure of the level of information transmission among the borrowers in
branch j throughout the learning period. This measure is consistent with the geographic
location of the lender’s branches, because borrowers are required to originate and service
their loans at the branch nearest to their home. In addition, because branches serve a
distinct set of neighborhoods, they are not simultaneously impacted by strategic default.
Next, I construct a measure of forgiveness policy strictness, Likelihood of F orgiveness,
to analyze the extent to which each branch uses debt forgiveness. This measure captures the
average number of loans in default that receive debt forgiveness at each branch. Branches
that offer forgiveness to a low fraction of defaulters are considered stricter.
26
To examine the relation between the level of information transmission and the level of
forgiveness policy strictness, I conduct the following cross-sectional test:
Likelihood of F orgivenessj,post = γ0 + γ1 Strategic Def aultj + γ2 T otal Def aultj ,
(5)
where Likelihood of F orgiveness is measured after the learning period, Strategic Def aultj
is the product αd
1,j ×F orgiveness Ratej estimated from equation 4 using the learning period,
and T otal Def aultj is the average default rate observed in branch j during the learning
period. In addition, I use simulations in all specifications to correct standard errors for the
presence of a generated regressor.
I find no clear relation between strategic default and likelihood of forgiveness within a
three-month learning period (Table 7, column 1). Thus, the branches are either unable to
detect strategic default or fail to react by tightening their forgiveness policy. Consistent
with the notion that learning occurs gradually, increasing the learning period to six months
strengthens the relation between strategic default and forgiveness policy strictness (column
3). Although not statistically different, the coefficient is larger in absolute value and has a
smaller p-value.
In a nine-month period (column 5), the branches tighten their forgiveness policies in
response to learning from contagion. Branches that experience higher strategic default are
less likely to grant forgiveness to defaulters. In terms of economic magnitude, within a
nine-month learning period, the branches reduce their use of forgiveness by 2.44%. This
corresponds to approximately a 4.7% tightening of the forgiveness policy. In columns 2, 4,
and 6, I include the average default rate of the branch as a control for any change in overall
default, which could drive the change in forgiveness. My results are robust to including this
additional control.
27
6.2
Origination Policy
Next, I examine how FA branches adjusted their origination policy within the three, six, and
nine months after the surge in defaults in September 2012. Branches can reduce strategic
default by improving the quality of their borrowers through stricter loan origination policies.
Alternatively, branches can be less strict in origination. Although the latter strategy may
seem counterintuitive, it could prove effective if the lender credibly commits to not grant
debt forgiveness. The idea is that borrowers would refrain from strategically defaulting once
they realize that delinquent loans will not be forgiven. Relatedly, managers may choose a
less strict origination policy due to their compensation incentives. FA pays an additional
bonus that increases as a function of the number of loans originated. Managers can therefore
substitute their loan performance bonus with the origination bonus. Specifically, when the
loan performance bonus decreases due to increased default, managers can lower origination
standards to increase their origination bonus.
Traditionally, FA’s loan granting process takes less than 24 hours and relies on the borrower’s credit score and internal origination score. The credit and origination scores are
compared to the company’s loan granting thresholds. The branches cannot grant loans to
individuals who fail to meet the minimum requirements; however, they can choose to deny
a loan to an individual that meets the minimum requirements. If a branch sets higher standards such that it rejects candidates that meet FA’s requirements, it is considered stricter.
To examine whether the level of information transmission during the learning period
affects the strictness of the origination policy for a given branch, I conduct the following
test:
Decision to Rejectj,post = γ0 + γ1 Strategic Def aultj + γ2 T otal Def aultj .
(6)
I calculate the dependent variable, Decision to Reject, as the number of qualifying loan
requests that were rejected, as a fraction of all qualifying requests. This measure is based on
branches’ actual decisions, as opposed to the average quality of new borrowers. Therefore,
28
the measure has the advantage of circumventing concerns regarding branches changing their
strictness because of loan requests from worse candidates. For example, using a variable that
measures the fraction of loans accepted or rejected could be biased if there are changes to the
quality of the candidates requesting loans from the branch. In contrast, Decision to Reject
captures the decision made by the loan officer. These decisions are frequently subjective and
will depend on the loan officers’ incentives.
The relation between level of contagion and origination strictness after a three-month
learning period is not statistically significant (Table 8, column 1). Although not statistically
different, the coefficients of strategic default are larger for six and nine-month learning periods
(columns 3 and 5). Therefore, as the learning period increases, the lender’s branches are
better able to detect strategic default and in turn tighten their origination policies. In
addition, branches that suffer from higher levels of strategic default are more likely to reject
loans that meet FA’s origination thresholds. After a nine-month learning period, a 1%
increase in strategic default leads to a 0.8% increase in loan rejections (column 5). This
is equivalent to a 10.1% increase in the likelihood that a branch rejects an acceptable loan
request.
In sum, I find that branches modify their operating decisions as they learn about strategic
default contagion. On average, branches that experience a higher level of contagion react
within nine months by granting less debt forgiveness and by rejecting more borrowers that
meet FA’s loan origination thresholds. Accordingly, lenders renegotiate less often as they
learn that renegotiations are suboptimal because of information externalities.
6.3
Robustness
In this section I conduct two robustness tests. Specifically, in Panel A of Table 9 I perform
a placebo test. I repeat the analyses in Tables 7 and 8 for a nine-month learning period
assuming that the exogenous change in FA’s forgiveness policy occurred on three different
placebo months: February, July, and December 2010. In these tests I find that the coefficients
29
of Strategic Def ault are not statistically significant. In some of the cases the coefficients
also have the opposite sign of an outcome consistent with FA learning over time.
I also test whether the learning results are due to different branches having different
strictness levels for their origination and forgiveness policies before FA changed its forgiveness
policy. For example, if a given branch has historically been stricter than other branches, it
is not clear that learning about strategic default contagion drives the strictness level after
the policy changed. In Panel B, I use changes in the level of Likelihood of F orgiveness and
Decision to Reject from September 2011, one year before FA changed its forgiveness policy.
The coefficients of Strategic Def ault are consistent with FA learning over time. However, in
contrast to previous results, the nine-month learning period coefficient within the forgiveness
regression is not significant at the 10% level (column 3).
7
Conclusion
This paper exploits a unique setting to examine the negative externalities companies face
when revealing private information to customers. Using a detailed data set of loans to
small business owners, I find strategic default contagion among customers within the same
neighborhood who communicate regarding the lender’s forgiveness policy. I also find that
the lender learns from strategic default contagion and attempts to mitigate it by tightening
its debt forgiveness and origination policies.
A distinct feature of my setting is that negotiations occur in a private setting. Therefore,
my findings shed light on the dynamic information environment of private negotiations and
information transfers between firms and their customers. While theoretical models demonstrate that customers can play the role of a firm’s strategic opponent, there is a paucity of
empirical evidence that quantifies the costs incurred when a firm shares private information
with customers. I am able to quantify the cost by estimating the level of strategic default
contagion after forgiveness is granted.
30
Customers are particularly important to study because they are a key determinant of a
firm’s success. Their relationship with firms allows them to gain access to information that
is not publicly available. As such, it could be especially damaging to a firm if customers
freely communicate the information they obtain through negotiations. It is also valuable to
understand how customers’ actions impact the firm. My results show that firms take into
account the possibility of customer communication and modify their operating decisions
accordingly.
Furthermore, my paper contributes to the foreclosure contagion literature by demonstrating the consequences of loan modifications. The objective of loan modifications is to curtail
default. However, in line with the economic predictions of Posner and Zingales (2009), I show
that loan modifications do not achieve an efficient outcome. Instead, due to borrower communication, loan modifications (i) only modestly reduce the likelihood that a loan in default
becomes uncollectible and (ii) cause the lender to incur expenses beyond that of granting
forgiveness to financially distressed borrowers, as financially able borrowers strategically default. Lastly, I show that lenders recognize strategic default as a cost associated with loan
modifications and therefore limit their renegotiation activities. This finding helps answer the
question of why there was a low incidence of loan renegotiations during the financial crisis.
31
Appendix
A.1 Definition of variables
Variable
Description
Adj. Concentration
Size-adjusted Herfindahl index measure of the zip code concentration of
loans held by each branch on August 2012.
P
1
, where hi,j is equal to
Adjusted Concentrationj = i h2i,j × log(area
i)
the number of loans originated by branch j in zip code i as a fraction of
the number of loans in the portfolio of branch; areai is the area of a
circle with radius equal to half the distance between the center of zip
code i and the center of the closest zip code.
Age
Borrower’s age in years.
Ch. in Likelihood
of F orgiveness
Difference in the Likelihood of F orgiveness in the post-learning period
and the Likelihood of F orgiveness in September 2011 (one year before
the change to FA’s forgiveness policy).
Ch. in Decision to Reject
Difference in the rate of Decision to Reject in the post-learning period
and the rate of Decision to Reject in September 2011 (one year before
the change to FA’s forgiveness policy).
Concentration
Herfindahl index measure of the zip code concentration of loans held by
P
each branch on August 2012. Concentrationj = i h2i,j , where hi,j is
equal to the number of loans originated by branch j in zip code i as a
fraction of the number of loans in the portfolio of branch j.
Credit Score
Borrower’s credit score at origination, as reported by Buró de Crédito.
Scores range from 400 to 850.
Commuting Distance
Number of miles between home and work zip codes. This distance is
calculated using the spherical law of cosines using latitude and longitude
coordinates from GeoData Ltd.
Decision to Reject
Number of loans that met the minimum origination requirements but
were rejected, divided by the total number of loans that met the
minimum origination requirements. During the testing period of my
sample, the minimum internal score required from borrowers is 599.
Def aultt
One if the borrower was current on his loan in month t − 1 and had one
or more missed payments at the end of month t; zero otherwise.
Def ault Ratehome,t
Total number of nonperforming loans in month t as a fraction of all
loans in the borrower’s home zip code.
32
Variable
Description
Def ault Ratework,t
Total number of nonperforming loans in month t as a fraction of all
loans in the borrower’s work zip code.
Education
The highest degree of education attained by the borrower. One if
primary school, 2 if secondary school, 3 if high school or technical
school, 4 if undergraduate degree, 5 if graduate degree, and 0 otherwise.
F orgiven in the P ast
One if the borrower has previously received forgiveness on his current
loan; zero otherwise.
F orgiveness Ratehome,t−1
Total number of loans that received forgiveness in month t − 1 as a
fraction of all loans in the borrower’s home zip code.
F raction of Lif e
On August 2012, the average fraction of life that the borrowers in a
given branch’s loan portfolio had been living in their current home.
Lif e at Home
Years at current home divided by age.
Lif e at Occupation
Years at current occupation divided by age.
Gender
One if the borrower is male, zero otherwise.
Inf lation in Clothes
Monthly inflation rate in clothing items, as reported by INEGI.
Inf lation in F ood
Monthly inflation rate in food, as reported by INEGI.
Inf lation in Housing
Monthly inflation rate in housing, as reported by INEGI.
Inf lation in T ransportation
Monthly inflation rate in transportation services, as reported by INEGI.
Internal Credit Score
FA’s internally generated borrower loan eligibility score, which
incorporates demographic information and credit score. Score ranges
from 0 to 680.
Likelihood of F orgiveness
Average number of loans in default that receive debt forgiveness at each
branch.
Loan Amount
Principal amount of the loan at origination in thousands of Mexican
pesos.
M ultiple Accounts
One if the borrower has multiple credit accounts at origination as
reported by Buró de Crédito, zero otherwise.
N atural Disaster
One if the state where the borrower lives used funds from the Federal
Natural Disaster Fund (FONDEN) during the year; zero otherwise.
N ew Def ault Ratework(−i),t−2
Fraction of borrowers that become delinquent in t − 2 and satisfy the
following two conditions: (i) they live in the same home neighborhood as
borrower i, and (ii) they do not work in the same zip code as borrower i.
N ot F orgiven
One if the borrower has defaulted on his current loan in the past, but
did not receive forgiveness; zero otherwise.
33
Variable
Description
P ayment F requency
One if the loan requires monthly payments, two if it requires biweekly
payments, and three if it requires weekly payments.
P ct P aid
Percentage of the loan that has been repaid at time t.
P ct P aid2
Square of the percentage of the loan repaid at time t.
Ref erral Rate
On August 2012, the fraction of borrowers in a given branch’s loan
portfolio that had been referred to FA through a family member or
friend.
Repeat Borrower
One if the borrower has previously had a loan with FA; zero otherwise.
Reported Income
Borrower’s income in thousands of Mexican Pesos as reported at
origination.
Strategic Def ault
Estimated marginal effect of forgiveness α
d
1,j times average forgiveness
rate.
T otal Def ault
Number of loans in default divided by the number of loans in the
portfolio of each branch.
W almart N ear Homet
One if the borrower lives at a distance smaller than two miles and works
at a distance greater than two miles from a Walmart store that was
opened in month t.
W almart N ear W orkt
One if the borrower works at a distance smaller than two miles and lives
at a distance greater than two miles from a Walmart store that was
opened in month t.
W inter
One if the month is between November and March; zero otherwise.
34
A.2 Correlations table
This table presents the Pearson correlations for the variables used in the strategic default regressions. Variable definitions are provided in Appendix
Table A.1.
35
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
(11)
(12)
(13)
(14)
Variables
Def aultt
F orgiveness Ratehome,t−1
Def ault Ratework,t
Def ault Ratehome,t
Internal Credit Score
Age
Gender
Education
Credit Score
F raction of lif e at current home
F raction of lif e at current occupation
Loan/Income
P ayment F requency
Repeat Borrower
(1)
1.000
0.057
0.053
0.062
-0.085
-0.068
-0.004
0.013
-0.086
0.013
-0.007
-0.041
0.034
-0.038
(2)
1.000
0.133
0.351
0.015
-0.023
0.027
0.041
-0.005
0.089
0.043
0.002
0.011
0.087
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
(11)
(12)
1.000
0.279 1.000
0.006 0.010 1.000
-0.010 -0.043 0.367 1.000
-0.026 0.002 0.071 0.075 1.000
-0.010 -0.005 0.015 -0.157 0.020 1.000
0.002 0.016 0.448 0.078 -0.010 -0.069 1.000
0.012 0.034 0.019 -0.254 0.022 0.176 -0.038 1.000
-0.033 -0.008 0.230 0.018 0.082 0.063 -0.013 0.203 1.000
-0.013 -0.014 0.049 0.012 -0.002 -0.010 0.131 0.031 0.007 1.000
0.031 0.018 -0.038 -0.059 -0.048 -0.108 -0.001 0.043 0.029 -0.035
-0.004 0.001 -0.009 0.095 0.039 -0.002 -0.036 0.010 0.093 0.166
(13)
1.000
-0.014
References
Adelino, M., K. Gerardi, and P. S. Willen (2013). “Why don’t Lenders renegotiate more home
mortgages? Redefaults, self-cures and securitization”. Journal of Monetary Economics 60,
pp. 835–853.
Agarwal, S., G. Amromin, I. Ben-David, S. Chomsisengphet, T. Piskorski, and A. Seru
(2013). “Policy intervention in debt renegotiation: Evidence from the Home Affordable
Modification Program”. NBER, Working paper.
Ali, A., S. Klasa, and E. Yeung (2014). “Industry concentration and corporate disclosure
policy”. Journal of Accounting and Economics, forthcoming.
Angrist, J. D. and J. S. Pischke (2008). Mostly harmless econometrics: An empiricist’s companion. Princeton University Press.
Arrow, K. J. (1962). “The economic implications of learning by doing”. The Review of Economic Studies 22, pp. 155–173.
Brock, W. A. and S. N. Durlauf (2001). “Interactions-based models”. Handbook of Econometrics 5, pp. 3297–3380.
Campbell, J. Y. (2006). “Household finance”. The Journal of Finance 61, pp. 1553–1604.
Cohen, L., A. Frazzini, and C. Malloy (2008). “The small world of investing: Board connections and mutual fund returns”. Journal of Political Economy 116, pp. 951–979.
Crawford, V. P. and J. Sobel (1982). “Strategic information transmission”. Econometrica:
Journal of the Econometric Society 50, pp. 1431–1451.
Dedman, E. and C. Lennox (2009). “Perceived competition, profitability and the withholding
of information about sales and the cost of sales”. Journal of Accounting and Economics
48, pp. 210–230.
Doner, R. F. and B. R. Schneider (2000). “Business associations and economic development:
Why some associations contribute more than others”. Business and Politics 2, pp. 261–
288.
36
Dye, R. A. (1986). “Proprietary and nonproprietary disclosures”. Journal of Business 59,
pp. 331–366.
Ellis, J. A., C. E. Fee, and S. E. Thomas (2012). “Proprietary costs and the disclosure of
information about customers”. Journal of Accounting Research 50, pp. 685–727.
Goodstein, R., P. Hanouna, C. D. Ramirez, and C. W. Stahel (2013). “Contagion effects in
strategic mortgage defaults”. George Mason University, Working paper.
Guiso, L., P. Sapienza, and L. Zingales (2013). “The determinants of attitudes toward strategic default on mortgages”. The Journal of Finance 68, pp. 1473–1515.
Guo, R., B. Lev, and N. Zhou (2004). “Competitive costs of disclosure by biotech IPOs”.
Journal of Accounting Research 42, pp. 319–355.
Hong, H., J. D. Kubik, and J. C. Stein (2005). “Thy neighbor’s portfolio: Word-of-mouth
effects in the holdings and trades of money managers”. The Journal of Finance 60,
pp. 2801–2824.
Jones, D. A. (2007). “Voluntary disclosure in R&D-intensive industries”. Contemporary Accounting Research 24, pp. 489–522.
Keefe, S.E. (1984). “Real and ideal extended familism among Mexican Americans and Anglo
Americans: On the meaning of" close" family ties”. Human organization 43, pp. 65–70.
Manski, C. F. (1993). “Identification of endogenous social effects: The reflection problem”.
The Review of Economic Studies 60, pp. 531–542.
— (2000). “Economic analysis of social interactions”. The Journal of Economic Perspectives
14, pp. 115–136.
Mayer, C. J., E. Morrison, T. Piskorski, and A. Gupta (2013). “Mortgage modification and
strategic behavior: evidence from a legal settlement with Countrywide”. Columbia University, Working paper.
Munroe, D. J. and L. Wilse-Samson (2013). “Foreclosure contagion: Measurement and mechanisms”. Columbia University, Working paper.
37
Pastor, L. and P. Veronesi (2009). “Learning in financial markets”. Annual Review of Financial Economics 1, pp. 361–381.
Pena, A. (2008). “Unregulated financial intermediation in Mexico-the new Sociedad Financiera De Objeto Multiple or SOFOM”. Law and Business Review of the Americas
14, p. 169.
Piskorski, T., A. Seru, and V. Vig (2010). “Securitization and distressed loan renegotiation: Evidence from the subprime mortgage crisis”. Journal of Financial Economics 97,
pp. 369–397.
Posner, E. A. and L. Zingales (2009). “A loan modification approach to the housing crisis”.
American Law and Economics Review 11, pp. 575–607.
Rodriguez, N., C. B. Mira, N. D. Paez, and H. F. Myers (2007). “Exploring the complexities
of familism and acculturation: Central constructs for people of Mexican origin”. American
Journal of Community Psychology 39, pp. 61–77.
Shue, K. (2013). “Executive networks and firm policies: Evidence from the random assignment of MBA peers”. Review of Financial Studies 26, pp. 1401–1442.
Stock, J. H., J. H. Wright, and M. Yogo (2002). “A survey of weak instruments and weak identification in generalized method of moments”. Journal of Business & Economic Statistics
20, pp. 518–529.
Towe, C. and C. Lawley (2013). “The contagion effect of neighboring foreclosures”. American
Economic Journal: Economic Policy 5, pp. 313–335.
Verrecchia, R. E. (1983). “Discretionary disclosure”. Journal of Accounting and Economics
5, pp. 179–194.
Verrecchia, R. E. and J. Weber (2006). “Redacted disclosure”. Journal of Accounting Research 44, pp. 791–814.
Wagenhofer, A. (1990). “Voluntary disclosure with a strategic opponent”. Journal of Accounting and Economics 12, pp. 341–363.
38
Figure 1: Identification Strategy
This figure presents a graphical exposition of the identification strategy. Borrowers A and B live in the
same tight-knit neighborhood. They commute to their respective working areas where they each run a small
independent business. Analogously, Borrowers C and D live in another neighborhood and they also commute
to the respective locations of their small businesses. I define strategic contagion in the following way: suppose
Borrower A and Borrower C default on their small consumer loans, because of an economic shock that affects
the whole work area. The lender offers Borrower A debt forgiveness as a means of getting him out of default
and back into repayment. Borrower A subsequently communicates with his neighbor, Borrower B, regarding
the lender’s forgiveness policy. Borrower B then chooses to strategically default with the objective of getting
a portion of his loan forgiven. Borrower D does not default, because he neither experiences an economic
shock nor has any interaction with Borrowers A or B.
39
Figure 2: Frequency and Total Amount of Debt Forgiveness over Time
These figures present the frequency and total amount of debt forgiveness that FA has granted over time.
Panel A plots the number of loans that received debt forgiveness during a given month. Panel B plots the
total amount of debt forgiveness in thousands of Mexican pesos (MXN) in a given month. In both graphs,
the vertical line is located on September 2012, the month when FA changed its debt-forgiveness policy.
Panel A: Frequency of forgiveness
Panel B: Total forgiveness amount
40
Figure 3: Branch Learning
This figure presents the time structure of the learning analysis conducted in section 6. Each of the three
learning periods begins on September 2012, the date that FA changed its debt-forgiveness policy. Learning
period 1 consists of three months after the policy change and the corresponding testing period begins thereafter. Learning period 2 consists of six months after the policy change and the corresponding testing period
begins thereafter. Learning period 3 consists of nine months after the policy change and the corresponding
testing period begins thereafter.
41
Table 1: Descriptive Statistics
This table presents the summary statistics for the loan sample. The loans in the sample were originated in
54 branches from January 2011 through March 2014. Panel A reports the borrower and loan characteristics
at origination for 14,649 loans. Panel B reports the performance of the loans in 110,649 loan months. Panel
C reports summary statistics at the branch level. Variable definitions are provided in Appendix Table A.1.
Panel A: Borrower and loan characteristics
Variable
N
Mean
Age
14,649 39.22
Gender
14,649
0.43
Education
14,649
2.69
Credit Score
9,412 691.87
Internal Credit Score
14,649 601.20
Lif e at Home
14,649
0.67
Lif e at Occupation
14,649
0.29
Loan/Income
14,649
0.85
P ayment f requency
14,649
2.92
Repeat borrower
14,649
0.38
Commuting distance
14,649
3.80
Std. Dev.
Q1
Median
12.24
29.00
38.00
0.50
0.00
0.00
0.75
2.00
3.00
33.62
671.00 696.00
20.02
587.00 604.00
0.31
0.39
0.69
0.16
0.17
0.27
0.37
0.62
0.80
0.31
3.00
3.00
0.48
0.00
0.00
4.38
1.25
2.38
Panel B: Loan-performance
Variable
Def ault
Def ault Ratework,t
Def ault Ratehome,t
F orgiveness Ratehome,t−1
Mean
0.16
0.29
0.29
0.20
Std. Dev.
0.36
0.11
0.12
0.12
Q1
0.00
0.23
0.23
0.13
Median
0.00
0.29
0.30
0.20
Q3
0.00
0.36
0.36
0.28
Mean
0.26
0.06
0.19
0.67
0.54
0.10
Std. Dev.
0.10
0.05
0.13
0.11
0.19
0.05
Q1
0.18
0.03
0.12
0.59
0.42
0.05
Median
0.25
0.04
0.14
0.67
0.51
0.09
Q3
0.33
0.07
0.21
0.75
0.62
0.13
statistics
N
110,649
110,649
110,649
110,617
Panel C: Branch-level statistics
Variable
N
Ref erral Rate
54
Concentration
54
Adj. Concentration
54
F raction of Lif e
54
Likelihood of F orgiveness
54
Decision to Reject
54
42
Q3
48.00
1.00
3.00
715.00
615.00
1.00
0.40
1.08
3.00
1.00
4.86
Table 2: Effect of Debt Forgiveness on Defaults
This table reports the estimation results from regressions of the following form:
Def aulti,t =α0 +α1 F orgiveness Ratehome,t−1 +Controls + T ime F.E. In this model, the dependent
variable is binary and takes the value of one if a loan i is in default at month t, but not at month t − 1, and
zero otherwise. The main explanatory variable is the forgiveness rate at month t − 1 in the zip code where
borrower i lives. Panel A reports a linear regression analysis. Panel B partitions the analysis at the median
of Fraction of Life at Current Home. Variable definitions are provided in Appendix Table A.1. Standard
errors are clustered at the branch level and are reported below the coefficient estimates. *, **, and ***
indicate significance at the two-tailed 10%, 5%, and 1% levels, respectively.
Panel A: Linear Regression Analysis
Variables
F orgiveness Ratehome,t−1
Predicted Sign
+
Def ault Ratework,t
+
Internal Credit Score
-
Age
?
Gender
?
Education
-
Credit Score
-
Lif e at Home
?
Lif e at Occupation
-
Loan/Income
+
P ct P aid
?
P ct P aid2
?
P ayment F requency
+
F orgiven in the P ast
+
N ot F orgiven
+
Repeat Borrower
-
F orgiveness Rate × Repeat Borrower
?
Month fixed effects
Industry fixed effects
Work zip code - month fixed effects
Home zip code fixed effects
Prior forgiveness
Repeat borrowers
Adj. R2
Observations
43
(1)
0.087∗∗∗
(0.023)
0.129∗∗∗
(0.031)
(2)
0.074∗∗∗
(0.019)
0.129∗∗∗
(0.021)
-0.001∗∗∗
(0.000)
(3)
0.053∗∗∗
(0.019)
0.122∗∗∗
(0.019)
-0.001∗∗∗
(0.000)
-0.002∗∗∗
(0.000)
-0.008
(0.006)
-0.002
(0.004)
-0.001∗∗∗
(0.000)
-0.018∗
(0.010)
-0.014
(0.015)
0.037∗∗∗
(0.008)
0.446∗∗∗
(0.034)
-0.337∗∗∗
(0.053)
0.035∗∗∗
(0.008)
0.051∗∗∗
(0.009)
0.458∗∗∗
(0.025)
-0.371∗∗∗
(0.035)
0.037∗∗∗
(0.007)
0.051∗∗∗
(0.008)
0.264∗∗∗
(0.024)
-0.305∗∗∗
(0.032)
0.029∗∗∗
(0.005)
0.317∗∗∗
(0.016)
0.174∗∗∗
(0.008)
Yes
Yes
No
No
No
No
0.035
32,415
Yes
Yes
No
No
No
No
0.026
64,222
Yes
Yes
No
No
Yes
No
0.068
66,339
(4)
0.095∗∗∗
(0.027)
(5)
0.086∗∗
(0.028)
-0.001∗∗∗
(0.000)
-0.001∗∗∗
(0.000)
0.036∗∗∗
(0.007)
0.227∗∗∗
(0.028)
-0.256∗∗∗
(0.037)
0.025∗∗∗
(0.006)
0.281∗∗∗
(0.014)
0.167∗∗∗
(0.007)
-0.043∗∗∗
(0.008)
-0.018
(0.036)
No
Yes
Yes
No
Yes
Yes
0.170
107,687
0.055∗∗∗
(0.008)
0.316∗∗∗
(0.027)
-0.293∗∗∗
(0.035)
0.026∗∗∗
(0.008)
0.241∗∗∗
(0.013)
0.134∗∗∗
(0.007)
-0.038∗∗∗
(0.011)
-0.013
(0.042)
No
Yes
Yes
Yes
Yes
Yes
0.179
107,687
Panel B: Linear regression analysis partitioned at the median of Lif e at Home
Lif e at Home
Variables
F orgiveness Ratehome,t−1
Predicted Sign
+
Def ault Ratework,t
+
Internal Credit Score
-
Loan/Income
+
P ct P aid
?
P ct P aid2
?
P ayment F requency
+
F orgiven in the P ast
+
N ot F orgiven
+
Repeat Borrower
-
F orgiveness Rate × Repeat Borrower
?
Month fixed effects
Industry fixed effects
Work zip code - month fixed effects
Home zip code fixed effects
Prior forgiveness
Repeat borrowers
Adj.R2
Observations
(1)
0.052∗
(0.029)
0.134∗∗∗
(0.031)
-0.001∗∗∗
(0.000)
0.043∗∗∗
(0.009)
0.453∗∗∗
(0.029)
-0.387∗∗∗
(0.037)
0.033∗∗∗
(0.008)
Yes
Yes
No
No
No
No
0.027
28,157
44
Low
(2)
0.036
(0.028)
0.127∗∗∗
(0.028)
-0.001∗∗∗
(0.000)
0.045∗∗∗
(0.009)
0.321∗∗∗
(0.028)
-0.348∗∗∗
(0.036)
0.028∗∗∗
(0.007)
0.304∗∗∗
(0.019)
0.128∗∗∗
(0.007)
Yes
Yes
No
No
Yes
No
0.057
29,056
(3)
0.065
(0.057)
-0.001∗∗∗
(0.000)
0.048∗∗∗
(0.013)
0.391∗∗∗
(0.044)
-0.335∗∗∗
(0.058)
0.024∗
(0.013)
0.209∗∗∗
(0.025)
0.081∗∗∗
(0.010)
-0.031∗∗
(0.014)
-0.035
(0.053)
No
Yes
Yes
Yes
Yes
Yes
0.096
46,497
(1)
0.090∗∗∗
(0.026)
0.148∗∗∗
(0.022)
-0.001∗∗∗
(0.000)
0.069∗∗∗
(0.010)
0.466∗∗∗
(0.030)
-0.351∗∗∗
(0.044)
0.037∗∗∗
(0.009)
Yes
Yes
No
No
No
No
0.028
28,157
High
(2)
0.079∗∗∗
(0.023)
0.141∗∗∗
(0.021)
-0.001∗∗∗
(0.000)
0.066∗∗∗
(0.009)
0.296∗∗∗
(0.025)
-0.285∗∗∗
(0.037)
0.033∗∗∗
(0.007)
0.281∗∗∗
(0.015)
0.137∗∗∗
(0.007)
Yes
Yes
No
No
Yes
No
0.057
29,056
(3)
0.105∗
(0.060)
-0.001∗∗∗
(0.000)
0.093∗∗∗
(0.013)
0.428∗∗∗
(0.047)
-0.355∗∗∗
(0.062)
0.051∗∗∗
(0.016)
0.187∗∗∗
(0.023)
0.074∗∗∗
(0.014)
-0.042∗
(0.022)
0.008
(0.099)
No
Yes
Yes
Yes
Yes
Yes
0.081
46,498
Table 3: Instrumental Variables Regressions
This table reports the coefficients from the instrumental variables regressions.
Panel A reports
the estimation results of the second-stage regression of the following form: Def aulti,t = α0 +
d Ratehome,t−1 )+Controls. In this model, the dependent variable is binary and takes
α1 (F orgiveness
the value of one if loan i is in default at month t, but not at montht − 1, and zero otherwise.
The main explanatory variable is estimated from the following first-stage regression:
F orgivenessRatehome,t−1 = γ0 + γ1 N ew Def ault Ratework(−i),t−2 + Controls, reported in Panel B. The
instrument, N ew Def ault Ratework(−i),t−2 , is defined as the fraction of borrowers that become delinquent
in t − 2 and satisfy the following two conditions: (i) they live in the same home neighborhood as borrower
i, and (ii) they do not work in the same area as borrower i. Robust standard errors are reported below
the coefficient estimates. *, **, and *** indicate significance at the two-tailed 10%, 5%, and 1% levels,
respectively.
Panel A: Second-stage regressions
Variables
F orgiveness Ratehome,t−1
Predicted Sign
+
Def ault Ratework,t
+
Internal Credit Score
-
Loan/Income
+
P ct P aid
?
P ct P aid2
?
P ayment F requency
+
Repeat Borrower
-
F orgiveness Rate × Repeat Borrower
?
F orgiven in the P ast
+
N ot F orgiven
+
F orgiveness Rate × F orgiven in the P ast
-
Month fixed effects
Industry fixed effects
Adj.R2
Observations
Panel B: First stage regressions
Instrument
N ew Def ault Ratework(−i),t−2
Predicted Sign
+
Second stage regressors
Angrist-Pischke partial R2
Angrist-PischkeF -stat
45
(1)
2SLS
0.214∗∗
(0.108)
0.130∗∗∗
(0.017)
-0.001∗∗∗
(0.000)
0.063∗∗∗
(0.006)
0.289∗∗∗
(0.024)
-0.181∗∗∗
(0.033)
0.040∗∗∗
(0.006)
(2)
2SLS
0.264∗∗∗
(0.100)
0.116∗∗∗
(0.013)
-0.001∗∗∗
(0.000)
0.039∗∗∗
(0.004)
0.306∗∗∗
(0.017)
-0.152∗∗∗
(0.024)
0.031∗∗∗
(0.004)
-0.078∗∗∗
(0.030)
0.095
(0.138)
(3)
2SLS
0.248∗∗∗
(0.080)
0.124∗∗∗
(0.012)
-0.001∗∗∗
(0.000)
0.021∗∗∗
(0.004)
0.141∗∗∗
(0.018)
-0.175∗∗∗
(0.026)
0.027∗∗∗
(0.003)
Yes
Yes
0.015
50,817
Yes
Yes
0.026
91,462
0.349∗∗
(0.144)
0.174∗∗∗
(0.004)
-0.222
(0.557)
Yes
Yes
0.062
91,462
(1)
0.198∗∗∗
(0.008)
Yes
0.023
545.03
(2)
0.199∗∗∗
(0.008)
Yes
0.023
888.44
(3)
0.193∗∗∗
(0.006)
Yes
0.022
883.34
(4)
2SLS
0.182∗
(0.099)
0.110∗∗∗
(0.012)
-0.001∗∗∗
(0.000)
0.039∗∗∗
(0.004)
0.106∗∗∗
(0.018)
-0.135∗∗∗
(0.027)
0.022∗∗∗
(0.003)
-0.088∗∗∗
(0.029)
0.143
(0.136)
0.358∗∗
(0.144)
0.173∗∗∗
(0.004)
-0.248
(0.558)
Yes
Yes
0.067
91,462
(4)
0.198∗∗∗
(0.008)
Yes
0.022
884.60
Table 4: Robustness Analysis and Alternative Hypotheses
This table reports the estimation results from regressions of the following form: Def aulti,t = α0 +
α1 F orgiveness Ratehome,t−1 + Controls + T ime F.E. In this model, the dependent variable is binary and
takes the value of one if loan i is in default at month t, but not at month t − 1, and zero otherwise. The main
explanatory variable is the forgiveness rate at month t − 1 in the zip code where borrower i lives. Columns
1-6 vary in the controls included in the regression. Variable definitions are provided in Appendix Table A.1.
Standard errors are clustered at the branch level and are reported below the coefficient estimates. *, **, and
*** indicate significance at the two-tailed 10%, 5%, and 1% levels, respectively.
F orgiveness Ratehome,t−1
Predicted Sign
+
Def ault Ratework,t
+
Def ault Ratehome,t
+
Inf lation in F ood P rices
+
Inf lation in T ransportation
+
Inf lation in Clothes
+
Inf lation in Housing
+
N atural Disaster
+
F orgiveness Rate × N at. Disaster
?
M ultiple Accounts
-
F orgiveness Rate × M ultiple Acc.
?
W inter
+
F orgiveness Rate × W inter
+
W almart N ear W ork
+
F orgiveness Rate × W almart W ork
-
W almart N ear Home
-
F orgiveness Rate × W almart Home
?
Borrower and loan characteristics
Month fixed effects
Industry fixed effects
Work zip code - month fixed effects
Adj.R2
Observations
(1)
0.049∗∗
(0.019)
0.126∗∗∗
(0.015)
0.095∗∗∗
(0.014)
(2)
0.098∗∗∗
(0.031)
(3)
0.084∗∗∗
(0.024)
0.143∗∗∗
(0.016)
(4)
0.077∗∗∗
(0.018)
0.145∗∗∗
(0.015)
(5)
0.078∗∗∗
(0.019)
0.146∗∗∗
(0.015)
Placebo
(6)
-0.004
(0.016)
0.140∗∗∗
(0.003)
0.053
(0.053)
-0.089
(0.239)
0.026
(0.051)
-0.077
(0.233)
Yes
Yes
Yes
No
0.024
89,943
Yes
Yes
Yes
No
0.025
89,943
0.021
(0.023)
0.019
(0.015)
0.072∗∗∗
(0.013)
0.149∗∗∗
(0.032)
0.074∗∗∗
(0.016)
0.001
(0.007)
-0.009
(0.026)
-0.031∗∗∗
(0.005)
-0.004
(0.024)
-0.051∗∗∗
(0.013)
0.002
(0.025)
Yes
Yes
Yes
No
0.025
89,943
46
Yes
No
Yes
Yes
0.045
89,943
Yes
Yes
Yes
No
0.026
89,943
Yes
Yes
Yes
No
0.024
89,943
Table 5: Loan-Performance Transition Matrices, pre and post September 2012
This table presents loan-performance transition matrices. In all panels, the rows indicate the delinquency
status of the loan at month t, ranging from non-delinquent to over 90 days delinquent. The columns indicate
the status of the loan at month t + 1, ranging from non-delinquent to over 90 days delinquent. In Panel A,
the transition probabilities are calculated for the six-month period before the change in forgiveness policy. In
Panel B, the transition probabilities are calculated for the six-month period as of the change in forgiveness
policy. Panel C reports the difference between Panel B and Panel A, with *, **, and *** indicating significance
at the 10%, 5%, and 1% levels, respectively.
Panel A: March 2012 to August 2012
Fraction of
loans
Status at
month t
Non-delinquent
1-29 days
30-59 days
60-89 days
90+ days
58.9%
29.9%
3.7%
3.0%
4.5%
Status at month t + 1
Non-delinquent 1-29 days
79.9%
18.7%
1.9%
0.3%
0.0%
20.1%
69.0%
8.8%
0.8%
0.1%
30-59 days
60-89 days
90+ days
0.0%
12.3%
3.8%
0.3%
0.0%
0.0%
0.0%
85.5%
1.6%
0.1%
0.0%
0.0%
0.0%
97.1%
99.8%
30-59 days
60-89 days
90+ days
0.0%
11.6%
2.2%
0.1%
0.0%
0.0%
0.0%
81.1%
1.9%
0.0%
0.0%
0.0%
0.0%
96.6%
99.7%
30-59 days
60-89 days
90+ days
0.00%
-0.68%∗∗∗
-1.64%∗∗∗
-0.17%∗∗
-0.02%
0.00%
0.00%
-4.45%∗∗∗
0.33%
-0.03%
0.00%
0.00%
0.00%
-0.47%
-0.13%
Panel B: September 2012 to February 2013
Fraction of
loans
Status at
month t
Non-delinquent
1-29 days
30-59 days
60-89 days
90+ days
52.7%
36.8%
4.4%
3.3%
2.9%
Status at month t + 1
Non-delinquent 1-29 days
77.6%
18.4%
5.0%
0.7%
0.2%
22.4%
70.0%
11.8%
0.7%
0.2%
Panel C: Changes in delinquency after September 2012 (Panel B − Panel A)
Fraction of
loans
Status at
month t
Non-delinquent
1-29 days
30-59 days
60-89 days
90+ days
∗∗∗
-6.19%
6.86%∗∗∗
0.65%
0.29%
-1.62%∗∗
Status at month t + 1
Non-delinquent 1-29 days
∗∗∗
-2.24%
-0.35%
3.13%∗∗∗
0.39%∗∗∗
0.12%∗∗
47
∗∗∗
2.24%
1.03%∗∗∗
2.96%∗∗∗
-0.08%
0.06%
Table 6: Ex-ante Measures of Contagion
This table reports the estimation results from the following regression: Strategic Def aultj =
β0 + β1 Ref erral RateAug0 12,j + β2 ConcentrationAug0 12,j + β3 Adjusted ConcentrationAug0 12,l +
β4 F raction of Lif eAug0 12,j .
In this model, the dependent variable is computed as α
d
1,j ×
F orgiveness Ratej during the three-month period after the change in forgiveness policy. α
d
1,j is estimated by running the strategic default regression, Def aultj,t = α0 + α1 F orgiveness Ratehome,t−1 +
Controls + T ime F.E., for branch j. Variable definitions are provided in Appendix Table A.1. Simulations
are used to correct the standard errors for the presence of a generated regressor. P-values are reported
below the coefficient estimates. *, **, and *** indicate significance at the two-tailed 10%, 5%, and 1%
levels, respectively.
Variable
Predicted Sign
Constant
Ref erral Rate
+
Concentration
+
Adjusted Concentration
+
F raction of Lif e
+
(1)
(2)
-0.073∗ 0.009∗ *
(0.061) (0.029)
0.316∗
(0.079)
0.044
(0.457)
(3)
0.030
54
48
0.002
54
(5)
0.0021 -0.110 -0.154∗
(0.163) (0.213) (0.086)
0.180∗
(0.091)
0.078
(0.189)
Adj. R2
Observations
(4)
0.008
54
0.052
(0.327)
0.192
0.172
(0.161) (0.216)
0.005
54
0.038
54
Table 7: Bank’s Adjustments to the Forgiveness Policy
This table reports the estimation results from the following regression: Likelihood of F orgivenessj = γ0 +
γ1 Strategic Def aultj + γ2 T otal Def aultj . In this model, the dependent variable is computed as the average
likelihood that a defaulter in branch j receives forgiveness. In columns 1 and 2, the analysis is performed
over a three-month learning period, in columns 3 and 4, a six-month learning period, and in columns 5 and
6, a nine-month learning period. In all Columns, the main explanatory variable is the level of strategic
default among borrowers serviced by branch j. Strategic default is measured as α
d
1,j × F orgiveness Ratej ,
estimated from: Def aultj,t = α0 + α1 F orgiveness Ratehome,t−1 + Controls + T ime F.E., columns 2, 4, and
6 include the level of total defaults in each branch as an additional control. Simulations are used to correct
the standard errors for the presence of a generated regressor. P-values are reported below the coefficient
estimates. *, **, and *** indicate significance at the two-tailed 10%, 5%, and 1% levels, respectively.
Learning period:
Prediction with learning:
Constant
Strategic Def ault
3 months
9 months
(1)
(2)
(3)
(4)
(5)
(6)
(-)
0.538∗∗∗
(<.001)
-0.165
(0.577)
(-)
1.287∗∗∗
(<.001)
-0.191
(0.612)
-7.475∗∗
(0.010)
(-)
0.509∗∗∗
(<.001)
-1.036
(0.103)
(-)
1.038∗∗∗
(<.001)
-0.891
(0.136)
-5.388∗∗
(0.034)
(-)
0.519∗∗∗
(<.001)
-1.397∗
(0.069)
(-)
1.011∗∗∗
(<.001)
-1.212∗
(0.086)
-5.191∗∗
(0.018)
1.08%
54
9.82%
54
2.15%
54
10.69%
54
2.43%
54
12.91%
54
-0.871
(0.903)
-0.700
(0.709)
-0.361
(0.305)
-0.321
(0.281)
T otal Def ault
R2
Observations:
6 months
Three-month difference:
t-test
49
Table 8: Bank’s Adjustments to the Origination Policy
This table reports the estimation results from the following regression: Decision to Rejectj = γ0 +
γ1 Strategic Def aultj + γ2 T otal Def aultj . In this model, the dependent variable is computed as the
number of loans rejected by branch j as a fraction of the number of loans that met FA origination
requirements. In columns 1 and 2, the analysis is performed over a three-month learning period, in
columns 3 and 4, a six-month learning period, and in columns 5 and 6, a nine-month learning period. In all columns, the main explanatory variable is the level of strategic default among borrowers serviced by branch j. Strategic default is measured as α
d
1,j × F orgiveness Ratej , estimated from:
Def aultj,t = α0 + α1 F orgiveness Ratehome,t−1 + Controls + T ime F.E., columns 2, 4, and 6 include the
level of total defaults in each branch as an additional control. Simulations are used to correct the standard
errors for the presence of a generated regressor. P-values are reported below the coefficient estimates. *, **,
and *** indicate significance at the two-tailed 10%, 5%, and 1% levels, respectively.
Learning period:
3 months
(1)
Prediction with learning:
Constant
Strategic Def ault
T otal Def ault
R2
Observations:
6 months
(2)
(+)
(+)
∗
0.096
0.153∗
(0.070) (0.070)
0.278
0.260
(0.357) (0.386)
-0.571
(0.492)
3.30%
54
3.93%
54
Three-month difference:
t-test
50
9 months
(3)
(4)
(5)
(6)
(+)
0.086∗∗∗
(<.001)
0.505
(0.227)
(+)
0.179∗∗
(0.013)
0.544
(0.188)
-0.945
(0.185)
(+)
0.076∗∗∗
(<.001)
0.799∗
(0.095)
(+)
0.150∗∗∗
(<.001)
0.843∗
(0.098)
-0.783
(0.175)
5.11%
54
8.56%
54
9.71%
54
13.28%
54
0.227
(0.268)
0.284
(0.336)
0.294
(0.310)
0.299
(0.311)
Table 9: Robustness Analysis of FA’s Learning
Panel A of this table reports the estimation results from the following regressions:
Likelihood of F orgivenessj = γ0 +γ1 Strategic Def aultj +γ2 T otal Def aultj and Decision to Rejectj = γ0 +
γ1 Strategic Def aultj + γ2 T otal Def aultj .The analysis is performed over a nine-month learning period that
starts on three placebo dates: February 2010, July 2010, and December 2010. Panel B reports the estimation
results from the following regressions: Ch. in Likelihood of F orgivenessj = γ0 + γ1 Strategic Def aultj +
γ2 T otal Def aultj and Ch. in Decision to Rejectj = γ0 + γ1 Strategic Def aultj + γ2 T otal Def aultj .
In all Columns, the main explanatory variable is the level of strategic default among borrowers
serviced by branch j. Strategic default is measured as α
d
1,j × F orgiveness Ratej , estimated from:
Def aultj,t = α0 + α1 F orgiveness Ratehome,t−1 + Controls + T ime F.E. Simulations are used to correct
the standard errors for the presence of a generated regressor. P-values are reported below the coefficient
estimates. *, **, and *** indicate significance at the two-tailed 10%, 5%, and 1% levels, respectively.
Panel A: Placebo changes to FA’s Forgiveness Policy
Dependent variable:
Likelihood of F orgiveness
Month of placebo change:
February 2010
(1)
Prediction with learning:
Constant
(-)
1.044∗∗∗
(<.001)
-0.119
(0.284)
-6.361∗∗∗
(0.003)
(-)
1.246∗∗∗
(<.001)
1.052
(0.131)
-7.900∗∗∗
(<.001)
(-)
1.400∗∗∗
<.001)
0.822
(0.161)
-8.789∗∗∗
(0.002)
(+)
0.130
(0.229)
0.124
(0.531)
0.121
(0.336)
(+)
0.126
(0.246)
0.627
(0.257)
0.030
(0.900)
(+)
0.135
(0.148)
0.353
(0.280)
-0.279
(0.757)
16.31%
54
21.7%
54
18.45%
54
1.51%
54
4.61%
54
3.57%
54
Strategic Def ault
T otal Def ault
R2
Observations:
July 2010 December 2010
(2)
(3)
Decision to Reject
February 2010
(4)
July 2010 December 2010
(5)
(6)
Panel B: Changes in likelihood of forgiveness and decision to reject
Dependent variable:
Learning period:
Prediction with learning:
Constant
Strategic Def ault
T otal Def ault
R2
Observations:
Three-month difference:
t-test
Ch. in Likelihood of F orgiveness
Ch. in Decision to Reject
3 months
(1)
6 months
(2)
9 months
(3)
3 months
(4)
6 months
(5)
9 months
(6)
(-)
1.278∗∗∗
(<.001)
-1.242
(0.114)
-8.079∗∗
(0.019)
13.47%
54
(-)
0.976∗∗∗
(0.002)
-0.800
(0.252)
-5.491∗
(0.083)
7.70%
54
(-)
0.974∗∗∗
(<.001)
-1.640
(0.102)
-5.553∗
(0.057)
9.88%
54
(+)
-0.219∗∗∗
(<.001)
0.140
(0.339)
1.013∗∗∗
(<.001)
1.24%
54
(+)
-0.178∗∗∗
(<.001)
0.464
(0.169)
0.508∗∗∗
(<.001)
1.21%
54
(+)
-0.246∗∗∗
(<.001)
0.792∗
(0.074)
0.526∗∗∗
(0.001)
2.25%
54
0.442
(0.363)
-0.840
(0.644)
0.324
(0.466)
0.328
(0.371)
51