Content uploaded by Carlos Madeira
Author content
All content in this area was uploaded by Carlos Madeira on Apr 01, 2015
Content may be subject to copyright.
Adverse selection, loan choice and default in the Chilean
consumer debt market
Carlos Madeira
Draft
October 2014
Abstract
Using household survey data I estimate a model of consumer loan choice and default behavior.
I show that households are sorted into di¤erent lenders, with higher income and education
being positively associated with choosing Bank loans and negatively associated with other
lenders. Debt amounts increase with income, unemployment risk and wage volatility, providing
evidence that consumer debt is used to smooth income shocks. Also, debt amounts increase
with household size and are quadratic in age, resembling the life-cycle consumption pro…le.
Default behavior decreases with income and increases with higher indebtedness, unemployment
and wage risk, con…rming the role of adverse selection.
JEL Classi…cation: E21; E24; E32; E51; G01; G21.
Keywords: Consumer credit; Default risk; Unemployment.
carlosmadeira2009@u.northwestern.edu. All errors are my own.
1
1 Introduction
Household debt increased consistently in the last decades, both in emerging economies (IMF, 2006)
and developed countries (Girouard, Kennedy, André, 2007). This evolution in the quantity of
household credit coincided with a period of strong …nancial innovation, with a great range of loan
products being available to consumers. Consumers are able to access credit from a variety of sources,
such as credit cards, auto loans, education loans, and for motives as diverse as health, vacations,
purchase of durable goods, or a renegotiation of previous debts. Also, the technological evolution
has allowed lenders to process larger and better databases on the characteristics of debtors, allowing
for an increased use of credit scoring and an heterogeneity of loan terms for each loan applicant
(Roszbach, 2004, Edelberg, 2006, Einav, Jenkins and Levin, 2012). Yet despite an increasing
availability of consumer credit, several families are still unable to access credit markets or obtain
lower loan amounts than desired (Adams, Einav and Levin, 2009). The factors behind consumers’
loan choice and their credit constraints have been documented in recent studies for the United
States (see Dynan and Kohn, 2007, Attanasio, Goldberg and Kyriazidou, 2008, Adams, Einav and
Levin, 2009); however, loan choice in developing countries is still understudied.
This paper studies the consumer credit access, lender choice and repayment behavior of families
in Chile. Consumer loans are particularly relevant in Chile, since around 60% of the households have
some consumer debt. Using data from the Chilean Household Finance Survey (EFH), I estimate
an econometric model in which families choose among a variety of lender types according to their
earnings, labor risk, demographics, and unobserved preferences. I …nd that families are sorted
into di¤erent lenders according to their labor market risk. Furthermore, household’s debt levels,
income, and labor market risk have a signi…cant impact on default behavior, which shows evidence
of adverse selection and moral hazard (Einav, Jenkins and Levin, 2012) in Chilean loan markets.
In Chile the market for consumer loans has several di¤erent providers and their credit o¤ers
represent imperfect substitutes for consumers. These loan providers access di¤erent customer
lists and information, besides being subject to di¤erent legal regulations, which a¤ects their loan
terms and the ability to target speci…c markets (see Marinovic, Matus, Flores and Silva, 2011,
for a review of the structure and legal framework of di¤erent credit providers). The industrial
organization literature argues that even small di¤erences across product providers - such as the
2
cost of screening applicants, brand preferences, marketing initiatives, search frictions, travel costs,
tied products and asymmetric information - can create substantial frictions for customers’decisions
(Nevo, 2011). Therefore it seems adequate to treat loan decisions as a di¤erentiated product model
where heterogeneous agents with unobserved preferences choose their preferred loan provider.
Household surveys are an ideal source of information for the study of the market of consumer
loans, since the study of problems such as adverse selection requires information on the heterogeneity
of agents. Aggregate data can hide the factors a¤ecting the decisions of di¤erent consumers. The
Chilean Household Finance Survey (EFH) collects detailed information on families’income, assets,
loans, debt repayment behavior, and demographic characteristics. From 2007 to 2011 the EFH
interviewed 12,264 families, which includes panel data of 2,739 families who were interviewed twice.
The model of loan choice and repayment behavior has three main components: i) a categorical
choice between having no debt, wanting debt but being credit constrained, and …ve di¤erent types
of lenders, ii) the choice of loan amount, and iii) a categorical outcome of whether the household
defaulted or not on at least one payment over the previous year. The …ve lender types in this
categorical model correspond to: Banks, Banks and Retail Stores, Retail Stores, Social Credit (i.e.,
loans provided by credit and labor unions), and Other Loans (which includes mainly auto loans,
educational debt, plus pawn shops and some informal lending). Banks and Retail Stores are the
two major lenders in Chile, therefore using both lenders is treated as a separate choice than the
option of using just one type of lender. Other types of lenders represent a small proportion of the
population and therefore I do not model the interaction of those lenders with other types of debts.
Furthermore, there are two options for the families that did not get a consumer loan, which are
"No Access to Debt" and "No wish to apply for Consumer Loans". The option of "No Access
to Debt" represents families with credit constraints. These are families who applied for consumer
loans but were denied credit, plus those who wished to apply for credit but did not do so because
they expected to be refused. "No wish for Loans" represents the outside option for all agents,
comprising the families who report not having consumer debt and no interest in applying for loans.
All three endogenous variables - the choice of type of lender, loan amount, and repayment
outcome - are a¤ected by both observable factors and unobserved preferences. The observable
factors include income, education, labor income risk, and demographic characteristics such as the
age of the household head and household size. Also, I consider that the choice of lender is a¤ected
3
by the motives behind the indebtedness and by unobservable preferences of the household. The
observable motives are broadly classi…ed in three categories: general consumption, the payment
of previous loans or debt consolidation, and health needs (Chatterjee et al., 2007, show that a
signi…cant part of unsecured debt in the USA is contracted for health reasons). Labor market
characteristics are measured in terms of three di¤erent variables for the working members of the
household: i) the unemployment rate, ii) the wage volatility, and iii) the replacement ratio of
income during an unemployment spell (the proportion of the working income that workers still
earn after losing a job). Also, as suggested by Shimer (2012), the unemployment rate can be
further decomposed in terms of two di¤erent mechanisms, the separation rate (the probability of
entering unemployment given that one had a job before) and the job …nding rate (how quickly
workers exit out of unemployment). All of these measures of households’labor market risk provide
interesting insights into the nature of their shocks over time.
Unobserved preferences include random e¤ects denoting the taste for each type of lender, a taste
for higher or lower loan amounts, and a propensity of each family to repay or not its loans. The
random-e¤ects for the taste of each choice are correlated, implying for instance that families with a
higher taste for certain lenders may have a higher propensity to default on their loans. Besides the
…xed unobserved tastes that are constant for each family, the model also takes into account there
are uncorrelated shocks for each time period, implying that agents’decisions may change over time
due to some unexplained shock. Assuming a parametric distribution for the unobserved terms, this
model can be estimated using Simulated Maximum Likelihood (Train, 2009).
The results show that households with di¤erent characteristics tend to sort themselves among
di¤erent lenders. Households with No desire for Consumer Loans have the lowest wage volatility and
the lowest unemployment and job separation rates among all groups. This result seems to con…rm
that consumer debt is related to smoothing income shocks, therefore households with few income
shocks have low demand for consumer debt. Banks are the institution that applies credit scoring
and customer speci…c interest rates on a wider basis, therefore Banks capture the households of
highest income and with the lowest unemployment rates among loan applicants. Also, households
with loans in Banks su¤er the lowest income falls during unemployment. Households with loans in
Bank plus Retail and Other debts have both the largest loans in absolute amount and the larger
ratios of debt relative to income. Households with No Access to Debt have the lowest income levels
4
and also su¤er the strongest income falls during unemployment spells.
Unemployment rates increase the probability of households opting for all loans, but their impact
is highest for the clients of Retail Stores, Social Credit and Other Loans. Wage volatility is more
strongly associated with households opting for Social Credit, Other Loans and No Access to Debt.
Loan amounts increase with income, unemployment risk and wage volatility, therefore it is possible
that households use consumer loans to smooth income shocks. The probability of default decreases
with income and increases with high levels of debt amount and debt service (debt service includes
both monthly amortization and interest payments) relative to income, unemployment risk and
wage volatility, which con…rms the existence of adverse selection and moral hazard among Chilean
debtors. Bank debtors have a signi…cantly lower probability of default even after controlling for
observable variables. Since banks resort more to credit scoring and risk-adjusted interest rates, then
one should expect banks to capture the customers with lowest risk (Edelberg, 2006), con…rming the
economic theory of equilibrium in loan markets with adverse selection (Ja¤ee and Russell, 1976,
Ja¤ee and Stiglitz, 1990). It is also interesting that Health needs are positively associated with
default behavior, which con…rms the predictions of economic models for health expenses that are
unpredictable and uninsurable for households (Chatterjee et al., 2007).
Finally, the probability of getting a loan and the choice of loan amount is increasing in the
number of household members and quadratic in age, …rst increasing with the age of the household
head and then falling in its later years. Therefore the demand for consumer debt has an age pro…le
that resembles the …ndings of life-cycle consumption in the literature (Attanasio and Weber, 2010).
In terms of unobservable factors, I …nd that households with higher income and education are less
heterogeneous in their tastes, and that their choice of loan amount is less persistent over time.
This paper is related to a recent and growing literature of empirical models of loan choice
and default behavior which measures the impact of observable risk factors and adverse selection
(Roszbach, 2004, Adams, Einav and Levin, 2009, Einav, Jenkins and Levin, 2012). It extends that
literature in three ways: i) it applies a similar framework for loan choice and default to a developing
economy such as Chile, ii) it introduces a wider range of loan options and unobserved preferences
by using tools from the applied product choice models in the …eld of industrial organization (Train,
2009, Nevo, 2000, 2011), and iii) it uses a more diverse characterization of labor income risk by
separating overall risk into di¤erent variables such as unemployment risk and wage volatility.
5
This paper is organized as follows. Section 2 describes the consumer credit environment in Chile
and the applied model of loan choice and default. Section 3 summarizes the Chilean Household
Finance Survey dataset (2007-2011) and the main characteristics of Chilean families. Section 4
describes the sorting of households across di¤erent types of lenders in terms of loan amount, income
and labor market risk. Section 5 presents the results of the joint model of lender choice, loan amount
and default. Finally, section 6 concludes with implications for policy and future research.
2 Credit environment and empirical model of consumer behavior
2.1 The structure of consumer loan providers in Chile
This section starts with a review of the structure of Chilean credit markets and the di¤erences
among lenders, whether caused by di¤erentiated product lines or by legal regulations (see Marinovic,
Matus, Flores and Silva, 2011, for a review). In Chile all lenders have public access to a commercial
registry of debtors who defaulted on payments1, however this public registry is limited only to agents
with negative histories and therefore lenders’information sets on the positive characteristics of loan
applicants di¤er substantially, implying agents’can have di¤erent relationships with each lender
(Ja¤ee and Stiglitz, 1990). Banks represent one type of consumer loan provider in Chile, as well
as in other countries. Chilean banks have access to a common credit registry with information on
all loan amounts and debt default within the banking system2, but they do not observe loans from
non-banking institutions. Banks also make a strong use of credit scoring, changing their loan o¤ers
according to agents’credit history, and may even tie their loan o¤ers to other banking products
that are signed by their customers, giving for instance preferential treatment to families that have
direct deposit of wages, automatic bill payment, mortgages and other …nancial accounts with them.
Retail stores are another kind of credit provider, with a strong brand image and their own credit
cards3, and which have access to their own private databases on customers’ loans and product
1See www.dicom.cl/.
2See the General Law of Banks of the Chilean Superintendency of Banks and Financial Institutions, www.sbif.cl.
3The norms for non-banking credit card providers are detailed in the Chapter III.J.1 of the Compendium of
Financial Norms of the Central Bank of Chile.
6
transactions. Another type of loan providers are credit unions (denoted as Savings and Loans’
Cooperatives4) and labor unions (denoted as Family Compensation Funds5) which are regulated
as providers of "social credit". By legislation all Chilean companies must register their workers in
one among several Family Compensation Funds, which provide social credit and other services to
their a¢ liates. These labor unions or Family Compensation Funds represent 67.6% of the aggregate
"social credit". Family Compensation Funds are chosen by each employer for all its workers and
therefore workers do not choose their institution directly. Social credit providers must o¤er the
same conditions to all of their a¢ liates, therefore they can change interest rates according to loan
size and maturity, but are unable to discriminate against characteristics of the debtors such as
their income. Also, Family Compensation Funds bene…t from being able to deduct loan payments
directly from their clients’wage payroll and therefore face little risk of default. Even in the case of
a debtor losing its job, its Family Fund is able to deduct a substantial payment from the worker’s
severance compensation and therefore the risk of default is limited even relative to unexpected
unemployment shocks. Finally, there are lenders with more speci…c goals, such as auto loans at car
dealers, education loans, pawn shops6, and consumer loans provided by insurance companies7.
In terms of the aggregate amount of consumer credit in Chile, banks represented 71.7% of the
total market, while social credit institutions represented 14.8% and retail stores 13.5%, respectively.8
However, market presence in terms of customers di¤ers from the aggregate loan amounts, since it is
estimated that there are around 3.5 million debtors with banking loans, while social institutions and
retail stores reach around 2.5 million and 7 million customers, respectively. Therefore retail stores
are actually the largest provider of small consumer loans and reach the widest number of customers.
Over the last half-decade the market size of each type of lender has di¤ered substantially. Consumer
loans in banks at the end of 2013 were 233% as large as their aggregate amount at the beginning of
4See the Chilean Government Department of Cooperatives, www.decoop.cl, the General Law of Cooperatives,
DFL 5 (2003), www.bcn.cl, and Chapter III.C.2 of the Compendium of Financial Norms of the Central Bank of Chile.
5These institutions are regulated by the Chilean Superintendency of Social Security. Each Family Compensation
Fund is associated with one of the …ve labor unions registered at the Confederation of Production and Trade. See
the General Statute of Family Compensation Funds, articles 29 to 31 of the Law N18.833 of 1989.
6See www.dicrep.cl.
7The regulation of credit by insurance companies is detailed in several norms of the Chilean Superintendency of
Assets and Insurance, such as norms NCG 152 of 2002, NCG 208 of 2007 and NCG 247 of 2009.
8The aggregate amount of other loans (such as automotive and informal lending) is not entirely known, since
credits of smaller and unregulated institutions do not need to be registered for statistical purposes.
7
2006 (Banco Central de Chile, 2013). Aggregate consumer credit by social institutions was 245%
as large in 2013 as in 2006, but credit by retail stores grew only by 57% in the same period.
2.2 A review of the economic theory on adverse selection and lending
In this section I develop a simple review of the main results from the theoretical models of loan
choice and credit constraints in equilibrium, along lines similar to Ja¤ee and Russell (1976), Ja¤ee
and Stiglitz (1990), plus some empirical works such as Edelberg (2004), Adams, Einav and Levin
(2009), and Einav, Jenkins and Levin (2012). For simplicity, let us think of two types of agent, one
of high risk (H) and one of low risk (L). Agents of high risk are willing to pay higher interest rates
for each loan amount, because either they are more credit constrained or they actually expect to
default in the future (and therefore avoid the payment due to the higher interest rate). If lenders
are unable to di¤erentiate the risk type of agents, then the zero pro…t curve of credit supply will
ask higher interest rates (i) from larger loan amounts (D), i.e. S(i; D)=0. This supply strategy
may lead di¤erent agents to reveal their type by choosing di¤erent contracts: i) a contract for low
risk agents with low amounts and low interest rates, ii) a contract for high risk agents with a higher
interest rate and a higher loan amount. Therefore the theory predicts that if lenders are unable
to di¤erentiate agents or apply credit scoring, then loan amounts will be positively correlated with
the risk type of agents. Figure 1.A makes a graphical description of this adverse selection case.
The information technology evolution in the last 3 decades has made it easier to di¤erentiate
loan o¤ers according to the characteristics and credit scores of applicants. The widespread use of
credit scoring makes it more realistic to assume there are di¤erent loan supply curves for agents
with observable low risk (L) and high risk (H) types. The lenders’credit supply curve will o¤er
lower interest rates (i) at each debt amount (D) for low risk types, S(i; D jL)=0, relative to the
high risk type isopro…t curve, S(i; D jH) = 0. Therefore the theory predicts that lenders able to
use credit scoring will o¤er larger loan amounts and lower interest rates to debtors of a low risk
type. This case of adverse selection with observable characteristics is illustrated in Figure 1.B.
In general, it is more realistic to assume that there are both observable variables for agents’risk
types and unobservable characteristics which credit scoring models are unable to include. Therefore
economic theory should predict that better loan conditions (such as larger loan amounts, longer
8
Figure 1: Theory of self-selection of debtors across di¤erent loan contracts
maturities for payment, and lower interest rates) are associated with observable characteristics of
lower risk (such as higher income and more secure jobs). However, unobservable risk characteristics
(such as a taste for higher loan amounts) may create adverse selection and will be associated with
larger loan demand and more frequent default. The standard theory of adverse selection and credit
market equilibrium predicts that some types of high risk agents will not be a pro…table market
segment, due to either legal restrictions (such as usury laws and interest rate ceilings) or …xed costs
for loan evaluation. For these high risk agents, lenders will be unable to o¤er pro…table loans,
therefore these agents will remain outside the credit market and will be credit constrained.
In summary, according to the economic theory of loan markets we should expect three results:
i) lenders will o¤er better and larger loans to agents with observable characteristics of low risk, ii)
unobservable characteristics of high risk will still be associated with both larger loan amounts and
default, and iii) agents with very high risk will be credit constrained and without access to loans.
2.3 An empirical model of choice of lender, loan amount and debt default
The consumer choice model considers three endogenous variables: i) a categorical choice between
having no debt, wanting debt but being credit constrained, and …ve di¤erent types of loans, ii) the
choice of loan amount, and iii) a categorical outcome of whether the household defaulted or not
9
on at least one payment over the previous year. The …ve lender types in this categorical model
correspond to the major loan providers described in the previous section: Banks, Banks and Retail
Stores, Retail Stores, Social Credit, and Other Loans (which includes mainly auto loans, educational
debt, plus pawn shops and informal lending). It is possible that some consumers have more than
one debt type, say debt at Banks and Other Loans (for example, an educational loan), but except
for retail store credit (which reaches around 7 million people in Chile) there are few observations
with such interactions. For simplicity, I classify the observed lender choice of each household as the
one corresponding to the largest loan amount reported by each family. Banks and Retail Stores are
the two major lenders in Chile, therefore using both lenders is treated as a separate choice when
the household has a positive amount of loans with both lenders.
Families with no consumer loans are classi…ed in two categories: "No Access to Debt" and "No
wish to apply for Consumer Debt". "No Access to Debt" represents families with credit constraints,
including those who applied for credit but were denied and the ones who did not apply for credit
because they expected to be refused. "No wish for Debt" is the outside option for all agents,
comprising the families who report no consumer debt and no interest in applying for loans. To be
succinct, these options from now on will be denoted simply as "No Access" and "No Debt".
The modelling of a multivariate choice model with several options and many periods incurs into
a problem of multidimensionality, since with Ppossible products there are PTpossible choices in
a panel of Tperiods (Nevo, 2011). Therefore it is useful to apply a parsimonious model that can
summarize the choice among the di¤erent options in terms of a restricted number of observable
and unobservable factors. This is done in terms of a fully speci…ed maximum likelihood model.
Let Ui;b;t denote the utility of household ifrom the option bin period t, with b2 f1 "Bank",
2 "Bank & Retail", 3 "Retail", 4 "Social", 5 "Other Loans", 6 "No Access"g. Furthermore, let
us standardize the utility of the outside option, "No wish for Debt", as zero, Ui;0;t = 0. This
standardization is made without any loss of generality, since all that matters for the agents’choice
is the di¤erence in utility from each option relatively to the outside option (Nevo, 2000). Consumer
chooses the option Yi;t =bof highest utility (max(Ui;0;t; Ui;1;t; ::; Ui;B ;t)) and then a loan-amount
Li;t, which are a¤ected by observable characteristics, xi;t, plus unobservable preferences for each
loan type b,"i;b;t, and loan-amount, i;t. For simplicity, let us assume the utility of each loan type
10
is both an additive and linear function of the observables and the error term:
1) Ui;b;t =b;t +bxi;t +"i;b;t.
If the consumer decides to have a loan (options 1 to 5) instead of either "No wish for debt"
(option 0) or "No Access to Debt" (option 6), then he chooses a log-loan amount which is again an
additive and linear function of the observable factors, xi;t , plus an unobservable preference i;t:
2) ln(Li;t) = t+xi;t +i;t.
The decision of defaulting at time t,Di;t 2 f0;1g, is then given by whether a latent propensity
to default is positive, di;t >0. The latent propensity for defaulting on loans is again given by an
additive and linear function of the observable characteristics, zi;t, plus an unobserved shock i;t:
3) di;t =t+zi;t +i;t.
Note that the vector of observable variables that explains default, zi;t, di¤ers from the vector
of observable variables that explains the choice of the type of loan and the loan-amount, xi;t. This
is an intentional feature of the model and it is necessary for identi…cation. The reason is because
choice models that include an endogenous variable (for example, default in this model) a¤ected
by sample selection into di¤erent groups (for example, the type of loan chosen by agents in this
model) are ill-identi…ed if the same exact vector of variables explains both the endogenous variable
choice and the sample selection choice (Vella, 1998). Therefore it is useful if there are at least
a few variables that a¤ect sample selection (the loan choice, in this case), but do not a¤ect the
default decision directly. In our application there are actually some valid candidates for this role of
instruments that a¤ect loan choice, but not default. Note that although for simplicity of exposition
all the variables are indexed as being observed at time t, in fact loans have a maturity of several
periods (typically, around 1 or 2 years) and therefore the decision of loan choice happened before
the repayment period. For this reason it is natural to use the lagged value of some variables as
an explanation for loan choice and loan amount (for example, unemployment in the past year),
but use the contemporary value of the same variables as an explanation for default. This choice of
instrumental variables for loan choice is quite intuitive in economic terms and the validity of this
identi…cation approach is often recommended for panel data estimators (Vella, 1998).
11
To estimate the model it is necessary to specify the distribution of the unobservable random
terms, which has some degree of subjectivity since there are several possible distributions that may
provide a plausible …t. However, it is desirable that the distribution of the error term satis…es
four characteristics: i) it allows for the unobserved preferences of each agent to be correlated over
time, with some families being persistent in their behavior; ii) it accounts for some loan types being
closer substitutes to each other, therefore the utilities of di¤erent options are correlated; iii) the
agents’choice of all the distinct outcomes such as lender type, loan amount and default must be
correlated, which is predicted by the theory of adverse selection of debtors (Einav, Jenkins and
Levin, 2012); and, iv) the distribution should allow for an appropriate degree of heteroscedasticity,
since groups are not equally a¤ected by the unobserved shocks. A ‡exible way for achieving these
desired properties is to assume the unobserved tastes for each option "i;b;t are given by the sum of
an independent extreme valued component plus a normal random-e¤ect that is heteroscedastic and
correlated over several choices and time periods (McFadden and Train, 2000, Nevo, 2000, 2011):
4.1) "i;b;t = "i;b + ~"i;b;t,
4.2) "i;b = 1(1 b5)i;1+ 1(1 b2)i;2+ 1(2 b3)i;3+!i;b,
with ~"i;b;t EV (0;1),i;a N(0; a(xi)) and !i;b N(0; !b(xi)).1(:)is the indicator function,
assuming the value 1 if the condition is met and 0 otherwise. "i;b is the random-e¤ect that
represents the time-invariant tastes of the agent for each choice. Equation 4.2) for "i;b has a simple
interpretation in terms of its distinct components, with i;1representing a random factor denoting
agent i’s taste for any type of loan, i;2being a random factor denoting agent i’s taste for both the
Bank and Bank plus Retail loan options, and i;3denoting his taste for the options of Bank plus
Retail or just Retail. Finally, the random e¤ect !i;b is agent i’s speci…c taste for option b. The
distribution of all the random-e¤ects is heterocedastic in the vector xi, which represents the time
invariant characteristics of the agent and di¤ers from xi;t which includes time-varying variables.
In a similar way, I assume the unobserved terms for loan amount, i;t, and the propensity to
default, i;t, are correlated with the unobserved tastes for loan type:
4.3) i;t =
i+"i;b +~
i;t,
4.4) i;t = i+["i;b;
i] + ~i;t,
with ~
i;t N(0; ~
(xi)),
iN(0;
(xi)),iN(0; (xi)), and ~i;t EV (0;1). The log-loan
12
amount is a continuous variable and for this reason the contemporary shock that each agent faces
can be heteroscedastic. Note that the unobserved propensity of default is correlated with both the
unobserved tastes for each loan type and the unobserved taste for loan amount
i.
The model includes random-e¤ects, which requires panel data to identify the parameters.
However, the EFH data contains some purely cross-sectional samples and it is ine¢ cient to ignore
such a observations. For this reason the likelihood function includes both the panel and the
cross-section samples, which is a speci…c case of a Full Information Maximum Likelihood (FIML)
model. Let Pi fYi;t; Li;t; Di;t ; Yi;t+s; Li;t+s; Di;t+sjxi; xi;t; zi;t ; xi;t+s; zi;t+sgbe the vector containing
agent i’s choices of type of loan, loan amount and default at both time tand t+s, conditional on
the observables of both years. Also, let "ii;1; i;2; i;3; !i;1; :::; !i;B ;
i;ibe the vector of all
the unobservable random-e¤ects. All the random-e¤ects in vector "iare independent of each other,
therefore the pdf of "iis given by f("i) = (i;1
1(xi))::(i;3
3(xi))::(!i;b
!b(xi))::(
i
(xi))(i
(xi)).
This is assumed without any loss of generality, since the same random-e¤ects a¤ect di¤erent
endogenous variables and therefore the endogenous variables are correlated with each other.
For simplicity of exposition it is easier to write the likelihood of the three endogenous variables
given in equations 1), 2) and 3) conditional on the …xed-e¤ects "iand then multiply it by the pdf
f("i). Let ~
Ui;b;t =b;t+bxi;t+"i;b,ln( ~
Li;t) = t+xi;t+
i+"i;b, and ~
di;t =t+zi;t+i+["i;b;
i],
represent the expected means for the latent variables of equations 1), 2) and 3), assuming "iis
known. The likelihood of observing Pican then be written as a simple product of the multivariate
probability of the observed loan option b(given by the traditional multivariate logit ratio), with
the probability of loan amount (Li;t) and subsequent default (Di;t) in both periods:
5) Pr(Pi) = R Rf("i)exp( ~
Ui;b;t)
Pdexp( ~
Ui;d;t)(ln(Li;t)ln( ~
Li;t)
~
(xi))1(Li;t>0) exp( ~
di;t)Di;t
1 + exp( ~
di;t)
exp( ~
Ui;b0;t+s)
Pdexp( ~
Ui;d;t+s)(ln(Li;t+s)~
Li;t+s
~
(xi;t+s))1(Li;t+s>0) exp( ~
di;t+s)Di;t+s
1 + exp( ~
di;t+s)@"i.
For the cross-sectional sample, let the vector Pi;t fYi;t; Li;t; Di;t jxi; xi;t; zi;tgrepresent agent
i’s choices at time t, conditional on the observables fxi; xi;t; zi;tg. If one assumes the panel and
cross-sectional samples have the same representation in the population, then the likelihood function
can be integrated for the same distribution of random-e¤ects as the panel data observations. Note
that this does not imply the model is unidenti…ed, since the panel sample allows the model to
identify the complete distribution of the unobservables. Therefore this approach is valid as long as
13
the panel data sample is large enough. The likelihood of vector Pi;t is therefore written as:
6) Pr(Pi;t) = R Rf("i)exp( ~
Ui;b;t)
Pdexp( ~
Ui;d;t)(ln(Li;t)ln( ~
Li;t)
~
(xi;t))1(Li;t >0) exp( ~
di;t)Di;t
1 + exp( ~
di;t)@"i.
The log-likelihood of the model is then given by the sum of the log-likelihood of the panel and
cross-sectional samples, where i2Pdenotes whether the observation is in the panel sample or not:
7) LL =
T1
X
t=1
Nt
X
i=1;i2P
B
X
b=1
ln(Pr(Pi)) +
T
X
t=1
Nt
X
i=1;i=2P
B
X
b=1
ln(Pr(Pi;t)).
Besides the time-varying error terms, this model has 11 unobserved random-e¤ects which form
the vector "iand in‡uence the correlation of di¤erent choices and periods. This implies that
the likelihood function of equations 5) and 6) is based on a high dimensional integral and it is
computationally di¢ cult to calculate precisely. For this reason the choice probabilities are not
calculated exactly, but rather based on an approximation which averages a limited number of
draws, R, from the distribution of f("i). This Simulated Maximum Likelihood (SML) method is
asymptotically consistent if Rincreases proportionally with N(Train, 2009). In this application I
use 100 draws to simulate the probability of each observation, with the multivariate draws chosen
by a Modi…ed Latin Hypercube Sampling (MLHS) method (Hess, Train and Polak, 2006).9In
general, the MLE asymptotic distribution is also valid for the SML method, but this asymptotic
distribution is invalid if the model is not exactly true and if the number of draws Rdoes not
converge to in…nity (Train, 2009). Therefore the model’s standard-errors are estimated from 100
bootstrap replicas, which is asymptotically valid under a general set of conditions (Horowitz, 2001).
9I choose the MLHS method, because it chooses pseudo-random draws equally spaced in each dimension of
the integral and then randomly paired across dimensions. The reason why MLHS can perform better than standard
uniform draws is because uniform draws can have too much randomness and there is a certain probability of obtaining
draws too close to each other, while some areas of the integral have few or no draws at all. In this sense MLHS
guarantees that all the areas of each dimension are represented with at least one draw and therefore the simulated
draws have a wider coverage. Some simulation studies show that 100 MLHS draws can be as e¢ cient as more than
1000 uniform draws (Hess, Train and Polak, 2006). The MLHS method to obtain Rmultivariate draws basically
starts with an equal spaced sequence of values, '(j) = j1
Rfor j= 1; :::; R, in each dimension. Then a scrambled
Halton pseudo-uniform number xis added to the draws of each dimension to get ~'(j) = '(j) + x
Rfor j= 1; :::; R. The
draws are then transformed using the inverse normal cdf and multiplied by the standard-deviation of the univariate
distribution of the integral, 1(~'(j)), to obtain an univariate normal draw. The draws of each dimension are then
randomly paired with the Rdraws from the other dimensions to obtain Rmultivariate normal draws.
14
This model of loan decisions has certain implicit assumptions into it, since it assumes choices
are well approximated by a function of known characteristics and randomly distributed unobserved
preferences. One could assume other models for debt choice, such as an explicit multi-period
optimization where agents choose the best option for maximizing expected lifetime utility based on
an explicit evaluation of uncertain future paths and punishment costs for defaulting (see for instance,
Chatterjee et al., 2007). However, an explicit lifetime optimization framework requires several
assumptions about the agents’utility functions, their discount rates relative to future consumption
and the knowledge agents have about their uncertain future outcomes. Empirical evidence of
agents’cognitive limitations disputes assumptions such as rational expectations, time-consistency
and revealed preference (Bertrand and Morse, 2009, Kahneman, 2011). Therefore simple behavioral
models are not necessarily less realistic than structural models based on complete optimization. For
this reason, the choice model in this paper is more closely related to other works who approximate
agents’decisions in a ‡exible way, such as Edelberg (2006) and Einav, Jenkins and Levin (2012).
3 Data
3.1 The Chilean Household Finance Survey (EFH)
The main source of information for the characterization of the …nancial behavior of Chilean households
is the Chilean Household Finance Survey (in Spanish, Encuesta Financiera de Hogares, hence on
EFH). The EFH is a representative survey with detailed information on households’assets, debts,
income and …nancial behavior, and is broadly comparable to similar surveys in the United States
and Europe (Eurosystem, 2009). In 2007 and 2011 the EFH interviewed 3828 and 4059 urban
families nationwide. In the years 2008 to 2010 the EFH was only implemented in the capital city
of Santiago (which represents over 40% of the total national population), therefore the sample
size is smaller for those waves. The EFH has a rotating sample, in which part of the sample is
re-interviewed. Therefore there are 1792 families which were interviewed both in 2007 and 2011,
while 947 families were interviewed both in 2008 and 2009. In total there are 6790 cross-sectional
observations (i.e., families interviewed only once) plus 2739 panel observations (Table 1).
15
Table 1: Panel and cross-sectional sample size of the Household Finance Survey (EFH)
EFH Panel Cross-Section Total
2007 1,792 2,036 3,828
2008 947 207 1,150
2009 947 243 1,190
2010 2,037 2,037
2011 1,792 2,267 4,059
Total 2,739 6,790 12,264
The EFH has a particularly detailed focus of the loans and debt commitments of each household.
It asks for the largest 3 debts that each household has for each type of loan, among a total of 13
categories of loans: Banking Credit Card Debt, Banking Line of Credit, Banking or Financial
Agency Consumer Credit Loan, Retail Store Credit Card, Retail Store Consumer Loan, Auto
Loans, Social Credit, Education Loans, Loans from relatives, Loans from usurers, Pawn shops,
Grocery and Shopping on credit (i.e., store tabs), and Other Debts. Therefore the survey may ask
up to a total of 39 debts that the household has at the moment, although obviously very few agents
will report having debts with all the possible categories of loans.
For two reasons it is easier to work with just 5 types of lenders (or 5 types of loans), therefore
my analysis is limited to options that sum all the loans for a given lender type and with each family
classi…ed discretely with the lender type representing the largest loan amount: Banks, Banks and
Retail Stores (for the families reporting the use of both kinds of loans), Retail Stores, Social Credit,
and Other Debts. The …rst reason is that it is desirable to eliminate the irrelevant alternatives from
the choice model (Train, 2009), with a classic example being the inclusion of options such as "red
bus" and "blue bus" for agents that do not care about the color of public transport. Several of the
13 types of loans elicited by the survey are similar products and are often o¤ered by lenders to the
same customers and for similar purposes (for instance, many customers use Credit Cards and Lines
of Credit for similar reasons, although their choices may depend on the speci…c convenience of the
occasion). This is a strong reason for aggregating all the options for credit cards, lines of credit and
consumer contracts for each lender, instead of treating them separately. The second reason is related
to the curse of multidimensionality, since the number of parameters in the model increases with the
number of options and it is di¢ cult to make a reliable analysis of too many options, particularly if
16
some options have few or no observations (for example, loans from usurers are reported by less than
0.07% of the families). For this reason, Other Debts represents the sum of Auto Loans, Education
Loans, Loans from relatives, Loans from usurers, Pawn shops, Grocery and Shopping on credit
(i.e., store tabs), and Other Debts. Note that this category is largely composed of Auto Loans,
Education Loans and Other Debts, with the remaining options representing negligible numbers.
Table 2 shows the proportion of households that chose each of the 5 lender types, plus households
with either No Consumer Debt (because the family does not want debt) and No Access to Debt
(if the family applied for loans, but was refused). The proportion of households without a wish for
consumer debt represents 27% of the Chilean population, while those with No Access to Debt
represent close to 13% of the population. Retail Stores are the most popular choice among
households, representing more than 40% of the population, with 29% being Retail Store only
users and 13% being users of both Bank and Retail Store Loans.
For each debt the EFH survey registers its loan amount (in Chilean pesos), maturity (in months),
and other details such as the motivation for contracting the loan (with possible motives including
vacations). The survey questionnaire also asks about the loan’s interest rates, but less than half
the respondents report to remember them.10 One important aspect of our study is how default of
consumer loans is measured. The question used for measuring default is "Approximately, in the last
12 months have you fallen into morosity or late payments for each one of your loans?". I consider
that default corresponds to a dummy variable denoting one or more events of morosity.
Table 2 shows the loan amount, maturity and morosity rates of di¤erent lender types using the
pooled EFH sample, that is all the cross-sectional samples available. I also report the average loan
interest rates of di¤erent lender types, from statistics of the Chilean Superintendency of Banks
and Financial Institutions and the Superintendency of Social Security. There is no information
on interest rates from users of several lenders, such as Banks and Retail or Other Debts. Table
2 also reports the share of the consumer loan destined for a given purpose of the household,
more speci…cally "Pay previous debts" and "Health needs". Other motivations such as "general
consumption" are not reported, since their classi…cation is too general to be interpreted. The main
10 This memory problem is explained by debtors tendency to remember the payment amount better than their
contract’s interest rate. Also, it is di¢ cult to recover an estimate of the implicit interest rate from the …nancial
formula for the present value of the payments of a loan. This is due to ommited variables in the …nancial formula,
such as loan commission fees, and measurement error in the reporting of the loan amounts, payments and maturities.
17
Table 2: Population*, Maturity (months), Loan amounts (thousands of Chilean pesos), motivation (share
of total consumer debt destined for a given purpose), interest rate and morosity rates (EFH)
Type of Debtor Population Maturity Morosity Loan amount: mean/percentiles Interest Pay debts Health
Mean Mean Mean p25 p50 p75 Mean Mean Mean
Bank 7.8% 25.28 10.2% 2,549 416 1,110 2,649 19% 14.3% 5.0%
Bank+Retail 12.9% 20.45 21.3% 3,015 859 1,703 3,520 16.9% 6.0%
Retail Store 28.9% 12.17 19.0% 492 102 216 486 47% 3.7% 3.3%
Social Debt 5.6% 27.07 11.4% 1,124 307 590 1,131 21% 18.1% 13.8%
Other Debts ** 4.6% 32.25 21.5% 4,101 1,207 2,425 4,274 8.0% 3.2%
No Consumer Debt 27.3%
No Access to Debt 13.0%
* % of the total Chilean households in urban areas. ** Maturity for Other Debts is for Auto Loans only.
conclusion is that households with Bank, Bank plus Retail, and Social Debt are more likely to have
a motivation of paying back previous debts or health needs. Loan consolidation and health needs
also motivate a signi…cant part of unsecured debt in the USA (Chatterjee et al., 2007).
Users of Bank credit only have a morosity rate of 10%, which is half the value reported by
users of both Bank and Retail credit (Table 2). Also, Bank users have much larger loan amounts
and longer maturities than the users of Retail Stores. In Chile neither Retail Stores or institutions
of Social Credit are able to o¤er heterogeneous interest rates to their customers, only Banks o¤er
customer speci…c interest rates (Marinovic, Matus, Flores and Silva, 2011), so the economic theory
predicts that Banks will get the best observable risk types by o¤ering better loan terms such as
lower interest rates, larger loan amounts and longer maturities. While Social Debt lenders are
unable to risk price their o¤ers, these institutions are able to garnish their clients’wages easily,
therefore this high punishment cost should explain their low morosity rates. However, households
with both Bank and Retail Store debt have morosity rates as high as the customers of Retail Stores
only. Perhaps this can be explained because such debtors have an unobservable taste for high loan
amounts. Table 2 shows that households with both Bank and Retail Store debt have much higher
loan amounts than the debtors of Bank and Retail Store separately, which could be a sign that these
are debtors with particularly high needs for liquidity. The households with Other Debts also have
high loan amounts and morosity rates, but perhaps this can be explained by special characteristics
of these debtors. For example, education loans are granted to younger agents, who may be more
18
Table 3: Population of debtors, loan amounts (thousands of pesos) and morosity over time (EFH)
Type of Debtor Population Loan amount (median) Morosity rate
2007 2011 2007 2011 2007 2011
Bank 6.5% 8.2% 968 1,176 8.8% 11.7%
Bank+Retail 13.6% 11.8% 1,435 1,826 18.9% 24.6%
Retail Store 31.9% 25.9% 232 177 21.1% 19.5%
Social Debt 3.8% 7.8% 484 748 12.1% 12.2%
Other Debts 4.6% 4.9% 1,511 2,866 25.2% 20.5%
No Consumer Debt 26.6% 28.7%
No Access to Debt 13.0% 12.7%
subject to unemployment risk and unstable income. Also, perhaps education and auto loans have
lower punishment costs for morosity, since lenders cannot deduct payments and punishment fees
from their clients’bank accounts (as Banks do) or their wages (as Social Credit institutions do).
Table 3 shows the percentage of the population, median loan amounts and morosity rates in the
years 2007 and 2011. The biggest changes observed between 2007 and 2011 are that users of only
Banks and Social Debt increased respectively to 8.2% and 7.8% of the population. Loan amounts
of users of Social and Other Debts increased substantially, while the median loan amount at Banks
increased less. It is also noticeable that the morosity rate of Bank users increased somewhat.
Tables 4 and 5 summarize the changes to income and use of consumer loans in Chilean households,
using information from the EFH panel sample (2007-2011). In Table 4 I report the transition
probabilities from one household income quintile (Qi;t) to another between 2007 and 2011, Pr(Qi;2011 =
qjQi;2007 =q0), where 1 denotes the families with the 20% lowest income. The conclusion is that
household income has some persistence, but there is substantial income volatility in Chile. The
probability that a household of the lowest income (quintile 1) will remain at the bottom of the
distribution is 40%, while the probability of a household remaining at the top income level (quintile
5) is 53%. Among the middle income levels (quintiles 2 to 4), mobility is even higher and there is
a high chance that households will move into either a higher or a lower income level.
In Table 5 I show the transition probability of a household changing from one lender type to
another or towards having either no consumer debt or no access to debt, Pr(Yi;2011 =bjYi;2007 =b0).
The last column in the table replicates the share of the population in each debt status over the
whole period of 2007 to 2011. If one compares the diagonal values of the transition matrix, which
19
Table 4: Transition of families across di¤erent income quintiles (EFH Panel, 2007-11)
Quintile 2011
Quintile 2007 1 2 3 4 5
1 40% 27% 19% 7% 7%
222% 29% 27% 15% 7%
311% 23% 26% 25% 15%
410% 14% 22% 30% 24%
57% 9% 12% 20% 53%
Table 5: Transition of households across di¤erent debtor types (EFH Panel, 2007-11)
Debt Status in 2011 Population
Debt in 2007 No Debt Bank Bank+Retail Retail Social Other No Access in 2007-11
No Debt 40.9% 7.0% 6.7% 27.8% 4.2% 2.3% 11.1% 27.3%
Bank 27.8% 18.1% 20.5% 19.1% 3.9% 2.1% 8.4% 7.8%
Bank+Retail 16.4% 18.2% 30.0% 25.1% 2.2% 1.5% 6.7% 12.9%
Retail Store 20.9% 7.5% 13.6% 39.1% 5.2% 0.4% 13.3% 28.9%
Social Debt 36.2% 8.1% 3.5% 34.0% 14.1% 0.0% 4.1% 5.6%
Other Debts 34.7% 12.7% 28.4% 12.6% 0.0% 5.5% 6.1% 4.6%
No Access 33.4% 4.9% 7.8% 28.9% 6.9% 0.6% 17.6% 13.0%
represent the probability of a debtor keeping the same status as previously, with the average debt
status of the population, then one gets an idea of how persistent agents are in their choices. It
is clear that the probability of an agent keeping the same debt status is above the average rate
in the total population and this happens for all categories, therefore choices tend to be persistent.
In particular, debtors of Social Debt, Banks or of Bank plus Retail Store are more than twice as
likely to keep their choices relative to the average probability in the population. Also, it is striking
that debtors of Retail or Bank plus Retail have a probability of only 20% and 16% respectively
of moving into a state of No Debt. Therefore these debtors are systematically in need of debt,
whether with the same lender or a di¤erent one. This con…rms the previous results that debtors of
Bank plus Retail appear to be agents with higher needs for liquidity relative to other households.
20
4 The sorting of income risk across di¤erent types of loans
The EFH survey collects detailed information on the income, education, age and other characteristics
of each household member, but it has limited data on some aspects, such as their income volatility
or stability of employment. For this reason I estimate the income and employment risks of the EFH
workers based on the mean statistics for workers with the same characteristics in another dataset.
Based on the quarterly Chilean Employment Survey, ENE, which covers 35,000 households,
Madeira (2014) estimated three measures of risk in employment status for the period 1990 to 2012:
the unemployment rate (uk;t = Pr(Uk;t = 1 jt; xk)), the separation rate (EU
k;t = Pr(Uk;t+1 = 1 j
t; Uk;t = 0; xk)) de…ned as the probability of being unemployed given that one was employed in the
previous quarter, and the job …nding rate (U E
k;t = Pr(Uk;t+1 = 0 jt; Uk;t = 1; xk)) de…ned as the
probability of being employed after being unemployed in the previous quarter. The vector xkis
composed of 540 mutually exclusive groups, given by xk=fSantiago Metropolitan city or Outside,
Industrial Activity (primary, secondary, terciary sectors), Gender, Age (3 brackets, 35,35 54,
55), Education (less than secondary schooling, secondary or technical education, college), and
Household Income quintileg. Madeira (2014) also computed these groups’ labor income volatility
even if no job is lost, ;t(xk) = pE[(Yk;t E[Yk;t jYk;t1; xk])2jt; Uk;t =Uk;t1; Yk;t; xk], and the
income loss caused by going into unemployment, Rk;t (xk) = E[Yk;t jt; Uk;t = 1; xk]
E[Yk;t jt; Uk;t = 0; xk].
Using these labor risk measures I calculate the expected income
Pi;t of each EFH household
ias the sum of their non-labor income, ai, and its expected labor income, Pi;t:
Pi;t =ai+
Pi;t, where Pi;t =PkPk;t is the sum of expected labor income of each household member k.
Pk;t =Wk;t(1 uk;t) + Wk;tRk;t(uk;t)is each worker k’s average labor income during the employed
(Wk;t =Yk;tRUk;t
k;t ) and unemployed states. The employment risk of each household is then given
by a weighted average of the rates of each member using their labor income relative to the total
household labor income: ui;t =Pk
Pk;t
Pi;t uk;t,
UE
i;t =Pk
Pk;t
Pi;t UE
k;t and
EU
i;t =Pk
Pk;t
Pi;t EU
k;t . Similarly,
the household’s weighted labor income volatility (even if no job is lost) and the replacement ratio
during unemployment are given by i;t =Pk
Pk;t
Pi;t ;t (xk)and
Ri;t =Pk
Pk;t
Pi;t RRk;t(xk).
Figure 2 shows the cumulative distribution function of the loan amounts (in logarithm) and
the consumer debt to annual income ratio ( Li;t
12
Pit , where
Pit is the expected monthly income) in
the pooled EFH survey (2007-11). Retail only debtors are the ones with the highest probability
21
Figure 2: The Cdf of the loan amounts chosen by debtors of di¤erent loan types
.2
.4
.6
.8
1
11.5 12 12.5 13 13.5 14 14.5 15 15.5
Ln(Loan)
Bank Bank-Retail Retail
Social Other
.2
.4
.6
.8
1
.1 .2 .3 .4 .5 .6
Debt to Income Ratio
of having low loan amounts, since their cdf is stochastically dominated by either Social and Bank
debtors. Bank plus Retail debtors and Other debtors have the greatest probability of having
high loan amounts (or the lowest probability of having low loan amounts). One question is if the
di¤erence in loan amounts is entirely explained by income, since higher income households may
also pay larger loans. The answer is given by the empirical cdf of the consumer debt to annual
income ratio. In Figure 2 it is shown that clearly Retail only debtors have lower debt to income
values in relation to both Social and Bank debtors. Also, Bank plus Retail and Other Debts users
have the highest debt to income ratios. Therefore the di¤erences in the sorting of loan amounts
across lender types remains even if we take into account household income.
Table 6 reports the mean values of the household’measures for the unemployment rate (ui;t), the
separation rate (
EU
i;t ) and the job …nding rate (
UE
i;t ) across di¤erent loan choices. The groups with
No Consumer Debt or only Bank loans are the ones with the lowest unemployment and separation
rates. Households with Other Debts are the ones with the highest average unemployment rates,
22
Table 6: Mean values of labor market risk and household earnings across debtor types (EFH)
Debtor Type ui;t
EU
i;t
UE
i;t ln(
Pi;t) i;t
Ri;t
Bank 4.8% 2.0% 33.8% 13.56 18.4% 25.8%
Bank+Retail 5.3% 2.3% 35.4% 13.46 18.2% 25.5%
Retail Store 5.5% 2.6% 36.6% 13.01 16.5% 23.5%
Social Debt 5.0% 2.0% 30.9% 13.14 17.4% 22.7%
Other Debts 6.1% 2.2% 34.2% 13.47 20.7% 26.1%
No Consumer Debt 4.2% 1.9% 30.6% 13.13 16.2% 23.0%
No Access to Debt 5.4% 2.2% 31.0% 12.77 17.6% 21.3%
perhaps because of their younger age. The mean job …nding rate is between 31% to 37% for all
groups. Table 6 also reports the means values for the log household expected income (ln(
Pi;t)), the
labor income volatility (i;t) and its replacement ratio of income during unemployment (
Ri;t). Bank
only customers are the group of highest income, while those with Retail Store loans or with No
Access to Debt have the lowest mean income. Unemployment represents a strong income reduction
for Chilean households, since the mean values of
Ri;t imply that agents only keep 21% to 26% of their
working income during an unemployment spell. The households with No Consumer Debt appear
to be the group least susceptible to shocks, since they are the group with lowest unemployment
rate, lowest separation rate and lowest labor income volatility. The permanent income theory
of consumption predicts that agents should use debt to smooth temporary income shocks (see
Chatterjee et al., 2007, or Dynan and Kohn, 2007), therefore it makes sense that households with
the lowest income risk also have the lowest demand for consumer loans.
While Table 6 reports the mean values of households’income, employment risks and income
volatility, it is also useful to analyze how heterogeneous households are and how each group deviates
from the mean. Figure 3 shows the cdf of the households’expected income (ln(
Pi;t)), unemployment
rate (ui;t ), labor income volatility (i;t, which can also be denoted as the standard deviation of wage
shocks) and replacement ratio of income during unemployment (
Ri;t) for debtors and non-debtors.
For simplicity, I use only 4 groups in the graphical comparison instead of the 7 groups used in Table
6 and the previous tables. Basically, I classify households in the same two options for non-debtors
as before (No Consumer Debt, No Access to Debt), but use only two classi…cations for the groups
of debtors: i) users of Retail Store loans only, which represent 29% of the household population
23
(Table 2) and are the largest group with consumer debt; and, ii) users of Bank, Social Debt and
Other Debts, which represent 30.9% of the Chilean population (this …gure is obtained by summing
the distinct categories of this group in Table 2). Another simpli…cation concerns the problem that
often households have a lot of heterogeneity at the extreme margins, but one is mostly concerned
with the heterogeneity that a¤ects most of the population and not its extreme points (which could
eventually be outliers due to measurement error). Therefore to make the graphs easier to read the
cdfs are plotted only in the range of 20% to 90% probability.
Figure 3 shows that in terms of income there is a clear stochastic dominance among the di¤erent
groups, with households with No Access to Debt having lower income than those with Retail loans
and those with Retail loans having lower income than both the households with No Consumer Debt
and the households with Bank, Social and Other Debts. Also, it is clear that households with No
Consumer Debt have the lowest unemployment rates, which is another con…rmation that a partial
motivation for consumer loans is to smooth temporary income shocks. Households with Bank,
Social and Other Debts also have lower unemployment rates relative to those with Retail loans
only or No Access to Debt. Labor income volatility (i;t) is highest for the households with Bank,
Social and Other Debts, which may imply that consumer debt is used for smoothing income shocks
in this group. The replacement ratio of income during unemployment is the lowest for those with
No Access to Debt, followed by the users of Retail loans only and those with No Consumer Debt.
Users of Banks, Social and Other Debts have the highest replacement ratios during unemployment,
therefore this is the group that su¤ers the lowest loss of income from job loss.
Figure 3 shows that income and labor experiences have a lot of heterogeneity in the population.
Unemployment rates can range from as low as 2% to as high as 8%. Labor income volatility has a
range between 11% to 27%, while replacement ratios can vary between 18% and 34%.
Besides analyzing unemployment rates, it is also appropriate to look at the employment separation
(
EU
i;t ) and job …nding (
UE
i;t ) rates. The reason is because unemployment rates has a di¤erent
interpretation if it is driven by high separation rates (lots of workers losing their jobs) or by low job
…nding rates (which implies that unemployed workers have di¢ culties …nding jobs and therefore
unemployment spells last a long time). Both of these employment transition rates play a role in
explaining labor market shocks in the United States (Shimer, 2012) and in Chile (Madeira, 2014).
Figure 4 shows the cdf of the separation (
EU
i;t ) and job …nding (
UE
i;t ) rates for Chilean households.
24
Figure 3: The Cdf of labor market characteristics of debtors versus non-debtors
.2 .4 .6 .8
12.6 12.8 13 13.2 13.4 13.6 13.8 14
Ln(Income)
Bank-Social-Other Retail
No Debt No Access
.2 .4 .6 .8
.02 .03 .04 .05 .06 .07 .08
Unemployment Rate
.2 .4 .6 .8
.11 .13 .15 .17 .19 .21 .23 .25 .27
Standard-deviation of wage shock
.2 .4 .6 .8
.18 .2 .22 .24 .26 .28 .3 .32 .34
Replacement Ratio of Income
25
Figure 4: The Cdf of employment transition probabilities for di¤erent loan types
.2 .4 .6 .8
.01 .02 .03 .04
Separation Rate
Bank-Social-Other Retail
No Debt No Access
.2 .4 .6 .8
.25 .3 .35 .4 .45 .5 .55
Jobfinding Rate
There is a lot of heterogeneity in these variables, with the separation rate ranging from as low as
1% to as high as 4% and the job-…nding rate varying between 25% and 55%. The separation
rate has the most clear di¤erences between debtor and non-debtor groups. Households with No
Consumer Debt have lower separation rates than users of Bank, Social and Other Debts, and these
last ones have lower separation rates than those with No Access and the users of Retail loans only.
The di¤erences in job-…nding rates are less clear. Users of Retail loans only have both the highest
separation rates and the highest job …nding rates, which implies that employment mobility is high
in this group. However, the groups with No Consumer Debt, No Access and users of Bank, Social
and Other Debts have a similar distribution for the job-…nding rate.
Figure 5 shows the di¤erences in income and labor market characteristics of di¤erent debtor
groups. Users of Retail and Social Debt are the ones with the lowest income, while the users of Bank
and Other Debts have the highest income. Also, Bank users have a lower unemployment rate than
all the other debts, with Social Debt users being the second group with the lowest unemployment
26
rates and users of Other Debts having the highest unemployment. Retail and Social Debt users,
however, have the lowest labor income volatility (or standard-deviation of wage shocks), while users
of Bank and Other Debts have the highest wage risk. Users of Retail and Social Debt are the ones
with the lowest replacement ratios and therefore su¤er the most during a jobless spell.
Overall, Figures 2, 3 and 4 portray a clear picture of di¤erent income and labor market
characteristics across non-debtors and di¤erent groups of debtors. Households with No Access
to Debt have the lowest income, highest unemployment rates and lowest replacement ratios of
income, therefore it is the group most subject to low income and income ‡uctuations. Households
with No Consumer Debt (because of a lack of demand for such loans) have the lowest unemployment
rates, separation rates and labor income volatility, therefore it is the group least subject to income
shocks. The users of just Bank loans are the ones with the highest income, highest replacement
ratio and lowest unemployment rate, but they su¤er from substantial wage volatility which may
create a demand for smoothing consumption. Users of Other Debts have high income and high
replacement ratios in the same way as Bank users, but they are the debtor group most subject
to both high unemployment rates and high labor income volatility, therefore it could be seen as a
riskier segment relative to Bank users. Finally, users of Retail loans are the ones with the lowest
income among debtors (although they have higher income than the group with No Access to Debt),
and also have a high unemployment rate and low replacement ratio, which could make them a riskier
debt segment. However, Retail users have a low standard-deviation of wage shocks, therefore their
income is relatively stable during their employment experience. Users of both Bank plus Retail
loans are a segment somewhat in between the exclusive users of either Banks or Retail loans.
5 Results
5.1 The role of demographics, income pro…le and unobserved preferences
Now I discuss the results from the consumer loan choice and default model exposed in section 2. As
explained before, the model requires some variables that a¤ect loan choice, but not default. Since
unemployment risk, labor income volatility are measured for several time periods (all the quarters
from 1990 to 2012) for each type of worker, then it is possible to create these variables for each
27
Figure 5: The Cdf of labor market characteristics of debtors of di¤erent loan types
.2 .4 .6 .8
12.6 12.8 13 13.2 13.4 13.6 13.8 14
Ln(Income)
Bank Bank-Retail Retail
Social Other
.2 .4 .6 .8
.03 .04 .05 .06 .07 .08
Unemployment Rate
.2 .4 .6 .8
.13 .15 .17 .19 .21 .23 .25 .27
Standard-deviation of wage shock
.2 .4 .6 .8
.18 .2 .22 .24 .26 .28 .3 .32 .34
Replacement Ratio of Income
28
EFH household at previous periods than the survey date. It is natural to assume that households
were driven by labor market e¤ects that happened at the time of the loan contract, which was a
substantial time before the current period t. Consumer loans typically have a maturity of 12 to 24
months, therefore it is reasonable to assume that the labor market conditions that in‡uenced loan
choice happened 4quarters or more before the current period. For this reason the vector a¤ecting
loan choice includes expected income (ln(
Pi;t4),), unemployment risk (ui;t4) and labor income
volatility (i;t4) with a lag of 4quarters (although a shorter or longer lag could be used). Note that
expected income is a weighted sum of each household member’s labor income, its unemployment
probability and its replacement ratio, therefore it can be estimated for previous time periods.
In addition the vector that a¤ects loan choice includes the education, age and structure of the
household (whether it is a couple or a family with many members), and the motivation of the loans
(which share of the debt was motivated by needs of "Health" or to "Pay Previous debts"):
xi;t =
8
>
>
>
>
>
>
>
>
>
<
>
>
>
>
>
>
>
>
>
:
ln(
Pi;t4), unemployment risk ui;t4, labor income volatility i;t4,
years of education of household head, age of the household head, age squared,
dummies for each year, dummy for residence being out of the Santiago capital,
dummy for 2 members and dummy for 3 or more members in household,
Share of debt justi…ed by "Pay Previous Debts", Share of debt justi…ed by "Health"
9
>
>
>
>
>
>
>
>
>
=
>
>
>
>
>
>
>
>
>
;
.
In a similar way I assume that the vector zi;t that explains loan morosity or default at time
tincludes some variables that do not necessarily a¤ect loan choice. One variable is the ratio
of consumer debt to the annual income (RDIi;t =Li;t
12
Pi;t
), which can be seen as a measure
of long-term solvency of the household. Households with larger loans may feel more stressed
about their long term commitments and choose to default on their loans. In the same way, some
households may be more worried about this month’s speci…c commitments instead of their long-term
expenses. For those households, the current monthly debt service (the debt service of a loan, DSi;t,
includes both the amortization and the interest payments) relative to this month’s income (Yi;t)
may provide a liquidity motive for defaulting or simply paying a loan with some delay. For this
reason I also include the ratio of monthly Debt Service to Income (RDSIi;t ) as a possible factor
a¤ecting household default. The overall vector of observables that explain default includes the
…nancial ratios RDIi;t ,RDS Ii;t, plus the current expected income (ln(
Pi;t)), unemployment risk
(ui;t) and labor income (i;t ), and the same demographic variables that a¤ect loan choice:
29
Table 7.1: Coe¢ cients for the mean value of the Utility of each type of Loan
1=Bank 2=Bank+ Retail 3=Retail Store 4 =Soc ial D ebt 5=O ther De bts 6=No-Access
2007 -9.706 (0.161) -6.357 (0.188) 4.783 (0.687) 3.499 (0.959) -5.347 (3.809) 9.737 (0.906)
2008 / 09 -9.912 (0.157) -6.924 (0.183) 4.623 (0.759) 3.126 (1.003) -6.055 (4.068) 9.928 (0.915)
2010 -9.904 (0.163) -7.204 (0.191) 4.855 (0.676) 3.904 (0.99) -5.225 (3.783) 10.228 (0.924)
2011 -9.524 (0.165) -6.569 (0.19) 4.868 (0.65) 4.462 (0.989) -4.819 (3.809) 10.251 (0.933)
Income: ln(
Pi;t4)0.346 (0.019) 0.095 (0.022) -0.5 (0.055) -0.449 (0.056) 0.086 (0.719) -0.831 (0.053)
Edu cation 0.073 (0.01) 0.046 (0.01) -0.109 (0.013) -0.077 (0.007) 0.047 (1.223) -0.068 (0.009)
Unemp loym ent ui;t42.668 (1.052) 3.139 (0.513) 4.618 (0.415) 4.108 (0.292) 4.344 (1.845) 2.402 (0.754)
Wage volatility i;t40.152 (0.257) 0.386 (0.232)-0.372 (0.38) 0.938 (0.376) 3.115 (0.599) 1.502 (0.527)
Out of Santiago -0.143 (0.007) -0.232 (0.011) -0.266 (0.03) -0.475 (0.034) -0.238 (1.486) -0.261 (0.014)
Age o f hom e he ad 0.115 (0.003) 0.148 (0.003) 0.108 (0.006) 0.001 (0.022) 0.025 (0.424) 0.026 (0.013)
Age squared -0.001 (0) -0.002 (0) -0.001 (0) 0 (0) -0.001 (0.002) 0 (0)
2 members in hom e 0.257 (0.029) 0.542 (0.02) 0.74 (0.029) -0.01 (0.052) 0.723 (0.498) 0.005 (0.009)
3 or more m emb ers 0.567 (0.019) 1.069 (0.014) 1.337 (0.071) 0.29 (0.071) 1.014 (0.31) 0.348 (0.018)
Pay previo us debts 5.034 (3.636) 5.286 (3.632) 3.126 (3.622) 4.892 (3.786) 0
Hea lth needs 5.093 (2.913)5.544 (2.917)4.398 (2.922) 6.119 (2.838) 0
zi;t =8
>
<
>
:
Li;t
12
Pi;t
,DSi;t
Yi;t
,ln(
Pi;t), unemployment risk ui;t , labor income volatility i;t ,
years of education of household head, age of the household head, ::: .
9
>
=
>
;
.
Finally, I need to specify the degree of heteroscedasticity in the unobserved tastes for loan choice,
loan amount and default, that is the standard-deviation of each element of the normally distributed
vector n~
i;t; "io, where the vector of random-e¤ects of tastes (i.e., tastes that are constant over time
for each agent) is given by "ii;1; i;2; i;3; !i;1; :::; ! i;B;
i;i. In this case I assume that all the
standard-deviations are exponential functions of a linear-index, = exp(xi), which guarantees
that all standard-deviations are positive. The vector xithat models heteroscedasticity includes
a constant, a dummy for the 2008/09 panel, the years of education of the household head, plus
the average labor market characteristics of the household (since income and labor market risk is
time-varying I apply the mean values over the period 2007 to 2011 for simplicity):
xi=nconstant, dummy for 2008/09, years of education of head, 1
TP2011
t=2007 ln(
Pi;t);ui;t ;i;to.
Table 7.1 shows the estimates for the coe¢ cients of loan choice, b. The coe¢ cients of a
multivariate logit model sometimes have a di¢ cult interpretation (Train, 2009), because the agents’
choice is made over a multivariate set with B+ 1 choices, max(Ui;0;t; Ui;1;t; ::; Ui;B ;t), with the …rst
choice being standardized to have value zero. Let us think of a generic variable xand its coe¢ cient
on choice b,b, which is assumed to be positive. Then b>0implies the odds ratio of the
probability of option brelative to option 0is increasing in x, meaning that larger xmakes bmore
30
likely to be chosen relative to option 0. However, at the same time there could be another option
cwhich has a larger coe¢ cient than b, implying xdecreases the chance of bbeing chosen relative
to option c. Therefore in the multivariate case b>0does not always increase the probability of
bbeing chosen with larger x. Such is the case only if bmax(1; ::; B). This interpretation of
the multivariate logit coe¢ cients must be kept in mind while reading Table 7.1.
The coe¢ cient for the lagged household expected income (ln(
Pi;t4)) is the largest (i.e., the
most positive) for the option of Bank loans, while it is the lowest (i.e., the most negative) for the
No-Access option (Table 7.1). This implies that larger income unambiguously increases the option
of a Bank loan and decreases the option of No-Access. The impact of income on the choice of
Retail Store loans and Social Debts is negative, therefore larger income increases the likelihood of
No Debt in relation to these options. The coe¢ cient of education is largest for the Bank option and
lowest for the Retail Store option, which implies that education increases the probability of a Bank
loan and decreases the option of Retail Store loans. The coe¢ cient of lagged unemployment (ui;t4)
increases the probability of all the loan options and the No-Access option in relation to No-Debt.
However, unemployment has the e¤ect of increasing more the probability of speci…c loans, such
as the Retail Store, Social Debt and Other Debts options. Being outside of the Santiago capital
city has the e¤ect of lowering the probability of all loan options, with its strongest e¤ect on Social
Debt. Higher age is positively related to choosing the Bank, Bank plus Retail, and Retail Store
options. Households with more members are more likely to choose the option of Retail Store loans.
Finally, the motivations for undertaking a loan (Pay previous debts or Health needs) have a special
standardization, because families who report positive values for the loan motivation must have
chosen one of the loan options 1 to 5 and therefore I standardize the loan motivation coe¢ cients for
the last option Other Debts as being 0. Pay Previous Debts does not have an e¤ect on a speci…c
loan type, but Health needs is associated to the choice of Bank, Bank plus Retail and Social Debt.
Table 7.2 shows the heteroscedasticity for the random e¤ects that denote the unobserved tastes
for each loan type. The strongest conclusions is that the heterogeneity for the tastes of Bank,
Bank plus Retail, and Social Debt, decreases with income and years of education. Similarly, Table
7.3 shows the heteroscedasticity of the random-e¤ects that a¤ect several loan options. The main
conclusion is that the heterogeneity for the unobserved taste for all loans (i;1) and of the taste
for Bank plus Retail and Retail only (i;3) is also decreasing in income and education. Therefore
31
Table 7.2: Coe¢ cients for the standard-deviation (in log) of the random-e¤ect of each type !i;b
1=Bank 2=Bank+Retail 3=Retail Store 4=Social Debt 5=Other Debts 6=No-Access
constant -0.327 (0.962) -0.296 (0.092) 0.519 (0.465) 1.001 (0.394) -0.126 (1.797) 0.12 (1.18)
2008/09 0.146 (0.682) 0.278 (0.157)0.388 (2.875) -2.715 (1.683) -0.213 (2.108) -0.001 (0.788)
1
TP2011
t=2007 ln(
Pi;t)-2.616 (1.342)-2.127 (1.126)-0.024 (0.34) -4.691 (1.604) -1.008 (2.387) -0.625 (1.534)
Education -2.307 (0.886) -1.566 (0.946)-0.003 (2.308) -9.046 (3.245) -0.693 (1.595) -1.246 (1.28)
1
TP2011
t=2007 ui;t -0.05 (1.206) -0.032 (1.002) -0.658 (0.6) 0.765 (1.179) -0.009 (1.132) 0.05 (1.127)
1
TP2011
t=2007 i;t -0.082 (0.736) -0.047 (0.769) -1.046 (0.706) -0.007 (1.271) 0.007 (0.422) 0.097 (2.303)
Table 7.3: Coe¢ cients for the standard-deviation (in log) of the factors a¤ecting several choices
factor 1, i;1factor 2, i;2factor 3, i;3
(choices 1 to 5) (choices 1 to 2) (choices 2 to 3)
constant -0.133 (7.117) -5.559 (1.522) -0.516 (2.898)
2008/09 0.566 (5.986) 0.9 (4.178) -0.008 (3.415)
1
TP2011
t=2007 ln(
Pi;t)-3.691 (1.921)0.481 (1.45) -6.718 (3.387)
Education -1.3 (0.706)-0.082 (3.156) -6.14 (3.041)
1
TP2011
t=2007 ui;t -0.003 (1.942) 3.034 (4.014) -0.02 (2.27)
1
TP2011
t=2007 i;t 0.093 (1.505) -2.87 (2.58) -0.088 (0.727)
higher income and more highly educated households have lower heterogeneity of unobserved tastes.
Table 8.1 shows the results for the choice of loan amount (ln(Li;t)) and default (Di;t). It is
worth noting that the expected income (ln(
Pi;t4),), unemployment risk (ui;t4) and wage volatility
(i;t4) that a¤ect the loan amount decision have a lag of 4quarters, while the variables a¤ecting
default correspond to the current period t. Basically, income, unemployment rates, wage volatility,
households with more members, and debt motivations (especially, the motive of Pay previous debts)
are positively related to loan amounts. Education is negatively related to loan amount, which may
denote a tendency of highly educated households to better manage their …nances over time and
resort less to expensive consumer debt (Table 2 showed that interest rates for consumer loans are
high in Chile). Also, the estimates show that households with strong unobservable tastes for either
Bank plus Retail and Other Debts are more likely to also have a taste for higher loan amounts.
The propensity to default is positively related to high levels of consumer debt relative to annual
income (RDIi;t), higher debt service (RDSIi;t ), unemployment risk, wage volatility, households
with more members, and to loan motivations (especially, the Health needs). In the year 2010 there
was a substantially lower rate of default, even after accounting for the other factors in the model,
32
which could have been due to more selective credit supply policies in the early aftermath of the
…nancial crisis. The choice of loan amount and default behavior are both quadratic in terms of
age, …rst increasing with age and then falling. Default is negatively related to income, but it is not
signi…cantly a¤ected by education, which seems to coincide with recent studies for the USA, which
show that education, apart from math skills, has no signi…cant impact on debt repayment behavior
(Brown, van der Klaauw, Wen and Zafar, 2013). Default is also negatively related to households
with a higher taste for Bank loans. This results justi…es the behavior of Chilean banks in terms
of giving preference to customers with a longer and more exclusive credit history in the banking
system, since those are the households of lower risk. Credit history is also related to default in
other countries as well (Gross and Souleles, 2002, Roszbach, 2004, Edelberg, 2006).
Both the motives "Pay previous debts" and Health needs have a positive e¤ect on loan amount
and the propensity to default. It is also interesting that Health needs has a larger impact on
default than the motive "Pay previous debts", since it con…rms the predictions of economic models.
Economic models of default decision assume that Health expenses are less predictable than other
expenses, since tastes for consumption and past loan commitments are already known to the
household. Therefore health expenses are an unpredictable shock for households and one that
often leads to default even for low amounts of debt (Chatterjee et al., 2007).
Table 8.2 summarizes the estimated heteroscedasticity of the unobserved tastes for loan amount
and default behavior, plus a contemporary loan amount shock which is independent over time.
Again, we can conclude that households of higher income and education are less heterogeneous
in their tastes for loan amount and default behavior. However, the heteroscedasticity of the
contemporary unobserved shock for loan amount is increasing with income and education. This
shows that higher income and highly educated households are less persistent in their indebtedness,
since their loan amounts depend more on contemporary shocks than constant tastes.
6 Conclusions
This paper shows how households’characteristics impact their choice of consumer loans and default
behavior. Low labor market risk (as measured by unemployment risk, job separation rates and wage
volatility) is correlated with a desire for not having consumer debt, while low income is the strongest
33
Table 8.1: Coe¢ cients for the mean loan amount (in log) and propensity for morosity
Exogenous variables Log-loan amount (t0=t4) Propensity to morosity (t0=t)
2007 0.074 (3.7) 1.262 (0.17)
2008 / 09 -0.048 (3.704) 1.326 (0.195)
2010 0.579 (3.34) 0.568 (0.171)
2011 0.089 (3.602) 1.369 (0.173)
Ratio of Debt to Income, RDIi;t 1.291 (0.022)
Ratio of Debt Service to Income, RDSIi;t 0.54 (0.052)
Log-Income: ln(
Pi;t0)6.992 (4.033)-0.385 (0.089)
Years of education of home head -6.598 (3.32) -0.062 (0.126)
Unemployment ui;t00.171 (0.103)2.818 (1.585)
Wage volatility i;t00.157 (0.092)0.854 (0.269)
Out of Santiago 0.067 (0.369) -0.028 (1.405)
Age of home head 0.65 (0.309) 0.109 (0.015)
Age squared -0.013 (0.006) -0.001 (0.005)
2 members in home 0.291 (0.155)0.095 (0.036)
3 or more members 0.317 (0.154) 0.407 (0.236)
Share of loan for "Pay previous debts" 0.588 (0.318)0.376 (0.213)
Share of loan for Health needs 0.184 (0.108)0.778 (0.249)
RE of loan type: "i;b
1 = Bank 1.245 (0.584) -1.674 (0.723)
2 = Bank+Retail 2.012 (0.78) 0.626 (0.726)
3 = Retail 1.013 (0.778) -0.179 (0.728)
4 = Social 1.634 (0.727) -0.162 (0.49)
5 = Other 1.808 (0.794) -0.342 (1.021)
6 = No-Access 0.002 (0.765) 0.261 (0.66)
RE of log-amount:
i0.341 (1.468)
Table 8.2: Coe¢ cients of the standard-deviation (in log) of the random-e¤ects of loan amount and morosity
Exogenous variables Log-loan amount: Log-loan amount: Propensity to morosity:
Random E¤ect
iContemporary shock ~
i;t Random E¤ect i
constant -0.694 (0.489) 0.137 (1.15) -0.125 (0.276)
2008/09 -1.026 (1.063) 0.331 (1.021) 0.045 (0.375)
1
TP2011
t=2007 ln(
Pi;t)-10.702 (3.656) 4.16 (1.251) -1.838 (1.065)
Education -7.032 (3.551) 2.552 (0.487) -0.885 (0.527)
1
TP2011
t=2007 ui;t 0.29 (1.479) 0.185 (0.424) 0.015 (0.934)
1
TP2011
t=2007 i;t 0.124 (0.054) 0.235 (0.663) -0.013 (0.845)
34
cause of a lack of access to debt and credit constraints. Unemployment rates increase the probability
of households opting for all types of consumer loans, but it has a greater impact on lenders who do
not apply credit scoring such as Retail Stores and institutions providing Social Credit.
Loan amounts increase with income, unemployment risk and wage volatility, therefore consumer
loans may help smooth income shocks. The default probability decreases with income and increases
with high levels of indebtedness relative to income, unemployment risk and wage volatility, con…rming
the existence of adverse selection among Chilean debtors. Bank debtors have the lowest risk levels,
which is expected from a lender that applies credit scoring extensively (Edelberg, 2006). Health
needs are positively associated with default behavior, which could denote these expenses are di¢ cult
to predict and insure (Chatterjee et al., 2007). Finally, the probability of getting a loan and the
choice of loan amount is increasing in the number of household members and quadratic in age,
resembling the pro…le of life-cycle consumption (Attanasio and Weber, 2010).
Finally, I show that households are heterogeneous in their loan tastes. This result has broad
implications for policy, since economic shocks or new regulations (or deregulation initiatives)
a¤ecting a certain lenders would have heterogeneous welfare impact across the population.
References
[1] Adams, William, Liran Einav, and Jonathan Levin (2009), “Liquidity Constraints and
Imperfect Information in Subprime Lending,” American Economic Review, 99(1), 49-84.
[2] Attanasio, Orazio, Pinelopi Koujianou Goldberg, and Ekaterini Kyriazidou (2008), "Credit
Constraints In The Market For Consumer Durables: Evidence From Micro Data On Car
Loans," International Economic Review, 49 (2), 401-436.
[3] Attanasio, Orazio and Guglielmo Weber (2010), "Consumption and Saving: Models of
Intertemporal Allocation and Their Implications for Public Policy," Journal of Economic
Literature, 48(3), 693-751.
[4] Banco Central de Chile (2013), "Financial Stability Report," 2nd Semester.
[5] Bertrand, Marianne and Adair Morse (2011), "Information Disclosure, Cognitive Biases, and
Payday Borrowing," Journal of Finance, 66, 1865-1893.
35
[6] Brown, Meta, Wilbert van der Klaauw, Jaya Wen, and Basit Zafar (2013), “Financial
Education and the Debt Behavior of the Young”, Federal Reserve Bank of New York Sta¤
Reports 634.
[7] Chatterjee, Satyajit, Dean Corbae, Makoto Nakajima and José-Víctor Ríos-Rull (2007), "A
Quantitative Theory of Unsecured Consumer Credit with Risk of Default," Econometrica,
75(6), 1525-1589.
[8] Dynan, Karen and Donald Kohn (2007), “The Rise in U.S. Household Indebtedness: Causes
and Consequences,”Finance and Economics Discussion Paper 2007-37, Federal Reserve Board.
[9] Edelberg, Wendy (2004), “Testing for Adverse Selection and Moral Hazard in Consumer Loan
Markets,”Finance and Economics Discussion Paper 2004-9, Federal Reserve Board.
[10] Edelberg, Wendy (2006), “Risk-based Pricing of Interest Rates for Consumer Loans,”Journal
of Monetary Economics, 53, 2283-2296.
[11] Einav, Liran, Mark Jenkins, and Jonathan Levin (2012), "Contract Pricing in Consumer Credit
Markets," Econometrica, forthcoming.
[12] Eurosystem Household Finance and Consumption Network (2009), "Survey Data on Household
Finance and Consumption: Research Summary and Policy Use", ECB Occasional Paper 100.
[13] Girouard, N., M. Kennedy, and C. André (2007), "Has the Rise in Debt Made Households
More Vulnerable?", OECD Economics Department Working Paper 535.
[14] Gross, David B. and Nicholas S. Souleles (2002), "An Empirical Analysis of Personal
Bankruptcy and Delinquency," Review of Financial Studies, 15 (1), 319-347.
[15] Hess, Stephane, Kenneth Train and John Polak (2006), "On the use of a Modi…ed Latin
Hypercube Sampling (MLHS) method in the estimation of a Mixed Logit Model for vehicle
choice," Transportation Research Part B: Methodological, 40(2), 147-163.
[16] Horowitz, Joel L. (2001), "The Bootstrap," in Handbook of Econometrics, edition 1, vol. 5,
chapter 52, 3159-3228, Elsevier.
36
[17] IMF (2006), "Household credit growth in emerging market countries", Global Financial
Stability Report, 46-73.
[18] Ja¤ee, Dwight M. and Thomas Russell (1976), “Imperfect Information, Uncertainty, and Credit
Rationing,”Quarterly Journal of Economics, 90(4), 651–666.
[19] Ja¤ee, D. and J. Stiglitz (1990), "Credit rationing," in B. M. Friedman and F. H. Kahn (eds.)
Handbook of Monetary Economics, vol. 2, Amsterdam: North Holland.
[20] Kahneman, Daniel (2011), "Thinking, Fast and Slow," Farrar, Straus and Giroux Press.
[21] Madeira, Carlos (2014), “Earnings dynamics of Chilean workers: employment transitions,
replacement ratio, and income volatility,”Central Bank of Chile, mimeo.
[22] Marinovic, Alejandra, José Matus, Karla Flores, Nancy Silva (2011), “Systemic risk of
households in Chile”, Revista de Economía Chilena, 14(3), 5-39.
[23] McFadden, Daniel and Kenneth Train (2000), “Mixed MNL models for discrete response,”
Journal of Applied Econometrics, 15(5), 447–470.
[24] Nevo, Aviv (2000), "A Practitioner’s Guide to Estimation of Random Coe¢ cients Logit Models
of Demand," Journal of Economics & Management Strategy, 9(4), 513-548.
[25] Nevo, Aviv (2011), "Empirical models of consumer behavior," Annual Review of Economics,
3, 51-75.
[26] Roszbach, K. F. (2004), “Bank Lending Policy, Credit Scoring and the Survival of Loans,”
Review of Economics and Statistics, 86(4), 946–958.
[27] Shimer, Robert (2012), "Reassessing the Ins and Outs of Unemployment," Review of Economic
Dynamics, 15(2), 127-148.
[28] Train, Kenneth (2009), "Discrete choice methods with simulation," Cambridge University
Press.
[29] Vella, Frank (1998). “Estimating Models with Sample Selection Bias: A Survey”, The Journal
of Human Resources, 33(1), 127-169.
37
