Accepted Manuscript

Risk-based loan pricing consequences for credit unions

Adam G. Walke, , Thomas M. Fullerton, Jr., , Robert J. Tokle, Jr.

PII: S0927-5398(18)30014-8
DOI: https://doi.org/10.1016/j.jempfin.2018.02.006
Reference: EMPFIN 1038

To appear in: Journal of Empirical Finance

Received date : 29 March 2017

Revised date : 8 August 2017
Accepted date : 25 February 2018

Please cite this article as: Walke A.G., Fullerton T.M., Tokle R.J., Risk-based loan pricing
consequences for credit unions. Journal of Empirical Finance (2018),
https://doi.org/10.1016/j.jempfin.2018.02.006

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form.
Please note that during the production process errors may be discovered which could affect the
content, and all legal disclaimers that apply to the journal pertain.
*Highlights (for review)

Highlights:
 Factors affecting credit union adoption of risk-based loan pricing are identified.
 A propensity score matching procedure is used to estimate the impacts of adoption.
 Adoption of risk-based pricing increases the availability of credit union loans.
 For some adopter cohorts, this practice also reduces loan delinquency rates.
Risk-Based Loan Pricing Consequences for Credit Unions

Adam G. Walke a, Thomas M. Fullerton, Jr. b *, Robert J. Tokle c

a
Department of Economics & Finance, University of Texas at El Paso, El Paso, TX 79968-0543,
USA, Telephone: 915-747-7775, Facsimile: 915-747-6282, Email: [email protected]
b
Department of Economics & Finance, University of Texas at El Paso, El Paso, TX 79968-0543,
USA, Telephone: 915-747-7747, Facsimile: 915-747-6282, Email: [email protected]
c
Department of Economics, Idaho State University, Pocatello, ID 83209-8053, USA, Telephone
208-282-3835, Facsimile 208-282-3186, Email [email protected]
* Corresponding author

Abstract: Risk-based pricing, in which interest rate offers vary according to individual borrower
risk levels, has been increasingly used to price credit union loans in the United States. The key
question examined in this research, given credit union not-for-profit objectives, is whether this
pricing strategy increases the availability of loans, particularly for high-risk borrowers. Data on
the number of loans per credit union member and loan delinquency rates are used to assess loan
access and average risk-levels, respectively. The results indicate that risk-based pricing adopters
increase the availability of loans relative to otherwise similar non-adopters. However, average
risk levels, as measured by delinquency rates, appear to be somewhat lower for adopters of risk-
based pricing compared to matched non-adopters. This suggests that lending increases primarily
for low-risk borrowers, which contrasts with claims that risk-based pricing chiefly benefits high-
risk borrowers who might otherwise be denied credit.

Keywords: Risk-Based Pricing; Credit Union Loan Rates; Delinquency Rates; Propensity Score
Matching; Kernel Density Plots; Balanced Covariates

JEL Categories: G21, Financial Institutions; O33 Technological Change: Diffusion Processes;
D22 Empirical Analysis of Firms; D82 Asymmetric Information

Acknowledgements: Financial support for this research was provided by El Paso Water, City of
El Paso Office of Management & Budget, UTEP Center for the Study of Western Hemispheric
Trade, and the Hunt Institute for Global Competitiveness at UTEP. Helpful comments and
suggestions were provided by Paul Hellman and two anonymous referees. Econometric research
assistance was provided by Ernesto Duarte and Omar Solis.
*Blinded Manuscript
Click here to view linked References

Risk-Based Loan Pricing Consequences for Credit Unions

Abstract: Risk-based pricing, in which interest rate offers vary according to individual borrower
risk levels, has been increasingly used to price credit union loans in the United States. The key
question examined in this research, given credit union not-for-profit objectives, is whether this
pricing strategy increases the availability of loans, particularly for high-risk borrowers. Data on
the number of loans per credit union member and loan delinquency rates are used to assess loan
access and average risk-levels, respectively. The results indicate that risk-based pricing adopters
increase the availability of loans relative to otherwise similar non-adopters. However, average
risk levels, as measured by delinquency rates, appear to be somewhat lower for adopters of risk-
based pricing compared to matched non-adopters. This suggests that lending increases primarily
for low-risk borrowers, which contrasts with claims that risk-based pricing chiefly benefits high-
risk borrowers who might otherwise be denied credit.

Keywords: Risk-Based Pricing; Credit Union Loan Rates; Delinquency Rates; Propensity Score
Matching; Kernel Density Plots; Balanced Covariates

JEL Categories: G21, Financial Institutions; O33 Technological Change: Diffusion Processes;
D22 Empirical Analysis of Firms; D82 Asymmetric Information

1. Introduction

With advances in computer technology, data-based procedures for evaluating loan

applications have become ingrained into the underwriting practices of most financial institutions.
The use of automated credit scoring technology to segment loan applicants into risk pools at first
mainly affected decisions regarding whether or not to issue a loan. However, beginning in the
1990s, a large number of banks, and also some credit unions, began to use such technologies to
assign different loan interest rates to applicants on the basis of credit risk (Edelberg, 2006). The
practice of charging different interest rates depending on the risk-level of the individual borrower
is known as risk-based pricing. Figure 1 shows the increased utilization of risk-based loan
pricing mechanisms by credit unions between year-end 2006 and year-end 2015, according to
data reported by the National Credit Union Administration (NCUA, 2016).

1
Figure 1. Percentage of credit unions using risk-based pricing

Asymmetric information problems, in particular when borrowers know their credit risk
but lenders do not, exist in most credit markets. The advances in credit scoring technology that
have facilitated risk-based pricing of interest rates can potentially reduce the severity of adverse
selection problems by providing greater information about the riskiness of individual loan
applicants (Adams et al., 2009; Riley, 1987; Stiglitz and Weiss, 1981). A number of empirical
studies show that risk-based pricing can increase the value of loans issued, especially loans
issued to low-risk applicants, as well as the profitability of financial institutions that employ this
practice (Edelberg, 2006; Einav et al., 2012, 2013).
A separate but related question is whether risk-based lending practices extend credit
access to a wider pool of potential borrowers, including those who might otherwise be denied
credit due to high risk levels. This question is particularly relevant to credit unions, which are
not-for-profit, member-owned financial institutions. Credit unions ostensibly exist to serve the
interests of members, many of whom are borrowers as well as depositors (McKillop and Wilson,
2011). A letter issued by the National Credit Union Administration (NCUA, 1995) indicates that
“credit unions should engage in risk-based lending, not as a means of re-pricing existing balance
sheets, but as a tool to reach out to the under-served and take a risk that might otherwise be
avoided.” A subsequent letter (NCUA, 1999) returns to the same theme: “a properly structured
risk-based lending program… should bring new higher-risk borrowers into the credit union.”
This study examines the track record of risk-based loan pricing toward achieving this objective.
Data from the National Credit Union Administration (NCUA) are used to examine the
effects of risk-based loan pricing on two outcome variables that help quantify access to credit
and outreach to high-risk borrowers: loans outstanding per credit union member and loan
delinquency rates. The data are assembled from call reports and span the period from 2006
through 2015. Credit unions that adopted risk-based loan pricing are paired with otherwise
similar credit unions that maintained traditional uniform loan pricing and differences between the
two groups in terms of the key outcome variables are then analyzed. A propensity score
matching strategy is employed to carry out the analysis.
The paper is organized as follows. The next section provides an overview of research on
risk-based pricing, particularly as it relates to asymmetric information in financial markets. The

2
empirical methodology is then discussed, followed by a summary of variable definitions, data
sources, and descriptive statistics. Empirical results are subsequently presented with practical
implications of the research findings addressed in the concluding section.

2. Risk-Based Loan Pricing

Stiglitz and Weiss (1981) notes that long-standing difficulties in distinguishing loan
applicants who are likely to repay loans from those who are unlikely to do so creates an adverse
selection problem. Only the relatively risky applicants are willing to pay interest rates above a
certain threshold and their propensity to default lowers expected returns to the bank. In this
context, raising interest rates to cover losses only worsens the problem. The authors argue that,
in part to circumvent the adverse selection problem, banks restrict interest rates and the number
of loans issued below the market-clearing level. Riley (1987) points out that, in a context in
which banks can sort loan applicants into distinct groups by level of risk and then charge
different interest rates to each risk pool, restrictions on the supply of credit are likely to be
minimal. Adams et al. (2009) find evidence that this practice, i.e. risk-based pricing, can help
mitigate the adverse selection problem in consumer credit markets.
A critical issue addressed by some of the early studies on risk-based pricing concerns the
extent to which it is actually practiced by financial institutions. Using data from the 1983 Survey
of Consumer Finances, Duca and Rosenthal (1994) find that mortgage interest rates do not vary
with household credit risk, suggesting that risk-based pricing was not widespread at that time.
Getter (2006) explores this question again using data from the 1998 and 2001 Surveys of
Consumer Finances. For the purpose of the latter study, households with records of bankruptcy,
delinquency, or rejected loan applications are defined to have poor credit risk. The study finds
that, by the turn of the millennium, interest rates did reflect differences in credit risk.
Technological innovations such as automated credit-scoring help account for this transition.
Edelberg (2006) documents that risk-based pricing of interest rates on consumer loans
became increasingly common in the 1990s. For mortgage, automobile, and credit card loans, the
spread between the interest rates paid by borrowers with high and low levels of bankruptcy risk
increased substantially after 1995. Lower interest rates for borrowers with relatively low
bankruptcy risk led, in turn, to increased debt levels among those consumers. The effects of risk-
based pricing on consumers with high levels of bankruptcy risk is less straightforward. Risk-
based pricing generally increased access to credit among high-risk borrowers, but higher interest
rates reflecting the elevated risk levels of that group meant that average borrowing levels
decreased for some loan types. In a similar study conducted for Italy, Magri and Pico (2011)
document how mortgage interest rates vary according to the probability of household mortgage
delinquency.
The studies cited in the two preceding paragraphs use data from large-scale household
surveys to study risk-based pricing of consumer loans. Financial institutions represent another
important source of information for the study of this topic. Using data from surveys of banks,
Berger et al. (2005) find that banks evaluating applications for small business credits under
$100,000 tend to have higher lending volumes if they utilize credit scoring technologies. The
use of credit scoring is also associated with higher average interest rates and higher risk levels,

3
suggesting that this technology increases access to credit primarily for riskier firms that pay
higher prices for credit. Tokle et al. (2015) examine the effects of risk-based lending and other
factors on credit union loan interest rates following the 2008 financial crisis. This pricing
strategy is found to result in higher interest rates for vehicle loans, but somewhat lower interest
rates for unsecured credit card loans.
Another group of studies uses data obtained from an auto finance company before and
after it began employing credit score information to implement risk-based pricing of down
payments (Adams et al., 2009; Einav et al., 2012). In one of these studies, Einav et al. (2013)
report that loan originations actually decreased after the lender implemented credit scoring,
although profit per applicant increased by approximately 42 percent. With risk-based pricing,
applicants with a high default risk faced substantially higher down payment requirements,
resulting in a marked drop in close rates for this group. Close rates remained relatively stagnant
for low-risk applicants although loan profitability increased because this group of applicants was
allowed to take on larger loans. This study illustrates that risk-based pricing may not necessarily
result in larger customer bases.
The implications of risk-based pricing of interest rates for different groups of consumers
depend partly on how ‘risk’ is defined. Deng and Gabriel (2006) find evidence that risk-based
pricing of mortgage loans could lead to lower housing finance costs for underserved borrowers if
the risk of early loan repayment (prepayment) is taken into account. That is because prepayment
risk is found to be comparatively low for borrowers with credit scores below 620 as well as for
Black and Hispanic mortgagors. Those groups of borrowers also have elevated probabilities of
loan default but that is more than offset by markedly lower prepayment propensities resulting in
lower overall risk of loan termination. When mortgages are stratified and priced on the basis of
risk, borrowers having a combination of high default risk and low prepayment risk are simulated
to pay reduced mortgage interest rates. However, these results are not generally consistent with
the empirical regularities documented in the studies cited above in which higher default risk is
correlated with higher interest rates under risk-based pricing.
The previous literature on risk-based pricing of interest rates generally indicates that it
can increase the value and profitability of loans issued by offering more attractive terms to loan
applicants with low default risk while raising interest rates for high-risk applicants. There is less
consensus on whether risk-based pricing facilitates access to credit for a larger number of
potential borrowers, especially for those with relatively high levels of credit risk. The main
purpose of this effort is to examine that question using data for United States credit unions. Prior
literature on risk-based pricing of loans is relatively scant (Magri and Pico, 2011), and this is
particularly true for loans issued by credit unions. Tokle et al. (2015) examine the effects of
risk-based lending on credit union interest rates. However, the existing academic literature does
not specifically examine the effects of credit union risk-based lending on two other key
variables: the number of loans and loan delinquency rates. This study attempts to at least
partially fill that gap in the literature.

3. Empirical Methodology

4
 For the purpose of this analysis, risk-based pricing adoption is considered a treatment and
its impacts are quantified as treatment effects. The binary treatment indicator D, is equal to one
for credit unions that adopt risk-based pricing in period t and zero for non-adopters. The value of
an outcome variable, such as the delinquency rate, under treatment is Y1 and the corresponding
value of the outcome in the absence of treatment is Y0. The average treatment effect on the
treated (ATET) can be defined as shown in Equation (1).

(1)

Equation (1) identifies the impact of adopting risk-based loan pricing as the difference between
the outcome observed for treated units and the outcome that would have resulted if the same
units had been left untreated. However, this formula cannot be directly applied for
computational purposes because the latter outcome, , is not observed (Caliendo and
Kopeinig, 2008). One approach to estimating ATET involves constructing a control group that is
similar to the treatment group in terms of key observed characteristics and then quantifying the
expected difference between outcomes for the treatment and control groups.
The fundamental difficulty in applying this approach to treatment effect estimation lies in
finding control units that are truly comparable with the treated units. It is relatively
straightforward to identify treated and control units with similar values of a single variable, such
as age. It is more complicated to match those units on the basis of several different
characteristics simultaneously. One solution is to match on the basis of a single propensity
score, which is a function of one or more characteristics, often called covariates (Rosenbaum and
Rubin, 1983; Rosenbaum, 1995). The propensity score matching estimator for ATET is shown
in Equation (2), where X is a covariate and P(X) is the propensity score (Caliendo and Kopeinig,
2008).

(2)

There are two key assumptions for identification of the average treatment effects on the treated.
First, the distribution of potential outcomes under the control treatment, Y0, is independent of
assignment into treatment and control groups conditional on the relevant covariates. One
implication of this assumption is that all the covariates that influence both treatment assignment
and the potential outcomes must be observed. The second assumption states that assignment to
treatment is not perfectly predicted. The latter helps ensure that there is overlap in the range of
propensity scores for treated and control units so that there are suitable potential matches for
each treated unit (Imbens, 2004).
Before implementing (2), a propensity score must be estimated. The propensity scores
utilized in this analysis are the predicted adoption probabilities obtained from probit regressions
of D on the set of covariates. The general probit model for estimating propensity scores is shown
in Equation (3), where Φ denotes the normal cumulative distribution function.

(3)

5
The goal of estimating a propensity score is not to predict treatment as accurately as possible, but
rather to efficiently control for confounding variables, those correlated with both the treatment
and the outcomes (Brookhart et al., 2006). As such, the selection of variables for inclusion in the
model is guided by knowledge of which variables are likely to be correlated with the treatment
and the outcomes (Stuart, 2010). It is also important that the covariates be unaffected by the
treatment. For this reason it is desirable to use covariates measured before the treatment was
chosen (Imbens, 2004). Once propensity scores have been estimated, they are then used to
construct a control group of non-adopters that are similar to the adopters.
Several alternative matching estimators can be employed to identify control units that are
comparable with treated units on the basis of propensity scores. In comparing these various
methods, there is sometimes a tradeoff between the goals of matching the units that are most
similar to one another on average and improving the efficiency of the estimates. One approach,
known as nearest neighbor matching, pairs treated units with the control units that are closest in
terms of the propensity score (Caliendo and Kopeinig, 2008). Huber et al. (2013) find that the
nearest neighbor matching estimator has relatively low bias but also higher variability than
competing matching estimators. Rosenbaum (2010) argues that reducing bias should be the
primary concern in choosing a matching approach. In this analysis, the nearest neighbor
algorithm is used to estimate baseline treatment effects and these are compared with results
obtained by alternative approaches. The nearest neighbor matching algorithm has previously
been used in the finance literature (Goenner, 2016; Hellman et al., 2008).
Once control units have been matched with treated units, the balance of covariates across
treated and control groups is assessed (Dehejia and Wahba, 2002). The aim of propensity score
matching is to balance the values of covariates so that the conditional distribution of the
covariates given the propensity score is the same for treated (D = 1) and control (D = 0) units.
Equation (4) describes the balancing property of the propensity score (Rosenbaum, 2010;
Rosenbaum and Rubin, 1983).

(4)

Several statistics can be utilized to evaluate covariate balance. One of these is a two-
sample t-test to determine if there are statistically significant differences in the covariate means
across the treated and control subsamples. Another statistic that can be used to assess matching
quality is the standardized bias, abbreviated SB, which is shown in Equation (5), where and
are the covariate means for adopters and non-adopters, respectively, while and are the
respective variances (Rosenbaum and Rubin, 1985).

(5)

There is no general agreement regarding the maximum acceptable standardized bias, but a
threshold of 0.10 or 0.25 is sometimes used (Austin, 2009; Garrido et al., 2014). Compared to t-
tests and other statistical hypothesis tests, the standardized bias may serve as a better indicator of
covariate imbalance in small samples. Finally, pseudo-R2 values can be re-calculated for only
the matched sample. Ideally, the pseudo-R2 for this subsample should be close to zero,

6
indicating that the covariates cannot help predict assignment into treatment and control groups
within the matched sample (Caliendo and Kopeinig, 2008).
The goal of the matching approach is to determine whether credit unions that adopt risk-
based lending have different values of the outcome variables than otherwise similar credit unions
that do not adopt this practice. The null hypothesis tested is that the treatment effects are equal
to zero. The outcome variables, loans per credit union member and loan delinquency rates, are
selected to shed light on the question of whether implementation of risk-based loan pricing
increases the number of borrowers, particularly those with elevated delinquency risk. If the
values of the outcome variables diverge significantly for adopters versus matched non-adopters,
this implies that the effect is most likely attributable to adoption.
Risk-based pricing is hypothesized to increase the number of loans issued by providing
better terms to low-risk borrowers and, possibly, by expanding access to loans for high-risk
borrowers who would otherwise be excluded from credit markets (Edelberg, 2006). The
literature is less conclusive regarding how loan delinquency rates will react to adoption of risk-
based pricing. Berger et al. (2005) find that the adoption of credit scoring for small business
loans is associated with higher interest rates and greater risk levels, suggesting that the main
beneficiaries of this practice are relatively high-risk customers who pay higher prices for credit.
If this is true of credit union loans under risk-based pricing, it would be expected that this
practice is associated with higher loan delinquency rates.
ATET is calculated as the mean difference for each outcome variable, Y, between treated
units and the matched control units. To determine the statistical significance of treatment effects,
standard errors are calculated using the method proposed by Abadie and Imbens (2006, 2011,
2016). In this approach, the variance of ATET is adjusted to account for the fact that, in the
initial step above, the propensity scores are estimated rather than being known.

4. Data

Credit union data are obtained from fourth quarter call reports published by the National
Credit Union Administration (NCUA) over the period from 2006 through 2015. The length of
the dataset is limited by the availability of consistent data on the key variable of interest, an
indicator equal to one for credit unions that practice risk-based lending and zero otherwise. This
variable, labeled account 879e, was first added to the call reports in 2006 and has been reported
on a consistent basis since that time. Because the key empirical question pertains to the effects
of adopting risk-based pricing, account 879e is used to develop new indicator variables that are
equal to one for credit unions that implement this loan pricing strategy in a given period t.
In order to separately analyze different cohorts of adopters, the ten-year sample period is
split into a series of overlapping sub-sample periods. Each sub-sample contains data on a group
of credit unions over the course of five years from the end of the year prior to adoption (t-1)
through the end of the third year after adoption (t+3). Credit unions without complete data over
the full five-year sub-period are excluded from the analysis. Because some credit unions are
established or dissolved each year, examining five year sub-samples allows a larger number of

7
observations to be included in the empirical analysis than the alternative strategy of only
analyzing credit unions with complete data over the full ten-year period.
Adopters are defined as credit unions that did not employ risk-based pricing at the end of
year t-1 but did report using this practice at the end of each of years t through t+3. Non-adopters
are defined as credit unions that did not report using this loan pricing strategy at the end of any
year from t-1 through t+3. The subsamples exclude credit unions that instated risk-based pricing
between year t+1 and year t+3, those that used the strategy consistently throughout the whole
five-year sub-period, and those that had such a policy but then discontinued it. Pana et al. (2015)
employ a similar strategy to define ‘adopters’ and ‘non-adopters’ in a study of transactional
website adoption by credit unions.
The impacts of adoption are assessed for two outcome variables. The first of these is the
ratio of delinquent loans to total loans (Pana et al., 2015). The second outcome variable, which
measures lending activity, is the ratio of loans outstanding to total credit union membership. The
NCUA (2013) calls this variable the borrowers-to-members ratio. The numerator of the ratio is
the number of loans, rather than the dollar amount of loans, because the critical question in this
analysis is whether risk-based pricing increases access to loans for credit union members.
Making credit available to members who might not otherwise qualify is a critical rationale for
the adoption of risk-based pricing by credit unions (NCUA, 1995, 1999). The number of loans
issued is an important measure of access to credit in general (Stiglitz and Weiss, 1981) and is
likely a more appropriate measure of loan availability than the dollar value of loans in the case of
credit unions. The denominator of the ratio is total membership because adopters of risk-based
pricing, with 10,504 members on average, tend to be substantially larger than non-adopters, with
4,466 members on average (see Figure 2). Dividing the number of loans by the number of
members helps control for credit union size. Average loan size is considered as an alternative
outcome indicator and results regarding that variable are briefly discussed in the next section.
The propensity score matching methodology used in this analysis requires selection of
covariates for inclusion in the probit models. Because implementation of risk-based loan pricing
is facilitated to a large extent by access to automated credit scoring technologies, the previous
literature on determinants of technology adoption in the financial sector serves as a guide for
selecting potential covariates. Observed correlations with the dichotomous adoption indicator
and the outcome variables serve as additional criteria for selecting covariates. The nine
covariates ultimately chosen include six characteristics of credit unions and three characteristics
of the markets in which they operate.

8
 .4
.3
Density

.2
.1
0

0 5 10 15
Ln(Total Credit Union Membership)

Risk-Based Pricing Adopters

Non-Adopters
kernel = epanechnikov, bandwidth = 0.2712

Figure 2. Kernel density plots of total membership for adopters and non-adopters

One key determinant of technology adoption is institutional size. Hannan and McDowell
(1984) point out that economies of scale along with differences in managerial attitudes and
relative risk exposure for institutions of different sizes may affect the likelihood of adopting new
technologies. As in several previous studies of technology adoption, size is measured by the
natural logarithm of total assets (Akhavein et al., 2005; Dow, 2007; Hernández-Murillo et al.,
2010; Pana et al., 2015). Similarly, the net worth ratio and the fee income ratio reflect other
relevant credit union characteristics. Pana et al. (2015) consider the net worth ratio as an
indicator of credit union conservativeness and expect it to be inversely related to technology
adoption. Similarly, Goenner (2016) argues that the net worth ratio can proxy for managerial
attitudes towards risk and return. Akhavein et al. (2005) consider ‘fee income,’ measured by
non-interest income as a share of operating revenue, to be a proxy for the aggressiveness of a
financial institution’s business strategy, which may relate to its openness to new practices such
as credit scoring.

Table 1. Summary statistics for pooled cohorts

Standard Mean for Mean for
Mean Difference
Deviation Adopters Non-adopters
Treatment Variable:
Risk-based pricing 0.11 0.31 1.00 0.00 1.00
Covariates:

9
ln (Total assets) 15.69 1.81 16.86 15.55 1.32
Net worth ratio 16.68 7.67 13.85 17.03 -3.19
Fee income ratio 8.50 9.46 15.49 7.64 7.85
Regular shares ratio 68.76 28.89 49.40 71.14 -21.74
Transactional website 0.35 0.48 0.67 0.31 0.36
Market share (local) 50.50 26.69 37.85 52.06 -14.20
Market share (state) 30.04 12.48 30.45 29.99 0.45
Loan interest rate 7.45 2.11 7.31 7.46 -0.15
Per capita income $40.929 $14.501 $39.672 $41.084 -$1.412
Outcome Variables:
Loans/Members Ratio t-1 37.55 20.57 43.18 36.86 6.32
Loans/Members Ratio t 37.45 21.53 44.25 36.61 7.64
Loans/Members Ratio t+1 37.33 20.79 44.29 36.48 7.82
Loans/Members Ratio t+2 37.41 20.82 44.72 36.51 8.22
Loans/Members Ratio t+3 37.41 21.18 44.94 36.48 8.46
Delinquency Rate t-1 2.84 5.61 1.87 2.96 -1.09
Delinquency Rate t 2.94 5.84 2.00 3.05 -1.06
Delinquency Rate t+1 2.97 5.74 1.95 3.09 -1.14
Delinquency Rate t+2 2.98 5.94 1.94 3.11 -1.17
Delinquency Rate t+3 2.95 6.14 1.84 3.08 -1.24
Note: Summary statistics are for the subsample with complete data for all variables listed. For
the sake of brevity, the data for all six cohorts are pooled together for the purpose of this table,
resulting in 15,673 observations (which is equal to the sum of the observations for all cohorts
reported in Table 2 and also equals the number of propensity scores that are estimated). Per
capita income data are reported in thousands of dollars.

The credit union’s ‘product mix’ has also been employed as a determinant of technology
adoption by banks (Hannan and McDowell, 1984). For credit unions, a relevant measure of
product mix is the ratio of regular member shares to total shares and borrowings (NCUA, 2013).
More traditional credit unions, with a higher proportion of regular shares as compared to money
market shares, IRA/KEOGH accounts, non-member deposits, and borrowed money, are
hypothesized to be less likely to adopt new lending technologies. Conversely, credit unions that
use a transactional website are hypothesized to be more inclined to adopt such technologies.
Finally, the loan interest rate is included as a covariate primarily because of its theoretical
relationship with outcome variables such as lending activity (Fase, 1995). As mentioned above,
covariates used to estimate propensity scores should include those correlated with the outcome
variables as well as with the risk-based lending adoption indicator.
The three covariates that reflect market characteristics are county-level per-capita
personal income plus local- as well as state-level market share variables. Hannan and McDowell

10
(1984) find that wages positively affect the probability of automatic teller machine adoption.
Similarly, income levels are likely to affect adoption of new loan pricing strategies. Previous
literature also indicates that the extent of competitive pressures present in financial markets
affects the propensity to adopt new technology (Bofondi and Lotti, 2006; Hannan and
McDowell, 1984; Hernández-Murillo et al., 2010). Competing theories make different
predictions regarding the direction of market concentration effects on technological adoption
proclivities (Akhavein et al., 2005). The local market share variable employed in this analysis is
the ratio of membership in an individual credit union to total potential membership. The other
market share indicator is statewide total membership in all credit unions as a percentage of state
population. The latter has previously been used as a measure of market structure in credit union
research (Feinberg, 2001; Feinberg and Rahman, 2006).
The above-mentioned explanatory variables are employed in separate analyses for each
cohort of adopters. The probability of adoption is predicted in years 2007 through 2012. The
choice of these years for the analysis is conditioned by data availability and the research design
which requires analyzing data from one year prior to adoption until three years afterwards. Prior
propensity score matching studies in the finance literature that use a similar strategy of
developing separate models for different years include Li and Zhao (2006), Behr and Heid
(2011), and Pana et al. (2015). The latter two studies also follow the strategy of analyzing the
null hypothesis that treatment effects are equal to zero over successive multi-year windows. A
null hypothesis of zero treatment effects has the advantage of being relatively easy to interpret
while directly addressing the question of whether adopters differ from matched non-adopters in
terms of the key outcome variables.
Data for the six variables representing credit union characteristics as well as the ratio of
actual to potential credit union membership are obtained from NCUA call reports. The average
loan rate is calculated by dividing calendar-year income earned from interest on loans (less
interest refunded to borrowers) by the total amount of all loans outstanding. State credit union
membership data are from the Credit Union National Association. Per-capita income data for the
county where each credit union is headquartered and state-level population are obtained from the
Bureau of Economic Analysis. All of the independent variables are lagged one year due to the
potential existence of feedback effects when contemporaneous data are employed (Caliendo and
Kopeinig, 2008; Pana et al., 2015). Lagged explanatory variables have also been used to avoid
endogeneity problems in the analysis of credit scoring adoption and lending behavior (Berger et
al., 2005). This is consistent with the convention that covariates should be measured prior to
treatment (Rosenbaum, 2010).
Table 1 presents summary statistics for the treatment variable, the nine covariates, and
the two outcome variables. The statistics shown are calculated for the subsample with complete
data on all variables. In addition, the last three columns of the table provide means for risk-
based lending adopters and non-adopters and the differences between these means. The net
worth ratio for non-adopters is 17 percent, on average across all cohorts, which is substantially
above the average of 13.9 percent for adopters and 14.4 percent for all credit unions. Adopters
have substantially higher fee income ratios and are more than twice as likely to have
transactional websites as compared with non-adopters (67 percent versus 31 percent). Adopters
also have lower local market shares, lower ratios of regular shares to total shares and borrowings,

11
and operate in counties with lower average per capita income levels ($39,672 versus $41,084 for
non-adopters). Finally, the adopters also tend to be much larger, with average assets of $97
million versus $46 million for non-adopters (values for this variable are expressed in natural
logarithms in Table 1).
Summary statistics for the outcome variables are shown for each pooled sub-sample
period t-1 through t+3 (for example, the t-1 data for the 2007-2012 cohorts of risk-based pricing
adopters are pooled over 2006-2011 and the t+3 data for the same cohorts are pooled over 2010-
2015). On average, across all cohorts of adopters, the ratio of credit union loans to members
fluctuates between 37.3 percent and 37.6 percent. Adopters of risk-based pricing offer more
loans in proportion to credit union membership even before the year of adoption. However, the
gap relative to non-adopters increases from 6.3 percentage points in the year prior to adoption to
8.5 percentage points in third year after adoption. Adopters also had lower mean delinquency
rates prior to adoption and these rates subsequently decline further with respect to non-adopters,
possibly suggesting lower average risk levels within the adopter group. Under risk-based
pricing, lower risk levels would tend to be associated with lower interest rates. It should be
noted, in this regard, that adopters do tend to charge somewhat lower average loan rates.

5. Empirical Results

Probit estimation results are presented in Table 2. Larger credit unions are significantly
more likely to adopt risk-based loan pricing, which is consistent with previous research on
technology diffusion in this sector (Dow, 2007). Credit unions that have lower net worth ratios
and are more dependent on fee income are also generally more likely to adopt risk-based pricing.
The negative coefficients for net worth are consistent with the argument that high net worth
ratios proxy for conservative attitudes towards risk and innovation (Pana et al., 2015). Similarly,
the positive relationship between the fee income ratio and the probability of adoption seems to
corroborate the argument that fee income proxies for the aggressiveness of the financial
institution’s business strategy (Akhavein et al., 2005).
The ratio of regular shares to total shares and borrowings has negative coefficients in all
cases, indicating that credit unions with a more traditional deposit mix, one dominated by regular
shares, are less likely to adopt risk-based loan pricing strategies. For the 2007 and 2008 cohorts,
having a transactional website helps predict adoption of risk-based lending technologies,
although this relationship becomes weaker for later cohorts. The positive coefficients for all but
one cohort suggest that having a transactional website may serve as a proxy for technological
sophistication. Credit unions with expertise in implementing new technologies may be better
equipped to deploy such resources in the service of innovative lending strategies.
The coefficients on local market shares indicate that credit unions with a larger presence
in the local market are less likely to adopt risk-based pricing. The other measure of market
share, statewide membership in credit unions as a percentage of state population, is also
inversely related to the probability of adoption for all but one cohort. These results run counter
to those documented in a number of studies which find that greater market concentration is
associated with increased propensity to adopt new technologies (Bofondi and Lotti, 2006;
Hannan and McDowell, 1984). On the contrary, the results suggest that credit unions facing high

12
levels of competition from other credit unions and from banks are those with the highest
proclivity towards new technology adoption. Competition from other financial institutions likely
provides credit unions with greater incentives for innovation.

Table 2. Probit regression results

Cohort 2007 2008 2009 2010 2011 2012
Constant -4.3782 -3.1576 -2.9137 -2.8165 -2.4168 -2.1640
(-7.96)** (-5.34)** (-5.49)** (-4.08)** (-2.71)** (-3.27)**
ln (Total assets) 0.2046 0.1200 0.0942 0.1046 0.0734 0.0907
(7.54)** (4.03)** (3.59)** (3.12)** (1.70)* (2.77)**
Net worth ratio -0.0159 -0.0082 -0.0146 -0.0072 -0.0090 -0.0155
(-2.58)** (-1.34) (-2.86)** (-1.05) (-0.97) (-2.12)**
Fee income ratio 0.0285 0.0243 0.0285 0.0217 0.0076 0.0128
(7.58)** (5.73)** (7.98)** (5.06)** (1.39) (2.97)**
Regular shares ratio -0.0017 -0.0042 -0.0047 -0.0081 -0.0051 -0.0029
(-1.05) (-2.45)** (-3.15)** (-4.21)** (-2.08)** (-1.50)
Transactional website 0.3062 0.2311 0.1099 0.1235 -0.0334 0.1495
(3.69)** (2.45)** (1.27) (1.13) (-0.24) (1.35)
Market share (local) -0.0059 -0.0045 -0.0029 -0.0003 -0.0069 -0.0056
(-4.78)** (-3.28)** (-2.42)** (-0.18) (-3.23)** (-3.42)**
Market share (state) -0.0044 -0.0020 -0.0023 -0.0123 0.0055 -0.0002
(-1.66)* (-0.75) (-0.96) (-3.11)** (1.41) (-0.06)
Loan interest rate 0.0823 0.0622 0.1040 0.0718 0.0477 0.0334
(4.40)** (3.19)** (6.03)** (2.88)** (1.54) (1.53)
Per capita income -0.0075 -0.0044 -0.0002 -0.0047 -0.0084 -0.0088
(-2.75)** (-1.74)* (-0.11) (-1.39) (-1.78)* (-2.59)**
Observations 3,264 2,956 2,850 2,449 2,105 2,049
2
LR χ 623.20 325.99 356.75 202.72 68.27 156.56
Probability value <0.0001 <0.0001 <0.0001 <0.0001 <0.0001 <0.0001
Pseudo R2 0.2332 0.1639 0.1429 0.1490 0.0942 0.1160
Notes: z-statistics are in parentheses; * surpasses 10% critical value of the standard normal
distribution; ** surpasses 5% critical value of the standard normal distribution

Credit unions with higher loan interest rates are also more likely to adopt risk-based loan
pricing. Previous studies have noted a positive correlation between bank profitability and
technology adoption (Akhavein et al., 2005; DeYoung et al., 2007). Those efforts indicate that
less profitable firms may face liquidity constraints that inhibit investment in new technologies. It
is possible that higher loan interest rates are associated with greater profitability and, indirectly,

13
greater investment in new lending technologies. It should be noted however, that loan interest
rates are included in the model mainly because they affect the outcome variables and not
primarily because of their relationship with risk-based pricing adoption. The same can be said of
per-capita income. The negative coefficients on the latter variable probably indicate that credit
unions in low income areas are more prone to seek mechanisms that allow identification of, and
differentiated pricing for, high risk customers.
The primary objective of developing propensity scores through the probit modeling
exercise is to identify non-adopters that are similar to the adopters in terms of the key covariates.
The baseline matching procedure utilized is nearest neighbor one-to-one matching with
replacement. In one-to-one matching, risk-based pricing adopters are paired with only the single
most similar control unit in terms of propensity scores. In matching with replacement, the same
non-adopter can be matched to more than one adopter, allowing treatment units to be matched to
the most similar control unit, even if that unit has already been matched with another treatment
unit. Dehejia and Wahba (2002) note that one-to-one matching and matching with replacement
generally tend to reduce bias, though at the cost of somewhat less precise estimates.
As a preliminary step, it is crucial to check the balance of covariates across treated and
control groups. Before matching, the standardized differences between the mean values of
covariates for risk-based pricing adopters and non-adopters frequently exceed 50 percent and, in
some cases, are close to 100 percent. Table 3 shows that the standardized differences between
adopters and the subset of non-adopters comprising the control group are much smaller. In 49
out of 54 cases, the standardized bias statistics are below the 10 percent threshold and in all cases
the bias is below 15 percent. The matched sample R2 and likelihood ratio chi-square statistics
both indicate that, within the matched subsample, the covariates do not help predict whether a
credit union will adopt risk-based pricing. This suggests that the treated and the matched control
units have, on average, similar values for the observed covariates.

Table 3. Standardized bias statistics for the matched sample

2007 2008 2009 2010 2011 2012
ln (Total assets) 0.1% -1.9% -4.2% 5.5% -6.7% 3.9%
Net worth ratio 7.8% 2.1% -2.5% 4.8% 0.6% 2.3%
Fee income ratio -8.9% 1.8% -2.2% -0.8% 12.1% 3.2%
Regular shares ratio 3.5% 5.8% 0.9% 0.9% 9.3% 1.8%
Transactional website 9.4% -1.4% 1.8% 4.4% 2.3% 4.0%
Market share (local) -8.0% -3.1% -1.9% -13.2% -2.8% 2.9%
Market share (state) -0.1% 1.3% -2.0% 10.2% 10.1% 4.8%
Loan interest rate 0.5% 8.9% 5.8% 1.4% 4.5% -6.6%
Per capita income -7.7% 11.7% -4.7% -5.8% 3.0% 2.1%
Matched sample R2 0.009 0.005 0.002 0.008 0.012 0.003
Matched sample LR χ2 11.12 4.29 3.04 4.57 2.97 1.66
Probability value 0.268 0.892 0.963 0.870 0.965 0.996

14
 To further check that the balancing property is satisfied, two-sample t-tests for equality of
covariate means are also conducted. Before matching, the t-statistics are, in most cases,
significantly different from zero, indicating that there are significant differences between the
covariate means for adopters versus non-adopters. However, the differences between adopters
and the control group of non-adopters constructed via the matching procedure are always
statistically insignificant, as shown in Table 4. This corroborates the results presented in Table 3
and indicates that an adequate degree of covariate balance has been achieved.
It is also important to verify that the overlap assumption is satisfied so that adopters can
be matched with non-adopters that are truly comparable in terms of propensity scores (Heckman
et al., 1997). To check the validity of this assumption, kernel density plots are developed
showing the distribution of estimated propensity scores for treated versus control units (the plots
are available from the authors upon request). For all cohorts, the density plots demonstrate a
sufficient degree of overlap, ensuring that there are adequate matches among the control units for
all of the treated units.

Table 4. Two sample t-tests for treated and control groups

2007 2008 2009 2010 2011 2012
ln (Total assets) 0.02 -0.24 -0.63 0.54 -0.44 0.40
(0.984) (0.810) (0.528) (0.589) (0.658) (0.687)
Net worth ratio 1.19 0.26 -0.38 0.47 0.04 0.23
(0.234) (0.793) (0.703) (0.638) (0.971) (0.815)
Fee income ratio -1.35 0.22 -0.33 -0.08 0.80 0.33
(0.176) (0.824) (0.743) (0.938) (0.425) (0.744)
Regular shares ratio 0.54 0.72 0.14 0.09 0.62 0.18
(0.592) (0.471) (0.890) (0.932) (0.539) (0.858)
Transactional website 1.43 -0.17 0.27 0.43 0.15 0.41
(0.154) (0.862) (0.784) (0.667) (0.880) (0.685)
Market share (local) -1.22 -0.39 -0.29 -1.31 -0.19 0.29
(0.221) (0.695) (0.774) (0.192) (0.851) (0.769)
Market share (state) -0.02 0.16 -0.30 1.01 0.67 0.49
(0.982) (0.870) (0.761) (0.314) (0.504) (0.622)
Loan interest rate 0.08 1.11 0.88 0.14 0.30 -0.67
(0.934) (0.267) (0.379) (0.888) (0.767) (0.500)
Per capita income -1.17 1.46 -0.70 -0.58 0.20 0.21
(0.242) (0.144) (0.481) (0.564) (0.844) (0.834)
Note: Probability-values are in parentheses

Having verified the quality of nearest neighbor one-to-one matching procedure, treatment
effects are computed for each cohort of adopters from the year prior to adoption until the third

15
year after adoption. The results are shown in Tables 5 and 6. For each cohort, risk-based
lending adopters account for half of the observations and the control group of non-adopters
comprises the other half. It is noteworthy that the number of observations decreases markedly
after the 2007-2009 recession, which reflects a large decline in the number of risk-based lending
adopters. Reluctance to adopt new technologies may result from more binding liquidity
constraints during periods of weak economic performance (Akhavein et al., 2005) or from
increased caution towards financial innovations in the wake of the banking crisis. This contrasts
with the observation of Magri (2015), in a study conducted for Italy, that risk-based pricing of
interest rates expanded after the economic crisis.
Table 5 reports the mean ratios of loans outstanding to total credit union membership for
adopters and for the associated control groups of non-adopters. The differences between these
two means are the estimated treatment effects. The overall differences in the mean loan-to-
member ratios between adopters and non-adopters are statistically significant at the 5-percent
level in a majority of cases. For all cohorts, adopters supplied more loans, in proportion to
membership, than paired non-adopters even before implementing risk-based pricing. However,
the differences between the loan ratios of these two groups grow in magnitude and statistical
significance over the course of the five year observation periods. To evaluate these differences
over time, another t-test is completed (not shown) replacing the null hypothesis of a zero
treatment effect with the hypothesis that the treatment effect will remain unchanged after year t-
1. The latter t-tests do not indicate a significant change with respect to year t-1. However, the
estimated differences with respect to the year prior to adoption are sometimes large in practical
terms, reaching 1.0 and 2.9 percentage points by the end of the third year after adoption.
The positive treatment effect estimates seem to corroborate the argument that knowledge
of the risk levels associated with individual borrowers allows financial institutions to expand
access to credit (Edelberg, 2006; Riley, 1987). However, the effects of adopting risk-based
pricing estimated using the matching procedure are smaller than might be assumed by simply
comparing the loan ratio differentials between adopters and all non-adopters (including those
that are highly dissimilar to adopters). The latter differentials are shown in Table 1 for the
pooled sample.

16
TABLE 5. Average treatment effect on the treated: Loan-to-member ratio
Cohort Year Treated Mean Control Mean ATET t-statistic Observations
t-1 43.9895 42.2369 1.7525 (1.29) 930
2007 t 44.3414 41.4621 2.8793 (2.15)** 930
t+1 44.4721 41.1731 3.2990 (2.43)** 930
t+2 45.8256 41.1441 4.6815 (2.96)** 930
t+3 45.4455 40.7986 4.6469 (3.01)** 930
t-1 43.9273 42.5384 1.3889 (0.90) 622
2008 t 44.4212 41.9982 2.4230 (1.44) 622
t+1 46.2849 43.2795 3.0054 (1.48) 622
t+2 45.9668 42.2302 3.7366 (2.05)** 622
t+3 45.5780 42.0777 3.5003 (2.01)** 622
t-1 40.8308 39.0951 1.7357 (1.30) 906
2009 t 43.4670 39.3786 4.0884 (1.77)* 906
t+1 42.3910 38.4218 3.9692 (2.44)** 906
t+2 41.3351 38.3454 2.9897 (2.23)** 906
t+3 41.7286 38.9507 2.7780 (2.00)** 906
t-1 45.4139 40.1873 5.2266 (2.71)** 390
2010 t 46.1958 40.3017 5.8941 (2.79)** 390
t+1 44.4132 39.7501 4.6631 (2.43)** 390
t+2 45.1936 39.9680 5.2256 (2.67)** 390
t+3 46.5379 40.3137 6.2241 (3.14)** 390
t-1 42.4432 39.0097 3.4334 (1.64) 174
2011 t 42.9009 38.3233 4.5776 (2.01)** 174
t+1 42.7246 38.7545 3.9701 (1.72)* 174
t+2 44.1115 39.0169 5.0946 (2.10)** 174
t+3 45.4620 39.9026 5.5594 (2.07)** 174
t-1 43.5882 40.9183 2.6699 (1.56) 418
2012 t 44.2750 40.6758 3.5993 (2.14)** 418
t+1 45.5946 41.0529 4.5417 (2.75)** 418
t+2 47.5880 43.0763 4.5117 (2.38)** 418
t+3 48.0926 42.7887 5.3039 (2.55)** 418
Notes: t-statistics are in parentheses; * surpasses 10% critical value of the t-distribution; **
surpasses 5% critical value of the t-distribution. ATET may not exactly equal the difference
between the treated and control means due to rounding error.

17
 Einav et al. (2013) point out that additional information on borrower credit risk may
induce lenders to change not only interest rates but also the size of loans offered. To examine
the possibility that risk-based loan pricing is correlated with changes in loan size, the analysis
summarized in Table 5 is repeated with the average real value of loans as the outcome variable
(results are available from the authors upon request). The treatment effect estimates are
generally negative and are only significantly different from zero, using a 5-percent significance
threshold, for one of the six cohorts. The negative signs of these effects suggest that credit
unions may restrict average loan amounts under risk-based pricing. This is somewhat akin to the
finding in Berger et al. (2005) that the adoption of credit scoring increases the volume of loans
under $100,000 issued to small businesses, while having a negative and statistically insignificant
effect on the volume of larger loans, those between $100,000 and $250,000. It is possible that
credit unions may have offered smaller loans to high-risk applicants after adopting risk-based
pricing and that this may have reduced overall average loan sizes even as the total number of
loans increased. However, the results suggest that the effects of risk-based pricing on loan sizes
are, in general, statistically indistinguishable from zero.
The foregoing suggests that the number of loans as a proportion of credit union
membership increases in response to adoption of risk-based loan pricing, but it does not provide
any insight into how borrowers of different risk levels are affected. The post-adoption evolution
of loan delinquencies, shown in Table 6, may shed some light on that question. One line of
reasoning holds that technologies that allow financial institutions to identify high-risk borrowers
may also enable them to offer more loans to that group by charging higher interest rates,
commensurate with greater risk of default, rather than denying loan applications outright (Berger
et al., 2005). Risk-based lending is hypothesized to increase loan delinquency rates by
increasing lending to high-risk borrowers who might otherwise be excluded from credit markets
as well as by raising interest rates for that group. That hypothesis is not borne out by the
estimates in Table 6. With few exceptions, loan delinquency rates for adopters of risk-based
pricing are lower than those for matched non-adopters. For half of the cohorts, the adoption of
risk-based loan pricing is accompanied by significantly lower delinquency rates.
One explanation for this finding is that the higher interest rates charged to relatively risky
borrowers may suppress loan demand within that group, thus discouraging loan applications
from potential borrowers with a high propensity to default. Einav et al. (2013) find that
implementing risk-based pricing allowed an auto finance company to screen out high-risk
borrowers resulting in lower loan default rates. A similar mechanism may be at work in the
credit union sample examined here. It is also quite possible that lower interest rates for relatively
low-risk applicants may have increased the proportion of total loans going to that group.
Edelberg (2006) documents that risk-based pricing expands debt levels primarily for low-risk
borrowers. Factors such as attractive interest rates for those with low propensities to default and
reduced loan demand within the high-risk tiers may help explain why loan delinquency rates are
generally lower for credit unions that adopt risk-based pricing. It should be noted that, for most
years, the effect of this loan pricing strategy on delinquency rates is not significantly different
from zero. Nonetheless, the results are not generally consistent with the hypothesis that risk-
based lending increases access to credit primarily for high-risk borrowers.

18
TABLE 6. Average treatment effect on the treated: Loan delinquency rate
Cohort Year Treated Mean Control Mean ATET t-statistic Observations
t-1 1.3753 1.5280 -0.1527 (-0.79) 930
2007 t 1.4596 1.5352 -0.0756 (-0.45) 930
t+1 1.7643 1.8705 -0.1062 (-0.52) 930
t+2 1.9646 2.1680 -0.2033 (-0.85) 930
t+3 2.0647 2.0648 -0.0001 (<0.01) 930
t-1 1.6078 1.8210 -0.2132 (-0.94) 622
2008 t 1.7468 1.8553 -0.1085 (-0.54) 622
t+1 2.0183 2.2109 -0.1926 (-0.75) 622
t+2 1.9697 2.1892 -0.2196 (-0.77) 622
t+3 1.6654 2.2543 -0.5889 (-2.46)** 622
t-1 2.3484 2.4473 -0.0989 (-0.36) 906
2009 t 2.6886 2.4678 0.2208 (0.68) 906
t+1 2.3944 2.4135 -0.0192 (-0.08) 906
t+2 2.2828 2.5247 -0.2419 (-0.93) 906
t+3 2.0828 2.0262 0.0566 (0.27) 906
t-1 2.2814 3.0784 -0.7970 (-1.33) 390
2010 t 2.1301 2.5131 -0.3830 (-1.12) 390
t+1 1.7714 2.4979 -0.7265 (-2.14)** 390
t+2 1.7021 2.4691 -0.7670 (-1.77)* 390
t+3 1.6604 2.4369 -0.7764 (-1.68)* 390
t-1 2.1153 1.7303 0.3849 (0.93) 174
2011 t 2.2977 1.7844 0.5134 (1.20) 174
t+1 1.8839 2.1485 -0.2647 (-0.67) 174
t+2 1.8786 2.3558 -0.4772 (-0.91) 174
t+3 1.8361 2.0246 -0.1885 (-0.39) 174
t-1 1.8073 2.4914 -0.6841 (-1.58) 418
2012 t 1.8242 2.4314 -0.6072 (-1.35) 418
t+1 1.5123 2.1370 -0.6247 (-1.94)* 418
t+2 1.3494 2.2848 -0.9355 (-2.73)** 418
t+3 1.2341 2.0120 -0.7779 (-2.24)** 418
Notes: t-statistics are in parentheses; * surpasses 10% critical value of the t-distribution; **
surpasses 5% critical value of the t-distribution. ATET may not exactly equal the difference
between the treated and control means due to rounding error.

19
 To gauge the robustness of the results, the baseline estimates reported in Tables 5 and 6
are compared with estimated treatment effects obtained by applying a caliper and by using an
inverse probability weighting approach rather than nearest neighbor matching. First, calipers are
applied to ensure that only very close matches are used to calculate the treatment effects
(Appendix A). The treatment effects are re-estimated using calipers equal to 0.25 times the
standard deviation of the propensity score for each year (Rosenbaum and Rubin, 1985; Stuart,
2010). The calipers range from 0.01 to 0.04 depending on the year. This results in eliminating
from the sample only five out of the 1,720 risk-based lending adopters across all cohorts. The
signs of the estimated treatment effects are the same as those documented in Tables 5 and 6 in all
cases and the magnitudes are only very slightly different for four of the six cohorts and exactly
the same for the other two cohorts.
Inverse probability weighting (IPW) is an alternative approach to treatment effects
estimation based on propensity scores. Instead of constructing a control group using only a
subset of the non-adopters in the sample, this method uses a weighted mean of all non-adopters.
The weights are based on the estimated propensity scores and the difference in weighted means
computed for adopters and non-adopters is the estimated treatment effect (Wooldridge, 2010).
With this weighting procedure, the balancing property is satisfied and, in fact, none of the
standardized bias statistics exceed 5.7 percent in absolute value, which is well below the 10
percent threshold commonly used (Appendix B).
The IPW estimates of the loan-to-member ratio differentials are all positive as in Table 5,
but exhibit somewhat lower variability than the matching-based estimates. The treatment effect
estimates for the first four years after adoption range from 3.0 to 5.7 percentage points and the
estimated t-statistics in the majority of cases surpass the 1 percent critical value of the t-
distribution. Similar to the matching results, the weighted mean loan-to-member ratios for
adopters generally diverge from the weighted means for non-adopters over the course of the five
year observation period within each cohort. The IPW estimates of the loan delinquency rate
differentials are similar to those presented in Table 6, though again with somewhat less
variability. In all cases but one, the estimated treatment effects are negative, indicating that the
weighted mean delinquency rate for adopters is lower than the weighted mean for non-adopters.
In seven cases, the negative treatment effect estimates are statistically distinguishable from zero
at the 5-percent significance level. Overall, the results are broadly similar to those obtained
using the nearest-neighbor matching algorithm.
As an additional robustness check, panel data fixed effects regressions are carried out
using the full data sample (2006-2015) for both the loan delinquency rate and loan-to-member
ratio outcome variables. Three specifications are examined. The first is a simple regression of
the outcome variables on a risk-based pricing adoption indicator along with year and credit union
fixed effects. The second is an augmented specification including the loan interest rate and the
county-level unemployment rate as additional explanatory variables. Several studies find that
these variables are related to loan demand and loan default rates (Agarwal et al., 2008; Archer et
al., 2002; Calza et al., 2003). The third specification includes lagged values of the explanatory
variables, rather than contemporaneous values.

20
 The results largely corroborate the outcomes of the propensity score analysis. The
impacts of adoption on the loan-to-member ratio, estimated using fixed effects regressions, range
from 1.02 to 1.51 percentage points across the three specifications. These estimates are
somewhat smaller in magnitude than those reported in Table 5, but are always significantly
different from zero using a one percent significance criterion. The estimated impacts of adoption
on the loan delinquency rate range from -0.14 to -0.19 percentage points. The negative sign of
these parameter estimates generally aligns with the results reported in Table 6. Furthermore, the
effect of adoption on the loan delinquency rate after a one-year lag is statistically distinguishable
from zero using a 5 percent significance criterion. It should be noted, however, that these results,
which are based on a 10-year sample period, are not directly comparable with those reported in
Tables 5 and 6, which are based on five-year subsamples and include impacts observed up to
three years after the year of adoption.
Risk-based pricing affects loan availability and loan delinquency primarily by raising
interest rates for high-risk borrowers and lowering them for low-risk borrowers. However, it is
not clear whether this pricing strategy has any effect on average loan interest rates. Average
interest rates are treated as covariates in this analysis rather than outcome variables. Thus, the
main results do not directly address the effect of risk-based loan pricing on average loan rates
and only tentative conclusions can be drawn in this regard. To control for factors that are
correlated with both the credit union’s decision to adopt risk-based pricing and the average
interest rate it charges on loans, a fixed effects analysis similar to the one described above is
conducted, using per capita income and the credit union’s market share as explanatory variables
(Feinberg and Rahman, 2006; Tokle et al., 2015). The estimated coefficients on the adoption
variable are all negative, but statistically insignificant. The negative signs of the coefficients are
consistent with the argument that adoption of risk-based pricing expands credit access primarily
for lower-risk consumers who will be charged lower interest rates.

6. Conclusion

Expanding access to credit for members is generally a priority for credit unions. Indeed,
the NCUA has suggested that the appropriateness of a risk-based loan pricing strategy for credit
unions is largely determined by whether it increases the number of new, higher-risk borrowers.
To answer this question, data are assembled for two outcome variables, loans per credit union
member and loan delinquency rates, as well as a dichotomous risk-based pricing adoption
indicator and a number of covariates. The estimated probabilities of adopting risk-based pricing
are used to pair adopters with otherwise similar non-adopters and differences in the outcome
variables are then analyzed over six successive five-year windows between 2006 and 2015.
The results indicate that adopting risk-based loan pricing does increase the number of
loans supplied as a proportion of total credit union membership. This aligns with predictions
regarding lending behavior when asymmetric information problems are mitigated through
increased information on individual borrowers’ default probabilities. When interest rates are set
according to individual risk levels, loan availability appears to increase, though the average loan
amount does not necessarily increase and, in some cases, appears to decrease. The latter effect,
which may be due to restrictions on loan size for high-risk borrowers, seems to be outweighed by

21
the positive and generally significant changes in the number of loans issued. Because credit
unions are member-owned and many members are either actual or potential borrowers, an
increase in the number of loans issued is likely to be considered a desirable outcome of risk-
based pricing adoption in most cases.
While risk-based lending does appear to allow credit unions to issue more loans, the
question remains as to whether new loans are offered primarily to low- or high-risk borrowers.
The data on loan delinquency rates do not provide a conclusive answer to this question.
However, they do suggest that average risk levels decreased, at least for some cohorts, after the
adoption of risk-based pricing. The reductions in loan delinquency rates are consistent with the
findings in previous studies that risk-based pricing primarily increases lending to low-risk
borrowers, who receive better terms, whereas lending to high-risk consumers either increases
less or even contracts due to higher borrowing costs for that group (Edelberg, 2006; Einav et al.,
2013).
Overall, the results suggest that risk-based pricing does increase the number of loans
issued, but that the main beneficiaries are most likely lower-risk borrowers. The reduction in
loan delinquency rates suggests that risk-based lending may also help improve credit union
profitability by reducing charge-offs. Given that, the upward trend shown in Figure 1 is likely to
continue to be observed. However, it does not appear that the practice benefits underserved
high-risk borrowers to the extent that some had hoped. Early guidance from the NCUA (1995)
implies that risk-based loan pricing is likely to increase credit access for high-risk borrowers and
that, consequently, credit unions adopting this practice could expect some increases in loan
delinquency rates. Indeed the NCUA guidelines emphasize that, before adopting risk-based
pricing, credit unions would need to ensure that capital is sufficient to absorb ensuing loan
losses. The guidelines even recommend capping lending in the highest-risk tiers, at least in the
early stages of program implementation.
One implication of the findings is that credit unions adopting risk-based loan pricing are
likely to increase the number of loans offered and sustain little, if any, adverse impact in terms of
higher delinquency and charge-off rates. Also, because the net worth ratios for non-adopters are
much higher than the average for all credit unions, those institutions probably have the capacity
to absorb modest increases in charge-off rates, should they implement a risk-based pricing
strategy. If higher levels of loan delinquency do not materialize after adoption, as seems to be
the case for most cohorts in the sample, credit unions may consider raising or eliminating any
caps that limit lending to high-risk borrowers. Routine review of rejected loan applications by
loan officers is also recommended for credit unions adopting risk-based pricing (NCUA, 1995).
It may also be useful to compare the distribution of borrowers’ credit scores before and
after adoption of risk-based pricing to shed additional light on how the practice affects lending to
high- and low-risk borrowers. The latter is a useful avenue for future academic research and
may also be helpful for in-house evaluation purposes at individual credit unions. Finally, future
research regarding the impacts of risk-based pricing adoption on loan interest rates for low- and
high-risk borrowers may provide insight into the mechanisms by which this lending strategy
affects debt levels and loan delinquency rates for both groups.

22
References

Abadie, A., Imbens, G.W., 2006. Large sample properties of matching estimators for average
treatment effects. Econometrica. 74(1), 235-267.

Abadie, A., Imbens, G.W., 2011. Bias-corrected matching estimators for average treatment
effects. Journal of Business & Economic Statistics. 29(1), 1-11.

Abadie, A., Imbens, G.W., 2016. Matching on the estimated propensity score. Econometrica.
84(2), 781–807.

Adams, W., Einav, L., Levin, J., 2009. Liquidity constraints and imperfect information in
subprime lending. American Economic Review. 99(1), 49-84.

Agarwal, S., Ambrose, B.W., Chomsisengphet, S., 2008. Determinants of automobile loan
default and prepayment. Federal Reserve Bank of Chicago Economic Perspectives. 32(3),
17-28.

Akhavein, J., Frame, W.S., White, L.J., 2005. The diffusion of financial innovations: An
examination of the adoption of small business credit scoring by large banking
organizations. Journal of Business. 78(2), 577-596.

Archer, W.R., Elmer, P.J., Harrison, D.M., and Ling, D.C., 2002. Determinants of multifamily
mortgage default. Real Estate Economics. 30(3), 445-473.

Austin, P.C., 2009. Balance diagnostics for comparing the distribution of baseline covariates
between treatment groups in propensity-score matched samples. Statistics in Medicine.
28(25), 3083–3107.

Behr, A., Heid, F., 2011. The success of bank mergers revisited. An assessment based on a
matching strategy. Journal of Empirical Finance. 18(1), 117-135.

Berger, A.N., Frame, W.S., Miller, N.H., 2005. Credit scoring and the availability, price, and risk
of small business credit. Journal of Money, Credit and Banking. 37(2), 191-222.

Bofondi, M., Lotti, F., 2006. Innovation in the retail banking industry: The diffusion of credit
scoring. Review of Industrial Organization. 28(4), 343–358.

Brookhart, M.A., Schneeweiss, S., Rothman, K.J., Glynn, R.J., Avorn, J., and Stürmer, T., 2006.
Variable selection for propensity score models. American Journal of Epidemiology.
163(12), 1149–1156.

23
Caliendo, M., Kopeinig S., 2008. Some practical guidance for the implementation of propensity
score matching. Journal of Economic Surveys. 22(1), 31–72.

Calza, A., Gartner, C., and Sousa, J., 2003. Modelling the demand for loans to the private sector
in the euro area. Applied Economics. 35(1), 107-117.

Dehejia, R.H., Wahba, S., 2002. Propensity score-matching methods for nonexperimental causal
studies. Review of Economics and Statistics. 84(1), 151-161.

Deng, Y., Gabriel, S.A., 2006. Risk-based pricing and the enhancement of mortgage credit
availability among underserved and higher credit-risk populations. Journal of Money,
Credit, and Banking. 38(6), 1431-1460.

DeYoung, R., Lang, W.W., Nolle, D.L., 2007. How the Internet affects output and performance
at community banks. Journal of Banking & Finance. 31(4), 1033–1060.

Dow, J.P., Jr., 2007. The adoption of web banking at credit unions. Quarterly Review of
Economics and Finance. 47(3), 435–448.

Duca, J.V., Rosenthal, S.S., 1994. Do mortgage rates vary based on household default
characteristics? Evidence on rate sorting and credit rationing. Journal of Real Estate
Finance and Economics. 8(2), 99-113.

Edelberg, W., 2006. Risk-based pricing of interest rates for consumer loans. Journal of Monetary
Economics. 53(8), 2283–2298.

Einav, L., Jenkins, M., Levin, J., 2012. Contract pricing in consumer credit markets.
Econometrica. 80(4), 1387-1432.

Einav, L., Jenkins, M., Levin, J., 2013. The impact of credit scoring on consumer lending. Rand
Journal of Economics. 44(2), 249–274.

Fase, M.M.G., 1995. The demand for commercial bank loans and the lending rate. European
Economic Review. 39(1), 99-115.

Feinberg, R.M., 2001. The competitive role of credit unions in small local financial services
markets. Review of Economics and Statistics. 83(3), 560-563.

Feinberg, R.M., Rahman, A.F.M.A., 2006. Are credit unions just small banks? Determinants of
loan rates in local consumer lending markets. Eastern Economic Journal. 32(4), 647-659.

24
Garrido, M.M., Kelley, A.S., Paris, J., Roza, K., Meier, D.E., Morrison, R.S., Aldridge, M.D.,
2014. Methods for constructing and assessing propensity scores. Health Services
Research. 49(5), 1701-1720.

Getter, D.E., 2006. Consumer credit risk and pricing. Journal of Consumer Affairs. 40(1), 41-63.

Goenner, C.F., 2016. The policy impact of new rules for loan participation on credit union
returns. Journal of Banking & Finance. 73, 198-210.

Hannan, T.H., McDowell, J.M., 1984. The determinants of technology adoption: The case of the
banking firm. Rand Journal of Economics. 15(3), 328-335.

Heckman, J.J., Ichimura, H., Todd, P.E., 1997. Matching as an econometric evaluation estimator:
Evidence from evaluating a job training programme. Review of Economic Studies. 64(4),
605-654.

Hellmann, T., Lindsey, L., Puri, M., 2008. Building relationships early: Banks in venture capital.
Review of Financial Studies. 21(2), 513-541.

Hernandez-Murillo, R., Llobet, G., Fuentes, R., 2010. Strategic online banking adoption. Journal
of Banking & Finance. 34(7), 1650–1663.

Huber, M., Lechner, M., Wunsch, C., 2013. The performance of estimators based on the
propensity score. Journal of Econometrics. 175(1), 1-21.

Imbens, G.W., 2004. Nonparametric estimation of average treatment effects under exogeneity: A
review. Review of Economics and Statistics. 86(1), 4-29.

Li, X., Zhao, X., 2006. Propensity score matching and abnormal performance after seasoned
equity offerings. Journal of Empirical Finance. 13(3), 351–370.

Magri, S., Pico, R., 2011. The rise of risk-based pricing of mortgage interest rates in Italy.
Journal of Banking & Finance. 35(5), 1277-1290.

Magri, S., 2015. Are lenders using risk-based pricing in consumer loan market? The effect of the
crisis. http://swfa2015.uno.edu/E_Speculation_&_Risk_Premium/paper_74.pdf (accessed
06.03.17).

McKillop, D., Wilson, J.O.S., 2011. Credit unions: A theoretical and empirical overview.
Financial Markets, Institutions & Instruments. 20(3), 79–123.

NCUA, 1995. NCUA Letter to Credit Unions. Letter no. 174. National Credit Union
Administration, Alexandria, VA.

25
NCUA, 1999. NCUA Letter to Credit Unions. Letter no. 99-CU-05. National Credit Union
Administration, Alexandria, VA.

NCUA, 2013. Financial Performance Report (FPR) Ratios for Federally Insured Credit Unions.
National Credit Union Administration, Alexandria, VA.

NCUA, 2016. Credit Union and Corporate Call Report Data. National Credit Union
Administration, Alexandria, VA.

Pana, E., Vitzthum, S., Willis, D., 2015. The impact of internet-based services on credit unions:
A propensity score matching approach. Review of Quantitative Finance and Accounting.
44(2), 329–352.

Riley, J.G., 1987. Credit rationing: A further remark. American Economic Review. 77(1), 224-
227.

Rosenbaum, P.R., 1995. Observational Studies. New York, NY: Springer-Verlag.

Rosenbaum, P.R., 2010. Design of Observational Studies. New York, NY: Springer.

Rosenbaum, P.R., Rubin, D.B., 1983. The central role of the propensity score in observational
studies for causal effects. Biometrika. 70(1), 41-55.

Rosenbaum, P.R., Rubin, D.B., 1985. Constructing a control group using multivariate matched
sampling methods that incorporate the propensity score. The American Statistician. 39(1),
33-38.

Stiglitz, J.E., Weiss, A., 1981. Credit rationing in markets with imperfect information. American
Economic Review. 71(3), 393-410.

Stuart, S.A., 2010. Matching methods for causal inference: A review and a look forward.
Statistical Science. 25(1), 1-21.

Tokle, R., Fullerton, T.M., Jr., Walke, A.G., 2015. Credit union loan rate determinants following
the 2008 financial crisis. Social Science Journal. 52(3), 364-373.

Wooldridge, J.M., 2010. Econometric Analysis of Cross-Section and Panel Data, 2nd ed. MIT
Press, Cambridge, MA.

26
Appendix A: Caliper Matching Results

Table A.1 Average treatment effect on the treated: Loan-to-member ratio (propensity score
matching with caliper)
T-1 T T+1 T+2 T+3 Obs.*

2007 1.7525 2.8793 3.2990 4.6815 4.6469 930/930

(1.27) (2.18) (2.44) (2.90) (3.17)

2008 1.3700 2.4093 3.0052 3.7354 3.5131 620/622

(0.91) (1.20) (1.34) (1.95) (2.10)

2009 1.7235 4.0938 3.9661 2.9917 2.7487 902/906

(1.33) (1.71) (2.46) (2.20) (2.02)

2010 4.9882 5.5969 4.4825 5.0447 6.0148 388/390

(2.59) (2.71) (2.35) (2.52) (3.00)

2011 3.4334 4.5776 3.9701 5.0946 5.5594 174/174

(1.48) (1.98) (1.64) (2.02) (2.15)

2012 2.6108 3.5051 4.4386 4.3916 5.1967 416/418

(1.26) (1.74) (2.07) (1.82) (2.04)
Notes: t-statistics are in parentheses. The caliper is equal to 0.25 × standard deviation of the
propensity score (Rosenbaum and Rubin, 1985; Stuart, 2010).
*The observations column shows the number of observations included in the calculation after the
caliper is applied (on the left) out of the total observations possible for one-to-one nearest
neighbor matching (on the right). If the two numbers are equal, that means no observations were
pruned.

27
TABLE A.2 Average treatment effect on the treated: Loan delinquency rate (propensity score
matching with caliper)
T-1 T T+1 T+2 T+3 Obs.*

2007 -0.1527 -0.0756 -0.1062 -0.2033 -5E-05 930/930

(-0.93) (-0.52) (-0.49) (-0.96) (-0.00)

2008 -0.1913 -0.1149 -0.1149 -0.2102 -0.5901 620/622

(-0.68) (-0.46) (-0.43) (-0.72) (-2.81)

2009 -0.1014 0.21579 -0.0613 -0.242 0.05044 902/906

(-0.35) (0.64) (-0.22) (-0.85) (0.20)

2010 -0.6758 -0.3376 -0.6804 -0.6909 -0.7658 388/390

(-1.09) (-0.91) (-1.85) (-1.91) (-1.95)

2011 0.38494 0.51336 -0.2647 -0.4772 -0.1885 174/174

(0.48) (0.55) (-0.37) (-0.95) (-0.37)

2012 -0.6853 -0.6083 -0.6306 -0.9349 -0.779 416/418

(-1.81) (-1.59) (-2.21) (-3.10) (-2.51)
Notes: t-statistics are in parentheses. The caliper is equal to 0.25 × standard deviation of the
propensity score (Rosenbaum and Rubin, 1985; Stuart, 2010).
*The observations column shows the number of observations included in the calculation after the
caliper is applied (on the left) out of the total observations possible for one-to-one nearest
neighbor matching (on the right). If the two numbers are equal, that means no observations were
pruned.

28
Appendix B: Inverse Probability Weighting Results

Table B.1 Standardized bias statistics (inverse probability weighting method)

2007 2008 2009 2010 2011 2012
ln (Total assets) -5.3% 0.7% 5.7% -0.9% 0.0% -1.1%
Net worth ratio -0.3% -2.4% -3.5% -1.6% -1.5% -0.6%
Fee income ratio -5.4% -1.9% -3.8% 0.5% -2.1% -1.9%
Regular shares ratio 0.6% -0.1% -4.4% 1.8% -1.1% 0.3%
Transactional website -0.5% 2.9% 4.5% 2.0% -0.2% 0.2%
Market share (local) -1.3% -2.6% -1.7% -1.5% -0.5% -2.9%
Market share (state) 1.5% 0.2% -0.9% -1.2% -1.4% 1.0%
Loan interest rate 3.7% -1.4% -4.8% 0.1% -1.0% -0.8%
Per capita income 1.5% 2.3% -1.6% 2.8% -2.9% 1.2%

TABLE B.2 Average treatment effect on the treated: Loan-to-member ratio (inverse probability
weighting method)
T-1 T T+1 T+2 T+3 Obs.

2007 2.3293 2.9996 3.0314 4.2240 4.6400 3,264

(2.52) (3.15) (2.95) (3.33) (3.88)

2008 3.2573 3.8449 5.3415 5.5950 5.5399 2,956

(3.03) (2.97) (3.21) (3.91) (4.14)

2009 1.8692 4.1791 3.7014 2.9512 3.1426 2,850

(1.78) (1.85) (2.65) (2.75) (2.81)

2010 3.5078 4.9167 3.5024 4.2345 5.1733 2,449

(2.53) (3.13) (2.58) (2.96) (3.37)

2011 2.4150 3.2721 3.1815 4.2476 3.4914 2,105

(1.45) (1.84) (1.76) (2.20) (1.41)

2012 3.6259 4.4558 5.4367 5.7097 5.7097 2,049

(2.58) (3.31) (3.51) (2.90) (2.93)
Note: t-statistics are in parentheses.

29
TABLE B.3 Average treatment effect on the treated: Loan delinquency rate (inverse probability
weighting method)
T-1 T T+1 T+2 T+3 Obs.

2007 -0.0535 -0.1027 -0.0711 -0.2136 -0.0822 3,264

(-0.62) (-1.08) (-0.63) (-1.37) (-0.55)

2008 -0.1903 -0.2345 -0.4660 -0.3615 -0.6529 2,956

(-1.47) (-1.81) (-1.66) (-1.92) (-4.23)

2009 -0.0356 0.0427 -0.3997 -0.2480 -0.3906 2,850

(-0.19) (0.16) (-1.25) (-1.30) (-1.74)

2010 -0.3442 -0.1834 -0.5082 -0.4446 -0.4033 2,449

(-1.33) (-0.95) (-3.11) (-2.44) (-2.20)

2011 -0.2876 -0.0655 -0.3551 -0.2420 -0.2209 2,105

(-0.98) (-0.14) (-0.95) (-0.89) (-0.57)

2012 -0.4038 -0.2075 -0.3856 -0.4817 -0.6356 2,049

(-2.14) (-0.90) (-2.50) (-3.32) (-3.94)
Note: t-statistics are in parentheses.

30