427639

The Diffusion of Financial Innovations: An Examination of the Adoption of Small Business Credit Scoring by Large Banking Organizations Author(s):
Jalal Akhavein, W. Scott Frame, and Lawrence J. White Source: The Journal of Business, Vol. 78, No. 2 (March 2005), pp. 577-596 Published by: The University of Chicago Press Stable URL: http://www.jstor.org/stable/10.1086/427639 . Accessed: 15/11/2013 07:02
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp
.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact support@jstor.org.
The University of Chicago Press is collaborating with JSTOR to digitize, preserve and extend access to The Journal of Business.
http://www.jstor.org
This content downloaded from 197.251.143.70 on Fri, 15 Nov 2013 07:02:22 AM All use subject to JSTOR Terms and Conditions
Jalal Akhavein
Moodys Risk Management Services
W. Scott Frame
Research Department, Federal Reserve Bank of Atlanta
Lawrence J. White
Stern School of Business, New York University
The Diffusion of Financial Innovations: An Examination of the Adoption of Small Business Credit Scoring by Large Banking Organizations*
One of the bedrocks of our nancial system is nancial innovation, the life blood of efcient and responsive capital markets. Van Horne (1985)
I. Introduction
Financial innovation has been described as the life blood of efcient and responsive capital markets. Yet, few quantitative investigations have studied nancial innovations and the diffusion of these new technologies. In this paper, we examine the diffusion of one such technology: credit scoring models for small business lending. Using data for large banking organizations, our hazard model indicates that banking rms with more branches innovate earlier, as do those located in the New York Federal Reserve district. Our Tobit model conrms these results and nds that organizations with fewer separately chartered banks but more branches innovate earlier.
It is widely recognized that innovation plays a crucial role in improving productivity. Equally important, however, is the rate at which innovations diffuse through an economy. Faster diffusion of innovations means a more immediate impact and thus a higher social return on the initial investment. There is general recognition of the particular importance of nancial innovation (e.g., Van Horne 1985; Miller 1986, 1992; Finnerty 1992; Merton 1992). However, until recently, the number of quantitative studies of nancial innovation
* We thank Robert DeYoung; William Greene; Felix Treptow; the seminar participants at the 2000 Atlantic Economic Association meetings, the 2001 Federal Reserve Bank of Chicago Conference on Bank Structure and Competition, the 2002 Federal Reserve Bank of Philadelphia Conference on Innovation in Financial Services and Payments; and an anonymous referee for helpful comments. The views expressed in this paper are those of the authors and not those of the Federal Reserve Bank of Atlanta, the Federal Reserve System, or Moodys. (Journal of Business, 2005, vol. 78, no. 2) B 2005 by The University of Chicago. All rights reserved. 0021-9398/2005/7802-0007$10.00 577
#04463 UCP: BN article # 780207 This content downloaded from 197.251.143.70 on Fri, 15 Nov 2013 07:02:22 A
All use subject to JSTOR Terms and Conditions
578
Journal of Business
has been modest.1 A recent survey by Frame and White (2004) documents this relative dearth of empirical studies of nancial innovation. Further, we have been able to nd only seven quantitative studies that examine the diffusion process for nancial innovations (Hannan and McDowell 1984, 1987; Sinha and Chandrashekaran 1992; Saloner and Shepherd 1995; Molyneux and Shamroukh 1996; Ingham and Thompson 1993; Gourlay and Pentecost 2002), the rst four of which use the same data on the introduction of automated teller machines (ATMs).2 This paper contributes to the empirical literature on the diffusion of nancial innovations. Specically, using a unique survey data set, we examine the introduction of credit scoring techniques for small business loan applications during the middle 1990s. This survey yielded information for 99 large banking organizations, including whether they had adopted credit scoring for small business lending by June 1997 and (if they had adopted) when they rst did so. We then relate this information to a number of market, rm, and managerial factors to uncover the important characteristics driving technological adoption decisions by large banking organizations. This is done by estimating a parametric hazard function with time-varying covariates.3 The benet of using this approach is that we can uncover the institutional factors that inuence both the probability and the rate of small business credit scoring adoption. Allowing these factors to evolve through time (via time-varying covariates) further enriches the model specication. We also employ Tobit analysis as an alternative means of capturing the relationship between the date of a banks adoption of credit scoring and the potential explanatory variables for its decision. Our paper proceeds as follows. Section II provides background information about small business credit scoring. Section III outlines our variables of interest. Section IV discusses the data. Section V presents our statistical tests and results. Conclusions are provided in section VI.
II. Credit Scoring
Credit scoring is the process of assigning a single quantitative measure, or score, to a potential borrower representing an estimate of the borrowers future loan performance ( Feldman 1997). While credit scores have been used for some time in the underwriting of consumer loans
1. This is in comparison with the much larger and much older set of quantitative studies that encompass innovation in manufacturing and agriculture. See Cohen and Levin (1989) and Cohen (1995) for broad reviews. It is striking that no articles concerning nancial services are cited in either survey. 2. Again, this is in contrast with a larger, older, and more varied set of diffusion studies that exist for the nonnancial sector. 3. The estimation of such hazard functions falls under the rubric of duration or survivor analysis. See Kalbeisch and Prentice (1980), Cox and Oakes (1984), and Kiefer (1988) for an introduction to duration analysis.
Diffusion of Financial Innovations
579
and residential mortgages, this technology has been routinely applied to commercial credit only since the mid-1990s. This is because commercial loans were thought to be too heterogeneous and documentation was not sufciently standardized either within or across institutions (Rutherford 1994/1995). However, credit analysts ultimately determined that the personal credit history of small business owners is highly predictive of the loan repayment prospects of their businesses, especially for loans under $100,000.4 Thus, personal information is obtained from a credit bureau and then augmented with basic business-specic data to predict repayment. Eisenbeis (1996) presents an overview of the history and application of credit scoring techniques to small business loan portfolios. Asch (1995) describes how the leading model of credit scoring for small business lendingthe model originated by Fair, Isaac and Companywas developed. The introduction of credit scoring may have profound effects on small business credit markets, including borrower-lender interactions, loan pricing, loan risk, and credit availability. First, credit scoring allows lenders to underwrite and monitor small business loans without actually meeting the borrower. This development contrasts with the perceived importance of a local bank-borrower relationship for small business lending (e.g., Petersen and Rajan 1994; Berger and Udell 1995). In fact, because of scoring systems, borrowers can obtain unsecured credit from distant lenders through direct marketing channels. Indeed, empirical evidence suggests that the distance between small rms and their lenders has increased in recent years (e.g., Petersen and Rajan 2002) and that credit scoring increases the probability that a large banking organization will make small business loans in certain areas (Frame, Padhi, and Woosley 2004). Second, small business credit scoring may affect the price of credit, although the direction of any change may depend on how a lender uses the technology. While credit scoring is a less-expensive way of underwriting loans (providing downward price pressures), it is not always used as a substitute technology. Indeed, if a lender generates credit scores to complement its usual underwriting approach, total costs may actually increase (providing upward price pressures). Furthermore, depending on the lenders perception of the quality of the information set, borrower risk may change. Empirical evidence by Berger, Frame, and Miller (2005) suggests that the net effect of small business credit scoring is higher average loan prices, especially for those lenders using the technology as a complement to traditional underwriting. The authors also nd that the loans funded by scoring institutions are classied as higher risk, on average, consistent with increased lending to riskier marginal borrowers.
4. Mester (1997) cites the use of information such as the applicants monthly income, outstanding debt, nancial assets, employment tenure, home ownership, and previous loan defaults or delinquencies.
580
Journal of Business
Third, credit scoring should increase credit availability for small businesses. Cheaper or better information about the repayment prospects of a small business applicant makes it more likely that a lender will price the loan based on expected risk, rather than denying the loan out of fear of charging too little.5 Empirical studies have found a positive relationship between small business credit scoring and credit availability (Frame, Srinivasan, and Woosley 2001; Frame, Padhi, and Woosley 2004; Berger, Frame, and Miller 2005). This paper examines a related but new and different issue: what factors inuence the probability and timing (i.e., the diffusion) of the adoption of this new technology by its rst movers, large banking organizations. To that end, the next section discusses a number of variables we believe may affect the conditional probability of a bank adopting the use of a small business creditscoring model.
III. Variables of Interest
We believe that a number of market, rm, and managerial factors may have inuenced the diffusion of credit-scoring technology for small business loans. Most important, the rate of diffusion should be related to the expected benets (protability) of adopting the new technology. Indeed, various theoretical models of this exist (Davies 1979; Reinganum 1981a, 1981b; Fudenberg and Tirole 1985; Karshenas and Stoneman 1993). However, in the spirit of agency theory, we also believe that variations in the quality (e.g., effectiveness, openness to innovation) of rm managers could also inuence adoption decisions. We discuss each of the covariates included in our analysis next.
A. Market Variables
The relative competitiveness of the marketplace may inuence technological adoption decisions. However, two competing theories exist. On the one hand, Schumpeter (1950) contends that rms with market power are most likely to advance industrial technology because of their superior access to capital and their ability to capture (internalize) the likely externalities that accompany innovation. On the other hand, Scherer (1984) suggests that rms in less-concentrated markets are relatively more likely to innovate than rms in more-concentrated markets, because of standard competitive motives. Generally, as an empirical matter, Cohen and Levin (1989) report that, if a relationship between
5. Moreover, the widespread use of credit scoring should increase future prospects for asset securitization by encouraging consistent underwriting standards that can reassure investors (and rating agencies) as to the quality of the underlying loans. To date, however, the securitization of small business loans has largely been limited to credits guaranteed by the U.S. Small Business Administration.
581
market concentration and innovation exists, it is fragile and likely to be economically unimportant. Nevertheless, in the area of nancial services innovation, Hannan and McDowell (1984) provide empirical evidence, based on the adoption of automated teller machines by banks, that is consistent with Schumpeters theory. Accordingly, to account for the possible relationship between market concentration and nancial innovation, we include an average Herndahl-Hirschman index (HHI) across all the local geographic markets in which the subsidiary banks operate weighted by their total deposits in each market. Considerable research suggests that technological adoption varies geographically ( Hagerstrand 1967; Thwaites 1982; and Rees, Briggs, and Oakey 1984). Baptista (2000) suggests that these regional differences in diffusion rates result from the geographical clustering of innovators. Similarly, we posit that the geographic distribution of individuals qualied to develop and maintain sophisticated econometric credit scoring models may be concentrated, specically in New York and San Francisco. This conjecture is based on the tremendous demand by Wall Street rms for so-called quants and the fact that the largest vendor of credit scoring models (Fair, Isaac and Co.) is based in the San Francisco area. To account for this, we include dummy variables indicating whether the banking organization is located in the New York ( NEWYORK) or San Francisco (SANFRAN) Federal Reserve districts.6
B. Firm Variables
Schumpeter (1950) predicts that larger rms (irrespective of market power) are likely to be more innovative, because of economies of scale in innovation and because of their ability to pool risks. Studies have found that larger rms are indeed likely to adopt new technologies before smaller rms (Manseld 1968a, 1968b; Manseld et al. 1977; Hannan and McDowell 1984; Rose and Joskow 1990). To account for the effects of size, we include the natural logarithm of total domestic banking assets ( LNASSETS).7 We believe that banking organizations that are more focused on small business lending are more likely to adopt credit scoring. Because credit scoring serves objectively to measure and rank the riskiness of
6. These location variables may also capture a competitive effect, whereby the adoption of small business credit scoring by one bank (or several banks) in an area may spur other banks in the same area to adopt also. 7. It has been suggested to us that an additional factor that might yield a positive relationship between bank size and early adoption of credit scoring would be the fact that larger banks were the original contributors of the data that underlay the creation of the Fair, Isaac model and they would therefore be more likely to adopt it. However, Aschs (1995, p. 13) description leaves this point open to question: 17 leading small business credit grantors contributed data to form the initial data pool. These . . . institutions represented different asset sizes.
582
Journal of Business
borrowers, nancial institutions are better able to manage credit risk and price loans accordingly. Therefore, for those institutions especially dependent on earnings from small business lending, credit scoring should be more valuable. As a result, we include the ratio of small loans (under $100,000) to businesses to total domestic banking assets (SBLRATIO).8 The underlying structure of the banking organization may also inuence the technology adoption decision. A more centralized organizationwith fewer separately chartered banks ( but possibly with more branches, as largely deposit-gathering instruments)may be more inclined to centralize its lending decisions and thus may be more receptive to credit scoring as a way to improve its from afar decisions.9 Conversely, a more decentralized organization, with more separately chartered banks (and fewer branches), may be more inclined to leave lending decisions in the hands of local loan ofcers, who make their judgments on less-formal criteria and would be less interested in adopting credit scoring.10 For these reasons, we include both the number of subsidiary banks (BANKS) and the number of branches (BRANCHES) in our model.11 Although both BANKS and especially BRANCHES are correlated with bank size ( LNASSETS), they play a different role in our model, as indicators of the underlying organizational structure (centralized or decentralized) of the banking organization, especially after we control for size through the inclusion of LNASSETS. The protability of an institution may inuence its propensity to adopt new technologies. Hannan and McDowell (1984) suggest that this might stem from a liquidity constraint: Less protable rms are less able to invest in new technologies.12 For this reason, we include return on equity (ROE), or net income as a percent of total equity capital, as a measure of protability. Identifying organizational willingness to implement new programs or technologies is difcult. While this may come from chief executives
8. There may be a simultaneous relationship here, whereby the adoption of credit scoring spurs more small business lending. We address this issue later. Also, as will be discussed, we tried the alternative of including just the level of small business lending, rather than the ratio. 9. A larger number of branches, relative to the number of banks, may suggest a more retail orientation, including a focus on small business lending, but nevertheless a centralized decision-making structure for loan approvals. 10. Empirical support for both predictions can be found in Cole, Goldberg, and White (2004). 11. Also, as will be discussed, we tried instead the ratio of BRANCHES/ BANKS. 12. This was one of Schumpeters (1950) arguments in favor of monopoly (with its larger prot ow) as a superior market structure for encouraging innovation. In the language of more modern economics, the asymmetric information problems surrounding a rms attracting outside nance for an investment are even greater when the investment involves uncertain new technologies that may have little liquidation value. Therefore, a rm with a greater capability for internal nance (from greater prots) has an advantage with respect to investing in new technologies.
583
(see later), it may also be related to a rms corporate culture. For example, Furst, Lang, and Nolle (2002) nd that the probability of federally chartered commercial banks to adopt internet banking is positively related to their reliance on noninterest income. That is, fee income may act as a proxy for the aggressiveness of the banks business strategy. Accordingly, we include the ratio of noninterest income to net operating revenue (NIINCOME).13
C. CEO Variables
Last, the personal characteristics of each banking organizations chief executive ofcer (CEO) may inuence the adoption decision. This focus on management has been explored in the technology diffusion literature, primarily by Manseld (1968a, 1968b; Manseld et al. 1977). The two CEO variables we selected are tenure and education.14 First, the length of time that an individual has held the CEO position (CEOTEN) may inuence decision making in opposing ways. On the one hand, the longer a CEO has been in control of a rm, the more he or she knows about its resources and capabilities and how to mobilize them. On the other hand, new managers may be more likely to shake up an organization and try new approaches, especially in formerly protected industries like banking. Second, we posit that executives with additional education beyond their college degrees are more inclined to adopt new technologies. The dummy variable ADVDEG accounts for this by taking a value of one for CEOs with graduate degrees.
IV. Data
The data used in this study come from three sources. First, we have data recording the month and year that a sample of large banking organizations rst used small business credit scoring. This information was obtained from a telephone survey of the 200 largest U.S. banking organizations conducted by the Federal Reserve Bank of Atlanta in January 1998. Ninety-nine institutions responded to the survey in which (among other things) respondents were asked whether their institution credit scored small business loans and, if so, when it was rst implemented. See Frame, Srinivasan, and Woosley (2001) for a complete discussion of the survey questions and results.
13. At the suggestion of a referee, we explored the possibility that an additional rm-level explanatory variable, the salaries of loan ofcers, might be important, since banks in areas where salaries are high might be more inclined to substitute credit scoring for the efforts of loan ofcers. We collected state-level data for 1997 from the Bureau of Labor Statistics on wages for loan ofcers and counselors and tried this variable in our regression models described later. Doing so added no explanatory power. 14. In preliminary efforts, we also included CEO age in our analysis. Because it was generally found to be unimportant (and correlated with CEO tenure), we drop it here.
584
Journal of Business
Second, we use several items reported on the commercial bank Call Reports (and aggregated to the holding company level). This information was obtained from the Wharton Financial Institutions Center (which, in turn, obtained the data from the Federal Reserve Board). Since information about banks small business lending activities is collected only on the midyear Call Report (and only since 1993), this is how we organize our analysis. As a result, we focus on annual, midyear observations for each large banking organization between 1993 and 1997. Next, we collected data on the personal characteristics of the chief executive ofcers of each large banking organization that responded to our survey. Specically, information on CEO tenure and education was obtained primarily from the Forbes magazine "500" lists.15 Because CEO information was unavailable for four institutions (all foreign owned), our analysis ultimately focuses on a total of 95 large banking organizations.16 Finally, because only about half of the large banking organizations to which the survey was sent responded, the potential for nonresponse bias is present. To correct for this, we employed Heckmans (1979) two-step procedure. We rst estimated a probit regression, where a dummy variable indicating a response to the telephone survey is the dependent variable and LNASSETS and SBLRATIO are the independent variables.17 From this regression, the inverse Mills ratio was computed and became an additional explanatory variable, LAMBDA, in our statistical models.18 Table 1 provides the descriptive statistics for our sample, as of June 1997. Of the 95 large banking organizations in our sample, 55 (58%) had adopted credit scoring for small business lending by June 1997.19 Of those that did adopt, nine had adopted before 1995, while the remainder adopted during the last few years of our sample period. On average, the large banking organizations in our sample had 1.7% of their portfolios devoted to small business lending. These institutions were quite protable (an average return on equity of almost 16%) and earned about 30% of their revenues from fees.20 These banking
15. This biographical information can be found on the Forbes website (www.forbes.com). For information on CEOs who were not in those lists, we turned to various Whos Who directories and news stories in the American Banker. 16. A sample size of 95 is smaller than the sample that was used for the ATM diffusion studies but larger than the sample sizes used in many nonnancial diffusion studies (e.g., Manseld 1968a; Manseld et al. 1977; Oster 1982). Also, these 95 organizations accounted for about half of all banking assets (as of June 1997). 17. Only the size variable is signicant in these regressions. 18. In the hazard model (discussed later), LAMBDA is computed annually over the 199397 period; for the Tobit model, LAMBDA is based on 1997 data. 19. This percentage is somewhat lower than, but not greatly different from, the 70% (38 out of 54 banks) that indicated (in a Federal Reserve Board survey) that they used credit scoring for small business lending in January 1997 (Federal Reserve Board 1997). 20. NIINCOME is measured over the 4 quarters ending in June.
585
TABLE 1
Sample Statistics for 95 Large U.S. Banking Organizations (data as of June 30, 1997 )
Average Std. Deviation Minimum Maximum
SBLRATIO LNASSETS BANKS BRANCHES HHI SANFRAN NEWYORK ADVDEG CEOTEN ROE NIINCOME
.0167 9.3859 6.9053 367.3789 .1880 .1263 .0737 .4737 9.1789 .1588 .2984
.0100 1.2660 8.6234 476.5632 .0602 .3340 .2626 .5020 7.8753 .3291 .0997
.0015 7.2364 1.0000 13.0000 .0709 .0000 .0000 .0000 .0000 .0728 .0951
.0510 12.3351 38.0000 2617.0000 0.5005 1.0000 1.0000 1.0000 40.0000 .2526 .6405
Note.Number of banking organizations that had adopted credit scoring for small business lending by June 1997: 55. Number that adopted in 1992: 1. Number that adopted in 1993: 2. Number that adopted in 1994: 6. Number that adopted in 1995: 9. Number that adopted in 1996: 28. Number that adopted in 1997: 9.
organizations averaged seven subsidiary banks and 367 bank branches. Almost 13% of the institutions were located in the San Francisco Federal Reserve district, while over 7% were in the New York district. On average, chief executives in the sample organizations had been in their positions for over 9 years, and about half (47%) had advanced degrees. Table 2 provides the simple correlation matrix among the variables listed in table 1. We have also included ADOPT, which is a dummy variable indicating whether the banking organization had adopted the innovation by June 1997, and YEARS, which is an integer equal to the number of years that the institution had used credit scoring as of June 1997. For example, YEARS is equal to one if the banking organization had used credit scoring for 112 months as of June 1997. As can be seen, ADOPT is positively and statistically signicantly related to YEARS, LNASSETS, and BRANCHES, while negatively related to SBLRATIO. YEARS is positively and statistically signicantly associated with LNASSETS, BRANCHES, NEWYORK, and ADVDEG. All these last four results are consistent with the expectations discussed previously. Among the independent variables, the following correlation coefcients are worth mentioning. First, SBLRATIO has signicant negative associations with LNASSETS, BRANCHES, and NEWYORK and signicant positive associations with HHI and CEOTEN. Therefore, small business lending tends to be relatively less important at larger banks, banks with more branches, and banks located in the New York Federal Reserve district. Also, small business lending is more important in more concentrated markets (which tend to be smaller,
586
#04463 UCP: BN article # 780207

* ***
TABLE 2
Simple Correlation Coefcients

ADOPT YEARS SBLRATIO LNASSETS BANKS BRANCHES HHI SANFRAN NEWYORK ADVDEG CEOTN ROE NIINC
ADOPT YEARS SBLRATIO LNASSETS BANKS BRANCHES HHI SANFRAN NEWYORK ADVDEG CEOTN ROE NIINC
*
1.00 *** .75*** ** .22** *** .46*** .05 *** .38*** .13 .07 .16 .17 .13 .7 .07
1.00 .16 *** .51*** .07 *** .50*** .10 .01 * .19* ** .25** .06 .12 .13
1.00 *** .46*** .04 ** .25** *** .40*** .03 ** .21** .08 ** .22** .12 * .19*
1.00 *** .27*** *** .80*** ** .21** .09 .12 *** .35*** .16 .02 * .18*
1.00 *** .34*** .06 .06 .14 .10 .13 .09 * .18*
1.00 .09 .14 .03 ** .22** .07 .13 *** .28***
1.00 .10 .03 .02 .10 .09 .04
1.00 .11 .04 .03 .14 .08
1.00 .03 .03 ** .21** .13
1.00 .07 * .18* ** .25**
1.00 .13 .12
1.00 .14
1.00
Journal of Business
* Statistically signicant at the 10% level. * Statistically signicant at the 5% level. *** Statistically signicant at the 1% level.
587
less urban, and populated with smaller banks) and where the CEO has been with the bank longer (which tends to be associated with smaller banks). Second, LNASSETS has signicant positive associations with BANKS, BRANCHES, ADVDEG, and NIINCOME and a signicant negative association with HHI. In other words, larger banking organizations tend to have more constituent banks and branches (as expected), a CEO with an advanced degree, and higher relative noninterest income (as expected), while they tend to operate in less concentrated (larger, urban) markets. We now turn to the detailed tests of our hypotheses.
V. Tests and Results
To examine the determinants of the likelihood and timing of small business credit scoring adoption we estimate two different statistical models, a hazard and a tobit. The hazard model is of primary interest because it utilizes information over the entire 199397 time frame. The tobit model, which relies on cross-sectional data from 1997 for the explanatory variables, provides a supplement to the results of the hazard model.21 The hazard model allows us to determine simultaneously the institutional factors that inuence both the probability and rate of small business credit scoring adoption. This approach views the adoption of small business credit scoring as the hazard and therefore includes only observations during which these models are not being used by the institution. The dependent variable in this model is the natural logarithm of the number of years until adoption occurs. A detailed description of hazard, or duration, analysis is provided in the next section. The dependent variable in the tobit model is the number of years that the banking organization had been using credit scoring as of June 1997 (YEARS). Thus, large banking organizations that had not adopted credit scoring for their small business loan portfolios are censored and coded with zero. The estimated coefcients in the tobit model tell us what variables inuence the timing of credit scoring adoption. Denoting Y as the relevant dependent variable (and suppressing the institution and time subscripts), the following relationship can be denoted for both statistical models:
Y f SBLRATIO; LNASSETS; BANKS; BRANCHES; HHI; SANFRAN; NEWYORK; ADVDEG; CEOTEN; ROE; NIINCOME; LAMBDA: 1
21. We used the 1997 data because it represents the realizations of how our sample institutions evolved over the 199397 period. Using the 1993 data, by contrast, may not be as good a predictor of future activity because future business decisions are unobserved or otherwise unaccounted for.
588
Journal of Business
Both models are estimated by maximum likelihood using LIMDEP, Version 7.0. A. Hazard Model
To examine both the likelihood and rate of small business credit scoring adoption, we estimate a hazard model. Using standard duration analysis, we estimate a hazard rate at which rms adopt small business credit scoring, conditional on time and on the market, rm, and CEO characteristics outlined in section IV. To this end, we begin by providing an overview of duration analysis, then turn to the results yielded by this procedure.
1. Overview
The variable of interest in the analysis of duration is the length of time, T, that elapses from the beginning of some event until its end or measurement is taken (Greene 1997). In our case, we are interested in the length of time that elapses before large banking organizations adopt credit scoring for underwriting small business loans. We then estimate a hazard rate, or the probability of an adoption in each year given that the institution has not introduced credit scoring by that time. Following Kiefer (1988), we describe a cumulative probability distribution as
F T ProbT t; 2
or the probability that the nonnegative random variable T is less than some value t . By denition, this cumulative probability can also be described in terms of a probability density function: f t dF t =dt F 0 t : Alternatively, one can specify a survival function, S t 1 F T ProbT t: 4 3
Given these denitions, eqq. (2)(4), we can then present a hazard function or the probability that an institution will adopt small business credit scoring at time T, given that it has not done so by that time: lt it =S t : 5
Furthermore, by substituting eqq. (3) and (4) into (5), it can also be shown that lt d ln S t =dt: 6
589
The hazard function may also be related to explanatory variables:

lt f Xt b; 7
where X is a matrix of variables inuencing the hazard rate and b is a vector of regression coefcients. The variables in X may also be updated through time, reecting what the duration literature refers to as timevarying covariates. Assuming a particular underlying probability distribution for l implies a likelihood function that can be maximized to generate parameter estimates for b and any other variables dictated by the distribution of l. However, as noted by Greene (1997), censoring is a pervasive and usually unavoidable problem in the analysis of duration data. As a result, the general form of the log-likelihood function is Lq
n X i 1
di ln f t ; q
n X 1 di ln S t ; q i1
where the indicator di equals one for an uncensored observation and zero otherwise. For our sample of 95 large U.S. banking organizations, we observe credit scoring adoption decisions annually over the 199397 period. Those institutions that had not adopted credit scoring by midyear 1997 are right censored and therefore have no failure times. We estimate this model, including time-varying covariates, assuming a Weibull distribution.22 The Weibull distribution is selected because it is consistent with the overall-increasing rate of credit scoring adoption observed during the sample period. 2. Results
Table 3 presents the estimation results for the parametric hazard function assuming a Weibull process for the hazard rates.23 We nd that BRANCHES and NEWYORK are statistically related to early adoption of small business credit scoring. Specically, the model indicates that both variables are negatively related to the probability of not adopting small business credit scoring through the sample period. Equivalently, banking organizations with more branches and those headquartered
22. Both Kiefer (1988) and Greene (1997) present the survivor, probability distribution, and hazard functions for the Weibull and other distributions. 23. Note that we must be careful in interpreting the parameters: A negative coefcient means a negative effect on duration and therefore a positive effect on adoption speed. Also, three banks had already adopted credit scoring for small business lending by June 1993, which is the beginning of our hazard model analysis; consequently, we use only 92 banks for our hazard analysis.
590
Journal of Business
TABLE 3
Hazard Model Estimates of the Length of Time to Adoption of Small Business Credit Scoring; Dependent Variable: ln(TIME)
Coefcient t-Statistic
Constant SBLRATIO LNASSETS BANKS BRANCHES HHI SANFRAN NEWYORK ADVDEG CEOTEN ROE NIINCOME LAMBDA l exp bX P 1=s Median
1.7859 5.0514 * .2871* .0007 ** .0005** .0006 .0237 ** .4147** .0579 .0070 .2808 .6698 *** 1.5091*** .1895 3.1907 4.7052
1.146 .711 1.817 .119 2.012 .001 .120 2.208 .519 .834 .217 1.058 3.330
Log-likelihood = 110.1814; chi-square = 51.7922 * * Statistically signicant at the 10% level. * * Statistically signicant at the 5% level. *** *** Statistically signicant at the 1% level.
in the New York Federal Reserve district tended to adopt the new technology sooner. As noted, these ndings are consistent with the presence of organizational inuences on technology adoption (also, BANKS has the expected sign but is statistically insignicant) and the geographic clustering of adopters. Bank size (LNASSETS) shows a surprising positive coefcient that is signicant at the 10% level; this sign (implying a negative relation between early adoption and size) is the opposite of what we hypothesized. However, the coefcient on the inverse Mills ratio ( LAMBDA) is also highly signicant (at the 1% level), and this variable is inversely related to bank size. Thus, the LAMBDA variable is likely capturing the effects of organization size and indicates that larger organizations tended to adopt earlier.24 There is little discernable relationship between the rate of credit scoring adoption and any of our other market, rm, or CEO variables. However, as measured by a log-likelihood ratio (chi-square) test, the model easily rejects the null hypothesis of no overall relationship at a 1% level. Next, we estimate a somewhat similar tobit model to supplement our results.
24. When the hazard equation is estimated in the absence of the LAMBDA variable, LNASSETS has the expected negative sign. This reinforces our conjecture that size does have the effect of encouraging earlier adoption but that the effect appears through LAMBDA in the Heckman-corrected results.
591
B. Tobit Model
Table 4 displays the results for the tobit analysis, in which YEARS is the dependent variable. Consistent with the hazard model results, we nd positive and statistically signicant coefcients on BRANCHES and NEWYORK. Further, the coefcient on BANKS is negative and signicant, consistent with our expectations concerning the effects of organizational structure.25 Again, bank size (LNASSETS) has the wrong sign (but is insignicant); also, again, the coefcient on LAMBDA is highly signicant and likely to be capturing the effects of organizational size.26 One surprise in the tobit model is that the coefcient on ROE is negative and signicant. Apparently, the organizations that were less protable during the 1990s were more inclined to adopt credit scoring earlier, which is counter to the Schumpeterian expectations.27 Perhaps being lean and hungry may be more important for early technological adoption than the nancial capability that accompanies protability.28 The remaining variables are insignicant; but, as was true for the hazard model, a log-likelihood (chi-square) test easily allows us to reject the null hypothesis of no overall relationship at the 1% level.
C. A Summing Up
The results of the hazard and tobit models are quite consistent: Organizational structure (which is proxied by the number of separately chartered banks and the number of branches within the organization) is important. More-centralized organizations (with fewer separately chartered banks and more branches) were more likely to adopt small business credit scoring earlier.29 Further, organizations headquartered in the New York Federal Reserve district were also more likely to be early adopters of this technology. Also, although organization size does not directly appear to be positively related to the adoption of small business credit scoring, we believe that this relationship is being indirectly captured in our models via the Heckman correction. Finally,
25. This result is consistent with results reported by Frame, Srinivasan, and Woosley (2001) for the probability of a banks having adopted small business credit scoring. 26. Again, when the tobit model is estimated without the Heckman correction (and therefore without LAMBDA), the coefcient on LNASSETS is positive and signicant at the 5% level. 27. Furst, Lang, and Nolle (2002) nd a similar result for rst movers into Internet banking. One possible explanation is that the technology may aid in improving performance going forward. 28. In the hazard model, the coefcient on ROE had a consistent ( positive) sign but was statistically insignicant. 29. As a robustness check, we tried the ratio of BRANCHES/BANKS as a replacement for the two separate variables in both models. The signs on the ratio were consistent with the organizational structure hypothesis, and the other inferences of the models were unchanged.
592
Journal of Business
TABLE 4
Tobit Estimates of Variables Inuencing the Rate of Adopting Small Business Credit Scoring; Dependent Variable: YEARS (June 30, 1997 data)
Coefcient t-Statistic
Constant SBLRATIO LNASSETS BANKS BRANCHES HHI SANFRAN NEWYORK ADVDEG CEOTEN ROE NIINCOME LAMBDA
9.8294* 13.1996 .5713 *** .0611*** *** .0018*** 2.1801 0.4745 ** 1.5762** .3288 .0011 ** 12.4343** 1.1707*** 3.9713***
1.727 .545 1.065 2.610 2.610 .590 .866 2.212 .860 .044 1.999 .553 2.790
Log-likelihood = 124.6548; chi-square = 54.2992 * * Statistically signicant at the 10% level. * * Statistically signicant at the 5% level. *** *** Statistically signicant at the 1% level.
the tobit model especially indicates that less-protable organizations were more likely to adopt credit scoring earlier. The remaining variables in our model failed to show statistical signicance in either regression framework. Perhaps the most surprising is the absence of signicance for SBLRATIO, since organizations for whom small business lending was relatively more important might be expected to adopt credit scoring earlier.30 However, unlike smaller banks, for whom small business lending is a larger fraction of their lending activity (e.g., Berger, Goldberg, and White 2001), small business lending constituted only 0.155.10% of any organizations assets, with a sample average of 1.67%.31 Therefore, the tendency for early
30. As was noted, there may well be a simultaneous relationship between small business lending and credit scoring, since the adoption of credit scoring may spur more small business lending. To deal with the possible bias in our results that might arise from this simultaneity, we created an instrument, called SBLHAT, that (following Frame, Srinivasan, and Woosley 2001) was the tted values from a linear regression for each of the years 199397 of SBLRATIO on a constant, a dummy variable indicating whether credit scoring was being used that year, the equity-to-assets ratio, the ratio of commercial loan charge-offs to commercial loans, the ratio of total loans to total assets, LNASSETS, BANKS, and BRANCHES. The inclusion of SBLHAT in our hazard and tobit models (instead of SBLRATIO) yielded regression results that are quite similar to our results in tables 3 and 4: The signs on the coefcients on SBLHAT were the same as those on the coefcients on SBLRATIO (and still insignicant), and the other inferences from the models were unchanged. Also, if SBLRATIO is simply excluded from the models, no changes are found in the resulting inferences. Therefore, simultaneity bias does not appear to be affecting our results. 31. In addition to SBLRATIO, we also tried as a variable just the (log of ) the level of small business lending by the organization (LNSBL). But the level of small business lending and the organizations overall size (as measured by assets) are (unsurprisingly) highly correlated (r 0:83). Accordingly, when LNASSETS was dropped from these regressions,
593
adoption of the technology could well have been driven more by other organizational and environmental factors than by the fraction of small business lending. The absence of any signicant effects from the HHI variable is another nding that is inconsistent with the Schumpeterian hypothesis. Although there might be confounding effects of market concentration (HHI), net income ( ROE), and noninterest income ( NIINCOME), the correlation coefcients between them are negative and small (0.09 and 0.04, respectively), as shown in table 2. Further, when ROE and NIINCOME are excluded from the hazard and tobit models, there is no appreciable change in the coefcients on HHI. The absence of signicance on the coefcient for SANFRAN indicates that geographic proximity to the source of the new technology (Fair, Isaac and Co.) was not important for early adoption. Apparently, inexpensive telecommunications were a good substitute for geographic proximity. Finally, the various CEO variables had no signicant explanatory power in our statistical tests, despite the signicant positive simple correlation (r 0:25) between ADVDEG and early adoption shown in table 2. The signicant positive correlation between ADVDEG and LNASSETS (r 0:35) may indicate that the effects of CEO education are being mufed by the effects of organization size.
VI. Conclusions
In this paper, we test a diffusion model of technology adoption for the nancial services industry. Specically, we examine the characteristics of the early adopters of small business credit scoring using two statistical tests. Using both hazard and tobit analysis, we nd that banking organizations that are more centralized in their organizational structure (i.e., that have more branches and fewer separately chartered banks) and those located in the New York Federal Reserve district adopted small business credit scoring before their peers. The former nding is consistent with the inuence of organizational structure on technology adoption, and the latter with theories of geography-dependent innovation diffusion. Although organization size did not directly inuence diffusion in the way expected, it appears to have done so indirectly, through the parameter that corrects for possible response bias, since the larger banking organizations were those most likely to have responded. The consistent nding that rm organizational structure is significantly related to the probability of early adoption has immediate
the coefcient on LNSBL had the same sign and signicance in the hazard model as for the coefcient on LNASSETS in our reported results, while the signs on the other important variables did not change. In the tobit model, the coefcient on LNSBL is positive but insignicant, while our inferences for the rest of the tobit model are similarly unchanged.
594
Journal of Business
implications with respect to the ongoing consolidation of the banking industry: The increased ease with which banks can branch across state lines and replace separately chartered banks with branches, brought about by the Riegel-Neal Interstate Banking and Branching Efciency Act of 1994, is likely to make them more receptive to adopting innovations earlier.32 Further work can surely be done to extend these analyses. For example, it would certainly be useful to know the extent to which small business credit scoring pervades the industry today. It would also be interesting to see if our ndings are consistent for other nancial innovations, particularly on the credit side of the banking business. In addition, studies of diffusion that encompass smaller banks would be worthwhile. Since nancial innovation is the lifeblood of the nancial sector, further efforts to understand the diffusion of this vital ingredient through the sector are surely warranted.
ReferencesREFERENCES
Asch, L. 1995. How the RMA / Fair, Isaac credit-scoring model was built. Journal of Commercial Lending 77 (June): 1016. Baptista, R. 2000. Do innovations diffuse faster within geographical clusters? International Journal of Industrial Organization 18 (April): 51535. Berger, A. N., W. S. Frame, and N. H. Miller. 2005. Credit scoring and the availability, price and risk of small business credit. Journal of Money, Credit, and Banking 37 (forthcoming). Berger, A. N., L. G. Goldberg, and L. J. White. 2001. The effects of dynamic changes in bank competition on the supply of small business credit. European Finance Review 5 (January): 11539. Berger, A. N. and G. F. Udell. 1995. Relationship lending and lines of credit in small rm nance. Journal of Business 68 (July): 35181. Cohen, W. 1995. Empirical studies of innovative activity. In Handbook of the economics of innovation and technological change, ed. P. Stoneman. Cambridge, MA: Blackwell. Cohen, W., and R. Levin. 1989. Empirical studies of innovation and market structure. In Handbook of industrial organization , vol. 2, ed. R. Schmalensee and R. Willig. Amsterdam: North-Holland. Cole, R. A., L. G. Goldberg, and L. J. White. 2004. Cookie cutter versus character: The micro structure of small business lending by large and small banks. Journal of Financial and Quantitative Analysis 39 (June): 227251. Cox, D., and D. Oakes. 1985. Analysis of survival data. New York: Chapman & Hall. Davies, S. 1979. The diffusion of process innovations. Cambridge: Cambridge University Press. Eisenbeis, R. 1996. Recent developments in the application of credit scoring techniques to the evaluation of commercial loans. IMA Journal of Mathematics Applied in Business and Industry 7: 271-90. Federal Reserve Board. 1997. Senior loan ofcer opinion survey on bank lending practices. (January). Feldman, R. 1997. Banks and a big change in technology called credit scoring. Federal Reserve Bank of Minneapolis The Region (September): 19-25.
32. Also, given the indirect results that show the favorable effects of size, even if our results can be applied only to the very large banks of our sample, the continuing rise in the average size of these banks bodes well for earlier adoptions of innovations.
595
Finnerty, J. 1992. An overview of corporate securities innovation. Journal of Applied Corporate Finance 4 (Winter): 2339. Frame, W. S., M. Padhi, and L. Woosley. 2004. The effect of credit scoring on small business lending in low- and moderate-income areas. Financial Review 39 (February): 3554. Frame, W. S., A. Srinivasan, and L. Woosley. 2001. The effect of credit scoring on small business lending. Journal of Money, Credit, and Banking 33 (August): 81325. Frame, W. S. and L. J. White. 2004. Empirical studies of nancial innovation: Lots of talk, little action? Journal of Economic Literature 42 (March): 116114. Fudenberg, D., and J. Tirole. 1985. Pre-emption and rent equalisation in the adoption of new technology. Review of Economic Studies 52 (July): 383401. Furst, K., W. W. Lang, and D. N. Nolle. 2002. Internet banking. Journal of Financial Services Research 22 (August): 95117. Gourlay, A. R. and E. J. Pentecost. 2002. The determinants of technology diffusion: Evidence from the UK nancial sector. Manchester School 70 (March): 185203. Greene, W. 1997. Econometric analysis. New York: Macmillan. Hagerstrand, T. 1967. Innovation diffusion as a spatial process. Chicago: University of Chicago Press. Hannan, T., and J. McDowell. 1984. The determinants of technology adoption: The case of the banking rm. Rand Journal of Economics 15 (Autumn): 32835. . 1987. Rival precedence and the dynamics of technology adoption: an empirical analysis. Economica 54 (May): 155171. Heckman, J. 1979. Sample selection bias as a specification error. Econometrica 47 (January): 15361. Ingham, H. and S. Thompson. 1993. The adoption of new technology in nancial services: The case of building societies. Economics of Innovation and New Technology 2: 26374. Kalbeisch, J., and R. Prentice. 1980. The statistical analysis of failure time data. New York: Wiley. Karshenas, M., and P. Stoneman. 1993. Rank, stock, order, and epidemic effects in the diffusion of new process technologies: An empirical model. RAND Journal of Economics 24 (Winter): 50328. Kiefer N., 1988. Economic duration data and hazard functions. Journal of Economic Literature 26 (June): 64679. Manseld, E. 1968a. The economics of technological change. New York: Norton. . 1968b. Industrial research and technological innovation: Econometric analysis. New York: Norton. Manseld, E., J. Rapoport, A. Romeo, E. Villani, S. Wagner, and F. Husic. 1977. The production and application of new industrial technology. New York: Norton. Merton, R. C. 1992. Financial innovation and economic performance. Journal of Applied Corporate Finance 4 (Winter): 1222. Mester, L. 1997. Whats the point of credit scoring? Federal Reserve Bank of Philadelphia Business Review (SeptemberOctober): 316. Miller, M. H. 1986. Financial innovation: The last twenty years and the next. Journal of Financial and Quantitative Analysis 21 (December): 45971. . 1992. Financial innovation: achievements and prospects. Journal of Applied Corporate Finance 4 (Winter): 412. Molyneux, P., and N. Shamroukh. 1996. Diffusion of nancial innovations: the case of junk bonds and note issuance facilities. Journal of Money, Credit, and Banking 28 (August): 50222. Oster, S. 1982. The diffusion of innovation among steel rms: The basic oxygen furnace. Bell Journal of Economics 13 (Spring): 4556. Petersen, M., and R. Rajan. 1994. The benets of lending relationships: Evidence from small business data. Journal of Finance 49 (March): 337. . 2002. Does distance still matter? The information revolution in small business lending. Journal of Finance 57 (December): 253370. Rees, J., J. Briggs, and R. Oakey. 1984. The adoption of new technology in the American machinery industry. Regional Studies 18 (December): 489504. Reinganum, J. 1981a. Market structure and the diffusion of new technology. Bell Journal of Economics 12 (Autumn): 61824.
596
Journal of Business
. 1981b. On the diffusion of new technology: A game theoretic approach. Review of Economic Studies 48 (July): 395405. Rose, N., and P. Joskow. 1990. The diffusion of new technologies: evidence from the electric utility industry. RAND Journal of Economics 21 (Autumn): 35473. Rutherford, R. 1994/1995. Securitizing small business loans: A bankers action plan. Commercial Lending Review 10 (Winter): 6274. Saloner, G., and A. Shepard. 1995. Adoption of technologies with network effects: An empirical examination of the adoption of automated teller machines. RAND Journal of Economics 26 (Autumn): 479501. Scherer, F. M. 1984. Innovation and growth: Schumpeterian perspectives. Cambridge, MA: MIT Press. Schumpeter, J. A. 1950. Capitalism, socialism, and democracy, 3d ed. New York: Harper & Brothers. Sinha, R. K. and M. Chandrashekaran. 1992. A split hazard model for analyzing the diffusion of innovations. Journal of Marketing Research 29 (February): 116127. Thwaites, A. 1982. Some evidence of regional variations in the diffusion of new industrial products and processes within the British manufacturing industry. Regional Studies 16 (October): 37181. Van Horne, J. 1985. Of nancial innovations and excesses. Journal of Finance 40 (July): 62136.

427639

Hochgeladen von

Dokumentinformationen

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

427639

Hochgeladen von

Copyright:

Verfügbare Formate

The Diffusion of Financial Innovations: An Examination of the Adoption of Small Business Credit Scoring by Large Banking Organizations Author(s):

Diffusion of Financial Innovations

Diffusion of Financial Innovations

Diffusion of Financial Innovations

Diffusion of Financial Innovations

#04463 UCP: BN article # 780207

Simple Correlation Coefcients

1.00 .09 .14 .03 .22 .07 .13 * .28*

1.00 .10 .03 .02 .10 .09 .04

1.00 .11 .04 .03 .14 .08

1.00 .03 .03 .21 .13

1.00 .07 * .18* .25

1.00 .13 .12

Diffusion of Financial Innovations

Diffusion of Financial Innovations

The hazard function may also be related to explanatory variables:

Diffusion of Financial Innovations

Diffusion of Financial Innovations

Diffusion of Financial Innovations

Das könnte Ihnen auch gefallen