A Bayesian approach for analyzing the services of banking institutions.

INTRODUCTION

Banking institutions today must create a perception of uniqueness in the mind of the customer to gain an advantage in the marketplace. The evolving customer base brought about by geographic, demographic, and life-style variances must be monitored to maintain and enhance market position.

The interrelationships of variables defining the antecedents and also the consequences of customer satisfaction have been studied extensively in the consumer research literature (e.g., Anderson and Sullivan 1993; Bearden and Teel 1983; Bolton and Drew 1991a, 1991b; Cardozo 1965; Churchill and Suprenant 1982; Cronin and Taylor 1992; La Barbera and Mazursky 1983; 6 and Peat 1979; Oliva, Oliver, and MacMillan 1992; Oliver 1977, 1980; Oliver and Bearden 1985; Oliver and DeSarbo 1988; Oliver and Swan 1989; Tse and Wilton 1988; Westbrook 1981; Yi 1990). However, there appears to be conflicting evidence as to the nature of the linkages between the antecedents and consequences of satisfaction. For example, Cronin and Taylor (1992) report structural modeling results which indicate perceived service quality is an antecedent to satisfaction. However, Bolton and Drew (1991a, 1991b) give conceptual representations where the satisfaction construct is an antecedent to a service quality construct. Yi's (1990) review reports some of the conflicting findings that exist in different studies.

One of the most important directions for future research is to shed more light on which paradigm may best model customer satisfaction judgments in various applications (Erevelles and Leavitt 1992). The use of methods that incorporate structural information in an adaptive manner, using or capitalizing on it to the extent that observed data are consistent with that information, may be especially useful in modeling customers' preferences. Prior structural information plays an important role in this context, even when relevant prior information is vague in form. In relation to structural equation systems in the literature which assume a priori knowledge concerning the linkages between constructs, adaptive methods are less likely to suggest knowledge where it may not exist, especially in situations where the investigation is genuinely exploratory. This is consistent with the recommendation by many researchers to use methods that account directly for the conditions that characterize a specific problem (cf. Laughlin 1986).

The purpose of this study is to present a Bayesian Structural Regression (BSR) paradigm for modeling customers' perceptions of banking services and identifying the important determinants of preference for a particular banking institution. Unlike previous research in the literature, this study uses adaptive structural methods to model customer preference data. These methods are based on conjugate Bayesian theory discussed by Dempster (1969) and made operational by Chen (1979) using the EM method (Dempster, Laird, and Rubin 1977). The Bayesian approach provides a mechanism for incorporating prior structural information into covariance estimation. This information can be either vague or specific and is used only to the extent that it reflects worthwhile information about as many interrelationships among variables as possible. Exploratory common factor models are used here to represent vague prior knowledge about covariance structure. It will be shown below that common factor models are particularly useful in applications where there are substantial measurement errors in the variables.

The current situation in the banking business is one of increased complexity in terms of customer, distribution, and product (Kimball and Gregor 1995). The industry can be characterized by changes in customers' preferences and rising competition among banks and between banks and nonbank service providers. Customer satisfaction with the services of banks is a moving target; banks must continually monitor the rapidly changing marketplace and strive to understand and become more responsive to their customers' needs and preferences (Chakravarty, Feinberg, and Widdows 1995; Glassman 1995). The role of adaptive methodology for modeling customers' preferences may be especially important for banks as they adopt a customer focus as the basis of operations in a target market.

The remainder of this article will proceed as follows. An overview of the BSR procedures is given. Then, sampling characteristics for an empirical application in a target market are discussed. Results of using the adaptive Bayesian methodology on data obtained from a customer preference survey are then presented. The final section summarizes the findings, discusses limitations, and provides a number of implications for research and practice.

OVERVIEW OF BAYESIAN STRUCTURAL REGRESSION

Chen (1979) developed a class of methods for stochastic multiple regression where the criterion and predictor variables are jointly random. The BSR approach uses adaptive smoothing procedures and maximum likelihood estimation to produce stable representations of the predictor-criterion covariance structure. For more information see the article by Pruzek and Lepak (1992) which discusses techniques in covariance and regression estimation that were motivated by Chen's work. However, Pruzek and Lepak developed adaptive smoothing and estimation techniques using frequentist principles where estimation is noniterative and generally does not involve maximum likelihood estimation.

In particular, adapting the conjugate Bayesian procedure for joint covariance and mean estimation (see Dempster 1969), the BSR methodology assumes that a system of n observation vectors (each composed of one criterion value, and j = p - 1 predictor values) represents a random sample of n p-dimensional values from a multivariate normal distribution with mean [Mu] and positive definite covariance matrix [Sigma]. To simplify discussion, the first observation in each vector is assumed to represent the criterion measurement so that the remaining observations correspond to measurements on random predictor variables. The derivation of the adaptive Bayesian methods is based on the properties of the Wishart distribution. For the conjugate form it is assumed that the inverse of [Sigma], [[Sigma].sup.-1], has a Wishart prior distribution. Specifically, [[Sigma].sup.-1] [similar to] W([(v[Omega]).sup.-1], v), with degrees of freedom v. It follows that the posterior distribution of [Sigma], given the observation vectors, has the inverse Wishart form [Sigma] [similar to] [W.sup.-1]([[Sigma].sup.*][(n + v).sup.-1], n + v). Without prior information for [Mu], the mean of the posterior distribution of [Sigma] takes the form [Mathematical Expression Omitted], a weighted average of the given prior [Omega] and the data-based matrix [Mathematical Expression Omitted], where [Mathematical Expression Omitted] is the usual maximum likelihood estimate (MLE) of [Sigma].

Chen's (1979) approach was to assume a given structural form for [Omega], the mean of the prior distribution of the population covariance matrix [Sigma], and to estimate the posterior mode (or mean, as a result of symmetry), given the prior structural model. In theory, the prior structure for [Omega] can take on any form; however, Chen focused on structure of general factor analytic form. Chen shows that the MLE [Mathematical Expression Omitted] of ([Omega], v) can be obtained by an iterative EM procedure (Dempster, Laird, and Rubin 1977) based on the marginal distribution of [Mathematical Expression Omitted]. Chen's main result is a Bayesian estimate of [Sigma], defined as the mode of the posterior density of [Sigma], of the form

[Mathematical Expression Omitted]. (1)

In equation (1), [Mathematical Expression Omitted] is the conventional MLE of [Sigma]; [Mathematical Expression Omitted] denotes a maximum likelihood common factor estimate of the original [Omega] and is based on the same data used to generate [Mathematical Expression Omitted]; and [Mathematical Expression Omitted] is a derived posterior estimate which indicates the degree to which structural information in the sample is in agreement with the covariance structural model assumed for [Omega]. If the prior common factor structural model for [Omega] is consistent with the data, [Mathematical Expression Omitted] will be large and more weight will be given to the structural covariance estimate [Mathematical Expression Omitted]. Otherwise, relatively more weight will be shifted to the conventional MLE [Mathematical Expression Omitted].

For any set of j predictor variables and a criterion, BSR equations can be derived from the covariance estimate [Mathematical Expression Omitted] in equation (1). Specifically, if y designates the criterion and x the predictors, the (1 [where] j) symmetrically partitioned estimate

[Mathematical Expression Omitted] (2)

can be used to compute the j x 1 vector of BSR coefficients

[Mathematical Expression Omitted] (3)

where [Mathematical Expression Omitted] represents the j x j covariance matrix for the predictor variables, and [Mathematical Expression Omitted] is the vector of j predictor-criterion covariances. This representation assumes that all variables in the system have been converted to deviation score form. Additionally, Chen (1979, 241242) uses Bayesian arguments to show that [Mathematical Expression Omitted] has a multivariate t-distribution and gives expressions to compute estimated standard errors for the BSR coefficients.

Several important points for motivating the use of this adaptive Bayesian methodology deserve mention. Ordinary least squares (OLS) has been widely used in consumer and marketing research to estimate parameters of linear models. If the criterion and predictor variables in a particular application are random variables measured with error (for example, variables measured by customer perception), then OLS regression coefficients are biased towed zero or attenuated by the measurement error; the squared multiple correlation also will be reduced; and there will be less power in testing whether OLS coefficients are significantly different from zero (Fuller 1987). Also, measurement error can lead to difficulty in interpreting OLS regression coefficients because of suppressor effects. For example, if a predictor variable having a positive or zero correlation with a criterion variable is associated with a negative recession coefficient, the predictor variable is a suppressor variable (Darlington 1990). One would certainly expect predictors which have positive correlations with a criterion to also have regression coefficients with positive signs. As a result of measurement error, OLS regression coefficients associated with suppressor variables are indeed very difficult to interpret in practice.

To facilitate discussion, equation (1) is rewritten as

[Mathematical Expression Omitted] (4)

where [Mathematical Expression Omitted]. If w in equation (4) is set arbitrarily at unity, then [Mathematical Expression Omitted] and [Mathematical Expression Omitted]. Thus, the Bayesian approach includes OLS estimation as a special case. Suppose that for a particular application, [Mathematical Expression Omitted] is estimated to represent how well the prior structural model is supported by the data. If the prior structural model is unsatisfactory, w [approaches] 1 and the BSR estimates converge to OLS estimates. However, if the prior structural model is strongly supported by the data, then w [approaches] 0 and the BSR estimates are derived primarily from [Mathematical Expression Omitted], i.e., the BSR estimates will depend strongly on the parameter estimates of the prior structural model. Clearly, if w in equation (4) is set to zero, then all BSR results can be generated from the parameters associated with the prior structural model. Of course, complete reliance on a particular structural model is unnecessary when adaptive procedures are available.

Thus, the adaptive BSR approach to estimation provides a means to circumvent problems with the use of OLS estimation, especially when prior structural models are chosen to accommodate measurement errors in the variables. Also, the use of [Mathematical Expression Omitted] in equation (1) provides a safeguard against the uncertainty associated with the prior structural model selected for [Mathematical Expression Omitted]. Clearly, the aim is to estimate w adaptively from the data, where [Mathematical Expression Omitted] uses the prior structural model (only) to the extent that observed data support that model. The index w can be viewed as a badness of fit index on a scale from zero to unity, indicating how poorly the prior structural model is supported by the observed data. Accordingly, the complement (1 - w) represents a goodness of fit index for the prior structural model.

Common factor analysis models have found useful application in virtually all applied sciences (see Lawley and Maxwell (1971) for a discussion of common factor theory). These models are extremely useful when it is impossible to obtain wholly reliable measures of constructs; for example, when eliciting customers' perceptions from survey instruments. As shown later, exploratory factor models represent a class of structural models which can facilitate covariance estimation in many situations, especially when there are substantial measurement errors in the variables. When common factor models are used to construct a covariance estimate of [Sigma], the prior structural model will generally take the form

[Omega] = [FF.sup.T] + [U.sup.2] (5)

where F is the p x m matrix of common factor coefficients for p = j + 1; [U.sup.2] represents the diagonal matrix of uniqueness variances; m is the number of common factors; and T denotes the transpose of a matrix. If m is much smaller than p for a population, and the parameters are identifiable, a common factor model may provide a highly parsimonious representation of an observed covariance matrix.

An alternate common factor representation for prior structure can be obtained by assuming the independent unique variables have the same variance. Using this assumption, equation (5) is rewritten as

[Omega] = [FF.sup.T] + [Sigma]I (6)

where I is the p x p identity matrix and a is the common uniqueness variance, [Sigma] [greater than] 0. This common factor form assumes that the smallest p - m eigenvalues of [Omega] are equal (Chen 1979, Case(II) 244). A useful motivation for this parsimonious structural form is that when the population common factor model with m factors is true, and the uniqueness diagonal is known, then the smallest p - m eigenvalues of interest will equal one another (Lawley and Maxwell 1971). Also, as discussed below, the use of the prior structure in equation (6) allows the Bayesian approach to include ridge regression as a special case in its general framework.

Chen uses the prior structure in equation (6) and applies the EM algorithm(1) to obtain the maximum likelihood estimates [Mathematical Expression Omitted] for use in equation (1). For this case, the adaptive Bayesian covariance estimate in equation (1) is shown to alter the conventional MLE [Mathematical Expression Omitted] by shrinking its smallest p - m eigenvalues toward a common quantity, but leaving its first largest m eigenvalues as well as its entire matrix of eigenvectors unchanged. Of course, these results depend on a suitable selection of m, the number of common factors. Furthermore, this Bayesian estimate of covariance structure is invariant under any orthogonal transformations.

The maximum likelihood common factor estimate of [Omega] in equation (6) takes the form [Mathematical Expression Omitted], for [Mathematical Expression Omitted] of order p x m with m orthogonal factors. It is noteworthy that given this form for [Mathematical Expression Omitted], the BSR coefficient vector in equation (3) can be expressed as

[Mathematical Expression Omitted] (7)

(Chen 1979, 243-244) where the matrix [Mathematical Expression Omitted] is partitioned to correspond to the criterion and predictor variables as

[Mathematical Expression Omitted] (8)

Ridge regression (Hoerl and Kennard 1970) has been extensively used in the marketing literature as a means of deriving stable regression weights from an observed sample covariance matrix. The BSR methods also include ridge regression as a special case. Equation (7) makes the ridge regression form explicit. When the number of common factors m = 0, each term with [Mathematical Expression Omitted] drops out, and [Mathematical Expression Omitted] is a scalar multiplier for the diagonal matrix [Mathematical Expression Omitted]. The ridge form is equivalent to assuming that all the eigenvalues of the prior mean of [Sigma] are equal. In this case, the Bayes covariance estimate in equation (1) alters the conventional MLE [Mathematical Expression Omitted] by shrinking all its eigenvalues toward a common value.

Simulations reported by Chen (1979) compare the BSR procedures (including the ridge form) and OLS regression. The simulations are based on a population predictor-criterion covariance (correlation) system with 12 variables (one variable was arbitrarily selected as the criterion variable). The correlations between predictors ranged from 0.12 to 0.77. The ordered eigenvalues of the population covariance matrix are 5.02, 1.63, 0.86, 0.77, 0.68, 0.59, 0.53, 0.51, 0.43, 0.42, 0.34, 0.22. Sample covariance matrices were simulated from this population and BSR(m) estimates were obtained for fixed m = 0,1,2,3, under the assumption that the smallest p - m eigenvalues of the prior mean of the population covariance matrix are equal - the same assumption used in this study. Note that the BSR(0) estimates are ridge regression estimates. Results indicated that the BSR(m) estimates for all values of m show substantial improvement over OLS estimates in terms of reliability for a sample of size n = 40, and moderate improvement for a sample of size n = 120. These sample sizes are not uncommon in market response studies. More importantly, this evidence suggests that the BSR(m) methods perform well when the number of variables under consideration is large relative to the number of observations, especially when the sample sizes are limited. The BSR estimates based on common factor models with one and two factors also performed slightly better than the BSR estimates based on three factors and the BSR estimates based on ridge regression. A prior structural model with one or two common factors may be appropriate here since the two largest eigenvalues of the population covariance matrix appear to be relatively larger than the remaining ten eigenvalues. Generally speaking, these results suggest that the analysis does not depend critically on the choice for the number of common factors in one's prior structural model. Chen reports that similar results have been obtained in other unpublished works. Additionally, he suggests that the BSR methods may not only be robust to violations of distributional assumptions, but also more resistant to outliers than OLS procedures.

SAMPLE FOR EMPIRICAL APPLICATION

This research effort is actually an extension of a study performed for one financial institution (the focal institution). With the focal bank's permission, the BSR techniques are used to examine the perceptions of financial services held by commercial customers in the focal institution's immediate target market.

In the initial stages of the study, management at the focal bank organized several customer focus group meetings in order to identify the factors used by commercial customers in choosing a bank. Customer information regarding the service offerings of banks also was obtained from meetings with panels of consumer opinion leaders in the target market. The information obtained from the customer focus groups and advisory panels was used to construct a survey instrument which reflected customers' needs, interests, and concerns. This survey form was then used to assess how customers perceived competitive banks in the target market.

A total of 750 questionnaires were mailed to the commercial customers of the focal bank and two other banking institutions. These three financial institutions comprised the target market. Mailing lists were secured from the focal bank, and the businesses were selected using random sampling techniques. A sample respondent was asked to rate each of the three banks on each of seventeen variates. Respondents were instructed to give their perceptions about a bank even if they were not totally familiar with that bank. These seventeen variates measured important bank selection criteria (as perceived by the customer), and comprised the predictor variables for the BSR analysis. The criterion variable in this study was a measure of the overall preference for a particular financial institution The criterion and predictor variables were quantified using six-point bipolar scales. For these scales, a six represented a favorable response whereas a one represented an unfavorable evaluation. Thus, the adaptive BSR procedures are used to derive the statistically significant determinants of overall customer preference for a banking institution.

Two hundred and fifty survey forms were returned; however, lack of information further restricted the useable sample size for each institution. Ratings on all eighteen variates for a particular bank by a sample respondent were required for that respondent to be included in the final sample for a bank. There was no attempt made to estimate missing data on a variate for a bank. Survey results indicated that each respondent in the final sample for a bank is a current commercial customer of at least one of the three institutions.

The entire survey was conducted in a Standard Metropolitan Statistical Area with a population of 500,000. It can be argued that this market is fairly characteristic of the national marketplace in terms of geographic, demographic, socioeconomic, and lifestyle variables. However, given the rapidly changing marketplace for financial services where customers are continually revising their preferences and expectations of banks, a major premise of this paper is that banks use adaptive methods for modeling customers' perceptions of financial services on a market-by-market basis. Therefore, interpretations of the results of this study are restricted to the target market used here.

EMPIRICAL RESULTS

Correlation matrices for the financial institutions, along with means and standard deviations for the selected image criteria, are given in Tables 1 through 3. Tables 4 through 6 report standardized regression coefficients resulting from the BSR and OLS procedures. Since the BSR coefficients have a multivariate t-distribution, and given the relatively large sample sizes for each bank, a parameter estimate is considered to be statistically significant in this study if it is more than twice its standard error.

Tables 1 and 4 give estimation results for Bank 1, the focal institution. Table 1 indicates that each predictor variable is positively correlated with the criterion of overall preference for this institution. All [TABULAR DATA FOR TABLE 1 OMITTED] [TABULAR DATA FOR TABLE 2 OMITTED] [TABULAR DATA FOR TABLE 3 OMITTED] [TABULAR DATA FOR TABLE 4 OMITTED] [TABULAR DATA FOR TABLE 5 OMITTED] [TABULAR DATA FOR TABLE 6 OMITTED] correlations between predictor variables are positive as well. Therefore, it is reasonable to expect the signs of the regression coefficients to be positive. Table 4 shows that a significant determinant of overall preference for Bank 1 based on OLS estimation is that the bank affords its customers with convenient hours of operation; however, the sign of the estimated regression weight is negative. This result is a suppressor effect and is quite troublesome for purposes of interpretation. The effects of measurement error are distorting the results of parameter estimation.

It is evident that more specialized methods of estimation are needed to account directly for the conditions that may characterize a model involving customers' perceptions of banking services. The adaptive BSR framework outlined here provides a random variable approach to estimation where all variables are, at best, assumed to be fallible measures of underlying (latent) variables. The usefulness of the methodology to model customers' perceptions of financial services is illustrated below.

BSR(m) estimates for Bank 1 were derived from a common factor model with m = 1,2,3 common factors and are shown in Table 4. The significant BSR weights for different choices of m are seen to be quite similar, although the standard errors for the BSR(3) estimates are smaller than the standard errors for the BSR(1) and BSR(2) estimates. Various methods for choosing a value for m in a common factor model are available to researchers (Akaike 1987; Bozdogan 1990; Joreskog 1969; Lawley and Maxwell 1971). In the context of Bayesian Structural Regression, where a common factor model is used as an approximate structural representation for the prior mean of [Sigma], the selection of m may not play a critical role in obtaining reliable regression estimates. As evidenced in the simulations discussed earlier, a reasonable choice for the number of common factors should suffice in most applications. Table 1 shows that the three largest eigenvalues of the joint predictor-criterion correlation matrix for Bank 1 exceed unity and appear to be relatively larger than the remaining 15 eigenvalues. Accordingly, a common factor model with three common factors is considered to be an adequate prior structural representation for use in the computation of BSR estimates. Thus, in modeling customers' perceptions for Bank 1 it is assumed that the 18 variates can be regarded as observed indicators of three latent measures - the three common factors.

The goodness of fit index (1 - w) computed from the data for the focal bank is 0.68. An index of this magnitude indicates that the BSR(3) estimates rely rather strongly on the parameter estimates associated with the prior common factor model. In this case it is reasonable to examine the latent measures corresponding to the structural model for the prior mean of [Sigma]. The fact that all the correlation coefficients between predictor variables and the criterion variable in Table 1 are positive indicates that commercial customers who rate a bank above average on any one of the predictor variates also tend to give that bank an above average overall preference rating. A varimax rotation of the factor loadings reveals that the first factor, on which all the loadings are positive, accounts for this characteristic of the data. The largest loadings on this factor are for V1 - Overall Preference for the Bank, V2 - Friendliness, V3 - Reputation, V4 - Progressiveness, V5 - Financial Stability, V6 - Professionalism (Customer Service), V7 - Experience (Loan Officers), V8 - Aggressiveness (Get and Keep Business), V11 - Extensiveness of Services, V15 - Quality of Services, V16 - Ability to Obtain Loans, V17 - Commercial Deposits, and V18 - Commercial Loans. On the second factor, the largest loadings are for V13 - Branch Locations (Convenience) and V14 - Hours of Operation (Convenience). The largest loadings on the third factor are for V9 - Interest Rates (Loans), V10 - Interest Rates (Deposits), and V12 - Service Fees. Clearly, the first factor represents the variates which are related to overall preference for the focal bank from the customer's perspective. One may argue that BSR estimates can be based on a prior structural model with this single factor. However, there is potentially useful structural information regarding the off-diagonal elements of [Mathematical Expression Omitted] contained in all three factors. For example, the structural information represented by the second and third factors conveys what commercial customers think about the focal bank, but that information does not necessarily reveal why those customers prefer doing business with the bank. One should bear in mind that the structural information contained in these three factors is used adaptively in the derivation of BSR estimates. The degree of reliance on the structural information depends on how well that information is supported by the data. The structural goodness of fit index is relatively high for the focal bank. Nevertheless, in this case, complete reliance on the prior structural model is not supported by the data. The adaptive BSR methodology will derive the significant determinants of overall preference for the focal bank by automatically discounting the structural information to the degree warranted by the data.

The regression results for Bank 1 are given in Table 4. The BSR(3) estimates in Table 4 indicate that the focal bank is a preferred choice for commercial loans and deposits. Professionalism in customer service is also a significant determinant of customer preference for the focal bank. The BSR(3) weights show no apparent irregularities and, unlike the OLS results, the important determinants of overall preference for Bank 1 have regression coefficients with positive signs. Thus, these empirical results show that marketing knowledge may be seriously misrepresented by using OLS to model customers' perceptions of financial institutions. For comparison purposes, BSR estimates based on the ridge form are included in Table 4. For this example, the BSR(0) estimates are similar to the BSR(3) estimates; however, the BSR(3) estimates have lower standard errors (and therefore higher t-values) than the BSR(0) estimates. Also, from a practical point, the significant determinants of preference for a bank may be ordered according to their absolute t-ratios. This may be useful when a bank is interested in identifying priorities regarding its service delivery approaches.

Similar analyses were conducted for each of the focal bank's primary competitors in its target market (financial institutions labeled Bank 2 and Bank 3). Results of the analyses are reported in Tables 2, 3, 5, and 6. For each of these competitors, substantive interpretations are made using a BSR coefficient vector derived from a common factor model with m = 3 factors. As before, this decision was based on the magnitude of the eigenvalues reported in Tables 2 and 3. The goodness of fit indices associated with the structural models used for Bank 2 and Bank 3 are 0.64 and 0.62, respectively. The factor structure for Bank 2 is similar to the structure obtained for the focal bank. The only notable difference is that V14 - Hours of Operation (Convenience) for Bank 2 has a high loading on the first factor, i.e., the factor where individual variates are related to overall preference for Bank 2. This particular variate did not have a high loading on the factor relating individual variates and overall preference for the focal bank. Also, there are no major differences in the factor structure for Bank 3 compared with the factor structure for the focal bank.

The BSR(3) estimates in Table 5 show that Bank 2 is a preferred choice for commercial deposits. Other significant determinants of overall preference for Bank 2 are its convenient hours of operation and general reputation. This bank also is a good choice for commercial loans. In addition, the ability to obtain loans through this bank is an important dimension of perceived image. Several of these significant image characteristics obtained from the BSR(3) model have much weaker effects in the OLS and ridge regression models. This is an indication of the power of the BSR procedures based on a prior common factor model to determine statistically significant effects.

The BSR(3) estimates in Table 6 show that the most significant determinant of overall preference for Bank 3 is that the bank is a good choice for commercial loans. Other important dimensions of preference for this bank are its professionalism in customer service, friendliness, and reputation. This institution also is a preferred choice for commercial deposits. Note that a distinctive feature of preference for Bank 3 in the OLS model is the high interest rates it pays on deposits. However, this effect is nonsignificant in both the ridge and BSR(3) models.

One final comment should be made regarding these results. Tables 4, 5, and 6 indicate a difference between the ridge regression and the BSR(3) models for Bank 2 and Bank 3 but not for the focal bank. This is especially apparent in the case of Bank 2 and may be due in part to the larger p/n ratio for Bank 2 and Bank 3 compared with the p/n ratio for the focal bank.

CONCLUSIONS AND IMPLICATIONS FOR RESEARCH

There are several interesting conclusions and research implications that may be drawn from these results. One of the primary causes of service design failures is the lack in understanding of the evolving needs and preferences of targeted customers (Bateson 1990). Also, important characteristics of a successful service firm are its ability to ascertain its competitive position within a target market and satisfy customers better than the competition (cf. Heskett 1990; Heskett, Sasser, and Hart 1990; Quinn 1992; Rust and Oliver 1994). Knowledge of perceived similarity and divergence in resources and capabilities may, and often will, be important for competitive advantage. The results of this study reveal a high degree of concordance between the focal bank and its competitors in terms of perceived image. In fact, all significant determinants of commercial customer preference for the focal bank are dimensions of preference that are common to its competitors in the target market. All three financial institutions are perceived as good choices for commercial loans and deposits. Professionalism in customer service is another perceived determinant of preference shared by the focal bank with one of its primary competitors. The focal bank may choose to reposition itself since its perceived market image is similar with the perceived image of its competitors for this target market (Porter 1985).

Khazeh and Decker (1992-93) argue that perceived differences among banks are the true determinants of customers' bank selection decisions. There are several significant dimensions of customer preference for Bank 2 and Bank 3 that differentiate those banks from the focal bank. Reputation, friendliness, hours of operation, and ability to obtain loans are significant determinants of preference for Bank 2 and Bank 3; however, these perceived dimensions of preference do not exist for the focal bank. The focal bank may, therefore, opt for a resource commitment to enhance its perceived image in these areas and improve its competitive position in the market.

Also, the focal bank may attempt to reposition itself by restructuring its delivery system to enhance perceived image in areas of customer preference that currently do not exist in this target market. Characteristics of banks such as progressiveness, financial stability, experience of loan officers, aggressiveness to get and keep business, interest rates on loans and deposits, extensiveness of services and service fees, convenience of branch locations, and general quality of services appear to be fertile areas for future inquiry. These characteristics are important to commercial customers; however, none of the variates measuring these particular dimensions of image were assessed as significant determinants of overall preference for banks in the target market. Since competition is intensifying in the banking industry, banks that create a differential advantage in terms of customer preference and image will undoubtedly be in a much stronger position to use that advantage in enhancing market performance.

The present study has some limitations. This research is based on a cross-sectional design, a paradigm which does not allow one to investigate how customers' perceptions of the services of banks are changing over time. A banking institution will be in a better position to monitor and evaluate the success of its consumer-directed strategies by measuring customers' preferences over time. Additionally, given the complex nature of the marketplace for financial services, any generalizations of the findings obtained here to other target market settings are not warranted. The adaptive BSR customer paradigm and its use of vague or diffuse prior structural information is especially designed for modeling preference data of targeted customers on a market-by-market basis. A related point is the environment of mergers and acquisitions which prevails in the mid-1990s. This has created several large regional banks (CitiCorp, Chase [the new Chase], NationsBank, First Union, Wells Fargo, etc.). The evaluation of targeted customer preference data on a market-by-market basis may not be meaningful to these mega-banks.

The adaptive BSR paradigm for modeling customers' preferences has provided useful and timely information on the potential areas for resource investment within the focal firm, information that may allow the firm to achieve a better relative competitive position in its market. The BSR methodology is a random variable approach that accommodates the existence of measurement errors in all observed variables. It includes OLS and ridge regression as special cases and is flexible enough to accommodate a wide variety of research needs. The adaptive methodology assumes that prior information takes the form of a structural model. If the observed data are not consistent with that structural model, the structural information will be automatically discounted, and estimation will revert to OLS estimation. As illustrated in this bank image study, this prior information will often enhance the stability of derived variable effects and provide a useful amount of valid customer information regarding the services of banking institutions.

One final point is that the BSR methods discussed in this paper use common factor structural models. The use of common factor models seems generally appropriate in marketing applications where one's prior structural knowledge is diffuse or vague, and when variables are measured with less than perfect reliability. The Bayesian framework discussed here can, in theory, accommodate a wide variety of structural equation models; for example, those recently developed structural equation systems that are intended to estimate parameters of prechosen structural models using very general modeling software. However, the adaptive BSR procedures using exploratory common factor models do not impose a rigid assumption about the structure of the population covariance system. Unlike other studies in the consumer research literature, the methods discussed here permit much flexibility in the use of structural information to test hypotheses and formulate theory.

In summary, the findings here give evidence of the appropriateness of the adaptive BSR procedures in marketing response applications. The effectiveness of the approach in both an absolute and a relative sense has been discussed. This author is not aware of any other study in the marketing literature that uses this adaptive BSR methodology. This study has concentrated on using the methodology to examine the significant determinants of preference for a banking institution from the customer's perspective. Results of this application indicate that serious misrepresentations of marketing knowledge can be avoided by using the adaptive BSR customer paradigm. It is expected that the methodology may improve the treatment of model building and applied linear prediction in future marketing response applications, especially in the general area of consumer research.

1 Details for implementing the EM algorithm for this prior structure are outlined by Chen (1979), 240). The EM algorithm is guaranteed to converge under the general conditions specified by Dempster, Laird, and Rubin (1977).

REFERENCES

Akaike, H. (1987), "Factor Analysis and AIC," Psychometrika, 52(3): 317-332.

Anderson, E. W. and M. W. Sullivan (1993), "The Antecedents and Consequences of Customer Satisfaction for Firms," Marketing Science, 12(2): 125-143.

Bateson, J. E. G. (1990), "Evaluating the Role and Place of Marketing in Service Firms," in Service Management Effectiveness, D. E. Bowen, R. B. Chase, and T. G. Cummings (eds.), San Francisco: Jossey Bass: 324-342.

Bearden, W. O. and J. E. Teel (1983), "Selected Determinants of Consumer Satisfaction and Complaint Reports," Journal of Marketing Research, 20(February): 21-28.

Bolton, R. N. and J. H. Drew (1991a), "A Longitudinal Analysis of the Impact of Service Changes on Customer Attitudes," Journal of Marketing, 55(January): 1-9.

Bolton, R. N. and J. H. Drew (1991b), "A Multistage Model of Customers' Assessments of Service Quality and Value," Journal of Consumer Research, 17(March): 375-384.

Bozdogan, H. (1990), "On the Information based Measure of Covariance Complexity and its Application to the Evaluation of Multivariate Linear Models," Communication in Statistics, Theory and Methods, 19(1): 221-278.

Cardozo, R. N. (1965), "An Experimental Study of Customer Effort, Expectation, and Satisfaction," Journal of Marketing Research, 2(August): 244-249.

Chakravarty, S., R. A. Feinberg, and R. Widdows (1995), "What Do Customers Want from Banks?" Journal of Retail Banking, 17(Summer): 15-19.

Chen, C-F. (1979), "Bayesian Inference for a Normal Dispersion Matrix and its Applications to Stochastic Multiple Regression Analysis," Journal of the Royal Statistical Society, Series B, 41: 235-248.

Churchill, G. A. and C. Suprenant (1982), "An Investigation into the Determinants of Customer Satisfaction," Journal of Marketing Research, 19(November): 491-504.

Cronin, J. J. and S. A. Taylor (1992), "Measuring Service Quality: A Reexamination and Extension," Journal of Marketing, 56(July): 55-68.

Darlington, R. B. (1990), Regression and Linear Models, New York: McGraw-Hill.

Dempster, A. P. (1969), Elements of Continuous Multivariate Analysis, Cambridge, MA: Addison-Wesley.

Dempster, A. P., N. M. Laird, and D. B. Rubin (1977), "Maximum Likelihood from Incomplete Data Via the EM Algorithm (with Discussion)," Journal of the Royal Statistical Society, Series B, 39: 1-38.

Erevelles, S. and C. Leavitt (1992), "A Comparison of Current Models of Consumer Satisfaction/Dissatisfaction," Journal of Consumer Satisfaction, Dissatisfaction and Complaining Behavior, 5: 104-114.

Fuller, W. A. (1987), Measurement Error Models, New York: John Wiley and Sons.

Glassman, C. A. (1995), "Industry Structure: Erosion of Banks' Franchise," Journal of Retail Banking Services, 17(Autumn): 53-58.

Heskett, J. L. (1990), "Rethinking Strategy for Service Management," in Service Management Effectiveness, D. E. Bowen, R. B. Chase, and T. G. Cummings (eds.), San Francisco: Jossey Bass: 17-40.

Heskett, J. L., W. E. Sasser, and C. W. Hart (1990), Service Breakthroughs: Changing the Rules of the Game, New York: Macmillan.

Hoerl, A. E. and R. W. Kennard (1970), "Ridge Regression: Biased Estimation for Non-orthogonal Problems," Technometrics, 12(1): 55-67.

Joreskog, K. G. (1969), "Efficient Estimation in Image Factor Analysis," Psychometrika, 34(1): 51-75.

Khazeh, K. and W. H. Decker (1992-1993), "How Customers Choose Banks," Journal of Retail Banking, 14(Winter): 41-44.

Kimball, R. C. and W. T. Gregor (1995), "How Distribution is Transforming Retail Banking: Changes Leading Banks are Making," Journal of Retail Banking Services, 17(Autumn): 1-9.

LaBarbera, P. A. and D. Mazursky (1983), "A Longitudinal Assessment of Consumer Satisfaction/Dissatisfaction: The Dynamic Aspect of the Cognitive Process," Journal of Marketing Research, 20(November): 393-404.

Laughlin, J. E. (1986), "Contrasting Alternatives to Least squares in Regression Using Diagnostics for Identifying Influential Data," Multivariate Behavioral Research, 21(1): 77-102.

LaTour, S. A. and N. C. Peat (1979), "Conceptual and Methodological Issues in Consumer Satisfaction Research," in Advances in Consumer Research, W. L. Wilkie (ed.), Ann Arbor, MI: Association for Consumer Research: 431-437.

Lawley, D. N. and A. E. Maxwell (1971), Factor Analysis as a Statistical Method, London: Butterworth.

Oliva, T. A., R. L. Oliver, and I. C. MacMillan (1992), "A Catastrophe Model for Developing Service Satisfaction Strategies," Journal of Marketing, 56(July): 83-95.

Oliver, R. L. (1977), "Effects of Expectations and Disconfirmation on Post exposure Product Evaluations," Journal of Applied Psychology, 62(April): 246-250.

Oliver, R. L. (1980), "A Cognitive Model of the Antecedents and Consequences of Satisfaction Decisions," Journal of Marketing Research, 17(November): 460-469.

Oliver, R. L. and W. O. Bearden (1985), "Disconfirmation Processes and Consumer Evaluations in Product Usage," Journal of Business Research, 13(June): 235-246.

Oliver, R. L. and W. DeSarbo (1988), "Response Determinants in Satisfaction Judgements," Journal of Consumer Research, 14(March): 495-507.

Oliver, R. L. and J. E. Swan (1989), "Consumer Perceptions of Interpersonal Equity and Satisfaction in Transactions: A Field Survey Approach," Journal of Marketing, 53 (April): 21-35.

Porter, M. E. (1985), Competitive Advantage: Creating and Sustaining Superior Performance, New York: Free Press.

Pruzek, R. M. and G. M. Lepak (1992), "Weighted Structural Regression: A Broad Class of Adaptive Methods for Improving Linear Prediction," Multivariate Behavioral Research, 27(1): 95-129.

Quinn, J. B. (1992), Intelligent Enterprise, New York: The Free Press.

Rust, R. T. and R. L. Oliver (1994), Service Quality: New Directions in Theory and Practice, London: Sage Publications.

Tse, D. K. and P. C. Wilton (1988), "Models of Consumer Satisfaction Formation: An Extension," Journal of Marketing Research, 25(May): 204-212.

Westbrook, R. A. (1981), "Sources of Consumer Satisfaction with Retail Outlets," Journal of Retailing, 57(Fall): 68-85.

Yi, Y. (1990), "Critical Review of Consumer Satisfaction," in Review of Marketing 1990, V. A. Zeithaml (ed.), Chicago: American Marketing Association: 68-123.

Greg M. Lepak is Professor, Business Administration, Le Moyne College, Syracuse, NY. The Journal of Consumer Affairs, Vol. 31, No. 1, 1998