European Journal of Marketing, Vol 28 Issue 10 Date 1994 ISSN 0309-0566

Using Neural Network Analysis to Evaluate Buyer-Seller Relationships

Barry Wray

University of North Carolina at Wilmington, Wilmington, USA

Adrian Palmer

Leicester Business School, De Montford University, Leicester, UK

David Bejou

University of North Carolina at Wilmington, Wilmington, USA.

Keywords: Consumers, Financial services, Neural networks, Relationship marketing

Type of Article: Survey

Conceptual arguments favouring a relational rather than a transactional approach to the study of buyer-seller relationships are now well understood. However,

attempts to quantify the factors contributing towards relationship quality have been held back by the complexity of the underlying factors and their interrelatedness.

Traditional regression techniques are not effective in analysing data with high levels of multi-collinearity and missing information, typical in many studies of buyer

behaviour. Makes use of a relatively new technique – neural network analysis – to try to quantify the factors contributing to buyer-seller relationship quality. The

technique uses a statistically-based learning procedure modelled on the workings of the human brain which quantifies the relationship between input and output

variables through an intermediate "hidden" variable level analogous to the brain. For this study, a neural network was developed with two outcome components of

relationship quality (relationship satisfaction and trust), and five input antecedents (the salesperson¢s sales orientation, customer orientation, expertise, ethics and the

relationship¢s duration). In a comparison of multiple regression and neural network techniques, the latter was found to give statistically more significant outcomes.

New applications within marketing for neural network analysis are being found. Contributes towards the development of the technique and suggests a number of

further possible applications.

Quality Indicators: Readability**, Practice Implications*, Originality**, Research Implications*

Introduction

Relationship marketing has been hailed by some as a paradigm shift which moves attention away from individual

buyer-seller transactions towards ongoing relationships. Instead of adopting a warfare approach to the bargaining of each

transaction, buyers and sellers become interrelated and achieve shared goals through patterned relationships with one

another[1]. In an ongoing relationship, promises are exchanged and relationships developed on the basis of trust[2].

Of course, there is nothing new in the way that organizations have sought to develop ongoing relationships with their

customers. Before the advent of large-scale production systems and mass marketing, producers of goods were able to know

each of their customers individually and able to suggest appropriate product offers. They also knew what level of credit to

trust them with. Mass production inhibited the development of ongoing personal relationships. More recently, the

development of powerful user-friendly databases has allowed organizations to recreate relationships with their customers –

a computer can now know what a business owner knew in his or her head, allowing relationship development

responsibilities to be given to a large number of staff within an organization. Some have argued that in markets where

companies offer similarly high levels of product quality, the quality of an ongoing relationship becomes a means of gaining

competitive advantage[3].

Relationship marketing strategies are not appropriate to all buyer-seller interactions. They are most appropriate where

purchases involve a high level of risk and a relationship acts as a manager of risk exposure[4]. Relationships are often a

necessity where the stream of service benefits is produced and consumed over a period of time or where they can reduce the

transaction costs associated with repeated purchase routines[5].

To suppliers of services, the development of strong relationships helps to build loyalty from customers whose loyalty is

challenged by competing suppliers. Retaining existing customers rather than expensively seeking new ones can have a major

impact on profitability[6].

Although the increasing importance of relational as opposed to discrete exchange has triggered considerable research into

the quality of buyer-seller relationships, measurement of the concept remains poor. Previous research has produced

conceptual models of the antecedents of relationship quality, or quantified simple relationships based on one or two

components at a time.

In reality, the components of relationship quality are complex, with significant interaction between the contributory

components. Attempts to isolate a few of these components at a time have generally failed to recognize interdependence and

to give clear indications of the relative importance of each factor in contributing towards a higher-order relationship quality

construct.

Much of the weakness of previous research can be attributed to the use of regression techniques which require a prior

specification of input and output relationship variables. A number of multivariate techniques, such as Canonical Correlation,

have been developed which do not presuppose relationships other than to identify which are the dependent and independent

variables. An alternative approach – which is adopted in this article – is the use of neural network analysis. Such a

methodology does not require this prior knowledge, since the network "learns" hidden relationships in data implicitly.

Linear, non-linear, interactions and multi-colinearity relationships are thus discovered and automatically assimilated into a

network of relationships. In the context of buyer-seller relationships, the model is capable of learning the relationship

between a salesperson¢s actions and customers¢ perceptions of relationship quality. The neural network is then able to map

any variety of seller¢s actions and link them to a customer response (i.e. it predicts outcomes on the basis of inputs).

This article describes an application of neural network analysis to the study of relationship quality in the financial services

sector, a sector which has seen significant development of relationship marketing strategies. First, it outlines previous

studies of relationship quality, then develops a neural network to analyse data from a survey of customers of financial

services¢ intermediaries. The results of the neural network analysis are finally compared with those obtained from a more

traditional regression analysis.

Previous Studies of Relationship Quality

There has now been considerable research into the factors that hold buyers and sellers together in an ongoing relationship.

Emerging from general models of buyer behaviour, a number of theoretical and empirically-based models have been

developed to explain the processes of interaction between buyers and sellers, both in the industrial sector[7,8,9,10], and

more recently for consumer markets[11,12]. Studies of relationship quality have drawn heavily on the social psychology

literature of interpersonal relationships, for example in explaining trust[13,14] and conflict resolution[15,16]. Analogies

have been drawn between buyer-seller relationships and relationships within families[17].

The development of ongoing relationships between an organization and its customers can involve a wide range of

operational and marketing staff and satisfactory delivery of a service has been seen as pre-requisite to the development of a

relationship[18,19]. However, particular attention has been given to the relationships which develop between buyers and

sales personnel. For some low-contact services, such as financial services, customers¢ most important contact with an

organization may be through its sales personnel.

Previous research into sales persons¢ effectiveness has concentrated on a static analysis of the behavioural, environmental

and organizational factors that influence performance. The relatively static nature of much of this analysis has been followed

by analyses which see the sales person as a dynamic processor of information, analyses in which interaction between

customers and sales personnel results in the development of long-term relationships[12].

A number of studies have identified individual components of relationship quality, using concepts and measurement devices

used in social psychology. A comprehensive model of relationship quality has been proposed by Lagace et al.[20] which

conceptualized two components of quality – trust in a salesperson and satisfaction in the relationship so far. From the

literature, a number of antecedents of these components have been identified, including sellers¢ customer

orientation/empathy, sellers¢ expertise, sellers¢ ethics, the degree of sales orientation shown by a salesperson, and the

duration to date of the relationship.

A number of studies have sought to conceptualize relationship satisfaction. Three dimensions have been attributed to it by

Crosby and Stevens[19]: satisfactory interactions with personnel; satisfaction with the core service (the extent to which a

service satisfies customers¢ needs), and satisfaction with the organization. In a study of life insurance customers, satisfaction

with the core service was found to be closely related to satisfaction with the contact person and the organization.

An important objective of relationship marketing strategies is the development of trust, which has been seen as having a

crucial function in a relationship in allowing tensions to be worked out[14]. Its development results in the exchange of

promises being perceived by both buyer and seller as more important than short-term transactional exchange. Trust

facilitates joint problem solving.

Some studies have sought to measure the extent to which sales personnel exhibit customer orientation[21,22]. Sales

personnel who are customer-oriented "practice the marketing concept at the level of the individual salesperson and

customer"[21]. Such sales personnel are able to empathize with customers and are concerned about satisfying their needs

better than would their competitors, in contrast to sales-oriented personnel who attempt to create demand for their services

with only a secondary regard for the needs of their customer. Saxe and Weitz[21] have analysed customer orientation in

terms of two factors: "relations" and "ability to help". The former refers to the abilities of sales personnel to develop

long-term relationships with customers on the basis of trust, co-operation and conflict resolution, while ability to help refers

to "the ability of salespeople to help their customers satisfy their needs".

The opposite of a salesperson¢s customer orientation is a sales orientation. Many salespeople prefer to sell hard what they

know best, rather than taking trouble to identify customer needs[21]. A sales orientation may be present where the

salesperson fails to diagnose a clients¢ product requirements[23] or the processes by which the client wishes to evaluate

alternatives[24]. Customers may perceive a sales orientation as a form of pressure, and relationship development may be

inhibited by customers¢ continuing suspicion that the seller is paying too much attention to his or her own selling needs,

rather than to their needs as customers.

The inability of many customers properly to evaluate complex high-credence goods and services can put them at the mercy of

sales personnel. Without the technical knowledge necessary to judge salespersons¢ claims, buyers seek sales personnel who

act in an ethical manner. A preoccupation of sales personnel with short-term goals may result in unethical behaviour which

could subsequently endanger the development of long-term buyer relationships[25]. The importance of customers¢

judgements of sales personnel¢s ethics in evaluating relationship quality has received much recent attention[20,29]. It has

been suggested that consumers¢ assessment of a seller¢s ethics are based on the seller¢s past ethical behaviour and their

expectation of future ethical behaviour, based on both personal and societal norms[29].

The effect of sales personnel¢s levels of expertise on sales performance has been researched extensively. Sellers¢ expertise

has had a number of elements attributed to it including: their measurable technical knowledge; their ability to demonstrate

such knowledge and competence; proof that they are expert in their field (e.g. through formal qualifications); and an explicit

statement of availability, ability and capacity to serve the customer[26]. Credibility, reliability, responsiveness and an

ability to get answers were seen as important determinants of a salesperson¢s competence by Hayes and Hartley[27], in

contrast to aggressiveness and persuasiveness, which detracted from it.

Finally, the duration of a relationship has been cited as a factor explaining the quality of a relationship. This is particularly

true in the development of trust, for which Swan and Nolan[28] identify three stages of development. In the first stage, there

has been no opportunity for exploration of each parties¢ credentials; therefore the level of trust between buyer and seller is at

a minimum. Once exchanges have occurred, trust development moves into the second stage, in which the buyer has the

opportunity to check the actual delivery of a service against the promises that the seller has made. Trust is established in the

third stage, where the perceived performance matches the promised performance.

Neural Network Approach to Measuring Relationship Quality

Previous studies of the factors contributing to relationship quality have produced mixed results in terms of the emerged

constructs and the levels of significance obtained. Furthermore, these have tended to give little indication of the relative

importance of each construct in contributing to overall relationship quality or the interaction between underlying factors.

Where regression analysis has been used, models can only be run by a prior specification of the relationship between

dependent and independent variables. By limiting a model to this set of relationships, other, possibly statistically significant,

relationships may be missed.

Many of the constructs which have emerged from previous studies are interdependent: for example ethical behaviour

contributes towards trust and relationship satisfaction, while trust itself can be seen as contributing towards the assessment

of sales personnel¢s ethical credibility. It would be more appropriate to see these constructs in terms of an interdependent

network of influencers of relationship quality.

In analysing ongoing buyer-seller relationships, neural network analysis has three primary advantages over regression

analysis:

1.Neural network development does not require knowledge of the underlying relationships between the input and output

variables (both linear and non-linear), since the network "learns" relationships hidden in the data. These complex

relationships are discovered and automatically assimilated into the weights connecting the nodes of the network. These

weights contain the "learned information" from the network training phase and are analogous to regression coefficients.

2.The associative abilities of neural networks make them more robust to missing and inaccurate data, since the

knowledge of relationships between variables is distributed across numerous network connections. Regression, on the

other hand, cannot tolerate missing data and works poorly with inaccurate data since all relationship knowledge is

stored in a single beta coefficient.

3.Neural networks¢ performance is not diminished by the multi-collinearity problem of regression analysis. Non-standard

conditions, violations of assumptions, high influence points, and transformations can all be handled by the neural

network model.

As well as learning the relationship between a salesperson¢s actions and customer¢s reactions, the neural network approach

to predicting satisfaction and trust also provides the user with the ability to identify factors which have a significant impact

on the customer¢s perception of the salesperson. In this manner, the neural network is both a predicting tool and a

factor-screening tool.

There are two principal phases in neural network analysis: "learning" and "predicting." During the learning, or training,

phase the network "learns" by adjusting the weights between its nodes. The input data must be presented to the network many

times. Data are split into two files. The first is used to train the network and the second file (the recall set) is used as a test

of the network¢s predictive ability. During the training phase the network weights are "saved" at many intervals and tested to

see how well the network can predict outcomes using the weights it has learned up to that point. Following thousands of

iterations, convergence occurs and the best weights for each element of the network can be derived.

Methodology

Sample Frame

The sample for this study was drawn from customers of financial services¢ intermediaries. The financial services sector

presents a good opportunity to study buyer-seller relationships and the sector has been at the forefront of the development of

relationship marketing strategies, for a number of reasons. Customers¢ perception of the riskiness of financial services

purchases can allow a relationship to be used as a means of managing their perceived risk exposure. The longevity of many

financial services entails some form of relationship continuing to exist between an organization and its customers and the

sales personnel are often the only people within the organization with whom a customer will have dealings. Relationships

are attractive to buyers because high transaction costs can occur where a portfolio of financial services is transferred from

one provider to another. Finally, the intangible and often incomprehensible nature of financial services encourages buyers to

judge the ethical credibility of sales personnel.

Instrument Development

A structured questionnaire comprising three sections was developed for this study. The first group of questions related to the

demographic and socio-economic characteristics of respondents; for example, their age, income level, gender and race. The

second group of questions sought information about the behaviour of respondents in relation to financial services; for

example, the types of financial services that they had purchased (e.g. stocks, bonds, etc.) from a financial broker and the

length of their relationship with the broker. The final group of questions sought to elicit respondents¢ attitudes towards their

broker. For this, a set of questions was developed based on the SOCO scale items[21] and other scales. The SOCO scale

has been used previously to measure the extent of personnel¢s sales orientation and customer orientation, and a number of

studies have replicated its results[22].

Seven emerged constructs were used in the neural network analyses:

1.the level of sales orientation/pressure the customer perceived in the broker (selling);

2.the perceived level of the broker¢s customer orientation (customer);

3.the perceived ethical standards of the broker (ethics);

4.the broker¢s perceived expertise (expertise);

5.the customer¢s trust in the broker (trust);

6.the relationship¢s duration (duration);

7.the customer¢s overall satisfaction (satisfaction).

Except in the case of relationship duration, response categories ranged from 7 "strongly agree" to 1 "strongly disagree" (for

relationship duration, response categories ranged from less than a year to three years and more). The selling, customer and

ethics scales were drawn from items of the SOCO scale. Some of the items were worded negatively to reduce response bias

and the scores of these items reversed in the data analysis. Simple unweighted summations of the scores of the scales were

used in the artificial neural network analyses. The reliability of the emerged constructs are shown in Table I . Reliability

Analysis of Scale Items of Emerged Constructs . The duration, expertise, satisfaction and trust scales were indicated by

single-item questions.

Data Collection

A telephone survey was conducted in 1992 in four south-eastern cities of the United States, using a randomly selected

sample based on telephone numbers. Interviewers were asked to interview the adult member of the household who was most

involved in the purchase of financial services. A pilot survey had achieved an acceptable level of response. The problems

of using telephone surveys to collect personal information were recognized, and respondents were asked to answer only

those questions that they felt comfortable with. Out of 1944 interviews, 564 usable questionnaires were available (a

response rate of 29 per cent). A total of 280 callbacks were made for verifications.

Development of the Neural Network

A neural network was developed using the InstaNet submenu in NeuralWorks Professional II. The standard back

propogation configuration suggested by Rumelhart[30] was used. The heteroassociative neural network uses generalized

delta rule learning with five nodes in the input layer, 12 nodes in the hidden layer and two nodes in the output layer. The

input layer had one node for each of the antecedent constructs described above (expertise, duration, selling, customer and

ethics).

The output layer had two nodes for relationship quality components:

1.customer satisfaction with their relationship (satisfaction);

2.customer¢s trust in the seller (trust).

The number of nodes in the hidden layer is based on the rule of thumb that a good guess is roughly twice the number of nodes

in the input layer plus two. A conceptual model of the neural network is shown in Figure 1 . Conceptual Model of the Neural

Network .

A total of 564 survey responses were available and these were randomly split into two files, each of 282 observations. The

first file was used to train the network and the second file (the recall set) was used as a test of the network¢s predictive

ability. During the training phase the network weights were "saved" at many intervals and tested to see how well the

network was able to predict the customer¢s satisfaction and trust based on the weights it has learned up to that point.

"Prediction" involves two separate measures – customer satisfaction and the customers¢ trust in their broker. The difference

between the predicted level of satisfaction and the actual observed level is calculated for every data point in the recall set.

For instance, if the neural network predicted the level of satisfaction to be 4.3 and the actual observed satisfaction is 5.2,

then the absolute difference for that one observation is 0.9. This absolute difference is averaged over all data points to

provide an overall measure to compare between networks. The same is done for the trust variable. The best performance for

predicting satisfaction for the network is an average difference of 0.5760 when the network is trained for 400,000

replications. The best performance for predicting trust for the network is an average difference of 0.5927 when the network

is trained for 400,000 replications. The 282 data points were shown to the network around 1,418 times each to get the best

performance possible. It should be reiterated that the network was predicting for 282 data points not used in network

training.

Validation of the Neural Network

To compare the performance of an artificial neural network to linear regression, a regression equation was computed from

the same data used for training the neural network. Two separate regression equations are necessary because of the two

different output variables – satisfaction and trust. The equations were then used to predict satisfaction and trust from the

same recall data set used to evaluate the neural network. The performance of each approach was tested to determine which

tool is the better predictor. Using the SAS REG procedure, the least squares multiple regression model computed for

predicting customer satisfaction was:

y = 3.120 + 0.3570 × expertise - 0.0333 × duration

- 0.2250 × pressure + 0.3174 customer - 0.1362 ethics

The multiple regression model¢s best performance is an average difference of 0.6098 from the actual rating for customer

satisfaction, whereas it was noted above that the neural network¢s performance was 0.5760. To test the significance of the

difference in predictive ability of the two models, a matched sample pairs statistical procedure was used to test the

hypothesis that the mean difference between the models is zero (i.e. there is no difference between the predictive abilities of

the two models). The test statistic for the t-test is -1.69. The probability of a greater absolute value for this statistic under the

null hypothesis that the population mean is zero (there is no significant difference between the predictive ability of the two

models) is 0.0928. This p-value is evidence that the neural network outperforms the regression model for predicting

customer satisfaction.

The least squares multiple regression model computed for predicting customer trust is:

y = 4.2509 + 0.3132 × expertise + 0.0364 × duration

- 0.3526 × pressure + 0.1599 customer - 0.1252 ethics

The multiple regression model¢s best performance is an average difference of 0.6459 from the actual rating for customer

satisfaction, compared to the neural network¢s performance which was 0.5927. A matched sample pairs statistical procedure

was again used to compare the predictive abilities of the two approaches. The test statistic for the t-test was -2.42 and the

probability of a greater absolute value for this statistic under the null hypothesis that the population mean is 0 is 0.0160.

Again, the small p-value is evidence that the neural network outperforms the regression model for predicting customers¢

level of trust in their broker.

Analysis of the Relative Importance of Factors Contributing to Satisfaction and Trust

The contribution of each of the input factors to relationship quality was determined in the following way.

First, if the neural network¢s ability to predict is observed to be unchanged by removing a factor, the factor should be

"trimmed" (left out of the process). Second, when the neural network¢s predictive ability is significantly reduced by

eliminating a single factor, the factor is contributing unique and valuable information and should be kept. If a factor is

critical in determining the level of customer satisfaction and/or trust, an investigation is warranted to determine why the

factor¢s influence is high. Third, factors having a negative impact on network performance, actually hindering the network¢s

ability to predict by presenting confounding information, should be excluded.

Results

A procedure analogous to step-wise regression was used to investigate the significance of each determinant of customer

satisfaction and trust, the two output components of relationship quality.

Impact of Seller¢s Actions on the Buyer¢s Perceived Level of Satisfaction

To investigate the importance of each of the factors – expertise, duration, selling, customer and ethics – to satisfaction, the

following six-step procedure was used:

1.A new neural network was constructed with one less input processing element using the same paradigm as the full

model (learning rule, transfer function, number of nodes in the hidden layer, etc.)

2.A column of data was eliminated, corresponding to a single factor from the training data set and recall data set.

3.The new network was trained using the "trimmed" file.

4.A recall was performed using the trimmed recall data set.

5.The performance of the trimmed network was computed.

6.Each of steps 2-5 were repeated for each of the five factors using the same basic model as step 1.

Initially each network was trained for 800,000 replications, pausing at intervals of 50,000 replications to save the network

weights. The decision to end the training process was based on the convergence of network learning as before. The optimum

performance of each network at the end of the training process (minimum average deviation from optimal cost) is used to

compare the difference of networks. In order to determine whether the difference in predictive ability of the six networks

(the full model with all five factors and the five models with one factor trimmed from each) is significant, a randomized

block design test was used. The treatments are the elimination of individual factors, while the blocks are the set of survey

results. Since there are six models to test, a two-way analysis of variance (ANOVA) without interaction was considered the

most appropriate statistical test (Table II . ANOVAs for Variables Satisfaction and Trust ). All six distributions exhibit very

similar non-normal characteristics. The result of a test for normality using PROC UNIVARIATE is a p-value < 0.0001.

Since the normality assumption is violated, it was necessary to use a technique to test for homogeneous variance that does

not rely on normality. A rank transformation approach[31] was used to deal with the normality problem. This procedure is a

valid, powerful, and easily implemented non-parametric alternative to the (parametric) ANOVA approach which does not

require homogeneous variance and alleviates the need for a test of homoscedasticity. A non-parametric analysis was created

by transforming the data into ranks and then using the ranks in a parametric ANOVA procedure.

In order to compare the predictive ability of different networks, a file was created containing one observation for each block

within each treatment. The deviations from optimum for all five trimmed networks as well as the full network were used.

The two-way ANOVA analysis revealed a significant difference between the predictive ability of at least one of the six

models (treatments) and at least one block. The F-value of 17.32 is strong evidence (p-value < 0.0001) of a treatment effect

(the treatment being the elimination of one factor from the full model). The significant F-test suggests that at least one of the

factors included in the study had a significant impact on the model¢s predictive ability. The F-test for the blocking factor, the

different survey results, also has a significant impact on predictive ability. The F-value of 17.49 is strong evidence (p-value

< 0.0001) that blocking is important.

In order to compare the performance of each model with all other models, the Walker-Duncan K-ratio T-test procedure was

used on the rank transformed data. As a result of the use of the rank transformed data, the test is more robust and has more

power[31] than a test using the raw data from non-normal populations. The results of the test are shown in Table III . Rank

Transformation One-way Analyses for Variables Satisfaction and Trust . In this table, the networks are ordered according to

the arithmetic mean of differences from optimum.

The most notable result is that the model with all five factors included performed significantly better than the models with

one missing factor, except the model with duration missing. This indicates that all factors except duration are providing

some information important to determining customer satisfaction. The relative importance (ranking) of each factor, from most

important to least important, is as shown in Table IV . Ranking of Factors for Customer Satisfaction .

A second observation is that selling orientation is significantly more important than any other factor included in the study.

Impact of Seller¢s Actions on Buyer¢s Perceived Level of Trust

The same procedure was used to assess the importance of individual determinants on the customers¢ trust in their brokers. A

file containing the difference between the actual trust rating and the predicted trust rating for the full model and for the five

models with one factor left out was used.

The results of the ANOVA procedure for customer trust are shown in Table II . ANOVAs for Variables Satisfaction and

Trust . The two-way ANOVA analysis revealed a significant difference between the predictive ability of at least one of the

six models (treatments) and at least one of the blocks. The F-value of 14.88 is strong evidence (p-value < 0.0001) of a

treatment effect (the treatment being the elimination of one factor from the full model). The significant F-test suggests that at

least one of the factors included in the study has a significant impact on the model¢s predictive ability. The F-value of 15.00

is strong evidence (p-value < 0.0001) that blocking is important. Duncan¢s multiple range test was again used on the rank

transformed data and the results are shown in Table II . ANOVAs for Variables Satisfaction and Trust .

The key observation from the test is that all factors are contributing information significant for determining customer trust.

The relative importance (ranking) of each factor from most important to least important is as shown in Table V . Ranking of

Factors for Customer Trust .

The ranking of the factors for customer trust, however, is different from the customer satisfaction ranking.

Conclusions

Many conceptual models have been developed in the attempt to explain the processes by which salespeople develop ongoing

relationships with their customers; but there have been relatively few attempts to measure and quantify the concept of

relationship quality.

This study has indicated the usefulness of a research methodology appropriate for the analysis of complex ongoing

buyer-seller relationships. In a comparison of regression analysis and neural network analysis, the latter was significantly

better able to explain the relationship between two indicators of relationship quality (relationship satisfaction and trust) and

five of its antecedents (level of salesperson¢s customer orientation, level of salesperson¢s sales orientation, salesperson¢s

ethics, salesperson¢s expertise and the duration to date of the relationship). The neural network analysis has been able to

show that each of these factors has a statistically significant effect on the level of perceived relationship quality.

Relationship satisfaction has been shown to be most influenced by the level of sales orientation customers perceive in their

salesperson. A high level of pressure was negatively related to satisfaction, whereas expertise – which was ranked the

second most important factor – was positively related to satisfaction.

In the case of trust, the primary ranking of relationship duration confirms theories of trust development which see its

development proceeding through a number of stages in which trust only develops after the time has elapsed for a buyer to

check out the seller¢s ability to honour sales promises.

This study has been restricted to customers of financial services organizations, and replication studies would be useful.

While it is quite likely that the methodology is appropriate to a wide range of buyer-seller situations, the input and output

constructs may not be of general applicability. It may be the case, for example, that ethical credibility is perceived as being

an important sought characteristic of financial sales personnel (because customers¢ ability to evaluate their product is low),

but it may be of less relevance where a salesperson is dealing with a tangible good.

Neural network analysis is a relatively new analytical tool which has only recently begun to find marketing applications. The

results of this study suggest that it may offer superior solutions to a wide range of marketing prediction problems

characterized by complex and interdependent, causative variables. The technique has been used in direct marketing to

develop a profile based on customer characteristics of the most profitable type of customer. The technique would also

appear to have potential for predicting consumer choice and the success of site locations.

While this research has offered further insights into the complex factors underlying buyer-seller relationships, further

research would be useful in order to model alternative networks. While the networks used in this analysis used only those

inputs which had been most frequently cited in the literature, it would be useful to build alternative networks which include

factors such as mutual disclosure of information which some research has suggested may contribute towards relationship

quality.

 

Table I . Reliability Analysis of Scale Items of Emerged Constructs

Figure 1 . Conceptual Model of the Neural Network

Table II . ANOVAs for Variables Satisfaction and Trust

Table III . Rank Transformation One-way Analyses for Variables Satisfaction and Trust

Table IV . Ranking of Factors for Customer Satisfaction

Table V . Ranking of Factors for Customer Trust

References

1.Levy, S. and Zaltman, G., Marketing, Society and Conflict, Prentice Hall, Englewood Cliffs, NJ, 1975.

2.Gronroos, C., "Defining Marketing: A Market-Oriented Approach", European Journal of Marketing, Vol. 23 No. 1,

1989, pp. 52-60.

3.Christopher, M., Payne, A. and Ballantyne, M., Relationship Marketing, Heinemann, London, 1991.

4.Lovelock, C.H., "Classifying Services to Gain Strategic Marketing Insight", Journal of Marketing, Vol. 47, Summer

1983, pp. 9-20.

5.Marshall, R.A.B., Palmer, B.A. and Weisbart, S.N., "The Nature and Significance of Agent-Policyholder

Relationships", CLU Journal, Vol. 33, January 1979, pp. 44-53.

6.Reichheld, F.F., "Loyalty Based Management", Harvard Business Review, Vol. 71 No. 2, 1990, pp. 64-73.

7.Ford, D., "The Development of Buyer-Seller Relationships in Industrial Markets", European Journal of Marketing,

Vol. 14, 1981, pp. 339-53.

8.Hakansson, H., International Marketing and Purchasing of Industrial Goods, IMP Group, John Wiley & Sons, New

York, NY, 1982.

9.Frear, C.R. and Metcalf, L.E., "Marketing and Purchasing of Industrial Goods: A Study of the US Commercial

Aircraft Engine Industry", in Turnbull, P. and Paliwoda, S. (Eds), Research Developments in International Marketing,

Manchester School of Management, UMIST, Manchester, 1988, pp. 198-221.

10.Jackson, B., Winning and Keeping Industrial Customers: The Dynamics of Customer Relationships, Lexington Books,

Lexington, MA, 1985.

11.Dwyer, F.R., Schurr, P.H. and Oh, S., "Developing Buyer and Seller Relationships", Journal of Marketing, Vol. 51,

April 1987, pp. 11-27.

12.Crosby, L., Evans, K.R. and Cowles, D., "Relationship Quality in Services Selling: An Interpersonal Influence

Perspective", Journal of Marketing, Vol. 54, July 1990, pp. 68-81.

13.Pruitt, D.G., Negotiation Behaviour, Academic Press Inc., New York, NY, 1981.

14.Sullivan, J. and Peterson, R., "Factors Associated with Trust in Japanese-American Joint Ventures", Management

International Review, Vol. 22, 1982, pp. 33-40.

15.Thomas, K.H., "Conflict and Conflict Management", in Dunnette, M.D. (Ed.), Handbook of Industrial and

Organisational Psychology, Rand McNally, Chicago, IL, 1976, pp. 889-935.

16.Assael, H., "Constructive Role of Interorganisational Conflict", Administrative Science Quarterly, Vol. 14,

December 1969, pp. 573-82.

17.Levitt, T., The Marketing Imagination, The Free Press, New York, NY, 1983.

18.Crosby, L.A., "Maintaining Quality in the Service Relationship", in Brown, S.W. and Gummesson, E. (Eds), Quality

in Services, Lexington Books, Lexington, MA, 1989.

19.Crosby, L.A. and Stevens, N., "Effects of Relationship Marketing on Relationship Satisfaction, Retention and Prices

in the Life Insurance Industry", Journal of Marketing Research, Vol. 24, November 1987, pp. 404-11.

20.Lagace, R.R., Dahlstrom, R. and Gassenheimer, J.B., "The Relevance of Ethical Salesperson Behaviour on

Relationship Quality in the Pharmaceutical Industry", Sales Management, Vol. 4, Fall 1991, pp. 39-47.

21.Saxe, R. and Weitz R.A., "The SOCO Scale: A Measure of the Customer Orientation of Sales People", Journal of

Marketing Research, Vol. 19, August 1982, pp. 340-51.

22.Michaels, R.E. and Day, R.L., "Measuring Customer Orientation of Salespeople: A Replication with Industrial

Buyers", Journal of Marketing Research, Vol. 22, November1985, pp. 443-6.

23.Churchill, G.A. Jr, Ford, N.M., Hartley, S.W. Jr and Walker, O.C. Jr, "The Determinants of Salesperson

Performance: A Meta-Analysis", Journal of Marketing Research, Vol. 22, May1985, pp. 103-18.

24.Symanski, D.M., "Determinants of Selling Effectiveness: The Importance of Declarative Knowledge to the Personal

Selling Concept", Journal of Marketing, Vol. 52, January 1988, pp. 64-77.

25.Ebejer, J.M. and Morden, M.J., "Paternalism in the Marketplace: Should a Salesman be his Buyer¢s Keeper?",

Journal of Business Ethics, Vol. 7, 1988, pp. 337-9.

26.Swan, J.E., Trawick, I.F. and Silva, D., "How Industrial Salespeople Gain Customer Trust", Industrial Marketing

Management, Vol. 14, August 1985, pp. 203-11.

27.Hayes, H.M. and Hartley, S.W., "How Buyers View Industrial Salespeople", Industrial Marketing Management, May

1989, pp. 73-80.

28.Swan, J.E. and Nolan, J., "Gaining Customer Trust: A Conceptual Guide for the Salesperson", Journal of Personal

Selling and Sales Management, Vol. 5, November 1985, pp. 39-48.

29.Whalen, J., Pitts, R.E. and Wong, J.K., "Exploring the Structure of Ethical Attributions as a Component of a

Consumer Decision Model: The Vicarious Versus Personal Perspective", Journal of Business Ethics, Vol. 10, 1991,

pp. 285-93.

30.Rumelhart, D. and McClelland, J., Parallel Distributed Processing, MIT Press, Cambridge, MA, 1986.

31.Perry, L.G. and Cronin, T.P., "A Note on Rank Transformation Discriminant Analysis: An Alternative Procedure for

Classifying Bank Commercial Paper Ratings", Journal of Banking and Finance, December 1986, pp. 605-10.