Satisfaction and Loyalty in a Machine Learning Context

ABSTRACT - Machine learning is introduced as a method of analysing satisfaction and loyalty data. This technique discretizises data and discovers rules for the obtained intervals. No linear relationships are presumed and different relationships between different intervals are allowed for. In an automobile context it is shown that additional insights in the relationship between satisfaction and loyalty can be gained from using machine learning. In particular, distinctions can be made between dissatisfiers, satisfiers, performers and irrelevant factors in explaining loyalty.

Citation:

JosTe M.M. Bloemer, Koen Vanhoof, and Koen Pauwels (1998) ,"Satisfaction and Loyalty in a Machine Learning Context", in AP - Asia Pacific Advances in Consumer Research Volume 3, eds. Kineta Hung and Kent B. Monroe, Provo, UT : Association for Consumer Research, Pages: 138-146.

Asia Pacific Advances in Consumer Research Volume 3, 1998      Pages 138-146

SATISFACTION AND LOYALTY IN A MACHINE LEARNING CONTEXT

JosTe M.M. Bloemer, Limburg University Centre, Belgium

Koen Vanhoof, Limburg University Centre, Belgium

Koen Pauwels, Limburg University Centre, Belgium

ABSTRACT -

Machine learning is introduced as a method of analysing satisfaction and loyalty data. This technique discretizises data and discovers rules for the obtained intervals. No linear relationships are presumed and different relationships between different intervals are allowed for. In an automobile context it is shown that additional insights in the relationship between satisfaction and loyalty can be gained from using machine learning. In particular, distinctions can be made between dissatisfiers, satisfiers, performers and irrelevant factors in explaining loyalty.

1. INTRODUCTION

In times of severe competition and rising customer expectations, firms are more and more interested in keeping existing customers instead of attracting new ones (Heskett et al, 1994). While chances of expanding the market are small, companies realize that luring consumers away from competitors by expensive promotions, product line extensions or price reductions may not be in line with long-term organizational goals (Reichheld, 1996; Fornell and Wernerfelt, 1987). Therefore, satisfying existing customers becomes a main objective, implicitly assuming a direct link with loyal behaviour demonstrated by these customers. However, the relation between satisfaction and loyalty is far from obvious and still requires additional research.

Over the years, research in the field of the relationship between satisfaction and loyalty has been confronted with a number of conceptual, methodological, analytical as well as operational drawbacks. This article will concentrate on some of the analytical difficulties.

First, most of the time, the distribution of satisfaction scores is rather skewed (Fornell, 1992; Reichheld, 1996). Many respondents indicate that they are 'very satisfied’ or at least 'satisfied’ and only a few indicate that they are 'not so satisfied’ or 'not satisfied at all’. Second, the most widely used analysing techniques assume linear relations between satisfaction and loyalty. However, there are indications that these types of relations may not be typical for this research domain. It has been suggested for instance that the relationship between low levels of satisfaction and loyalty may differ from the relationship between high levels of satisfaction and loyalty. Coyne (1989) hypothesis the existence of thresholds of satisfaction that have to be met in order to affect customer behaviour. Whereas, Heskett et al (1994) and Jones and Sasser (1995) argue that only extremely satisfied customers demonstrate loyal behaviour.

This picture becomes even more complicated if not just one type of loyalty is taken into account but different types. There are a number of studies that report on store loyalty in general or dealer loyalty in particular (Cunningham, 1956, 1961; Carman, 1970; Tranberg and Hansen, 1985; Bloemer et al, 1990 and Bloemer and Lemmink, 1992). These studies generally indicate that store or dealer loyalty is an intervening variable between satisfaction with the product and brand loyalty. However, the exact nature of the relationship remains unclear.

Concentrating on these analytical problems within the field of the relationship between customer satisfaction and loyalty, we introduce in this article an inductive method of analysis, which helps to resolve some of the above mentioned issues. The technique is called machine learning. It discretizises data and discovers rules for the obtained intervals. No linear relationships are presumed and different relations between different intervals are allowed for. Moreover, machine learning allows to take into account the skewness of the original data-set. Therefore, it helps to provide better insight into the nature of the relationship between satisfaction and loyalty. Building on previous research concerning brand and dealer loyalty, the relationship between satisfaction with the car, satisfaction with the dealer (sales and after-sales), brand loyalty and dealer loyalty (sales and after-sales) will be investigated, using machine learning.

This paper is structured as follows. First of all the method of analysis will be introduced. Second, the research setting will be brought to the fore. An overview of the literature related to the relationship between customer satisfaction and brand loyalty within an automobile context will be provided. Third, the research design will be discussed in terms ofthe method of data collection and the measurement instruments used. Fourth, results of the data analysis with help of the machine learning technique will be given. Finally, conclusions will be drawn and managerial implications will be provided.

2. METHOD OF ANALYSIS

The primary goal of machine learning is the discovery, representation and analysis of 'interesting’ data regularities and dependencies between attributes (such as satisfaction) and the concept or goal variable (such as loyalty). In this application 'interesting’ means that we are able to confirm or reject propositions from previous research. Therefore, we use class-sensitive discretization and a context-sensitive relevance measurement technique developed for Knowledge Discovery Systems (KDS), also known as machine learning techniques. In a first step, we will construct an appropriate representation of the data set in the context of the knowledge to be acquired. The discretization process constructs an abstract representation of the data. Such an abstraction enables us in the second step to identify attributes as satisfiers, dissatisfiers or performers. In a third step classification rules, describing the specific constructs, are analysed. The classification rules are evaluated by a confidence and a coverage percentage which deliver additional insights in the relationships between the satisfaction and loyalty constructs.

2.1 Discretization

For discretizing the concept (goal variable) the monothetic contrast criterion (van de Merckt 1993) has been used :

Contrast (N₁, N₂, A) =    N₁N₂   (m_A1-m_A2)²

                                     N₁ + N₂

where N₁, N₂ are the number of cases of the resulting binary split and m_Ai is the mean value for attribute A of N_i instances. The desirable split is the cut point that produces the highest contrast. The resulting intervals were labelled as low, medium, high or maximl loyalty.

The second stage consists of a discretization of the attributes. When the discretization algorithm is concept-sensitive, it is possible to construct intervals of consecutive attribute values which feature a concept distribution that is uniform and homogeneous but, at the same time, contrasts the distributions of adjacent intervals significantly. A top-down method for discretizing continuous attributes based on a minimal entropy heuristic, presented in Catlett (1991) and Fayyad & Irani (1993), is used here.

2.2 Contextual merit and determination of satisfiers,dissatisfiers and performers

We choose contextual merit (which takes this context into account) as the base of our pattern detection scheme. The contextual merit measure (Hong 1994, Vanhoof 1995), is able to capture the relative importance of attributes in distinguishing the concept values in the context of the other attributes. The contextual merit of an attribute can be calculated for each concept value. Before introducing a formal definition of the relevance measure, let us consider an example set that has only symbolic attributes. Suppose a case with concept value Y₁ and a counter case with concept value Y₂, which have all identical attribute values except for two attributes, say X₂ and X₇. These two attributes are the sole contributors in distinguishing the concept values for the two examples, and hence are indicative. Therefore, their relevance measure will be increased. The number of attributes showing a difference between the two cases is 2. We can define the magnitude of the similarity between the cases as the inverse of the distance, 1/2, which should be shared equally by X₂ and X₇, i.e. shared in proportion to their component distances. This is the key observation. The rationale is that counter cases that are far from the case under consideration will only contribute noise to the process. They would be distinguished from the case easily anyway since the distance is relatively large. The philosophy is similar to that of k-nearest neigh bour schemes. Each of the counter cases chosen for the example contributes its 1/D distributed to the attributes in proportion to their component distance. The algorithm to compute the contextual merit (Hong, 1994) for a given concept level is presented below:

MER := null vector

Total_dif := 0;

FOR all cases with a given concept level

DO begin

find N nearest difference counter cases

FOR every counter case

DO BEGIN

Total_dif := Total_dif + dif(case,counter)

FOR all_attributes

DO BEGIN

MER[attr] := MER[attr]

+ difattr(attr,case,counter)/dif(case,counter));

END

END

MER := MER / Total_dif;

with dif(case,counter) the summation of the attribute distances difattr(attr,case,counter), which is always a positive number. For discrete attributes the difference difattr (attr,case,counter) is either 1 (the values are different) or 0 (the values are equal). The normalisation has the advantage that the sum of the relevance measures equals one. Therefore, in a study with for instance five independent attributes, a contextual merit higher than 0.20 (1/5) is considered as high, otherwise as low.

This approach allows us to define the 'interesting’ patterns as follows. An attribute A is called a dissatisfier when it has a high contextual merit for a low loyalty level an a low contextual merit for a high loyalty level. An attribute B is called a satisfier when it has a low contextual merit for a low loyalty level an a high contextual merit for a high loyalty level. An attribute C is a performer when it is both a satisfier and a dissatisfier. An attribute D is a less-relevant attribute when it is neither a satisfier nor a dissatisfier. Figure 1 shows the typical attribute relevance patterns of the different attributes.

Attribute A is only relevant when the concept value is low. Therefore, A can be interpreted as a dissatisfier. When attribute A has not reached a certain threshold, the concept value is low and A is relevant. When A has reached the threshold level, it does not have a strong influence anymore and it becomes less relevant. In contrast, attribute B is only relevant when the concept value is high. Thus, B is seen as a satisfier: other attributes explain low values of the concept, but the value for B is important in explaining high concept values (compare Herbert’s two factor theory (see also Madhouse, 1981)). Finally, the relevance of attrbutes C and D are more or less independent of the concept value. However, the interpretation differs due to the different levels of attribute relevance. Attribute C has a high merit and influences the concept at every level. Therefore, C is a performer. In contrast, attribute D has a low merit and is classified as a less relevant attribute. Using the contextual merit in combination with these descriptions has one main advantage: we have a common selection criterium to classify the most interesting attributes in different contexts.

2.3 Analysis of the corresponding classification rules

The previous classification of attributes in satisfiers, dissatisfiers and performers can be detailed by analysing the corresponding classification rules. A corresponding rule is a rule where the level of the independent attribute is the same as the level of the concept. A case is said to match a rule if it satisfies all the premises in the promise of the rule. The confidence in the rule then is the number of matching cases that actually have the same value for the concept (goal variable) as appears in the conclusion of the rule. The coverage of the rule is the number of matching cases that actually have the same value for the concept (goal variable) divided by the number of all cases that has the value of the concept. So, if there are 160 cases with a specific value of the concept, 100 cases which match the rule, 80 cases which match the rule and have the specific value, then the confidence percentage is 80 and the coverage is 50. The confidence percentage is a measure for the discriminating power of the rule, while the coverage is a measure for the characteristic power of the rule.

For a dissatisfier A, the corresponding rule is 'IF A=low Then loyalty=low '. When this rule shows a high confidence and a low coverage level, this attribute is considered as a penalty attribute. We have high confidence that low scores on such an attribute are 'penalised’ by low loyalty values. When the confidence level is low and the coverage level is high, the attribute is a characteristic descriptor. Although low values on this attribute are often found together with low concept values (high coverage), a low score on this attribute alone is not sufficient to produce low loyalty values (low confidence). Probably, the combination of low values on A and low values on some other attribute causes low concept values. When both confidence and coverage level are high, we may conclude that the attribute level and the concept level are strongly related. Low values on A and low concept values correlate strongly.

For a satisfier B the most interesting corresponding rule is 'IF B=high Then loyalty=high’. The rule is replaced by ' IF B=high Then loyalty=maximal’ for concepts with maximal as a separate concept value. When this rule shows a high confidence and a low coverage level, this attribute is considered as a reward attribute. A high score on such an attribute is almost always 'rewarded’ by high loyalty values. When the confidence level is low and the coverage level is high, the attribute is a characteristic descriptor. When both levels are high, we may conclude that the attribute level and the concept level are strongly related.

FIGURE 1

CONTEXTUAL MERIT OF 4 HYPOTHETICAL ATTRIBUTES

For a performer C, the corresponding rules are: 'IF C=low Then loyalty=low’ and 'IF C=high Then loyalty=high’. Attribute C is called a basic attribute when the first rule indicates a penalty and the second rule a characteristic pattern. Low scores on such an attribute are penalised by low concept values, but high attribute scores do not necessarily lead to high concept values; although they are often found ogether. Attribute C is called an excitement attribute when the first rule shows a characteristic and the second rule a reward pattern. High scores on such an attribute are rewarded by high concept values, but low attribute scores do not necessarily lead to low concept values (although they are often found together). Finally, the combination of a penalty pattern and a reward pattern indicates a strongly related performer.

In summary, the discretization process and the contextual merit relevance measure are used for detecting the interesting patterns of the data set. These patterns are further analysed by interpreting the discrimination and characteristic power of the corresponding rules. A similar technique of firstly using a pattern detection mechanism and secondly qualifying or quantifying the patterns has been also successfully applied in other studies (Vanhoof & Swinnen, 1996).

3. RESEARCH CONTEXT: THE RELATIONSHIP BETWEEN SATISFACTION AND LOYALTY

A literature review reveals several satisfaction-loyalty studies in recent years. Newman & Werbel (1973) indicate that brand loyalty appears to vary directly with the satisfaction with the brand, but a closer look shows that the correlation is not perfect. Not all satisfied customers will be brand loyal, while not every customer who is not fully satisfied will be non-loyal. Similar results are reported by La Barbera & Mazursky (1983), Bearden & Teel (1983), Garfein (1987), Kasper (1988), Woodside et al (1989) Oliver & Swan (1989) and Bloemer and Kasper (1995). Furthermore, using data from the Swedish Customer Satisfaction Barometer, Fornell (1992) and Anderson et al (1994) have discovered industry differences concerning the correlation between satisfaction and loyalty, where low correlations are supposed to be caused by the existence of important switching costs. However, in the context of relatively low switching costs (like automobiles), satisfaction indeed seems to explain a large part of customer loyalty.

In contrast, recent empirical studies (Zeithaml et al, 1996) show the growing doubt of both managers and researchers in the assumed positive relationship between satisfaction and loyalty. The automobile industry, which pioneered the use of satisfaction research and spent twenty years and a lot of money refining its measures (Reichheld, 1996), is an excellent illustration of this concern. Whereas satisfaction scores have skyrocketed, to the point that more than 90% of the customers today report that they are satisfied/very satisfied, repurchase rates remain stuck in the 30- to 40- percent range. Moreover, Heskett et al (1994) report that high levels of measured satisfaction sometimes go hand in hand with a continuous decline of turnover and profits. Research by Jones and Sasser (1995) confirms that the relation between satisfaction and loyalty is nor linear nor simple. Therefore, additional research in this area has become an issue of high priority (Zeithaml et al, 1996).

In fact, the relationship between brand loyalty and store or dealer loyalty has been the subject of a number of studies (Cunningham 1956, 1961; Carman 1970; Tranberg & Hansen 1985, Bloemer et al. 1990, Rust and Zahorik, 1993). However, the picture that emerges from these studies is neither fully transparent, especially in terms of the causal direction of the relationship. For our study, the research by Bloemer and Lemmink (1992) is particularly relevant. It deals with brand loyalty and dealer loyalty as well as with brand satisfaction and dealer satisfaction with a Japanese brand of ca. In their model, they include the relations between four constructs: satisfaction with the product, satisfaction with the dealer (sales and after-sales), (intended) brand loyalty and (intended) dealer loyalty. As expected, satisfaction with the sales service and satisfaction with the after-sales service by the dealer are found to be the major determinants of dealer loyalty. Furthermore, dealer loyalty and satisfaction with the car are the major determinants of brand loyalty. Finally, dealer loyalty appears to be an intervening variable between satisfaction with the sales service, satisfaction with the after-sales service and brand loyalty.

From this overview and also taking into account the analytical problems mentioned earlier, it can be concluded that additional research is needed on the relationship between satisfaction and loyalty to provide better insight, The inductive perspective of machine learning allows us to shed a new light on the problems involved and to 'test’ the following propositions:

P1:  The relationship between low levels of satisfaction and loyalty may differ from the relationship between high levels of satisfaction and loyalty (Coyne, 1989; Zeithaml et al, 1995, Jones and Sasser, 1995);

P2:  Only extremely satisfied customer demonstrate loyalty (Heskett et al, 1994; Jones and Sasser, 1995).

P3:  There exist cross relationships between brand- and dealer related satisfaction on the one hand and brand- and dealer related loyalty on the other hand (Bloemer et al, 1990; Bloemer and Lemmink, 1992)

Satisfaction is defined here as 'the outcome of the subjective evaluation that the chosen alternative meets of exceeds the expectations’ (Engel et al. 1990, p. 481).

Loyalty is defined as 'the biassed (i.e. non-random), behavioural response, expressed over time by some decision-making unit, with respect to one or more alternatives out of a set, which is a function of psychological (decision making, evaluative) processes resulting in brand commitment. This definition is based on the definition by Jacoby and Chesnut (1978, p. 80-81) and it incorporates both the behavioural as well as the attitudinal part of loyalty.

The influences retained in the conceptual model are confirmed by the machine learning tests when the corresponding contextual merits show the highest values. The shape and the relative position of the relevance patterns found is new knowledge for complementing and validating the model or for constructing a new model.

4. RESEARCH DESIGN

Since we choose the automobile market as the research setting, the concepts included in the analysis are: satisfaction with a car, satisfaction with the sales service, satisfaction with the after sale service, brand loyalty, dealer sales loyalty and dealer after-sales loyalty. In comparison with Bloemer and Lemmink (1992), dealer after-sales loyalty is considered to be a separate construct on both conceptual as well as managerial grounds. Symmetry recommends that if one measures the satisfaction with the sales service and the after-sale service separately, the same distinction should be made for the loyalty measures.

The respondents in the empirical part of the study are customers of automobile dealers (n= 407) of a German brand in the Netherlands. The brand is generally regarded as a exclusiveand expensive. Because the respondents had to express their feelings about the sales service, the car had to be bought less than two years before. Furthermore, the customer had to have some experience with the after sales service, which leads us to impose a minimum of a one year ownership. Previous research (Bloemer and Lemmink, 1992) found significant differences with respect to loyalty between new and used cars and between automobiles for private and for business use. In order to avoid difficulties here and prevent them from confusing our findings, we decided to concentrate on new cars for private use. Therefore, a homogeneous group of car owners will be researched here.

In fact, the population consists of the Dutch customers, who bought a new car for private use from an official dealer of the German brand between 1 and 2 years before the study. The RAI Data centrum, where all the cars and their owners in the Netherlands are registered, provided the necessary name and address data which enabled us to draw a random sample and contact the respondents.

In 1994, 1000 owners of the brand were contacted and asked to complete a mail questionnaire. The initial response rate was 30.8%. Several respondents, however, did not qualify to the selection criteria mentioned earlier. After their elimination, we ended up with 205 respondents, which constituted the final research set. This means a final response rate of 20,5 %.

We operationalised the loyalty variables of our model as the combination of a behavioural and an attitude component (Jacoby & Chestnut, 1978) in accordance to the operationalization of Bloemer and Kasper (1995). [The validity of the operationalisation of loyalty is extensively addressed in this article.] Brand loyalty (BL) is measured as the likelihood of a repeat purchase of the brand, taking into account the degree of brand commitment (repeat purchase * brand commitment). The repeat purchase measure ranged from zero (no chance at all to buy the same brand) to ten (absolutely certain to buy the same brand). Commitment is measured with a 6-item validated 5 points Likert-type commitment scale [Sample items for measuring commitment are: If my preferred brand (dealer) is not available right away, it would make little difference to me if I had to choose another brand; If my preferred brand (dealer) is not available, I will choose another brand (dealer). Reliability of the commitment scales was measured with Cronbach alpha which was .79 for the car, .77 for the dealer service and .73] (see also Bloemer and Kasper, 1995). Dealer sales loyalty (DSL) is measured as the likelihood of buying again at that dealer, taking into account the degree of dealer commitment (repeat purchase * dealer commitment). This repeat purchase measure also ranges from zero (no chance at all to buy from the same dealer) to ten (absolutely certain to buy from the same dealer) and commitment, again, is measured with a 6-item validated 5 points Likert-type commitment scale. Dealer after-sales loyalty (DSAL) is measured in a comparable way as dealer sales loyalty.

Eventhough we are aware of the many problems with a valid operationalisation of satisfac- tion (see for instance Peterson and Wilson, 1992) we choose to use a rather simple one here. Satisfaction with the car (SC); satisfaction with the sales service (SS) and satisfaction with the after-sales service (SA) are measured with an open-ended question asking for the percentage of each type of satisfaction. This percentage could range from zero (not at all satisfied) to hundred (fully satisfied).

FIGURE 2

CONTEXTUAL MERIT FOR BRAND LOYALTY

5. RESULTS AND ANALYSIS: CONTEXTUAL MERIT AND CORRESPONDING RULES FOR BRAND LOYALTY

Since we seek explanation for the loyalty level of customers, our analysis alternatively treats brand loyalty, dealer sales loyalty and dealer after-sales loyalty as goal variables. For each type of loyalty, we present the discretization of the loyalty variable and the contextual merit of the other variables in a figure. The corresponding rules are intgrated in the text.

5.1. Brand Loyalty

As demonstrated in figure 2, dealer sales loyalty, satisfaction with the car and satisfaction with the sales service are the most important determinants of low brand loyalty.

It is obvious that the satisfaction with the car must be high for brand loyalty to occur. Our findings show that satisfaction with the car is a performer factor for brand loyalty. However, satisfaction with the car is neither the only nor the most important attribute for brand loyalty. Customers, who are loyal to the dealer, also demonstrate brand loyalty. Another attribute influenced by dealers, satisfaction with the sales service, is a penalty factor. A customer dissatisfied with the sales service is not likely to stay loyal to the brand. However, when the minimum requirements for brand loyalty are met, dealer after-sales loyalty starts to play a role. Customers who return to the same dealer for after-sales service are likely to buy the same kind of brand on the next purchase.

FIGURE 3

CONTEXTUAL MERIT FOR DEALER SALES LOYALTY

5.2. Dealer sales loyalty

High satisfaction scores are at best minimum requirements for dealer sales loyalty to occur. In order to buy again from the same dealer, the customer has to be satisfied with the (exclusive and expensive) car. But this is only the starting point on the road to loyalty. Brand loyalty and dealer after-sales loyalty have to be stimulated too. Confirming the customer in his brand choice and informing him about new exciting models could help to accomplish the first task. The second task required from the dealers is that they have to make sure that the customers return for after-sales service. This objective is somehow problematic in the automobile business, in which, on the average, dealers retain only 30 to 40 percent of the post-warranty service dollars spent on the automobiles they sell (Reichheld, 1996).

5.3 Dealer after sales loyalty

Satisfaction with the after-sales service does not seem to be an important factor in explain ing after-sales loyalty. Probably, the typical customer does not feel capable of judging the after-sales performance of the dealer and searches for other clues to assert the dealer’s qualities (like sales service and car performance). This observation indicates that a relationship approach to this customer is very suitable.

6. CONCLUSION AND MANAGERIAL IMPLICATIONS

Overviewing our findings, the mutual dependence between the loyalty constructs becomes quite clear. Especially brand loyalty and dealer sales loyalty areclosely linked. Table 1 summarizes the findings about the attributes. The detailed classification of the attributes based on the contextual merit analyses also indicated.

Summarizing, satisfaction measures are mainly dissatisfiers or less relevant. Generally, high satisfaction is not enough for loyalty to occur. The pivotal role of brand loyalty and dealer sales loyalty, appearing as performer factors for after-sales loyalty and brand loyalty and dealer after-sales loyalty appearing as performer factors for sales-loyalty, offer a first explanation for these observations. The level of involvement of the buyers could provide another explanation. We refer to Bloemer, Pauwels & Kasper (1996, forthcoming) for this possibility. From a managerial perspective this leads to the conclusion that it is important to stress that the manufacturer of this brand should devote special attention to the dealer network and that dealers have everything to gain from high brand loyalty (through its effect on dealer sales and after-sales loyalty). Nevertheless finding ways to stimulate satisfaction with the car will help to overcome low levels of loyalty, since this attribute is a basic performer for brand loyalty and a penalty dissatisfier for dealer sales loyalty. If a customer is not satisfied with the car the dealer provides, he is not likely to return. Furthermore, satisfaction with the sales service is a dissatisfier for two out of three loyalty constructs. Therefore, dealers should pay extra attention to the crucial period in which the sales service is perceived by the customer. When the customer is not satisfied with the sales service, loyalty is out of the question. A special treatment of the (potential) buyer, in correspondence to the exclusivity of the brand, is very important here.

FIGURE 4

CONTEXTUAL MERIT FOR DEALER AFTER-SALES LOYALTY

Comparing the results of our analysis with the propositions derived from previous research, the findings support the belief that the relationship between low levels of satisfaction and loyalty differs from the relationship between high levels of satisfaction and loyalty. In other words, propostion 1 is supported. However, our findings do not support proposition 2: extremely high satisfaction scores are not a sufficient condition for high loyalty. As for dealer sales and after-sales loyalty, the corresponding satisfaction attributes are dissatisfiers or even less relevant variables. And for brand loyalty, satisfaction with the car plays a role in explaining high loyalty scores, but even in this case it is only a basic factor and not a sufficient condition for high loyalty. Furthermore, additional information is gained about a cross-relationship between brand- and dealerrelated satisfaction and brand- and dealerrelated loyalty construct. Therefore the third proposition can be supported.

Numerous studies have attempted to address whether data mining is superior than statistical analysis methods, starting with the ESPRIT STATLOG-project in 1990-1993. There is no uniform or simple conclusion and most studies advice to use both methods complementary. However, there are some guidelines to follow in selecting an analytical approach. If a high degree of interaxtion exist among varibles, linear modeling may not be appropriate. As the level of interaction increases, so does data mining’s utility. When the underlying frequency distribution is not known, models that assume a data distribution looses some robustness. Most statistical models assume a certain distribution. Data mining does not. Therefore we have chosen for a data mining method and applied a two-stage-least-squares analysis in parallel. Comparing the results, we noticed that identical conclusions can be dwawn concerning the relationships between satisfaction and loyalty concepts. The 2SLS-method lacks the richness of attribute classification according to the proposed conceptual framework. On the other hand, the machine learning method does not provide direct insight into the strength of identified relations. The absence of correlation coefficients and significance test is a disadvantage. In the context of identified mutual dependence between loyalty constructs, neither 2SLS nor machine learning can identify the direction of the relationship. Nevertheles, machine learning does break down these relations in terms of satisfiers and performes, proving additional insight in the nature of relationships. We conclude that the proposed method offers several advantages in comparison with traditional types of regression such as 2SLS. However, additional research is needed to address its weaknesses and explore its strengths in other research settings.

TABLE 1

SUMMARY TABLE OF THE FINDINGS

REFERENCES

Anderson, E., C. Fornell and D. Lehmann (1994), "Customer Satisfaction, Market Share and Profitability: Findings from Sweden", Journal of Marketing, 58, 53-66.

Bearden, W.O. and J.E. Teel (1983), "Selected Determinants of Consumer Satisfaction and Complaint Reports", Journal of Marketing Research, 10, 309-313.

Bloemer, J.M.M. and J.D.P. Kasper (1995), "The complex relationship between consumer satisfaction and brand loyalty", Journal of Economic Psychology, 16, 311-329.

Bloemer, J.M.M., J.D.P. Kasper and J.G.A.M. Lemmink (1990), "The relationship between overall dealer satisfaction, satisfaction with the attributes of dealer service, intended dealer loyalty and intended brand loyalty: a Dutch automobile case", Journal of Consumer Satisfaction, Dissatisfaction and Complaining Behavior, 3, 42-47.

Bloemer, J.M.M. and J.G.A.M. Lemmink (1992), "The importance of customer satisfaction in explaining brand and dealer loyalty", Journal of Marketing Management, 8, 351-364.

Bloemer, J.M.M. and K. Pauwels (1995), "Explaining brand loyalty, dealer sales loyalty and dealer after-sales loyalty: the influence of satisfaction with the car, satisfaction with the sales service and satisfaction with the after-sales service", Internal Research Paper.

Bloemer, J.M.M., K. Pauwels and J.D.P. Kasper (1996), "The underpinnings and implications of an ongoing relationship within the automobile market: the impact of satisfaction and involvement on brand loyalty and dealer loyalty", Internal Research Report, Diepenbeek, Limburg University Centre.

Coyne, Kevin P. (1989), "Beyond Service Fads: Meaningful Strategies for the Real World", Sloan Management Review (Summer), 69-76.

Carman, J.M. (1970), "Correlates of brand loyalty; some positive results", Journal of Marketing Research, 7, 67- 76.

Catlett, J. (1991), "On changing continuous attributes into ordered discrete attributes", In: European working session on learning.

Cunningham, R.M. (1956), "Brand loyaltyBwhat, where, how much?", Harvard Business Review, 34, 116-128.

Cunningham, R.M. (1961), "Customer loyalty to store and brand", Harvard Business Review, 39, 127-128.

Engel, J.R., R.D. Blackwell and P.W. Miniard (1990), Consumer Behavior (6th ed.), Chicago, IL: The Dryden Press.

Fayyad, U.M. and K.B. Irani (1993), "Multi-interval discretization of continuous-valued attributes for classification learning", In: M. Kaufmann (ed.) Proceedings of the 13th Internaltional Joint Conference on Artificial Intelligence, 1022-1027.

Fornell, C. (1992), A national customer satisfaction barometer: the Swedish experience. Journal of Marketing, 56, 6-21.

Fornell, C. and B. Wernerfelt (1987), "Defensive Marketing Strategy by Customer Complaint Management:A Theoretical Analysis", Journal of Marketing Research, 14, 337-346.

Garfein, R. (1987), "Evaluating the impact of customer service delivery systems", Marketing Review , 42, 11-15.

Heskett, J., T. Jones, G. Lovemer, W. Sasser and L. Schlesinger (1994), "Putting the Service-Profit Chain to work", Harvard Business Review (March-April), 164-171.

Hong, S. J. (1994), "Use of contextual information for feature ranking and discretization", Technical report RC 19664, IBM Research Division.

Jacoby, J. and R.W. Chesnut (1978), Brand loyalty measurement and management, New York, John Wiley and Sons.

Jones, T.O. and W.E. Sasser (1995), " Why Satisfied Customers Defect" Harvard Business Review (November- December), 88-99.

Kamber, M. And R. Shinghall (1996), "Evaluating the Interestingness of Characteristic Rules", Proceedings of the Second International Conference on Knowledge Discovery and Data Mining, E. Simoudis, J. Han and U. Fayyad (eds.), AAAI Press, Menlo Park, California.

Kasper, J.D.P. (1988), "On problem perception, dissatisfaction and brand loyalty", Journal of Economic Psychology, 9, 387-397.

LaBarbera, P.A. and D. Mazursky (1983), "A longitudinal assesment of consumer satisfaction/dissatisfaction", Journal of Marketing Research, 20, 393-404.

Maddox, R.N. (1991), "Two-factor theory and consumer satisfaction: replication and extention", Journal of Consumer Research, 8, 97-102.

Merckt, T. van de (1993), "Decision trees in numerical attributes spaces", In: M. Kaufmann (ed.), Proceedings of the 13th International Joint Conference on Artificial Intelligence, 1016-1021.

Newman, J.W. and R.A. Werbel (1973), "Multivariate analysis of brand loyalty for major household appliances", Journal of Marketing Research, 10, 404-409.

Oliver, R.L. and J.E. Swan (1989), "Consumer Perceptions of interpersonal Equity and Satisfactin in transac tions: a field survey approach", Journal of Marketing, 53, 21-35.

Peterson, R.A. and W.R. Wilson (1992), Measuring custome satisfaction: fact and artifact, Journal of the Academy of Marketing Science, 20, 61-71.

Reichheld F. (1996), "The Loyalty Effect. The Hidden Force Behind Growth, Profits, and Lasting Value. Bain Company Inc. Harvard Business School Press. Boston, Massachusetts.

Rust R. T. and A. J. Zahorik (1993), "Modelling the impact of Service Quality on Profitability: A review. Advances in Service Marketing and Management", T. Schwartz (ed.). Greenwich, CT: JAT Press. 247-276.

STATLOG, "Comparative testin and evaluation of statistical and logical learning algorithms for large scale applications in classification, prediction and control". European Commission, 12/93.

Tranberg, H. and F. Hansen (1986), "Patterns of brand loyalty: their determinnats and their role for leading brands", European Journal of Marketing, 30, 81-109.

Vanhoof, K. and G. Swinnen (1995), "Comparing different measures of relevance", In: New techniques and technologies for statistics (Eurostat seminar), 520-533, Bonn.

Vanhoof, K. and G.Swinnen (1996), "Attribute importance. Assessing nonlinear patterns of factors contributing to customer satisfaction", ESOMAR/EMAC Symposium. Research methodologies for the New Marketing (ISBN 92- 831-1248-2), 171-185.

Woodside, A.G., L.L. Frey and R.T. Daly (1989), "Linking service quality, customer satisfaction and behavioral intention", Journal of Health Care Marketing, 5-17.

Zeithaml V. A., L. L. Berry and A. Parasuraman (1996), "The Behavioral Consequences of Service Quality", Journal of Marketing, 60. 31-46.

----------------------------------------