Customer Satisfaction Modelling and Analysis: A Case Study. Total Quality Management & Business Excellence, 18(5), 545-554. doi:10.1080/14783360701240337 (attached below) Write a summary analysis and determine if they used the correct tools to conduct the analysis. Use APA format and adhere to the writing rubric. 34 pages in length (excluding cover page, abstract, and reference list) ECSI Customer Satisfaction Modelling and Analysis: A Case Study ENRICO CIAVOLINO & JENS J. DAHLGAARD Dipartimento di Filosofia e Scienze Sociali, Universita` degli Studi di Lecce, Italy; Division of Quality Technology, Linkoping University, Sweden ABSTRACT In this paper we analyse the European Customer Satisfaction Index (ECSI) model measured on five different customer groups. In the analysis we use Structural Equation Modelling, which is a general and convenient framework for statistical analysis, including traditional multivariate procedures such as factor analysis or multiple regression modelling. To estimate the structural coefficients, a Multilevel Regression Model is used, which facilitates the analysis of hierarchical data, where observations may be nested within higher levels of classification. In this analysis the hierarchical structure is represented by the customers nested in the companies, so we have two levels of analysis, a micro level nested in a macro level. KEY WORDS: European Customer Satisfaction Index, structural equation modelling, hierarchical data, multilevel regression model Introduction This paper focuses on the application of the European Customer Satisfaction Index (ECSI) model with respect to customers in several context companies. The aim of the paper is to analyse the relations of customers belonging to different customer/company groups taking into account the influence that company structure has on customer satisfaction. The ECSI is represented by a path analysis diagram, which is a graphical representation of a` priori specified structures and assumptions. Ordinary Least Square (OLS) estimates of regression coefficients may be used to estimate the strengths of the structural relationships specified in the diagram. In general, Path Diagrams are used to display graphically the prior hypothesized structure among the variables in a model. For any two variables, X and Y, we have several kinds of relationships, but in our model we consider the relationship X ! Y, where X might structurally influence Y, but not vice versa. This unidirectional straight Total Quality Management Vol. 18, No. 5, 545 554, July 2007 Correspondence Address: Enrico Ciavolino, Dipartimento di Filosofia e Scienze Sociali, Universita` degli Studi di Lecce, Lecce, Italia. Email: enrico.ciavolino@ateneo.unile.it 1478-3363 Print/1478-3371 Online/07/05054510 # 2007 Taylor & Francis DOI: 10.1080/14783360701240337 arrow in a path diagram represents the influence of one variable to another. Arrows pointing to the variables from outside the model represent the collection of all other unmeasured influences, which usually are referred to as error or disturbance. Independent variables that are exclusively influenced by factors lying outside the model are called exogenous, while variables that are hypothesized to be influenced from inside the model are called endogenous. As we said, the data are hierarchical, so we use a statistical method that measures the multilevel effect. Many problems in social and political science can be studied at multiple levels combining these levels of analysis into a single analytical approach. The general concept is that individuals interact with the social contexts to which they belong, meaning that individual persons are influenced by the social groups or contexts to which they belong (Hox, 1995). Generally, the individuals and the social groups are conceptualized as a hierarchical system of individuals, and groups/variables can be defined for each group. ECSI Model and Data Collected Data were collected from five customer groups in Sweden and represent the answers of 89 customers. The customers can be classified into five fields of activities, such as the following: Auto Industry, Service Industry, Motor Vehicle Industry, Other Manufacturing Industry, and Remaining Industry. To measure customers quality perceptions we used the ECSI (European Costumer Satisfaction Index) model which is a model for measuring and analysing customer satisfaction data. In the model we define as exogenous variables the following four variables: Image, Expectation, Hardware, Software. As endogenous variables we used the following three variables: Perceived Value, Customer Satisfaction and Loyalty. These seven variables are latent variables, and are estimated by measuring other kinds of variables, called manifest variables. Table 1 shows all the variables and the relationships among them. The evaluations are made on an ordinal scale from 1 (minimum) to 5 (maximum). Missing data for the manifest variables have been replaced by the mean of the respective variable. A brief description of the latent variables can help to make the discussion easy. . Endogenous Variables Customer Satisfaction (CS) can be defined as an overall evaluation of the service performances or utilization we are analysing. The CS, representing the core of the ECSI model, is in the middle of a cause and effect system running from the antecedent variables (Image, Expectation, Hardware, and Software) and consequence variables (Loyalty). Perceived Value is the perceived level of product quality relative to the price paid or the value for money aspect of the customer experience (Coenders & OLoughlin, 2002). Value is defined as the ratio of perceived quality relative to price (Anderson et al., 1994). Loyalty is the last dependent variable in the model and can be seen as a measure for profitability (Reichheld & Sasser, 1990). Loyalty is measured by repurchase intention and by the intention to recommend products or services to others. 546 E. Ciavolino & J. J. Dahlgaard Table 1. The Variables used in the ESCI Model estimation Exogenous variables Endogenous variables Latent variables Manifest variables Latent variables Manifest variables 2. Image 1. The Company are known to offer priceworthy services 1. Perceived Value 27. The price on The Companys services is reasonable in relation to the delivered value 2. The Company are known to be a reliable supplier of competence 3. The Company are known to offer excellent customer service 4. The Company are known to be environment conscious 3. Expectation 5. General expectations were high when you turned to The Company 25. On the following scale, state your general satisfaction with The Company 6. Expectations were mainly due to information from The Company themselves 4. Hardware 7. Language used in The Companys reports is easy to understand 8. The technical content of the reports is of high level 9. The reports contain straight forward recommendations that are easy to use 6. Customer Satisfaction 10. The Companys reports are easy to grasp 11. It is easy to get hold of reports when needed 12. The Company gives oral presentations when needed 26. Based on your initial expectations, state the result as perceived 16. The Companys consultants are technically competent within their fields 17. In contacts with The Company we have no problems understanding each other 18. The Company reports continuously (Table continued) Customer Satisfaction Modelling and Analysis 547 . Exogenous Variables Image, refers to the brand name and the kind of associations customers get from the product/brand/company. Expectation is referred to the level of quality that customers expect to receive and is the result of prior consumption experience with a firms products or services. Hardware and Software are variables referring to the measure of the service quality, divided in its main aspects. Hardware comprises the attributes of the core product and Software comprises the services related to the core product. The specified ECSI model is shown in Figure 1. The latent variables are estimated by using the average of each manifest variable belonging to the respective latent variable. The arrows in the path diagram show relationships between the variables, and these relationships are estimated by a series of regression analyses (Goldsteen & Ross, 1989). We might estimate the structural coefficients by using the Ordinary Least Square estimation method, but we also have to consider an influence of the classification factor because, Table 1. (Continued) Exogenous variables Endogenous variables Latent variables Manifest variables Latent variables Manifest variables 19. The Company delivers on time 5. Software 13. The Company is supportive when unexpected problems occur 28. If additional needs for expertise would come up, would your company engage The Company again? 14. The Company takes initiative by rationalizing activities when needed 15. The Company takes initiative by making activities more environmentally friendly 20. It is easy to get in contact with The Company 7. Loyalty 21. One is always received with kindness by The Company 29. Would you recommend The Company? 22. The Companys quotations are easy to comprehend 23. The Companys invoices contain all information needed 24. Delivery time changes are dealt with in a satisfactory way 548 E. Ciavolino & J. J. Dahlgaard as said above, the data collected are grouped into five customer/company groups, so we need a statistical method that considers the effect of this classification. The classification of the data structure can be viewed as two levels of hierarchical structure (Companies Customers, Figure 2). The statistical method used to estimate a hierarchical structure is known as multilevel regression model, but is also known under a variety of other names, such as hierarchical model, random coefficient model, and variance components analysis. Although these methods have some small differences, the common element is that the dependent variable is measured at the lowest level of analysis and is a function of predictors measured at this same level as well as predictors measured at one or more higher levels of analysis Jones & Steenbergen (1997). The basic assumption of multilevel modelling is that the dependent variable variation is a function of both lower and higher level predictors. For the regression model we used as dependent variables (endogenous variables) the Perceived Value (PV), Customer Satisfaction (CS), Loyalty (LT), and as independent variables (exogenous variables), the Image (IM), Expectation (EX), Hardware (HD), Software (SW). As we can see we have more then one dependent variable, so for the multilevel model we need for each dependent variable to check the context effect. Figure 1. The ECSI model Figure 2. Hierarchical structure Customer Satisfaction Modelling and Analysis 549 Statistical Model Before analysing the questionnaire we transformed the data set from the ordinal scale to a quantitative scale by using the Thurstone Approach (Zanella, 2001). A path diagram presents a set of structural equations, that is, a set of regression equations with a specified structure among the variables in the model. The structural coefficients are the regression coefficients in these equations and indicate two types of hypostasised relationships: (a) The structural effects of some endogenous variables on other endogenous variables denoted by b, and (b) The structural effects of some exogenous variables on endogenous variables, denoted by g. The first three b are simply the intercepts. The algebraic model in Figure 1, using the variables labels instead of algebraic symbols, can be represented by the following three equations: PVij b10 g11IMij g12EXij g13HDij g14SWij 61ij CSij b20 g21IMij g22EXij g23HDij g24SWkj b21PVkj 62ij LTij b30 g31IMij g34SWij b32CSij 63ij (1) These equations are the analytical representation of the path model. In fact, as we can see, there are, for each endogenous variable, a set of explicative variables, and for the second and the third equations among the explicative variables there are also endogenous variables (PV, for the second and CS, for the third). yij b0j X P p1 b pjx pij 1ij b0j g00 X Q q1 g0qzqj d0j b pj g p0 X Q q1 g pqzqj d pj (2) The general multilevel model used is represented in the system of equations (2). To select the multilevel model and to analyze if there are significant differences between the companies we used an explorative procedure (Hox, 2002). Starting from the simplest possible model, the empty model (in which all the variables are set to zero, the resulting model is like an ANOVA model), we add the parameters (variables and residuals) step by step. The empty model is useful because we can estimate the intraclass correlation (r s2 0=s2 0 s2 1) which shows the part of variance explained by the hierarchical structure. We calculated the intraclass correlation for each equation of the three-equations (1). The results show that only the first equation has a significant intraclass correlation (equal to 0.148), which indicate a group effect (Snjders & Bosker, 2000), so it makes sense to use a multilevel model only for the first equation (the equation that represents the relationship between the PV and the IM, EX, HD, SW). 550 E. Ciavolino & J. J. Dahlgaard The final multilevel model is shown in equations (3) and it is called Random Intercept Model. PVij b0j g11IMij g12EXij g13HDij g14SWij 61ij b0j g00 d0j (3) In this model the regression intercept is assumed to vary across the groups, but the regression slopes are assumed fixed (Snjders & Bosker, 2000), so a company with a high intercept is predicted to have customers with a better perceived value than a company with a low value of intercept. The structural coefficients of the other two equations are estimated by considering the intercepts and the slopes fixed, so it is like estimating the coefficients of the multiple regression model. For recursive path models, regression coefficients based on OLS criterion can be used to estimate structural coefficients, but the estimation method used in the multilevel approach is the maximum likelihood estimation method, which is also a typical estimation method used by the structural equation modelling methodology Lisrel. Results The path modelling has the advantage of showing a complete vision of the phenomenon you are analyzing, but we want to underline the context influence regarding the Perceived Value of the customers. As we said, for the first equation, the model used is the random intercept model. Table 2 shows the results of parameters estimated and the standard deviations for the random intercept and empty model. The empty model is useful as a null-model, which means it is used like a benchmark with which other models are compared. The variance at the customer unit level (s2 ) is estimated equal to 0.481, and the variance at the company level (s2 0) is estimated equal to 0.084. The intraclass correlation is r 0.148, which means that 14.8% of the Perceived Value is explained by the group level. The deviance estimated, equal to 173.596 for the empty model, is a measure of misfit, so for the next model, when we add the explicative variables, it is expected to go down. The Table 2. Results of multilevel analysis Empty model Random intercept Model Coefficients St. Dev Coefficients St. Dev Fixed Part g11 IM 0.453 0.106 g12 EX 20.143 0.094 g13 HD 0.236 0.200 g14 SW 0.332 0.100 Random Part s2 0.481 0.693 0.314 0.561 s2 0 0.084 0.290 0.021 0.147 Deviance 173.596 141.744 Customer Satisfaction Modelling and Analysis 551 Random Intercept model includes micro level variables (IM, EX, HD and SW). The latent variables IM, HD and SW increase the Perceived Value respectively of 0.453, 0.236 and 0.332 points, when all parameters are fixed, which means that when these variables are evaluated high, on average, we expect a high estimated Perceived Value. The structural coefficient of the micro level variable Expectation is equal to 0.143 points, which means that increasing EX variable by 1.0 point, the average of Perceived Value reduces by 0.143 points. The estimated deviance decreases with more covariates (from 173.596 to 141.744), which means that the Random Intercept model fits better the data than the Empty Model, so by introducing the explicative variables and the random effect intercept we reduced the unexplained variance. The causality model of Figure 3 summarizes the various structural regressions of the ECSI model. The structural coefficients are, as we said, the multilevel regression coefficients. Inside the circles the estimated latent variables are shown in brackets ( the average value of the manifest variables), and the values beside the arrows are the structural coefficients. These coefficients are shown in different colours for each endogenous variable (blue for PV, red for CS and black for LO). The values of PV are the same as we saw in Table 2. These values represent the structural equation coefficients estimated by the maximum likelihood method in the multilevel regression. The most import impact on Customer Satisfaction is the Hardware (0.443), but relatively high impacts are shown also from Perceived Value and Software (with the values equal to 0.166). Image and Expectation have less impact on Customer Satisfaction (0.104). Loyalty is a very important factor for the ECSI model. It depends strongly upon Software (0.310) and Customer Satisfaction and less on Image. For this analysis we didnt take into account the significance of the t-tests, because we were interested in analyzing just the descriptive phenomenon, and also because it does not make sense to extend the results to other kinds of enterprises. Figure 3. Results 552 E. Ciavolino & J. J. Dahlgaard Discussion The first interesting consideration is about the hierarchical structure. In fact, the hierarchical structure is influenced only the way in which the customers perceive the value of the Enterprise. The hierarchical structure has no significant direct influence on customer satisfaction or on the loyalty. For the last two structural equations we consider the random coefficients as fixed. The multilevel results for the Perceived Value give us a desirable reduction of variability, so the introduction of the random intercept model facilitates the analysis of a hierarchical structure where the customers are nested within companies. The very interesting aspects are the negative influence of Expectation (-0.143) and the very high positive values of Image (0.453) and Software (0.332). The variable Expectation gives a decreasing effect on the evaluation of the Perceived Value. It seems as if the Company should be careful not to raise customers expectations too high. A better strategy is to reduce the expectations a little and then try to deliver more than expected. For Customer Satisfaction there are indications from the literature suggesting that perceived value is the most significant predictor (e.g. Anderson & Sullivan, 1993; Churchill & Suprenant, 1982; Johnson & Fornell, 1991). However, Martensen et al. (2000) noted that the drivers of both customer satisfaction and loyalty are industry specific. In fact, as we can see from Figure 3 the most important structural coefficient related to customer satisfaction is Hardware (0.332), and the second most important coefficients are Perceived Value (0.166) and Software (0.166). Image (0.104) and Expectation (0.104) have the lowest coefficients. In this analysis, both product Software (0.310) and Customer Satisfaction (0.291) had a significant impact on Loyalty, in fact the product itself gives the main influence on the Loyalty in industries such as fast food or soft drinks, whereas in industries with more highly competitive markets (mobile phone) or with multiple outlets (supermarkets), Loyalty is more image driven (Martensen et al., 2000). We agree with Martensen et al. (2000) that loyalty is in fact the most important outcome measure in the ECSI model and is a truer measure of quality compared with satisfaction. To recommend a product or service to others has greater consequences and requires more commitment than just indicating that one is more or less satisfied compared to an ideal or a competitive product or service (Coenders & OLoughlin, 2002). References Anderson, E. W & Sulliwan, M. W. (1993) The antecedents and consequence of customer satisfaction for firms, Marketing Science, 12(2), Spring, pp. 125 143. Anderson, E. W. et al., (1994) Customer satisfaction, market share and profitability: findings from Sweden, Journal of Marketing, Luglio, pp. 53 66. Churchill, G. A. & Suprenant, C. (1982) An investigation into the determinants of customer satisfaction, Journal of Market Research, 19 (November), pp. 491 504. Coenders, G. & OLoughlin, C. (2002) Application of the European Customer Satisfaction Index to postal services. Structural equation models versus partial least squares, Girona, September, available at: http:// www.udg.es/fcee/economia/n4.pdf. Dahlgaard, J. et al. (2002) From defect reduction to reduction of waste and customer/stakeholder satisfaction (understanding the new TQM metrology), Total Quality Management, 13(8). Goldsteen, K. & Ross, C. E. (1989) The Perceived burden of children, Journal of Family Issues, 10(4), pp. 504526. Customer Satisfaction Modelling and Analysis 553 Goldstein, H. (1995) Multilevel Statistical Models (London: Edward Arnold), available at: http://multilevel.ioe.ac.uk/index.html. Hox, J. J. (1995) Applied Multilevel Analysis (Amsterdam: TT-Publikaties), available at: http://www.fss.uu.nl/ ms/jh/. Hox, J. J. (2002) Multilevel Analysis, Techniques and Applications(Mahwah, NJ: Lawrence Erlbaum Associates). Hox, J. J. & Bechger, T.M. (1998) An introduction to the structural equation modeling, Family Science Review, 11, pp. 354373, available at: http://www.fss.uu.nl/ms/jh/. Johnson, M. D. & Fornell, C. (1991) A framework for comparing customer satisfaction across individuals and product categories, Journal of Economic Psychology, 12(2), (Fall), pp. 267 286. Jones, B. S. & Steenbergen, M. R. (1997) Modelling multilevel data structures. Fourteenth Political Methodology Society Conference, Columbus, OH, July 25. Kreft, I. & De Leeuw, J. (1998) Introducing Multilevel Modeling (London: Sage). Longford, N. (1993) Random Coefficient Models (Oxford: Clarendon Press). Martensen, A. et al. (2000) The drivers of customer satisfaction and loyalty, cross-industry findings from Denmark, Total Quality Management, 11, pp. 85448553. Mueller, R.O. (1996) Basic Principles of Structural Equation Modeling (New York: Springer). Reichheld, F. F. & Sasser, W. E., Jr. (1990) Zero defections : quality comes to service, Harvard Business Review, SeptemberOctober, 105111. Snijders, T. & Bosker, R. (2000) Multilevel Analysis (London: Sage). Tenenhaus, M. (2003) Comparison between PLS and LISREL approaches for SEM: application to the measure of customer satisfaction. Third International Symposium on PLS and Related Methods, Lisbon, Portugal. Zanella, A. (2001) Measures and models of customer satisfaction: the underlying conceptual construct and a comparison of different approaches. Sixth TQM World Congress, Saint Petersburg. 554 E. Ciavolino & J. J. Dahlgaard