# Structuring Latent Consumer Needs Using Lisrel

ABSTRACT - A LISREL (Linear Structural Relationships) model is formulated according to the hierarchical division of customer needs presented in the literature on Quality Function Deployment (QFD). The purpose is to evaluate the relative importance of firstBhand impression and taste experience as regards food products. The empirical illustration is due to a cooperation with a Danish butter cookie company but the authors' opinion is that the model has a much wider application which is supported by the fact that a similar structure has been found in a dataset on peas.

##### Citation:

*
Hans Jorn Juhl, Kai Kristensen, and Anne C. Bech (1995) ,"Structuring Latent Consumer Needs Using Lisrel", in E - European Advances in Consumer Research Volume 2, eds. Flemming Hansen, Provo, UT : Association for Consumer Research, Pages: 298-303.
*

A LISREL (Linear Structural Relationships) model is formulated according to the hierarchical division of customer needs presented in the literature on Quality Function Deployment (QFD). The purpose is to evaluate the relative importance of firstBhand impression and taste experience as regards food products. The empirical illustration is due to a cooperation with a Danish butter cookie company but the authors' opinion is that the model has a much wider application which is supported by the fact that a similar structure has been found in a dataset on peas.

INTRODUCTION

Product quality is a broad concept which possesses the possibility for different interpretations depending on the actual context. In the modern world of quality we speak about two waves (Senge (1992)). In the first wave the focus was on the improvement of tangible work processes and on obtaining measurable improvements in production costs and customer satisfaction. In the second wave of quality we have a new perspective of the consumer. The perspective has now moved from satisfying the consumer's expressed needs to meeting his latent needs. This means that a new set of measurement techniques and a new set of tools for analysis have to be employed. We feel that the LISREL technique has a very promising future in this setting. It allows for a very flexible handling of latent variables and furthermore it makes it possible at the same time to work with structural relationships between the variables. This means that the use of the LISREL technique makes it possible in an easy way to combine technical aspects of food quality with other aspects e.g. psychologic or economic, and this makes it possible to break down barriers between the departments of the company at the product development stage. Such a breakdown will undoubtedly be to the benefit of the whole company.

In relation to food products the technical aspects covers nutritional, hygienic, technological and sensory attributes among others Martens (1984) and Molnar (1993). Packaging and brand name are other product characteristics that the producer is able to change in an attempt to increase customer satisfaction. This study is limited to evaluate the importance of different sensory attributes as regard the perceived quality of butter cookies. We assume the sensory attributes are observed sequentially. Before the food is tasted a consumer may form a firstBhand impression and when the product is eaten the consumer has a taste experience. A LISREL model is formulated and estimated and possible generalizations to other food products are discussed.

THE MODEL

Many factors influence the overall preference of food products. In Shepherd (1989) a framework is given that interrelates the factors related to the food with factors related to the consumer and those related to the social and economic environment. They include the sensory characteristics of the food and the attitudes of the consumer as regards price, nutritional requirements of the food and the sensory attributes. Other psychological factors such as personality, experience or mood are also included. In this study we limit ourselves to study the relationship between preference and the sensory characteristics of the food assuming that all other factors are kept equal.

Customer needs are the necessary input to efficient development of food products and QFD (Quality Function Deployment) is a process to deploy these needs throughout design, manufacturing and service. In Bech et al (1994) we present the "House of Quality" especially designed for food products, here reproduced as Figure 1. The general framework has been modified to illustrate more clearly the specific problems of developing new food products. Traditionally, three sets of data are available when proactive new food products are considered: (1) market research data on customer's preferences, (2) descriptive sensory data from a sensory panel, and (3) the recipe for making the products. However more often than not the three data sets are treated as separate sources of information that may be linked two by two, but hardly ever are seen as an integrated entity. The risk is suboptimization in terms of either customer benefits or production costs.

In the revised House of Quality, the center of transformation has been partitioned to reflect the physicalBchemical dimensions as given in the recipe, along with the sensory dimensions as measured by the panel. By this, we indicate that descriptive sensory analysis is seen as part of the transformation process rather than the goal in itself. Sensory panels are seen more as tools for measurement than indicators of customer demands. It appears that the fact that humans are used as instruments in sensory panels to produce measures in some instances leads to the false conclusion that they may replace customer measurements. By placing sensory analysis on a footing equal to the technical dimensions the proper perspective is retained. Rather than substituting market research sensory panels complement it by facilitating the process of translating customer requirements into new food products.

To make the House of Quality operational it is necessary to develop the various matrices of transformation and requirements somewhat more. This task may be seen as having two aspects, a structural and a quantitative part. By structural we mean a numeration of the important concepts involved and their mutual relationship. By quantitative we mean a measure of the strength of these associations. Methods have been developed to cover both these areas. The purpose is to increase the understanding of the processes, thereby making the crossBfunctional communication easier and more effective. In this study we focus on the left block in the House of Quality and apply LISREL in order to structure customer needs. It is the natural starting point for the whole development process. An identification of the structure in the needs leads to a more thorough understanding of the market and hence improved possibilities to transform customer needs to technical specifications.

It is widely recommended to divide customer needs hierarchically into primary, secondary and tertiary needs. The primary needs may be regarded as the strategic needs. The secondary needs are known as the tactical needs and describe how the product development team satisfies the strategic needs. Finally the tertiary needs correspond to the tactical needs and provide detail so that R&D can develop solutions to satisfy the secondary needs. As a consequence of the hierarchical structure no interaction between needs at a given level are normally assumed. In our model we start with this assumption, but the LISREL estimation procedure allows for possible interactions.

As mentioned earlier we limit ourselves to perceived sensory quality and see it as a function of taste and first-hand impression. To some extent it resembles Steenkamp and van Trijp (1989) who operate with the constructs quality expectation and quality experience. Quality expectation is the impression formed by the consumer when the product is bought, while quality experience relates to actual consumption of the product. Although Steenkamp and van Trijp (1989) assume a possible dependency between expectations and experience this is not in accordance with the use of the constructs in service quality and related areas where quality of a given service or good is defined as the difference between expectations and experience. This is the main reason that we have chosen the constructs firstBhand impression and taste which do not demand an assumption of independence.

In Figure 2 we illustrate the structure of our preliminary model. The product category is butter cookies. The firstBhand impression of a butter cookie is assumed to be a function of crispness and colour. The reason for treating colour as part of the firstBhand impression should be obvious. Generally colour signals something about the ingredients and the baking process. As an example a yellow colour may leave an impression of butter. Crispness as a first impression may be more difficult to accept. It may by justified by the fact that crispness is sensed before the cookie is tasted. Hence, the sequental process describing consumpBtion of a cookie is our argument for including it as part of firstBhand impression. The other part of the secondary needs is taste. It is assumed to be a function of the tertiary needs butter, salt and sweetness.

INTRODUCTION TO LISREL

Since researchers in sensory analysis cannot be expected to be familiar with the LISREL modelling framework, this section contains a brief introduction to the area. For more extensive expositions, see the references. LISREL is an acronym for LInear Structural RELations. Strictly speaking, it is the name of a computer program that appeared in 1972. Its purpose is estimation and testing of covariance structure models, but the program's name and the modelling approach has become synonymous.

The importance of the approach is derived from several sources. Firstly, epistomologically it represents the very important distinction between theory construction and operations of measurement. Theories contain hypotheses on relations between theoretical constructs, while empirical testing involves translating these constructs into (imperfect) measures and relationships.

Secondly, it bridges the gap between two scientific disciplines, econometrics and psychometrics. Econometrics deals with modelling economic theory by way of systems of equations, reflecting 'cause and effects'Brelations. Psychometrics deals with measuring unobservable or latent variables as they are reflected in observed or manifest variables. Combining the two traditions yields a potential very powerful approach that allows the researcher to model structural or causal relations between latent and/or manifest variables. Hence, wellBknown statistical models like regression analysis, canonical correlation, factor analysis, path models, and others are all contained in the general framework of LISREL as special cases.

HIERARCHICAL PRESENTATION OF CONSUMER NEEDS FOR A BUTTER COOKIE

Thirdly, the approach has been made available as a practical tool to the researcher in a series of versions of the computer program LISREL, now in its eighth enhancement on the PC. With the recent introduction of the command language SIMPLIS and graphical capabilities (J÷reskog & S÷rbom (1993)), it has become userBfriendly and more interpretable in its output.

The factor analytic model

The factor analytic model assumes that one or more latent variables, corresponding to the constructs of interest, generate the observed variables that are measured with errors. In its general form, the relationships between the manifest and latent variables are given by the matrix equation

x = L_{x}x + d (1)

where x is a vector of observed variables, L_{x} a matrix of factor loadings, x is a vector of common factor, and d is a vector of errors in measurement. Statistically, the task is to explain the interrelationships among the observed variables, in terms of relationships among the observed and the latent variables defined by the equation (1).

The structural equation model

This model is widely used in the social and behavioural sciences. A simple case is the regression of a single measurable variable y on a set of measurables variables, x=(x_{1},x_{2},...,x_{k}). In matrix notation,

y = b'x + e (2)

where y is a dependent variable, x is a vector of independent variables related to the dependent variable by the slope coefficient vector b, and e is an error in equation, denoting measurements error as well as omitted independent variables.

In (2) it is usually assumed that the xBvector is measured without error. If this is not the case, these 'errors in variables' must be taken into account in order to avoid biased conclusions. Allowing for errors in xBvariables as well as in y leads to the following specification, using standard notation in LISREL:

h = Gx + z (3)

where h is a vector of dependent latent variables, x denotes the vector of latent independent variables related to h by the coefficient matrix G, and z are errors in equations.

More general and realistic models allows for possible reciprocal causation between dependent variables, i.e. one or more elements of h appears as explanatory variables on the right hand side of (3). This leads to the formulation:

h = Bh + Gx + z (4)

where B is the matrix of coefficients relating the dependent variables to each other.

Since we have assumed in (4) that dependent as well as independent variables are measured with errors, the structural equation model has to be complemented with two models of measurement, one for the hBvariables, and one for the xBvariables, cf. (2):

x = L_{x}x + d (1) repeated

y = L_{y}h + e (5)

The general LISREL model

Equations (1), (4) and (5) form the basis of the covariance structure model LISREL. However, in order to be able to test the model and confront it with empirical data, additional statistical assumptions are needed. These assumptions involve the levels of measurement of the observed variables and the distributional properties of error terms in the structural and measurement models. In this regard the most fundamental assumption in structural equations models is that the error term in each relationship is uncorrelated with all the independent constructs.

DESCRIPTIONG OF COOKIES INCLUDED IN THE ANALYSIS

In effect, the equations (1), (4), and (5) along with the statistical assumptions impose restrictions on the covariance matrix S between the observed variables x and y. The statistical inference involves (1) estimation of the parameters q, unspecified by the model, and (2) testing, whether the model fits the data.

A number (seven) of methods for estimation is available in LISREL8, e.g. instrumental variables, twoBstage least squares, unweighted least squares, and maximum likelihood. In general terms, the estimation step provides estimates of q by minimizing some 'distance' function F=F(S,S(q)). The methods differ with respect to computations and to the particular function they try to minimize, but under general assumptions they all provide consistent estimates of the parameters.

Testing the model involves comparing two matrices: S(q), the covariance matrix restricted by the model, and S, the unrestricted covariance matrix, computed from the data. When the estimation step has provided estimates of q by minimizing F(S,S(q)), testing means to evaluate whether the obtained minimum distance between model and data are in fact tolerable, indicating an acceptable fit, or the model has to be rejected. For this purpose, one may use c=N*F. If the model holds and the sample size N is sufficiently large, c will be approximately distributed as c^{2} with degrees of freedom determined as the difference between the number of independent data points in S and the number of independent estimated parameters.

In practice, it is more useful to consider c as a measure of fit rather than a test statistic. The model may not be expected to hold exactly in the population, and the researcher is merely looking for some guidance in model assessment and modification. In addition, since c is proportional to sample size, the researcher may be unable to fit any models to his data merely because of large samples. One way of going about this problem is to evaluate the fit based on moderate sample sizes like N=100 or 200.

THE DATA

In Denmark a wide range of cookies exist. The main ingredients are usually flour, sugar, fat and eggs and some kind of flavour such as vanilla, chocolate, nuts etc. In order to examine consumer use, habits and preferences for butter cookies a market analysis was performed. Ten products were selected for the analysis and a description is given in the table below.

The investigation was carried out as a hall test with an incomplete block design. Each of the 210 consumers was asked to evaluate their unique combination of 4 out of 10 cookies on 5 sensory attributes and to indicate overall preference and to the degree in which the cookie tasted and looked like a homeBmade cookie. The four cookies were presented and served one at a time in random order. A dish with a sample of the product was presented and colour of the cookie was assesessed by the consumer. Hence the consumer was asked to select one of the cookies from the dish and assess the intensity of crispness, saltness, sugar and butter taste. Finally, the consumer evaluated homeBmadeness. After the four cookies had been assessed on these 6 attributes all the four products were presented and the consumer was asked about overall preference for each of the products.

In the model formulation we have assumed a sequential experience with the butter cookie. In the empirical part crispness is evaluated together with butter, salt and sweetness and maybe the joint evaluation give rise to reversible effects. Hence we model equilibrium and not the process. The sensory attributes were measured on a 9Bpoint relativeBtoBideal scale. With this scale "5" represents ideal , "9" indicates much more than ideal and "1" indicating much less than ideal. Homemade and preference are measured on a monotone 9Bpoint scale and 9 indicating an extremely good cookie. The measurements of each of the sensory attributes were transformed to deviations from ideal by the following formula: (Dev = 16 B (Observed value B 5)^{2}). The transformation treats positive and negative deviations from ideal equivalently and ensures high values are closer to ideal than low values. Subsequently the polychoric correlations given in table 2 between the 7 variables were calculated and applied as input to the LISRELBanalysis.

ESTIMATION

A LISREL model corresponding to the structure in Figure 1 is formulated and estimated using weighted least squares as recommended by J÷reskog and S÷rbom (1993), because of the ordinal measurement level. Perceived sensory quality is operationalized by the variables preference and homemade. In the first place the model was formulated according to the hierarchy in Figure 2, but a significant improvement in fit was obtained changing the structure according to Figure 3 below. There has been an extensive discussion of the insufficency of the cBstatistic due to its dependence upon sample size (Browne and Cudeck (1993)). We have decided to accept th emodel, if the model is acceptable when the number of observations are set to 100. The c-statistic is equal to 5.02 with a number of degrees of freedom equal to 12 and hence acceptable at any reasonable level of significance.

CORRELATIONS BETWEEN VARIABLES INCLUDED IN THE ANALYSIS (n=840)

The numbers in figure 3 are the estimated coefficients in the underlying measurement models and the structural model. These coefficients may be interpreted as the importances of each of the consumer needs, derived as a part of the LISREL formulation. As regards the secondary needs the most striking result is the high importance of first-hand impression. It can be seen that 84% of the variation in taste is explained by first-hand impression and 74% of the variation in perceived sensory quality is explained by the sequential influence of first-hand impression and taste. This result is in good accordance with the slogans "You never get a second chance to make a first impression", "First impressions last". It also reflects the fact that reintroduction of a product is a very difficult task with a low success rate. It may be of considerable interest to evaluate if this particular structure can be identified in other product groups. For the time being the authors work with a similar dataset on peas but of course with other measurement variables. It seems as if the same structure may be found in this very different product group.

As regards the tertiary needs in our butter cookie example crispness is the most important determinant for first-hand impression and butter is not surprisingly the most important taste determinant.

A natural extension of our model is to include the sensory profiles of the butter cookies obtained from a panel in order to give R&D a more comprehensive guide to develop butter cookies with the secondary needs demanded by the consumer. Finally, the tertiary needs and the sensory profiles have to be connected to the technical specifications of the butter cookies covering such variables as sugar percent, butter percent and fat percent. It may be done in a LISREL setting by adding more levels to the structure in figure 3.

CONCLUSION

The QFD model has been used to systematize consumer needs for a butter cookie into the three hierarchical levels. The application of LISREL results in estimated importances for each of the consumer needs. It has been shown that first-hand impression is a very important direct determinant for taste experience and hence indirectly for perceived sensory quality of the butter cookie. It may be of considerable interest to evaluate if and under which conditions this particular structure can be identified in other product groups. For the time being the authors work on a similar dataset on peas but of course with other measurement variables. Preliminary analysis on the very different group point to a structure -similar to the one found in that very different product group. Both datasets include a sensory evaluation of the products by an expert panel and some technical data. Hence it is possible later to widen the analysis as described in the former section.

REFERENCES

Bech, A., Engelund, E., Juhl, H.J., Kristensen, K. & Poulsen, C.S. (1994). QFooD - Optimal design of food products,

MAPP Working Paper no. 19, Arhus, Denmark.

Browne, M.V. & Cudeck, R. (1993). Alternative ways of assessing model fit. In Bollen, K.A. & Long, J.S:(eds.) *"Testing Structural Equation Models", *SAGE Publications.

Hauser, J. & Clausing, D. (1988). The House of Quality. *Harvard Business Review, 66,66-73.*

Joreskog, K. & Sorbom, D. (1993). *LISREL 8: *Structural Equation Modeling with the SIMPLIS Command Language. Scientific Software International, Chicago, USA.

Molnar, P.J. (1993). Quality assurance and quality control. In R. Macrae, R.K.Robinson & M.J. Sadler (eds.) *Encyclopedia *of *Food Science Food Technology and Nutrition, pp *38463847. Academic Press, London.

Martens, M. (1984). Quality and Quality Evaluation. *Acta Horticulturae *no. 163, 15-30.

Senge, P. (1992). Building Learning Organizations. *Journal For Quality and Participation, *March.

Shepherd, R. (1989). Factors affecting food preferences and choice. In R. Shepherd (ed.) *Handbook *of *the Psychophysiol*ogy of *Human Eating, *John Wiley & Sons Ltd. C .

Steenkamp, J.B.E-M. & VanTrijp, H.C.M (1989). Quality Guidance: A consumer-based approach for product quality improvement. In Avlonitis, G.J, Papavasilliou, N.K. & Kouremenos, A.G. *Marketing Thought and Practice in the *1990's, pp 1191-1217.

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

##### Authors

Hans Jorn Juhl, Aarhus School of Business

Kai Kristensen, Aarhus School of Business

Anne C. Bech, Aarhus School of Business

##### Volume

E - European Advances in Consumer Research Volume 2 | 1995

##### Share Proceeding

## Featured papers

See More#### Featured

### Material Gifts as Relationship Mnemonics: Why More Material Gifts Are Given Than Wanted?

Adelle Xue Yang, National University of Singapore, Singapore

Minjung Koo, Sungkyunkwan University

JAEWON HWANG, Sejong University

#### Featured

### Proximity Bias: Interactive Effects of Event Valence and Event Nearness on Probability Estimates

Jennifer Seok Hwa Hong, New York University, USA

Chiara Longoni, Boston University, USA

Vicki G. Morwitz, New York University, USA

#### Featured

### Cues to Sincerity: How People Assess and Convey Sincerity in Language

Alixandra Barasch, New York University, USA

Juliana Schroeder, University of California Berkeley, USA

Jonathan Zev Berman, London Business School, UK

Deborah Small, University of Pennsylvania, USA