Chapter 19

Multivariate Analysis III: Factor Analysis, Cluster Analysis, Multidimensional Scaling and Correspondence Analysis

LEARNING OBJECTIVES

Upon completion of this chapter, you will be able to:

• Understand the concept and application of factor analysis
• Interpret the factor analysis output obtained from the statistical software
• Use Minitab and SPSS to perform factor analysis
• Understand the concept and application of cluster analysis
• Interpret the cluster analysis output obtained from the statistical software
• Use SPSS to perform cluster analysis
• Understand the concept and application of multidimensional scaling
• Interpret the multidimensional scaling output obtained from the statistical software
• Use SPSS to perform multidimensional scaling
• Have an introductory idea about correspondence analysis

19.1 FACTOR ANALYSIS

Factor analysis is a widely used multivariate technique in the field of marketing research. It has been observed that many new researchers are using this technique without understanding the prerequisites for using it. The following section focuses on the use and application of factor analysis in the field of marketing research.

19.1.1 Introduction

The techniques described in this chapter, namely, factor analysis, cluster analysis, multidimensional scaling, and correspondence analysis are often referred as ‘interdependence analysis.’ As different from the regression and discriminant analysis, in the factor analysis, there is no concept of predicting the dependent variables through the independent variables. The factor analysis is most widely applied to multivariate technique of research, especially in the field of social and behavioural sciences. The techniques described in this chapter (factor analysis, cluster analysis, and multidimensional scaling) mainly examine the systematic interdependence among the set of observed variables, and the researcher is mainly focused on determining the base of commonality among these variables.

19.1.2 Basic Concept of Using the Factor Analysis

The factor analysis is a proven analytical technique that has been studied extensively by statisticians, mathematicians, and research methodologists (Raven 1990). It is a very useful technique of data reduction and summarization.

The main focus of the factor analysis is to summarize the information contained in a large number of variables into a few small number of factors. Generally for conducting any kind of research, a researcher has to collect information for a large number of variables. These variables are generally gathered either from the literature or from the experience of a researcher or an executive provides it. Most of these variables are correlated and can be reduced into fewer factors.

The factor analysis allows a researcher to group the variables into a number of factors based on the degree of correlation among the variables. For example, if a company wishes to examine the degree of satisfaction the consumers are deriving from a launch of a new product, then it can collect variables from different sources such as literature, experience of the company executives, and so on. In fact, the list of these variables may be sufficiently large because this is the discretion of a researcher to include a large number of variables in the research. These variables may then be factor analysed to identify the underlying construct in the data. The name of the new factor (group of correlated variables) or factors can be subjectively defined by the researcher, which can be used for further multivariate analysis (may be regression or discriminant analysis). Thus, the factor analysis is a statistical technique used to determine the prescribed number of uncorrelated factors, where each factor is obtained from a list of correlated variables. The factor analysis can be utilized to examine the underlying patterns or relationships for a large number of variables and to determine whether the information can be condensed or summarized into a smaller set of factors or components (Hair et al. 2009).

19.1.3 Factor Analysis Model

It has already been discussed that factor analysis is a statistical technique used to transform the original correlated variables into a new set of uncorrelated variables. These different groups of uncorrelated variables are referred as factors. Each factor is a linear combination of correlated original variables. Mathematically, it can be expressed as

where F_i is the estimate of the ith factor, a_i the weights or factor score coefficients, and n the number of variables.

It is important to note that for forming a factor F₁, each variable X_i, which is correlated with other variables, shares some variance with these other variables. This is referred as communality in the factor analysis.

The variable X_i may then be correlated with another group of variables to form another factor F₂. It may be possible that another group of variables may not be significantly correlated with the variables, which formed Factor F₁. This situation is explained in Figure 19.1.

FIGURE 19.1 Joining of various variables to form different factors

From Figure 19.1, it can be seen that Variable X₁ is highly correlated with Variables X₂, X₃, and X₄ to form Factor F₁. It also clearly shows that Variable X₁ is correlated with Variables X₅ and X₆ to form Factor F₂. However, Variables X₅ and X₆ may not be correlated with Variables X₂, X₃, and X₄. The covariance among the variables can be explained by a small number of common factors and the unique factor for each variable. Such Variable X_i may then be defined as

where X_i is the ith standardized variable, A_ij the standardized multiple regression coefficient of Variable X_i on common Factor j, F the common factor, V_i the standardized regression coefficient of Variable X_i on U_i, U_i the unique factor in X_i, and m the number of common factors.

19.1.4 Some Basic Terms Used in the Factor Analysis

The following is the list of some basic terms frequently used in the factor analysis:

Correlation matrix: It is a simple correlation matrix of all the pairs of variables included in the factor analysis. It shows a simple correlation (r) between all the possible pairs of variables included in the analysis. In correlation matrix, the diagonal element is always equal to one, which indicates the correlation of any variable with the same variable.

Kaiser-Meyer-Olkin (KMO) measure of sampling adequacy: This statistic shows the proportion of variance, for variables included in the study is the common variance. In other words, this is the common variance, attributed to the underlying factors. A high value of this statistic (from 0.5 to 1) indicates the appropriateness of the factor analysis for the data in hand, whereas a low value of statistic (below 0.5) indicates the inappropriateness of the factor analysis.

Bartlett’s test of sphericity: This statistic tests the hypothesis whether the population correlation matrix is an identity matrix. This is important to note that with an identity matrix, the factor analysis is meaningless. Using significance level, the degree of relationship among the variables can be identified. A value less than 0.05 indicates that the data in hand do not produce an identity matrix. This means that there exists a significant relationship among the variables, taken for the factor analysis.

Communality: It indicates the amount of variance a variable shares with all other variables taken for the study.

Eigenvalue: It indicates the proportion of variance explained by each factor.

Percentage of variance: It gives the percentage of variance that can be attributed to each specific factor relative to the total variance in all the factors.

Scree plot: It is a plot of eigenvalues and component (factor) number according to the order of extraction. This plot is used to determine the optimal number of factors to be retained in the final solution.

Factor loadings: Also referred as factor–variable correlation. These are a simple correlation between the variables.

Factor matrix: Factor matrix table contains the factor loadings for each variable taken for the study on unrotated factors.

Factor score: It represents a subject’s combined response to various variables representing the factor.

Factor loading plot: It is a plot of original variables, which uses factor loadings as coordinates.

19.1.5 Process of Conducting the Factor Analysis

The factor analysis can be performed using the seven steps as shown in Figure 19.2: problem formulation, construction and examination of correlation matrix, selection of an appropriate factoring technique, a decision regarding the number of factors to be retained in the final solution, factor rotation for enhancing the interpretability of the solution, substantive interpretation, and check the model fit.

Figure 19.2 Seven steps involved in conducting the factor analysis

The following section describes a detailed discussion of these steps with the help of an example that is included in Step 1 as problem formulation.

19.1.5.1 Problem Formulation

The first step in conducting the factor analysis is to formulate the problem of the factor analysis. As discussed earlier, the main focus of the factor analysis is to reduce data. For this purpose, a researcher has to select a list of variables that will be converted into a new set of factors based on the common essence present in each of the variables. For selecting variables, a researcher can take the help of literature, past research, or use the experience of other researchers or executives. It is important to note that the variables should be measurable on an interval scale or a ratio scale. Another important aspect of the factor analysis is to determine the sample size, which will be used for the factor analysis.

As a thumb rule, the sample size should be four or five times of the variable included in the factor analysis.

For understanding the factor analysis, let us consider an example of a garment company that wishes to assess the changing attitude of its customers towards a well-established product, in light of many competitors presence in the market. The company has taken a list of 25 loyal customers and administered a questionnaire to them. The questionnaire consists of seven statements, which were measured on a 9-point rating scale with 1 as strongly disagree and 9 strongly agree. The description of the seven statements used in the survey is given as follows:

X₁: Price is a very important factor in purchasing.

X₂: For marginal difference in price, quality cannot be compromised.

X₃: Quality is OK, but competitor’s price of the same product cannot be ignored.

X₄: Quality products are having a high degree of durability.

X₅: With limited income, one can afford to spend only small portion for cloth purchase.

X₆: In the present world of materialism and commercialization, people are evaluated on the basis of good appearance.

X₇: By paying more if we can get good quality, why not to go for it.

Table 19.2 gives the rating scores obtained by different consumers for the seven statements described earlier.

TABLE 19.2 Rating scores obtained by different consumers for the seven statements to determine the changing consumer attitude

Figures 19.3(a)–19.3(m) shows the SPSS factor analysis output.

Figure 19.3 (a) Descriptive statistics, (b) Correlation matrix, (c) KMO and Bartlett’s test, (d) Communalities, (e) Total variance explained

Figure 19.3 (f ) Scree plot, (g) Component matrix, (h) Reproduced correlations

Figure 19.3 (i) Rotated component matrix, (j) Component transformation matrix (k) Component plot in rotated space, (l) Component score coefficient matrix, (m) Component score covariance matrix

19.1.5.2 Construction and Examination of Correlation Matrix

The following discussion is based on the SPSS output, which is given in the form of Figures 19.3(a)–19.3(m). Figure 19.3(a) is a simple table that shows the descriptive statistics for the variables taken into study. In this figure, the second column shows the mean value for each item for 25 customers, the third column the degree of variability in scores for each item, and the fourth column the number of observations (sample size).

Figure 19.3(b) is the SPSS-produced correlation matrix for the descriptor variables. This is the initial stage of the factor analysis, which gives some initial clues about the patterns of the factor analysis. It is important to note that for the appropriateness of the factor analysis, the variables must be correlated.

If there is no correlation among the variables or if the degree of correlation among the variables is very low, then the appropriateness of the factor analysis will be under serious doubt. In the factor analysis, a researcher expects that some of the variables are highly correlated with each other to form a factor. Figure 19.3(b) shows that a few variables are relatively highly correlated with each other to form factors.

Figure 19.3(c) shows two important statistics: the KMO measure of sampling adequacy and the Bartlett’s test of sphericity for judging the appropriateness of a factor model. KMO statistic compares the magnitude of the observed correlation coefficient with the magnitude of the partial correlation coefficient. As discussed earlier, a high value of this statistic (from 0.5 to 1) indicates the appropriateness of the factor analysis.

Kaiser has presented the range as follows: statistic >0.9 is marvellous, >0.8 meritorious, >0.7 middling, >0.6 mediocre, >0.5 miserable, and <KMO statistic is computed as 0.741, which indicates the value in the acceptance region of the factor analysis model.

Bartlett’s

test of sphericity tests the hypothesis whether the population correlation matrix is an identity matrix. The existence of the identity matrix puts the correctness of the factor analysis under suspicion. Figure 19.3(c) shows that chi square statistic is 135.592 with 21 degrees of freedom. This value is significant at 0.01 level. Both the results, that is, the KMO statistic and Bartlett’s test of sphericity, indicate an appropriate factor analysis model.

19.1.5.3 Selection of an Appropriate Factoring Technique

After deciding the appropriateness of the factor analysis model, an appropriate technique for analysing the data is determined. Most of the statistical software available these days present a variety of methods to analyse the data. The principal component method is the most commonly used method of data analysis in the factor analysis model. When the objective of the factor analysis is to summarize the information in a larger set of variables into fewer factors, the principal component analysis is used (Aaker et al. 2000).

The main focus of the principal component method is to transform a set of interrelated variables into a set of uncorrelated linear combinations of these variables. This method is applied when the primary focus of the factor analysis is to determine the minimum number of factors that attributes maximum variance in the data. The obtained factors are often referred as the principal components. Another important method is the principal axis factoring, which will be discussed in the following section. Apart from the principal component method, a variety of other methods are also available. Description of all these methods is beyond the scope of this book.

Figure 19.3(d) shows the initial and extracted communalities.

The communalities describe the amount of variance a variable shares with all other variables taken into study. From the figure, it can be seen that the initial communality value is equal to 1 (as can be seen, the unities are inserted in the diagonal of correlation matrix) for all the variables taken into the factor analysis model. The SPSS, by default, assigns a communality value of 1 to all the variables. The extracted communalities as shown in the third column of Figure 19.3(d) is the estimate of variance in each variable, which can be attributed to factors in the factor solution. Relatively small value of the communality suggests that the concerned variable is a misfit for the factor solution and can (should) be dropped out from the factor analysis.

Figure 19.3(e) presents the initial eigenvalues (total, % of variance, and cumulative %), extraction sums of squared loadings (total, % of variance, and cumulative %), and the rotation sums of squared loadings (total, % of variance, and cumulative %). In this figure, the “Total” column gives the amount of variance in the variable attributed to the concerned component or factor.

The“% of variance” column indicates the percentage of variance accounted for by each specific factor or component. It is important to note that the total variance accounted for by all the seven factors is equal to 7. This is equivalent to the number of variables. The variance attributed to factor 1 is 3.592/7 = 51.31%. The total variance attributed to factor 2 is 2.200/7 = 31.43%. Similarly, the total variance attributed to all the factors can be computed. The second part of the figure is the extraction sums of squared loadings that gives information related to the extracted factors or components. If a researcher has adopted the “principal component method” as the method of analysis in a factor model, then the values will remain the same, as shown under the heading “initial eigenvalues.” As can be seen from the third part of the figure, rotation sums of squared loadings, the variance accounted for by the rotated factors or components is different from those indicated in the second column of extraction sums of squared loadings. It is important to note that for rotation sums of squared loadings, the cumulative percentage for the set of components (factors) will always remain the same.

19.1.5.4 Decision Regarding the Number of Factors to be Retained in the Final Solution

It is possible to have a number of factors as the number of variables available in the factor analysis. If this is the case, the rationale of applying factor analysis is questionable because the primary objective of the factor analysis is to summarize the information contained in the various variables into a few factors. For this purpose, the basic question in the mind of a researcher while applying factor analysis is that how many factors should be abstracted from the factor analysis. Fabrigar et al. (1999) stated that determining how many factors to include in the model requires the researcher to balance the need for parsimony (i.e., a model with relatively few common factors) against the need for plausibility (i.e., a model with a sufficient number of common factors to adequately account for the correlations among the measured variables).

A number of approaches are available to decide the number of factors to be retained in the factor analysis solution. In this section, we will focus on three commonly used criteria for determining the number of factors: eigenvalue criteria, Scree plot criteria, and percentage of variance criteria.

Eigenvalue Criteria

An eigenvalue is the amount of variance in the variable taken for the study that is associated with a factor. According to the eigenvalue criteria, the factors having more than one eigenvalue are included in the model. A factor that has an eigenvalue of less than 1 is not better than a single variable because due to standardization each variable has a variance of 1.

Scree Plot Criteria

One of the most popular guides for determining how many factors should be retained in the factor analysis is the Scree test (Cattell 1966).

Scree plot is a plot of the eigenvalues and component (factor) number according to the order of extraction [as shown in Figure 19.3(f)]. The shape of the plot is used to determine the optimum number of factors to be retained in the final solution. The objective of the Scree plot is to visually isolate an elbow, which can be defined as the point where the eigenvalues form a liner descending trend (Bentler and Yuan 1998). For an appropriate factor analysis model, this plot looks like an intersection of two lines (Figure 19.4).

Figure 19.4 Scree plot shown as an intersection of two lines

Figure 19.5 clearly shows that the factors on the steep slope should be retained in the model and the factors on the shallow slope can be excluded from the model (as these factors contribute relatively little to the factor model).

Figure 19.5 Scree plot indicating the number of factors to be retained in the model

Percentage of Variance Criteria

This approach is based on the concept of cumulative percentage of variance. The number of factors should be included in the model for which cumulative percentage of variance reaches a satisfactory level. Now there is a question that what should be that satisfactory level.

However, the general recommendation is that the factors explaining 60%–70% of the variance should be retained in the model.

Considering Figures 19.3(e) and 19.3(f), all the three approaches suggest that the number of factors to be extracted should be only two. According to the eigenvalue approach, only two factors have eigenvalues more than 1. The Scree plot for the problem in hand also shows that only two factors on the steep slope should be retained in the model. The cumulative percentage of variance for the first two factors is 82.751%, which is well within the prescribed limits, which is why only two factors must be retained in the final factor analysis solution. Above are some criteria; however, the number of factors to be retained is a highly subjective matter. The scope of interpretation can be enhanced by rotating the factors and will be described in the following section.

Figure 19.3(g) shows the component matrix table that presents the factor loading for each variable on unrotated factors (components). As can be seen from the figure, each value under the heading component (1 or 2) represents the correlation between the concerned variable and the unrotated factors. The values given under Factor 1 represent the correlation between the concerned variable and Factor 1, and the values given under Factor 2 represent the correlation between the concerned variable and Factor 2. From this figure, it can be seen that Variables X₁, X₃, X₅, and X₇ are relatively highly correlated (with values 0.965, 0.939, 0.944, and −0.928) with Factor 1, and Variables X₂, X₄, and X₆ are relatively highly correlated with Factor 2 (with values 0.815, 0.830, and 0.919).

Figure 19.3(h) shows the reproduced correlation and residuals for the factor analysis solution. For an assumed appropriate factor analysis solution, the reproduced correlation shows the predicted pattern of relationship.

The residuals are the difference between the predicted and observed values. If a factor analysis solution is good enough, then most of the residual values are small.

19.1.5.5 Factor Rotation for Enhancing the Interpretability of the Solution

After selection of factors, the immediate step is to rotate the factors. The rotated simple structure solutions are often easy to interpret, whereas the originally unextracted (unrotated) factors are often difficult to interpret (Reise et al. 2000).

A rotation is required because the original factor model may be mathematically correct but may be difficult in terms of interpretation. If various factors have a high loading on the same variable, then interpretation will be extremely difficult. Rotation solves this kind of interpretation difficulty. The main objective of rotation is to produce a relatively simple structure in which there may be a high factor loading on one factor and a low factor loading on all other factors. Similar to correlation, factor loading varies between +1 and −1 and indicates the degree of relationship between a particular factor and the particular variable. It is interesting to note that rotation never affects the communalities and the total variance explained [from Figure 19.3(e), it can be seen that after rotation, the cumulative percentage of variance is not changed (82.751%)].

Figure 19.6 shows asterisks on the unrotated factor lines, and Figure 19.7 shows asterisks on the rotated factor lines. If we compare both the figures, the rationale of rotation can be easily explained.

Figure 19.6 Asterisks on the unrotated factor lines

Rotation enhances the interpretability in terms of association of factor loadings with the concerned factor. When compared with Figure 19.6, Figure 19.7 clearly explains the high factor loading on one factor and the low factor loading on all other factors.

Figure 19.7 Asterisks on the rotated factor lines

Figure 19.3(i) represents the rotated component matrix that is often referred as the ‘pattern matrix for oblique rotation.’ The columns in this figure represent the factor loading for each variable, for the concerned factor, after rotation. The figure clearly shows the interpretability importance of the rotation.

Earlier, the rotation was done manually by the researchers, whereas nowadays, a variety of statistical softwares present a variety of methods of rotation such as Varimax procedure, Quartimax procedure, Equamax procedure, Promax procedure, and so on. However, the different rotation procedures more or less reflect the same result about the data; by performing different procedures and later comparing the results of the different schemes, obtaining the ‘most interpretable’ solution will always be beneficial for researchers. The widely applied method of rotation is the ‘Varimax procedure.’ Although a number of rotation methods have been developed, varimax has been generally regarded as the best orthogonal rotation and is overwhelmingly the most widely used orthogonal rotation in psychological research (Fabriger et al. 1999). This section focuses on varimax procedure of rotation.

The rotation is often referred as ‘orthogonal rotation’ if the axes are maintained at right angles. Varimax procedure is an orthogonal rotation procedure. Orthogonal rotation generates the factors that are uncorrelated. On the other hand, in oblique rotation, axes are not rotated at right angle and the factors are correlated. If the factors are truly uncorrelated, orthogonal and oblique rotation produce nearly identical results (Costello and Osborne 2005).

The rotated factor matrix as shown in Figure 19.3(i) shows that the Variables X₁, X₃, X₅, and X₇ are having high loadings on Factor 1 and the Variables X₂, X₄, and X₆ are having high loadings on Factor 2.

19.1.5.6 Substantive Interpretation

At this stage, there is a point to identify the factors with concerned variables as the constituents of the factor. Figure 19.3(k), as the component plot in rotated space, provides this opportunity. Before that, it is important to interpret Figure 19.3(j). Figure 19.3(j) shows the component transformation matrix. The component transformation matrix is used to construct the rotated factor matrix from the unrotated factor matrix by using a simple formula (unrotated factor loading = rotated factor loadings). In the case in which off-diagonal elements are close to zero indicates a relatively smaller rotation and off-diagonal elements are large (greater than ±5) indicates a relatively larger rotation.

As described in the previous paragraph, Figure 19.3(k) is the component plot in rotated space. This plot is a pictorial representation of the factor loadings of the variables taken in the study on the first few factors. The variables near or on the axis show a high factor loading on the concerned factor. Figure 19.3(k) shows that the Variables X₁, X₃, X₅, and X₇ are highly loaded on Factor 1. It is interesting to see here that the Variable X₇ has got the high negative value, and hence, it is in the opposite direction of the three Variables X₁, X₃, and X₅. If a variable is not a part of any factor (neither close to horizontal axis nor close to vertical axis), then it should be treated as an undefined or a general factor. Similarly, the Variables X₂, X₄, and X₆ have high factor loadings on Factor 2.

Factor 1 consists of Variables X₁ (price is important), X₃ (competitor’s price cannot be ignored), X₅ (limited affordability), and X₇ (get good quality by paying more). Negative factor loading value of Variable X₇ on Factor 1 indicates that this segment of customers is more concerned about the price even at the cost for compromising quality. Hence, Factor 1 can be named as ‘Economy Seekers.’ Similarly, Factor 2 consists of Variables X₂ (quality cannot be compromised), X₄ (quality products are having a high degree of durability), and X₆ (societal evaluation of an individual is based on quality products). Thus, Factor 2 can be named as ‘Quality seekers.’ After applying the factor analysis, the original seven variables are categorized into two factors economy seekers and quality seekers.

Figure 19.3(l) shows the component score coefficient matrix. For each subject, the factor score is calculated by multiplying the values of variable (can be obtained from Table 19.2) by factor score coefficients. For example, for Subject 1, factor score can be computed as follows:

For Factor 1:

(0.269) × (8) + (−0.030) × (4) + (0.261) × (7) + (−0.005) × (5) + (0.263)

× (7) + (0.033) × (4) + (−0.258) × (2) = 4.211.

For Factor 2:

(−0.015) × (8) + (0.370) × (4) + (0.003) × (7) + (0.377) × (5) + (−0.010)

× (7) × (0.418) × (4) + (−0.012) × (2) = 4.844.

Similarly, using the factor score coefficient, factor scores for other subjects can also be computed. Instead of all the original variables, the factor scores can be used in the subsequent multivariate analysis.

Sometimes instead of computing a factor score, researchers select the substitute variable commonly known as surrogate variable by selecting some of the original variables for further multivariate analysis. It is important to note that a researcher can also select the variables on a discretionary basis. This means that on the basis of experience or literature or any other logical basis, if a researcher believes that the variable with a high factor loading is less important than the variable with a low factor loading, then he or she can select the relatively important variable with a low factor loading as the surrogate variable.

19.1.5.7 Check the Model Fit

The last step in the factor analysis is to determine the fitness of the factor analysis model. In factor analysis, the factors are generated on the basis of observed correlation between the variables. The degree of correlation between the variables can be reproduced as shown in Figure 19.3(h).

For an appropriate factor analysis solution, the difference between the reproduced and observed correlation should be small (less than 0.05). As can also be seen from Figure 19.3(h), only seven residuals are greater than 0.05, which indicates an appropriate factor analysis model.

19.2 CLUSTER ANALYSIS

This section focuses on a popular multivariate technique known as cluster analysis. It deals with all important dimensions related to the application of cluster analysis. Use of SPSS output for explaining the concept will make its use very convenient for a researcher.

19.2.1 Introduction

In every field of the study, researchers or scientists or marketing executives will like to group the objects on the basis of some similarity.

Specifically, in the field of marketing, mangers try to identify a similar group of customers, so that marketing strategies can be chalked out for the concerned similar group of customers. These customers can be grouped or clustered on the basis of a variety of common features such as the benefits the consumers seek from the product, lifestyle of the consumers with special reference to its impact on their purchase behaviour, and so on. The main objective of the marketing executives is to find out a similar group of customers to develop products or services according to their specific need. Thus, the researchers or executives face a common problem in terms of availability of some scientific technique to group the objects on the basis of some similarity. Cluster analysis provides a solution to this kind of problem. It has become a common tool for the marketing researcher (Punj and Stewart 1983). In some aspects, the factor analysis and cluster analysis are similar. In both, a researcher tries to reduce a larger number of variables (in factor analysis) or cases (in cluster analysis) into a smaller number of factors (in factor analysis) or clusters (in cluster analysis) on the basis of some commonality that within-group members share with each other. However, the statistical procedure to do this and the interpretations are different in these two methods.

19.2.2 Basic Concept of Using the Cluster Analysis

Cluster analysis is a technique of grouping individuals or objects or cases into relatively homogeneous (similar) groups that are often referred as clusters. The subjects grouped within the cluster are similar to each other, and there is a dissimilarity between the clusters. In other words, we can say that the individuals or objects or cases when correlated with each other form a cluster.

The main focus of the cluster analysis is to determine the number of mutually exclusive and collectively exhaustive clusters in the population, on the basis of similarity of profiles among the subjects. In the field of marketing, cluster analysis is mainly used in market segmentation, understanding purchase behaviour of the consumers, test marketing, and so on. A typical use of cluster analysis is to provide market segmentation by identifying subjects or individuals who have similar needs, lifestyles, or responses to marketing strategies (Zikmund 2007). For example, in the process of launching a new product, the CEO of a company wants to group cities of the country on the basis of age group and education of the consumers. If a company has collected data from 15 customers of various cities, then a simple way to group the subjects is to plot the result on two variables, that is, age and education of the consumers. Figures 19.8(a) and 19.8(b) present two different situations: before clustering and after clustering. In most of the cases, the researchers or marketing executives find situations as explained in Figure 19.8(a). This is a situation when there is no clustering and the subjects are scattered with respect to the two variables: age and education. Figure 19.8(b) shows after clustering situation in which subjects join to form different clusters based on some similarity.

Figure 19.8 (a) Before clustering situation, (b) After clustering situation

19.2.3 Some Basic Terms Used in the Cluster Analysis

The following list presents some basic terms that are commonly used in the cluster analysis:

Agglomeration Schedule: It presents the information on how the subjects are clustered at each stage of the hierarchical cluster analysis.

Cluster Membership: It indicates the cluster to which each subject belongs according to the number of cluster requested by a researcher.

Icicle plot: It provides a graphical representation on how the subjects are joined at each step of the cluster analysis. This plot is interpreted from bottom to top.

Dendrogram: Also referred as tree diagram. It is a graphical representation of relative similarities between the subjects. It is interpreted from left to right in which the cluster distances are rescaled, so that the plot shows the range from 0 to 25.

19.2.4 Process of Conducting the Cluster Analysis

The process of conducting the cluster analysis involves six steps as shown in Figure 19.9: problem formulation, choice of an appropriate proximity or similarity measure, selection of an appropriate clustering technique, a decision regarding the number of clusters to be retained in the final solution, substantive interpretation, and check the model fit.

Figure 19.9 Six steps involved in conducting the cluster analysis

The following section describes a detailed discussion of these steps with the help of an example that is included in Step 1 as problem formulation.

19.2.4.1 Problem Formulation

The first and most important step in the cluster analysis is to formulate the problem properly in terms of properly defining the variables on which the clustering procedure is to be performed. For understanding the procedure of conducting the cluster analysis, let us formulate a problem. A famous motorbike company wants to launch a new model of motorbike with some additional features. Before launching the product to an entire country, the company has to test market it in select cities. The company is well aware that the majority of the target consumer group consists of urban youth (aged below 45). For understanding the future purchase behaviour of the consumers, the company administered a questionnaire to the potential future customers. The questionnaire consists of seven statements that were measured on a 9-point rating scale with 1 as strongly disagree and 9 as strongly agree. The description of the seven statements used in the survey is given as follows:

X₁: I am trying very hard to understand the loan schemes offered by different banks to purchase the product when it will be coming in the market.

X₂: I am still in confusion whether to sell my old bike or not.

X₃: My old bike is old fashioned; I want to get rid of it.

X₄: My company has promised that it will be releasing a much awaited new incentive scheme; purchase of a new model is based on this factor.

X₅: I will not wait for the product’s performance in the market, I know the company is reputed and will certainly launch a quality product.

X₆: Companies always claim high about its product, let the product come in the market only then I will be taking any decision about the purchase.

X₇: I want to purchase a new bike but my kids are growing up and are demanding to purchase a car instead of a bike, which I already have.

Table 19.3 gives the rating scores obtained by different potential consumers for the seven statements, as described earlier, to determine the future purchase behaviour.

TABLE 19.3 Rating scores obtained by different consumers for the seven statements to determine the future purchase behaviour

For explaining the procedure of conducting cluster analysis, only 25 subjects are selected, but in practice, the cluster analysis is performed on more than or equal to 100 subjects.

19.2.4.2 Choice of an Appropriate Proximity or Similarity Measure

It has already been discussed in the beginning of the chapter that in the cluster analysis, the subjects are grouped on the basis of similarity or commonality. Thus, there is a need to have a measurement of similarity method by which similarity between the subjects can be identified, to group similar subjects in a cluster.

The method that is widely used for measuring the similarity is the Euclidean distance or its square. The Euclidean distance is the square root of the sum of squared differences between the values for each variable. The squared Euclidean distance is the sum of squared differences between the values for each variable. This distance measure is used for the interval data. Varieties of other measures are also available, but this section will mainly focus on the squared Euclidean distance method. The Euclidean distance measure also has one disadvantage in terms of varying result with varying units. This disadvantage can be overcome by expressing all the variables in a standardized form with a mean value of zero and standard deviation one.

19.2.4.3 Selection of an Appropriate Clustering Technique

There are two approaches of clustering: hierarchical clustering approach and non-hierarchical clustering approach. Hierarchical clustering starts with all the subjects in one cluster and then dividing and subdividing them till all the subjects occupy their own single-subject cluster. As different from hierarchical clustering, non-hierarchical clustering allows subjects to leave one cluster and join another in the cluster forming process if by doing so the overall clustering criterion will be improved. The various clustering procedures are shown in Figure 19.10.

Figure 19.10 Different methods of clustering

In the hierarchical clustering method, various approaches are available to cluster the subjects in different clusters. It can mainly be divided into two categories: agglomerative and divisive. In agglomerative clustering, initially each subject occupies a separate cluster. In this process of clustering, clusters are created by grouping the subjects into bigger and bigger clusters until all the subjects join a single cluster. In divisive clustering, as clear from the name itself, initially all the subjects or members are grouped into a single cluster and then the clusters are divided until each subject occupies a single cluster.

As discussed, varieties of clustering methods are available, but in the field of marketing, the agglomerative methods are widely used. From Figure 19.10, we can see that agglomerative methods are categorized into linkage method, variance method, and centroid method. Linkage method is based on how the distance between the two clusters is defined. On the basis of the procedure of measuring the distance, it can be categorized into three categories: single linkage method, complete linkage method, and average linkage method.

The single linkage method also referred as the nearest neighbour approach is based on the principle of the shortest distance. This method measures the shortest distance between the two individual subjects and places them into the first cluster. Then on the basis of the next shortest distance, either the third subject joins the first two or a new subject cluster is constituted [Figure 19.11(a)]. This process continues until all the subjects become a member of one cluster. The complete linkage method also referred as the farthest neighbour approach is based on the principle of the longest distance. This method is similar to the single linkage method except the measurement pattern in which the distance between the two clusters is measured as the distance between their two furthest points [Figure 19.11(b)]. As the name indicates, in the average linkage method, the distance between the two clusters is defined in terms of the average of the distance between all pairs of the subjects, in which one subject of the pair is from each of the clusters [Figure 19.11(c)]. This method is preferred over the single linkage and complete linkage methods because it avoids taking two extreme members while taking the distance between the two clusters. In fact, this method considers all the members of the cluster rather than the two extreme members of the clusters.

Figure 19.11 (a) Single linkage method, (b) Complete linkage method, (c) Average linkage method

The second type of agglomerative method is the variance method. This method is mainly based on the concept of minimizing within the cluster variance. The most common and widely applied variance method is the Ward’s method. In Ward’s method, the distance is measured as the total sum of squared deviation from the mean of the concerned cluster to which the subject is allocated. In other words, for each subject, the squared Euclidean distance to cluster means is computed and then these computed distances are summed up for all the subjects. The formation of new clusters tends to increase the total sum of the squared deviations (error sum of squares). At each stage, two clusters are joined to produce the smallest increase in the overall sum of the squared within the cluster distances [Figure 19.12(a)]. The third type of agglomerative method is the centroid method. A cluster centroid is a vector containing one number for each variable, in which each number is the mean of the variable for the observations in that cluster. In this method, the distance between the two clusters is measured as the distance between their centroid. The two clusters are combined on the basis of the shortest distance between their centroids [Figure 19.12(b)].

Figure 19.12 (a) Ward’s method, (b) Centroid method

The second type of clustering method is the non-hierarchical clustering method that is also referred as k-means cluster analysis or iterative partitioning. As clear from Figure 19.10, this method can be classified into three categories: sequential threshold, parallel threshold, and optimizing partitioning.

In the sequential threshold non-hierarchical clustering method, first a cluster centre is selected and all the subjects within a pre-specified threshold value are grouped together. Next, a new cluster centre is selected, and the process is repeated for the unclustered subjects. It is important to note that once a subject is selected as a member of a cluster or as an associate of a cluster centre, it is not considered for clustering with the subsequent cluster centre or it is removed from further processing.

The mode of operation in the parallel threshold method is the same except one difference in terms of the selection of several cluster centres simultaneously and then the subjects within a threshold level are clubbed to the nearest cluster centre. Then, this threshold level can be adjusted to accommodate some more subjects to the concerned cluster. In the optimizing threshold method, the subjects once assigned to a cluster can later be reassigned to another cluster to optimize an overall criterion measure such as the average within the cluster distance.

The non-hierarchical clustering method suffers from two major drawbacks. First, one has to pre-specify the number of clusters, and second, the cluster centre selection is arbitrary. This method is recommended when the number of observations is large. It has been recommended to first obtain the initial clustering solution by applying the hierarchical methods (may be Ward’s). Next, the number of clusters or cluster centroids can be used as an input material to the optimizing partitioning method. Figures 19.13(a)–19.13(f) present the SPSS output for the problem already defined in the problem formulation stage. To understand the concept of the cluster analysis completely, it is important to understand the interpretation of Figures 19.13(a)–19.13(f).

Figure 19.13(a) is the simple statement of the number of cases, missing cases, and total number of cases. Figure 19.13(b) presents the Proximity matrix and is the matrix of proximity between subjects. The values in the table represent dissimilarities between each pair of items. The distance measure used in Figure 19.13(b) is the measure of dissimilarities (squared Euclidean distance). In the case of dissimilarities, the larger values indicate items that are very different. This figure gives some initial clues about clustering of subjects.

Figure 19.13(c) presents the agglomeration schedule. This figure gives the information of how subjects are being clustered at each stage of the cluster analysis. The cluster analysis starts with 24 clusters at Stage 1, where Case 11 and Case 17 are combined (being minimum squared Euclidean distance). The coefficient column in Figure 19.13(c) indicates the squared Euclidean distance between the two clusters (or subjects) joined at each stage. After “coefficient column” “stage cluster first appears column” is shown. When we started the analysis, single cases existed, so it is indicated by zero at the initial stage. The last column “Next stage” indicates the subsequent stage at which the case combined in this cluster will be combined to another case. In the first row of this Column ‘2’ can be seen. Its interpretation can be done in Stage 2, where Case 7 is clustered with Case 11. Remember that Case 11 has already been clustered with Case 17 in Stage 1. On the basis of the squared Euclidean distance, this process continues until all the cases are clustered in a single group.

Figure 19.13 (a) Case processing summary, (b) Proximity matrix

Figure 19.13 (c) Agglomeration schedule

Figure 19.13 (d) Cluster membership

Figure 19.13 (e) Vertical icicle plot, (f) Dendrogram

The process described in the earlier paragraph can be well explained with the help of three examples in terms of clustering of Case 11 and Case 17 at Stage 1, Case 5 and Case 15 at Stage 11, and Case 3 and Case 13 at Stage 14, and Case 1 and Case 5 at Stage 20.

Cases 11 and 17 are grouped at Stage 1 with minimum squared Euclidean distance. These cases are not clustered before (see two zeros in “stage cluster first appears” below Clusters 1 and 2). In the “Next stage” column, Stage 2 is mentioned. This shows that Case 11 that was grouped with Case 17 in the first stage is grouped again with Case 7, making it a three-case cluster (Case 11, Case 17, and Case 7). This situation is shown in Figure 19.14.

Figure 19.14 Joining of the cases or clusters at different stages

Cases 5 and 15 are grouped at Stage 11. In the “stage cluster first appears,” 3 and 0 are mentioned. It means that before clustering at this stage (Stage 11), cluster or Case 5 is first clubbed with Case 25 at Stage 3. Against Stage 11 in the “Next stage” column, Stage 15 is mentioned. This means at Stage 15, Case 5 will be grouped with Case 9 making it a three-case cluster (Case 5, Case 15, and Case 9). This situation is also shown in Figure 19.14.

At Stage 14, in the case of grouping of Case 3 and Case 13, it can be seen from Figure 19.14 that in the column “stages cluster first appears,” 9 and 5 are mentioned. It means that before clustering at this stage (Stage 14), cluster or Case 3 is first clubbed with Case 23 at Stage 9. Similarly, before clustering at this stage (Stage 14), cluster or Case 13 is first grouped with Case 21 at Stage 5. Thus, at Stage 14 there is a four-case cluster (Case 3, Case 13, Case 21, and Case 23). Against Stage 14 in the “Next stage” column Stage 17 is mentioned. This means at Stage 17, Case 3 will be grouped with Case 10 making it a five-case cluster. Interpretation of the other parts of the agglomeration schedule can be done on the same lines.

Figure 19.13(d) indicates ‘Cluster membership’ for each case. The SPSS has presented “two- to five-cluster” solutions as requested (is described in using the SPSS for Cluster analysis section). For example, if we take a three-cluster solution, Cases 1, 5, 6, 9, 12, 15, 19, 22, and 25 are the members of ‘Cluster 1.’ Cases 2, 4, 8, 14, 16, 18, 20, and 24 are the members of ‘Cluster 2.’ Cases 3, 7, 10, 11, 13, 17, 21, and 23 are the members of ‘Cluster 3.’ Similar interpretation can be done for other cluster solutions as well.

Figure 19.13(e) exhibits ‘Vertical icicle plot’ that displays some important information graphically. In fact, this plot generates the same information as generated by agglomeration schedule except that the values of distant measures are not shown in this plot. As can be seen from the first column of Figure 19.13(e), rows indicates “number of clusters.” Most importantly, this plot is interpreted from ‘bottom to top.’ In this problem, there are 25 subjects, and in the beginning, each subject is considered as individual cluster, so there are 25 initial clusters. At the first stage, two nearest subjects are combined (here these objects are Case 11 and Case 17) that results in 24 clusters. The bottom line of Figure 19.13(e) indicates that 24 clusters (as the plot is interpreted from “bottom to top”). As can also be seen from the agglomeration schedule, in the second stage two subjects Cases 7 and 11 are combined. Figure 19.13(e) shows that the column between 7 and 11 is of maximum length after the column between 11 and 17. In the next stage, Cases 5 and 25 are combined. See Figure 19.13(e), column between 5 and 25 is the third in length (from bottom to top) after “7 and 11” and “11 and 17”. In a similar manner, interpretation of other stages can be obtained.

Figure 19.13(f) is another graphical display as an output part of cluster analysis from the SPSS. This graphical plot is referred to as “dendrogram.” This tree diagram is a critical component of hierarchical clustering output (Lehmann, Gupta, and Steckel 1998).The dendrogram exhibits a relative similarity between subjects considered for cluster analysis. Dendrogram is interpreted from “left to right” in the form of “branches” that merge together as can be seen from Figure 19.13(f). On the upper part of the plot, one can see “Rescaled Distance Cluster Combine.” This indicates that cluster distances are rescaled to get the range of the output from “0 to 25” with 0 representing no distance and 25 representing the highest distance. Cases or clusters that are joined by nearest vertical line (from the left) are very similar, and cases or clusters that are joined by relative distant vertical line (from the left) are very dissimilar (Figure 19.15).

Figure 19.15 Joining of cases or clusters in a rescaled cluster distance

19.2.4.4 A Decision Regarding Number of Clusters to be Retained in the Final Solution

A very important question in the cluster analysis is to determine the number of clusters. Virtually all clustering procedures provide little if any information as to the number of clusters present in the data (Milligan and Cooper 1985).

In fact, there is no strict rule to adhere to determine the number of clusters in cluster analysis. There are few guidelines that can be followed when determining the number of clusters. A brief discussion of these guidelines is given below:

• On the basis of theoretical considerations or experience of an executive, number of clusters can be prespecified.
• Agglomeration schedule and dendrogram can also be used to determine the required number of clusters in cluster analysis in terms of the distance measurement at which the clusters are combined. While reading agglomeration schedule, in a good cluster solution, a sudden jump appears. The stage just before the sudden jump point indicates the stopping point for merging of clusters. In the example taken for this chapter, at Stage 22 a sudden jump can be seen. So, this is a stopping point, and three-cluster solution will be a good solution in our case (Figure 19.16).
  

  Figure 19.16 Determining number of clusters through agglomeration schedule

Dendrogram can also be used to determine the number of clusters in cluster analysis. Drawing an imaginary vertical line through dendrogram will present different cluster solutions. The first line in Figure 19.17 presents a three-cluster solution, one for each cut point where the branch of the dendrogram intersects the imaginary drawn vertical line. By considering different cut points, different number of cluster solutions can be obtained (as can be seen that second vertical line presents a two-cluster solution). A good cluster solution can be obtained by considering small within-cluster distances and large between-cluster distances. By tracing backward down the branches, cluster membership can be obtained.

Figure 19.17 Determining number of clusters through dendrogram

Number of clusters can also be determined by simply considering relative cluster size. From Figure 19.13(d) if we take simple frequency count of the cluster membership, then the number of clusters can be determined. Table 19.4 presents different number of cluster solutions with the number of subjects in the concerned cluster.

TABLE 19.4 Different number of cluster solutions with number of subjects in the concerned cluster

So, it can be seen from the table that a three-cluster solution presents a meaningful cluster solution with relative equal distribution of the subjects in concerned clusters.

19.2.4.5 Substantive Interpretation

After clustering, it is important to find the meaning of the concerned clusters in terms of finding some natural or compelling structure in the data. For this purpose, cluster centroid can be a useful tool. Remember that cluster centroid can be obtained through discriminant analysis. Figure 19.18 exhibits these cluster centroids. From the figure, it can be seen that Cluster 1 is associated with high values on Variables X₁ (I am trying very hard to understand the loan schemes offered by different banks to purchase the product when it will be coming in the market), X₃ (my old bike is old fashioned; I want to get rid of it), and X₅ (I will not wait for product’s performance in the market, I know company is reputed and will certainly launch quality product). This cluster (Cluster 1) can be named as ‘desperate consumers.’ Cases 1, 5, 6, 9, 12, 15, 19, 22, and 25 are the members of ‘Cluster 1’ [see Figure 19.13(d)]. Cluster 2 has relatively low values on Variables X₁, X₃, and X₅ and relative high value on Variable X₆. Variable X₆ is “companies always claim high about its product, let the product come in the market then only I will be taking any decision about the purchase.” This cluster (Cluster 2) can be named as ‘patient consumers.’ Cases 2, 4, 8, 14, 16, 18, 20, and 24 are the members of ‘Cluster 2’. Cluster 3 has relative high values on Variables X₂ (I am still in confusion whether to sell my old bike), X₄ (my company has promised that it will be releasing a much awaited new incentive scheme; purchase of new model is based on this factor), and X₇ (I want to purchase a new bike but my kids are growing up and are demanding to purchase a car instead of bike, which I already have). This cluster (Cluster 3) can be named ‘perplexed consumers.’ Cases 3, 7, 10, 11, 13, 17, 21, and 23 are the members of ‘Cluster 3’ [see Figure 19.13(d)].

Figure 19.18 Cluster centroids obtained through discriminant analysis

19.2.4.6 Check the Model Fit

Formal procedures for testing statistical reliability of clusters are not fully defensible. Following are some ad hoc procedures to put a rough check on the quality of cluster analysis.

• Perform the cluster analysis on same data using different distance measures and compare the results across distance measures. In addition, different methods of clustering for the same data can be used and the result can be compared.
• Split the data into halves and perform the cluster analysis on the halves. Obtained cluster centroid can be compared across subsamples.
• Delete various variables from the original sets of variables and perform the cluster analysis on remaining set of variables. Obtained result should be compared with the result obtained from the original set of variables.

19.2.5 Non-Hierarchical Clustering

As discussed earlier, to get the optimum cluster solution, hierarchical clustering procedure must be used first to obtain the number of clusters and then these number of clusters can be used as the basis of initial information to perform k-means cluster analysis (non-hierarchical clustering procedure). The k-means cluster analysis typically involves minimizing the within-cluster variation or, equivalently, maximizing the between-cluster variation (Magidson and Vermunt 2002). In the hierarchical clustering method, a three-cluster solution is obtained for the motorbike company example discussed in the beginning of this section. Using this information, the SPPS output for the problem at the beginning of this section through k-means cluster analysis (non-hierarchical clustering procedure) is presented in the form of Figure 19.19(a) to 19.19(g).

Figure 19.19 (a) Initial cluster centers, (b) Iteration history

Figure 19.19 (c) Cluster membership, (d) Final cluster centers, (e) Distance between final cluster centers, (f) ANOVA table, (g) Number of cases in each cluster

Figure 19.19(a) exhibits Initial cluster centers. These initial cluster centers are the values of three randomly selected cases. The SPSS by default selects the cases that are dissimilar and then cases join these initial values (based on similarity) to make distinct clusters. Each subject joins nearest classification cluster center and the process continues until the stopping criteria is reached.

Figure 19.19(b) presents Iteration history and is the progress of estimation process for each iteration. In each iteration, cases join different clusters, as a result cluster center changes. Any number in the second row of Figure 19.19(b) against “iteration 2” indicates the distance of new cluster center (after joining of cases) from the previous cluster center. When a zero appears in the rows against iteration, change in distance is very small, and stopping criteria is achieved to obtain final solution.

Figure 19.19(c) presents Cluster membership. As the name indicates, this figure shows the cluster membership information for each case. In the “Distance” column, distance between the subject and cluster center is given. This distance can be used to obtain some knowledge about the “representativeness” of the case for the concerned cluster. A small distance indicates that the cases are good representative of the cluster and a large distance indicates the cases are not good representative of the cluster.

Figure 19.19(d) shows ‘Final cluster centers’. Values in the figure indicate means for each variable within each final cluster. It can be seen that first cluster has high cluster-center values on Variables X₁, X₃, and X₅. This cluster has also a relatively low value for Variable X₆. Second cluster has relatively high cluster-center value for Variable X₆. Third cluster has relatively high cluster-center value on Variables X₂, X₄, and X₇.

Figure 19.19(e) shows Distance between final cluster centers and the “Euclidean Distances” between the final cluster centers. Relatively large values indicate that clusters are different from each other. However, relatively small values indicate that clusters are not so different from each other.

Figure 19.19(f) exhibits ANOVA table. This figure indicates the variables that are most important in cluster solution and cluster analysis. Mean square in the Cluster column indicates the variance in the variable that can be attributed to clusters. Mean square in the Error column indicates the variance in the variable that cannot be attributed to clusters. F ratio indicates the ratio of cluster variance to error variance.

Large F value indicates variables that are important for segregating the clusters. Small F value (close to 1) indicates variables that are not very useful in identifying the membership of the clusters F statistic should not be interpreted in the traditional way and hence the significance value associated with F value cannot be interpreted in the traditional manner, in terms of acceptance or rejection of null hypothesis.

Figure 19.19(g) presents Number of cases in each cluster. This figure shows number of cases assigned to each cluster. A big mismatch in the proportional size of the cluster is an indication of some problem in the cluster solution. For example, if Cluster 1 has got 18 members and Clusters 2 and 3 have only 3 and 4 members, respectively, then there is some problem with the cluster solution. In this kind of situation, the number of clusters requested should be rechecked, and included list of variables in the cluster analysis must be re-examined in light of its importance in the cluster analysis.

19.3 MULTIDIMENSIONAL SCALING

This section deals with a famous multivariate technique known as multidimensional scaling. It presents some basic understanding about the multidimensional scaling procedure and deals with the use of SPSS for performing multidimensional scaling.

19.3.1 Introduction

In the field of marketing, companies are generally concerned with the issue of positioning a product. Management of a company is always interested in knowing the position of its products as compared with the position of competitor’s product in the market.

Multidimensional scaling is an attempt to answer such questions.

Multidimensional scaling commonly known as MDS is a technique to measure and represent the perception and preferences of respondents in a perceptual space as a visual display. The goal of multidimensional scaling is to represent the relationships among objects by constructing a configuration of n points in low dimension from pair-wise comparison of similarities/dissimilarities among a set of n object (Huang et al. 2006). Multidimensional scaling handles two marketing decision parameters. As a first case, the dimension on which respondents evaluate objects must be determined. As a convenient option, only two dimensions are worked out as the evaluation objects are graphically portrayed. As a second case, objects are to be positioned on these dimensions.

The output of multidimensional scaling happens to be in the form of location of objects on the dimensions and is termed as spatial map or perceptual map.

As discussed, cluster analysis groups individuals or objects or cases into relatively homogeneous (similar) groups on the basis of similarity. Subjects grouped within the cluster are similar to each other and there is dissimilarity between the clusters.

Multidimensional scaling attempts to infer the underlying dimensions from the preference judgment provided by the customers. This is done by assigning responses of the respondents to a specific location in a perceptual space in a manner that the distances in the space match the given dissimilarity as closely as possible. In fact, cluster analysis ends with grouping subjects into clusters, whereas multidimensional scaling ends with the construction of a graph in which the distance between objects can visually and quantitatively be examined. Data obtained from the respondents can be metric or non-metric. As a case of metric data, rating of respondent’s preference can be obtained and as a non-metric data, ranking of the respondent’s relative preference can also be obtained. Multidimensional approaches are available for analyzing metric as well as non-metric data.

19.3.2 Some Basic Terms Used in Multidimensional Scaling

Following list presents some basic terms quite commonly used in multidimensional scaling:

Stress: Stress measures lack of fit in multidimensional scaling. A higher value for stress is an indication of poorer fit.

R-square (squared correlation): R² value indicates how much of the variance in the original dissimilarity matrix can be attributed to multidimensional scaling model. Higher value for R² is desirable in multidimensional scaling model. In fact, R² is a goodness- of-fit measure in multidimensional scaling model.

Perceptual map: Perceptual map is a tool to visually display perceived relationship among various stimuli or objects in a multidimensional attribute space.

19.3.3 The Process of Conducting Multidimensional Scaling

For conducting multidimensional scaling, a researcher usually follows six steps as exhibited in Figure 19.20. These six steps are as follows: problem formulation, input data collection, selection of multidimensional scaling procedure, determining number of dimensions for perceptual map, substantive interpretation, and check the model fit. A step-by-step description of performing multidimensional scaling is presented in the following section.

Figure 19.20 Steps involved in conducting multidimensional scaling

19.3.3.1 Problem Formulation

As a first step of conducting multidimensional scaling, brand or stimuli that are to be compared are selected. The number of brands that are to be selected is a matter of researcher’s discretion, but as a matter of understanding, too few brands like three of four will not be able to produce the desired perceptual map and using too many brands will be difficult to interpret through perceptual map. So there exists a question as to how many brands should be included in multidimensional scaling procedure. A minimum of 8 to 10 brands can be included to construct a well-defined perceptual map, and a maximum of 25 to 30 brands can be included in multidimensional scaling model. In fact, the number of brands or stimuli to be included is based on some factors such as research objective, past researches, decision of researchers, or requirement of management.

We will take a hypothetical example of 10 edible oil brands for better understanding the concept of multidimensional scaling with special reference to obtaining a perceptual map. These 10 edible oil brands are Fortune, Sundrop, Saffola, Gemini, Nutrela, Dhara, Ginni, Maharaja, Vital, and Nature Fresh. Step 2 of performing multidimensional scaling provides discussions on how we can collect data for theses 10 brands of edible oil.

19.3.3.2 Input Data Collection

The input data used for multidimensional scaling may be connected with the similarity data or the preference data. As a first case, similarity data are described as follows:

Similarity Data: Similarity data are collected through the respondents by just noting the perceived similarity between the two brands or objects. While collecting similarity data, the respondents are not provided with a set of attribute list to judge the similarity or dissimilarity, rather they make their own assessment about similarity or dissimilarity between two objects. These data are often referred to as similarity judgment. Figure 19.21 provides respondent’s similarity judgment between two pairs (Fortune–Saffola) of edible oil brands.

Figure 19.21 Respondent’s similarity judgment between two pairs of edible oil brands

Number of pairs required to be judged based on similarity judgment are . Hence, in the edible oil example, number of pairs to be judged are This approach of data collection is referred to as the direct approach.

As a second way, derived approach for data collection in terms of conducting multidimensional scaling can also be used. Using this approach, the respondents are supposed to rate the brands for identified attributes on a rating scale. Responses obtained from a single respondent are summarized in Table 19.5 above.

TABLE 19.5 Similarity score data for different brands of edible oil (data that will be used for multidimensional scaling)

Preference Data: In some cases, a researcher may be interested in knowing the respondent’s preference for a stimuli or object. In such situations, the respondent is asked to provide rank order for all the objects or stimuli as per their preference. The configuration derived from similarity data and preference data are not the same but are different. Two objects can be perceived differently through a similarity data produced map but may be close together for a preference data. For taking care of this dimension, an ideal object approach is considered.

An ideal object is the object that a respondent will like to prefer over all other objects included in the study. It is very interesting to note that this ideal object is actually a hypothetical object. This object can be conceptualized in the map but does not really exist. Respondents may be having similar perception about objects but their preference may vary considerably. Ideal objects are of two types. These two ideal objects can be explained with the help of two examples. As a first example, the respondent is required to rate his preference on an attribute scale related to the “smell of the edible oil when boiling.” As his response, suppose the respondent has not selected extreme points but selected some middle point in the scale. The scale is given below:

Good smelled — — — — — — — Bad odoured

As a second example, a respondent is required to rate his preference for edible oil on a scale (for three attributes) given below:

Very good packaging — — — — — — — Not at all good packaging

Health friendly — — — — — — — Not health friendly

Fat free — — — — — — — Containing fat

In the second case, suppose the respondent has preferred an extreme point of the scale as very good packaging. This case is different from the first case. In the first case, when the respondent has selected some middle point, the ideal object lies within the perceptual map. In the second case, when the respondent has selected an extreme point as his point of preference, the ideal point is reflected by a direction or vector in the perceptual map.

19.3.3.3 Selection of Multidimensional Scaling Procedure

While selecting multidimensional scaling procedure, a researcher should focus on two important issues. The first issue is related to the nature of the data and the second point is related to using multidimensional scaling procedure for average similarity ratings. The input data play a key role in determining multidimensional scaling procedure. Non-metric multidimensional scaling procedure is based on the ordinal nature of input data, whereas metric multidimensional scaling procedure is based on the assumption that the input data are interval scaled. In multidimensional scaling procedure, ordinal or non-metric information is preferred. It is important to learn that non-metric multidimensional scaling procedure results in a metric output. Obviously, metric multidimensional scaling procedure produces a metric output. Here, it is important to note that metric and non-metric multidimensional scaling procedure both produce similar type of results.

As a second issue, a researcher has to determine whether multidimensional scaling procedure should be performed on an individual or an aggregate data is required. Individual-level multidimensional scaling procedure results in a perceptual map for each respondent. While performing multidimensional scaling procedure on aggregate data, a perceptual map on the basis of average similarity rating can be obtained very easily. When using multidimensional scaling procedure for aggregate data instead of an individual’s ranking, an aggregate ranking for various individuals is obtained. Aggregating of multidimensional scaling input data is based on the assumption that all the respondents included in the study use the same dimension to evaluate the objects, but these common dimensions are weighted differently by different respondents. Our discussion of multidimensional scaling is based on the edible oil brands example for which data are rank ordered (ordinal) and the adopted procedure is non-metric.

19.3.3.4 Determining Number of Dimensions for Perceptual Map

The focus of multidimensional scaling is to develop a perceptual map with smallest number of dimensions for which there is a “best-fit” between the similarity ranking as input data and resulting perceptual map. In the light of the visual interpretation objective of multidimensional scaling procedure, a two-dimensional or at most a three-dimensional perceptual map is desirable. As discussed in the beginning of the discussion of multidimensional scaling, a statistic “stress” is used. Stress measures lack of fit in multidimensional scaling as the higher value for stress is an indication of poorer fit. Stress measure indicates the proportion of the variance of the disparities (differences in distances between objects on the perceptual map and the similarity judgment of the respondents) not accounted for by multidimensional scaling model (Hair et al. 2009). A widely used criteria to determine the number of dimensions in multidimensional scaling is to construct a plot between stress values (obtained as the SPSS output) and dimensionality. An elbow in the plot indicates number of dimensions to be included in the study to construct a perceptual map.

19.3.3.5 Substantive Interpretation

For generating a perceptual map, we will use the SPSS. The SPSS output for edible oil data is presented from Figures 19.22(a) to 19.22(d). Figure 19.22(a) is the SPSS output exhibiting iteration for stress-value improvement, stress value, and R² value. Figure 19.22(b) is the SPSS output exhibiting stimulus coordinates. Figure 19.22(c) is a SPSS-produced two-dimensional perceptual map. Figure 19.22(d) is a SPSS-produced three-dimensional perceptual map.

Figure 19.22 (a) SPSS output exhibiting iteration for stress value improvement, stress value, and R² value, (b) SPSS output exhibiting stimulus coordinates

Figure 19.22 (c) SPSS-produced perceptual map (two dimensional), (d) SPSS-produced perceptual map (three dimensional)

As discussed, Figure 19.22(a) is the SPSS output exhibiting iteration for stress-value improvement, stress value, and R² value. As discussed, this stress index indicates lack of fit in multidimensional scaling. This stress value is commonly known as S-stress or Kruskal’s stress. This value ranges from the worst fit (stress value as 1) to best fit (stress value as 0). In fact, stress value is based on the type of multidimensional scaling procedure adopted and the data on which multidimensional scaling is performed. Now, we will analyse the acceptable stress value in multidimensional scaling. The acceptable stress value is suggested by Kruskal (1964) as given below:

As a second step, we are required to examine the value of R². As discussed, R² value indicates how much of the variance in the original dissimilarity matrix can be attributed to multidimensional scaling model. Higher value for R² (close to 1) is desirable in multidimensional scaling. An R² value greater than or equal to 60% is considered acceptable.

Figure 19.22(b) exhibits the stimulus number (brands) and their scores on different dimensions. Figure 19.22(c) is the desired SPSS-produced perceptual map (two dimensional). Spatial map may be interpreted by examining the coordinates of the map and relative position of the brands with respect to these coordinates. The labelling of horizontal axis (X-axis) and vertical axis (Y-axis) is a matter of researcher’s judgment and depends on factors such as researcher’s insight, obtained information parameters, and so on. In some cases, the respondents are often asked to provide the base of similarity they have used for judging the different brands or objects.

Figure 19.23 exhibits perceptual map (two dimensional) with labelling of dimensions. Brands located close to each other may have competitive nature on related dimension. A brand located in isolation may have unique image. This perceptual map is based on the similarity judgment of a single respondent. If we take an aggregate score of the responses, a perceptual map based on multiple responses (when taken as aggregate) can also be constructed and is very helpful for marketing managers to assess the positioning of their own brand as compared with different brands on some defined attributes.

Figure 19.23 Perceptual map (two dimensional) with labelling of dimensions

19.3.3.6 Check the Model Fit

After performing multidimensional scaling, it is very important for a researcher to assess the reliability and validity of multidimensional scaling model. As a first step of checking reliability and validity of the model, the value of R² must be examined. As discussed, an R² value greater than or equal to 60% is considered acceptable. In edible oil multidimensional scaling model, R² value comes to 0.9707 (97.07%), which is very close to 1 and hence the model is very well acceptable. As a second step, stress value must be examined. In the previous section, the interpretation of an S-stress or Kruskal’s stress is already presented. In edible oil multidimensional scaling model, stress value comes to 0.0746 (close to 5%). This is an indication of a good-fit multidimensional scaling model. As a usual approach for performing multidimensional scaling, the original data should be divided in two or parts and obtained results must be compared. As another measure, input data must be gathered at two different points of time and test–retest reliability must be computed.

19.4 CORRESPONDENCE ANALYSIS

Discriminant analysis and factor analysis both are based on the assumption of interval-scaled data, which generally is rated on a 1- to 7-point rating scale. However, there are situations when the input data are only binary or categorical when the respondents are simply asked to state which attribute apply to a list of several brands. In this case, respondents are supposed to present their answer in “Yes” and “No” only. “Yes” indicates that the attribute is applicable to the concerned brand and “No” indicates that the attribute is not applicable to the concerned brand. Correspondence analysis is a technique that looks like multidimensional scaling and is used to scale qualitative data in the field of marketing research. Correspondence analysis also generates a perceptual map in which both attribute elements and object or stimuli are positioned. As a matter of difference from multidimensional scaling, correspondence analysis generates perceptual map from nominal or categorical scaled data. Correspondence analysis has the capacity to position products or brands with respect to any type of data (e.g., attitude, usage occasions). Both multidimensional scaling and correspondence analysis are based on the concept of similarity. Correspondence analysis defines similarity in terms of sharing the same level of categorical variables.

References

Aaker, D. A., V. Kumar and G. S. Day (2000). Marketing Research, 7th ed., p. 554. Wiley.

Bentler, P. M. and K. H. Yuan (1998). “Tests for Linear Trend in the Smallest Eigenvalues of the Correlation Matrix”, Psychometrica, 63(2): 131–44.

Cattell, R. B. (1966). “The Screen Test for the Number of Factors”, Multivariate Behavioural Research, 1(2): 245–76.

Costello, A. B. and J. W. Osborne (2005). “Best Practices in Exploratory Factor Analysis: Four Recommendations for Getting the Most from Your Analysis”, Practical Assessment, Research & Evaluation, 10(7): 1–9.

Fabrigar, L. R., D. T. Wegener, R. C. MacCallum and E. J. Strahan (1999). “Evaluating the Use of Exploratory Factor Analysis in Psychological Research”, Psychological Methods, 4(3): 272–99.

Hair, J. F., W. C. Black, B. J. Babin, R. E. Anderson and R. L. Tatham (2009). Multivariate Data Analysis, 6th ed., p. 125. Pearson Education.

Huang, J., C. Ong and G. Tzeng (2006). “Interval MultiDimensional Scaling for Group Decision Using Rough Set Concept”, Expert System with Applications, 31(3): 525–30.

Kruskal J. B. (1964). “Multidimensional Scaling by Optimizing Goodness of Fit to a Nonmetric Hypothesis”, Psychometrica, 29: 1–27.

Lehmann, D. R., S. Gupta and J. H. Steckel (1998). Marketing Research, pp. 575. Addison-Wesley.

Magidson, J. and J. K. Vermunt (2002). “Latent Class Models for Clustering: A Comparison with K-Means”, Canadian Journal of Marketing Research, 20: 37–44.

Milligan, G. W. and M. C. Cooper (1985). “An Examination of Procedures for Determining the Number of Clusters in a Data Set”, Psychometrica, 50(2): 159–79.

Punj, G. and D. W. Stewart (1983). “Cluster Analysis in Marketing Research: Review and Suggestions for Applications”, Journal of Marketing Research, 20: 134–48.

Raven, M. R. (1990). “The Application of Exploratory Factor Analysis in Agriculture Education Research”, Journal of Agriculture Education, 35(4): 9–14.

Reise, S. P., N. G. Waller and A. L. Comrey (2000). “Factor Analysis and Scale Revision”, Psychological Assesment, 12(3): 287–97.

Zikmund, W. G. (2007). Business Research Methods, 7th ed., p. 589. South-Western Thomson Learning.

Summary

The techniques described in this chapter, namely, factor analysis, cluster analysis, multidimensional scaling, and correspondence analysis are often referred to as “interdependence analysis.” Factor analysis is a very useful technique of data reduction and summarization. The main focus of factor analysis is to summarize the information contained in a large number of variables into few small numbers of factors. Factor analysis is performed using the following steps: problem formulation, construction and examination of correlation matrix, selection of an appropriate factoring technique, a decision regarding the number of factors to be retained in the final solution, factor rotation for enhancing the interpretability of the solution, substantive interpretation, and check the model fit.

Cluster analysis is a technique of grouping individuals or objects or cases into relatively homogeneous (similar) groups. These homogeneous groups are often referred to as clusters. The subjects grouped within the cluster are similar to each other, and there is dissimilarity between the clusters. Cluster analysis is performed using six steps. These six steps are as follows: problem formulation, choice of an appropriate proximity or similarity measure, selection of an appropriate clustering technique, a decision regarding the number of clusters to be retained in the final solution, substantive interpretation, and check the model fit. To get the optimum cluster solution, first hierarchical clustering procedure must be used to obtain the number of clusters and then this number of clusters can be used as the basis of initial information to perform k-means cluster analysis (non-hierarchical clustering procedure).

Multidimensional scaling commonly known as MDS is a technique to measure and represent perception and preferences of respondents in a perceptual space as a visual display. Multidimensional scaling handles two marketing decision parameters. As a first case, the dimension on which the respondents evaluate objects must be determined. As a convenient option, only two dimensions are worked out as the evaluation objects are graphically portrayed. As a second case, the objects are to be positioned on these dimensions. The output of multidimensional scaling happens to be in the form of location of the objects on the dimensions and is termed as spatial map or perceptual map. Multidimensional scaling is performed using six steps. These six steps are as follows: problem formulation, input data collection, selection of multidimensional scaling procedure, determining number of dimensions for perceptual map, substantive interpretation, and check the model fit.

Correspondence analysis is a technique that looks like multidimensional scaling and is used to scale qualitative data in the field of marketing research. Correspondence analysis also generates a perceptual map in which both attribute elements and object or stimuli are positioned. As a matter of difference from multidimensional scaling, correspondence analysis generates perceptual map from nominal or categorical scaled data. Correspondence analysis has the capacity to position products or brand with respect to any type of data (e.g., attitude, usage occasions). Both multidimensional scaling and correspondence analysis are based on the concept of similarity. Correspondence analysis defines similarity in terms of sharing the same level of categorical variables.

Key Terms

% of variance, 476

Agglomeration schedule, 483

ANOVA table, 500

Average linkage method, 487

Bartlett’s test of sphericity, 470

Centroid method, 487

Cluster analysis, 468

Cluster membership, 483

Communality, 470

Complete linkage method, 487

Correlation matrix, 470

Correspondence analysis, 468

Dendrogram, 483

Dissimilarity matrix, 501

Distance between final cluster centers, 500

Euclidean distance, 485

Eigenvalue, 470

Factor loading plot, 471

Factor loadings, 470

Factor matrix, 470

Factor score, 469

Factors, 469

Final cluster center, 500

Hierarchical clustering approach, 486

Icicle plot, 483

Initial cluster centres, 498

Interdependence analysis, 468

Iteration history, 499

Kaiser-Mayer-Olkin measure of sampling adequacy, 470

Linkage method, 487

Multidimensional scaling, 468

Non-hierarchical clustering approach, 486

Number of cases in each cluster, 500

Oblique rotation, 480

Orthogonal rotation, 480

Parallel threshold optimizing partitioning, 488

Preference data, 502

Principal axis factoring, 476

Rotation, 471

R-square (squared correlation), 501

Scree plot, 470

Sequential threshold, 488

Similarity data, 502

Single linkage method, 487

Spatial map or perceptual map, 501

Stress, 501

Sudden jump, 495

The principal component method, 476

Variance method, 487

Variance methods and centroid method, 487

Varimax procedure, 479

Ward’s method, 488

Notes

1. www.indiastat.com, accessed September 2009, reprinted with permission.
2. http://www.bata.in/page.php?kon=5_2_1, accessed October 2009.
3. http://www.telegraphindia.com/1060331/asp/others/print.html.

Discussion Questions

1. What is the conceptual framework of factor analysis? Under what circumstances a researcher should apply factor analysis.
2. What are the steps used in performing factor analysis?
3. Write short note on following topics related to factor analysis:
   1. Estimation or analysis sample
   2. Communality
   3. Correlation matrix
   4. Kaiser-Mayer-Olkin measure of sampling adequacy
   5. Bartlett’s test of sphericity
   6. Eigenvalue, % of variance
   7. Scree plot, factor loadings
   8. Factor matrix and factor score
   9. Factor loading plot
   10. The principal component method
   11. Principal axis factoring
   12. Varimax procedure
   13. Orthogonal rotation
   14. Oblique rotation
4. What is the conceptual framework of cluster analysis? Under what circumstances a researcher should apply cluster analysis.
5. What are the steps used in performing cluster analysis?
6. Explain the difference between hierarchical clustering technique and non-hierarchical clustering technique. What is the use and application of clustering technique in making managerial decisions.
7. Write short note on following topics related to cluster analysis:
   1. Agglomeration schedule
   2. Cluster membership
   3. Icicle plot
   4. Dendrogram
   5. Euclidean distance
   6. Linkage method
   7. Variance methods and centroid method
   8. Single linkage method
   9. Complete linkage method
   10. Average linkage method
   11. Variance method
   12. Ward’s method
   13. Centroid method
   14. Dissimilarity matrix
   15. Cluster membership
   16. Final cluster centre
   17. Distance between final cluster centres
   18. ANOVA table in clustering
8. What is the conceptual framework of multidimensional scaling? Under what circumstances a researcher should apply multidimensional scaling.
9. What are the steps used in performing multidimensional scaling?
10. Write short note on following topics related to multidimensional scaling:
    1. Spatial map or perceptual map
    2. Stress, R² (squared correlation)
    3. Similarity data
    4. Preference data
11. What do you understand by correspondence analysis? Under what conditions a researcher should be applying correspondence analysis.

Case Study

Case 19: Britannia Industries Ltd: A leading player in Indian Bakery Industry

Introduction: An Overview of Indian Bakery Industry

Indian bakery industry is mainly dominated by small-scale sector. In India, bakery products primarily include biscuits and bread. The two major bakery products in India are biscuits and bread that accounts for 82% of all bakery production. For all the leading companies producing bakery products, biscuits have got the lion’s share in total sales. In the year 2001–2002, demand for biscuits was 1188 thousand metric tonnes, which is estimated to increase to 2758 thousand metric tonnes in the year 2014–2015. Market segmentation wise organized and unorganized market secure 50% of the total market size. In the time span of 2009–2010 to 2014–2015, the market is estimated to grow at a rate of 6.2%. Biscuit market is also segmented with respect to product line. Glucose, Milk, Marie, Cream, Crackers, and others occupy 60%, 10%, 10%, 5%, 7%, and 8% of the total market, respectively. Britannia and Parle are two major players in the biscuit market. In the year 2001–2002, demand for bread was 12.85 billion rupees, which is estimated to increase to 26.90 billion rupees in the year 2014–2015. For bread, only 15% market is occupied by the organized players and 85% market is captured by unorganized players. In the time span of 2009–2010 to 2014–2015, bread market is estimated to grow at a rate of 5.0%. Milk bread has got the lion’s share with 85% of share of the total market. Modern food and Britannia are two major producers of bread.¹

Britannia Industries Limited: A Leading Player in Indian Bakery Industry

Britannia Industries Ltd is a leading bakery products company in India with a predominant focus on the sale of branded biscuits. The company’s history goes back to 1892, when it was incorporated in Kolkata under the name of Britannia Biscuit Company. BIL got its current name in 1979. In 1993, the Nusli Wadia group acquired a stake in BIL’s parent company, Associated Biscuits International Ltd, UK and became an equal partner in BIL with the French Major Groupe Danone. The company is engaged in the business of biscuits, bread, cakes, and rusks. It caters to diverse needs and tastes of the Indian consumer across age groups through its optimum range of biscuit brands. Some of its popular brands are Tiger, Good Day, Marigold, Milk Bikis, Treat, 50:50, Little Heart, Bourbon, Pure Magic, Snax, Premium Bake, and Nutrichoice. The company’s strategy of consistently renovating its existing brands and launching new ones has helped it to garner larger share of the Indian biscuits market every year. In the year 2005–2006, the company’s market share stood at 37.2%. Britannia is the pioneer in the sliced bread business in India. Today, it caters to almost half the branded bread market in India. Its cake and rusk business, albeit small, has been growing at a healthy pace. However, the bread, cakes, and rusk business together accounts for only 8.9% of the company’s total turnover. It is essentially a biscuit company with 90.2% revenue coming from the sale of biscuits.²

The company has also been following the inorganic route of expansion quite actively in the recent past. It made its first international acquisition in March 2007 by buying 70% stake in Strategic Food International Co LLC, Dubai, a leading company in the biscuits and cookies segment in the GCC (Gulf Cooperation Council) markets. In July 2005, it acquired 50% stake in Bangalore-based Daily Bread. The company believes that this acquisition will help increase its presence in select markets with a range of gourmet sold under the brand names Daily Bread and Deluca. In addition to the bakery business, Britannia runs its dairy operations through its joint venture Britannia New Zealand Foods Pvt. Ltd. This joint venture with Fonterra of New Zealand was formed in 2002 by demerging the dairy products business from the Britannia Industries Limited.² TABLE 19.01 below exhibits sales and profit after tax (in million rupees) of the Britannia Industries Limited from 1994–1995 to 2007–2008. The company has registered encouraging profit after tax figure in the financial year 2007–2008.