Segmentation Analysis Segmentation Assign people to groupsclusters on

Segmentation Analysis

Segmentation • Assign people to groups/clusters on basis of similarities in characteristics, attributes • Bases of segmentation: - Demographic - Behavior - Geographic - Benefit - Product/ Product usage - Purchase • Here we focus on two techniques used in segmentation

Segmentation: Factor Analysis • Defines underlying structure of data matrix • Summarizes a large no. of var. by identifying a few common underlying dimensions called Factors • Two primary uses: Summarization; Data reduction • It is an interdependence technique • Can be: Exploratory; Confirmatory. • We consider Exploratory view-point.

Segmentation Method: Factor Analysis(contd. ) • Six stage model building for F. A. • Stage 1: Research Problem a. Exproratory or confirmatory b. Data summarization or data reduction • Stage 2: Three basic decisions a. Calculation of input data b. Variable selection and measurement issues c. Sample size

Segmentation: Factor Analysis(contd. ) • Stage 3: assumptions of F. A. a. Normality b. Linearity c. Ensure data matrix has enough correlations d. Homogeneity of sample e. Conceptual linkages

Segmentation: Factor Analysis(contd. ) • Stage 4: Deriving factors and assesing fit. a. Method of extracting factors i. Common F. A. ii. Component F. A. Different types of variances i. Common variance ii. Specific (unique) variance iii. Error

Segmentation: Factor Analysis(contd. ) • Stage 4: Deriving factors and assesing fit (contd. ). b. Number of factors to extract i. Latent root/eigenvalue ii. %-age of variation iii. Scree test iv. A priori criterion

Segmentation: Factor Analysis(contd. ) • Stage 5: Interpreting the factors a. Initial unrotated factors are obtained b. Rotation of factors for better interpretation i. Orthogonal methods - Varimax ii. Oblique methods c. Assess need to respecify factor model

Segmentation: Factor Analysis(contd. ) • Criteria for significance of factor loadings - Practical significance - Statistical significance • Interpreting the factor matrix - Examine matrix of loadings - Identify highest loading for each variable - Assess communalities for acceptable levels - Label the factors

Segmentation: Factor Analysis(contd. ) • Stage 6: Validation of F. A. a. Move to confirmatory perspective and assess replicability with split-sample or a new sample.

Example: Factor Analysis • A direct mail company sent mailings for an upscale good, based on a commercially available cluster analysis package • Contrary to expectation, response came from middle/lower socioeconomic clusters, not from the top • To verify clustering, the company used census data at the zip code level • 40 census variables were chosen and each of the 2000 New York state zip codes was rated on these • 10 factors (those with eigenvalue>1)were used for further analysis

Segmentation: Cluster Analysis • Groups objects into clusters • Only multi variate technique that does not estimate the variate empirically but uses a researcher-specified variate • Focuses on comparison of objects based on variate not on estimation of variate itself • Groups objects whereas Factor Analysis groups variables

Segmentation: Cluster Analysis(contd. ) • Six stage model building process • Stage 1: Research problem a. Select objectives i. Taxonomy description ii. Data simplification iii. Relationship identification b. Select clustering var. Include only var. that i. Characterize objects being clustered ii. Relate to objects of C. A.

Segmentation: Cluster Analysis(contd. ) • Stage 2: Research design a. Can outliers be detected b. Similarity measures i. Metric data (correlational; distance) ii. Non-metric data (association) c. Standardize data? • Stage 3: Assumptions of C. A. a. Ensure representativeness in sample b. Examine is variables are highly correlated

Segmentation: Cluster Analysis(contd. ) • Stage 4: Deriving clusters and assessing fit a. Hierarchical algorithms i. Agglomerative(single linkage, complete linkage, avg. linkage, Ward’s method etc) ii. Divisive(agglomerative in reverse order) b. Non-hierarchical c. Combination

Segmentation: Cluster Analysis(contd. ) • Algorithm for K-means clustering 1. Partition items into K clusters 2. Assign items to cluster with nearest centroid mean 3. Recalculate centroids both for cluster receiving and losing item 4. Repeat steps 2 and 3 till no more reassignments

Segmentation: Cluster Analysis(contd. ) • How many clusters to form - Stop when similarity measure exceeds certain value - Stop when successive values between steps makes sudden jump - Adopt some statistical test or rule - Researcher’s empirical/theoretical basis

Segmentation: Cluster Analysis(contd. ) • Stage 5: Interpretation of clusters a. Examine each cluster w. r. t cluster variate b. Use cluster’s centroid as a measure • Stage 6: Validation and profiling of clusters a. Validation(generalizability, test results on split-sample, predictive validity) b. Profiling(Discriminant analysis with identified clusters as dep. var. & demographics, psychographics as indep. var. )

Example: Cluster Analysis • Stage 1: Objectives To segment HATCO’s customers into groups having similar perceptions of HATCO. Perceptions are based on X 1 through X 7 • Stage 2: Research Design - No outliers - All 7 variables are metric, so Euclidean distance used as similarity measure - No correlations among variables - All variables on same scale, so no standardization needed

Example: Cluster Analysis • Step 3: Assumptions All assumptions are met • Step 4: Deriving clusters and assessing fit a) Hierarchical method - Both a 2 -cluster and a 4 -cluster solution selected for further analysis - Tested for distinctiveness of the clusters b) Non Hierarchical method • Step 5: Interpretation of clusters From Factor analysis of X 1 through X 7 we know that X 1, X 2, X 3 , X 7 constitute one dimension (with X 1, X 3 inversely related to X 2, X 7 ) and X 4 , X 6 another dim

Example: Cluster Analysis From Profile Diagram, Cluster 1 is higher on X 1 and X 3 and lower on X 2 and X 7. Thus cluster 1 is high on dim 1 (Basic value). Also, cluster 2 is high on X 4 , X 6 i. e on dim 2 (Image) • Step 6: Validation and profiling Clusters profiled on additional variables