Bruce Cooil, Lerzan Aksoy, and Timothy L. Keiningham proposed a study titled “Approaches to Customer Segmentation” published in the Journal of Relationship Marketing in 2008. The study focuses on how a multi-national firm has developed effective segmentation strategies and gained a conceptual understanding of how inventive and novel approaches can provide more value. In addition, the research has reviewed various general approaches to customer segmentation, with an emphasis on practical and flexible analytical and statistical approaches.
The primary theme of the research is to provide a comprehensive review of the most efficient analytical methods that can offer firms more value and market-related information. The authors suggest that a priori and custom-designed post-hoc approaches are practical for firms, and they should be implemented commonly. The researchers also highlighted various statistical methods and powerful tools for analysis segmentation results, such as clustering algorithms, as they provide a novel approach to view the segmentation framework.
Furthermore, latent analysis as a descriptive method offers insights into the objectives of a segmentation task. However, the findings of the research highlight a problem wherein the best statistical methods may provide substantial and statistically assessed and identifiable segmentation results, but they are not always accessible or actionable. Therefore, the authors pinpoint the need for analytical insights to provide the necessary statistical evaluation to generate significant and clearly defined segments that are responsive to the marketing necessities of the firms to target individual customer groups.
Market segmentation comprises different groups with diverse needs, characteristics, and behavior, which requires different marketing efforts. Therefore, effective segmentation necessitates each segment to be evaluated on various criteria such as stability, growth potential, size, accessibility, responsiveness, and consistency. An effective segmentation ensures companies determine the customer groups and position their products and services according to the demands of each group. To achieve appropriate segmentation results, the research suggests statistical methodologies that can provide valuable insights. Firstly, the research focuses on the a priori segmentation method, which is considered a critical component of market segmentation. Such methods are useful for customer segmentation before collecting data using factors like customer characteristics and product-related information. Furthermore, the goals of such analyses are considered to be descriptive, which allows the development of models based on predefined segments to predict one or more dependent variables.
Furthermore, the authors suggest a post-hoc method that segments primarily through data analysis. The post-doc approach aims to study the customer groups and effectively develops a predictive model for various dependent variables. However, the author illustrates that studies can implement a hybrid approach of combining a-priori and post-hoc analysis depending on the market requirements. Moreover, the authors illustrated that latent class models are more promising for effective post-hoc descriptive and predictive analysis. For the experimental analysis, a framework was devised that includes an a-priori definition of the customer segments with data that has supervised labels, and the goal of the model is to classify new customers based on the labels. In addition, general clustering procedures were not based on any statistical models for the descriptive analysis, and predictive clustering approaches were briefly summarized. Finally, the study explored the various components of the latent class model that possess the ability to provide appropriate descriptive and predictive post-hoc analytical results.
The study emphasizes the need for customer perspectives to achieve a successful customer segmentation method. To attain a conceptual understanding, the authors performed a case study to analyze the customer segmentation approaches of Migros Turk T.A.S, which is the largest grocery chain in Turkey. The purpose of considering Migros is because of the rapid expansion strategies that saw them increase the number of stores in Istanbul and various other regions in Turkey. In addition, the company introduced various store formats based on the size and the diversity of the product ranges and demands. Moreover, the company established several stores under different brand names to cater to different market segments.
Furthermore, the research evaluates the effectiveness of Migros’s different segmentation methods for their customer base, including values like behavior, lifestyle, life cycle, and activity-based segmentation approaches. The authors suggest that these factors can be used individually or in conjunction in a hybrid approach. Firstly, the value-based segmentation technique was assessed, including household as one of the primary units of analysis. Next, the base is divided into profitability group and frequency of visits by customers. These two measures form the productivity index to determine customer loyalty in six segments (see figure 1).
Figure 1: The Value Pyramid
The value pyramid signifies the six different segments based on different household products purchased, such as fruits, vegetables, and dairy over 6 months. Partitions 1,3, and 5 define the loyal segments, whereas 1 and 2 are considered the most valuable customers in the pyramid.
On the other hand, behavioral-based segmentation includes individuals as part of the analysis. The cluster analysis technique is utilized to determine the customer base with similar behavioral patterns regarding the number of purchases, food and non-food purchase, location, seasonal purchase, communications, and products. However, the fundamental difference of value-based segmentation is that it is not limited to the amount spent, food and non-food purchases, or communication level.
The lifestyle-based segmentation also includes individuals as the primary factor of the analysis. However, contrary to behavior-based segmentation, this approach prioritizes customers based on a priori categorization, including the profile with the list of products these customers are expected to buy. Therefore, the segmentation comprises 14 predetermined lifestyle groups, wherein the gourmet purchases such as cheese, fine wine, and other ethnic foods were considered high profile customers. Therefore, the existing customers were compared and included as either high profile customers or potential candidates to one of the 14 segments depending on the average and standard deviation evaluation criteria.
Finally, the activity-level-based segmentation considers individuals and a 15-month longitudinal data period for customer classification as active, regular, or passive. These customers are classified in terms of purchase amounts and frequency. Furthermore, the average from the 12 months is calculated to determine the standard deviation to identify the recent purchases. From the segmentation schemes’ findings, the author highlighted how Migros improved productivity while aiding suppliers about various analytical factors to evaluate the performance against the competitors. The study also suggests that the segmentation schemes allowed Migros to organize collaborative campaigns with banks, gas stations for successfully acquiring new customers.
Additionally, the authors reviewed the segmentation approaches based on supervised classification. The research emphasized the most commonly used linear discriminant analysis methods and many other classification methods such as tree-based, neural networks, and support vector machines. Although, boosting allows to improve the classification performance, the logistic model is more flexible for K classes, where each class K is estimated using logistic regression. Besides, the classification rule constructed from the maximum likelihood estimation of the logistic regression model is equally capable as the best available boosting procedure. Additionally, logistic regression is considered a more robust procedure than linear discriminant analysis.
Future Work Suggestions and Implications for Practitioners
Customer segmentation is a complex process when multiple variables and multiple segmentations are to be performed. Therefore, latent class models provide a segmentation framework within which the general linear models with predictors can vary according to the segments when representing complex variables in building segmentation relationships. In addition, these models offer flexible classification functions involving covariates and offer general features to include within and across segments with dependencies among different components of the multivariate dependent variable. Future studies should accommodate a latent class cluster model as it holds the possibility of combining continuous, ordinal, and nominal variables.
In contrast, these models allow for multivariate normal distribution when dealing with segments, which is among the most accessible and most statistically manageable approaches to model dependencies among variables. Therefore, it is essential to incorporate a flexible approach with every plausible covariate for the latent class when selecting models. Furthermore, this two-step procedure can be implemented repeatedly until there is no scope for improvement. Therefore, one must include the covariate after identifying the appropriate predictors and covariance before eliminating the less influential covariates. Thus, applying this approach offers the starting point of finding the best and consistent model for segmentation. However, it is critical to understand that adding random effects allows generalizing the classification function and provides a way to include additional dependencies among observations.
Besides, there are numerous exceptional cases of latent cluster models where the dependent variable results from a single outcome measure. This measure allows dealing with substantially complicated segmentation models with many coefficients that vary according to the segments. Although, such features are not practical in cluster models as there are many parameters required to include dependent variables and dependencies among multiple components. However, for continuous dependent measures, one must use a general linear model. Thus, a general selection procedure for the latent class regression model should include a general cluster model focused on selecting predictors and is easier to execute.
In the context of conjoint analysis for segmentation, the procedure focuses on estimating customer preferences for various product attributes, where the preferences are based on how the customers evaluate different product attribute profiles. The attributes can include functional and physical features, product characteristics such as brand and price, and more. In such scenarios, these attributes can be presented as benefits or act as a qualitative measure. Therefore, to work with conjoint analysis, a priori segmentation effectively studies the selected product attribute profiles or individual attributes to define appropriate segments. In addition, logistic regression or supervised classification methods are more helpful in studying selected profiles related to other variables. Finally, post-hoc approaches allow the preference data to determine the appropriate segments to target for product design strategies, where the latent class model provides a flexible framework to work with such analysis.