The Information Edge

In this increasingly competitive marketing environment, it is critical to husband resources to be effective. Successful businesses need to focus marketing efforts on potentially profitable customers, new customer opportunities and those customers for which the firm has competitive advantage. Therefore, identifying potentially profitable interested groups of customers is critical for the formulation of any effective marketing strategy. This is the chore of identifying market "segments."

Identifying market segments serve multiple functions and therefore may be defined in multiple ways. Segmentation should be viewed as perspectives on the market rather than inherent characteristics. Depending on what they are to be used for, there may be several sets of effective segmentation schemes. Individuals may fit into different segments depending on what is intended those segments to represent. The "best" scheme is that which allows the marketer to identify an opportunity and implement a plan. As such, the segments should have the following characteristics:

Market segments must be fairly homogeneous in regards to key characteristics. By this we seek groups of customers who have more in common than members of other market segments. Ultimately, segmentation is intended to help focus programs on things that will change behavior. As such, the segments should be predictors of behavior.

The groups of respondents defined as segments must be identifiable in means other than their beliefs and perspectives. One must be able to target programs to them. For that purpose they must be reachable or targetable. And finally, effective segments must be useful. That is, the segments should allow for modifying behavior. They must be approachable and allow a means of conveying change. Often this requires some degree of "isolation" where independent programs can be applied.

There are two general methods of identifying segments from market and marketing research data based on how the criteria are selected: (1) prior segments are based on imposing specific criteria on the data, and (2) derived segments are identified from the research data itself.

Prior segmentation is usually based on convenient and historically accepted criteria of segmentation. These usually follow traditional definitions of the market and therefore, usually represents the current perspective of the market and the business organization. Derived segmentation focuses on commonality of the market based on research information. These segments are identified using statistical analysis of quantitative research data.

The key in derived segmentation is the use of many variables. The goal is to identify groups of respondents with common characteristics. Because we can select different characteristics to use in forming segments, there are many types of derived segmentation. None are necessarily "correct," nor is there often a single overwhelming form of segmentation that describes a market. The key goal is to identify homogeneous groups of respondents for which a strategy can be directed.

Types of segmentation that should be examined include:

As previously noted, identifying derived segments is as much of an art form as it is statistics. The resulting segments must have some degree of "face" validity; that is, the resulting segments must be "reasonable." There are a number of different methods that can be employed as well as choices in variables to use.

The most common multivariate statistical tools for segmentation are referred to as cluster analyses where the segments are identified as clusters of respondents with common characteristics. Cluster analysis tools are at present not available on standard spreadsheets packages such as Microsoft EXCEL or Lotus 1-2-3 but are available on separate statistical packages, such as SYSTAT, SPSS, and SAS. Traditional methods identify clusters by assigning respondents uniquely to specific groups based on various rules. These segments are referred to as hard clusters since they are unique and mutually exclusive segments. Strategically hard clusters are used for sales force assignments. The figure below shows the average values of variables within hard clustering example..

Alternatively respondents can be considered to participate in a number of overlapping clusters. These are referred as soft clusters. Soft clusters involve identifying appropriate "archetype" descriptors of the sample. Each respondent is then assigned a percentage participation in these archetypes based on a "distance or loading" on them. We have found soft clusters to be more effective for communications and new product strategy formulation than hard clusters. Soft clusters can be computed from any hard cluster structure. Selection of which to use depends on the applications and the distinction among clusters..

There are four sets of common methods used for clustering (1) K-Means, (2) hierarchical, (3) regression, and (4) Q-factors. Traditionally K-Means and hierarchical have been used with survey research data. Regression and Q-factor methods are used for special applications and are not typically used with standard business research data. K-Mean (also referred to as "Quick Cluster" in SPSS and "FastClus" in SAS) is based on assigning clusters by averaged attribute values and is capable of handling large databases. Until recently, it was the only practical clustering procedure for marketing research data. However, K-Means clustering requires normalized data as well as other constraints common to all methods of clustering. This has become a particularly difficult problem with rating data where normalization loses the distinction between uniformly poor and high ratings.

Hierarchical clustering procedures are highly robust methods based on identifying clusters by distances between points and among clusters. Until the advent of high speed personal computers these methods were mainly used by academics. However, today it is feasible and desirable to use them on standard marketing research data. The results are similar to K-Means but do not require normalization of data and allow for exploring various ways of distinguishing clusters.

In general, the selection of the number of clusters is "arbitrary" in that it is chosen by the analyst and is part of the art. A useful tool for identifying the uniqueness of the clusters and selecting the appropriate number is mapping the results. This can be done by using triangle graphs of the principal variable groups for three or four clusters as shown below.

Alternatively reduced dimensional plots based either on Factor Analysis or Multiple Dimensional Scaling (MDS) can be used. We recommend MDS for these applications even though it is significantly more difficult because the method uses the same distance measures as Clustering. Typical results are shown below.