!DOCTYPE html PUBLIC "-//W3C//DTD XHTML 1.0 Strict//EN" "http://www.w3.org/TR/xhtml1/DTD/xhtml1-strict.dtd"> Streamline Training & Documentation: Customer Segmentation via MaxDiff and Latent Class Analysis

Saturday, April 04, 2009

Customer Segmentation via MaxDiff and Latent Class Analysis

Since I'm currently working with colleagues on a product designed to assist outsource companies in doing a good job of satisfying the firms with which they contract, a short piece on identifying customer preferences in the April issue of the Harvard Business Review caught my eye.

Eric Almquist, a partner at Bain & Company, and Jason Lee, a manager in Bain's Customer Insights Group, describe the Maximum Difference technique for homing in on the relative importance to customers of different aspects of the customer experience a vendor delivers (or that the vendor is considering delivering, in the case of products and services under development).

As Almquist and Lee explain, the MaxDiff technique was developed by Jordan Louviere, currently a professor at the University of Technology Sydney. The technique forces customers to discriminate among product attributes, as opposed to rating them all as more or less equally desirable or undesirable.

There are three basic steps to MaxDiff:
  1. List the product/service attributes whose relative utility to customers you are investigating.

  2. Present respondents with sets (one at a time) of three to six attributes, asking them to select which atribute in each set they prefer most and which least. The sets of attributes are not mutually exclusive, i.e., any given attribute appears as often as necessary in different mixed groupings to allow inference of the attributes' relative utility to the customers surveyed.

    Note that the MaxDiff technique is valid in cross-cultural studies because it does not use a rating scale (Likert scale). Thus, the issue of people in some countries tending to choose higher ratings on average than people in other counries does not arise.

  3. Using statistical analysis, the details of which I won't attempt to describe, draw up a list of the attributes ranked according to the customers' responses.
A likely follow-on, which Almquist and Lee only touch on in their piece, is identifying customer segments that reflect different patterns of preferences. For guidance on segmentation, you can turn to Latent Class Analysis (LCA), described in a 2003 technical paper published by Sawtooth Software, a company in Washington State that develops software that marketers can use for interviewing and data analysis. The classes identified — in this context, customer segments — are "latent" because they are not directly observed. Instead, statistical analysis is used to infer the characteristics of classes within the population under study.

Once the classes have been defined, one can determine to which class any particular individual is most likely to belong by looking at the individual's characteristics.

The paper concludes with a well-presented example of combined use of MaxDiff and LCA by a multinational company that wanted "to identify key leverage points for new product design and marketing messaging."

A detailed FAQ on LCA put together by John Uebersax is here.