Analytical harakiri-Ignoring latent class models

My black books are out and here are some notes that were jotted down on projects;

• Scenario 1: You're working on an analysis to study drivers of purchase intent and you keep feeling that different sets of of customers may have different drivers. The client however, has not asked for segmenting the customer base.
• Scenario 2: Your data is a mix of continuous, ordinal and nominal variables, yet you continue to try and segment your customers by using some distance metrics. You wonder why two customers are slotted in different segments though there is not much difference in their average spend or other variables?
• Scenario 3: You know the major themes that your client's brand stands for but are unable to fully break them out from the data into meaningful, actionable sets and subsets for understanding or positioning purposes.

The reason why researchers and analysts continue to grapple with these issues in market research data is because they are not getting the most from their analysis. One reason I love statisticians in social science is because they are the 'early adopters' for most new and emerging techniques in statistics. They quickly work on and distill new learning's in the area to their everyday problems and thus boldly go where the rest of us are timid or lazy to venture.

How do we get more from our analyses in the areas of regression, choice, factor and segmentation problems? Latent class or finite mixture models are the answer. These set of models differ from the traditional models in their structure due to the inclusion of one or more discrete latent(unobserved) variables into the model relationship along with the observed. Thus categories/classes of these unobserved variables are interpreted as latent segments.

Some key advantages of latent class models are:

1. They are less affected by data not conforming to modeling assumptions(linear relationship, normal distribution, homogeneity)
2. They work with mixed scale(continuous, ordinal, nominal, count) type variables in the same analysis
3. They are able to simultaneously do two analyses together i.e. segment and predict thus eliminating the need for two steps in an analysis

In the area of segmentation, these models bring in a model based approach and an ability to accommodate categorical and continuous data and predictive and descriptive segmentation under a common modeling framework. This leads to far superior and insightful results in testing and estimation of market size and structure and profile of market segments. The probability based segment prediction criteria provides a more realistic picture of market reality since consumers can belong to more than one segment at a time. Some areas that latent class segmentation models should be used to 'do more' with the data are classical segmentation problems including descriptive ones, global segmentation and studying change in segments over time.

When working with factor analyses, latent class factor models are better able to make composites out of variables because they handle non-continuous data in a more elegant way. Plots and perceptual maps generated from the analysis score over the traditional technique because the factor scores have probabilistic interpretations. The categorical nature of the composites allows a more holistic extraction of themes and sub themes and helps in developing a more precise brand positioning strategy. Use of fewer variables to form factors is an added plus. Applying latent class factor models to attribute data(with ordinal and categorical scales) is an absolutely delightful exercise once the results are compared to traditional factor analysis. The new models provide a more accurate and vivid picture of the brand/product.

Very often analysts pass off the low predictive power of a regression model to a 'lack of all the right explanatory variables' and tell clients that the model may have been better if they had more data. What they don't check is that the same model may not hold for all customers. Latent class regression modeling allows for a simultaneous segmentation and regression of the data thus unearthing latent segments that may have different regression equations and estimates. This makes the estimation more precise and gives clients a more informative way to look at a drivers analysis. Applications of latent class models in regression lie in the area of conjoint analysis, customer satisfaction studies, purchase intent drivers or any traditional regression model that benefits from explanation of unobserved variance in the data.

Latent class models thus represent powerful improvements in model building, prediction and insight generation over traditional approaches to segmentation, factor and regression analysis. They truly allow the data to talk much more and analysts need to take them mainstream by learning to use them and unleashing them on projects which can benefit from the same.

Segmentation-making it more science than art

The reason for delay in posting this has been because I've been toying with whether I should write on segmentation or not. So much has been written on this subject that it makes me a little hesitant about revisiting this space.

What got me to finally pen this was the title of a paper at an upcoming conference that said 'How statistics get in the way of actionable segmentation'. I don't know what the presenters have to say (must source a copy after the conference) but the title made me laugh. The two words that stuck in my head were 'statistics' and 'actionable segmentation' and whether the twain will ever meet.

I have undertaken enumerable segmentation projects in my career, some simple, others complex, yet others that go nowhere. All of them in the end have the same things in common:

1. Too many bases variables


2. Over reliance on cluster analysis as the primary tool for segmentation


3. Use of subjective judgement to evaluate results of the cluster solution


4. Lack of reliability and validity tests on the solution


5. Recreation of the scientific solution into a more 'creative and arty' one

What the above means for managers who implement segmentation solutions is that they could be formulating strategy and targeting segments that are unstable and unreal. There exists a body of research that calls for a deeper look at the statistics and data that go into cluster analysis and segmentation(I will be happy to provide the references).

The real issue continues to be an inability of both analysts and practitioners to put together a common road map for segmentation that takes into account statistical robustness of the technique along with creation of actionable segments that can be targeted through focused marketing programs. In my experience, the science of segmentation gets lost in the art.

Here are the five key things that analysts and practitioners must do to create better, robust and more scientific segments-

1. Choose bases variables for segmentation that tie in with the end goal of segmentation and keep their number not more than 8-10. Build a set of good profiling variables that tie into the bases variables(there is no restriction in number here).
2. Explore other tools for segmentation(sometimes simple business rules work just as well). Latent class analysis offers excellent alternatives for both survey and crm data and is still a highly underused technique. Try two techniques, if possible and compare and contrast results.
3. Use a variety of statistical parameters to evaluate a solution instead of relying on one or two or on subjectivity. Decide which metrics you want to look at before the study. For example-dendograms, change rate, psuedo Rsq, hotelling's Tsq can be some metrics for evaluating no. of clusters in a cluster solution. The BIC, p-value, parsimony(no. of parameters) and the bootstrap p-value can be the parameters to nail number of segments in a latent class segmentation. Reliance on 'many' statistics vs. 'few' should be the mantra.
4. Test reliability through hold out samples and validity through looking at profiling variables and how they differentiate the solution. The holdout sample results must match those of the developmental sample in terms of the number of segments and profiles. Most of the picked profiling variables must adequately differentiate the final segment solution. If there is an issue with reliability and validity-the solution may have a problem. Going back and reworking the same is the best way out.
5. Don't use the art of segmentation to sidestep the science for a solution, use it if you will to add to the same.