Cluster Analysis, a common tool for segmentation analysis, is a widely used, often misused, and widely misunderstood statistical technique.
The basic theory is very simple and has a long history, indeed it is probably a good idea to start with the history. Back in the 1930s Ronald Fisher (a polymath who is responsible for many of today’s statistical techniques, including analysis of variance, the F distribution, and maximum likelihood estimation) used cluster analysis and discriminant analysis with a set of data about the iris flower to show how measurements could be used to identify the number of species and which individual plants belonged to which species. This was the birth of what we now know as numerical taxonomy.What we are seeking to do with cluster analysis is to find groups within the data.
We want these groups to have two characteristics
- That all the members of the group are ‘similar’ to each other
- That the groups are ‘different’ to each other
However Cluster Analysis is different from many other marketing science techniques in two important ways
- What it is seeking to measure is almost never there
- It is not really a scientific, quantitative technique
In the rest of the post I explore these two points and then provide some thoughts and advice on using cluster analysis.
Groups Don’t Exist (Usually)
The image below shows what we would like to believe cluster analysis delivers us, where each of the dots represents a respondent, and where the respondents represent the population.
However, this almost never happens in market research. In market research, unlike biology, real groups tend not to exist and a better metaphor for our data is the diagram below:
The clustering software finds three clusters (it almost always finds cluster, irrespective of whether they exist), if we ask it to, and the means of the three clusters are different, but the borders are clearly arbitrary and some people could easily be in a different cluster.
However, the fact that ‘real’ clusters do not normally exist in market research is not necessarily a problem. We group things because it makes them easier to think about. In the case of the data above if we left them as one group we would produce one product or one service targeted at the mid-point, Henry Ford’s any colour you want as long as it is black. By creating three groups, even though they are artificial, we can tailor our product offering so that more people have something closer to what they want.
A great example of this sort of structuring is the music industry. There can be long discussions and disagreements about which genre a particular artist is in, are they kerang or heavy metal, are they folk or country etc. However, the structure does enable the music industry and the music buying public to ‘home in’ on artists they are more likely to appreciate.
The Qualitative Nature of Cluster Analysis
When we think of cluster analysis we think of large sample size, complicated software (from respected software house such as SPSS) and it seems the epitome of a quantitative, scientific procedure. However, it falls well short of that aspiration and is best thought of as a qualitative process, much as say grounded theory is a qualitative process of analysing discussions. In this sense we are using the term qualitative in its constructionist and reflexive sense of using the knowledge of the researcher to construct an interpretation of the data that is helpful, without assuming that the ‘right’ answer can be found, or even that there is a ‘right’ answer.
The first issue with cluster analysis is that it is very unstable. If you re-run the data with a few cases added or deleted the solution can be very different. Indeed with some software simply re-ordering the cases in the data file will produce a different result.
The second issue is that there are no definitive statistics to tell you how many clusters there are and how good the solutions are. There are stats but they are all rule of thumb and very little weight can truly be put in them. In the classic textbooks there is often a reference to a scree chart and instructions to look for the knee. The scree chart shows how much improvement each additional group has added, usually measuring some ratio of within group similarity to between group differences. However, in my experience the scree charts rarely show nice knees.
The third big issue is the choice of which attributes to use in the clustering. If you use demographic questions (e.g age and sex) you will get back groups that are demographically differentiated, but did you need a computer to tell you that young women and old men are different demographically? If you overload one type of attribute then it will dominate the solution. Deciding which attributes are predictors of group membership (and should therefore be used) and which are descriptors (and therefore should not be used) is more about experience and knowledge of the market than science.
Some thoughts and tips
So, here are a few thoughts and tips on what I would advise when running cluster analyses.
- Don’t forget who is the boss, the software is there to help you find a plausible and useful solution. Use concepts such as triangulation to see if the clusters produced ‘work’. Think about people you know, could you put them into the clusters you are suggesting. Look at the open-ended responses by cluster and see if they ring true.
- The thinking starts when the survey is being designed. With cluster analysis you would like your variables to have been collected on a similar scale (adjusting later is not as good) and you do not want missing data. If you have missing data you are either going to leave those respondents out (bad) or you are going to have ‘guess’ what they would have said (e.g. apply means) and this makes the solution even greyer.
- Do not factor analyse the data and then cluster factor scores (this used to be a recommended process). For the cluster analysis to work it needs what marketing scientist Jon Pinnell describes as ‘lumpiness’ in the data. Factor scores smooth that lumpiness out. By all means run a factor analysis to choose a subset of attributes to use, but don’t use factor scores
- Concentrate on how many clusters your client can use, if the 3 group and the 8 group seem equally good, then normally you will want to favour the 3 group as it will in most cases be easier to use in the market place.
- You will usually get quite a large group who seem to have not particularly strong views, for or against. This is real, some people barely know what brand they buy and have very few feelings about the category.