Doing a research project in my summer internship with a telecom major, and having collected data through surveys I’m now left with the most important step i.e. data analysis. With the objective to make the best sense of the collected data in a way that adds value to the organization, recently I was attracted to a number of statistical methods. To be honest the first few days were pretty intimidating with lots of technical words and relatively tougher language than I am used to. Having managed to read some part of it, I thought it would be a good idea to put a few things in a way that may be helpful for those who want to take a dive into these statistical methods.
A detailed analysis of each method needs quite a lot of extensive reading and this article in no way will help you in actually using these methods, they may only help to get a basic idea and how these methods can be applied to take marketing decisions.
Cluster Analysis:
Cluster analysis is a technique to group similar objects (an object is the entity whose attributes are being measured). A simple example to understand what cluster analysis does is to see the people around us. Take for example your city. There are several ways in which you can group people in your city. Age can be one, so you will have children, youth, middle-aged, old people as groups of people. Similarly you can have height, income, location (measurable parameters or variables); and degree of happiness, fashion-sense (non-measurable or categorical or nominal variables) on which you can group people. That’s essentially what clustering does for you.
You must be wondering that if cluster analysis is just grouping of objects on the basis of some criterion, what’s the big deal about it (even MS Excel can do it). Imagine a situation where you want to group people on the basis of three variables, let’s say income, age and preference to a particular brand. If you think this is complex then that’s the beauty of cluster analysis.
The idea is simple. Decide the number of variables on which cluster analysis is to be done. Then plot the objects in an n-dimension space with each variable as one of the dimensions. Using the distance (Euclidean) between two points, find out which points lie close to each other in the n-dimension space. Accordingly, cluster the objects into various groups. If the variables are nominal instead of distance, matching is used to determine the closeness.
The next important question is: how do we decide what is the ideal the number of clusters or groups? Well, there is no rule to this decision. It solely depends on the researcher’s discretion and the problem at hand. In fact there are two ways of doing it.
One way is to decide the number of clusters beforehand (k-means cluster in SPSS). For example, you are using income as the grouping criterion and you want the people to be grouped in only three groups, i.e. low, medium and high. In such cases, you fix the number of groups or clusters before doing the analysis.
A word of caution here is that there are various algorithms for cluster analysis and they give different results for the same data set, so as said you have to be careful to make sure that the results make sense.
For those who are familiar with Discriminant analysis and Factor analysis, there are reasons to think that these methods are similar with cluster analysis. Although all these techniques have a common impact and that is grouping of similar objects / variables, there is significant difference between them.
Discriminant analysis is a technique to group objects on the basis of a known criterion. It's objective is to find a relationship among variables to group the objects as per the criterion.
Factor analysis on the other hand, though being a grouping technique has the main objective of data reduction i.e reducing a large number of variable into a lesser number of factors .
Cluster analysis, on the contrary is about finding a criteria of grouping the objects / variable unlike discriminant analysis and the objective is never to reduce the data but view them as groups with similar objects / variables.
Now as we move ahead with multidimensional scaling, you may try to figure the ways in which this method can be used to simplify our marketing decisions. At the end, we will discuss these applications.
Multidimensional Scaling:
Multidimensional scaling (MDS) is a very powerful tool to graphically understand the differences / similarities or preferences of objects / variables. As the name suggests, it’s plotting of objects / variable on more than one dimension in a way that represents their differences. They are often called perceptual maps because they represent the perceptions of various objects graphically. Here dimensions are not simply axes in space but they are variables (attributes) on which the comparison is being made.
A simple example to illustrate this technique would be to make paired comparison between say 5 brands of soft drinks. Say we ask 45 respondents to take two brands at a time and comment which brand they prefer. As a result, we get 5 X 5 matrix; with each element in the matrix representing the number of times one brand is preferred over the other.
| Brand 1 | Brand 2 | Brand 3 | Brand 4 | Brand 5 | |
| Brand 1 | - | 23 | 12 | 15 | |
| Brand 2 | 22 | - | 34 | 33 | |
| Brand 3 | 33 | 11 | - | 19 | |
| Brand 4 | 30 | 12 | 26 | - | |
| Brand 5 | | | | | - |
The element in red color indicates that 23 respondents prefer brand 1 over brand 2.
Now what MDS does is that it converts the preferences of brands into distances and plots the various brands on a multidimensional space what we call as perceptual map.
I can understand that there are a lot of questions that must have come to your mind. We will try to answer each of them one by one.
The first question surely would be how we get a multidimensional space and how is the distance representative of the preference of one brand over the other. When a respondent prefers one brand over the other, subconsciously he uses some variables (attributes of soft drinks) on the basis of which he makes a comparison and chooses one brand over another. These variables sometimes may not be even perceived by the researcher. For example in case of soft drinks, sweetness can be perceived differently for differently flavored drinks by consumers though they may have the same content of sugar.
Whatever the variable (attribute) is, the bottom line is that the various variables over which the respondents differentiate the various soft drink brands become the dimensions of the perceptual map.
The next question would be how do we decide the correct number of dimensions of the perceptual map? In MDS we choose the minimum number of dimensions through which we can explain the 5 X 5 matrix completely i.e. when we plot each brand on the perceptual map, the position of each brand and their distances should confirm the results of the paired comparison of brands preferences. Generally more than 3-dimensions should be avoided as they are very difficult to interpret.
Ok! So how do we decide what each dimension of the perceptual map stands for? As we discussed each dimension stands for an attribute of the brand. To find out what these attributes are, the researcher has to use his judgment. In fact it’s a good way to find out what are the attributes over which a consumer differentiates the soft drink and forms a perception of that brand.
We can also use ordinal data (rank) of different brands to carry MDS and get similar results.
This is one of the basic applications of MDS. According to Green, Tull and Albaum in the book “Research for Marketing Decisions”, there are two more ways in which MDS can be used.
The other two are Unfolding type-joint space model and Projection type-joint space model.
Unfolding type, Joint space model does exactly what the basic model discussed just now do with an added feature. It also plots objects (in our example respondents) on the perceptual map such that the objects lie close to the brands of their preference. As you can make out, this exercise does two things for you. It not only helps you understand the image of your brand vis-à-vis the other brands on those attributes, it also helps you identify the segment of the market that prefers your brand and target them.
Projection type, Joint space model is one more step ahead of the two models. We add one more angle to our experiment of soft drinks. We ask the respondents to rate / rank the various brands on four attribute namely sweetness, tanginess, fizz and taste in the order of preference. Projection type MDS would result in a perceptual map which would indicate how preferred each brand is on the four attributes. Each brand would be a point in the map with the four attributes as vectors and the distance of each brand from vector would indicate the preference of that brand on that attribute.
Application in Marketing Decisions
I am sure by now you would have understood the beauty of these techniques in solving the marketing issues and taking decisions.
Let’s reinforce our understanding by a simple example which is very close to what I did for my summer internship. We are talking about the offering corporate solutions to SMEs and corporate by telecom players. Suppose we asked customers to:
1. Rank six parameters namely Network, Price, Services, Sales Executives, Company’s reputation and Product line in order of preference as drivers to choose a service provider
2. Rank the various telecom service providers in order of preference on these six parameters
Using the data collected from second question, we can apply MDS technique (Projection type, Joint space model) to draw a perceptual map describing how each brand fared on the six parameters discussed. We will be able to know how my company is perceived on these parameters vis-à-vis the other service provider. We will be able to know the parameters on which my company fares better than other service providers. We can also identify space in the perceptual map that is free and can be used as effective positioning of your brand. We can also identify those parameters on which our company is less preferred vis-a-vis other service providers and improve on those.
Now suppose my company has done very well in network and product line. We would definitely want to identify those customers that consider these parameters as important preference driver for choosing a service provider i.e. they consider network and product line as most important parameters while choosing a service provider.
There are two way of achieving this objective.
1. One way would be to use MDS (Unfolding type, Joint space model) technique on the data obtained from first question. Using this technique you would be able to plot the six parameters on perceptual map with the customers plotted in such a way that the customers are closer to those parameters they consider more important. We can identify those customers and then run a cluster analysis taking revenue, location, sector etc as variables to identify segment(s) that can be targeted.
2. Alternatively, we can run a Discriminant analysis straight on the data obtained from first question with variables “network, product line and sector” or “network, product line and location” to identify groups of customers who give most importance to network and product line and are in the same sector or location. We can target these groups to get successful results.
We see that Cluster analysis and Multidimensional scaling are very effective techniques for positioning and segmentation decisions. For those who want to take a deeper dive into the techniques can go through the following books:
1. Research for Marketing Decisions – Green, Tull, Albaum
2. Multivariate Data Analysis - Joseph F. Hair, Bill Black, Rolph E. Anderson, Barry J. Babin, Ronald L. Tatham
3. SPSS 17.0 Guide to Data Analysis- by Marija J. Norusis
Happy reading!
Posts
Nicely written....
ReplyDelete