Business Drivers Multidimensional scaling and factor analysis are a couple of different statistical techniques that can be used to understand the drivers of a business. MDS is a technique that renders multiple data points in a relational manner -- data points that end up closer to one another are expressed visually as such, while those that are fairly far from each other in the key variable are expressed as being far. Factor analysis seeks to learn look at a set of different variables and identify the latent variable. This is the underlying variable that is driving the other variables, something that has a lot of value for a business.

Multidimensional Scaling

Multidimensional scaling is "a set of data analysis techniques that display the structure of distance-like data as a geometrical picture" (Young, 1985). There are a number of different techniques as part of multidimensional scaling, and this paper will outline them, as well as some other statistical techniques that are used in business analysis.

The most classic form of multidimensional scaling is the city grid, which outlines the distances between different cities. For example:

Source: Borgatti (1997)

The basic principle of multidimensional scaling is to produce an image that illustrates multiple dimensions at once. So on the above chart, the distances between these nine cities is reflected on the chart. The nine cities have 36 pairings, and each is rendered on the table. So multidimensional scaling is typically used to convey multiple data points in this sort of way, so that there can be easy comparisons between them. It is just as easy to render the distance between Seattle and LA as it is between New York and Miami on this table.

MDS would be able to put this on a map. The map tells you nothing - it's a map of the U.S. without lines -- but the more common business application of the concept is the perceptual maps used in marketing. Consider the following example of computers. In this example, the key variables are the perceptions of price and quality. Consumers would be surveyed with respect to their perceptions, using some sort of numeric value or Likert scale. That information would be translated into numbers for input onto the map. The value of the perceptual map shows the perceptions that consumers have of these different brands on a scale that illustrates the closeness of brands to each other. Brands can be clustered when they are similar, or set completely apart of all other brands if they are outliers.

One thing about this technique is that it can be used without quantitative inputs. For example, there is a "badness of fit" technique that can be used to translate qualitative data into quantitative for input into the perceptual map. In the computer example, price can be objective, but quality is subjective, and the map allows for the comparison between the two. It is important, however, to ensure that there is a specific technique that will be used to translate the qualitative data into quantitative.

There are some similarities between this and factor analysis. Factor analysis is another technique that looks at relationships between variables, but with factor analysis the objective is to examine the relationship that variables have with an underlying variable. In marketing, this would be something like trying to pin down a target market. There are a number of demographic and psychographic variables that the company would use, but there could be a single underlying variable that explains this. So what factor analysis does is to try and explain the relationship between a set of different correlated variables.

Think about the market for craft beer. A map of a city will show geographic hot spots for craft beer, particular areas that are positive outliers in terms of sales. The similarity between those neighborhoods will be found somewhere in the demographic, which would be the latent variable. Those neighborhoods are all outliers compared with nearby areas because they have more of the craft beer demographic. By determining the latent variable, a business can then investigate other areas with the same latent variable, and use that to predict maybe the next hot spots to emerge. This could be, say, an area in London, or Tokyo, or Sao Paulo with a similar demographic. One could use this information to get to the target market before the competition does, making it a good form of business intelligence to understand how to identify key latent variables...

There might also be other factors that are related -- determining what factors to test for is important. In some countries it might be a combination of the right demographics along with a large population of Western ex-pats who act as early adopters, just as an example.
But there are significant differences between multidimensional scaling and factor analysis. As a means of investigating business drivers, factor analysis is powerful because it can help identify the latent variables that are really driving business. This is different from what multidimensional scaling offers, which is essentially a way to visualize the way that the world is, or the way the world is perceived. Such perceptions are informative in their own way -- you can learn that customers perceive a branch much differently than you may have thought, for example. This information can be used to help make business decisions. For example, maybe the HP purchase of Compaq could be justified in the sense that the company did not want to use the HP name on low-end computers, but it wanted to capture part of the low end market. That is not why they bought Compaq, but that is an example of the sort of decision that a company can make with multidimensional scaling. Factor analysis would teach something different, perhaps identifying why some computers are perceived as low quality. It might be because of the price, but it might be because those companies have less pedigree in the market.

Cluster analysis is more similar. With cluster analysis, data points are displayed on a two-dimensional scale. There are clusters where there are large groups of similar items. For example, in a map of restaurants, there will be one cluster with fast food chains, another nearby for fast casual, then a bigger gap towards fine dining. The different types of restaurants will be clusters, and within the cluster there will be many different restaurants.

Cluster analysis is different, however, is that it has more data points. It can also be used to identify outliers, however, so in that sense it can be used in much the same way. But cluster analysis is specifically looking to identify clusters that exist within the data, identify what data is in the cluster, and just as important what data points are outside of the cluster. The visual rendering is similar, but the overall effect is slightly different because of the role of clusters, and there is also the need for more data points to make the clustering effective.

Real World Examples

The most common example is in product positioning. Multidimensional scaling is typically used by companies to identify target areas for new products. Ideally this means finding ways to enter businesses where opportunity lies. Getting back to the craft beer example, AB InBev is buying craft breweries frequently of late. A competitor might want to understand why. A perceptual map can help with that. It could be said that many of these share fairly similar characteristics -- a good brand in their region, relatively strong market share, and good reputation for quality. These characteristics just happen to be ideal for scaling up the brand, which is the relative strength of the large multinational brewer. If an MDS map of the different craft breweries in each city was put together, the ones that AB InBev bought would basically fit within the same area of each map. An executive in the industry might be able to not only figure out who the next buyout targets are, but also figure out what the end game is with respect to where that strategy leads. This can be a powerful type of business intelligence when it leads to unique insights.

The technique can be applied for new products, too. I would consider using this technique for a new product launch, in particular to identify a niche that has not been exploited adequately. As an example, the cell phone maker Xiaomi is a relatively new company from China. In a market where two companies seek out the high end of the market and everybody else is fighting for share with cheap phones, Xiaomi saw room at the very low end. This would be at the lower left of the perceptual map. They were smart enough to recognize that there are a lot of consumers who would not accept low quality, but who could afford nothing better. So they build their business model pursuing that market, and are now one of the top five cell phone producers. The cell phone market has those giant gaps that can be exploited. HTC basically hangs out in the middle…