Conjoint Analysis Technical Notes I

Conjoint analysis is a technique for measuring, analyzing, and predicting customers’ responses to new products and to new features of existing products. It enables companies to decompose customers’ preferences for products and services (typically provided as descriptions or visual images) into “part-worth” (or utilities) associated with each level of each attribute of the product. They can then recombine the part-worths to predict customers’ preferences for any possible combination of attribute levels, and the likely market share or revenue that a new product is likely to achieve when introduced into a market in which other competing products may already be available. They can also use conjoint analysis to determine the optimal product concept or to identify market segments that value a particular product concept highly.

There are three stages in a typical conjoint study:

(1) Design of a data collecting instrument,

(2) Collecting data from consumers, and

(3) Analyzing the data and simulating market response.

Designing a conjoint study

Conjoint Analysis starts with the premise that a product category could be described as a set of attributes. For example, pizzas could be considered to have the following attributes: size, brand, type of crust, topping, amount of cheese, type of sauce, price, etc. Every pizza could then be described as a combination of levels of those product attributes; for example, large Papa John's thick crust pepperoni pizza with extra cheese and tomato sauce, priced at $12.95.

The objective of the design stage is to specify a set of product bundles for which we obtain customers' overall evaluations, in such a way that those evaluations could then be decomposed into the part-worth value that each customer attaches to each level of each attribute. To develop such a design is not a simple task. For example, if there are 6 attributes, each with four possible levels, then we could create 46 (= 4096) different products. It is not reasonable to ask each customer to evaluate all of those bundles. Instead, in this case, if a customer rates as few as 25 product bundles, that is sufficient for estimating the part-worths for the attribute levels. One way to select the product bundles is to ensure that they satisfy an orthogonality constraint. This means that across the selected product bundles, each level of an attribute combines in roughly the same proportion with the levels of other attributes. In other words, if we select any two attributes A and B, then the probability of finding the attribute level Bi in a product bundle is the same irrespective of the particular attribute level of Aj found in that product bundle. One of the common methods of finding such orthogonal combinations is through the "Addelman" designs The number of independent parameters to be estimated is equal to:

where N is the number of attributes and ni is the number of levels of attribute i. For each product attribute we can arbitrarily set the lowest utility value (say, equal to zero). We can also arbitrarily set the maximum total utility from any product (say, equal to 100).

In some circumstances orthogonal designs can result in unrealistic products, such as when respondents perceive some of the attributes used in the study to be correlated—automobile horsepower (hp) and gas mileage (mpg) typically have a high negative correlation, but orthogonal designs could result in hypothetical products that combine high hp with unrealistically high levels of mpg. If a product is unrealistic in an orthogonal combination, there are several possible remedies: (1) We can combine the attributes and develop a new set of levels for the combined attribute. (For example, hp and mpg might be combined into a “performance” attribute with high performance associated with high hp and low mpg, and low performance associated with low hp and high mpg.) (2) We can replace unrealistic products by substituting other combinations (perhaps generated randomly, but not duplicating the retained combinations). While this approach compromises orthogonality, it will rarely affect the estimated utility functions significantly if we replace only a few bundles (say, less than five percent). (3) We can select other orthogonal combinations (although this remedy requires special expertise).

An additional consideration in developing a suitable design is the exact nature of the data collection instrument. There are several options here, including, (1) asking respondents to sort and rank-order a set of cards, each containing a description of a product bundle, (2) asking respondents to rate each product bundle, say on a scale of 0 to 100, to reflect their likelihood of buying that product, (3) presenting respondents a sequence of product bundles, two at a time, and asking them to assign 100 points between them, and (4) offering respondents a sequence of sets of product bundles and asking them to choose one product from each set. Each of these data collection options has an associated set of costs and benefits.

#Marketing #ConjointAnalysis #MarketingAnalytics