Conjoint demonstration - technical notes
The Conjoint Demonstration is a simplified example of a conjoint exercise. It uses simplified calculations to estimate the "utility" values, or part-worths associated with the choices being made as shown on the bars. There are actually several "flavours" of conjoint analysis depending on the task at hand.
The conjoint demonstration on this site is a very basic version of conjoint analysis simply to show the main principles of conjoint analysis.
The form of conjoint analysis used relies on a pairwise comparison (ie Pizza A v Pizza B) rather than selecting from more than two products, using just two attributes at a time. This is known as a limited profile set and is similar to Adaptive Conjoint Analysis (ACA) from Sawtooth Software, the most popular software for conjoint analysis.
Other forms of conjoint analysis include full-profile and choice-based conjoint analysis which use different way of presenting the description of the products and allow for more products to be seen at one time.
The demonstration here is heavily reliant on consistent choices being made, as utility values (the values shown in the graphs) are simplistically calculated based on the last choice made, rather than a probabilistic average of all answers. Consequently inconsistent choices can produce strong variations in value scores. Full conjoint analysis (eg ACA from Sawtooth Software, or choice-based conjoint) are more robust in handling such inconsistent choices.
Secondly, because the demonstration only has 3 attributes (topping, size and base type) no special schema is necessary to select the pairs of attributes used to make up each choice. A core challenge in using conjoint analysis is in minimizing the number of choices per respondent, yet still getting good estimates of utilities. This is most commonly achieved using what are known as orthogonal designs.
In a conjoint analysis with 5 attributes each of 4 levels for instance there would be 4x4x4x4x4 possible combinations of products to show to a respondent. By using an orthogonal design a minimal set of profiles is needed (typically 12-24) by chosing the profile designs so that levels are shown in combination in such a way that the effect of each level can be estimated without needing to show all combinations. This is a key part of any large scale experimental design seeking to examine underlying factors.
ACA (Adaptive Conjoint Analysis) helps this minimization by asking for the levels in each attribute to be placed in rank order and using this information to help in selecting pairs. However, with so few levels here, there is no need to pre-order the levels for each attribute as there is in ACA.
For further information about how the demonstration works or the technical issues involved, email email@example.com