If the stimuli (cards) have been rated on an interval scale, such as 0 to 100 (or if 100 points has been allocated across all cards) then ordinary least squares (OLS) regression analysis can be used to compute utilities. If cards have been ranked, then a special algorithm such as monotonic analysis of variance (MONANOVA) or LINMAP is used to transform the ordinal rank order data. OLS usually gives satisfactory results with ranked data. Models are usually developed for each respondent.
With choice modelling, the multinomial logit model is usually used, estimated by maximum likelihood (see "Discrete Choice Analysis" by Ben-Akiva and Lerman for full details of the formulas used). Models are usually developed with aggregate data.
The coding employed for the design variables can be either part-worth or vector (ideal point is a compromise between the two).
Continuous variables such as price are most efficiently represented as a vector (eg $1.00 as -1.0, $1.50 as 0.0, $2.00 as 1.0).
Nominal variables are best represented as a set of binary variables or dummies. For example, three brands could be represented as three binary variables (eg Kellogg as 0 0 1, Uncle Tobys as 0 1 0, Sanitarium as 1 0 0). If using regression with a constant, use one less dummies than levels to avoid multicollinearity (eg Kellogg as 0 1, Uncle Tobys as 1 0, Sanitarium as 0 0 - Sanitarium is then equal to the constant).
To conduct market simulations, a choice rule is needed.
The maximum utility rule assumes that each consumer chooses the product with the highest utility to them out of the products available. This is most appropriate for a high involvement product decision.
The share of utility rule allows that consumers do not always choose the product with the highest utility (they may seek variety, for example). Each product is given a share proportional to the individuals utility for each product.
The logit choice rule is similar to the share of utility rule but allows for random error in translating utilities into choice.
When the number of attributes and/or levels of the attributes is high, the
number of profiles (or cards) in the design becomes too cumbersome for
respondents. Adaptive Conjoint Analysis (ACA), developed by Sawtooth
Software uses self-explicated ratings on attributes to weed out the less
important ones - this is different for each respondent. A reduced set of
attributes is then used to design the profiles for each respondent. One
criticism of this approach is that it may not be realistic of the way consumers
make choices. Another is that one attribute, price for example, may not be
important relative to others and so be omitted for some respondents - but that
attribute could be decisive when everything else is equal, as is often the case.
Sawtooth Software
Choice modelling is often criticised for being based on what people say they would do, rather than what they actually do. Usually, choice modelling is employed when there is no actual data to study - for example a new product - so there are few alternatives. However, it has been possible in some studies where the new product has been launched to validate choice modelling - and a well-designed study has been proven to predict very accurately.
Another validation is comparing the price elasticity from a choice model to that from time series analysis. Again, where the comparison has been possible choice modelling has been validated.
The secret is to make the respondent task as realistic as possible. Also, we need to remember that as part of the experiment we made respondents aware of the alternatives and made them consider the alternatives. So the model predictions should be considered as "potential". Adjustment of potential to actual involves discounting for the likely levels of awareness and consideration to be achieved in practice.
"Conjoint Analysis in Marketing: New Developments With Implications for Research and Practice" by Paul E. Green & V. Srinivasan, Journal of Marketing, October 1990. A classic overview.
"Conjoint Analysis and Choice Modeling Consideration" Marketing Research, Spring 1997. A good forum comprising three articles:
"Using the Diagnostics of Dynamic Choice Models to Manage or Defend Against New Product Launches" by John Roberts and Charlie Nelson, Australasian Journal of Market Research, January 1996. A case study based on an actual application.
"Marketing Engineering" by Gary L. Lilien and Arvind Rangaswamy, published by Addison-Wesley is an excellent book for learning about a wide range of marketing models including conjoint analysis. The book comes with a software package that allows students to use limited versions of these models, again including conjoint analysis.
| Back | Home |