Using experimental auctions for marketing applications: A discussion

Journal of Agricultural and Applied Economics, Aug 2003 by Lusk, Jayson L

So how might one handle this problem? One solution is to only auction a single good or attribute in a single bidding round (e.g., Lusk et al., 2001b). However, this solution is overly restrictive, in that it eliminates within-subject comparisons and does not allow the researcher to easily examine the value of multiple product attributes. The easiest way to handle this problem is to use random drawings to determine a binding bidding round and/or a binding good. For example, Hayes et al. had subjects participate in 20 bidding rounds in which only one was drawn as binding, and Lusk, Feldkamp, and Schroeder auctioned five different types of steaks and randomly drew only one of the steak types as binding. Under the assumption that a subject's expected utility is linear in probabilities, randomly drawing a binding auction should produce the same results as conducting a single independent auction. Roosen et al. and Lusk, Feldkamp, and Schroeder provide results to suggest that individuals' behavior in these sorts of applications is consistent with this assumption.

Multiple Attribute Valuation6

Many researchers use EA because they are interested in a particular product characteristic or attribute. For example, one might be interested in the relative value of an organic tomato versus traditional tomato, the relative value of a tender versus tough steak, the relative value of a bag of non-GM versus GM corn chips, etc. As such, EA are often constructed to estimate the value of a single product characteristic. However, it is important to recognize that foods are composed of numerous attributes in addition to the one attribute that may be of research interest. Thus, there may be a number of potential problems with the approach of valuing single product attributes.

First, the product attribute of interest may have a number of substitutes. If not properly accounted for, this substitutability can create incorrect inferences about the valuations one is attempting to measure. For the sake of discussion, assume that an EA was conducted to estimate the value of GM versus non-GM corn chips. Using Lancaster's approach of specifying utility for a good as a function of its attributes (see also Louviere, Hensher, and Swait), assume that an individual's utility for a bag of corn chips can be given by the following:

where PRICE is the price of the bag of corn chips, GM is a dummy variable that takes the value of 1 if corn chips are genetically modified and 0 if the corn chips are not genetically modified, BRAND is a dummy variable that takes the value of 1 for some well-known brand name such as Tostitos and 0 for some lesser-known brand or generic bag of chips, and [beta] represents the marginal utility of each of the chip attributes. With this formulation, it is straightforward to show that the individuals' willingness to pay for GM over non-GM corn chips is -[beta]^sub 2^/[beta]^sub 1^, if BRAND = 0 and -([beta]^sub 2^ [beta]^sub 4^)/[beta]^sub 1^ if BRAND = 1.

Now, suppose an EA was conducted in which chips were taken out of their original packages and repackaged in generic bags, such that the subjects could not identify the brand of the chips. Because subjects do not know the brand of the chips, they must make some assumption about the brand name, but, unfortunately, the researcher cannot ex post determine what assumption was made. Because different subjects likely made different assumptions about the brand of the chips, we end up with a classical omitted-variable bias problem.