7.5 Measurement Error in Conjoint Studies 63
Adaptive Conjoint Methods
Adaptive Conjoint Analysis (ACA) and Adaptive Choice-Based Conjoint (ACBC)
methods result in a sets of utilities for each individual. We want conjoint mea-
surements for each individual in the study to be as accurate as possible.
Of the conjoint methods discussed in this book, ACA and ACBC are perhaps
the best at reducing measurement error. These interviews adapt to the respondent,
asking questions designed to be relevant and efficient for refining utility estimates.
If your sample size is particularly small and the number of attributes to mea-
sure is large, ACA or ACBC may be better tools to use. In fact, it is possible
to have an entire research study designed to learn about the preferences of one
respondent, such as an important buyer of an expensive industrial product. As
we discussed in chapter 5, there are many considerations for determining whether
ACA or ACBC is appropriate for a study.
Traditional Conjoint Studies
Traditional full-profile conjoint (such as Sawtooth Software’s CVA or SPSS’s
conjoint module) usually leads to the estimation of individual-level part-worth
utilities. Again, the minimum sample size is one.
One should include enough conjoint questions or cards to reduce measure-
ment error sufficiently. Sawtooth Software’s CVA manual suggests asking enough
questions to obtain three times the number of observations as parameters to be es-
timated, or a number equal to 3(K −k + 1), where K is the total number of levels
across all attributes and k is the number of attributes.
Respondents sometimes lack the energy or patience to answer many ques-
tions. We need to strike a good balance between overworking the respondent (and
getting noisy data) and not asking enough questions to stabilize the estimates.
Choice-Based Conjoint
Though generally considered more realistic than traditional conjoint, choice-based
questions are a relatively inefficient way to learn about preferences. As a result,
sample sizes are typically larger than with adaptive or traditional ratings-based
conjoint, and choice-based conjoint (CBC) results have traditionally been ana-
lyzed by aggregating respondents. Since the late 1990s, hierarchical Bayes has
permitted individual-level estimation of part-worth utilities from CBC data. But
to compute individual-level models, HB uses information from many respondents
to refine the utility estimates for each individual. Therefore, one usually does not
calculate utilities using a sample size of one. It should be noted, however, that
logit analysis can be run at the individual level, if the number of parameters to be
estimated is small, the design is highly efficient, and the number of tasks is large.
There are rules-of-thumb for determining sample sizes for CBC if we are will-
ing to assume aggregate estimation of effects. Like proportions, choices reflect
binary data, and the rules for computing confidence intervals for proportions are
well defined and known prior to collecting data.
Copyright 2010, 2019 © Research Publishers LLC. All rights reserved.