Travel Reference
In-Depth Information
respondents to select a 'none' option, which may reveal some non-compensatory preference
information about the cut-off point regarded by respondents. For example, I would not choose
any option within this set because prices of all offered products are too high. But information
overload is a key problem for full-profi le choice-based tasks since respondents need to deal with
lots of information to select one profi le with all attributes described before giving a single
answer for each choice set, which is even harder than rating each stimulus. As a result, partial-
profi le choice-based conjoint studies were adopted later by researchers, which only provide a
subset of the total number of attributes in each choice question. Because of attributes' omission,
the information gathered by this method is not enough for estimating the part-worth of each
individual respondent assigned to attribute levels. Data from groups of respondents are normally
aggregated for analysis so that part-worth of a target group can be investigated.
Parameter (attributes) estimation is normally the last step in conjoint analysis. During this
step, the part-worth utilities of each attribute are calculated so that the product with maximized
utility can be predicted. According to the literature review of Green and Srinivasan (1978) there
are three kinds of estimation methods which are:
1 non-metric estimation methods such as MONANOVA and LIMAP, which assume that the
dependent variable is, at most, ordinal scaled;
2 metric estimation methods such OLS, which assume that the dependent variable is interval
scaled and compute part-worth utilities by minimizing the squared sum of deviations
between estimated and observed metric values;
3 methods that relate paired-comparison data to a choice probability model or parametric
estimation methods. Methods in this class are the logit and probit models.
Nowadays, conjoint analysis is used as a prevalent tool in marketing research. In a survey among
market research institutes, 65 per cent of the institutes indicated having used conjoint analysis
within the last 12 months, and growing usage frequency was forecasted (Hartmann and Sattler
2002). Compensatory models with conjoint analysis are so popular because they not only forecast
decision-making processes of compensatory preferences but also approximate the outcomes of
other kinds of decision rules (Wahab
et al
. 1976). For instance, a weighted additive model can
theoretically reproduce a non-compensatory decision process if, in the ordered set of weights,
each weight is larger than the sum of all weights to come. Therefore, fl exibility in assigning
weights is the biggest advantage of conjoint analysis.
However, these methods to measure decision-making processes, which are based on utility
maximization, have been questioned by scholars since the 1970s (Payne 1976; Beach and
Mitchell 1978; Gigerenzer and Todd 1987; Rieskamp and Otto 2006). Some simple non-
compensatory heuristic models such as conjunctive, disconjunctive and lexicographic heuristics
were introduced and proved to be more or at least equally accurate in predicting consumer
behaviour in some situations (Czerlinski
et al
. 1999). Besides, the time required to complete
surveys and information overload for respondents is another disadvantage of conjoint tasks with
a relatively large set of attributes. How to increase respondent rate and prevent unreliable answers
caused by the complexity of the task remains a key problem to be solved. And this issue takes us
to the application of greedy algorithms to these decision-making problems.
Greedoid analysis
Greedoid analysis based on a greedy algorithm was developed by Kohli and Jedidi (2007) and
Yee
et al
. (2007) to infer non-compensatory heuristics including: conjunctive heuristic;
