Terry Elrod, Jordan J. Louviere, Krishnakumar S. Davey
Journal of Marketing Research, Vol. 29, No. 3 (Aug., 1992), pp. 368-377
Published by: American Marketing Association
Stable URL: http://www.jstor.org/stable/3172746
Accessed: 29/07/2010 01:46
Objectives
- compare two approaches to conjoint analysis in terms of ability to predict shares in a holdout choice task
- comparing rating-based and choice-based approaches on ability to predict choice shares from holdout choice sets
* versus Louviere and Gaeth (1988) who compared ratings-based to choice-based, but only coefficients and not predictive ability
* Hagerty (1986) suggests the best utilization of each method seems likely to different model specifications
* Bateson, Reibstein and Boulding (1987) only used OLS for analysis, but at individual level
Leigh, MacKay, and Summers (1984) used TRICON to infer a partial rank ordering for analysis by MONANOVA (discards information available in original choice data)
- introduce and evaluate new choice-based conjoint model specification
* inclusion of "generic cross-effects" allows and tests for departures from Independence of Irrelevant Alternatives (IIA)
Methods/Approach
Methods/Approach
- 3 models fit to individual-level ratings of full profiles vs. four multinommial logit models fit to choice shares for set of full profiles
* individual-level models fit to ratings of full profiles and choice simulators used to predict shares for sets of alternatives (most frequently applied method) (Wittink and Cattin 1989)
* aggregate multinomial logit model using choice data (Louviere 1988a,b; Louviere and Woodworth 1983) = respondents choose one alternative from several sets and aggregate logit model is fit to choice shares by maximum likelihood, choice shared can be predicted directly by aggregate model
- task involved student evaluations of rental apartments
* attributes chosen by examining aprtment listings in university's off-campus housing
* fit normal distribution to listed rent levels, selected levels to vary that corresponded to nearest $10 for first, third and fifth sextiles of distribution (median of lower, middle and upper thirds)
* levels: one bedroom | two bedroom ("All apartments have one bedroom large enough to accommodate 2 twin-sized beds, two bedrooms have 2nd large room as study or extra lounge area"), .5 miles | 1.5 miles | 2.5 miles (by shortest way by road for driving or cycling), very safe | fairly safe
* reduced task artificiality and response bias caused by repetition of attribute level by using random sampling procedure to allow rent and distance from the university vary slightly about a design value for each level (Louviere 1988b)
^ rent could be design level $20 more or $20 less each with 1/3 probability
^ distance was design level .2 mile more or .2 mile less each with 1/3 probability
^ exposed respondents to three times number of values for continuous variables, better representing variability to be found in real-world alternatives
^ thought to also help sustain respondent involvement in task
^ controlled effect on ratings and choices of attributes not included in study by saying apartments were very similar to where they currently lived in every respect not included in profile description
- 3 task preformed by respondents:
1) rating
2) calibration choice
3) holdout choice
- 7 models (specifications, goodness of fit, how predict choice for arbitrary choice sets, how assess ability of models to predict holdout choices):
Results/Conclusions
- both predict holdout shares well with neither ratings-based nor choice-based dominant, though some predict better than others
- new aggregate model that captures departures from independence of irrelevant alternatives (IIA)
Opinion
This paper has been cited 162 times according to Google Scholar. I am still a little confused on how holdout shares are predicted, especially with rating. I assume it is a “ringer” choice which is why I’m not sure how it would be used in a rating situation. Sawtooth (http://www.sawtoothsoftware.com/download/techpap/inclhold.pdf) recommends not including a “None” option. If the “None” option is what it is the JIMAR study did use “ I will not chooise either A or B.”
In regards to the seafood preference study, this paper has only been useful in explaining the disadvantages of choice, which is rare to find.
Useful Information
- advantages of choice over traditional ranking/rating for conjoint analysis:
* usually behavior of ultimate interest, presumably have advantage in predicting choice behavior
* designed to study effect of choice set composition on choice, such as departures from independence of irrelevant alternatives (IIA)
* allow direct prediction of choice shares avoiding conjoint simulators which require questionable assumptions to translate predicted ratings into choices
- disadvantages of choice over traditional ranking/rating for conjoint analysis:
* choice data more amenable to maximum likelihood multinomial logit (MNL) analysis, yet MNL biased
* for small samples there is finite probability that they will be infinite (estimation problems disappear in aggregate level)
* interpretation of aggregate models is more difficult because they confound choice process operating at individual level with heterogeneity in process across individuals
- ratings data allow model estimation by ordinary least squares (OLS) which yields unbiased estimates of parameters
* individual-level estimation is therefore possible which allows arbitrary heterogeneity in coefficients across respondents
* individual-level estimates unstable so observed variability in estimates will overstate variability in true coefficients
* problem of predicting choices from ratings data
Useful References
Hagerty, Michael R. (1986), "Cost of Simplifying Preference Models," Marketing Science, 5 (Fall), 298-319.
Louviere, Jordan J. and Gary J. Gaeth (1988), "A Comparison of Rating and Choice Responses in Conjoint tasks," in Proceedings of the Sawtooth Software Conference. Ketchum, ID: Sawtooth Software, 59-73.
Louviere, Jordan J. (1988) "Conjoint Analysis Modeling of Stated Preferences: A Review of Theory, Methods, Recent Developments and External Validity," Journal of Transport Economics and Policy, 22 (January), 93-119.
Green, Paul E. and V. Srinivasan (1990), "Conjoint Analysis in Marketing Research: New Developments and Directions," Journal of Marketing, 54 (October) 3-19.
Bateson, John E.G., David Reibstein, and William Boulding (1987), "Conjoint Anlysis Reliability and Validity: A Framework for Future Research," in Review of Marketing, Michael J. Houston, Ed. Chicago: American Marketing Association, 451-81.
Louviere, Jordan J and Gary J. Gaeth (1988), "A Comparison of Rating and Choice Responses in Conjoint Tasks," in Proceedings of the Sawtooth Software Conference. Ketchum, ID: Sawtooth Software, 59-73.
Hagerty, Michael R. (1986), "Cost of Simplifying Preference Models," Marketing Science, 5 (Fall), 298-319.
Reibstein, David, John E.G. Bateson, and William Boulding (1988), "Conjoint Analysis Reliability: Empirical Findings," Marketing Science, 7 (Summer), 271-86.
Leigh, Thomas W., David B. MacKay, and John O. Summers (1984), "Reliability and Validity of Conjoint Analysis and Self-Explicated Weights: A Comparison," Journal of Marketing Research, 21 (November), 456-62.
Green, Paul E., Kristiaan Helsen, and Bruce Shandler (1988), "Conjoint Internal Valudity Under Alternative Profile Presentations," Journal of Consumer Research, 15 (Dcember), 392-7.
No comments:
Post a Comment