A DISCRETE CHOICE AND ELICITED CHOICE
PROBABILITY COMPARISON THROUGH TIME:
AN APPLICATION TO E-BIKES.
by
INTRODUCTION
Getting insights into the influence of uncertainty on conjoint valuations:
• Uncertainty in choice
• Two conjoint analysis methods
• Uncertainty over time
• Two time periods
Applied to E-bikes
LITERATURE REVIEW
• Choice-Based Conjoint: subjects chose one single alternative as preferred choice (Louviere and Woodworth, 1983)
• Assumption: Subjects face no uncertainty over the preferred product choice.
• Elicited Choice Probability Conjoint: subjects set probability to each alternative in choice set (Manski, 1999).
• Permits subjects to express uncertainty.
Hypothesis choice model
• a: The estimated mean utilities differ significantly between CBC and ECPC.
LITERATURE REVIEW
• Durable good trade-off (Dubé et al., 2014)
• Inordinately optimistic about future (Kahneman and Lovallo 1993; Zauberman and Lynch 2005).
• Small influence of durable good trade-off due to time period, but immediate effects have less uncertainty.
Hypothesis choice over time
• a: There is a significant difference in the means of the utilities of the model with immediate vs. future effects.
LITERATURE REVIEW
• Willingness to pay (WTP) quantifies the maximum price customers are willing to pay.
• ECPC models are more accurate (Manski, 2004).
• No inordinately optimistic about future effect (Kahneman and Lovallo, 1993; Zauberman and Lynch, 2005).
Hypothesis willingness to pay
H5a: The estimated WTP differs between CBC and ECPC models.
METHODOLOGY
• 2x2 within-subjects experimental design
• Decision fatigue (Baumeister, 2003)
• Holdout tasks for reliability (Orme, 2015)
• Number of levels effect (Currim et al 1981; Wittink et al 1989)
• Levels determined via websites (Fietsenonline.com; saturn.de)
Attributes Number of levels Levels
Price 3 €1750, €2000, €2250
Range 3 80km, 100km, 120km
Battery charging time 3 2 hours, 4 hours, 6 hours
METHODOLOGY
Data description
• N = 115 completed surveys.
• Average age of 32.92 years.
• 77.4 % is working and 19.2% is student.
• Sample representative for young professionals.
Distribution of probabilities
• Lower probabilities more information.
• Subjects are aware and willing to answer with probabilities (Manski 2004, Manski et al. 2010).
RESULTS
• Individual estimation conducted via Hierarchical Bayes with 10,000 iterations and 10,000 draws.
• Non-parametric Welch’s two-sided T-test.
Choice model
• Enough evidence of a significant difference between CBC and ECPC in both time periods.
Choice over time
RESULTS
Willingness to pay
• WTP differ across CBC and ECPC.
• Highest WTP for first incremental improvement in performance.
• Models with effect in one month receive higher WTP.
Demographics effects
• Variables possession of E-bike, interest in E-bike and willingness to buy E-bike have greatest influence on cluster formation
1) Working enthusiasts
DISCUSSION
• ECPC analysis outperforms the CBC analysis in terms of efficiency.
• No significant differences across time dimensions; immediate effect vs. effects with a future narrative.
• Contradicting results in the predictive accuracy.
• Willingness to pay show that subjects are willing to pay extensively more for first higher attribute level.
CONCLUSION AND RECOMMENDATION
• Expression of uncertainty in each choice task allows for more coherence of the customer’s preference.
• Extra efficiency does not translate in higher ECPC hit rates and MAE scores.
• Contradicting evidence of time dimension effects.
Management recommendation
LIMITATIONS AND FUTURE RESEARCH
Limitations
• Sample size and sample bias.
• One hold-out task per conjoint model.
• Rounding error in ECPC probabilities.
• Paucity of an incentive alignment.
Future research
• Understanding of the time dimension in less complex settings.
