1

Using Discrete Choice Experiments to Estimate Willingness to Pay Intervals

Christian Schlereth*
Christine Eckert
Bernd Skiera

Marketing Letters (2012), 23(3), 761-776.

* Christian Schlereth, Assistant Professor, Goethe University Frankfurt, Faculty of Business and Economics,
Department of Marketing, Grueneburgplatz 1, 60323 Frankfurt am Main, Germany, Tel.: ++49-69-798-34638,
Fax: ++49-69-79-35001, E-Mail: schlereth@wiwi.uni-frankfurt.de
Christine Eckert, Lecturer, School of Marketing, University of Technology, Sydney, Australia, POBox 123,
Broadway 2007, NSW, Tel.: ++61 2 9514 3538, Fax: ++61 2 9514 3535, E-Mail: christine.eckert@uts.edu.au
Bernd Skiera, Chaired Professor of Electronic Commerce, Goethe University Frankfurt, Faculty of Business and
Economics, Department of Marketing, Grueneburgplatz 1, 60323 Frankfurt am Main, Germany, Tel.: ++49-69798-34649, Fax: ++49-69-79-35001, E-Mail: skiera@skiera.de

2

Using Discrete Choice Experiments to Estimate Willingness-to-Pay Intervals
Abstract
Willingness-to-pay has always been conceptualized as a point estimate, frequently as the
price that makes the consumer indifferent between buying and not buying the product. In
contrast, this article estimates willingness-to-pay (WTP) as an interval based on discrete
choice experiments and a scale-adjusted latent class model. The middle value of this interval
corresponds to the traditional WTP point estimate and depends on the deterministic utility;
the range of the interval depends on price sensitivity and the utility’s error variance (scale).
With this conceptualization of WTP, we propose a new measure, the attractiveness index,
which serves to identify attractive consumers by combining knowledge about their price sensitivities and error variances. An empirical study demonstrates that the attractiveness index
identifies the most attractive consumers, who do not necessarily have the largest WTP point
estimates. Furthermore, consumers with comparable preferences can differ in their purchase
probability by an average of 16%, as reflected in differences in their WTP intervals, which
yields implications for more customized target marketing.

Keywords: Willingness-to-pay, willingness-to-pay intervals, discrete choice experiments,
scale heterogeneity

Venkatesh. and the price at which he or she would be rather certain to refuse to purchase the product. see Shaikh. Wertenbroch and Skiera 2002). and Chatterjee (2007) and Dost and Wilken (2012) have argued that WTP should be understood not as a single point but as an interval. which implies that an interval estimate would more meaningfully and accurately represent WTP. which represents the lower end of the WTP interval (also called the floor reservation price). they draw from a distribution of WTP to make decisions. Venkatesh. Their idea is that consumers rarely know with certainty the amount they are willing to pay. Wang. The authors therefore conclude that a definition of WTP should be linked to the consumer’s probability of purchase. which represents the upper end of the WTP interval (ceiling reservation price). Miller et al. and Talukdar 1997). traditionally by regarding WTP as a point estimate that reflects the reservation price that makes consumers indifferent between buying and not buying a product (Eggers and Sattler 2009. instead.. Sun. Jedidi and Zhang 2002.g. Introduction To set sensible prices.3 1. 2011. Moorthy. a price at which the consumer is indifferent between buying and not buying (indifference reservation price). Ratchford. It is therefore not surprising that a lot of research has focused on WTP determinants and estimations (e. and Chatterjee (2007) further propose to estimate WTP intervals using a procedure that directly asks consumers for three different kinds of WTP (reservation prices): the price at which he or she would buy with some certainty. marketers require detailed knowledge about the willingness-to-pay (WTP) of their consumers. Yet Wang. and Cornelis van . Similar methods that incorporate consumer uncertainty in WTP measures appear in environmental economics literature as closed-ended contingent valuations (for an overview.

our results enable researchers to obtain WTP intervals directly. with additional questions asked during the data collection process. which may be problematic. we motivate the consideration of WTP intervals with a stylized example. Because we use a well-established discrete choice method. Section 3 outlines our approach for measuring WTP intervals and introduces the attractiveness index. . these approaches do not disentangle the size of the WTP intervals into separate effects of choice uncertainty versus price sensitivity. respondents must give probability statements about their purchase likelihoods (e. and Rimer 2001). Yet these methods all focus on stated preferences. In addition. considering that prior research has shown that even highly educated people perform poorly when they must solve a basic probability problem or convert a percentage to a proportion (Lipkus..g. We present an empirical study in Section 4 to demonstrate the value of using WTP intervals. Section 5 concludes.4 Kooten 2007). We show that such disentanglement enables marketing managers to decide which segmentation or communication strategy to use. Samsa. We organize the remainder of this article as follows: In Section 2. Accordingly. and (3) develop an attractiveness index that combines knowledge about WTP point estimates and WTP interval estimates and helps identify the most attractive consumers in the market. this article aims to (1) link WTP intervals to utility theory and choice models through a framework that disentangles preferences and choice uncertainty. “Please reveal a price at which you are only 10% likely to buy the bar of chocolate”). without having to modify the data collection process. (2) present an alternative approach to estimating WTP intervals that can be applied to stated and revealed preferences. In particular.

p) p = $54 90% 10% A A B WTP: $50 WTP: $52 Consistency : 0.$53.0. Usefulness of WTP Intervals We describe a simple numerical example with two consumer segments to outline the usefulness of WTP intervals. A traditional interpretation of WTP would suggest that a marketing manager.9 However. with its higher WTP.5 Consistency : 2 WTP Interval: $50.10 WTP Interval: $45. as well as the insights that individual differences in consumers’ choice uncertainty may provide.10 Attractiveness Index K: . respectively.61 .5 2.39 Attractiveness Index K: . Figure 1b reveals that a traditional understanding of WTP fails to capture differences in choice uncertainty between the segments.$54.90 . Figure 1: Illustrative Example of Willingness-to-Pay Intervals a) b) Willingness to Pay A WTP: $50 < A B WTP: Choice Uncertainty $52 B WTP: $50 WTP: $52 Consistency : 0. and that segment B is four times more consistent in its choices.5 > A is more attractive than B B WTP: $52 Consistency : 2 WTP Interval: $50. Assume that consumer segments A and B have a WTP of $50 and $52.$53.$54.5 WTP Interval: $45. (see Figure 1a).39 WTP: $50 Consistency : 0. WTP Interval.90 . will choose segment B.5 Consistency : 2 B is more attractive than A c) d) WTP Interval Attractiveness Index K = f(WTP. Segment B’s purchase probability den- . who has restricted targeting capabilities and can only target one of the two segments.61 .0.

capture these differences in choice uncertainty. which means that segment B is less likely to buy the product if its price exceeds. We illustrate our method using discrete choice experiments. The attractiveness index K also provides valuable insights for product managers who want to understand the relation between preferences and choice randomness and their effects on purchase probabilities. Method Our proposed method is suitable for any type of choice data and can be applied to both stated preference data. for example.. . in form of household purchase histories. its average WTP. the attractiveness index K is larger in the case of low randomness and thus should be preferred. Figure 1d depicts the application of an attractiveness index K (defined in Section 3). prices at which the segment buys with purchase probabilities in the range [90%. Segment B’s WTP interval (i. In this example. due to the higher consistency of segment B. as illustrated in Figure 1c. Willingness-to-pay intervals. For high-preference consumers. 10%]) is four times smaller than that of segment A. it is sufficient to know that larger values of the index K indicate more attractive consumers. for low preference consumers. e. which combines the knowledge of WTP intervals and WTP point estimates. In contrast. frequently offered by market research companies. and revealed preference data. even by very small amounts. the attractiveness index is calculated on the basis of the choice probabilities for a particular product versus choosing not to purchase any product.e.g.. segment A is actually more attractive as a target than is segment B. obtained from online surveys. for example.1 With this attractiveness index.6 sity function is much steeper than that of segment A. we can see that at a price of $54. even though their WTP point estimate is below the asking price 3. it is larger in the case of high randomness. in which consumers consider different choice sets in which they must choose among different choice alternatives 1 For these illustrative purposes. because they are more likely to purchase.

3.i . and Papageorgiou 1994). where ϖ h can be interpreted as the price parameter. (2002) and Palma et al.e. where smaller variance implies lower choice uncertainty (Palma. (2008) note a range of factors that could contribute to consumers’ choice uncertainty.i ~ G (1/ λh ) ). The utility of product i is usually obtained by multiplying the design vector Xi by the vector of preferences βh of consumer h. Assuming Gumbel-distributed stochastic error terms a ( ε h. The scale (or error variance) λh of these Gumbel-distributed error terms reflects the uncertainty of individuals in their choices. Assuming that a consumer fully spends this budget.i . i∈I).i units of another product that is not observed (i.i (h∈H.i with product i and that the consumer has a budget Yh. added to the otherwise deterministic utility in Equation (3) to capture the various. The utility of the outside good is assumed to be given by ϖ h ⋅ z h. often unpredictable choices people make. Willingness-to-Pay Point Estimator in the Presence of Scale Heterogeneity When analyzing discrete choice data.7 as well as a no-purchase option. advances in choice modeling also propose different models to disentangle the impacts of preferences and error vari- .i = X i ⋅ β h + ϖ h ⋅ z h.1.i (h∈H. the associated budget constraint is: (2) Yh= pi + z h. a researcher typically assumes that consumer h associates utility uh. Myers. The budget Yh can be spent on product i at a price pi and on zh. In addition. discrete choice modelers acknowledge the existence of an error term. i∈I).i units of the outside good with a price normalized to 1. outside good). The direct utility function is then: (1) u h. Louviere et al.. The consumer splits that budget between product i and zh. we obtain a closed form solution for the choice probability Prh.

namely. i '∈Ca The choice probability for a product i thus depends on the utility the product provides to consumer h. i∈I). Fiebig et al. Moorthy. and Chatterjee 2007). such that uh.8 ances on choices (e. it is possible to derive WTP intervals instead of point estimates. By considering this distinction between preferences and choice uncertainty. as well as the utilities of the competing products in the choice set.i = uh.0 ) + ∑ exp ( λh ⋅ u h. This level of uncertainty is represented by the probability with which a consumer chooses the product instead of the no-purchase option. and Jedidi and Zhang (2002).2. we derive the following choice probabilities in a choice set Ca of products: a Prh.i ' ) (h∈H.0. This equality implies that the utility of buying a product at a price equal to the consumer’s WTP (or WTPh.i = (4) Xi ⋅ β h ϖh (h∈H. The estimation of WTP as a point estimate in discrete choice experiments builds on the ideas offered by Small and Rosen (1981). The underlying idea for calculating WTP intervals is that choices take place with uncertainty and that consumer h is “rather certain” or “rather uncertain” about buying a particular product i at a given price (Wang. Willingness-to-Pay Intervals In line with the preceding considerations. which is independent of a respondent’s error variance (or scale-parameter λh). that WTP is the price at which a consumer is indifferent between buying and not buying a product. Venkatesh. Magidson and Vermunt 2007).g. It also implies that the probability of a purchase at a price equal to the WTP is exactly 50% and that the WTP point estimate can be expressed as: WTPh.i ) exp(λh ⋅ u h. Ratchford. 3. and Talukdar (1997). i∈I). 2010.i = (3) exp ( λh ⋅ u h. we can calculate the prices at which the consumer ..i) is as high as the utility of the no-purchase option. Therefore.

it is worth noting that our conceptualization of WTP intervals differs fundamentally from previous uses of the expression “WTP interval”. Greater choice uncertainty—that is. a lower scale λ h —leads to larger WTP intervals. which represents the WTP point estimate. At this point.i λh ⋅ϖ h  1 − Pr UB  λh ⋅ϖ h  1 − Pr LB    (h∈H. We provide details about the derivation of Equation (5) in Appendix A. in theoretical support of the observation by Wang. as a function of the choice probability WTPh. In particular. For our current discussion. we note that if PrLB < 50% the term − LB 1   ln  Pr LB  is positive and represents λh ⋅ϖ h  1 − Pr  the additional sum of money that a product may cost. as well as their price parameter. Similarly. The size of the WTP interval is determined by consumers’ price parameter and the degree of consumer uncertainty. the term − UB 1   ln  Pr UB  becomes negative and reflects the discount necλh ⋅ϖ h  1 − Pr  essary to obtain the choice probability PrUB > 50%. and Chatterjee 2007).i h. beyond the WTP point estimate. Venkatesh. such that it is purchased with a probability of PrLB < 50%. given a choice probability PrUB > 50%. which then describe the upper and lower boundaries of the WTP interval: ' WTPh. i∈I).i). and Chatterjee (2007) that higher choice uncertainty increases WTP intervals. Venkatesh. We finally note that the WTP interval has axial symmetry properties at the choice probability of 50%—a property also empirically observed in the direct WTP interval elicitation by Wang. WTP ln − −   h. Pr LB] = (5)  1 1  Pr UB   Pr LB   WTP ln . such as . Venkatesh. we follow the .9 ' rather certainly starts or stops buying.i (Prh. depends on consumers’ preferences. The midpoint of the WTP interval. The phrases “rather certain” and “rather uncertain” need further specification on the basis of concrete choice probabilities. PrUB = 90% and PrLB = 10% (Wang.i [Pr UB. and Chatterjee (2007).

. $20]. . For example.3. We introduce a new measure. and Loomes's (1994). Venkatesh. Dubourg. Therefore. the attractiveness index K. Jones-Lee. that outlines the attractiveness of individual consumers by summarizing the effects of differences in their utility assessments of competing products and differences in the WTP interval sizes (the formal derivation is in Appendix B). and Chatterjee (2007). We note that standard deviations may depend on the number of respondents or the number of observations per respondents though. hypothetical and non-hypothetical decisions) to estimate consumers' WTP in different situations.10%] interval for this particular product would be [$10. Our conceptualization differs from Bohm's (1984). the WTP [90%. Their resulting WTP intervals come from sampling the prices at which the purchase probability equals 50% in the distributions of the preference parameters and calculating the standard deviations of these prices. such that the boundaries of our WTP interval are defined according to different purchase probabilities for different prices. our conceptualization differs from Swait and Erdem's (2007) effort to capture choice uncertainty using standard deviations of the estimated preference parameters. 3. a price of $10 might be associated with a 90% purchase probability. a price of $15 with a 50% purchase probability. However. But different data sources rarely are available. that example cannot offer guidance about how companies should react to differences among WTP interval sizes and preferences. then recommend mixing those distributions. and a price of $20 with a 10% purchase probability. such that the measured uncertainty mixes with the technical aspects of the data collection. Furthermore.10 conceptualization used by Wang.g. Attractiveness Index Our stylized example in Section 2 demonstrates the usefulness of WTP intervals for managerial decision making. or Park and MacLachlan's (2008). These authors suggest the use of different data sources (e.

the attractiveness index K can be exploited for targeting purposes. we conclude that for a positive numerator (i.h = ( CS h. Rangeh denotes the size of the WTP interval and equals WTPh.h means higher market share for product A. Appendix B reveals that this probability is a continuous function of the attractiveness index KAB. that is. the more market share product A can gain. Therefore.A(PrUB). Equation (7) offers a formal proof of the result we demonstrated numerically in Section 2: Because a larger attractiveness index KAB.B – ph.. The attractiveness index can prioritize targeting of individual consumers. the provider of product A should target consumers with a rather high WTP range. who are rather certain in their decisions or have a high price sensitivity). the higher is its probability of being chosen.e.11 We define the attractiveness index by the ratio Pr(A)/(Pr(A) + Pr(B)). The higher the consumer surplus for product A compared with product B. namely purposefully influencing the consumer’s uncertainty.A – ph.B = WTPh.h.A and CSh. product A is better than product B).A(PrLB) – WTPh.B are the consumer surpluses generated by products A and B. which reflects the probability of choosing product A or B. where CSh. This consumer-specific WTP interval size is defined by the consumer’s price sensitivities and choice uncertainty.e..A = WTPh. A − CS h . Beyond an enhanced understanding of available options in terms of a marketing strategy. the attractiveness index KAB. In contrast. It provides a clear and simple decision rule to identify the most attractive consumers/segments. defined as: (6) K AB . if the consumer surplus of product A is less than product B and the numerator is negative. B ) Rangeh (h∈H). The numerator of Equation (7) indicates monetary differences in the utility of both products (including price).h also points to another strategy that can lead to market share gains. However. according to which . the provider of product A should target consumers with a low WTP range (i. respectively.

and their self-stated likelihood to purchase a netbook in the next 12 months. .6 inch). and price (four levels. Empirical Study Setup The products considered in our study belong to the category of netbooks. we demonstrate the managerial insights that can be gained from the use of WTP intervals and the attractiveness index.1 or 11. their knowledge about netbook prices.1. and the fictitious Texxus). hard drive (160 or 300 GB). as proposed by Bettman. Finally. we asked for demographic information. their familiarity with netbooks. such as gender and age. Empirical Study In this empirical study. which can be described by five attributes: brand (Acer. 4. Samsung. which is suitable for illustrating the benefits that WTP intervals and the attractiveness index provide. 4. ranging from €229 to €399). Each choice set included three netbooks and a no-purchase option. using seven-point Likert scales. “I would not buy any of the three netbooks. which used a D-efficient (3 ∙ 23 ∙ 4)–fractional factorial design with 13 choice sets.12 ones have the highest index value (assuming equal purchase quantities across consumers). The questionnaire consisted of three sections: The first asked about the respondents’ existing experience with computers and netbooks.” We then asked respondents to rate the difficulty of the choice task on a semantic differential scale. and the concept of WTP intervals provides important insights for segmenting and targeting consumers. main memory (1 or 2 GB). The second section featured the discrete choice experiment. Using online survey software. John. in the third section. we obtained 122 completed questionnaires. Relying only on WTP point estimates can be misleading for pricing decisions. display size (10. and Scott (1986).

as is common in the mixed logit model. (2002) propose estimating individual-level models to circumvent the need to specify a heterogeneity distribution for the product of the preferences and scale.scale . we cannot identify the individual scale and can only estimate the products λh∙βh and λh∙ωh. which enables calculations of WTP intervals in a wide range of contexts. a discrete version of the GMNL that offers an easier interpretation of parameter estimates by providing a segment solution.2. (2010) introduce the generalized multinomial logit model (GMNL). Both SALC model and GMNL can be applied to revealed and stated preference data.pref and v cov. Estimation Without specific assumptions about the distribution of preferences or scales. Fiebig et al.13 4. However. as Swait and Louviere (1993). can lead to a seriously misspecified model. (2010) argue. Louviere et al. which together form L∙S classes. v cov. The SALC model assumes that the distributions of preferences and scales can be described by a finite number of preference classes L and a finite number of scale classes S. which in turn allows analysts to test the outcomes of different marketing strategies and then apply the most promising one. To overcome this limitation. Different covariates. Covariates can link consumers to preference or scale segments. explain memberh h ship in each class. and Fiebig et al. which can estimate continuous consumer heterogeneity both for scale and preference parameters. though doing so poses stringent requirements on the data and cannot be used in every empirical setting. The general probability density function associated with respondent h’s choices across all Ah choice sets is given by . Swait and Andrews (2003). ignoring differences in the distributions of preferences versus error variances and assuming a normal mixing distribution. We use the scale-adjusted latent class (SALC) model proposed by Magidson and Vermunt (2007).

pref h ) ∏ ∏ ( Pr ) a∈A h i∈Ca a l.a . respectively.85%. 13.i (h∈H).16%. Only the self-stated task difficulty is weakly significant (p < .57%.a. Swait and Adamowicz (2001) argue that perceived task difficulty influences choice consistency. where C= C*a ∪ {no-purchase option} is choice set a that includes all alternatives (as well as a the no-purchase option). 4.27.s. so we refrain from reporting them.3.pref h ) ∑ exp (δ L cov. however. Results We estimate the SALC model with respondents’ choices from 13 choice sets. the latter can identify consumers who are more price sensitive.52% and 45. All levels are effects coded. with the exception of the price. for which we assume a linear relationship. We identify four preference classes (sizes: 49.i is a binary variable that indicates whether alternative i was chosen by respondent h in choice set a.scale h ∑ exp (γ S ⋅v ) ( exp δ l ⋅ v cov. the sole purpose of our analysis is to demonstrate that WTP intervals can be used to segment and target consumers effectively. The former covariate may affect the general preference structure.i d h . There might be other factors that influence scale and preference class membership. dh. the latter two factors have no significant impacts (p = . respondents’ familiarity with the category. according to the Bayes- . and 12.scale s' h = s' 1 = l' 1 l' ⋅v ) cov.42%) and two scale classes (sizes: 54. and the parameters γs and δl denote the multinomial logit parameters that drive scale and preference class membership. We allow for an influence of sociodemographic measures on consumers’ choices and link preference class membership with respondents’ ages and knowledge about netbooks’ current market prices.1). We therefore test whether self-stated difficulty.48%) as the optimal solution.14 Prh = ∑∑ (7) s∈S l∈L ( exp γ s ⋅ vcov. 24.71) on scale class membership. p = . or the time it takes to complete choice tasks may explain differences in respondents’ choice uncertainty.

and the second. Both scale classes vary substantially in their error variances. Covariates can partly explain both memberships in specific preference classes and in specific scale classes. the more likely he or she belongs to the scale class with the high error variance (see Table C.3) and Table 1). . smaller scale class is almost three times as consistent as the first ( λ2 / λ1 = 2. lead to eight WTP intervals that vary substantially in size.88 . The substantial differences in the sizes of the two scale parameters and the four preference classes. which also assigns the greatest importance to price among all attributes.15 ian information criterion.1 in Appendix C). detailed parameter estimates appear in Table C. Scale class membership can be predicted partly by the perceived difficulty of the task: The more difficult a respondent perceives the task to be.17€ to 499. with their specific price sensitivities. from 75.37€ (see Equation (B. respondents with high knowledge of market prices are more likely to belong to preference class 1.1 in Appendix C).

such as legal constraints or consumers’ objections to price discrimination.15 €) Rank Purchase (Attractiven Probability ess Index) 66% 0.43 -1.21 € 0. followed by preference class 3 (3L+3H) (see the third column of Table 1).69 € 200. we further assume that the company engages in no price discrimination and charges the profit-maximizing price of 333.49 € 5 248. Table 1: Comparison of Different Targeting Approaches WTP Point Estimates Rank (WTP Point Estimates) Range of Attractiveness WTP Index K Interval (price: 333.59 3 2H: Preference Class 2.27 € 1 173. high scale 627. It also provides the best sequence for prioritizing consumers. low scale 369.96 € -0. 300 GB storage capacity.79 € 3 75.39 € 1.37 € 0.15€ (see Appendix D). Yet this targeting sequence fails to take into account the differences in the degree of choice uncertainty. high scale 369. We assume the variable costs are 100€.25 7 8 86% 93% 100% 72% 94% 13% 0% + 21% + 7% +22% -13% As Table 1 illustrates.42 4 2L: Preference Class 2.16 4. high scale 200.27 € 1 499. conflicting recommendations result from the differences in the WTP point estimates and the attractiveness index K. we consider a situation in which the manufacturer ACER might offer a netbook with 2 GB memory.27 € 1H: Preference Class 1.69 € 7 7 305.17 € 0. the range of the WTP interval. For simplicity and to reflect market realities. as reflected in the size of the WTP intervals and the associated attractiveness index (see Figure 1 in Section 2).6".48 € 0. low scale 627. and thus balancing between price sensitivity and . In a simplified example. and the attractiveness index K in this example. low scale 380. Targeting based on differences in WTPs would identify preference class 2 (2L+2H) as the most attractive.4.63 2 4L: Preference Class 4. Managerial Benefits of WTP Intervals We use these results to outline the usefulness of WTP intervals and the attractiveness index. low scale 4H:Preference Class 4. and a display size of 11. By taking those differences into account.49 € 5 86.22 5 3H: Preference Class 3. high scale 380. using both a naive ranking that only considers WTP point estimates and the corrected ranking using the attractiveness index K.15 6 1L: Preference Class 1. In Table 1 we report the WTP for each segment.17 € 105.79 € 3 216.70 1 3L: Preference Class 3.

We further derive an attractiveness index K that summarizes the attractiveness of each consumer as a single number. The authors construct these WTP intervals using three questions to assess the prices at which respondents are almost certainly going to purchase the product. we obtain a different targeting sequence (sixth column of Table 1): Consumers with a high scale in preference class 2 (2H) should be targeted first. and so on. the increase in purchase probability by addressing 1H instead of 1L). and at which they are almost certainly not going to purchase. This identification increases purchase probabilities in preference class 1 by 21% (i. how discrete choice models enable estimates of WTP intervals.. we formally illustrate the impact of differences in consumers’ error variances on the size of their WTP intervals. at which they are indifferent between buying and not buying. Reviewing recent discussions on scale heterogeneity in choice modeling. in class 2 by 7%. with the average absolute increase in purchase probability equaling 16%. and Chatterjee (2007) and Dost and Wilken (2012) have recently argued that WTP intervals. and in class 3 by 22%. the attractiveness index can identify more attractive consumers within each preference class. then consumers with a high scale in preference class 3 (3H). 5. for the first time. in that respondents do not need to answer additional questions. We show. Summary Wang.e. and the approach is applicable for both stated and revealed preferences. because we trade off between the strength of a product and the .17 choice uncertainty. Our proposal provides a substantial advantage. can better depict the uncertainty with which consumers make purchase decisions. In addition to providing a full ranking of all preference/scale class combinations. Venkatesh. rather than WTP point estimates. and in class 4 by 13%.

this level of uncertainty is represented by the probability with which a consumer chooses the product over the no-purchase option. To do so. Marks and Kamins 1988). They also could strategically influence opinions about a product with a targeted product sample campaign in social networks (e. the company benefits from greater uncertainty. 6.18 amount of choice uncertainty.i can be derived analogously to Equation (3) as follows: (A.1) Prh. Nordgren and Dijksterhuis (2009) demonstrate that it is possible to purposely increase uncertainty by adding new.i = ( ( exp λh ⋅ X i ⋅ β h + λh ⋅ϖ h ⋅ (Yh − p h. Different levels of error variance in consumers’ choices lead to more or less extreme choice probabilities.i. leads . In a logit model. The insertion of Equation (4) into Equation (A.1). are better than competitive offers should try to reduce consumers’ uncertainty. irrelevant information. The results of our empirical study suggest that this attractiveness index can easily identify the most attractive consumers. they might rely on advertising campaigns that clearly outline the benefits of the product over competitive options.i ) ) ( ) exp λh ⋅ϖ h ⋅ Yh + exp λh ⋅ X i ⋅ β h + λh ⋅ϖ h ⋅ (Yh − p h. i∈I). The attractiveness index also shows that companies with products that. solving for price ph. if the product is associated with lower preferences. which distracts consumers’ attention from relevant information. Appendix A Derivation of Willingness-to-Pay Intervals The underlying idea for calculating WTP intervals is that choices are made with uncertainty and that a consumer h is “rather certain” or “rather uncertain” about buying a particular product i at a given price. In contrast.. The choice probability Prh.i ) ) (h∈H.g. at the respective prices. which increases purchase probability by an average of 16%.

B ) − λh ·ωh ·( ph .3) ' = WTPh. we consider the ratio Pr(A)/(Pr(A) + Pr(B)).i). B Derivation of the Attractiveness Index We formally derive a new measure. Given Equation (A.i 1 ln  λh ⋅ϖ h  1 − Prh.i    (h∈H.  1 1  Pr UB   Pr LB   − − WTP ln .i  Prh. B ) Prh ( A) = Prh ( A) + Prh ( B ) exp λh ·ωh ·(WTPh .i (Prh.i (Prh. the attractiveness index K. A − ph . such as PrUB = 90% and PrLB = 10% (Wang. Pr LB] = (A. These choice probabilities denote the prices at which the consumer rather certainly starts or stops buying. the range of the WTP interval for consumer h can be calculated as: . A − ph . to depict the attractiveness of individual consumers by summarizing the effects of differences in the utility of competing products and different sizes of WTP intervals.2) = p h. The price ph.i −  Prh.i is then a function of the choice prob' ability WTPh. and Equation (A.2) yields: (A.19 to: (A.i − h. To determine WTP intervals. WTP ln   h.3) and the axial symmetry property. i∈I). A − WTPh . we can define a WTP interval as follows: ' WTPh. A − WTPh .4) (h∈H. i∈I).1) ( ) exp λh ·ωh ·(WTPh . B ) − λh ·ωh ·( ph . Therefore.i [Pr UB. i∈I). which is the probability of choosing product A or B and rearrange it as: (B. B ) + 1 ( ) (h∈H).i ) p= WTPh.3).i WTPh . and Chatterjee 2007).i h.i 1 ln  λh ⋅ϖ h  1 − Prh.i λh ⋅ϖ h  1 − Pr UB  λh ⋅ϖ h  1 − Pr LB    where Pr UB(LB) is the probability for the upper (lower) limit of the WTP interval. From Equation (A. Venkatesh. we solve for prices that yield the corresponding choice probabilities of the interval boundaries. The phrase “rather certainly” needs further specification with concrete choice probabilities.i    (h∈H.

B ) Rangeh (h∈H).5) K AB . If we then substitute Equation (B. such that a higher attractiveness index is associated with a higher purchase probability for product A: (B. we obtain: (B. Table C.3) into Equation (B. according to the Bayesian information criterion (BIC). A − ph . we also simplify this range: (B. A − WTPh .2) 1 1  PrUB   Pr LB  − Rangeh = ·ln ·ln λh ·ωh  1 − PrUB  λh ·ωh  1 − Pr LB  (h∈H). Using the axial symmetry property we discussed in Section 3.2). C Parameter Results We next report the details of the parameter estimates of the models.4) Prh ( A) = 1− Prh ( A) + Prh ( B ) 1 UB    (WTPh . . The scale-extended version substantially improves the BIC from a value of 2604 using the traditional finite mixture choice model to 2567 with the SALC model. B ) − ( ph . B )  exp  2·ln  Pr h UB ·  +1   1 Range − Pr h   h   (h∈H). We identify four preference classes and two scale classes as the optimal solution.3. A − ph . A − WTPh .h = (WTP h. the difference in probabilities between two products A and B is influenced only by the attractiveness index (KAB.1 reports the estimates of the SALC model.3) Rangeh = UB 2   ·ln  Pr UB  λh ·ωh  1 − Pr  (h∈H). Thus.h). The first term in the exponent of the denominator is constant for all consumers.20 (B. B ) − ( ph .

2011).77 0.27 0.51 0.45** 1.03*** Covariates Constant 3.33*** -1.83 0.26*** 1.05.91*** 160 GB 0.53* -0. Because Prs.51 -0.84** -0.00 113.01 0.94*** 0. Profit is the difference between the price and the variable costs. **=p<.. Miller et al.03 12.42% 21. using the attractiveness index to determine the optimal price will not change the optimal price compared with previous approaches (e.73 -0. multiplied by the segment s–specific probability Prs.53* 0. .08 -0.00 18.57% 13.16% Attribute importances Brand 18.10% Main Memory 47.88*** -0.01 Preference Class 4 1.37 54. multiplied by the probability in each segment of purchasing at price pi.56 12.00 48.00 0.49% 3.46* Age -0.g.77 0.10 Wald p-Wert Wald(=) p-Wert 120.00 0.42% 15.79*** 1.85% 24.63% 23.88*** 2.80% 35.31 11.45** 0.03* 0.35*** 0.01 0.69** -0.77 0. and attractiveness indices.1) max π= i ∑( p s∈S i − cv ) ⋅ Prs .66% 14.27% 2.00 0.03 Wald 63.30*** -0.08 40.09 10.06* Segment Size 49.70% 18.30*** 0.6" 0.47 0.21 0.31 Storage Capacity -0.39*** Texxus 0.76% 35.00 33.91*** 300 GB Price (per 100 €) 1.93% 25.66 0.21 Table C.00 14.25* 45.1: Scale Parameter Covariates Constant Difficulty Segment Size Results of Scale-Adjusted Latent Class (SALC) Model Scale Class 1 Scale Class 2 1.1.79*** -1.00 13.i contains all information inherent to the calculation of WTP point estimates.04 0.1" -0.34 0.34 0.82% D Profit Maximizing Price Assuming no price discrimination across segments.62** 0.64 -0.04 -0.33*** Display Size 10.10** -0.36 Price knowledge -0.***=p<.05 -0.05*** -0.55% Significance based on z-values: *=p<.25* 54.48*** Preference Class 3 3.82% Display Size 1.56% 5.06 -0. 2011): (D.58 2.71*** Constant Brand Acer 0.05% Price (per 100 €) 40.00 2.48*** Main Memory 1 GB -0.00 283.20 0.10 0.35*** -0.04 0.01 0.00 47.00 Wald(=) p-Wert 1. WTP intervals.00 48.21% 13.57 0.14 Samsung -0.30 0.48% Preference Class 2 3.60 0.53% 29.26*** -1.68% Storage Capacity 12. the profit maximizing price pi* can be calculated as the price that maximizes the profit in each segment.39% 20.84** 0.64% 13.77*** 0.46*** 0.i of purchasing product i (see also Miller et al.77 p-Wert 0.44*** -1.04 0.06 0.77*** -1.00 114.i ( pi ) (i∈I).94*** 2 GB 1.52% Preference Class 1 5.

Bohm P (1984) Revealing Demand for an Actual Public Good. Jones-Lee MW. Louviere J. . Medical Decision Making 21(1):37-44. Jedidi K and Zhang ZJ (2002) Augmenting Conjoint Analysis to Estimate Consumer Reservation Price. Belief Confidence. California. Marks LJ and Kamins MA (1988) The Use of Product Sampling and Advertising: Effects of Sequence of Exposure and Degree of Advertising Claim Exaggeration on Consumers' Belief Strength. Fiebig DG. Revisited: When Should We Care? International Journal of Research in Marketing:forthcoming. Deshazo JR. and Scott CA (1986) Covariation Assessment by Consumers. . but it differs from the model used in Jedidi and Zhang (2002). and Zhang ZJ (2011) How Should We Measure Consumers’ Willingness to Pay? An Empirical Comparison of State-of-the-Art Approaches. a purchase of product B would be allowed. International Journal of Research in Marketing 26(2):108-118.2) also is consistent with random utility theory. Street D. 2007. Journal of Public Economics 24(2):135–151. Journal of Marketing Research 25(3):266281. but the probability of making the purchase would be less than 50%. Krohmer H. Journal of Marketing Research 48(1):172-184. Magidson J and Vermunt JK (2007) Removing the Scale Factor Confound in Multinomial Logit Choice Models to Obtain Better Estimates of Preference. Ainslie A. Hensher D. References Bettman JR. Carson R. Journal of Risk and Uncertainty 9(2):115-133. John DR. 7. Management Science 48(10):1350-1368. Marketing Science 29(3):393-421. Paper presented to the Sawtooth Software Conference Proceedings. They state that a respondent will choose product A over product B if and only if the difference between his or her WTP for product A and its price is greater than that same difference for product B.22 Equation (D. Louviere J. Santa Rosa. Miller KM. and Attitudes. Keane MP. and Wasi N (2010) The Generalized Multinomial Logit Model: Accounting for Scale and Coefficient Heterogeneity. Lipkus IM. Dost F and Wilken R (2012) Measuring Willingness to Pay as a Range. Cameron T. and Marley T (2002) Dissecting the Random Component of Utility. Eggers F and Sattler H (2009) Hybrid Individualized Two-Level Choice-based Conjoint (HIT-CBC): A New Method for Measuring Preference Structures with Many Attribute Levels. Samsa G. Marketing Letters 13(3):177-193. Kohn R. Dubourg WR. Journal of Consumer Research 13(3):316-326. With random utility theory. and Loomes G (1994) Imprecise Preferences and WTP-WTA Disparity. Hofstetter R. and Rimer BK (2001) General Performance on a Numeracy Scale among Highly Educated Samples.

Econometrica 49(1):105-130. Palma Ad. Journal of Consumer Research 23(4):263-276. and Papageorgiou YY (1994) Rational Choice Under an Imperfect Ability To Choose. McFadden D. . and Walker J (2008) Risk. Marketing Letters 19(3-4):269-285. Myers GM. Wang T. Shaikh SL. Sun L. Uncertainty and Discrete Choice Models. Swait J and Erdem Tl (2007) Brand Effects on Choice and Choice Set Formation Under Uncertainty. Holt C. Palma Ad. Nordgren Loran F and Dijksterhuis A (2009) The Devil is in the Deliberation: Thinking Too Much Reduces Preference Consistency. and Cornelis van Kooten G (2007) Treating Respondent Uncertainty in Contingent Valuation: A Comparison of Empirical Treatments. Journal of Consumer Research 28(1):135-148. Picard N. Ratchford BT. Journal of Marketing Research 44(2):200-213. Wakker P.23 Moorthy S. Moffatt P. Magnac T. Marketing Science 22(4):442-460. American Economic Review 84(3):419-440. Small KA and Rosen HS (1981) Applied Welfare Economics with Discrete Choice Models. Swait J and Andrews RL (2003) Enriching Scanner Panel Models with Choice Experiments. Park JH and MacLachlan DL (2008) Estimating Willingness to Pay with Exaggeration BiasCorrected Contingent Valuation Method. Marketing Science 26(5):679-697. Swait J and Adamowicz W (2001) The Influence of Task Complexity on Consumer Choice: A Latent Class Model of Decision Strategy Switching. Journal of Consumer Research 36(1):39-46. Venkatesh R. and Talukdar D (1997) Consumer Information Search Revised: Theory and Empirical Analysis. Ecological Economics 62(1):115-125. Marketing Science 27(4):691-698. Brownstone D. Ben-Akiva M. Swait J and Louviere J (1993) The Role of the Scale Parameter in the Estimation and Comparison of Multinomial Logit Models. and Chatterjee R (2007) Reservation Price as a Range: An Incentive Compatible Measurement Approach. Journal of Marketing Research 30(August):305314. Train K.