Two New Features in Discrete Choice Experiments to Improve - PowerPoint PPT Presentation

Two New Features in Discrete Choice Experiments to Improve Willingness to Pay Estimation that Result in SDR and SADR: Separated (Adaptive) Dual Response Forthcoming in Management Science Christian Schlereth WHU – Otto Beisheim School of Management Bernd Skiera Goethe University Frankfurt

Choice-based conjoint nowadays one of the most important method to measure willingness to pay Choice-based conjoint Do not Attribute 1 purchase A B C Attribute 2 any of the Attribute 3 three …  Free-choice questions only, i.e., each choice-sets contains a no-purchase option  No-purchase option provides: - Clear reference point - Realistic experimental setting - Allows prediction of market penetration Carson et al. (1994); Dhar (1997); Louviere et al. (2000); Haaijer et al. (2001); Vermeulen et al. (2008) 2

Information gained when choosing a product Choice-based conjoint Do not Attribute 1 purchase A B C Attribute 2 any of the Attribute 3 three … Selection decision Purchase decision B > A; C B > 0 Product B provides sufficient utility for a purchase 3

Information gained when choosing no-purchase option Choice-based conjoint Do not Attribute 1 purchase A B C Attribute 2 any of the Attribute 3 three … Selection decision Purchase decision 0 > A; B; C None of the products provide sufficient utility No information about relative attractiveness for a purchase of attributes  Not enough data to learn about individual preferences 4

Dual response: Selection decisions are also observed when no- purchase option is chosen Dual response Attribute 1  Dual Response A B C Attribute 2 Forced choice question Attribute 3 … Do not purchase most Free choice question Purchase most preferred preferred Purchase decision Selection decision 0 > A; B; C B > A; C From free choice question From forced choice question  Selection decision is always observable; thus: more accurate estimation of preferences  But higher cognitive effort for a respondent due to double amount of questions Dhar & Simonson (2003); Dhar & Nowlis (2004); Brazell et al. (2006) 5

Shortcoming: Context effects in choice-based conjoint Do not Attribute 1 purchase Attribute 2 A B C any of the Attribute 3 three … Examples:  Attraction Effect:  No-purchase option is chosen less frequently, if a dominant product alternative exists  Similarity Effect:  No-purchase option is chosen more frequently, if similar attractive product alternatives exists, as an “easy way out” • Whether products provide sufficient utility for a purchase is not the only reason for a respondent to pick the no-purchase option • Context effects typically neglected in estimation Huber, Payne, and Puto (1982); Tversky and Shafir (1992); Dhar (1997); Rooderkerk, Van Heerde, and Bijmolt (2011); 6

Shortcoming: Context effects in dual response Attribute 1 A B C Attribute 2 Attribute 3 … Purchase most Do not purchase preferred most preferred Other empirical findings - Higher share of chosen no-purchase option (Dhar and Simonson 2003; Dhar and Nowlis 2004; Brazell et al. 2006) - Artificial time delay between selection decisions and purchase decisions reduces no-purchase share (Dhar and Simonson 2003) • Context effects also exist for dual response • As a result, willingness to pay estimates are substantially lower compared to choice-based conjoint 7

Shortcoming: Extreme response behavior Extreme response behavior (Gensler et al. 2012) • Respondent always chooses no-purchase option •  No information when respondent will start buying •  WTP might be estimated too low • Respondent never chooses no-purchase option •  No information, when respondent will stop buying •  WTP might be estimated too high Extreme response behavior in previous studies (if reported) : Choice-Based Conjoint Dual Response • 58% in Gensler et al. (2012) • 64% in Parker and Schrift (2011) • • 22% in Wlömert and Eggers (2014) 31% in Wlömert and Eggers (2014) • • Up to 56% in our studies Up to 36% in our studies 8

Shortcoming: Impact of purchase probability on measurement accuracy If a choice-set contains more than one alternative, likelihood increases that a respondent compares a rather attractive alternative against the no-purchase option  purchase decisions are less informative for a respondent with high purchase probability Implication: Companies estimate willingness to pay more accurately for a respondent who does not intend to buy their product 9

Aims of paper  Development of SDR: “ Separated Dual Response “, which 1. Avoids context effects by imposing a strict separation between all forced and free choice questions  Development of SADR: “ Separated (Adaptive) Dual Response “, which also 2. Avoids extreme response behavior by imposing a strict separation between all forced and free choice questions through an adaptive mechanism that captures heterogeneity in willingness to pay Ensures similar accuracy in measured willingness to pay, independent of a respondent‘s purchase probability 3. 10

Agenda  Mechanism of SDR and SADR  Simulation study to analyze dependence between willingness to pay accuracy and purchase probability  Tests of statistical and behavioral endogeneity  Insights from three empirical studies 11

Two new features for discrete choice experiments resulting in SDR and SADR Feature 1: Feature 2: Strictly separating forced and Adaptive mechanism to select fewer, but more free choice questions informative, free choice questions SADR (Separated Adaptive Dual Response): SDR (Separated Dual Response): Separated Dual Separated Adaptive Dual Response Response (SDR) (SADR) A 1 B 1 C 1 A 1 B 1 C 1 A 2 B 2 C 2 A 2 B 2 C 2 ... ... Buy Do not Do not Buy A 1 Selected 1 buy buy Buy Do not Do not Buy A 2 Selected 2 buy buy ... ... In addition to feature 1 („ strict separation “), we use We ask all forced choice questions first and then all free choice questions. Thus, we introduce a time decisions in forced choice questions to adaptively identify fewer, but more informative free choice delay between a forced and a free choice question questions 12

Adaptive mechanism of SADR – Separated Adaptive Dual Response SADR 1. Forced choice block Select A, Select B, Select C  Use efficient choice design for all respondents (e.g. D-optimal) Select A, Select B, Select C  Use linear probability model to approximate individual preference order (Heckman & Snyder 1997) …  Simulate preference order of all products in full factorial design Estimate preliminary preference order 2. Free choice block  "Smartly“ select n products for inclusion in purchase questions Purchase Presented, Purchase None  Use binary logit model to select next area, which provides most information … about purchase decision making  Iterate m times Probability of a purchase 100% (= Purchase) Information gap 0% (= No Purchase) 25% 50% 75% Most preferred product Least preferred product Preference order 13

A respondent‘s perspective - screenshots of SADR Forced Choice Block: ( “pick one of the products“ ) j forced choice questions A 1 B 1 C 1 A 2 B 2 C 2 A 3 B 3 C 3 A 4 B 4 C 4 A 5 B 5 C 5 ... Free Choice Block: ( “buy or not buy product“ ) Block 1 of n free choice questions Buy D 1' Do not buy D 1' Buy D 2' Do not buy D 2' Block 2 of n free choice questions Buy ... Do not buy ... Buy ... Do not buy ... ... 14

Summary of studied discrete choice experiments Separated Dual Separated Adaptive Dual Choice-Based Conjoint Dual Response Response Response (CBC) (DR) (SDR) (SADR) Do not A 1 B 1 C 1 A 1 B 1 C 1 A 1 B 1 C 1 A 1 B 1 C 1 buy Buy Do not Do not A 2 B 2 C 2 A 2 B 2 C 2 A 2 B 2 C 2 buy Selected 1 buy ... ... A 2 B 2 C 2 ... Buy Do not Buy Do not Do not Buy A 1 Selected 1 buy Selected 2 buy buy Buy Do not Do not Buy A 2 Selected 2 buy buy ... ... Mechanism Shortcomings X X -- -- Context effects Extreme response X X X -- behavior Impact of purchase probability X X X -- on accuracy of WTP 15

Estimation Scale-extended model  DR-2Max-model (Diener, Orme, and Yardley 2006) Attribute 1 Attribute 1 Attribute 2 Attribute 1 A B C Attribute 2 A B C Attribute 3 Attribute 2  Extended to account for differences in consistency between selection and A B C Attribute 3 … Attribute 3 … … purchase decisions (Swait and Andrews 2003) d   h,i,j       d     h,i ',j'  exp V  exp V       1 h,i 2 h,i' L            h        exp V exp V  exp V        j J i C j' J' i' C 1 h,j 2 0 2 h,i' j   j' Do not buy product D Buy product D Do not buy product D J Buy product D Do not buy product D Buy product D Force choice questions Free choice questions Estimation  Multinomial logit model using Hierarchical Bayes  All models implemented in Matlab 16

Agenda  Mechanism of SDR and SADR  Simulation study to analyze dependence between willingness to pay accuracy and purchase probability  Tests of statistical and behavioral endogeneity  Insights from three empirical studies 17