University of Southern California Dissertations and Theses

Essays in information economics and marketing

(USC Thesis Other)

Essays in information economics and marketing

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content ESSAYS IN INFORMATION ECONOMICS AND MARKETING
by
Jisu Cao
A Dissertation Presented to the
FACULTY OF THE USC GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(ECONOMICS)
August 2020
Copyright 2020 Jisu Cao
Dedication
To my parents, Guangming Cao and Yingxin Chang,
for their unconditional love and constant support,
and to my husband, Chao Ma,
for the strength and laughter he brings me.
ii
Acknowledgements
I have a long list of people to thank for in this long journey. I could not go this far without their
help and support. Thank you, Sha Yang, for being such a great advisor. Without your patient
guidance and help, this dissertation would not have been possible. My deep gratitude also goes
to my co-advisor, Geert Ridder, for his mentoring and encouragement. I would like to thank my
other committee members, Roger Moon, Fanny Camera, Guofu Tan, and Yu-Wei Hsieh for their
insightful comments on my dissertation. My appreciation extends to my collaborators, Xin Zheng,
and Xingyao Ren. They taught me that the most important principle in research is persistence.
Thanks also go to James Polk for making English writing not so hard. Finally, I would like to
thank all of my friends especially Yu Cao, Ambuj Dewan, Eunjee Kwon, Mengnan Fan, Brian
Finley, Grigory Franguridi, Rachel Lee, Yanyan Li, Yiwei Qian, Simon Reese, Yuqi Song, Lidan
Tan, Jingbo Wang, Shichen Wang, Xiangqing Wang, Yingfei Wang, Siqi Wu, Josie Xiao, Weining
Xin, Alex Yao Yao, Jeonghwan Yun, and Nan Zhou. Thank you for making my time at USC so
enjoyable!
iii
Table of Contents
Dedication ii
Acknowledgements iii
List of Tables vi
List of Figures vii
Abstract viii
Chapter 1: Understanding the Impact of Consumer Reviews on Demand under Negoti-
ated Pricing 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Empirical Context . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.2.1.1 Product choice and transaction price . . . . . . . . . . . . . . . 12
1.2.1.2 Review characteristics . . . . . . . . . . . . . . . . . . . . . . . 13
1.2.1.3 Other control variables . . . . . . . . . . . . . . . . . . . . . . 16
1.2.1.4 Online reviews and negotiated price . . . . . . . . . . . . . . . . 18
1.2.1.5 Online reviews and product choice . . . . . . . . . . . . . . . . 18
1.3 Model and Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.3.1 Proposed model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.3.1.1 Stage two: bargain . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.3.1.2 Stage one: product choice . . . . . . . . . . . . . . . . . . . . . 28
1.3.2 Identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.3.2.1 Theoretical identification . . . . . . . . . . . . . . . . . . . . . 30
1.3.2.2 Empirical identification . . . . . . . . . . . . . . . . . . . . . . 32
1.3.2.3 Identification using marginal cost . . . . . . . . . . . . . . . . . 33
1.3.2.4 A simulation study . . . . . . . . . . . . . . . . . . . . . . . . . 37
1.3.2.5 Functional-form identification . . . . . . . . . . . . . . . . . . . 39
1.4 Empirical Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
1.4.1 Existence of bargaining power mechanism . . . . . . . . . . . . . . . . . 40
1.4.2 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
1.4.3 Parameter estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
1.4.4 Robustness check . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
1.4.4.1 Correcting for endogeneity . . . . . . . . . . . . . . . . . . . . 53
1.4.5 Counterfactual analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
1.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
iv
Chapter 2: Consumer Purchase Timing and Product Returns in Daily Deal E-commerce 63
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
2.2 Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
2.3 Data and Reduced-form Regression Analysis . . . . . . . . . . . . . . . . . . . . 69
2.4 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
2.5 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
2.5.1 Empirical identification . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
2.5.2 The Log-likelihood function . . . . . . . . . . . . . . . . . . . . . . . . . 91
2.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
2.7 Counterfactual Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
2.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
Reference List 99
Appendix A
Appendix to Chapter 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
A.1 Implementation of Kernel-smoothed Frequency Simulator . . . . . . . . . . . . . 104
A.2 Additional Tables and Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
A.3 Simulation Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
A.4 Additional Proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
v
List of Tables
1.1 Textual Review Contents on Price Negotiation and Bargaining Discount . . . . . . 14
1.2 Summary of Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.3 Variable Definition for Review Characteristics . . . . . . . . . . . . . . . . . . . . 17
1.4 Reduced-form Regression Results of Negotiated Price on Online Reviews . . . . . 19
1.5 Reduced-form Regression Results of Product Choice on Online Reviews . . . . . . 21
1.6 Regression Results of Negotiated Price on Marginal Cost . . . . . . . . . . . . . . 38
1.7 Regression Results of Negotiated Price on Marginal Cost . . . . . . . . . . . . . . 41
1.8 Model Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
1.9 In-sample Model Fit Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
1.10 Estimates on Product Attributes . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
1.11 Model Estimates with IV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
2.1 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
2.2 Reduced-form Regression Results of Number of Sales Events on Purchase Timing . 79
2.3 Reduced-form Regression Results of Number of SKUs on Purchase Timing . . . . 79
2.4 Reduced-form Regression Results of Purchase Timing on Return Probability . . . . 81
2.5 Utility Primitives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
2.6 State Transition Probability Parameters . . . . . . . . . . . . . . . . . . . . . . . . 95
2.7 Model Fit: Purchase and Return Probabilities . . . . . . . . . . . . . . . . . . . . 95
A.1 Variable Definition for Product Attributes . . . . . . . . . . . . . . . . . . . . . . 106
vi
List of Figures
1.1 An Example of A Consumer Review . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.2 Evolution of Review Characteristics Across Time . . . . . . . . . . . . . . . . . . 15
1.3 Illustration of Mechanims of Online Reviews on Product Choice . . . . . . . . . . 23
1.4 Model Identification Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
1.5 Impact of Online Reviews on Negotiated Price w/ or w/o Bargaining Mechanism . 36
1.6 Negotiated Price w/ or w/o Bargaining Power Mechanism . . . . . . . . . . . . . . 39
1.7 Model Fit Comparison w/ or w/o Review . . . . . . . . . . . . . . . . . . . . . . 48
1.8 Estimated Relative Consumer Bargaining Power . . . . . . . . . . . . . . . . . . . 51
1.9 Estimated Relative Consumer Bargaining Power by Product . . . . . . . . . . . . 52
1.10 Difference in Negotiated Price when Historical Price Not Available . . . . . . . . . 56
1.11 Difference in Negotiated Price by Product when Historical Price Not Available . . 57
1.12 Difference in Negotiated Price when Discount Frequency Not Available . . . . . . 58
1.13 Difference in Negotiated Price by Product when Discount Frequency Not Available 59
2.1 Number of Purchases Across Time . . . . . . . . . . . . . . . . . . . . . . . . . . 74
2.2 Number of Purchases by Purchase Date . . . . . . . . . . . . . . . . . . . . . . . 74
2.3 Number of Returns by Purchase Date . . . . . . . . . . . . . . . . . . . . . . . . 75
2.4 Return Rate by Purchase Date . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
2.5 Competition Intensity Across Time . . . . . . . . . . . . . . . . . . . . . . . . . . 77
2.6 Consumer Purchase Timing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
2.7 An Illustrative Figure of Rolling Sales Events . . . . . . . . . . . . . . . . . . . . 84
2.8 A Consumer’s Product Purchase and Return Decisions . . . . . . . . . . . . . . . 85
vii
Abstract
This dissertation includes two chapters with a focus on different topics in the field of Informa-
tion Economics and Marketing. The objective of the first chapter is to understand the impact of
consumer reviews on product choices when purchase price is negotiated. Although prior studies
have analyzed the impact of reviews on consumer baseline preferences, none has explored such
an impact on demand when the purchase price is negotiated. Leveraging a comprehensive data set
on consumer reviews, this study shows a double-edged sword effect of reviews on demand in such
a setting. On the one hand, good reviews increase demand by raising consumer baseline prefer-
ences; on the other hand, they also increase the sellers bargaining power, and this along with the
increased consumer baseline preferences leads to a higher purchase price, thereby hurting demand.
We develop a structural model to uncover the multiple roles reviews have on demand through
baseline preferences, bargaining power, and negotiated price. Ignoring the impact of reviews on
the negotiated price leads to a biased estimation of the effectiveness of reviews. In counterfac-
tuals, we quantify the economic value of different review characteristics and provide managerial
implications for platforms interested in monetizing this information.
The second chapter studies consumer purchase timing and product returns for daily deal e-
commerce where products are often sold in a short window of time (usually one to three days).
Leveraging a unique proprietary data set from a leading Chinese daily deal website, we find two
interesting patterns: (1) consumers generally buy earlier rather than later in sales events; and (2)
product return rates are higher for consumers who purchase earlier. These empirical patterns may
viii
suggest a potential problem for daily deal merchants: a sales promotion to encourage consumers to
buy earlier may actually increase their return probabilities and possibly hurt profits. To understand
such tradeoffs, we develop an integrative model of consumer purchase timing and product return
decisions. In a post-purchase stage, consumer knowledge of the product fit gets realized, and the
consumer can return the product at some cost. In a purchase (order) stage, the consumer decides
to purchase based on her expected utility considering the return probability. A forward-looking
consumer solves an optimal stopping problem for a finite-period game to decide when to purchase
in a sales event. Delaying the purchase allows consumers to see newly posted offers and to have
more time to consider their purchases. We estimate our structural model based on sample panel
data of the purchase and return histories of 5,000 consumers of womens clothing from January to
June 2017. We find that the proposed model fits the data well and that competition significantly
increases the probability that a consumer will buy later. In the counterfactual analysis, we adjust
the product price over days of a sales event and compare the merchants profits under different
pricing schedules. Our counterfactual results reveal important managerial insights that can help
daily deal merchants select a pricing schedule to improve profit.
ix
Chapter 1
Understanding the Impact of Consumer Reviews on Demand under
Negotiated Pricing
1
1.1 Introduction
Online reviews, with the easy accessibility and richness in contents, have radically changed how
consumers make their every day decisions: from seeking a job to researching a new car purchase.
Consumers trust these shared experiences from anonymous reviewers and rely on them to decide
which product to buy or which job to take. From a firm’s perspective, the persuasiveness of online
reviews have affected how they can leverage e-WOM as a marketing tool to affect customer acqui-
sition and business growth. This importance has been well documented by researchers: they have
found that online reviews are not only welfare-enhancing to consumers, but also closely related
to firms’ sales and profits. One hypothesis strongly supported by researchers is that consumers’
preferences and product choices are likely to be associated with review characteristics, such as
valence, volume, and textual contents (e.g., M. Anderson and Magruder, 2012; Archak, Ghose,
and Ipeirotis, 2011; Liu, 2006). Although these contexts vary by study, most of them involve a
fixed-price setting. It is largely underexplored how online reviews affect demand when purchase
price is negotiable.
1
The work in this chapter is joint with Sha Yang (University of Southern California).
1
As the well-known saying goes, “everything can be negotiated.” The rule no price is set can
indeed be applied to a wide range of contexts, such as automobiles, electronic appliances, home
furniture, etc. A national survey suggests that almost half of consumers try to negotiate deals with
salespeople on everyday offerings. Among consumers who choose to negotiate, 89% of them are
rewarded at least once (Consumer Reports, 2013). The monetary value saved from negotiation can
be large: For example, in the furniture market, successful hagglers can save 10% to 20% of the tag
price (White, 2012). In the automobile market, the bargaining room for a new car lies lies around
$500 to $1,000 (Federal Trade Commission, 2019).
In this situation, one important incentive for consumers to read online reviews is to seek a more
favorable bargaining outcome: according to Consumer Reports (2013), one common strategy used
by successful hagglers is to check online reviews to see what others paid. Potential shoppers use the
historical price reported by reviewers to set up a target range that they are willing to accept. Other
review characteristics such as ratings and textual comments can also help consumers to negotiate
a smart deal. Ratings make consumers well-informed on the pros and cons of a product and its
competitors; textual comments frequently contain haggling experiences reviewed by recent buyers.
From reading reviews, consumers can learn every phase of negotiation process and learn useful
tactics. They can counter the seller’s offer more confidently and save sizeable money through a
successful negotiation. In this context, the purpose for consumers to read online reviews is not
limited to evaluate the quality of a product or to learn if the product fits their needs (e.g., Chen and
Xie, 2008; Kwark, Chen, and Raghunathan, 2014). They also use information contained in reviews
to evaluate their bargaining powers and to negotiate lower purchase prices.
When price is negotiable, online reviews may affect consumer’s product choice more than one
way. On the one hand, online reviews may affect a consumer’s liking for a product, and directly
2
affect the choice probability; on the other hand, online reviews may affect a consumer’s willing to
pay for a product and her relative bargaining power with the seller, and indirectly affect the choice
probability through the negotiated price. As good reviews are likely to be associated with a high
consumer’s willingness to pay and a lower consumer’s bargaining power, they can potentially hurt
demand through affecting the negotiated price. Firms should be aware of this potential double-
edge-sword effect of good reviews when managing their online reputation. From a platform’s
perspective, this association between online reviews and the negotiated price offers a lucrative
opportunity for them to monetize their information. One business strategy that has been widely
adopted by Internet companies is to charge a price premium for critical information that offers
great value to consumers: Linkedin charges a premium membership fee for consumers who want
to know which employer has viewed their profiles; Truecar requires new users to register in order
to view full contents of the website, a way to impose shadow costs. If certain review features offer
consumers great value in price negotiation, they have the potential for commercialization. The
essential element of this strategy lies in identifying the most profitable information contents and
in quantifying how much a consumer is willing to pay for it. In this way, a platform understands
what to nurture and strategically set price for these contents.
The objective of this paper is to understand how reviews affect demand when the price is nego-
tiable. We aim to shed some light on the following research questions: (i) How do online reviews
affect a consumer’s baseline preference? (ii) Do online reviews affect a consumer’s bargaining
power? If so, by how much? (iii) How do online reviews affect a consumer’s negotiated price and
product choice? (iv) What is the economic value of different review characteristics to consumers?
To investigate these questions, we collaborate with a leading Chinese automobile platform.
This platform provides automotive information to consumers on all aspects of car purchases and
3
ownership. With approximately 20 million daily active users, it has the largest online review forum
for consumers to read and share reviews of their purchases and products. The platform has taken
several restrictive measures to ensure that these reviews come from actual car owners. We focus
on consumers’ new car choices as price negotiation is prevalent in this market, and consumers
are likely to invest time reading reviews before a high-stakes purchase. We obtain comprehensive
data on consumer reviews from January 2012 to December 2015, ever since the first review was
posted on the platform. For each review, a consumer posts the product information, the date and
location where she made her purchase, and a before-tax price. We build our sample on consumers’
product choices and purchase prices using new car reviews posted from January 2014 to December
2015. We treat revealed purchases as actual purchase choices and the revealed prices as actual
transaction prices. We identify four review characteristics that may affect a consumer’s product
choice and the purchase price: average rating, volume, historical price, and discount frequency.
Among these characteristics, historical price is the average reviewer-reported price for a product,
and discount frequency is a variable measuring how many times reviewers mention discount-related
information in their textual comments. We complement our data with the dealer’s marginal costs,
competition intensity, and the market share during our sampling period. Our final sample consists
of 40,515 observations of consumers’ product choices and transaction prices. We find that, on
average, a consumer receives a 9.6% bargain discount on top of the MSRP. The bargaining discount
is heterogeneous among different products.
Our reduced-form regression results provide evidence that online reviews may affect demand
more than one way: after controlling for product fixed effects and time fixed effects, we find a
strong correlation between review characteristics and the negotiated price. We also find that the
4
association between online reviews and consumers’ product choices is statistically and economi-
cally significant. We propose two potential mechanisms for how reviews may affect demand: by
affecting a consumer’s baseline preference for a product and by affecting a consumer’s relative
bargaining power. Either mechanism can cause the correlation between online reviews and the
negotiated price: a consumer’s preference for a product is related to her willingness to pay, and
the negotiated price is jointly determined by the consumer’s willingness to pay for a product and
her bargaining power. As a result, these two mechanisms cannot be separately identified by simply
regressing the negotiated price on review characteristics.
To disentangle these two mechanisms, we adopt a structural modeling approach. We incorpo-
rate online reviews into a Nash bargaining framework to explain a consumer’s price negotiation
outcome and her product choice (Chen, Yang, and Zhao, 2008). We allow online reviews to affect
both a consumer’s baseline preference for a product and her bargaining power. To ensure the model
can separately identify these two mechanisms, we closely scrutinize the conditions for identifica-
tion. We find that the empirical identification is achievable if we observe dealers’ marginal costs.
We conduct a simulation study to further illustrate this point, and our simulation results suggest
that if the bargaining power mechanism exists, we should observe a correlation between the ne-
gotiated price and the interaction term between the marginal costs and the review characteristics.
Although this reduced-form regression can be used to test our hypothesis if the bargaining power
mechanism exists, the estimates for coefficients turn out to be biased. The regression results on
our sample confirms the existence of the bargaining power mechanism.
To obtain the unbiased coefficients on the bargaining power, we estimate our model using a
simulated maximum likelihood approach. Our parameter estimates are robust to different model
5
specifications and a correction for endogeneity. From a model fit perspective, we find that incor-
porating online reviews into both the bargaining power and baseline preference increases model
fit. Our parameter estimates also provide support for our hypothesis that reviews may affect de-
mand in more than one way. We find that average rating is positively associated with a consumer’s
preference for a product, but is negatively associated with her relative bargaining power with the
dealer. Our empirical results confirm a potential double-edged-sword effect of online reviews on
product demand: Review characteristics such as good ratings increase a consumer’s bargaining
power, thereby increasing the negotiated price and lowering the choice probability. Ignoring this
effect may lead to biased estimation of the impact of online reviews on demand. For the bargain-
ing side, we find that the estimated average consumer bargaining power is 0.546, suggesting that in
our empirical context, consumers are in a slightly stronger bargaining position. We also find strong
heterogeneity in bargaining powers among different products. We find that review characteristics
historical price and discount frequency are strongly associated with a consumer’s relative bargain-
ing power. In particular, a consumer’s bargaining power is positively related to discount frequency
but negatively related to historical price.
As the bargaining power is a key determinant factor for the negotiated price, we postulate that
these two review characteristics offer great value to consumers. From a platform’s perspective,
knowing the economic value of these review characteristics is an essential step for them to under-
stand what contents to commercialize and how much to charge for these contents. The objective for
our counterfactual analysis is to quantify the economic value for different review characteristics.
We focus on historical price and discount frequency as we believe these two review characteris-
tics save consumers the largest amount of money and have the potential to be monetized by the
platform. We define the economic value of online reviews as the price difference when consumers
6
have access to this review information or not. We quantify the economic value for historical price
to be 236 CNY and that for discount frequency to be 939.8 CNY , respectively. These findings offer
great managerial insights for the online review platform.
Although it is well established in the fields of economics, marketing, and information systems
that online reviews are of great merit to modern consumers, it is hard to quantify such a value. For
one thing, reviews are multi-dimensional (e.g., Zheng, Ren, Hong, Cao, and Yang, 2018). Previ-
ous studies have found that important numerical features such as valence and volume may affect
a consumer’s product choice (see e.g., Chevalier and Mayzlin, 2006; Liu, 2006). Other papers
emphasize the inclusion of textual content of user reviews as an important factor (see e.g., Archak
et al., 2011; Netzer, Feldman, Goldenberg, and Fresko, 2012). In our paper, we combine both the
numerical measures with a textual mining approach to obtain a comprehensive understanding of
the economic value of different review features. Among different review features, we find that
historical price and discount frequency have the most economic value to consumers.
The other reason is that reviews may affect demand in more than one way. One mechanism
that has been widely examined is that online reviews may consumers’ product choices through
willingness to pay: consumers use reviews to minimize quality uncertainty or infer their liking
for a product (e.g., M. Anderson and Magruder, 2012; Friberg and Gr¨ onqvist, 2012; Zhao, Yang,
Narayan, and Zhao, 2013). When price is fixed, this willingness to pay is not associated with
price, and online reviews only affect demand in one way. For other goods for which the price
varies, this willingness to pay affects price. This indirect effect on demand via price renders that
online reviews affect demand more than one way. Only a handful of papers have looked into
how reviews affect price. For example, Lucking-Reiley, Bryan, Prasad, and Reeves (2007) find
an association between the auction price and a seller’s feedback rating for collectible one-cent
7
coins on eBay; Li and Hitt (2010) propose an analytical model to illustrate how online reviews
can influence on a firm’s optimal pricing strategy; Ghose, Ipeirotis, and Sundararajan (2007) find
the price premium charged by sellers in an online used-good market is affected by user feedback;
Pavlou and Dimoka (2006) suggest that the addition of text comments and benevolence have great
explanatory power on the variation of price premiums in eBay’s online auction marketplace; Liu,
Feng, and Liao (2017) build a two-period duopoly model to illustrate how online reviews affect
firms’ pricing strategies. A common finding is that when price is not fixed, positive reviews are
likely to be associated with a higher price. This finding applies to different price settings: price is
either set by the seller or is determined by an online auction. None of these papers considers price
negotiation, an important context in durable goods markets. Our paper complements this stream
of literature by studying how different review characteristics can help consumers achieve a more
advantageous position in bargaining with the dealer. We compute the economic value of these
characteristics to be the difference in the negotiated price when we hide this information from
the consumer. Ignoring the impact of online reviews on bargaining power may lead to a biased
estimation of the impact of online reviews on demand.
Our work is also closely related to bargaining literature. Ever since it was developed in Nash Jr
(1950), the Nash bargaining framework has been used to characterize the price negotiation in
various contexts, such as the labor market, the medical device market, the automobile market, etc.
(e.g., Ellis and Fender, 1985; Grennan, 2013; Zhang, Manchanda, and Chu, 2017). Although it is
an important concept in this model, few studies clearly define bargaining power or explore factors
that may affect it. As stated in Binmore, Rubinstein, and Wolinsky (1986), bargaining power
captures the asymmetry in the two parties’ beliefs about the negotiation environment. If we follow
such a definition, the information processed by both parties is an important factor that may affect it.
8
This hypothesis has been supported by past studies. For example, Ratchford and Srinivasan (1993)
find that a more knowledgeable consumer is more likely to get a favorable bargaining outcome
in the automobile market; in a similar spirit, Zettelmeyer, Morton, and Silva-Risso (2006) argue
that it might be easier for consumers with Internet access to negotiate a lower price. We bridge
this concept by exploring one of the most credible and trustworthy information sources of modern
shoppers: user-generated content. We explore if it may affect the bargaining power in a business-
to-consumer setting, and if so, by how much. We provide empirical evidence that several key
review features may have such an impact on shaping automobile shoppers’ beliefs about their
relative bargaining positions with the dealers.
We organize the rest of this study as follows. We first discuss our empirical context and de-
scriptive evidence. Then we present our model in Section 1.3 and the empirical analysis in Section
1.4. We conclude in Section 1.5.
1.2 Empirical Context
We discuss our empirical context in this section. We focus on consumers’ car choices for two
reasons. First, price negotiation is prevalent in this market. It is common for dealers to offer
a discount on top of the list price to secure the deal with consumers. Usually, this amount of
discount depends on how experienced the consumer is and how much information she has for the
product (see e.g., DHaultfœuille, Durrmeyer, and F´ evrier, 2019; G. Huang, 2010; Ratchford and
Srinivasan, 1993). Second, online reviews are an important source of information for consumers
to evaluate a product, to learn bargaining tactics, or to seek the best deal (J.D. Power, 2015).
According to a consumer survey of the Chinese automobile market conducted by the consulting
9
firm Capgemini, 80% of respondents said their car purchase decision had been affected by social
media, in particular by consumer reviews from online forums (Capgemini, 2015).
1.2.1 Data
We obtain our data by collaborating with a leading online automobile platform in China. This
platform is an online resource for automotive information, providing information such as vehicle
prices, dealerships, and inventory listings, on all aspects of car purchases and ownership. This
platform is not only the go-to source for consumers to search for automotive information but also
one of the largest online forums for car-owners to share their car shopping and driving experience.
With approximately 20 million active users, this user review section is one of the most popular
sections in the platform among potential shoppers. To ensure reviews are coming from actual
buyers, this platform has taken stringent measures to ensure the authenticity. A review will not
be released to the general public unless critical proofs of ownership have been provided. For
example, for a review to be approved by the platform, reviewers need to provide photos of the
invoice, receipt, or car key as proofs and to post photos of themselves with their cars. They are
also asked to provide a detailed description up to the trim level of their purchased products. Posts
will be deleted if reviewers fail to complete these requirements.
We collect a comprehensive data set on consumer reviews from the platform from January
2012, when the first review was posted, to December 2015. On average, consumers post reviews
71 days after their purchases. For each review, the consumer reveals purchase-related information
in great detail. She reports the product she bought, when and where she made the purchase, and
a before-tax price. The platform also asks reviewers to rate the purchased product on a one to
10
five-star scale along with eight different attributes: comfort, horsepower, exterior design, interior
styling, fuel consumption, handling, space, cost to own. The reviewer also writes textual comments
to justify her numerical ratings. In addition, we observe the post date of each review and how
helpful other consumers believe this review is. We present an illustrative example of a review in
Figure 1.1. We use this data set to construct samples on consumers’ product choices and review
characteristics.
Figure 1.1: An Example of A Consumer Review
Notes: This figure presents an example of a consumer review from the automobile platform.
11
1.2.1.1 Product choice and transaction price
We treat each reviewer as a consumer and each review as their purchase record. Prior studies
have used similar methods infer a consumer’s product purchase decision: for example, Y . Huang
(2019) uses information contained personal photographs from an online community to infer a
consumer’s camera replacement history. We construct our sample on consumers’ product choices
and transaction prices using reviews in which consumers report a new car purchase during January
2014 to December 2015. We choose this time period to match our data on dealers’ marginal costs,
which is an important source of identification that we will discuss later. We treat the reported
purchase dates as the actual purchase dates and the reported before-tax prices as the transaction
prices. We find that the reported transaction price is on average lower than the product’s MSRP and
are heterogeneous among reviewers. We only consider new car purchases as these constitute the
majority of the Chinese automobile market during our sampling period. According to a consumer
survey conducted by Capgemini, only 15% of potential consumers ever consider buying a used car,
as compared to the 80% who only consider buying new cars (Capgemini, 2015). We only include
observations in which a consumer reports a purchase of the latest model during this sampling
period.
We focus on compact-level sedans to minimize the bias from consumers’ heterogeneous tastes
in automobile segments. To reduce the computational burden, we identify sixteen of the most
popular products and use them as the main sample for estimations. We identify a product as a
unique combination of year, make, and model. If two cars are the same in year, make, and model,
and have different trim levels, we identify them as the same product. We do not disclose the make
12
and model for confidentiality reasons and use numbers 1 to 16 as identifiers to represent these
products.
To make sure our sample is representative of the Chinese automobile market, we re-sample
our choice data with a sampling weight equal to their actual market share. After adjusting for
their actual market share, product 2 has the largest market share of 17.9% among these sixteen
products, while product 1 has the smallest market share of 1.8%. For the rest of the purchases in
this segment, we construct them as the outside option and label the products bought as product 0.
1.2.1.2 Review characteristics
We construct several variables measuring different review characteristics. First, we include av-
erage rating and volume, as prior studies provide evidence that these review characteristics are
important to consumers’ product choices (e.g., Chevalier and Mayzlin, 2006; Liu, 2006). The plat-
form provides the aggregated average rating and the volume for a product on their website and
update them on a monthly basis. The average rating is computed base on the reviewer ratings
and the volume is computed base on the cumulative number of reviews. To mimic the environ-
ment viewed by each consumer, we construct average rating and volume in the same way as this
platform does.
In addition, we include historical price as hagglers are likely to check out the prices paid by
others as a reference point. We measure historical price using the cumulative average reviewer-
reported price. Also, reviewers sometimes post detailed negotiation process and discount-related
information in their textual comments. We postulate that these textual comments allow future
shoppers to evaluate how much bargaining room they have with the seller. We therefore construct
13
the variable discount frequency to measure the frequency that textual comments contain discount-
related information. We adopt a text-mining approach: for each review, if the consumer mentions
the negotiation process or the discount offered by the dealer, we set the dummy to be 1, otherwise
we set it to be 0. We find that 6.7% textual comments contain the word “discount.” This indicates
that it is common for reviewers to brag that a successful haggling experience or to mention nitty-
gritty in the price negotiation in their textual comments. We further present a few examples of
the textual comments that contain discount-related information in Table A.2. In these reviews,
consumers reveal great details regarding their price negotiation: from the best negotiation timing
to which color is harder to bargain for a low price. The richness in textual contents justifies our
hypothesis that the incentives for consumers to read reviews are not limited to evaluate product
quality, but also to acquire information related to price and negotiation. The variable definition of
all four review characteristics are summarized in Table 1.3. We present a time series plot of review
characteristics across time in Figure 1.2.
Table 1.1: Textual Review Contents on Price Negotiation and Bargaining Discount
Reviewer Product Textual Consumer Reviews
Reviewer 1 1 Great discount. The dealer offered a fat deal once I arrived
at the dealer. After close negotiation, they offered a few
thousand less on top of it.
Reviewer 2 2 Great CP ratio. I got a very good discount from the dealer.
Now it is approaching the lunar new year so it is hard to
have such a nice bargain.
Reviewer 3 3 The discount was not bad. If you choose pearl white as
the color, you have to pay an extra of 2,000 CNY . I had to
choose a regular white as a replacement.
Notes: This table offers three examples of textual consumer reviews that are related to price
negotiation and bargained discounts. We choose three representative reviewers, and we do not
disclose their identities to maintain anonymity.
14
Figure 1.2: Evolution of Review Characteristics Across Time
Notes: This figure presents the evolution of review characteristics across time for a representative product.
As we have learned from past studies, shopping for a car is a high-stakes purchase, and the
decision period is usually long for such a purchase. In an ideal world, we would observe each
purchase made by a consumer, when she makes her decision, and the source of information that
influences her choice and the transaction price. However, as is the challenge with many researchers,
these measures are not available to us. To accommodate it, we leverage the fact that we observe
the post date of each review and the purchase date. We assume that the date that consumers decide
which product to buy is 30 days before she actually purchases the product. We then match this date
with each review’s post date. If a review’s post date is before it, then we assume that the consumer
reads this review and include this review in our aggregation.
15
1.2.1.3 Other control variables
To tease out other confounding factors, we collect data on product-specific information and dealer-
specific information. By scraping the product information section on this platform, we collect
data on several important product attributes, including the manufacturer’s suggested retail price
(MSRP), country-of-origin, size, max horsepower, fuel consumption, etc. On the same platform,
we also observe dealer-side information, such as the list price for a product, their shop addresses,
phone numbers, etc. In addition, we have data on the dealers’ marginal costs to acquire a car from
the manufacturer for each product and a monthly market share for these products from January
2014 to December 2015. We match them with our choice sample based on the product identifiers
and names.
Next, we discuss the descriptive statistics, which we present in summary form in Table 1.2.
The sixteen products we include in our sample come from four countries of origin: Germany, the
United States, Japan, and Korea. These products are relatively homogeneous in terms of product
attributes such as size, max horsepower, and fuel consumption. We present a detailed variable
definition in Table A.1 in the Appendix A. They have the following price range: product 14 has
the highest MSRP of 239.90 thousand CNY
2
while product 10 has the lowest MSRP of 74.26
thousand CNY . We measure the competition intensity in selling the focal product in an area using
the variable num. dealers. This variable counts the number of dealers who are selling the same
product in the same city.
We observe a few interesting patterns in our data. We find that the transaction price is system-
atically lower than the dealer’s list price, while the list price is systematically lower than its MSRP.
2
CNY refers to Chinese Yuan.
16
The difference between the transaction price and list price justifies the anecdotal evidence that con-
sumers routinely negotiate prices with dealers in such a market. On average, consumers receive a
9.5% discount on top of the MSRP, and the discounts are heterogeneous among consumers.
Table 1.2: Summary of Statistics
Statistic Mean St. Dev. Min Max
Price 118.758 367.027 61.90 229.000
MSRP 133.525 22.431 74.261 239.900
List price 122.762 23.476 64.900 234.900
Marginal costs 113.281 22.190 59.877 216.719
Average rating 3.723 0.882 2.375 4.750
V olume 426.903 986.550 1.000 5,579.000
Historical price 147.098 44.438 72.772 212.175
Discount frequency 0.086 0.102 0.000 1.000
Num. dealers 5.887 5.099 1.000 35.000
Length 4,533.342 105.571 4,220.833 4,632.800
Width 1,772.839 38.099 1,687.667 1,831.957
Height 1,469.462 18.755 1,425.000 1,505.867
Weight 1,288.831 57.663 1,122.410 1,374.030
Max horsepower 91.560 9.189 72.104 114.429
Fuel consumption 782.074 52.058 705.201 876.425
Country-of-origin: Germany 0.250 0.433 0 1
Country-of-origin: Japan 0.424 0.494 0 1
Country-of-origin: U.S. 0.301 0.459 0 1
Country-of-origin: Korea 0.025 0.156 0 1
Notes: Price, MSRP, list price, marginal costs, and historical price are measured in
1,000 CNY . Variable definitions are in Table 1.3 and Table A.1
.
Table 1.3: Variable Definition for Review Characteristics
Variable Definition
Average rating The average star-rating of the focal product across reviewers.
V olume The cumulative number of reviews of the focal product.
Historical price The cumulative average historical reviewer-reported price of the
focal product.
Discount frequency The cumulative number that reviewers mention discount informa-
tion in their textual comments.
17
1.2.1.4 Online reviews and negotiated price
First, we conduct the following reduced-form regression analysis to test if the negotiated price is
affected by these review characteristics. The model specification is as follows:
Price
ijt
R
j,t1
β
R
Xβ
X
η
ijt
, η
ijt
,N,p0, 1q, (1.1)
where X are the covariates that may affect the negotiated price, such as num. dealers. We
present our reduced-form regression results in Table 1.4. We consider different model specifi-
cations in which we add product fixed effects and time fixed effects. The results are robust in
terms of signs and statistical significance. We find that average rating and historical price have
a positive impact on the negotiated price (coef. 0.251,pvalue 0.01 andcoef. 0.015,
pvalue 0.05), while discount frequency is negatively correlated with the negotiated price
(coef. 0.568, pvalue 0.01). The regression results suggest that review characteristics
are strongly correlated with the negotiated price, indicating that online reviews may affect demand
indirectly through the negotiated price.
1.2.1.5 Online reviews and product choice
Next, we provide empirical evidence of the effects of review characteristics on a consumer’s prod-
uct choice. We conduct the following multinomial logistic (MNL) regression as stated in Equation
1.2,
Prty
jt
1u
exppR
j,t1
β
R
αp
L
j
X
j
β
X
q
Σ
J
j
1
0
exppR
j
1
,t1
β
R
αp
L
j
1X
j
1β
X
q
, (1.2)
18
Table 1.4: Reduced-form Regression Results of Negotiated Price on Online Reviews
Dependent variable:
Negotiated price
(1) (2)
Average rating 0.329

0.251

(0.037) (0.041)
V olume 0.203

0.159

(0.008) (0.011)
Historical price 0.023

0.015

(0.007) (0.007)
Discount frequency 0.662

0.568

(0.112) (0.118)
Num. dealers 0.0001 0.0005
(0.002) (0.002)
Constant 7.255

7.671

(0.205) (0.227)
Product fixed effects Yes Yes
Time fixed effects No Yes
Observations 27,525 27,525
R
2
0.695 0.696
Adjusted R
2
0.695 0.695

p 0.1;

p 0.05;

p 0.01
Notes: OLS regression results of the negotiated price on online reviews.
Standard errors are in parentheses. This sample includes observations
with a transaction price available. We report the results in Column 2.
19
wherePrty
jt
1u is the probability for a consumer to choose productj in timet. R
j,t1
are
the corresponding review characteristics up to timet1. Note that the review characteristics are in
one period lag. This indicates that a consumer’s product choice decision at timet is determined by
the review characteristics she observes in timet1. We purposefully allow this time lag to capture
the time a consumer spent in deliberation when she buys a car. p
L
j
is the productj’s list price. Xs
are other covariates such as product attributes. We include these control variables to minimize the
heterogeneity in consumers’ product choices caused by attribute performances.
We present the estimation results in Table 1.5. We include four review characteristics: average
rating, volume, historical price, and discount frequency. We consider different specifications to
make sure the results are robust: in one specification, we control for product attributes as in Berry,
Levinsohn, and Pakes (1995); in the other specification, we control for product fixed effects. Con-
sistent with past studies, we find a strong correlation between different review characteristics and a
consumer’s product choice. In particular, we find that average rating is positively associated with
a consumer’s product choice (coef. 0.946,pvalue 0.01). We find that historical price is
positively associated with a consumer’s product choice (coef. 0.063, pvalue 0.01), and
discount frequency is negatively correlated with a consumer’s product choice (coef. 0.874,
pvalue 0.01). These regression results indicate that online reviews may affect a consumer’s
product choice.
20
Table 1.5: Reduced-form Regression Results of Product Choice on Online Reviews
Dependent variable:
Product choice
(1) (2)
Average rating 1.400

0.946

(0.028) (0.029)
V olume 0.102

0.092

(0.005) (0.008)
Historical price 0.143

0.063

(0.005) (0.005)
Discount frequency 2.470

0.874

(0.114) (0.098)
Num. dealers 0.062

0.038

(0.002) (0.002)
List price 1.587

1.842

(0.009) (0.012)
Weight 0.055

(0.001)
Height 0.048

(0.001)
Fuel consumption 0.053

(0.001)
Max horsepower 0.030

(0.003)
Width 0.011

(0.0003)
Length 0.006

(0.0001)
Country-of-origin:Japan 1.490

(0.036)
Country-of-origin:U.S. 1.220

(0.100)
Country-of-origin:Korea 1.049

(0.070)
Product fixed effects No Yes
Observations 40,515 40,515
Log Likelihood 83,220.860 79,203.650

p 0.1;

p 0.05;

p 0.01
Notes: OLS regression results of product choice probability on online
reviews. Standard errors are in parentheses. We report the parameter of
estimates based on results in Column 2 of this table.
21
1.3 Model and Identification
These reduced-form regression results provide suggestive evidence that review characteristics (1)
may affect demand; (2) may affect the negotiated price. If this is the case, we cannot separately
identify these impacts and achieve an unbiased estimation by simply regressing demand on the
review characteristics. Therefore, to quantify the economic value of online reviews in such a
market, it is necessary to build a structural model and provide an empirical identification strategy
for these mechanisms. In this section, we proposed a model that can be used to separate the
potential mechanisms reviews have on demand, and then we present an explanation for why the
identification is achievable both theoretically and empirically.
1.3.1 Proposed model
Our model builds on the framework proposed by Chen et al. (2008) with a different emphasis.
Our model is similar to theirs in the spirit of incorporating the price negotiation feature into a
consumer’s product choice, but our focus is to reveal the multiple roles that online reviews have on
a consumer’s product choice. Although it is well established in previous studies that online reviews
impact demand by affecting a consumer’s baseline preference (e.g., Lu and Rui, 2018; Seiler, Yao,
and Wang, 2017), our model is the first to illustrate their potential impacts via bargaining power
and the negotiated price.
Our proposed model allows us to decompose the impact of online reviews from different mech-
anisms. As illustrated in Figure 1.3, our model suggests that when the price is negotiable, online
reviews may affect demand through three potential mechanisms: First, as previous studies have
found, online reviews may affect a consumer’s baseline preference for a product, and thereby the
22
Figure 1.3: Illustration of Mechanims of Online Reviews on Product Choice
Baseline
Preference
δ
Bargaining Power
λ
Review
R
Product Choice
Negotiated Price
˜ p
WTP
Traditional Mechanism
New Mechanisms
Notes: WTP stands for willingness to pay.
choice probability (e.g., the traditional mechanism). Second, since a consumer’s willingness to
pay is associated with a consumer’s baseline preference and is one of the determinant factors of
the negotiated price, online reviews may enter a consumer’s utility function in this way. Third,
online reviews may affect a consumer’s bargaining power, and thereby the negotiated price and the
product choice probability. To the best of our knowledge, the second and the third mechanisms,
despite their importance, are not covered in previous papers. Without a structural model, it is im-
possible to disentangle these mechanisms and accurately quantify the impact of reviews on the
negotiated price, leading to a biased estimation of the impact of online reviews on demand.
Next, we discuss our proposed model in detail. We chose the automobile market to be the
context to be consistent with our empirical analysis, but our model can be easily extended to other
durable goods settings, such as home furniture, electronic appliances, housing, etc. In our model,
a consumer purchases a product from a dealer. She anticipates the price negotiation ex-ante and
23
reads online reviews to achieve a more favorable bargaining outcome. Specifically, we characterize
her purchase decision as a two-stage game as follows.
We start by introducing the timing of this game: In stage one, a consumer reads online reviews
and decides which product to purchase. In stage two, conditional on deciding what to buy, she uses
the valuable information from online reviews to negotiate a deal with the corresponding dealer. De-
spite its restrictiveness in timing, this assumption allows us to accommodate a consumer’s product
choice rather than to let a consumer just make a “purchase or not” decision. If the negotiation is
successful, the consumer makes the purchase; otherwise, the negotiation fails, and she chooses the
outside option. A forward-looking consumer anticipates the price negotiation in the next stage and
incorporates it into her utility function in the first stage. We solve the equilibrium of this game
using backward induction. We therefore discuss the details of these two stages in reverse order.
1.3.1.1 Stage two: bargain
In stage two, after consumeri has decided to purchase productj at timet, with knowledge about
the product and the bargaining tactics she acquired from online reviews, she visits a local dealership
and negotiates the purchase price ˜ p
ijt
with the dealer (him)
3
.
This bargaining process is modeled by the classic Nash bargaining framework (Nash Jr, 1950),
in which the equilibrium is characterized as a split in payoffs between the two parties. In our
context, this split is between a dealer and a consumer. To accommodate the disparities in parties’
beliefs about the market, we allow an asymmetry, namely the bargaining power, in splitting profits
between the dealer and the consumer. Mathematically, if consumer i purchases product j, the
equilibrium negotiated price ˜ p
ijt
solves the Nash bargaining problem:
3
We use she to represent a consumer and him to represent the local dealer.
24
˜ p
ijt
argmax
p
pw
ijt
pq
1λ
ijt
ppmc
j
q
λ
ijt
, (1.3)
where w
ijt
is consumer i’s maximum willingness to pay for product j at time t, and mc
j
is
the marginal cost of product j for the dealer (e.g., invoice price to pay to the manufacturer to
acquire the product). The intuition behind Equation 1.3 is two-fold. First, a successful negotiation
is achieved if and only if a dealer’s marginal cost is less than a consumer’s maximum willingness
to pay (i.e., mc
j
¤ w
ijt
). Second, the price negotiation outcome should lie between these two
variables, depending on how large the relative consumer bargaining power is.
Hereλ
ijt
Pp0, 1q is a scalar capturing the relative bargaining power of consumeri with dealer
j at timet. Traditional IO papers argue that this factor captures parties’ beliefs about the relative
bargaining position. We postulate that online reviews, as the most valuable information source to
modern consumers, may affect such a belief. We will further illustrate which review characteristics
may affect it in a later part of this section. Solving the maximization problem, we get the asym-
metric Nash bargaining solution (Binmore et al., 1986) of equilibrium price negotiation outcome
as a linear function betweenw
ijt
andmc
j
:
˜ p
ijt
p1λ
ijt
qw
ijt
λ
ijt
mc
j
. (1.4)
The equilibrium negotiated price depends on three parts: (1) a consumer’s willingness to pay
w
ijt
, (2) the dealer’s marginal costmc
j
, and (3) the relative bargaining powerλ
ijt
. Among these
three parts, online reviews may affect bothw
ijt
andλ
ijt
. The joint impact on negotiated price is
therefore complex, and can only be identified by estimating a structural model. For example, since
consumers naturally prefer a higher rated product,w
ijt
should be higher for a higher rated car than
25
for a lower rated one. On the other hand, dealers are likely to leverage this high rating to achieve
a more favorable position in a price negotiation. The relative bargaining power for consumerλ
ijt
should therefore be lower.
Explicitly, the relative bargaining power of consumeri with dealerj at timet,λ
ijt
, is given by
λ
ijt

exppRγZκ
C
Dκ
D
q
1exppRγZκ
C
Dκ
D
q
. (1.5)
We impose a logit function form on the bargaining power to ensure thatλ
ijt
lies in the range
between 0 and 1, i.e., 0 λ
ijt
1. We propose that the relative bargaining power can be
explained by three parts: the review featureR, consumer-specific characteristicsZ
4
, and dealer-
specific characteristicsD. γ,κ
C
, andκ
D
are the corresponding parameters of interest that we
want to estimate.
Then we discuss how we model the consumer’s willingness to payw
ijt
. To capture the stylized
fact that a rational consumer is never willing to pay more than the list price p
L
j
, we imposed an
upper bound on a consumer’s willingness to pay to bep
L
j
. Specifically, we assume a consumer’s
maximum willingness to payw
ijt
is given by
w
ijt

$
'
'
'
&
'
'
'
%
r
ijt
if r
ijt
¤p
L
j
,
p
L
j
otherwise,
(1.6)
where p
L
j
is the list price of a product j, and r
ijt
is the consumer’s reservation value. The
equation indicates that a consumer’s willingness to pay w
ijt
for a product will be equal to her
4
We drop the subscripts here for the sake of simplicity.
26
reservation valuer
ijt
if and only if her reservation value for the product is less than or equal to the
corresponding list price. Otherwise, it will be equal to the list price.
In this stage, we assume that consumeri has finalized her product choice, and she is deciding
between buying or not. In other words, if the negotiation fails, her next best alternative is to
choose the outside option. We define a consumers’ reservation valuer
ijt
as the price that makes
her indifferent between acquiring the product and her next best alternative, choosing the outside
option. That is,
Upr
ijt
qU
i0t
, (1.7)
where Upr
ijt
q denotes the utility the consumer gains from buying a product with price r
ijt
, and
U
i0t
denotes the utility of choosing the outside option. We normalizeU
i0t
to be
i0t
. Solvingr
ijt
from Equation 1.7, we have
r
ijt

U
r
ijt
m
j

ijt

i0t
α
. (1.8)
We can then rewrite Equation 1.4 as
˜ p
ijt
λ
ijt
mintr
ijt
,p
L
j
up1λ
ijt
qmc
j
. (1.9)
We accommodate for an idiosyncratic errorη
ijt
for each transaction price we observe.
5
Hence,
consumeri’s price for productj in timet,p
ijt
6
is given by
5
Alternatively, we can have this error term on the bargaining power. We perform the analysis for robustness check,
and the estimation results remain consistent.
6
We use p
ijt
to represent the transaction price we observe in the data and ˜ p
ijt
to represent the model predicted
equilibrium price.
27
p
ijt
˜ p
ijt
η
ijt
, (1.10)
where the unobserved errorη
ijt
is assumed to be i.i.d. distributed withNp0,σ
2
η
q.
1.3.1.2 Stage one: product choice
We then discuss a consumer’s decision in stage one. In this stage, consumer i decides which
product to buy, while anticipating the next stage price negotiation. She evaluates a product by
its performance in different attributes as well as consumer reviews. Mathematically, she solves a
utility-maximization problem. We present her utility function in the following equation:
U
ijt
Rβ
R
Xβ
X
α˜ p
ijt

ijt
, (1.11)
The utility function of customeri in purchasing productj consists of the baseline preference
determined by the product attributesX and online reviewsR, the disutility of paying price ˜ p
ijt
, and
an idiosyncratic consumer characteristic
ijt
that is observable to consumeri, but not to researchers.
We assume
ijt
is distributed i.i.d. with Type-I Extreme Value distribution. We useα,β
R
andβ
X
to denote the corresponding parameters of interests for estimation. Recall that we normalize the
utility of choosing the outside option to be
i0t
and construct consumeri’s reservation valuer
ijt
in
Equation 1.7. We can rewrite Equation 1.7 as:
Rβ
R
Xβ
X
αr
ijt

ijt

i0t
. (1.12)
Solvingr
ijt
from Equation 1.12 gives us
28
r
ijt

Rβ
R
Xβ
X

ijt

i0t
α
. (1.13)
We can then re-write Equation 1.11 as
U
ijt
Rβ
R
Xβ
X
αrp1λ
ijt
qw
ijt
λ
ijt
mc
j
s
ijt
, (1.14)
which is the utility function that the consumer maximizes in our model. From Equation 1.14
we know that online reviews may (1) directly affect demand by affecting a consumer’s baseline
preference for a product, (2) indirectly affect demand via affecting the negotiated price.
Essentially, we propose a two-stage model in which a consumer decides on a product while
anticipating she will bargain with the dealer in the next stage. Therefore, in stage one, she compares
the utility for products with the negotiated price and chooses one that maximizes her utility; in stage
two, price negotiation outcome gets realized. Online reviews are an important information source
for a consumer to choose a product: they affect both how much she likes the product and whether
she can negotiate a more favorable outcome when she bargains with the dealer.
1.3.2 Identification
As the key to this study is understanding how online reviews affect demand in multiple ways, a
clean identification strategy to separate these mechanisms is essential. In this section, we discuss
the theoretical and empirical identification strategy. We aim to show theoretically how the source
of variation in our data allows us to identify our parameters of interests; empirically, we aim to
29
explain how the separation of these different impacts of online reviews can be achieved by the
availability of the dealers’ marginal costs.
We start by laying out the theoretical identification strategy, which follows the strategy of Jindal
and Newberry (2018).
1.3.2.1 Theoretical identification
Letθ pθ
B
,θ
P
qpα,β,γ,κ,σ
η
q denote the set of parameters of interests. θ
P
pα,βq stands
for preference-related parameters andθ
B
pγ,κ,σ
η
q denotes bargaining-related parameters. We
separateθ
B
fromθ
P
as their identifications are achieved from two different sources of variations:
(1) the transaction price and (2) individual product choice.
We first illustrate intuitively how these two variables allow us to identify each parameter and
then provide a formal proof for identifying the bargaining parameters. We identify the bargain-
ing parameters using the covariation on the negotiated price and the covariates. The correlation
between the transaction price and online reviews allows us to identifyγ, which is the magnitude
of the bargaining mechanism. For example, suppose that when the historical price is higher, we
observe a higher transaction price on average, holding other factors constant. Then we can identify
the bargaining coefficientγ
historicalprice
to be negative. Following the same logic, we can use this
variation with dealer-specific and consumer-specific characteristics to identifyκpκ
D
,κ
C
q. The
magnitude of the difference between the model-predicted price and the observed transaction price
allows us to identifyσ
η
.
We then discuss the identification strategy on the bargaining parameters in a more formal way.
Suppose there exist a fraction of consumers whose reservation value for a product is greater than
30
the list price. Their maximum willingness to pay is by definition the product’s list price. Note that
under this condition, Equation 1.9 can be written as
˜ p
ijt
p1λ
ijt
qp
L
j
λ
ijt
mc
j
. (1.15)
If we observe both the transaction price and the dealer’s marginal cost, we can then identify
the bargaining powerλ
ijt
, thereby identifyγ andκ. We have formally proved in Theorem 1 that,
under assumption A1 to A5,θ
B
can be identified.
Theorem 1. Suppose
A1. Exist a fraction of consumers such that their reservation value is greater thanp
L
A2.Erη|Xs 0
A3.Erp
L
˜ p|Xs¡ 0
A4.ErXX
1
s is non-singular.
A5.Erc
1
|XsErc|Xs is non-singular.
Thenθ
B
is identified.
Proof. See Appendix A.4.
As we have identified the bargaining power of each consumer, we can compute the anticipated
negotiated price for each product the consumer considers. The marginal utility of incomeα and
consumer preferenceβ are identified from a consumer’s product choice probability. We can iden-
tify the marginal utility of income α from the change in the likelihood of choosing this product
associated with a change in the negotiated price. We can identifyβ
R
s in a similar way. If the rating
of a product increases, for example, consumers are more likely to purchase the product. The coeffi-
cient for rating should then be identified as positive. By imposing functional-form assumptions on
31
the distributions of unobserved heterogeneity
ijt
and the price errorη
ijt
, we can then fully identify
our parameters of interest.
1.3.2.2 Empirical identification
We then shift our focus to separately identifying different impacts of online reviews. We start
by showing how online reviews may affect the negotiated price in more than one way. As we
mentioned in our models, online reviews may affect the negotiated price via both the bargaining
power and the willingness to pay. For any consumer withr
ijt
p
L
j
, we can rewrite Equation 1.9
as
˜ p
ijt
p1λ
ijt
qr
ijt
λ
ijt
mc
j
. (1.16)
Consider a simple case in which
δ
ijt
R
jt1
β
R
(1.17)
and
λ
ijt

exppR
jt1
γq
1exppR
jt1
γq
. (1.18)
That is, a consumer’s baseline preference for a productδ
ijt
and her relative bargaining power
λ
ijt
are solely decided by online reviews
7
. Then we know from Equation 1.8 that
r
ijt

R
ijt1
β
ijt

i0t
α
. (1.19)
7
We consider only the average rating as the representative review feature in the identification section for the sake
of simplicity. In other words,R
ijt1
is an one-dimensional variable in this section. The identification of other review
features can be achieved in a similar way.
32
If we plug Equation 1.19 to Equation 1.18, we have
˜ p
ijt
p1
exppR
jt1
γq
1exppR
jt1
γq
qr
ijt

exppR
jt1
γq
1exppR
jt1
γq
mc
j
p1
exppR
jt1
γq
1exppR
jt1
γq
q
R
jt1
β
R

ijt

i0t
α

exppR
jt1
γq
1exppR
jt1
γq
mc
j
R
jt1
β
R
α

β
R
α
exppγR
jt1
q
1exppγR
jt1
q
R
jt1

exppR
jt1
γq
1exppR
jt1
γq

ijt

i0t
α

exppR
jt1
γq
1exppR
jt1
γq
mc
j

ijt

i0t
α
. (1.20)
Suppose we letγ 0, that is, we shut down the bargaining power mechanism, then Equation
1.20 becomes
˜ p
ijt

1
2
R
jt1
β
R
α

1
2
mc
j

1
2

jt

i0t
α
. (1.21)
This equation indicates that even if online reviews do not affect the bargaining power, they
may affect the negotiated price via the willingness to pay. This suggests that we cannot separately
identify these two mechanisms by simply regressing the negotiated price on online reviews. Al-
though the reduced-form regression results provide evidence that online reviews may affect the
negotiated price, it cannot specify whether the impact comes from the willingness to pay or from
the bargaining power. The separation is achieved by estimating our structural model.
1.3.2.3 Identification using marginal cost
We will first show that we can separately identify these two mechanisms if we observe the dealers’
marginal costs. If the marginal increase in negotiated price caused by an increase in the marginal
33
cost does not vary with respect to online reviews, it indicates that the bargaining power mechanism
does not exist.
We illustrate the intuition for our identification strategy in Figure 1.4. Suppose we observe the
negotiated price in the following four cases: good reviews, bad reviews, high marginal cost, and
low marginal cost. If online reviews affect both the baseline preference and the bargaining power,
the marginal increase in the negotiated price from low marginal cost to high marginal cost should
be lower for good reviews than for bad reviews. This corresponds to the case of Figure 1.4.a. If
online reviews only affect the baseline preference, the marginal increase in the negotiated price
from low marginal cost to high marginal cost should be the same for a product with good reviews
and with bad reviews. This corresponds to the case of Figure 1.4.b.
34
Figure 1.4: Model Identification Strategy
Notes: WTP stands for willingness to pay, MC stands for marginal cost. In this example, WTP is 20 when reviews
are good while WTP is 15 when reviews are bad. In each panel, the left two columns refer to the case in which the
marginal cost is high, and the right two columns refers the case in which the marginal cost is low. MC is 10 when the
marginal cost is high, and MC is 5 when the marginal cost is low. Panel a corresponds to the case that online reviews
affect both a consumer’s baseline preference and her bargaining power. Panel b corresponds to the case that online
reviews only affect a consumer’s baseline preference.In Panel a, a consumer’s bargaining power is 0.8 when reviews
are bad and is 0.2 when reviews are good. In Panel b, a consumer’s bargaining power is 0.8 no matter reviews are
good or bad.
Mathematically, if we take the first-order partial derivative of negotiated price on the marginal
cost in Equation 1.20, we have
B˜ p
ijt
Bmc
j

exppR
jt1
γq
1exppR
jt1
γq
, (1.22)
while that of with Equation 1.21 is
B˜ p
ijt
Bmc
j
1{2. (1.23)
35
These two different partial derivatives indicate that when online reviews do not affect bargain-
ing power,
B˜ p
ijt
Bmc
j
is constant while it has a logit form with respect to R
ijt1
if online reviews do
affect bargaining power. We further illustrate this in Figure 1.5.
Figure 1.5: Impact of Online Reviews on Negotiated Price w/ or w/o Bargaining Mechanism
Notes: This figure presents the relationship between the negotiated price and online reviews when the marginal cost
takes different values. On the left panel, we haveγ0, and on the right we haveγ1. The marginal costc takes
values from the range [1, 2].
This figure illustrates how we can use the partial derivative of negotiated price with respect to
online reviews to separate these two mechanisms. On the left panel, we observe that the marginal
increase in the negotiated price, when we increase marginal cost, is constant with respect to an
increase in online reviews. This corresponds to instances in which online reviews affect the ne-
gotiated price only via the baseline preference. On the right panel, the marginal increase declines
as the online reviews increase, referring to instances in which online reviews affect negotiated
price via both mechanisms. We argue that the key identification therefore lies in the availability of
36
the marginal cost. We can provide suggestive evidence by regressing the negotiated price on the
interaction between the marginal cost and online reviews as we illustrate in our simulation study.
1.3.2.4 A simulation study
We use a simulation study to further illustrate this point. We simulated two cases in whichγ 0
orγ 0.05, corresponding to the instances in which online reviews affect the negotiated price
via only the baseline preference or via both the baseline preference and the bargaining power,
respectively. Marginal costs are generated byNp0, 1q. We lay out the details for our experimental
design in Appendix A.3. To illustrate our point, we regress the following model:
price
ijt
β
R
R
j,t1
βmc
j
˜ γmc
j
R
j,t1
η
ijt
,η
ijt
Np0, 1q. (1.24)
We present our estimation results in Table 1.6. By comparing Column 1 and Column 3, we
can see that online reviews have a statistically significant explanatory power on the negotiated
price, regardless of whether the bargaining power mechanism exists or not. This indicates that
the identification cannot be achieved via by simply regressing the negotiated price on the online
reviews. However, we find that the coefficient on the interaction term is significant with a 95%
confidence level in Column 2 when γ 0, but not statistically significant in Column 4 when
γ 0. These OLS regression results confirm our findings that the interaction terms of the marginal
cost and online reviews can help us separately identify both mechanisms. We also find that the
estimates in all four columns deviate from the true value, suggesting a bias from using reduced-
form regressions to estimate the impact of online reviews. This signals a need for developing and
estimating a structural model to achieve an unbiased estimation.
37
Table 1.6: Regression Results of Negotiated Price on Marginal Cost
Dependent variable:
Negotiated price
(1) (2) (3) (4)
Online reviews 1.101

1.122

1.069

1.081

(0.012) (0.017) (0.012) (0.017)
Marginal cost 0.718

0.718

0.733

0.733

(0.012) (0.012) (0.012) (0.012)
Online reviews Marginal costs 0.021

0.012
(0.012) (0.012)
Constant 0.344

0.344

0.286

0.285

(0.017) (0.017) (0.017) (0.017)
γ = 0 Yes Yes No No
Observations 10,000 10,000 10,000 10,000
Adjusted R
2
0.535 0.535 0.540 0.540

p 0.1;

p 0.05;

p 0.01
Notes: OLS regression results of negotiated price on the marginal cost, online
reviews and the interaction term. Standard errors are in parentheses. γ is
defined in Equation 1.18. γ 0 in Column 1 and 2,γ .05 in Column 3
and 4.
38
1.3.2.5 Functional-form identification
In addition to the availability of marginal cost, we can also rely on the functional form assump-
tions imposed in our model to separate these two mechanisms. We plot Equation 1.20 forγ 0
and γ 0 in Figure 1.6. It is clear from the figure that when bargaining power is not affected
by online reviews (i.e., γ 0), the negotiated price is linear in review rating. However, when
bargaining power is affected by online reviews (i.e.,γ 0), the relationship becomes non-linear.
This indicates that the non-linearity in negotiated price with respect to online reviews can also help
us separate the mechanisms.
Figure 1.6: Negotiated Price w/ or w/o Bargaining Power Mechanism
Notes: This figure presents the patterns of the negotiated price whenγ equals to 0 andγ does not equal to 0. The
cyan line refers to the predicted negotiated price in Equation 1.20 when online reviews do not affect bargaining
power, and the pink line refers to the predicted negotiated price when online reviews affect bargaining power.
β.5,RUr1,5s,mc1.
39
1.4 Empirical Analysis
1.4.1 Existence of bargaining power mechanism
In this section, we start by providing some empirical evidence for the existence of the bargaining
power mechanism, then we discuss how we construct the likelihood for estimation. Next, we
present the parameter estimates and the counterfactual analysis. Following the logic from the
previous section, we can provide empirical evidence for the existence of the bargaining power
mechanism by regressing the negotiated price on the interaction terms between the marginal cost
and online reviews. Our reduced-form regression model follows the specification in Equation 1.24
but includes more review features. In particular, we regress the following equation
price
ijt
βmc
j
β
1
averagerating
j,t1
β
2
volume
j,t1
(1.25)
β
3
historicalprice
j,t1
β
4
discountfrequency
j,t1
β
5
num.dealers
j,t
˜ γ
1
mc
j
averagerating
j,t1
˜ γ
2
mc
j
volume
j,t1
˜ γ
3
mc
j
historicalprice
j,t1
˜ γ
4
mc
j
discountfrequency
j,t1
˜ γ
1
mc
j
num.dealers
j,t
m
j
η
ijt
,η
ijt
Np0, 1q.
We present the regression results of negotiated price on the interactions of online reviews and
the marginal cost in Table 1.7. We find that the interaction terms of marginal costs and review
characteristics have a statistically significant explanatory power on the negotiated price, indicating
that online reviews may affect a consumer’s bargaining power as well.
40
Table 1.7: Regression Results of Negotiated Price on Marginal Cost
Dependent variable:
Negotiated price
(1) (2) (3)
Marginal cost 1.351

0.937

0.942

(0.023) (0.034) (0.035)
Average rating 1.273

0.152

0.235

(0.064) (0.073) (0.077)
V olume 0.090

0.261

0.274

(0.014) (0.014) (0.015)
Historical price 0.039

0.012 0.011
(0.008) (0.011) (0.011)
Discount frequency 4.187

2.950

3.164

(0.243) (0.261) (0.267)
Num. dealers 0.0002 0.005

0.005
(0.003) (0.003) (0.003)
Average rating Marginal cost 0.095

0.009 0.013

(0.005) (0.006) (0.006)
V olume Marginal cost 0.004

0.011

0.012

(0.001) (0.001) (0.001)
Historical price Marginal cost 0.001 0.003

0.003

(0.001) (0.001) (0.001)
Discount frequency Marginal cost 0.284

0.198

0.209

(0.019) (0.021) (0.021)
Num. dealers Marginal cost 0.001

0.001

0.001

(0.0003) (0.0003) (0.0003)
Constant 3.876

1.950

1.695

(0.261) (0.361) (0.378)
Product fixed effects No Yes Yes
Time fixed effects No No Yes
Observations 27,525 27,525 27,525
R
2
0.928 0.938 0.938
Adjusted R
2
0.928 0.937 0.938

p 0.1;

p 0.05;

p 0.01
Notes: OLS regression results of negotiated price on covariates. Stan-
dard errors are in parentheses.
41
1.4.2 Estimation
We estimate our model using a simulated maximum likelihood approach. We will first show the
probability of each observation and then present how we use these probabilities to construct the
likelihood for the entire sample. Note that the outcome for each observation consists of two parts:
the transaction price and the product choice. In other words, for each consumer who does not
choose the outside option, we observe her product choice as well as the transaction price in this
purchase.
LetO
it
tO
i0t
,O
i1t
,...,O
iJt
u denote the potential outcomes that a researcher observes for
consumer i in time t. We use j 0 to denote the outside option. For each of the products
j 0, the outcome variableO
ijt
consists of two parts: (1) her decision of purchasing productj
and (2) the actual transaction price p
ijt
. Recall that in our model, we include error terms in the
predicted price and utility functions to accommodate unobserved factors that may affect consumer
behaviors and make parametric assumptions on their distributions accordingly. We will first derive
the probability of an outcome conditional on knowing these unobservables and then integrate over
their distributions to obtain the likelihood that we estimate.
Specifically, conditional on knowing the unobserved consumer characteristics
it
and error term
on priceη
ijt
, the probability that a researcher observes a consumeri buying productj and with a
transaction pricep
ijt
is given by
PrpO
ijt
1|θ;
it
,η
ijt
qPrpU
ijt
p˜ p
ijt
q¥max
j
1
j
U
ij
1
t
p˜ p
ij
1
t
q X ˜ p
ijt
η
ijt
p
ijt
|
ijt
,η
ijt
q,
(1.26)
42
where ˜ p
ijt
is the predicted negotiated price. Conditional on the unobserved characteristics
it
and error term η
ijt
, the likelihood that consumer i chooses to buy product j and with price p
ijt
is equal to the likelihood that (1) the consumer anticipates bargaining, and productj maximizes
her utility; (2) the addition of model-predicted price ˜ p
ijt
and an error term η
ijt
equals the actual
transaction pricep
ijt
.
Note that we do not observe the transaction price when consumers choose the outside option. In
this case, the conditional probability that a researcher observes a consumeri choosing the outside
optionj 0 is given by
PrpO
i0t
1|θ;qPrpU
i0t
¥max
j
1
0
U
ijt
p˜ p
ijt
q|
it
q. (1.27)
We account for the fact that researchers do not observe
it
andη
ijt
, and solve this problem by
integrating over the distribution of
it
and η
ijt
. Recall that we assume that
ijt
follows an i.i.d.
Type-I Extreme Value distribution with mean equal to 0 and that variance equal toπ{6,η
ijt
follows
an i.i.d. distribution withNp0,σ
2
η
q. Integrating over
it
andη
ijt
, the unconditional probability of a
researcher observing consumeri choosing the productj with a purchase pricep
ijt
is given by
PrpO
ijt
1|θq
»

»
η
PrpU
ijt
p˜ p
ijt
q¥max
j
1
j
U
ij
1
t
p˜ p
ij
1
t
q X ˜ p
ijt
η
ijt
p
ijt
|
it
,η
ijt
q (1.28)
fpqfpηqddη.
43
The unconditional probability of observing the consumer choosing outside option is given by
PrpO
i0t
|θq
»

PrpU
i0
¥max
j0
U
ijt
p˜ p
ijt
q|
it
qfpqd. (1.29)
The log-likelihood of observing a series of outcomes by consumeriPI t1, 2,...,Iu from
productjPJ t0, 1, 2,...,Ju in timetPT t1, 2,...,Tu can be written as
LLpθq
¸
iPI
¸
jPJ
¸
tPT

PrpO
ijt
|θq

. (1.30)
One challenge we face is that as the idiosyncratic error terms enter the consumer willingness
to pay as we illustrate in Equation 1.8, the integral in Equation 1.28 does not have a closed-form
solution even with the simplest logit distribution. To overcome this challenge, we simulate the
demand side shock and compute the choice probability based on the simulated outcomes. This
simulation approach helps us to obtain a numerical integral for our unconditional likelihood.
Specifically, the simulated likelihood is derived as follows. Let a dummy variableo
r
ijt
denote
the product choice a consumer makes in simulation drawr;o
r
ijt
equals 1 if modelj maximizes her
utility, and 0 otherwise. The simulated likelihood function can be written as
SLLpθq
¸
iPI
¸
jPJ
¸
tPT

1
R
¸
r
Prpη
ijt
p
ijt
˜ p
ijt
qo
r
ijt

. (1.31)
Notice that the simulated likelihood is not smooth because of the discreteness of the dummy
variableo
r
ijt
. We therefore cannot use a gradient-search algorithm to search for the maximum value
for our log-likelihood function. A remedy to this problem is to take numerous simulation draws,
but to do so is computationally costly. As illustrated in Honka (2014), an alternative method to
44
solve this problem is to smooth the dummy variable using a kernel-smoothed frequency simulator
(McFadden, 1989).
We smooth the dummy variables with a multivariate-scaled logistic cumulative distribution
function (CDF) as in Gumbel (1961)
Fpω;sq
1
1exppsωq
, (1.32)
where s is the tuning parameters and ω is calculated based on the simulated outcomes. Fol-
lowing Honka (2014), we set s 15 in our application. For any given draw of , ω
r
ijt

U
r
ijt
max
j
1
j
tU
r
ij
1
t
u. The intuition behind this kernel smoother is to use a logit form to ap-
proximate the indicator function to make it smooth. Note in Equation 1.32, the kernel smoother
is positive in ω and is asymptotically equal to 1 when ω goes to infinity. This indicates that the
probability is larger when the chosen product maximizes the utilities among alternatives, and is
smaller otherwise. The tuning parameter is chosen based on the level of smoothess we want to
achieve: the larger the tuning parameter is, the less smooth our CDF is, but the closer it is to the
indicator function.
Replacing dummyo
r
ijt
with a kernel-smoothed frequency simulatorFpω;sq, our log-likelihood
function becomes
SLLpθq
¸
iPI
¸
jPJ
¸
tPT

1
R
¸
r
Prpη
ijt
p
ijt
˜ p
ijt
qF
r
pω;sq

, (1.33)
which is the main objective function for us to maximize in our estimation.
45
1.4.3 Parameter estimates
Next, we present the estimation results from our structural model. We will first show how our
proposed model increases model fit, and then present the estimates and discuss the robustness
checks. Lastly, we present the counterfactual analysis in this section.
We find that incorporating online reviews into estimating bargaining power increases the overall
model fit and in particular the fit in predicting the negotiated price. In Table 1.9, we present the
in-sample model-fit comparison for different models. In model 1, we do not include online reviews
in estimating both the baseline preference and the bargaining power; in model 2, we do not include
them in estimating the bargaining power. We use these two models as benchmarks for comparison
with our proposed model, in which we include online reviews into estimating both of them. We
find a decrease in both the Akaike information criterion (AIC) and Bayesian information criterion
(BIC), indicating that our proposed model is better in terms of model fit. The likelihood ratio test
further confirms that our proposed model has a better fit than model 1 and model 2, respectively.
In Figure 1.7, we present the predicted negotiated price with or without including online reviews
into the estimation. This figure provides evidence that incorporating this information increases the
model fit on the negotiated price.
We report our parameter estimates with their standard errors in Table 1.8. Our parameter es-
timates provide evidence that online reviews may affect both the bargaining power and the base-
line preference. Consistent with previous studies (e.g., Chevalier and Mayzlin, 2006; Wu, Che,
Chan, and Lu, 2015), we find that the average rating positively affect a consumer’s baseline pref-
erence: a more positive rating is associated with a higher baseline preference (coef. 0.802,
pvalue 0.01). We do not find such a consistent positive effect for volume (coef. 0.007,
46
Table 1.8: Model Estimates
Estimates
(1) (2) (3)
Bargaining parameters (γ)
Average rating 0.947

1.146

1.142

(0.029) (0.011) (0.014)
V olume 0.120

0.153

0.170

(0.005) (0.005) (0.007)
Historical price 0.088

0.102

0.091

(0.004) (0.004) (0.002)
Discount frequency 2.745

4.667

4.578

(0.106) (0.004) (0.032)
Num. dealers 0.007

0.009

0.007

(0.002) (0.004) (0.003)
Constant 4.244

4.882

4.653

(0.144) (0.005) (0.023)
σ
η
0.549

0.559

0.565

(0.004) (0.009) (0.014)
Product choice parameters (β,α)
Price coefficientα 0.200

0.326

0.323

(0.011) (0.011) (0.009)
Average rating 0.863

0.802

0.913

(0.012) (0.051) (0.022)
V olume 0.002 0.007 0.046

(0.002) (0.012) (0.008)
Historical price 0.093

0.113

0.127

(0.002) (0.007) (0.003)
Discount frequency 0.701

0.244

0.051
(0.065) (0.103) (0.049)
Num. dealers 0.010

0.033

0.015

(0.001) (0.003) (0.004)
Product attributes No Yes No
Product fixed effects No No Yes
Observations 40,515 40,515 40,515
Log likelihood 144,530.1 137,482.5 134,922

p 0.1;

p 0.05;

p 0.01
Notes: Structural estimation results on parameters of interests. Standard er-
rors are in parentheses. We obtain the standard errors using a bootstrapping
method. We include a detailed variable description for product attributes in
Table A.1. We report the parameter of estimates based on results in Column
2 of this table.
47
Table 1.9: In-sample Model Fit Comparison
Include reviews in
bargaining
Include reviews in
product choice
AIC BIC
Model 1 No No 316,567.6 316,688.1
Model 2 No Yes 311,696.4 311,851.4
Model 3 (proposed) Yes Yes 275,009 275,198.4
Figure 1.7: Model Fit Comparison w/ or w/o Review
48
pvalue n.s.). This result is likely due to the durable good context of our paper: unlike ex-
perience good, a consumer’s preference for a durable good is less likely to be affected by review
volume (Liu, 2006). We find that the coefficients for historical price and discount frequency are
significant at a 1% significance level (coef. 0.113, pvalue 0.01 and coef. 0.244,
pvalue 0.01), indicating a potential quality signalling effect for these review characteristics.
Our estimates also suggest that consumers who have more num. dealers available in an area have
a higher preference for a product (coef. 0.033,pvalue 0.01).
Aside from review characteristics, we provide a few empirical findings on other covariates,
such as competition intensity and product attributes. We present the estimation results in Ta-
ble 1.10. Our results are consistent with what has been found in the economic and marketing
literature (e.g., Berry et al., 1995): we find that consumers on average prefer cars with better
fuel consumption (coef. 0.926, p value 0.01), higher horsepower (coef. 3.386,
pvalue 0.01). We find a significant effect on the variable country-of-origin: consumers on
average prefer Japanese cars over American cars, German cars, or Korean cars.
More importantly, our estimates suggest that there is a strong correlation between review fea-
tures and a consumer’s relative bargaining power. In particular, we find that a dealer with a higher
average rating is more likely to stay in a more advantageous position when bargaining with con-
sumers (coef.1.146,pvalue 0.01). This result indicates that a good rating may be associ-
ated with higher bargaining power for the dealer and thereby with a higher negotiated price. Good
reviews may hurt demand by indirectly affecting the negotiated price. We also find that historical
price is negatively associated with the relative bargaining power of a consumer (coef.0.102,
pvalue 0.01). Online reviews containing “dealer discount” information are positively corre-
lated with the relative bargaining power of a consumer (coef. 4.667,pvalue 0.01). Aside
49
Table 1.10: Estimates on Product Attributes
Estimates
Weight 5.208

(0.055)
Height 1.512

(0.185)
Fuel consumption 0.926

(0.008)
Max horsepower 3.386

(1.197)
Width 0.283

(0.012)
Length 0.783

(0.013)
Country-of-origin: Japan 0.292

(0.004)
Country-of-origin: U.S. 0.005

(0.0002)
Country-of-origin: Korea 0.092

(0.003)
Notes: Structural estimation results on parame-
ter of interests. Standard errors are in parenthe-
ses. We obtain the standard errors using a boot-
strapping method. Fuel consumption is mea-
sured as the fuel consumed per hundred kilo-
meters. We include a detailed variable descrip-
tion for product attributes in Table A.1. The
variable Country-of-origin can take four levels:
Germany, Japan, U.S. or Korea. The dummy
variable country-of-origin: Germany is normal-
ized to be the base level in our estimation.
50
Figure 1.8: Estimated Relative Consumer Bargaining Power
Notes: This histogram presents predicted relative bargaining power for consumers. The blue color area is the density
plot. The red dashed line refers to the predicted average bargaining power for the entire sample.
from online reviews, we find that competition is an important factor that is positively correlated
with a consumer’s relative bargaining power (coef. 0.009,pvalue 0.01).
We estimated the average bargaining power of this market to be 0.546, indicating that con-
sumers on average are in a more advantageous position when bargaining with the dealer. We
present a histogram of the relative bargaining power in Figure 1.8. We find that there is strong
heterogeneity in bargaining powers for consumers who have bought different products. We show
the histogram for the bargaining power by product in Figure 1.9. Among the 16 products, product
16 has the highest average bargaining power among consumers of 0.710, while product 14 has the
lowest of 0.371.
51
Figure 1.9: Estimated Relative Consumer Bargaining Power by Product
Notes: These density plots present predicted relative bargaining power for consumers by product. The red dashed line refers to the predicted average bargaining
power for the consumers who have brought such a product.
52
1.4.4 Robustness check
To ensure that our results are robust, we estimate our model with a few different settings. We
consider settings in which we add product attributes and product fixed effects, and present the
results in Column 2 and 3 of Table 1.8. The majority of results remain consistent in terms of signs
and significances.
1.4.4.1 Correcting for endogeneity
We are aware of the potential endogeneity in our data and adopt the control function approach
in Petrin and Train (2010) to mitigate this problem. In particular, we consider the endogeneity
that comes from the potential correlation between the unobserved consumer characteristics and the
number of dealers in the area. The underlying intuition is that when a dealership is set up, the
dealer may account for consumer characteristics that are not available to researchers. Therefore,
the number of dealerships we observe in our data might be an equilibrium rather than exogenous.
Not considering this correlation may cause the estimation results to be biased. The control function
approach in Petrin and Train (2010) allows us to correct for this type of endogeneity. To adopt this
approach, we need a proper instrumental variable. Following Nevo (2001) and Hausman, Leonard,
and Zona (1994), we construct the instrumental variable for num. dealers to be num. nearby
dealers, which is the average number of dealers selling the same product in nearby cities. This is a
Hausman-type of instrumental variable: the underlying assumption is that the cluster of dealerships
is driven by supply-side variables rather than demand-side variables.
53
Table 1.11: Model Estimates with IV
Estimates
(1) (2)
Bargaining parameters (γ)
Average rating 1.138

1.171

(0.021) (0.018)
V olume 0.145

0.159

(0.011) (0.009)
Historical price 0.101

0.098

(0.005) (0.003)
Discount frequency 4.663

4.730

(0.006) (0.014)
Num. dealers 0.007

0.009

(0.005) (0.003)
Constant 4.878

4.886

(0.003) (0.002)
σ
η
0.560

0.553

(0.014) (0.009)
Product choice parameters (β,α)
Price coefficientα 0.336

0.295

(0.013) (0.011)
Average rating 0.744

0.802

(0.03) (0.027)
V olume 0.025

0.063

(0.002) (0.007)
Historical price 0.114

0.137

(0.009) (0.006)
Discount frequency 0.053

0.157

(0.007) (0.061)
Num. dealers 0.037

0.024

(0.005) (0.005)
Product attributes No Yes
Observations 40,515 40,515
Log likelihood 142,851.3 137,514

p 0.1;

p 0.05;

p 0.01
Notes: Structural estimation results on parameter of interests. Stan-
dard errors are in parentheses. We obtain the standard errors using the
bootstrapping method. We include a detailed variable description for
product attributes in Table A.1. We report the parameter of estimates
based on results in Column 2 of this table.
54
1.4.5 Counterfactual analysis
So far, we have provided evidence on the multiple mechanisms online reviews have on demand.
We have uncovered the double-edge-sword effect on good reviews on demand. One privilege our
structural model has is to perform a counterfactual analysis. The objective of our counterfactual
analysis is to quantify the economic value of different review features. Understanding this ques-
tion offers great managerial insights to platforms. One popular business model among the Internet
based companies is the freemium pricing model – offering both free and paid for packages to con-
sumers: For instance, Linkedin charge a premium membership fee to job seekers who want to view
which employers have viewed their profiles. To view the full contents in Truecar.com, consumers
have to sign up for the website, a way to impose consumers with shadow costs. Platforms are
interested in commercializing their contents on consumers as well. In our context, we ask if the
online review platform has the potential to adopt this business model to monetize their contents.
Since monetary reward from a successful negotiation is large, the platform might have the potential
to charge consumers on the contents that offer them great value. In our counterfactual analysis,
we ask if it is possible to charge consumers on important review characteristics, and if so, by how
much?
Our counterfactual analysis allows us to quantify how much a consumer is willing to pay for
each review characteristics. The implementation is as follows. We define the value of each review
characteristic as the difference in the negotiated price when we hide these review characteristics
from consumers. In particular, we focus on two review features: historical price and discount
frequency. We choose focus on these two characteristics as (1) our estimates suggest that they are
55
Figure 1.10: Difference in Negotiated Price when Historical Price Not Available
Notes: This figure presents the density plots for the change in the negotiated price when historical price is not available
to consumers. Consumers assume the historical price to be the list price in this exercise.
strongly associated with a consumer’s relative bargaining power; (2) they are relatively underex-
plored by other studies.
In the counterfactual situation, we consider the worst-case scenario in which a consumer does
not observe any information on these two variables: they assume the historical price to be the list
price and discount frequency to be 0. We then compute the difference in the negotiated price using
our structural model. We quantify the economic value of historical price and discount frequency
to be 236 CNY
8
and 939.8 CNY
9
respectively. We present the economic value of these two review
features by product in Figure 1.11 and Figure 1.13 respectively.
8
approximately $37.46 with 2015 CNY to U.S. dollar exchange rate.
9
approximately $149.17 with 2015 CNY to U.S. dollar exchange rate.
56
Figure 1.11: Difference in Negotiated Price by Product when Historical Price Not Available
Notes: This figure presents the density plots for the change in the negotiated price by product when historical price is not available to consumers. Consumers
assume the historical price to be the list price in this exercise.
57
Figure 1.12: Difference in Negotiated Price when Discount Frequency Not Available
Notes: This figure presents the density plots for the change in the negotiated price when discount frequency is not
available to consumers. Consumers assume the discount frequency to be 0 in this exercise.
58
Figure 1.13: Difference in Negotiated Price by Product when Discount Frequency Not Available
Notes: This figure presents the density plots for the change in the negotiated price by product when discount frequency is not available to consumers. Consumers
assume it to be 0 in this exercise.
59
1.5 Conclusion
Although it is well documented that online reviews may affect a consumer’s product choice, their
impact on price is less well understood. Past studies have found that sellers will adjust their price
based on how good the reviews are; other papers investigate this question under an auction mech-
anism. None of these studies consider a setting that involves price negotiation.
In a context in which the purchase price is negotiable, the impact of online reviews on demand
can be complex. For one thing, the product-related information helps consumers to better evaluate a
product. At the same time, consumers may set the historical price as the acceptable price range and
read textual comments to learn bargain tactics. Reviews are not only correlated with a consumer’s
liking for a product, but also affect her negotiation power for it. The latter mechanism is largely
ignored by past studies.
In this paper, we aim to complement this stream of studies by examining how online reviews
affect demand when consumers can negotiate a purchase price. Our hypothesis is that online re-
views may affect both a consumer’s baseline preference for a product and her bargaining power.
We focus on the automobile market and collect data from a leading Chinese automobile review
platform. Our sample consists of 40,515 observations on consumers’ product choice and transac-
tion price. We identify four important review characteristics: average rating, volume, historical
price, and discount frequency. Our reduced-form regression results support our hypothesis as we
find an association between online reviews and the negotiated price.
As online reviews might affect demand in more than one way, we cannot separately identify
these mechanisms by simply using reduced-form regressions. To disentangle these potential mech-
anisms, we develop and estimate a structural model. We carefully lay out our identification strategy
60
to show that our structural model and data can help us identify the existence of these mechanisms.
In particular, we reveal a three-fold impact of online reviews on demand. (1) They affect baseline
preference directly; (2) they affect the consumer willingness to pay; (3) they affect the bargaining
power. The latter two mechanisms indirectly affect demand via the negotiated price. We are, to
the best of our knowledge, the first to uncover these indirect mechanisms by which online reviews
may affect demand.
Our structural analysis also reveals several other unique findings. We find that an increase in
average rating increases a consumer’s baseline preference, but lowers a consumer’s relative bar-
gaining power. This finding reveals the dual-sword-edge impact of good reviews on demand. On
the one hand, good reviews increase a consumer’s baseline preference for a product, and thereby
increase demand; on the other hand, they can hurt demand by increasing the negotiated price. Sell-
ers may take non-optimal marketing strategies to hurt profit if they are not aware of this potential
harm of good reviews on demand. We estimate the average bargaining power of the Chinese auto-
mobile market to be 0.546, which is slightly in favor of the consumer side. We discover substantial
heterogeneity in bargaining powers for consumers who purchase different products with the low-
est average bargaining power equal to 0.371 and the highest bargaining power equal to 0.710.
Consumers’ bargaining power is also affected by the competition intensity in the local market.
We find that historical price and discount frequency have strong impacts in affecting a con-
sumer’s bargaining power. Specifically, a higher historical price is correlated with a higher rela-
tive bargaining power for consumers, while a higher discount frequency is associated with a higher
relative bargaining power for dealers. This change in the bargaining power leads to a change in the
negotiated price paid by the consumer, saving the consumer a sizeable amount of money.
61
As these two review characteristics offer great value to consumers, we aim to quantify the
economic value of them in our counterfactual analysis. We compute the difference in the negotiated
price when the platform provides this information or not to consumers. In a scenario in which a
consumer cannot observe the historical price and discount frequency, we assume that the consumer
uses the list price to replace the historical price and sets the discount frequency to be zero in her
decision making. Our results suggest that these review characteristics greatly help consumers to
negotiate a lower price. In particular, we find that the average economic value for the historical
price is 236 CNY and that for the discount frequency is 939.8 CNY . We discuss related managerial
insights. For platforms that are interested in monetizing their contents to improve their profits, this
counterfactual exercise can serve as a pricing guideline to understand how much a consumer is
willing to pay for this information.
Our paper, to the best of our knowledge, is the first to study the multiple impacts of reviews
on demand when the price is negotiable. Our paper is not without limitations. First, although our
model is flexible enough to accommodate other settings, we restrict our attention to the automobile
industry due to data limitations. Future research can consider other durable good markets such
as real estate, home furniture, electronic appliances, etc. Second, we do not observe consumer
demographics such as the age or gender of the consumer in our data. If researchers have data
available, it will be interesting to see which segments of consumers are more likely to benefit from
reading online reviews during price negotiation. These limitations offer great opportunities for
future studies, and we hope our research can stimulate more research to explore this interesting
area.
62
Chapter 2
Consumer Purchase Timing and Product Returns in Daily Deal
E-commerce
1
2.1 Introduction
Daily deal, also known as “deal of the day” or “flash sale”, refers to an online business model
in which merchants offer a limited variety of products or services at steep promotional discounts
for a given period of time (usually one to three days). Since the launch of the first daily deal
site “woot.com” in 2004, this novel business idea quickly gained popularity among global market-
places: covering more than 300 markets and 35 countries in the world (Dholakia, 2011). Featured
daily deal sites in the U.S. include Groupon, LivingSocial, and Giltcity, among others. Addition-
ally, daily deals have been widely recognized as an effective marketing tool. By partnering with
daily deal sites, business owners have successfully boosted their revenues and created quick cash
flows.
A common strategy of many daily deal merchants is to encourage consumers to buy early.
Merchants often try to create a sense of urgency by stimulating customer excitement and building
the desire for their products by means of countdown timers, rewards for the first few purchases,
etc. These business tactics may persuade some consumers to take action early and can be therefore
successful in selling non-refundable goods and services. However, for returnable products, their
1
The work in this chapter is joint with Sha Yang (University of Southern California) and Chunmian Ge (South
China University of Technology).
63
effectiveness in encouraging consumers to buy earlier may come with a cost: consumers who buy
earlier may have a higher probability of returning the purchased product for a variety of reasons.
For example, consumers who take a longer time to consider a purchase may be more deliberate and
therefore they are less likely to reverse their purchase decisions afterward (R¨ ollecke and Huchzer-
meier, 2016).
Product returns are common in traditional retailing and occur even more often for online trans-
actions. The return rate is 8.89% in brick-and-mortar stores whereas that of e-commerce is around
30%. Retail categories with high return rates include clothing, electronics, and footwear. Prod-
uct returns are known to be costly to e-commerce merchants: billions of dollars have been paid
to reverse logistics and restocking warehouses, not yet accounting for the notorious depreciation
costs open-box products need to be sold at a lower price. However, as a lenient return policy may
encourage consumers to buy again (see e.g., Bower and Maxham III, 2012), most e-commerce
websites allow “free returns” for consumers at the expense of paying high return costs themselves.
Prior papers have studied return policies as a means to ensure consumers against ex-post utility
loss and quantified its option value to consumers (E. T. Anderson, Hansen, and Simester, 2009;
Che, 1996). However, existing literature has not fully examined the relationship between purchase
timing and product return decisions. This is important in daily deal transactions as it can help
merchants improve the effectiveness of their marketing campaigns. For example, in our empirical
context, merchants keep prices constant over the sales window. By exploiting the relationship
between purchase timing and return probability, we can design a more optimal pricing schedule
that increases their profits.
We leverage a unique proprietary data set from a leading Chinese daily deal website. On this
website, products are usually sold at steep discounts for a short window of time (usually three
64
days) and are often refundable. Consumers subscribe to sales events of their interests, and receive
notifications at the start of each new event. Popular product categories include women’s and mens
clothing, baby products, and electric appliances. As this website has a generous return policy,
allowing “unconditional returns” within 7 days for most categories and products, it provides us
with an excellent data set to examine the research questions above. We choose women’s clothing
as our focus as this category is vulnerable to product returns: a fit shock coming from physical
(e.g., size) or experiential (e.g., color, style) reasons drives a high return rate among consumers.
We collected a detailed record of consumer purchase and return histories of 5,000 consumers of
women’s clothing, from January 1st, 2017 to June 30th, 2017. We find the average return rate to
be 27.3%, calling for a careful examination of this issue. In the data, we observe two interesting
patterns: (1) consumers generally buy earlier rather than later in sales events; and (2) product
return rates are higher for consumers who purchase earlier. In an examination of the related factors
to consumer purchase timing, we find suggestive evidence that a more competitive environment is
associated with consumers’ decisions to delay their purchases.
Motivated by these empirical patterns, we develop an integrative model of consumer purchase
timing and product return decisions in the context of daily deal sites. In our model, a consumer’s
decisions to purchase a product and to keep/return the product are separate, but related through a
“fit shock.” When a consumer purchases (orders) a product, her knowledge of the product fit is not
perfect. Such a fit shock only gets realized in the post-purchase stage and she can return the product
with some cost. She makes her purchase decision based on her expected utility, considering the
next-stage return probability.
In our model, the consumer also makes decisions on when to purchase. She receives notifi-
cations of the sales events that are of interest to her. She can purchase the product right away or
65
delay her purchase decision to the next period. She is willing to consume it sooner but delaying
the purchase allows her to see new offers posted. This is a institutional feature that aligns with the
daily deal sites — new sales events come out to replace the expired ones on a daily basis. We pos-
tulate that this competition effect from newly posted offers serves as the main reason for consumer
to delay their purchase. In this way, we characterize a consumer’s purchase timing decision as an
optimal stopping problem for finite-horizon time periods.
Next we estimate the proposed structural model. With the estimated utility primitives, we suc-
cessfully replicate the empirical patterns in the data. We find the competition coefficient parameter
to be positive and significant, confirming our assumption that consumers are forward-looking and
that competition plays a role in consumers’ strategic delay decisions. Consumers who are facing
a more competitive environment may choose to delay and wait until the next day to purchase. We
also replicate the decreasing trend of return rates with respect to purchase timing: our parameter es-
timation suggests that consumers who buy later may become more deliberate about their purchases
and less prone to return what they have bought.
A structural model allows us to examine the counterfactual profitability of different marketing
strategies. With the parameter estimates, we recover consumers’ purchase and return probabilities
and their return probabilities. We find that ignoring product returns will significantly bias the
estimation of merchants’ profits. In our counterfactual experiment, we aim to understand the profit
implications for merchants by allowing for dynamic pricing schedules. We adjust the price of a
product over days of a sales event and compute the change in profits under these pricing schedules.
We find that a sales promotion at the first day of a sales event may have a potential negative effect
by increasing the product return rates, thereby possibly decreasing merchants’ profits. This has
managerial implications for merchants for selecting a pricing schedule that may improve profit.
66
The rest of this paper is organized as follows. In Section 2.2, we review the literature. Section
2.3 describes the data collection and empirical context, and provides reduced-form analyses. We
present the structural model in Section 2.4. Section 2.5 describes the identification strategy and
estimation method. Section 2.6 provides results. We demonstrate our counterfactual analysis in
Section 2.7. Finally, we conclude and discuss future research directions in Section 2.8.
2.2 Literature
Our paper contributes to marketing literature on product returns and daily deal e-commerce. First,
our work is closely related to papers that focus on consumer product returns. Early papers focus
on the validity of offering generous return policies and designing optimal return policies across
selling channels: Che (1996) argues that return policy serves as an insurance against “ex-post”
lost of experience good for risk-averse consumers, allowing monopoly sellers to charge more than
otherwise; Padmanabhan and Png (1997) discuss a provision of a return policy can be profitable
to manufacturers if demand is uncertain and retailing is competitive; Shulman, Coughlan, and
Savaskan (2010) compare the equilibrium return policies under different channel structures.
Ever since the growth of catalog orders and the advent of e-commerce websites, the division
between purchase stage and return stage has created more problems for merchants in terms of
product returns (Wood, 2001). As returns and refunds have become more prevalent under such a
purchase environment, more papers have looked into this issue by examining the return policies
and their consequences on post-return sales. Ofek, Katona, and Sarvary (2011) find that offering
an online store in addition to “brick-and-mortar” shops for a retailer may not be profitable under a
less competitive environment when we take product returns into consideration; E. T. Anderson et
67
al. (2009) quantify the option value of product returns using a structural approach in an empirical
context of catalog orders; Bower and Maxham III (2012) compare the consequences of post-return
spending on fee return and free return policies and discover that consumer post-return spending
declines sharply under a fee return policy, but increases dramatically under a free return policy;
Inderst and Tirosh (2015) suggest that refund can be viewed as a “meter device” for consumers to
learn their valuation about a product; a good conclusion would be the argument by Petersen and
Kumar (2009) that “product returns might be inevitable but not necessarily evil”.
Focusing on the consumer side, another stream of literature aims to understand product returns
from a behavioral perpective and engages in designing an online shopping environment to miti-
gate this costly issue. For example, Shulman, Cunha Jr, and Saint Clair (2015) incorporate the
behavioral theory of reference dependence and find that pre-purchase information may not always
be beneficial to merchants, as they may raise consumers’ expectations of a product’s value; Con-
lin, O’Donoghue, and V ogelsang (2007) examine the projection bias when consumers make return
decisions under different weather conditions and find that consumers are more likely to make a pur-
chase of a winter jacket on a low-temperature day and return a winter jacket on a high-temperature
day. Petrikait˙ e (2018) proposes an analytical model to explain consumers’ sequential shopping
behaviors and product return decisions. R¨ ollecke and Huchzermeier (2016) find evidence that a
free return policy reduces consumers’ motivation to deliberate; Dzyabura, El Kihal, and Ibragimov
(2018) apply a machine learning algorithm to predict consumer return rates using product images.
To the best of our knowledge, our work is among the first attempts to study the relationship
between purchase timing and product returns. We draw a similar conclusion to R¨ ollecke and
Huchzermeier (2016) while taking a different approach: R¨ ollecke and Huchzermeier (2016) use an
experimental method whereas our work takes a structural modeling approach. It has allowed us to
68
not only justify this relationship between purchase timing and product returns, but also to quantify
this relationship and provide profit implications to merchants. Our work is also closely related to
E. T. Anderson et al. (2009): we extend their model from a static setting to a dynamic setting as
our research context is daily deal e-commerce and this setup better aligns with the institutional
features of a daily deal site.
Our study also contributes to the existing literature on daily deal or group-buying industries.
A stream of literature focuses on the interesting phenomenon of “consumer fatigue”: for example,
in a survey-based study, Dholakia and Kimes (2011) find evidence against “daily deal fatigue”;
using search history data on a daily deal site, Hu, Dang, and Chintagunta (2019) find a decreasing
click-through rate, but an increasing purchase rate conditional on clicking. Other papers focus on
social interaction among consumers: Jing and Xie (2011) provide an analytical model comparing
the profitability of a group-buying model with a traditional selling website; Luo, Andrews, Song,
and Aspara (2014) focus on the effect of deal popularity on the two-phase decisions made by con-
sumers under the empirical context of Groupon.com; Wu, Liang, and Chen (2015) use a structural
approach to study the economic value associated with daily deal websites. None of these papers
investigates consumer behaviors in product returns. Our unique proprietary data set, along with
our structural model, has allowed us to fill this gap.
2.3 Data and Reduced-form Regression Analysis
We start by introducing our data. We obtain our data through collaboration with a leading Chinese
e-commerce company, which is the main marketplace for daily deals. As the company wishes to
69
remain anonymous in the study, we will call it “the company” or “the website” throughout the
paper.
Launched in 2008, the website started with the business idea of selling selected luxury brands
at a steep promotion for a fleeting period of time, the so-called “sales events.” They started their
business by selling women’s clothing and soon expanded to other categories, such as cosmetics,
baby products, and home accessories. A few American counterparts that operate in a similar man-
ner include Gilt, Rue La La, and Woot. Up to 2018, the website had organized sales events for
more than 20,000 brands. They have more than 300 million enrolled users, resulting in a yearly
order number of about 400 million and a yearly revenue of about 80 billion Chinese Yuan (hence-
forth, CNY). The website has about 3 million daily active users. In 2009, the website launched its
mobile application, allowing consumers to have accesses to deals from their smartphones.
The website organizes sales events by category. In their home page, they list a few categories
for consumers to choose from, including women’s clothing, baby products, and electronic appli-
ances. If consumers are interested in any of the listed categories and click on their links, they enter
a web page organized by sales events. In each of the sales event, the website lists the lower bound
of percentage of promotions in that sales event (e.g., up to 80% off). For each sales event, the
website demonstrates its brand name, a brand photo, a brief description of what products are on
sale, and the highest percentage of promotional discounts (e.g., up to 80% off) of the sales event.
Consumers have the option to subscribe to their favorite brands. If they subscribe to a brand,
they receive a notification from the mobile application or through email at the start of a sales event.
If they are interested in a particular product from a sales event, they have the option to put the
product into their shopping carts for further consideration before the sales event starts, but they do
not have the access to purchase it until the sales event starts.
70
When consumers further click through to the web page of a product, the website displays one
photo associated with it and lists several important product attributes, such as price and promotional
discount. Moreover, they list how much time is left until a sales event ends.
The website implements a generous product return policy: they accept free returns uncondi-
tionally within 7 days after ordered products get delivered. Although the website does not directly
cover the shipping costs, they reimburse consumers with store credits. Each time a consumer re-
turns a product, she obtains a 10 CNY reimbursement of store credits. It covers a large portion of
the return shipping costs (about 10 to 20 CNY). Return policies vary by product categories: for
example, the website requires returned clothes to be unworn, unwashed, and undamaged, and to
include its original tags and free gifts associated with the purchases. Refunds are issued within 3
business days of the delivery of returned product. In general, it takes 10-15 days for the refund to
be posted to the consumer’s financial account (e.g., credit card or Ali pay). Consumers are likely
to stay informed of this return policy when they make their purchases. The website highlights its
free return policy on every web page. Besides its free return policy, the website also advertises
that there are no fake products sold, as well as their selling strategy: offers are available for only a
limited time.
We choose women’s clothing to be the empirical context as we perceive this category to be
more vulnerable than others to the issue of product returns. For example, when a consumer decides
which t-shirt to buy, she needs to “touch and feel the product as opposed to just viewing the digital
image from a computer or smartphone screen. This means that a consumer experiences a larger
fit uncertainty when she buys women’s clothes online than from a brick-and-mortar store. The fit
shock comes from physical considerations, e.g., size, or experiential reasons, e.g., color or style.
We obtain a sample panel of 5,000 consumers who have made at least one purchase of women’s
71
clothing from the website from Jan 1st, 2017 to Jun 30th, 2017. For each consumer, we obtain
a detailed purchase and return history. In particular, we collect the date and time they made the
purchase, which sales events they participated in, the start and end date and time of all sales events,
the price paid for the product, the market price, and the listed promotional discount. In most cases,
sales events last for three days on this website. For sales events that last over three days, over 80%
of the purchases happen within the first three days. Hence, we choose sales events that last for
three days to be our research focus.
We provide the summary of statistics in Table 2.1. There are five categories within “women’s
clothing”: tops; bottoms; accessories; skirts and dresses; and matching sets. These five categories
contain 32 sub-categories, such as jackets, jeans, shorts, skirts, etc. We control for category fixed
effects and sub-category fixed effects in our reduced-form analyses. We find that the percentages
of promotional discounts offered on the website are very steep: on average, a product is sold at a
72% discount of its market price. The average price charged by the website on women’s clothing
is 264 CNY (approximately $35), and there are 17 Stock Keeping Units (henceforth, SKUs) that
are on sale for each sales event. On average, consumers purchase 1.4 items per order, in order to
reduce shipping costs. We find that the return rate on this website is high, but it is consistent with
anecdotal evidence on the return rate of e-commerce platforms in general.
We collect a few measures on the competition intensity of our sample period. We include
three measures: the number of competing sales events (#SalesEvents); the number of SKUs
(#SKUs); and average percentage of promotional discounts (average %promotion). We define
the the number of sales events or SKUs as the number of sales events or SKUs which sell products
in the women’s clothing category on the same day. We useaverage %promotion to represent the
average percentages of promotional discounts of these products. The average number of competing
72
Table 2.1: Summary Statistics
Statistic Mean St. Dev. Min Max
Price (CNY) 264.387 300.047 12.000 6,990.000
Promotions(%) 0.720 0.156 0 0.980
Return 0.273 0.427 0 1
# Items 1.407 0.693 1 6
Purchase timing 1.593 0.755 1 3
# SKUs 17.579 16.544 1 81
Competing # SalesEvents 104 185.816 0 516
Competing # SKUs 969.4 16.239 1 7,069.0
Competing promotions(%) 0.447 0.194 0.025 0.788
sales events is 104 per day and the average number of competing SKUs is 969 per day. We draw
a histogram plot on the number of purchases across time in Figure 2.1. We find the number of
purchases varies across time, but the density plot does not exhibit a strong seasonal pattern. The
number of purchases spreads relatively evenly both within each month and across month.
In Figure 2.2, we provide a histogram plot on consumers’ purchase timing. We find that con-
sumers buy earlier than later: about 50% of sales volume is concentrated on the first day of sales
event. From the second day and on, the sales volume starts to diminish, and the third day of sales
events generates the lowest sales volume. We also find that the number of product returns declines
with day of purchase, as we show in Figure 2.3. More importantly, we find that consumers who
buy earlier are on average more likely return a product than those who buy later, as suggested by
the declining return rates across days. We demonstrate this pattern in Figure 2.4. As we can see
from the figure, the return rate is 30% on the first day of sales events, dropping to 27% on the
second day, and to 23% on the third day.
We provide a time-series plot on the competition intensity across time in Figure 2.5. All time-
series plots fluctuate over days with some lag effects from the past day’s pattern, suggesting that
73
Figure 2.1: Number of Purchases Across Time
Figure 2.2: Number of Purchases by Purchase Date
74
Figure 2.3: Number of Returns by Purchase Date
Figure 2.4: Return Rate by Purchase Date
75
an auto-regression model with a one-period lag may be a good characterization of the change of
competition pattern over time. From the plot, we can see that the number of new sales events and
number of new SKUs co-move across time with a strong correlation.
76
Figure 2.5: Competition Intensity Across Time
77
To test the hypothesis that a more competitive daily deal environment will make the consumer
delay her purchase, we conduct the following regression analyses. Lettingy
i,j,t
1 if a consumer
delays her purchase conditional on not purchase from the previous period andy
i,j,t
0 otherwise,
we estimate
y
i,j,t

$
'
'
'
&
'
'
'
%
1, ify

i,j,t
0;
0, ify

i,j,t
¥ 0,
(2.1)
where
y

i,j,t
κ
1
comp
j,t
κ
2
average %promotion
j,t
seasonalityweekofday
t

i,j,t
. (2.2)
We control for weekofday
t
effect, which is a dummy variable that controls which day it
is in a week. We include three measures of competition level comp
j,t
in our regression model:
the number of new sales events, #SalesEvents, the number of new SKUs, #SKUs, and the
average percentage of promotional discounts for the new products, average %promotion. As
#SalesEvents
j,t
and #SKUs
j,t
exhibit strong collinearity in our data, we run two regressions
separately. We present our regression results in Tables 2.2 and 2.3.
Our results suggest that there is a positive correlation between a consumer’s decision to delay
her purchase and a more competitive environment on the daily deal website: #SalesEvents
j,t
,
#SKUs
j,t
, andaverage %promotion all have positive and significant explanatory powers for the
78
Table 2.2: Reduced-form Regression Results of Number of Sales Events on Purchase Timing
Dependent variable:
If delay purchase
(1) (2) (3)
#SalesEvents 0.023

(0.004) 0.023

(0.004) 0.023

(0.004)
Average(%)promotion 0.227

(0.098) 0.215

(0.098) 0.242

(0.104)
Constant 0.810

(0.079) 0.798

(0.078) 0.820

(0.087)
Category fixed effects No Yes Yes
Subcategory fixed effects No No Yes
R
2
0.094 0.094 0.094
Adjusted R
2
0.093 0.093 0.092
Note:

p 0.1;

p 0.05;

p 0.01
Table 2.3: Reduced-form Regression Results of Number of SKUs on Purchase Timing
Dependent variable:
If delay purchase
(1) (2) (3)
#SKUs 0.016

(0.003) 0.016

(0.003) 0.014

(0.003)
Average(%)promotion 0.175

(0.098) 0.180

(0.099)
Constant 0.773

(0.078) 0.778

(0.079) 0.631

(0.030)
Category fixed effects No Yes Yes
Subcategory fixed effects No No Yes
R
2
0.094 0.094 0.093
Adjusted R
2
0.093 0.093 0.092
Note:

p 0.1;

p 0.05;

p 0.01
79
consumer’s decision to delay her purchase. Our regression results are robust to controlling for
category fixed effects and sub-category fixed effects.
To better understand factors related to consumer product returns, we propose the following
binomial logit regression model:
r
i,j,t

$
'
'
'
&
'
'
'
%
1, ifr

i,j,t
0;
0, ifr

i,j,t
¥ 0,
(2.3)
where
r

i,j
αp
j
β
1
promotionp%q
j
γ#Items
j,t
daysofEvent
j,t

i,j,t
, (2.4)
r
i,j
is a dummy variable measuring a consumer’s decision on whether to return a product;
p
j
is the price charged by the merchant on the daily deal site; promotionp%q
j
is the percentage
promotional discount of this product, measured by dividing the difference between p
j
and the
market price by the market price; #Items
j,t
is the number of items ordered by a consumer on
this purchase, which may be correlated to consumers’ product return decisions;daysofEvent
j,t
refers to the day of a sales event the consumer makes her purchase on. We further control for
category fixed effects, and sub-category fixed effects. We report the regression results in Table 2.4.
80
Table 2.4: Reduced-form Regression Results of Purchase Timing on Return Probability
Dependent variable:
If return
(1) (2) (3)
Price 0.0002

(0.00002) 0.0001

(0.00002) 0.0001

(0.00002)
Promotion (%) 0.100

(0.032) 0.041 (0.032) 0.041 (0.034)
Purchase timing 0.019

(0.007) 0.013

(0.007) 0.012

(0.007)
# Items 0.051

(0.006) 0.050

(0.006) 0.050

(0.006)
Constant 0.258

(0.028) 0.202

(0.029) 0.062 (0.052)
Category fixed effects No Yes Yes
Subcategory fixed effects No No Yes
Observations 7,536 7,536 7,536
R
2
0.028 0.047 0.055
Adjusted R
2
0.028 0.046 0.051
Residual Std. Error 0.425 (df = 7531) 0.421 (df = 7527) 0.420 (df = 7500)
F Statistic 54.489

(df = 4; 7531) 46.071

(df = 8; 7527) 12.513

(df = 35; 7500)
Note:

p 0.1;

p 0.05;

p 0.01
81
Our regression results are broadly consistent in terms of signs and significance levels across
different models. As expected, our regression results suggest that there is a positive association
between the number of items purchased per order and the consumer’s decision to return a product.
We find that the price of a product is positively correlated with a consumer’s return decision while
the percentage of promotion is negatively correlated with it. These results suggest that consumers
are on average more likely to return products with high prices and lower percentages of promotion.
Most importantly, our regression results suggest that there is a negative significant relationship
betweendaysofEvent
j,t
and product returns. That is, consumers who buy earlier are more likely
to return products relative to consumers who buy later.
2.4 Model
Motivated by these reduced-form regression results, we build an integrative model to explain con-
sumer purchase timing and return decisions in the context of daily deals sites. In our model, a fully
rational consumer is informed about the start of a sales event. She decides whether to purchase
from this sales event, and if so, when to make her purchase. Conditional on purchasing, she re-
ceives the product with some time delay. After receiving it, she decides whether to keep or return
the product. She is forward-looking and solves an optimal stopping problem to decide her purchase
timing. When she makes her purchase, she fully anticipates the potential return possibility.
We begin presenting our model by introducing the notation and the timing of the game. Let
i 1, 2,...,I index consumers, andj 1, 2,...,J index different sales events. Each sales event
j has equal length ofT 8 time periods. At the beginning of the initial period, i.e.,t 1, all
consumers receive a notice regarding the start of a sales eventj. In all time periods except the last
82
onet T , the consumer decides between buying it now and waiting for the next period. In the
last period, the consumer decides to buy the product or not and the sales event ends. Whenever
a consumer decides to buy a product, she exits the market. After she receives the product at the
periodT 1, she makes the decision to keep it or to return it. We illustrate a consumer’s purchase
timing decision in Figure 2.4. In our illustrative example, the consumer decides to purchase in the
2nd time period, and therefore exits the market at that period and enters a post-purchase evaluation
stage (T 1). Note that we model a consumer’s purchase timing decision as an optimal stopping
problem; that is, the game does not stop until she makes the purchase or it is the final period.
Whenever she makes her purchase, she exits the market and we assume she does not come back.
The fit shock gets realized in the post-purchase stage, and she can return the product with some
cost.
Figure 2.6: Consumer Purchase Timing
t = 1 t = 2
delay purchase
purchase
exit market
post-purchase
evaluation
delay purchase
t = 3
event
ends
exit
market
The consumer has the desire to purchase the product right away to consume it early. However,
there are two incentives for a consumer to make a strategic delay in a sales event: first, she can view
more incoming sales events by delaying her purchase, i.e., the competition effectC
j,t
; second, she
can be more sure about her purchases by thinking twice, i.e., the deliberation effect D
j,t
. Since
83
Figure 2.7: An Illustrative Figure of Rolling Sales Events
sales events are coming out on a daily basis, a consumer can wait and see if new sales events are
better at matching her tastes compared with the one she currently considering. We use Figure 2.7
to illustrate rolling sales events. For example, in period 1, sales event 1 starts. Notice that sales
event 1 lasts for three periods. As we can see, sales event 2 and 3 get scheduled to start in period
2. If the consumer who is interested in sales event 1 chooses to delay in period 1, she can enjoy the
new sales events 2 and 3. Also, when she is considering for a longer time, she is more deliberate
about her purchase. Later in this section, we use mathematical symbols to show these two driving
forces make a consumer delay her purchase.
In Figure 2.8 we show a consumer’s decision tree for purchasing and returning a product.
Specifically, the consumer faces two shocks: an order shock and a fit shock. The order shock gets
realized when the consumer makes her purchase; the fit shock gets realized when the consumer
receives the product she has purchased. We assume the two shocks are independent of each other.
When a consumer purchases a product from the daily deal site, she does not know how the ex-post
fit shock gets realized and therefore makes her purchase decision based on her expected utility.
84
Figure 2.8: A Consumer’s Product Purchase and Return Decisions
purchase
decision
purchase
keep
good fit shock
return
bad fit shock
good order shock
not purchase
bad order shock
We model a consumer’s decisions in reverse order. First, we consider consumer i’s return
decision. Conditional on purchasing, consumeri makes a decision to keep it or to return it. Let
k 0, 1 denote a consumer’s return or keep decision, respectively. Consumer i’s alternative-
specific pay-off function at timet is given by
W
k
i,j,t

$
'
'
'
&
'
'
'
%
μ
j
D
i,j
ptqe
1
i,j,t
, if keep (k = 1);
α
i
pp
j
cqe
0
i,j,t
, if return (k = 0).
(2.5)
where μ
j
is an intrinsic value this consumer gains from consuming this product; D
i,j
ptq is a
function with respect tot capturing how deliberate the consumer’s purchase decision is, specifi-
cally, we assume
D
i,j
ptqζ
1
1pt 1qζ
2
1pt 2q;
ζ
1
andζ
2
are deliberation coefficient parameters for day 1 and day 2; α
i
is a heterogenous price
coefficient representing the marginal utility of income;c¥ 0 is a positive return cost; the idiosyn-
cratic error termse
k
i,j,t
are assumed to be i.i.d. across consumers, sales events, and time periods;
85
we further assume the idiosyncratic error terms have type-I extreme value distribution of scaling
parameter 1. Equation (2.5) captures the post-purchase evaluation that this consumer makes of the
product: she obtains a utility from consuming it if she keeps the product and receives a full-price
refund except the return costs if she returns it. The consumer makes the decision that maximizes
her utility, therefore the ex-post decisionk

is given byk

argmax
kt0,1u
!
W
k
i,j,t
e
k
i,j,t
)
.
Let EW
i,j,t
denote the ex-ante expected utility function and W
k
i,j,t
denote the corresponding
alternative-specific expected utility function, then we have a closed-form expression for EW
i,j,t
given by
EW
i,j,t
E
e
max
kt0,1u
!
W
k
i,j,t
e
k
i,j,t
)
.58ln
!
¸
kt0,1u
exp

W
k
i,j,t

)
.58ln
!
exp

μ
j
D
i,j
ptq

exp

α
i
pp
j
cq

)
,
(2.6)
where.58 is the well-known Euler-Mascheroni constant. Now let us consider a consumer’s pur-
chase decision. Recall that we assume the consumer does not fully anticipate how the fit shock gets
realized except that she knows the distribution of it. As a result, she makes her purchase decision
based on the expected utility of purchase EW
i,j,t
. We start with a single-period utility and then
incorporate the next-stage keep or return utility into this utility function. The single-period utility
function for consumeri at timet is given by
86
u
b
i,j,t

$
'
'
'
&
'
'
'
%
α
i
p
j
Φ
P
ptq
1
i,j,t
, if purchase (b = 1);
HpC
j,t
;κq
0
i,j,t
, if not purchase (b = 0),
(2.7)
whereα
i
p
j
is a disutility for consumeri to pay pricep
j
; Φ
P
ptq captures the freshness deteri-
oration effect, and we assume
Φ
P
ptqξ
1
1pt 1qξ
2
1pt 2q;
ξ
1
andξ
2
are freshness-based deterioration parameters for day 1 and day 2;HpC
j,t
;κq captures the
competition effect, i.e., the new-coming sales events. We use the same reduced-form specification
as in Ishihara and Ching (2012):
HpC
j,t
;κqκlnpC
j,t
1q, (2.8)
whereC
j,t
is the number of incoming sales events that are competing with sales eventj until time
t
2
. Here we do not allow the consumer to make a decision to purchase multiple items at a time
instead use a variable number of sales events to capture the potential outside options this consumer
is facing. We believe this variable effectively captures the competition effect and significantly
reduces computational costs.
A forward-looking consumer’s problem is to decide whether to purchase, and if so, when to
purchase. Subsequently, she makes an optimal keep/return decision. Note that she has the option
2
Note t refers to days of sales event. As the competition pattern may differ by calendar dates, we include a
subscriptionj.
87
to delay her purchase at every period except the last one. Hence we have to derive her value
function of continuation of the game. We solve this dynamic game through backward induction.
LetV
i,j
ps
j,t
q be the ex-ante integrated value function for consumeri at timet andV
b
i,j
ps
j,t
q be the
corresponding alternative-specific value functions of actionb. We first start with the last period,
i.e.,tT , we specify the alternative-specific value function as follows:
V
b
i,j
ps
j,T
q
$
'
'
'
&
'
'
'
%
α
i
p
j
δEW
i,j,T
, if purchase (b = 1);
HpC
j,T
;κq, if not purchase (b = 0).
(2.9)
whereδ is a discount factor. A consumer who stands in the final period of the game faces the
decision of whether or not to purchase the product. If she purchases it, she will receive it in the
next stage for post-purchase evaluation. Therefore, the ex-ante alternative-specific value function
of purchasing is given by the single period utility plus a discounted expected utility of purchase
EW
i,j,T
. Otherwise she decides not to purchase and gains utility from consuming her outside
option and there is no continuation of the game. Then we consider the consumer’s problem in all
periods other than the last, i.e.,t T :
V
b
i,j
ps
j,t
q
$
'
'
'
&
'
'
'
%
α
i
p
j
δ
Tt1
EW
i,j,t
, if purchase (b = 1);
HpC
j,t
;κqδEV
i,j
ps
j,t1
|s
j,t
q, if wait (b = 0).
(2.10)
A consumer is allowed to delay their purchase to the next period here and therefore the dis-
counted value function of the next period enters the utility for the waiting option. Notice that the
integrated-value functionV
i,j
ps
j,t
q is given by
88
V
i,j
ps
j,t
q max
bt0,1u
!
V
b
i,j
ps
j,t
q
b
i,j,t
)
. (2.11)
and
EV
i,j
ps
j,t1
|s
j,t
q
»
V
i,j
ps
j,t1
qfps
j,t1
|s
j,t
qds
j,t
. (2.12)
Lets
j,t
pC
j,t
,t,
i,j,t
,e
i,j,t
q be the state variables that are relevant to the value function. The
state variables
j,t
can break into two parts: observed state variables
o
j,t
pC
j,t
,tq and unobserved
state variables
u
j,t
p
i,j,t
,e
i,j,t
q. Among the two observed state variables,t is deterministic.
We assume the transition of the state variable C
j,t
follows a first-order Markov process, that
is, a consumer’s prediction about future state variablesC
j,t1
only depends on her previous beliefs
and current state variableC
j,t
. Specifically, we follow Liu and Ishihara (2017) and assume
lnpC
j,t1
q|lnpC
j,t
qNpablnpC
j,t
q,σ
2
C
q. (2.13)
The i.i.d. and parametric assumption on the idiosyncratic error terms allows us to have an
analytical solution for the purchase and return probability. Conditional on not purchasing from the
previous period, the conditional choice probability for consumeri to purchase from sales eventj
at timet is given by
Prpb
i,j,t
1|s
o
j,t
,b
i,j,t1
0q
exppV
1
i,j
ps
o
j,t
qq
°
dt0,1u
exppV
d
i,j
ps
o
j,t
qq
(2.14)
Conditional on purchasing, the probability of consumeri keeping a productj is given by
89
Prpk
i,j,t
1|s
o
j,t
,b
i,j,t
1q
exppW
1
i,j
ps
o
j,t
qq
°
kt0,1u
exppW
k
i,j
ps
o
j,t
qq
(2.15)
Because the number of sales events is large, the number of observations gets large if we take
consumers who do not purchase into consideration. There are more non-purchase observations
than purchase observations, resulting in data set with many zeroes. To cope with this challenge, we
only use consumers who have purchased in sales events to perform our estimation. We normalize
the purchase probabilities accordingly, and they are then given by
Prpb
i,j,t
1|s
o
j,t
,b
i,j,t1
0,
¸
t
b
i,j,t
1q
Prpb
i,j,t
1|s
o
j,t
,b
i,j,t1
0q
Prp
°
t
b
i,j,t
1q
. (2.16)
In this way, we ensure the sum of the probabilities of observing a purchase in at least one day of a
sales event is 1.
2.5 Estimation
In this section, we lay out our empirical identification strategy and derive the likelihood function
that we use for estimation.
2.5.1 Empirical identification
Let ΘpΘ
D
, Θ
C
q denote the set of parameters to be estimated. Note that the set of parameters
Θ consists of two parts: Θ
D
and Θ
C
. Θ
D
refers to the parameters related to observed consumer
decisions and Θ
C
refers to the parameters that are related to the competition pattern over time.
90
Let Θ
C
pa,b,σ
C
q denote the set of parameters related to the state transition probability
of the state variable competition. Relying on the first-order Markov process assumption and the
time-series on how competition changes over time, we can identify parametera,b, andσ
C
.
Now we discuss the how the model’s structural features and data variations can help us iden-
tify the parameters Θ
D
pμ,κ,γ,β,α,ζ,ξq. We re-parameterize the return cost c and use the
number of items per order to estimates return costs associated with it. Specifically, we assume
cγ #Items. We include the discount parameterβ in the utility of keeping a product to capture
the fact that consumers in general like products with high discounts in daily deal transactions. The
competition parameterκ is identified from the variation in consumer purchase timing. For exam-
ple, if we observe consumers are on average more likely to delay the purchase timing under a more
competitive environment, then the competition parameter κ is likely to be positive. We use the
variation in the return probability to recover the return cost parameterγ. If we observe consumers
with more number of items per order are more likely to return a product, conditional on all else
equal, then the return cost parameterγ is likely to be negative. The price sensitivity parameterα
and the promotional parameterβ are also recovered from a consumer’s product return decision. If
consumers on average are more likely to return products with a higher price and lower percent-
age of promotion, then the price sensitivity parameter α is likely to be negative and promotion
parameterβ is more likely to be positive.
2.5.2 The Log-likelihood function
Now we derive the log likelihood function we used in our estimation. We start by introducing
some notation. LetC
o
tC
o
j,t
,@j,@t¤ Tu be the observed set of competition levels, andD
o
i

91
tD
o
i,j,t
,@j,@t¤Tu denote the observed set of decisions made by consumeri. With Θ denoting the
parameters to be estimated, we construct the log-likelihood function for the sample as follows:
LLpΘq
¸
i
logpL
d
pD
o
i
|C
o
; ΘqqlogpL
c
pC
o
|Θ
C
qq, (2.17)
withL
d
pD
o
i
|C
o
; Θq denoting the likelihood we obtain from consumer decisions andL
c
pC
o
|Θ
C
q
denoting the likelihood we obtain from the transition probability of competition. Note that these
two components are additively separable, and only parameters related to transition probability Θ
C
enter the likelihoodL
c
pC
o
|Θ
C
q.
The individual decision likelihood L
d
pD
o
i
|C
o
; Θq consists of all the decisions made by con-
sumeri from the first periodt 1 to the last periodt T . Then we can derive the likelihood in
the following equation:
L
d
pD
o
i
|C
o
; Θq
¹
t
L
d
pD
o
i,j,t
|C
j,t
,D
i,j,t1
; Θq. (2.18)
The multiplication of likelihood for consumeri in each individual time periodt comes from
the conditional independence assumption we impose on the order-shock error term
i,j,t
across
time periodst. Recall that in our model, the individual decisionD
o
i,j,t
consists of two components:
a purchase decision b
o
i,j,t
and a return decision r
o
i,j,t
. The likelihood function of each individual
decision is specified as follows:
92
L
d
pD
o
i,j,t
|C
o
; Θq
¹
bt0,1u
Prpb
o
i,j,t
b|C
j,t
,b
j,t1
; Θq
1pb
o
i,j,t
bq

¹
rt0,1u
Prpr
o
i,j,t
r|b
i,j,t
1; Θq
1pr
o
i,j,t
rq
.
(2.19)
Our assumptions on the independence fit shock errore
i,j,t
across consumers, sales events, and
time and the two-stage decision show that the probability of return is independent of the probability
of purchase conditional on the purchase decision in time periodt.
Given the first-order Markov process assumption, the likelihood function for the transition
probability of competition levelC
j,t
can be decomposed as follows:
L
c
pC
o
|Θ
C
q
¹
j
¹
t
PrpC
j,t
|C
j,t1
; Θ
C
q. (2.20)
We estimate the parameters through a two-stage procedure to reduce computational costs as
in Rust (1994). We first estimate the parameters that enter the transition probability function for
competition and then use the estimates to recover the rest of them. We apply a maximum simulated
likelihood to obtain the parameter estimates. As discussed in Magnac and Thesmar (2002) that
a discount factor cannot be identified generically with a Rust (1994) type of model setting, we
therefore assume the discount factorδ to be 1 in our estimation.
2.6 Results
We apply our estimation method to our data and present our parameter estimates in this section.
We further assume the discount factor to be 1, that is, the consumer is indifferent between making
93
a purchase today and tomorrow, conditional on all else being equal. We set the terminal period
T to be 3 as it aligns with our empirical context. Note that what we observe in our sample data
is that: (1) conditional on purchase from a deal j, which day the consumer makes her purchase
on; (2) conditional on making a purchase, whether or not the consumer returns a product. As we
discussed in the previous section, we have modified our likelihood function to accommodate the
data.
Because T is small, we can solve the model through backward induction, and because of the
additive separability of the likelihood (see Equation (2.17)), we estimate the transition process
separately. We plug the parameters estimates from the transition probability into the likelihood of
consumer decisions and estimate parameters related to purchase timing and product return deci-
sions. We use a limited memory BFGS algorithm to conduct the outer-loop optimization (Byrd,
Lu, Nocedal, and Zhu, 1995). We also use multiple starting values to ensure that the algorithm
does not converge to local maxima that are not global maxima. The parameter estimates of the
structural model are reported in Table 2.5. The parameters of the competition transition process
are reported in Table 2.6.
Table 2.5: Utility Primitives
Parameters Notations Estimates Std. Err.
Product intercept parameter μ 1.190

0.061
Promotion parameter β 0.009 0.028
Price coefficient parameter α 0.440

0.029
Return cost parameter γ 0.143

0.026
Deliberation coefficient parameter t = 1 ζ
1
0.204

0.069
Deliberation coefficient parameter t = 2 ζ
2
0.122 0.075
Freshness-based deterioration parameter t = 1 ξ
1
0.59

0.014
Freshness-based deterioration parameter t = 2 ξ
2
0.29

0.015
Competition coefficient parameter κ 0.126

0.019
Note:

p 0.1;

p 0.05;

p 0.01
94
Table 2.6: State Transition Probability Parameters
Parameters Estimates Std. Err.
a 0.572

0.003
b 0.636

0.004
σ
C
0.423

0.006
Note:

p 0.1;

p 0.05;

p 0.01
The signs and magnitudes of our parameter estimates are aligned with our reduced-form regres-
sion results. Our parameters of estimates on state transition probability suggest that competition
level evolves over time with an AR(1) process with a positive coefficient. For the utility prim-
itives, all parameters except the promotion parameter β have exhibited statistical significance at
1% level, suggesting the variation can help explain the dependent variable. Note we find a nega-
tive return cost parameterγ, suggesting that there is a negative relationship between the number
of item purchases and return costs. We find a positive competition parameter κ, suggesting that
consumers who face a large number of competing merchants are more likely to wait to the next
period. We also find that deliberation coefficient parameterζ has positive explanatory power of the
data, suggesting that as consumers may think more deliberately if they decide to buy later.
In Table 2.7, we compare model fit by demonstrating a comparison between the probability of
purchase and probability of return from both the data and the model. We compute the predicted
probability of purchase and return using our estimated primitives. We show that our predicted
probabilities broadly align with the pattern of data, suggesting the validity of our proposed model.
Table 2.7: Model Fit: Purchase and Return Probabilities
Time Data Buy Est. Buy Data. Return Est. Return
1 0.512 0.580 0.293 0.347
2 0.234 0.221 0.249 0.241
3 0.155 0.169 0.232 0.151
95
2.7 Counterfactual Analysis
A common selling strategy among e-commerce merchants is to urge consumers to buy early. Mer-
chants are willing to give a small additional discount to consumers who buy early. In our counter-
factual analysis, we aim to understand the profitability of such a pricing schedule when consumers
can return a product.
Before we discuss our counterfactual simulation exercise in detail, we need to specify the profit
function for a merchant. Our profit function is different from past literature in the sense that we
incorporate consumers’ product return decisions in it. In particular, we assume a merchant’s profit
function to be
π
j
M
j
Σ
t
Prpb
i,j,t
1qPrpk
i,j,t
1|b
i,j,t
1qp
j,t
, (2.21)
whereM
j
refers to how many consumers are interested in productj,Prpb
i,j,t
1q is their prob-
ability of purchase at time periodt, andPrpk
i,j,t
1|b
i,j,t
1q is the probability of keeping that
product conditional on purchase, and p
j,t
is the price of product j charged at time t. The profit
function is given by the probability of purchase multiplies by the probability of keeping a product
and the price charged at timet,p
j,t
. Note that the merchant successfully sells a product if and only
if: (1) the consumer decides to purchase; (2) she decides to keep the product after receiving it.
Therefore, ignoring product returns may lead to a biased estimation of the profit function.
With the primitives estimated from our structural model, we conduct the following counterfac-
tual analysis: we allow a different pricing schedule, i.e., varying the price of a product by different
days of a sales event, and see how the profit of the merchant changes accordingly, keeping the
number of consumers who are interested in this product unchanged.
96
In our benchmark model, we assume merchants employ a uniform pricing schedule, that is,
p
j,t
p
j
,@t. This is consistent with our empirical context: in our data, merchants charge the same
price on a product from the beginning to the end of a sales event. In our first experiment, we allow
merchants to take a price drop at the beginning of a sales event. In this way, we have a increasing
pricing schedule, i.e., p
j,t
p
j,t1
,@t. Next, we manipulate the pricing schedule to allow for a
decline in price with respect to time, i.e.,p
j,t
¡p
j,t1
,@t.
2.8 Conclusion
In this paper, we study consumer purchase timing and product returns in the context of daily
deal sites. We find that consumers generally buy earlier rather than later in sales events, and that
consumers who buy earlier are more likely to return their purchases. Motivated by these interesting
patterns, we develop an integrative model to explain consumer decisions on purchase timing and
product returns. There are two interesting features in our model. First, consumer knowledge of
product fit gets realized in the post-purchase stage, and she can return the product with some
cost. She makes the purchase (order) decision based on her expected utility, considering the next-
stage return possibility. Second, she is forward-looking and solves an optimal stopping problem
in deciding when to purchase. Delaying a purchase allows her to see newly posted offers from
the website and have a longer time consider it. We take the proposed model to a panel of 5,000
consumers who made purchases of women’s clothing on a popular daily deal website from Jan 1st,
2017 to Jun 30th, 2017. Our parameter estimation successfully replicates the interesting empirical
patterns. We find that consumers are in general price sensitive and that competition is positively
related to their decisions to delay purchases. Our counterfactual simulation exercise suggests that
97
ignoring consumer product returns will lead to biased estimation of merchants’ profits. We quantify
the tradeoffs merchants face in terms of urging consumers to buy early and the increase in product
return rates and explore the profit implications for different pricing schedules.
To the best of our knowledge, our work is among the first to link consumer purchase timing
with their product return decisions. We show that this relationship is important to merchants and
has important implications for designing a more optimal pricing schedules. Without incorporating
consumer product returns, merchants may obtain a biased estimation of profits and implement a
sub-optimal pricing schedule.
Our paper is not without limitations. For now, we have not yet considered consumer het-
erogeneity. Allowing consumer heterogeneity may allow merchants to design coupon-targeting
programs to improve their profits. We plan to incorporate it into our future work. Second, we have
not taken the change in purchase incidence into consideration when we perform our counterfactual
analysis. We only consider the shift in timing and its impact on return rate. These limitations offer
great opportunities for extending the present work.
98
References
Anderson, E. T., Hansen, K., and Simester, D. (2009). The option value of returns: Theory and
empirical evidence. Marketing Science, 28(3), 405–423.
Anderson, M., and Magruder, J. (2012). Learning from the crowd: Regression discontinuity
estimates of the effects of an online review database. The Economic Journal, 122(563),
957–989.
Archak, N., Ghose, A., and Ipeirotis, P. G. (2011). Deriving the pricing power of product features
by mining consumer reviews. Management Science, 57(8), 1485–1509.
Berry, S., Levinsohn, J., and Pakes, A. (1995). Automobile prices in market equilibrium. Econo-
metrica: Journal of the Econometric Society, 841–890.
Binmore, K., Rubinstein, A., and Wolinsky, A. (1986). The nash bargaining solution in economic
modelling. The RAND Journal of Economics, 176–188.
Bower, A. B., and Maxham III, J. G. (2012). Return shipping policies of online retailers: normative
assumptions and the long-term consequences of fee and free returns. Journal of Marketing,
76(5), 110–124.
Byrd, R. H., Lu, P., Nocedal, J., and Zhu, C. (1995). A limited memory algorithm for bound
constrained optimization. SIAM Journal on Scientific Computing, 16(5), 1190–1208.
Capgemini. (2015). Capgemini’s 2015 automotive online research report reveals consumers’ in-
creasing demand for a more personalized car buying experience. Retrieved fromhttps://
www.capgemini.com/cn-zh/news/capgemini-cars-online-2015report
-highlights-growing-consumer-demand-for-a-more-personalized
-car-buying-experience/#
Che, Y.-K. (1996). Customer return policies for experience goods. The Journal of Industrial
Economics, 44(1), 0021–1821.
Chen, Y ., and Xie, J. (2008). Online consumer review: Word-of-mouth as a new element of
marketing communication mix. Management Science, 54(3), 477–491.
Chen, Y ., Yang, S., and Zhao, Y . (2008). A simultaneous model of consumer brand choice and
negotiated price. Management Science, 54(3), 538–549.
Chevalier, J. A., and Mayzlin, D. (2006). The effect of word of mouth on sales: Online book
reviews. Journal of Marketing Research, 43(3), 345–354.
Conlin, M., O’Donoghue, T., and V ogelsang, T. J. (2007). Projection bias in catalog orders.
American Economic Review, 97(4), 1217–1249.
Consumer Reports. (2013). Tips on how to bargain and save money - consumer reports. Retrieved
99
from https://www.consumerreports.org/cro/magazine/2013/08/how
-to-bargain/index.htm
Dholakia, U. M. (2011). What makes groupon promotions profitable for businesses? Available at
SSRN 1790414.
Dholakia, U. M., and Kimes, S. E. (2011). Daily deal fatigue or unabated enthusiasm? a study of
consumer perceptions of daily deal promotions. A Study of Consumer Perceptions of Daily
Deal Promotions (September 11, 2011).
Dzyabura, D., El Kihal, S., and Ibragimov, M. (2018). Leveraging the power of images in predict-
ing product return rates.
DHaultfœuille, X., Durrmeyer, I., and F´ evrier, P. (2019). Automobile prices in market equilibrium
with unobserved price discrimination. The Review of Economic Studies, 86(5), 1973–1998.
Ellis, C. J., and Fender, J. (1985). Wage bargaining in a macroeconomic model with rationing.
The Quarterly Journal of Economics, 100(3), 625–650.
Federal Trade Commission. (2019, Sep). Buying a new car. Retrieved from https://www
.consumer.ftc.gov/articles/0209-buying-new-car
Friberg, R., and Gr¨ onqvist, E. (2012). Do expert reviews affect the demand for wine? American
Economic Journal: Applied Economics, 4(1), 193–211.
Ghose, A., Ipeirotis, P., and Sundararajan, A. (2007). Opinion mining using econometrics: A case
study on reputation systems. In Proceedings of the 45th annual meeting of the association
of computational linguistics (pp. 416–423).
Grennan, M. (2013). Price discrimination and bargaining: Empirical evidence from medical
devices. American Economic Review, 103(1), 145–77.
Gumbel, E. J. (1961). Bivariate logistic distributions. Journal of the American Statistical Associ-
ation, 56(294), 335–349.
Hausman, J., Leonard, G., and Zona, J. D. (1994). Competitive analysis with differenciated
products. Annales d’Economie et de Statistique, 159–180.
Honka, E. (2014). Quantifying search and switching costs in the us auto insurance industry. The
RAND Journal of Economics, 45(4), 847–884.
Hu, M., Dang, I., and Chintagunta, P. K. (2019). Search and learning at a daily deals website.
Marketing Science.
Huang, G. (2010). Posted price and haggling in the used car market.
Huang, Y . (2019). Learning by doing and the demand for advanced products. Marketing Science,
38(1), 107–128.
Inderst, R., and Tirosh, G. (2015). Refunds and returns in a vertically differentiated industry.
100
International Journal of Industrial Organization, 38, 44–51.
Ishihara, M., and Ching, A. (2012). Dynamic demand for new and used durable goods without
physical depreciation: The case of japanese video games. Available at SSRN.
J.D. Power. (2015). 2015 china automotive user media research. Retrieved from https://
japan.jdpower.com/content/2015%E5%B9%B4%E4%B8%AD%E5%9B%BD%E6%
B1%BD%E8%BD%A6%E7%94%A8%E6%88%B7%E5%AA%92%E4%BD%93%E7%A0%94%
E7%A9%B6ams
Jindal, P., and Newberry, P. (2018). To bargain or not to bargain: The role of fixed costs in price
negotiations. Journal of Marketing Research, 55(6), 832–851.
Jing, X., and Xie, J. (2011). Group buying: A new mechanism for selling through social interac-
tions. Management Science, 57(8), 1354–1372.
Kwark, Y ., Chen, J., and Raghunathan, S. (2014). Online product reviews: Implications for
retailers and competing manufacturers. Information systems research, 25(1), 93–110.
Li, X., and Hitt, L. M. (2010). Price effects in online product reviews: An analytical model and
empirical analysis. MIS quarterly, 809–831.
Liu, Y . (2006). Word of mouth for movies: Its dynamics and impact on box office revenue. Journal
of Marketing, 70(3), 74–89.
Liu, Y ., Feng, J., and Liao, X. (2017). When online reviews meet sales volume information: is more
or accurate information always better? Information Systems Research, 28(4), 723–743.
Liu, Y ., and Ishihara, M. (2017). A dynamic structural model of endogenous consumer reviews in
durable goods markets. Available at SSRN 2728524.
Lu, S. F., and Rui, H. (2018). Can we trust online physician ratings? evidence from cardiac
surgeons in florida. Management Science, 64(6), 2557–2573.
Lucking-Reiley, D., Bryan, D., Prasad, N., and Reeves, D. (2007). Pennies from ebay: The
determinants of price in online auctions. The Journal of Industrial Economics, 55(2), 223–
233.
Luo, X., Andrews, M., Song, Y ., and Aspara, J. (2014). Group-buying deal popularity. Journal of
Marketing, 78(2), 20–33.
Magnac, T., and Thesmar, D. (2002). Identifying dynamic discrete decision processes. Economet-
rica, 70(2), 801–816.
McFadden, D. (1989). A method of simulated moments for estimation of discrete response models
without numerical integration. Econometrica: Journal of the Econometric Society, 995–
1026.
Nash Jr, J. F. (1950). The bargaining problem. Econometrica: Journal of the Econometric Society,
101
155–162.
Netzer, O., Feldman, R., Goldenberg, J., and Fresko, M. (2012). Mine your own business: Market-
structure surveillance through text mining. Marketing Science, 31(3), 521–543.
Nevo, A. (2001). Measuring market power in the ready-to-eat cereal industry. Econometrica,
69(2), 307–342.
Ofek, E., Katona, Z., and Sarvary, M. (2011). bricks and clicks: The impact of product returns on
the strategies of multichannel retailers. Marketing Science, 30(1), 42–60.
Padmanabhan, V ., and Png, I. P. (1997). Manufacturer’s return policies and retail competition.
Marketing Science, 16(1), 81–94.
Pavlou, P. A., and Dimoka, A. (2006). The nature and role of feedback text comments in online
marketplaces: Implications for trust building, price premiums, and seller differentiation.
Information Systems Research, 17(4), 392–414.
Petersen, J. A., and Kumar, V . (2009). Are product returns a necessary evil? antecedents and
consequences. Journal of Marketing, 73(3), 35–51.
Petrikait˙ e, V . (2018). A search model of costly product returns. International Journal of Industrial
Organization, 58, 236–251.
Petrin, A., and Train, K. (2010). A control function approach to endogeneity in consumer choice
models. Journal of Marketing Research, 47(1), 3–13.
Ratchford, B. T., and Srinivasan, N. (1993). An empirical investigation of returns to search.
Marketing Science, 12(1), 73–87.
R¨ ollecke, F., and Huchzermeier, A. (2016, 01). Shopping fast and returning plenty: How returns
policies affect online product choices. SSRN Electronic Journal. doi: 10.2139/ssrn.2997567
Rust, J. (1994). Structural estimation of markov decision processes. Handbook of Econometrics,
4, 3081–3143.
Seiler, S., Yao, S., and Wang, W. (2017). Does online word of mouth increase demand?(and how?)
evidence from a natural experiment. Marketing Science, 36(6), 838–861.
Shulman, J. D., Coughlan, A. T., and Savaskan, R. C. (2010). Optimal reverse channel structure
for consumer product returns. Marketing Science, 29(6), 1071–1085.
Shulman, J. D., Cunha Jr, M., and Saint Clair, J. K. (2015). Consumer uncertainty and purchase
decision reversals: Theory and evidence. Marketing Science, 34(4), 590–605.
White, M. C. (2012, Aug). 12 things you should always haggle over. Time. Retrieved
fromhttps://business.time.com/2012/08/13/12-things-you-should
-always-haggle-over/slide/the-art-of-negotiating/
Wood, S. L. (2001). Remote purchase environments: The influence of return policy leniency on
102
two-stage decision processes. Journal of Marketing Research, 38(2), 157–169.
Wu, C., Che, H., Chan, T. Y ., and Lu, X. (2015). The economic value of online reviews. Marketing
Science, 34(5), 739-754.
Wu, C., Liang, Y. S., and Chen, X. (2015). Is daily deal a good deal for merchants? an empirical
analysis of economic value in the daily deal market. An Empirical Analysis of Economic
Value in the Daily Deal Market (June 30, 2015).
Zettelmeyer, F., Morton, F. S., and Silva-Risso, J. (2006). How the internet lowers prices: Evi-
dence from matched survey and automobile transaction data. Journal of Marketing Research,
43(2), 168–181.
Zhang, X., Manchanda, P., and Chu, J. (2017). ‘meet me halfway’: The value of bargaining.
Zhao, Y ., Yang, S., Narayan, V ., and Zhao, Y . (2013). Modeling consumer learning from online
product reviews. Marketing Science, 32(1), 153–169.
Zheng, X., Ren, X., Hong, Y ., Cao, J., and Yang, S. (2018). Understanding the impact of multi-
dimensional ratings on product sales: An information-inconsistency perspective.
103
Appendix A
Appendix to Chapter 1
A.1 Implementation of Kernel-smoothed Frequency Simulator
The implementation of the kernel-smoothed frequency simulator follows Honka (2014), the de-
tailed procedure is listed as follows:
1. We taker 1, 2,...,R draws of
ijt
, where
ijt
follows a Type-I Extreme Value distribution
with mean equal to 0 and scale equal to 1, and label each draw as
r
ijt
.
2. For each
r
ijt
, we calculate the simulated utilityu
r
ijt
. Based on the simulated utility, we define
ω
r
ijt
as the difference in simulated utility between the consumer chosen product and the product
(including the outside option) maximizes her utility among the rest. In particular,
ω
r
ijt

ˇ
U
r
ijt
max
j
1
j
ˇ
U
r
ij
1
t
.
3. For each
r
ijt
, we calculate the conditional purchase probability for productj using scaled
logistic CDF (Gumbel, 1961):
Prt
ˇ
U
r
ijt
pˇ p
ijt
q¥max
j
1
j
ˇ
U
r
ij
1
t
pˇ p
ij
1
t
q|
r
u
1
1exppsw
r
ijt
q
Note that if in our data, we observe product j is chosen by consumer, and under simulation
drawr and estimated parameters, productj maximizes her utility among other choices,w
r
ijt
should
be positive. Under a positivew
r
ijt
and a positive tuning parameters, the kernel density is above
1{2. If under simulation draw r with estimated parameters, product j is not chosen, then the
kernel density is below 1{2. This density function smooths the log-likelihood function and makes
a gradient search method available in our estimation.
4. For each
r
ijt
, we calculate the probability of observing a transaction price p
ijt
. Under
a normally distributed reporting error term, we can write down the probability of observing a
particular reported price as
Pr
r
p˜ p
ijt
η
ijt
p
ijt
qφ
r
p
˜ p
ijt
p
ijt
σ
η
q.
5. We calculate the simulated probability of observing a particular outcome by averaging the
probability in each draw over R draws, i.e.,

PrpO
ijt
|θq
1
R
¸
r
φ
r
p
˜ p
ijt
p
ijt
σ
η
q
1
1exppsw
r
ijt
q
.
6. The simulated log-likelihood function is
SLLpθq
¸
iPI
¸
jPJ
¸
tPT

logp

PrpO
ijt
|θqq

, (A.1)
104
which is the main objective function that we maximize in our estimation.
105
A.2 Additional Tables and Figures
In this section, we present the additional tables and figures.
Table A.1: Variable Definition for Product Attributes
Variable Unit Definition
Weight kg The curb weight of a car
Height mm The length from the wheel center cap to the fender edge of
a car
Fuel Consump-
tion
L Fuel consumed per hundred of kilometers driving
Max Horsepower Ps The max engine horsepower of a car
Width mm The widest point of a car without its mirrors
Length mm The length of a car from front bumper to rear bumper
Country of origin NA The country in which the product is manufactured, takes
value of Germany, Japan, Korea, and United States in our
data.
A.3 Simulation Design
The experimental design for our simulation study in Section 1.3.2.4 is outlined below. There are
10,000 purchase occasions. We consider two cases in which γ
1
0.05 or γ
2
0. The true
parameters are
¯
β 1, α 0.5 and β
R
2. The idiosyncratic error terms are generated
by Gumbelp0, 1q. We generated the marginal cost mc Np1, 1q, and online review as
RNp0, 1q. The bargaining powerλ is generated asλ
expp
¯
βRγq
1expp
¯
βRγq
. We generate consumer
willingness to pay asw
β
R
R
α
and negotiated price aspp1λqwλmcη, whereη
is drawn fromNp0, 1q.
A.4 Additional Proofs
Proof.
˜ ppI Λqp
L
Λτcη, (A.2)
where
ΛfpX
1
βq
exppX
1
βq
1exppX
1
βq
(A.3)
106
Er˜ p|XsErpI Λqp
L
|XsErΛτc|XsErη|Xs
Er˜ p|XspI ΛqErp
L
|Xs ΛτErc|Xs
Er˜ p|XsErp
L
|XsΛErτcp
L
|Xs
lnpErp
L
˜ p|XsqlnpΛqlnpErp
L
τc|Xsq
Er˜ pτc|XspI ΛqpErτcp
L
|Xsq
lnpEr˜ pτc|XsqlnpI ΛqlnpErp
L
τc|Xsq
If we subtract Equation 6 from Equation 8, we have
lnpErp
L
˜ p|XsqlnpEr˜ pτc|XsqlnpΛqlnpI Λq
lnpErp
L
τc|XsqlnpErp
L
τc|Xsq
lnpErp
L
˜ p|XsqlnpEr˜ pτc|XsqX
1
β
lnpErp
L
˜ p|XsqX
1
βlnpEr˜ pτc|Xsq
lnpErp
L
˜ p|XsqX
1
βgpτq
M
X
lnpErp
L
˜ p|XsqM
X
X
1
βM
X
gpτq
M
X
lnpErp
L
˜ p|XsqM
X
lnpEr˜ pτc|Xsq
e
M
X
Erp
L
˜ p|Xse
M
X
Er˜ pτc|Xs
e
M
X
Erp
L
˜ p|Xse
M
X
Er˜ p|Xsτe
M
X
Erc|Xs
whereM
X
IX
1
pXX
1
q
1
X.
Then we have
τ pe
M
X
Erc|Xsq
1
pe
M
X
Erc|Xsq
1
te
M
X
Erp
L
˜ p|Xse
M
X
Er˜ p|Xsu.
τ pErc
1
|Xse
M
X
1
e
M
X
Erc|Xsq
1
te
M
X
Erp
L
˜ p|Xse
M
X
Er˜ p|Xsu.
107

Abstract (if available)

Abstract This dissertation includes two chapters with a focus on different topics in the field of Information Economics and Marketing. The objective of the first chapter is to understand the impact of consumer reviews on product choices when purchase price is negotiated. Although prior studies have analyzed the impact of reviews on consumer baseline preferences, none has explored such an impact on demand when the purchase price is negotiated. Leveraging a comprehensive data set on consumer reviews, this study shows a double-edged sword effect of reviews on demand in such a setting. On the one hand, good reviews increase demand by raising consumer baseline preferences

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Behavioral approaches to industrial organization

PDF

Essays on the luxury fashion market

PDF

Essays on econometrics

PDF

The EITC, labor supply, and child development

PDF

Essays on development and health economics: social media and education policy

PDF

Essays on price determinants in the Los Angeles housing market

PDF

Essays on development economics

PDF

Essays on firm investment, innovation and productivity

PDF

Essays on revenue management with choice modeling

PDF

A structural econometric analysis of network and social interaction models

PDF

Essays on economics of education

PDF

Essays on causal inference

PDF

Measurement error in income and consumption data and economic mobility

PDF

Marketing strategies with superior information on consumer preferences

PDF

Essays on economics of education and private tutoring

PDF

Competitions in two-sided markets

PDF

Essays in health economics and provider behavior

PDF

Essays on consumer returns in online retail and sustainable operations

PDF

Two essays in economic policy: The influence of perceived comparative need on financial subsidy requests; &, Unintended consequences of the FOSTA-SESTA legislation

PDF

Essays on urban and real estate economics

Asset Metadata

Creator Cao, Jisu (author)

Core Title Essays in information economics and marketing

School College of Letters, Arts and Sciences

Degree Doctor of Philosophy

Degree Program Economics

Publication Date 06/30/2020

Defense Date 04/30/2020

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag bargaining,daily deal,OAI-PMH Harvest,online reviews,product returns,structural model

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Ridder, Geert (committee chair), Yang, Sha (committee chair), Moon, Roger (committee member)

Creator Email caojisu0319@gmail.com,jisucao@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c89-323503

Unique identifier UC11673136

Identifier etd-CaoJisu-8627.pdf (filename),usctheses-c89-323503 (legacy record id)

Legacy Identifier etd-CaoJisu-8627.pdf

Dmrecord 323503

Document Type Dissertation

Format application/pdf (imt)

Rights Cao, Jisu

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA