Content uploaded by Monic Sun
Author content
All content in this area was uploaded by Monic Sun on Nov 15, 2016
Content may be subject to copyright.
MANAGEMENT SCIENCE
Vol. 58, No. 4, April 2012, pp. 696–707
ISSN 0025-1909 (print) ISSN 1526-5501 (online) http://dx.doi.org/10.1287/mnsc.1110.1458
© 2012 INFORMS
How Does the Variance of Product Ratings Matter?
Monic Sun
Graduate School of Business, Stanford University, Stanford, California 94305; and
Marshall School of Business, University of Southern California, Los Angeles, California 90089, monic.sun@usc.edu
This paper examines the informational role of product ratings. We build a theoretical model in which ratings
can help consumers figure out how much they would enjoy the product. In our model, a high average rating
indicates a high product quality, whereas a high variance of ratings is associated with a niche product, one that
some consumers love and others hate. Based on its informational role, a higher variance would correspond to a
higher subsequent demand if and only if the average rating is low. We find empirical evidence that is consistent
with the theoretical predictions with book data from Amazon.com and BN.com. A higher standard deviation of
ratings on Amazon improves a book’s relative sales rank when the average rating is lower than 4.1 stars, which
is true for 35% of all the books in our sample.
Key words: information transmission; product ratings; social media; user-generated content
History : Received May 29, 2010; accepted August 23, 2011, by Pradeep Chintagunta, marketing. Published
online in Articles in Advance December 2, 2011.
1. Introduction
Consumers often seek others’ opinions about a prod-
uct before making a purchase decision. They read
magazines such as Consumer Reports, browse special-
ized review websites such as Yelp.com, check con-
sumer ratings posted by previous patrons, and ask
their family members and friends for recommenda-
tions. According to Kee (2008), 64% of the respon-
dents in Forrester Research’s online survey want to
see user ratings and reviews on the e-commerce web-
sites they visit, edging out those who want special
offers or coupons (61%), personalization (37%), games
or quizzes (29%), and videos (44%). According to the
same report, 68% of online shoppers check at least
four reviews before buying, and almost one quarter
of respondents check at least eight reviews.1
In response to consumers’ desire to read mul-
tiple reviews, many leading consumer-review and
e-commerce websites, including Yelp.com, Walmart
.com and Amazon.com, are making the distribution
of ratings salient to consumers by putting up bar
charts that demonstrate the percentage of reviews that
are associated with each level of ratings.2When a
bar chart is offered, it often appears in a prominent
location on the product’s introduction page, over and
1Branco et al. (2011) study the problem of when a consumer should
stop collecting more product information as he tries to make a deci-
sion on whether to purchase.
2See Sun (2011) for a discussion on how firms would directly dis-
close information on multiple product attributes. See also Sun and
Zhu (2011) for a discussion on how advertising-sponsored business
models can use a portion of the ad revenue to induce bloggers to
write on popular topics such as the stock market.
above any further breakdown of reviews. As evident
from the bar charts, reviewers often hold different
opinions toward the same product. For example, the
book Breaking Dawn (part of the Twilight series) had
received 5,132 Amazon reviews as of May 2, 2010,
with an average rating of 3.5 out of 5 stars. All of the
three reviews that were considered the most helpful
by Amazon visitors featured a one-star rating.3Com-
paring the bar charts for two popular books with sim-
ilar average ratings, Breaking Dawn and The Lord of
the Rings—The Fellowship of the Ring,4one can see that
readers clearly disagree about the first book, whereas
most readers give the second book high ratings.5
Although consumers consistently check out mul-
tiple online reviews and can easily observe the dif-
ferences in the distribution of ratings, prior research
(e.g., Chevalier and Mayzlin 2006) has focused on
establishing the causal impact of the average rating
on sales, and little is known about how the rating dis-
tributions would matter.
3
See http://www.amazon.com/Breaking-Dawn-Twilight-Saga-Book/
dp/031606792X (accessed May 2010).
4
See http://www.amazon.com/Lord-Rings-Fellowship-Platinum
-Extended/dp/B000067DNF (accessed May 2010).
5As of May 2, 2010, the rating distribution for Breaking Dawn was
2,460 five-star ratings, 673 four-star ratings, 441 three-star ratings,
499 two-star ratings, and 1,059 one-star ratings. The distribution
for Fellowship of the Ring was 3,292 five-star ratings, 359 four-star
ratings, 172 three-star ratings, 124 two-star ratings, and 216 one-star
ratings.
Sun et al. (2011) document how users on a large Chinese social
network site generally prefer to diverge from the most popular
choice even among their friends.
696
Sun: How Does the Variance of Product Ratings Matter?
Management Science 58(4), pp. 696–707, © 2012 INFORMS 697
To understand the role of the distribution of rat-
ings, a useful starting point would be to consider the
variance of ratings. As a statistical concept, variance
is a natural measure to capture the heterogeneity in
consumer opinions. From a managerial perspective,
variance is also an easy measure to obtain: Market
researchers can easily calculate the variance by sum-
marizing online ratings for their products. Given that
the variance of ratings is fast and almost costless to
obtain, it would be beneficial to the managers if they
could understand how the variance informs potential
customer’ purchase decisions and how to incorporate
the measure for better demand forecasts.
Although it would be interesting to study how
firms may potentially manipulate online word of
mouth (e.g., Mayzlin 2006), we assume truthful rat-
ings and focus our attention on how consumers inter-
pret and make use of the these ratings upon seeing
them. In particular, we seek to address three ques-
tions. First, what message does a higher variance of
product ratings communicate to the consumers? Sec-
ond, what is the impact of the level of variance on
a product’s subsequent price, demand, and profit?
Third, is there any interaction effect between the aver-
age rating and the variance of ratings?
When asked about the role of the variance of rat-
ings, some argue that because consumers are risk
averse, inconsistent opinions should have a negative
impact on demand; others think that the big differ-
ences in ratings might trigger curiosity, which could
lead to a higher demand. Although these are interest-
ing arguments, we abstract away from the psycholog-
ical underpinnings of how consumers would react to
risk, and focus on the informative role of the variance.
Our game-theoretical model features one seller and
consumers with differentiated tastes. To characterize
consumers’ taste space, we use a variation of the
linear-city model (Hotelling 1929). The product in our
model is characterized by two attributes: quality and
mismatch cost, which is similar to the “transportation
cost” in the Hotelling (1929) model. We use mismatch
cost to capture aspects of the product that would have
an influence on how much consumers would differ
in their enjoyment of the product. A low mismatch
cost, for example, suggests that it is easy for all con-
sumers to enjoy the product, regardless of how well
their tastes match with the product. In other words,
the product is mainstream—it is designed to cater to
a broad range of tastes. A high mismatch cost, on the
other hand, indicates that a consumer would enjoy
the product only if her taste matches well with the
product. The product in this case is a niche one: it is
designed to cater to only a small group of consumers.
As one can imagine, both product attributes, qual-
ity and mismatch cost, affect ratings. Consumers who
arrive late at the market can therefore infer the two
product attributes from earlier ratings. A high average
rating in our model communicates to the consumers
that the product has high quality, which increases
subsequent price, demand, and profit. The role of the
average rating in our paper is hence consistent with
prior literature (e.g., Chevalier and Mayzlin 2006).
A high variance, on the other hand, is a double-edged
sword: it communicates to consumers that the prod-
uct has both a high quality and a high mismatch cost.
Upon seeing a high variance, consumers infer that
the product is a niche one that some people love and
others hate.
Given the message contained in the variance of
ratings, how does a higher variance relate to equi-
librium price, demand, and profit? Interestingly, the
answer depends on the level of the average rating.
When the average rating is high, perceived quality is
already above a certain threshold. In this case, most
consumers are interested in the product. As a result,
the dominant effect of a higher variance is to drive
away marginal consumers. When the average rating
is low, on the other hand, few consumers are inter-
ested in the product. A higher variance then helps
the seller to secure demand from well-matched con-
sumers. Putting these two scenarios together, we find
that equilibrium demand will increase with the vari-
ance of product ratings if and only if the average rat-
ing is low, which is a key result of our model. In
an extension, we give early consumers the option to
defer their purchase decisions until after they have
read the product ratings, and find a similar role of
variance.
We provide empirical evidence in the context of
online book sales that is consistent with our theo-
retical predictions. For a random set of bestselling
books, we collect from Amazon.com and BN.com6
the consumer ratings, price, sales rank, and ship-
ping information for each book. We then employ a
difference-in-differences (DID) approach to identify
the causal effect of product ratings. Consistent with
previous empirical research on product ratings (e.g.,
Chevalier and Mayzlin 2006), we find that a higher
average rating on Amazon always increases book
sales. As an important new insight consistent with
our theoretical framework, we also find that a higher
standard deviation of ratings on Amazon increases a
book’s relative sales if and only if the average rating
is lower than approximately 4.1 stars, which is true
for 35% of all the books in our sample.
Our findings provide important managerial impli-
cations and suggest that managers should realize the
important role of the rating distribution. They should
keep in mind that a product with a low average rat-
ing may still turn out to be profitable if the variance
6The URL is the same as http://www.barnesandnoble.com.
Sun: How Does the Variance of Product Ratings Matter?
698 Management Science 58(4), pp. 696–707, © 2012 INFORMS
of ratings is sufficiently high. The seller of such a
niche product should therefore make sure that con-
sumers can easily observe the high variance of ratings
and provide detailed product information that further
facilitates the matching between consumers and the
product. A truthful high variance of ratings coupled
with detailed product information can help a seller to
skim the market by selling to the best matched con-
sumers at a premium price. A mainstream-product
seller, on the other hand, should make sure that con-
sumers can easily observe the low variance of ratings
and limit the disclosure of detailed product informa-
tion that may drive away marginal consumers. The
low variance would then lead to purchases from con-
sumers with a wide range of tastes, and the seller can
profit from selling across the board.
Our paper fits into the marketing literature of con-
sumer reviews and, more generally, user-generated
content. In the theoretical literature, Chen and Xie
(2005, 2008) study whether firms should allow con-
sumer reviews to be posted on their sites, and
how they should adjust their marketing strategies
accordingly. Mayzlin (2006) examines firms’ incen-
tives to post fake reviews and finds that even with
fake reviews, consumers will still be able to extract
some information on product quality in equilib-
rium. Bhardwaj et al. (2008) look at how the choice
between seller-initiated and consumer-initiated infor-
mation revelation affects the equilibrium level of
product quality. In a different context, Kuksov and Xie
(2010) study firms’ incentive to offer frills in a two-
period model where the average rating can help late
consumers infer early consumers’ utility.7Most theo-
retical research to date focuses on how firms react to
the possibility of showing consumer reviews as a new
information revelation mechanism.
In the empirical and experimental literature, the
most closely related paper is Chevalier and Mayzlin
(2006). They were the first to use a difference-in-
differences approach to identify the causal effect of
online consumer ratings on sales. The focus of their
paper is the average rating, and they also found that
the impact of a one-star rating is greater than that of a
five-star rating. To expand our understanding beyond
how the first moment of the rating distribution mat-
ters, we examine the average rating as well as the
variance of ratings, and particularly their interaction.
There is a small literature that directly studies the
distribution of ratings. Meyer (1981) shows that con-
sumers discount the average critic rating to adjust for
critic disagreement. Martin et al. (2007), in contrast,
survey individuals choosing between two movies
7The rating is either 0 or 1 in their model, so the average rating
captures the entire distribution of ratings, which is a key difference
from our model.
with pregiven ratings and find that consumers pre-
fer the high-variance movie. Along the same lines,
Clemons et al. (2006) find that beer brands with
higher variances of ratings grow fastest in terms
of sales. West and Broniarczyk (1998) also consider
how others’ opinions influence consumers’ evalua-
tions of product attributes. They examine consumer
attitudes toward critic consensus and find experi-
mental evidence that is consistent with the current
paper: a higher variance increases purchase likeli-
hood if and only if the average rating is below an
aspiration level. The major difference between the
paper by West and Broniarczyk (1998) and the current
paper is that they ground their study in the prospect
theory framework (Kahneman and Tversky 1979),
focusing on how consumers respond to uncertainty
(Jaccard and Wood 1988) and how their risk attitudes
are reference dependent, whereas we model risk-
neutral consumers making inferences on the under-
lying product characteristics through product ratings.
Finally, Zhang (2006) finds that the variance of movie
reviews does not play a significant role in determin-
ing box-office revenues altogether. (Interested read-
ers can see Table 1 for a summary of theoretical and
empirical studies of consumer reviews.)
Unlike previous studies, we allow the variance of
product ratings to capture the extent to which con-
sumers differ from each other in their enjoyment of
a particular product and attribute this difference to
the underlying product characteristics. By exploring
the interaction of the variance of ratings and the aver-
age rating, we are able to provide a theoretical model
that reconciles the mixed evidence on the role of the
variance. To our best knowledge, we also provide the
first empirical demonstration of how the interaction
of the variance of ratings and the average rating is a
significant determinant of product sales.
The rest of the paper is organized as follows. Sec-
tion 2 presents our theoretical framework in which
consumers learn about product characteristics from
earlier ratings. We also discuss in this section the pos-
sibility of consumers’ choosing to defer their purchase
until after seeing the ratings. Section 3 presents empir-
ical evidence on the role of the variance of ratings
using data from Amazon.com and BN.com. Section 4
concludes.
2. A Model of Product Ratings
The baseline model in this section features a monopoly
seller and risk-neutral consumers with heterogeneous
tastes. The seller’s product has two attributes: quality
and mismatch cost.
The higher the product’s quality, the more every
consumer enjoys the product. Examples of quality-
related product attributes abound. Book readers pre-
fer better prose. Movie goers enjoy a better story line.
Sun: How Does the Variance of Product Ratings Matter?
Management Science 58(4), pp. 696–707, © 2012 INFORMS 699
Table 1 Previous Research on Consumer Reviews
Theoretical studies
Chen and Xie (2008) Whether firms should publish consumer reviews
Chen and Xie (2005) How firms adjust marketing strategies given reviews
Mayzlin (2006) and Dellarocas (2006) Firms’ incentives to post fake consumer reviews
Awad and Etzion (2006) Firms’ incentives to filter consumer reviews
Jiang and Chen (2007) Firms’ incentives to manipulate early period reviews
Empirical studies
Godes and Mayzlin (2004) Dispersion of conversations across communities has explanatory
power in a model of TV ratings
Godes and Silva (2006) Product ratings tend to decrease over time
Zhu and Zhang (2010) Product ratings are more influential for nonsuperstars
Gao et al. (2006) Online consumer reviews exhibit remarkable community features
Chevalier and Mayzlin (2006) A higher average rating leads to higher book sales, and impact of
one-star reviews is bigger than impact of five-star reviews
Dellarocas et al. (2007) Total box-office revenue can be predicted from user reviews in the
first week of a movie’s release
Duan et al. (2008) Ratings of online user reviews have no significant impact on
movies’ box office revenues
Liu (2006) Word of mouth offers significant explanatory power for both
aggregate and weekly box office revenue, especially in early
weeks after a movie’s release
Chintagunta et al. (2010) It is the valence that seems to matter and not the volume
Digital camera buyers want higher resolution. Car
drivers like more safety features. A higher quality
simply increases every consumer’s willingness to pay
for the product.8
There are often other aspects of a product over
which consumers disagree (Lancaster 1966). For
example, consumers may want different colors when
it comes to purchasing a car, a piece of clothing, or
a digital camera. They may like different categories
of books and different genres of movies. We use the
second product attribute, mismatch cost, to capture
how niche the product is: mismatch cost is high when
the product is a niche one and caters to only a small
group of consumers.
Formally, think of consumers’ taste space as a
straight line of length 2 on which the product is
located at the midpoint. Consumers are uniformly
distributed on the line: A consumer’s location repre-
sents her ideal product in the taste space. If a con-
sumer with distance xfrom the product buys the
product at price P, her utility is
v−t·x−P1
where v > 0 is the product’s quality, and t > 0 is the
mismatch cost. A consumer buys at most one unit of
the product. If she decides not to buy the product, her
utility is zero.9
8See Desai et al. (2010) for an interesting model of digital rights
management, in which they show how consumers with different
quality sensitivity choose to steal, buy restricted copies, or buy
unrestricted copies of music.
9For ease in writing, we refer to a consumer as “she” and the seller
as “he.”
A high mismatch cost suggests that consumers with
different tastes derive very different utility from the
product; that is, whereas consumers located near the
product enjoy the product a lot, consumers further
away do not like it at all. Therefore, we call a high t
product a “niche product.” A low mismatch cost, in
contrast, suggests that all consumers derive more or
less the same utility from the product. In the extreme
case of t=0, all consumers derive the same level of
utility from the product. Therefore, we call a low t
product a “mainstream product.”
To understand the difference between the two prod-
uct attributes, one can think of books for example.
A book with high quality is generally well written
and is characterized by features that most readers
would enjoy, such as an interesting plot or the use
of exquisite language. A book with high mismatch
cost, on the other hand, have features that some peo-
ple love and others hate, such as violent or salacious
content.
Although the taste parameter xdiffers across con-
sumers, quality and mismatch cost are inherent to
the product. When facing a new product, a consumer
knows her own taste (i.e., her distance from the prod-
uct), but she may not know the product’s quality or
mismatch cost. For example, a book reader may have
some idea of how much she likes history books in
general (her distance), but she does not know exactly
how much she would enjoy a particular history book
without further information.
With the observation above, we assume that at the
beginning of the game, neither the seller nor con-
sumers know the realizations of vand t, whereas
the joint probability density distribution f 4v1 t 5 is
Sun: How Does the Variance of Product Ratings Matter?
700 Management Science 58(4), pp. 696–707, © 2012 INFORMS
common knowledge.10 We treat the levels of quality
and mismatch cost as exogenously given: The seller is
learning the perceptions of his product together with
early consumers. It is natural to assume that many
new product sellers do not quite understand their
consumers, as 95% of new products fail each year
(Burkitt and Bruno 2010).
To avoid discussing corner solutions, we assume
that the market is never fully covered11 and normal-
ize the total cost of production to zero. The extensive
form of the game is as follows:
Period 1. A unit mass of early consumers enter the
market. Their distance from the product xis uni-
formly distributed in 60117. The seller chooses price
4P15, and consumers decide whether to buy a unit of
the product. Every consumer who buys the product
consumes it and publishes a rating s4x5 =v−t·x.
Period 2. A unit mass of late consumers enter the
market. Their distance from the product is also uni-
formly distributed in 60117. Late consumers and the
seller observe first-period demand,12 the average rat-
ing, and the variance of ratings. The seller chooses
price 4P25, and each consumer decides whether to buy
the product.
Two features of the game are noteworthy. First,
we assume that a consumer’s rating equals her con-
sumption utility.13 In particular, when giving ratings,
an unsatisfied consumer does not take into account
whether it is a low quality or a high mismatch cost
that has led to her low utility. We observe many
reviews on Amazon that fit this assumption. For
example, the book Because She Can received a three-
star review saying, “if you have a mean female boss
then this is a good book to read.” The reviewer real-
izes that some well-matched readers would love the
book but nonetheless gives a low rating, suggesting
that his mismatch with the book has hindered his
enjoyment of the book.14
Second, because uniformly distributed consumer
tastes would lead to uniformly distributed ratings,
the average rating and the variance of ratings in our
model perfectly describe the entire rating distribution.
Admittedly, ratings are often not uniform, and some
capable consumers might be able to process abundant
10 Note that we do not require vand tto be independent.
11 A sufficient condition for incomplete coverage of the market is
v∈6v1 ¯
v7,t∈6t1 ¯
t7, and ¯
v < t.
12 For example, consumers can learn prior demand by looking at
the product’s sales rank on a retailer’s website.
13 Incorporating price into ratings would not affect our analysis as
long as all the consumers understand how price enters the rating
formula.
14 Our results would hold, however, even if the reviewers are more
lenient toward high mismatch than low quality, that is, changing
the rating formula to s=v−·tx, where 0 < < 1, would not
substantially change the results as long as is common knowledge.
information from a bar chart (e.g., skewness). Because
our main goal is to expand our understanding of rat-
ings to include the variance, we leave considerations
of higher moments of the rating distribution for future
research.
We solve for the subgame perfect equilibrium of
the game. In the first period, there is no information
on quality or mismatch cost. Consumers make pur-
chase decisions based on their expectations of vand t,
denoted by E4v5 and E4t5, respectively. Because the
joint distribution f 4v1 t 5 is common knowledge, E4v5
and E4t5 are also common knowledge. When the seller
chooses price P1, the indifferent consumer with dis-
tance D1from the product is given by
E4v5 −E4t5 ·D1−P1=00(1)
Consumers located with distance x∈601 D17derive
higher utility than the indifferent consumer. They
would purchase the product, and the first-period
demand is D1.
Now consider the distribution of ratings.15 The
early consumer with x=0 has a perfect match with
the product and gives the highest rating, v. Similarly,
the marginal consumer with distance D1gives the
lowest rating, v−t·D1. Although the marginal con-
sumer is indifferent when purchasing the product, her
rating can be either positive or negative depending
on whether the product exceeds or falls short of her
prior expectation.
Ratings are uniformly distributed in 6v −t·D11 v7,
which suggests that both product attributes play crit-
ical roles in the distribution. When product quality v
increases, all ratings become higher. When the mis-
match cost tincreases, two effects occur. First, all rat-
ings become lower as consumers experience a larger
utility reduction from their taste mismatch. Second,
the difference across ratings becomes larger, which
captures the idea that the difference in consumers’
utility is larger for niche products.
The average rating and the variance of ratings can
be computed, respectively, as
M=v−1
2t·D1and V=1
124t ·D1520(2)
These two equations suggest that the average rating
and the variance of ratings are not independent. In
particular, upon seeing the highest possible average
rating, given any D1, one can infer that the variance of
rating is zero. Similarly, upon seeing the lowest pos-
sible average rating, one can infer that the mismatch
cost, and hence the variance, is the highest possible.
Nevertheless, for all other levels of the average rating,
the variance is uncertain.
15 All analysis remains qualitatively unchanged if the rating score
also reflects price: s4x5 =v−t·x−P.
Sun: How Does the Variance of Product Ratings Matter?
Management Science 58(4), pp. 696–707, © 2012 INFORMS 701
When consumers observe a relatively low average
rating, for example, they make the following infer-
ence: The low average may result from either low
quality or high mismatch cost. A high variance in this
case helps them to figure out that the product has
both high quality and a high mismatch cost, and well-
matched consumers would actually love it.
Mathematically, the late consumers infer the real-
izations of vand tby solving (2):
v=M+√3Vand t=2√3V
D1
0(3)
In sum, a high average rating indicates a high level
of quality and suggests that all early consumers enjoy
the product to a reasonable degree. A high variance,
on the other hand, indicates that early consumers
either love or hate the product depending on how
well their tastes match with it.
Given Equations (3), there is no uncertainty left
regarding the two product attributes in the sec-
ond period. Late consumers can perfectly infer the
product’s quality and mismatch cost of the prod-
uct. In other words, there is complete information
in the second period. The game is therefore equiv-
alent to one in which all second-period consumers
observe every single rating. Given the uniform distri-
bution of ratings, our game is also equivalent to one
in which (1) second-period consumers read only the
highest and the lowest ratings, or (2) each second-
period consumer reads only one rating, the one from
the first-period consumer who has the same taste x.
Scenario (1) may be descriptive of consumers who
visit review websites with a wide score distribution
(e.g., Yahoo! Movies, Metacritics) and read only the
most drastic reviews, whereas scenario (2) would hold
for consumers that are loyal to certain critics.
Given complete information, the indifferent con-
sumer in the second period is given by D2=4v −P25/t,
and hence the seller solves
max
P2
P2
v−P2
t0
Equilibrium levels of second-period price, demand,
and profit can be derived as
P∗
2=v
21 D∗
2=v
2t1and ç∗
2=v2
4t0
Based on the correspondence between ratings and the
product attributes given in Equations (3), the second-
period equilibrium outcomes can be rewritten as
P∗
2=M+√3V
21 D∗
2=D1
4M
√3V+11and
ç∗
2=D1
8M2
√3V+√3V+2M0
(4)
Figure 1 Equilibrium Price, Demand, and Profit in Period 2
D*
2
P*
2
Var
Π*
2
M2
3
Notes. The average rating is positive in this figure. The horizontal axis is
the variance of product ratings in the first period. As the variance of rat-
ings increases, late consumers infer both higher quality and higher mismatch
cost. The top, middle, and bottom curves are, respectively, equilibrium levels
of demand, price, and profit in the second period.
As suggested by (4), the average rating and the vari-
ance of ratings each plays an important informational
role in determining the second-period market out-
comes, as summarized in the following propositions.
Proposition 1. Second-period price, demand, and
profit all increase with the average rating.
Existing literature has provided empirical evidence
that is consistent with this prediction. Cao and Gruca
(2004) and Bruce et al. (2004), for example, find that
a seller raises the price when past consumers have
high satisfaction. Elberse and Eliashberg (2003) and
Chevalier and Mayzlin (2006), on the other hand, find
that a higher average rating leads to more sales.
Now we turn to the central proposition of the
paper, which examines the impact of the variance of
ratings on the equilibrium levels of subsequent mar-
ket outcomes.
Proposition 2. In the second period, price increases
with the variance of ratings, demand increases with the
variance if and only if M≤0, and profit increases with the
variance if and only if M≤√3V.16
Figure 1 provides an illustration of Proposition 2.
The proposition suggests that the seller should charge
a high price when the mismatch cost tis high, which
is reflected through a high variance of ratings. A high
price in this case helps the seller fully exploit well-
matched consumers’ high willingness to pay. In con-
trast, equilibrium demand and profit increase with the
variance only when the average rating is relatively
16 Although we find zero to be a threshold in the proposition, one
can easily do a linear transformation of ratings to make the thresh-
old positive. Therefore, zero as a threshold should not be taken
quantitatively.
Sun: How Does the Variance of Product Ratings Matter?
702 Management Science 58(4), pp. 696–707, © 2012 INFORMS
low.17 The intuition is as follows: When the aver-
age rating is low, a high variance crucially improves
consumers’ perception of the product’s quality, and
hence increases demand and profit. When the aver-
age rating is high, consumers are already confident of
the product’s quality. The dominant effect of a high
variance is to signal a high mismatch cost. It therefore
drives away marginal consumers and hurts demand
and profit.
Given the price and demand patterns, equilibrium
profit in the second period turns out to be U-shaped
in the variance (see Figure 1), suggesting that profit is
the highest when the variance is either extremely low
or extremely high. It is noteworthy that Johnson and
Myatt (2006) examine demand transformation that
results from changes in consumer taste dispersion.
In a static model, they also find that firms have pref-
erences for extreme dispersions of consumer tastes,
where high dispersion is complemented by niche
production, and low dispersion is complemented by
mass-market supply. Although the intuition behind
our arguments are quite similar, we formulate the
problem in the easily testable context of product rat-
ings with the additional focus on how the disper-
sion of consumer utility, as measured by the variance
of ratings, should interact with the average rating in
affecting subsequent demand.
To complete the characterization of the equilibrium,
consider the seller’s strategy in the first period. He
maximizes the expected total profit:
max
P1
P1
E4v5 −P1
E4t5 +Ev2
4t0
The second term above is determined by the joint dis-
tribution f 4v1 t 5 and hence independent of P1. There-
fore, equilibrium price, demand, and profit in the first
period are
P∗
1=E4v5
21 D∗
1=E4v5
2E4t51and ç∗
1=E4v52
4E4t 5 1
respectively; that is, the first-period equilibrium out-
comes depend on the prior expectations of quality
and mismatch cost. When E4v5/E4t5 is higher, con-
sumers have a more favorable expectation of the
product, and the first-period demand is higher.
An interesting observation one can make is that the
equilibrium price is higher in the second period than
in the first period if and only if
v=M+√3V > E4v53
17 In an empirical setting, one can often observe and control for the
price of a product. The impact of the rating statistics on subsequent
demand in this case can be derived by looking at the indifferent
consumer: D∗
2=4v −P∗
25/t =64M −P∗
25/√3V+174D1/250 Therefore,
the impact of the average rating and the variance of ratings in a
context with observable prices remains similar to that in Proposi-
tions 1 and 2, except for a different threshold.
that is, the seller should raise the product’s price over
time if it receives a favorable average rating and a
high variance of ratings. In a different context, Berge-
mann and Välimäki (2006) examine dynamic price
patterns of new experience goods and find a similar
result: equilibrium price increases over time for niche
products and decreases over time for mass-market
products.
2.1. Deferring Purchase Decisions
An important dimension in consumers’ use of prod-
uct ratings is that they can be strategic in the timing of
the purchase; that is, whether a consumer purchases
the product early or late can be an endogenous deci-
sion (e.g., Guo and Villas-Boas 2007). Because a big
fraction of online shoppers seek out product ratings
and reviews, it is quite conceivable that some con-
sumers choose to defer purchase decisions until after
they see the product ratings. Waiting behavior turns
out to have interesting consequences. In particular,
the impact of the variance of ratings on the second-
period price is more negative, as we show below.
Consider two changes to the baseline model. First,
consumers in the first period can choose whether or
not to defer their purchase decision with a discount
factor ∈40115. If they choose to wait to see the rat-
ings, they are allowed to purchase the product in the
second period. Second, late consumers of mass n > 0,
rather than 1, enter the market in the second period.
The indifferent consumer D1in the first period is now
given by
E4v5 −E4t5 ·D1−P1=·E 4max8v −t·D1−P210950
Similar to the baseline model, early consumers with
x∈601 D17choose to buy the product, and others
choose to wait; that is, consumers who are almost per-
fectly matched would purchase right away, whereas
consumers located further away in the taste space pre-
fer to wait to see the ratings. Intuitively, consumers
located further away are more likely to decide not
to buy the product when given full information, and
hence they are more motivated to wait until the sec-
ond period.18
In the modified game, the formulas of the average
rating and the variance of ratings remain the same as
in (2) in the baseline model. As before, late consumers
observe the average and the variance of ratings and
18 Theoretically, the seller can charge an extremely high first-period
price so that all early consumers choose to wait. If he does this,
however, there would be no ratings from early consumers, and
hence no way for late consumers to learn about how great the
product is. We assume that if the first-period price is so high that
there is no purchase in the first period, all consumers in the second
period believe that the product has the lowest possible quality and
highest possible mismatch cost.
Sun: How Does the Variance of Product Ratings Matter?
Management Science 58(4), pp. 696–707, © 2012 INFORMS 703
have complete information on the product attributes.
The second-period demand D2is now given by
D2=n·v−P2
t+max01v−P2
t−D11
where the first term on the right-hand side is demand
from late consumers, and the second term is demand
from early consumers who choose to wait. If the
second term is zero, second-period equilibrium out-
comes are proportional to those in the baseline model,
and the impact of Mand Vare simply multiplied
by n.
The seller in the second period chooses P2to
maximize P2·D2. To solve this maximization problem,
one simply needs to compare the highest possible
profit when the seller serves only the late consumers
with that when the seller serves both late consumers
and early consumers. Based on this comparison, one
of the following two equilibrium prices will emerge
in the second period. First, P∗
2=v/2. In this case, no
early consumer purchases in the second period, and
the equilibrium outcomes are proportional to those
in the baseline model. Second, P∗
2=1
24v −4t ·D15/
4n +155 =1
24M +44n −15/4n +155√3V 5. In this case,
some early consumers purchase the product in the
second period. The equilibrium demand becomes
D∗
2=1
26v4n +15/t −D17=64n +154M /4√3V+1
45
−1
27D1,19 and the equilibrium profit in the second
period is ç∗
2=44n +15/4t54v −4t ·D15/4n +1552=
44n +15D15/8√3V 4M +44n −15/4n +155√3V 52.
Based on these equilibrium outcomes, we can make
the following observations. First, as in the baseline-
model, equilibrium demand in the second period
always increases with the average rating. Second,
equilibrium demand increases with the variance of
ratings if and only if M≤0. Therefore, the interaction
effect of the variance of ratings and the average rat-
ing continue to hold when early consumers are given
the option to defer their purchase decisions. Third, a
higher variance increases second-period price if and
only if n > 1. The intuition for this result is as fol-
lows. When n > 1, many consumers come to the mar-
ket late and the second-period demand comes mostly
from late consumers. Variance of product ratings in
this case has a similar impact to that in the baseline
model. When n < 1, early consumers who choose to
wait form a big portion of the second-period demand.
These consumers are located further away from the
product and react more negatively to the variance
than the typical late consumer. As a result, the equilib-
rium price in the second period decreases, rather than
19 If price is controlled for, demand can be written as D∗
2=
4n +1544v −P∗
25/t5 −D1=64n +15/244M −P∗
25/√3V+15−17·D1,
which always increases with the average rating M, but increases
with the variance of ratings Vif and only if M≤P∗
2.
increases, with the variance of ratings. When n=1,
the two pricing incentives associated with a high
variance, exploiting well-matched late consumers and
trying to keep mismatched early consumers, balance
each other out. The variance hence does not affect the
second-period price. Finally, second-period profit still
always increases with the average rating, but the con-
dition under which the profit increases with the vari-
ance is quite different from before. In general, a high
variance is more likely to be profitable for the seller
when most consumers can observe some product rat-
ings when they first come across the product. When
n≤1, most second-period demand comes from the
early consumers that are sensitive to mismatch, and
hence the profit decreases with the variance. When
n > 1, profit increases with the variance if and only if
√3V > 44n +15/4n −155M . In this case, most second-
period patrons are late consumers, and the impact
of ratings on profit is similar to that in the baseline
model.
3. Evidence from Online
Book Retailers
In this section, we provide empirical evidence that is
consistent with our theoretical predictions. In an ideal
empirical setting, we would have data on sales and
prices over two periods, product ratings from only
the initial period, and control variables on product
characteristics that the consumers could observe with-
out reading any product ratings. Such variables could
include, for example, reputation of the brand, adver-
tising, and promotions.
Our actual data were obtained from two leading
booksellers on the Internet: Amazon and Barnes &
Noble. There are three reasons for employing the
data. First, book ratings are commonly sought after
by potential book buyers. Second, the fact that there
are two retailers means that we can use a differencing
approach to control for unobserved book characteris-
tics that may influence both sales and ratings. Third,
the possibility of tracking the two websites over time
provides an opportunity to perform a DID analysis to
eliminate any potential book-website effects, as dis-
cussed by Chevalier and Mayzlin (2006).
To collect data, we first created a list of 3,828
random ISBNs from the bestseller section of Global
Books in Print.20 All of the books in our list were
released in 2002–2006. For each book, we recorded the
number of reviews, numerical values of its ratings,
price, sales rank, availability, and shipping informa-
tion from Amazon.com and BN.com (henceforth, BN).
There were 892 books with complete data in Jan-
uary 2009. To use the DID approach, in May 2009
20 See http://www.GlobalBooksinPrint.com.
Sun: How Does the Variance of Product Ratings Matter?
704 Management Science 58(4), pp. 696–707, © 2012 INFORMS
Table 2 Summary Statistics of Books
January 2009 May 2009
Variable Website Mean Std. dev. Mean Std. dev.
Sales rank Amazon 201,696 307,726 267,085 350,943
BN 113,952 146,965 130,309 163,633
Price Amazon 1203 607 1206 608
BN 1502 901 1408 808
Number of reviews Amazon 11001 13405 10102 14005
BN 2902 5301 3005 5500
Average rating Amazon 306 006 402 005
BN 404 006 404 006
Std. dev. of ratings Amazon 104 003 100 004
BN 007 006 007 005
Note. Observations: 667.
we collected a second round of data including the
same information for every book.21 As a result of the
changes in Amazon’s and BN’s selections of book
offerings, we have a total of 667 books available for
the DID analysis, with summary statistics presented
in Table 2.
One can see from the table that a book on BN
typically has a higher price, fewer ratings, a higher
average rating, and a lower standard deviation of
ratings than it does on Amazon. The books in our
sample generally become less popular during our
data collection period (January–May 2009), as the
average sales rank increases on both Amazon and
BN. Although both websites are known to constantly
adjusted the prices of their books, the average prices
remain largely the same during the five months, with
a slight decrease on Amazon and a slight increase
on BN.
Table 2 also shows another interesting trend: While
a book on average earns 1.3 more reviews on BN dur-
ing the five months, the average number of reviews
on Amazon goes down, possibly because of prun-
ing.22 Moreover, whereas the average rating on BN
remains largely the same, the average rating on Ama-
zon increases significantly, suggesting that the prun-
ing tends to concentrate on reviews with low ratings.
Because previous research repeatedly demonstrates
the positive effect of a higher average rating, we focus
on examining the role of the standard deviation of
ratings. Based on our theoretical framework, we for-
mulate the following hypothesis.
Hypothesis 1. A higher standard deviation of ratings
for a book leads to higher sales if and only if the average
rating is low.
21 BN, in general, ships faster than Amazon.
22 Chevalier and Mayzlin (2006) find the same pattern.
We adopt a DID estimation approach from
Chevalier and Mayzlin (2006). Denote Amazon vari-
ables by superscript Aand BN variables by super-
script B, and use subscript ias an index for books. The
underlying data generating process is assumed to be
A02log4RankJ
it 5
=J
i+vi+X×âJ+ìJ
i×éJ+J
it 1 J ∈8A1 B90
Consistent with previous studies (e.g., Chevalier and
Mayzlin 2006, Brynjolfsson et al. 2003), we use log
sales rank on the left-hand side as a linear proxy
for log sales. If log sales quantity were used directly
as the dependent variable, the estimated coefficients
would be scaled by a constant. Brynjolfsson et al.
(2003), for example, find a scaling coefficient of −00871
using sales quantity and rank data from Amazon.23
On the right-hand side, J
iis an unobservable
book-site effect that captures any possible interaction
between characteristics of book iand preferences of
consumers on site J. For example, compared with BN
users, Amazon users may like to buy computer books
more and also give them higher ratings. If this is the
case, A
iwould be positive for the computer books.
The second term, vi, is a book-level fixed effect that
captures certain aspects of book ithat can directly
influence sales, such as the reputation of the author
and the publishing company, newspaper reviews of
the book, author events, and other forms of adver-
tising. Vector Xcontains rating variables from both
Amazon and BN. In our specification, a book’s ratings
on Amazon can affect its sales ranks on both Amazon
and BN. Similarly, its ratings on BN can also affect its
sales ranks on both sites. Control variables in ìare
price, log number of reviews, and shipping dummies.
Finally, J
it is a normally distributed random error.
To control for book-level fixed effects and book-site
effects, we take the difference of A0 across the two
sites and across time:
ã6log4RankA
i5−log4RankB
i57
=A
1·ãMA
i−B
1·ãMB
i+A
2·ãSDA
i−B
2·ãSDB
i
+A
3·ã4MA
i·SDA
i5−B
3·ã4MB
i·SDB
i5
+ãìA
i×éA−ãìB
i×éB+i1(5)
where MJ
iand SDJ
idenote, respectively, the average
and standard deviation of ratings of book ion site
J,J∈8A1 B9. Given the fact that sales rank is nega-
tively correlated with sales, Hypothesis 1 would be
confirmed if J
2<0 and J
3>0. As proposed by Zhu
and Zhang (2010), the sales of a popular product may
23 The specification they use is log4quantity5=1+2·log4r ank5 +,
and the intercept 1is estimated to be 10.526.
Sun: How Does the Variance of Product Ratings Matter?
Management Science 58(4), pp. 696–707, © 2012 INFORMS 705
Table 3 The Effect of Five-Month Changes in Ratings on Changes
in Sales
(1) (2) (3)
Amazon ãln(price) 10521∗∗∗ 10507∗∗∗ 10552∗∗∗
4002015 4002865 4002005
BN ãln(price)−10917∗∗∗ −20050∗∗∗ −10876∗∗∗
4005845 4007585 4005835
Amazon ãln(no. of reviews)−00835∗∗∗ −00858∗∗∗ −00847∗∗∗
4000925 4001055 4000945
BN ãln(no. of reviews) 00522∗∗∗ 00465∗∗ 00560∗∗∗
4001945 4002245 4002145
Amazon ãaverage rating −00196∗∗∗ 00103 −10014∗∗∗
4000675 4002545 4003665
BN ãaverage rating −00415 00240 00633
4003975 4008725 4009175
Amazon ãstd. dev. of ratings 00485 −20562∗∗∗
4003955 4008735
BN ãstd. dev. of ratings 00540 30786
4005685 4204265
Amazon ã(average rating ×std. dev.) 00627∗∗∗
4001675
BN ã(average rating ×std. dev.)−00753
4005215
Shipping dummies Yes Yes Yes
Observations 667 667 667
Adjusted R20.174 0.175 0.192
Notes. For a variable x,ãx =xMay 2009 −xJan 2009 . The dependent variable in
all three specifications is ã{ln(Amazon sales rank5−ln(BN sales rank)}.
∗∗p < 0005; ∗∗∗ p < 0001.
react less to online consumer ratings as other sources
of information become abundant. If this is true in
our context, we would expect the error term in (5)
to be heteroscedastic, because its variance increases
with a book’s popularity. A White test confirms het-
eroscedasticity (p < 0001), and we use a two-step
feasible weighted least squares estimation approach
(Greene 1999), with each book’s number of reviews
on Amazon and BN in January and May 2009 as pre-
dictors for the variance of errors in the first stage.24
Table 3 presents our estimation results. To high-
light the importance of the standard deviation of
ratings and how it interacts with the average rat-
ing in affecting demand, we compare three specifi-
cations. In the first specification, the rating variables
include only the average rating. In the second one,
we also include the standard deviation of ratings.
In the third specification, we further add the interac-
tion of the average rating and the standard deviation
of ratings. In all the three specifications, a lower price
24 To be more exact, our weights are obtained by regressing the
squared error terms on the numbers of book ratings in January and
May 2009 on Amazon and BN. Our results are robust to alterna-
tive choices of the weight. In particular, a prespecified weight that
equals the inverse of the total number of ratings across the two
websites yields similar results.
and a higher number of reviews lead to higher sales.
Regarding the role of consumer ratings, we make the
following observations. First, a higher average rat-
ing on Amazon increases the book’s relative sales on
Amazon in columns (1) and (3), which is consistent
with our baseline model and the previous literature
(Chevalier and Mayzlin 2006). BN average rating is
not significant, which might be due to the fact that
BN ratings remain largely unchanged during our data
collection period (see Table 2).
Second, column (3) suggests that a higher standard
deviation of Amazon ratings leads to relative higher
sales if and only if the average rating on Amazon is
low, confirming Hypothesis 1. The three coefficients
on Amazon’s rating variables are all significant with
p < 0001. Based on these estimates, a higher standard
deviation of Amazon ratings increases the book’s rel-
ative sales when the average Amazon rating is lower
than 4.1 stars, which is true for 35% of all the books
in our sample. Meanwhile, a higher average rating on
Amazon increases the book’s relative sales when the
standard deviation is lower than 1.6 stars, which is
true for almost all (96%) of the books.
Notably, if a researcher ignores the interaction and
uses the specification in column (2), he may reach
an imprecise conclusion that neither the average rat-
ing nor the standard deviation of ratings affect a
book’s relative sales rank. Comparing across the three
columns, one can see that incorporating the inter-
action term significantly changes the coefficients of
the rating variables, while increasing the adjusted R2
value. At the same time, incorporating the interaction
term does not have a big impact on the estimated
effect of nonrating variables such as price and the
number of reviews. This suggests that the explanatory
power of the interaction term comes within the rat-
ings, as suggested by our theory, rather than potential
correlation between the interaction term and nonrat-
ing variables.
Overall, the empirical evidence we find is con-
sistent with the hypothesis that a higher standard
deviation of ratings on Amazon improves the book’s
relative sales on Amazon when the average rating is
low, and hurts its relative sales when the average rat-
ing is high.
4. Concluding Remarks
In this paper, we examine the informational role of
the distribution of product ratings by focusing on the
variance of ratings. We find that the interaction of
the average rating and the standard deviation of rat-
ings plays a significant role on subsequent market
outcomes. For a product with a low average rating,
a higher variance of ratings communicates to poten-
tial buyers that well-matched consumers would love
Sun: How Does the Variance of Product Ratings Matter?
706 Management Science 58(4), pp. 696–707, © 2012 INFORMS
the product, which in turn increases demand. For a
product with a high average rating, a higher vari-
ance of ratings drives away marginal consumers and
reduces demand. We provide empirical evidence with
data from two leading Internet book retailers, Ama-
zon and Barnes & Nobel, that is consistent with our
theoretical predictions.
Two directions of future research are promising.
First, this paper opens up the possibility for managers
to use the variance of ratings as an additional mea-
sure in demand forecasts. The measure can be partic-
ularly useful in a competitive environment. Consider
a motion-picture studio trying to determine a movie’s
release date. The studio may try to forecast the open-
ing box-office revenue by mapping the locations of
its own movie and the competing movies into con-
sumers’ taste space by solving for quality and mis-
match cost from the distribution of prerelease critic
ratings.
Second, there are many other aspects of consumer
reviews that are worth exploring. For example, some
websites offer multidimensional ratings. It is remark-
able, for example, that Best Buy (bestbuy.com) not
only asks consumers to rate its products along multi-
ple dimensions, but also caters the set of dimensions
to the product category. For a global positioning sys-
tem, consumers need to rate the product along four
dimensions: value for price, durability, ease of use,
and features. For a TV, the dimensions become picture
quality, sound quality, and features. It would be inter-
esting to study the matching between the dimensions
and the product category, as well as to measure the
different weights that consumers put on these dimen-
sions. Specialized consumer-review websites such as
Yelp.com also publish the time trend of consumer rat-
ings. It would also be interesting to examine how such
trends influence consumers’ decisions to purchase a
particular service.25
Acknowledgments
The author thanks department editor Pradeep Chintagunta,
the associate editor, and three anonymous referees at Man-
agement Science for their thoughtful suggestions. She is
deeply indebted to Albert Ma, Jacob Glazer, Marc Rysman,
and Juanjuan Zhang for their advice on this paper. For help-
ful conversations, the author is also grateful to J. Miguel
Villas-Boas, David Godes, Wesley Hartmann, Jim Lattin,
Tilman Börgers, Philip Choné, Iván Fernández-Val, Chunyu
Ho, Panle Jia Barwick, Ginger Jin, Xiaofeng Li, Barton
Lipman, Michael Manove, Preston McAfee, Dilip Mookher-
jee, In-Uck Park, Larry Samuelson, Jean Tirole, Ram Rao,
Al Roth, Michael Zhang, Dazhuang Zhu, Xiaomei Zhu,
25 See Yoganarasimhan (2010) for an interesting discussion on how
the network structure among YouTube users affects the diffusion of
videos. See also Zhang and Zhu (2011) for a study on how audience
size would affect the incentives for users to generate content.
Feng Zhu, and seminar participants at Boston University
economics workshops, Summer Institute of Competitive
Strategy, Stanford Graduate School of Business, Harvard
Business School, UT Dallas, UC Davis, HEC Paris, and Sin-
gapore Management University.
References
Awad, N., H. Etzion. 2006. Stay out of my forum! Evaluating
firm involvement in online ratings communities. Presentation,
Workshop on Information Systems and Economics (WISC),
December 9–10, Northwestern University, Evanston, IL.
Bergemann, D., J. Välimäki. 2006. Dynamic pricing of new experi-
ence goods. J. Political Econom. 114(4) 713–743.
Bhardwaj, P., Y. Chen, D. Godes. 2008. Buyer-initiated versus seller-
initiated information revelation. Marketing Sci. 54(6) 1104–1114.
Branco, F., M. Sun, J. M. Villas-Boas. 2011. Optimal search for prod-
uct information. Working paper, Stanford University, Stanford,
CA. http://ssrn.com/abstract=1720122.
Bruce, N., E. Haruvy, R. Rao. 2004. Seller rating, price, and default
in online auctions. J. Interactive Marketing 18(4) 37–50.
Brynjolfsson, E., Y. Hu, M. D. Smith. 2003. Consumer surplus in
the digital economy: Estimating the value of increased product
variety. Management Sci. 49(11) 1580–1596.
Burkitt, L., K. Bruno. 2010. New, improved000and failed.
Forbes.com (March 24), http://www.msnbc.msn.com/id/
36005036.
Cao, Y., T. S. Gruca. 2004. The influence of pre- and post-purchase
service on prices in the online book market. J. Interactive
Marketing 18(4) 51–62.
Chen, Y., J. Xie. 2005. Third-party product review and firm market-
ing strategy. Marketing Sci. 24(2) 218–240.
Chen, Y., J. Xie. 2008. Online consumer review: A new ele-
ment of marketing communications mix. Management Sci. 54(3)
477–491.
Chevalier, J., D. Mayzlin. 2006. The effect of word of mouth online:
Online book reviews. J. Marketing Res. 43(3) 345–354.
Chintagunta, P. K., S. Gopinath, S. Venkataraman. 2010. The effects
of online user reviews on movie box office performance:
Accounting for sequential rollout and aggregation across local
markets. Marketing Sci. 29(5) 944–957.
Clemons, E. K., G. G. Gao, L. M. Hitt. 2006. When online reviews
meet hyperdifferentiation: A study of the craft beer industry.
J. Management Inform. Systems 23(2) 149–171.
Dellarocas, C. 2006. Strategic manipulation of internet opinion
forums: Implications for consumers and firms. Management Sci.
52(10) 1577–1593.
Dellarocas, C., X. M. Zhang, N. Awad. 2007. Exploring the value of
online product reviews in forecasting sales: The case of motion
pictures. J. Interactive Marketing 21(4) 23–45.
Desai, P., D. Purohit, D. Vernik. 2010. Digital music set free: The
flip side of DRM protection. Working paper, The Fuqua School
of Business, Duke University, Durham, NC.
Duan, W., B. Gu, A. B. Whinston. 2008. Do online reviews matter?—
An empirical investigation of panel data. Decision Support
Systems 45(4) 1007–1016.
Elberse, A., J. Eliashberg. 2003. Demand and supply dynamics for
sequentially released products in international markets: The
case of motion pictures. Marketing Sci. 22(3) 329–354.
Gao, G. G., B. Gu, M. Lin. 2006. The dynamics of online consumer
reviews. Presentation, Workshop on Information Systems and
Economics (WISE), December 9–10, Northwestern University,
Evanston, IL.
Godes, D., D. Mayzlin. 2004. Using online conversations to study
word-of-mouth communication. Marketing Sci. 23(4) 545–560.
Godes, D., J. Silva. 2006. The dynamics of online opinion. Working
paper, University of Maryland, College Park.
Greene, W. H. 1999. Econometric Analysis, 4th ed. Prentice Hall,
Upper Saddle River, NJ.
Sun: How Does the Variance of Product Ratings Matter?
Management Science 58(4), pp. 696–707, © 2012 INFORMS 707
Guo, L., J. M. Villas-Boas. 2007. Consumers stockpiling and price
competition in differentiated markets. J. Econom. Management
Strategy 16(4) 827–858.
Hotelling, H. 1929. Stability in competition. Econom. J. 39(153)
41–57.
Jaccard, J., G. Wood. 1988. The effects of incomplete information
on the formation of attitudes toward behavioral alternatives.
Econometrica 54(4) 580–591.
Jiang, B.-j., P.-y. Chen. 2007. An economic analysis of online prod-
uct reviews and ratings. Working paper, Tepper School of
Business, Carnegie Mellon University, Pittsburgh.
Johnson, J. P., D. P. Myatt. 2006. On the simple economics of adver-
tising, marketing, and product design. Amer. Econom. Rev. 96(3)
756–784.
Kahneman, D., A. Tversky. 1979. Prospect theory: An analysis of
decision under risk. Econometrica 47(2) 263–292.
Kee, T. 2008. Majority of online shoppers check at least four reviews
before buying. Online Media Daily (February 19), http://www
.mediapost.com/publications/?fa=Articles.showArticle&art
_aid=76727.
Kuksov, D., Y. Xie. 2010. Pricing, frills, and customer ratings. Mar-
keting Sci. 29(5) 925–943.
Lancaster, K. J. 1966. A new approach to consumer theory. J Political
Econom. 74(2) 132–157.
Liu, Y. 2006. Word of mouth for movies: Its dynamics and impact
on box office revenue. J. Marketing 70(3) 74–89.
Martin, J., G. Barron, M. I. Norton. 2007. Choosing to be uncer-
tain: Preferences for high variance experiences. Working paper,
Harvard Business School, Boston.
Mayzlin, D. 2006. Promotional chat on the internet. Marketing Sci.
25(2) 155–163.
Meyer, R. J. 1981. A model of multiattribute judgments under
attribute uncertainty and informational constraint. J. Marketing
Res. 18(4) 428–441.
Sun, M. 2011. Disclosing multiple product attributes. J. Econom.
Management Strategy 20(1) 195–224.
Sun, M., F. Zhu. 2011. Ad revenue and content commercializa-
tion: Evidence from blogs. Working paper, Stanford University,
Stanford, CA.
Sun, M., X. (M.) Zhang, F. Zhu. 2011. To belong or to be different?
A large-scale experiment in a chinese social network. Working
paper, Stanford University, Stanford, CA.
West, P. M., S. M. Broniarczyk. 1998. Integrating multiple opinions:
The role of aspiration level on consumer response to critic con-
sensus. J. Consumer Res. 25(1) 38–51.
Yoganarasimhan, H. 2010. Impact of social network structure on
content propagation—A study using youtube data. Working
paper, Graduate School of Management, University of Califor-
nia, Davis, Davis.
Zhang, X. (M.) 2006. Tapping into the pulse of the market:
Essays on marketing implications of information flows. Ph.D.
dissertation, MIT, Cambridge, MA.
Zhang, X. (M.), F. Zhu. 2011. Group size and incentives to con-
tribute: A natural experiment at Chinese Wikipedia. Amer.
Econom. Rev. 101(4) 1601–1615.
Zhu, F., X. (M.) Zhang. 2010. Impact of online consumer reviews
on sales: The moderating role of product and consumer char-
acteristics. J. Marketing 74(2) 133–148.
