Close
About
FAQ
Home
Collections
Login
USC Login
Register
0
Selected
Invert selection
Deselect all
Deselect all
Click here to refresh results
Click here to refresh results
USC
/
Digital Library
/
University of Southern California Dissertations and Theses
/
Essays on the role of entry strategy and quality strategy in market and consumer response
(USC Thesis Other)
Essays on the role of entry strategy and quality strategy in market and consumer response
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
ESSAYS ON THE ROLE OF ENTRY STRATEGY AND QUALITY STRATEGY
IN MARKET AND CONSUMER RESPONSE
by
Sajeev Vijayakumari Krishnan Nair
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
(BUSINESS ADMINISTRATION)
May 2022
Copyright 2022 Sajeev Vijayakumari Krishnan Nair
ii
Dedicated to
My Amma and Achchan
iii
Acknowledgments
They say it takes a village to produce a Ph.D. graduate. I want to acknowledge the key
people who directly or indirectly supported me in completing this dissertation.
I have been incredibly fortunate to work with my advisor Professor Gerard J. Tellis. I met
Gerry as a Master’s student, and he instilled in me the confidence to venture into academic
research. I am extremely grateful to him for his constant support, guidance, and motivation to
overcome all the challenges I faced in the Ph.D. journey. His passion for research is infectious,
and his positivity kept me going despite the hurdles I encountered in the program.
I am sincerely thankful to my dissertation committee members – Professors S. Siddarth,
Dinesh Puranam, and Nan Jia – who always offered their time and expertise to sharpen my
dissertation research. I also thank other faculty members of the Marketing department,
particularly Gary Frazier, Shantanu Dutta, Anthony Dukes, Lan Luo, Kristin Diehl, Max Wei,
and Linda Hagen, for their support at various stages in my stay at USC Marshall.
I would like to thank Elizabeth Mathew for helping me with all administrative matters,
Doris Meunier, the go-to person for numerous requirements, and Jennifer Lim for the timely
processing of payments that made life much easier. I also thank Julie Phaneuf for her help in
facilitating the Ph.D. program.
My Ph.D. life would have been much more difficult without the support of my friends in
the program. I am grateful to Jihoon Hong for the countless conversations we had over the years
on coursework, research, and life. I thank Wensi Zhang for selflessly helping me with
challenging coursework, Gizem Ceylan for offering me support when I coped with personal
losses, and Alex Yao for his feedback on my papers. I am also thankful to Mengxia Zhang,
Chaumanix Dutton, and Isamar Troncoso for their support in the program.
iv
I could not have completed my dissertation without the support of my family. My parents
would have been the happiest to witness me complete my dissertation. They have always
encouraged me in all my pursuits, and I dedicate this dissertation to them. My brother Sudeep
has been a great source of support all through. I am grateful to my in-laws for their support,
blessings, and prayers to keep me going. I am thankful to all my extended family members in
India for their love and prayers and those in the U.S. who never let me feel that I was away from
home.
Finally, I am indebted to my wife Sreelakshmy for her unflinching support and limitless
patience while juggling her work and our kids’ affairs. I am thankful to my kids, Akshay, who
grew up so much to discuss machine learning techniques with me, and Akash, who was bemused
to see me do schoolwork yet often reminded me to complete it on time. They both have kindly
forgiven my absence from many of their school events. I couldn’t have accomplished this feat
without my beautiful family’s support.
v
TABLE OF CONTENTS
Dedication………………………………………………………………………………………....ii
Acknowledgments………………………………………………………………………………..iii
List of Tables……………………………………………………………………………………..vi
List of Figures…………………………………………………………………………………...viii
Abstract……………………………………………………………………………………………x
Chapter 1: The Impact of Price and Reviewing Frequency on Ratings of Restaurants…………...1
1.1 Introduction……………………………………………………………………………1
1.2 Theory…………………………………………………………………………………7
1.3 Research Design and Models………………………………………………………...13
1.4 Results………………………………………………………………………………..27
1.5 Discussion……………………………………………………………………………42
Chapter 2: New Product Entry for Long Term Success: Waterfall, Sprinkler, or Niche?............ 47
2.1 Introduction…………………………………………………………………………..47
2.2 Theory………………………………………………………………………………..51
2.3 Data and Method.……………………………………………………….....................57
2.4 Results………………………………………………………………………………..64
2.5 Discussion……………………………………………………………………………87
Bibliography……………………………………………………………………………………..93
Appendices
Appendix A: Lists of words used to measure expectation, disconfirmation, and mentions
of restaurant-specific attributes in text analysis………………………….100
Appendix B: Generalized Synthetic Control using difference in ratings……………….101
Appendix C: The fifty geographic markets in the IRI data set…………………………102
Appendix D: Two stage model with Heckman Correction……………………………..103
vi
List of Tables
Table 1.1 Classification of the relevant literature on the effect of price or
reviewing frequency/expertise on online ratings
6
Table 1.2 Description of variables used in regression analysis 18
Table 1.3 Comparison of Los Angeles and Las Vegas cities 20
Table 1.4 Number of restaurants and percentage of reviews by price level
in Los Angeles
30
Table 1.5 OLS regression estimates for testing hypotheses H4, H5, and H6 35
Table 1.6 Estimates of the triple differences model 36
Table 1.7 Average treatment effect on ratings for Los Angeles (vs. Las
Vegas) zip codes using the Generalized Synthetic Control
37
Table 1.8 Average treatment effect on ratings for Los Angeles (vs. Las
Vegas) zip codes for different measures of reviewing frequency
using the Generalized Synthetic Control
41
Table 1.9 Average treatment effect on ratings for Los Angeles (vs. Las
Vegas) zip codes for restaurants operational throughout the
analysis period using the Generalized Synthetic Control
41
Table 2.1 Comparison of current study with past literature on launch
strategies for new products
52
Table 2.2 Definition of independent variables used in Equation 1 and 3 66
Table 2.3 Estimates of the first stage multinomial logit model 67
vii
Table 2.4 Raw and standardized estimates of the second stage hazard
model
72
Table 2.5 Recommended vs. actual strategies used – percentage of
products
82
Table 2.6 Predictive performance of the proposed model with launch
strategy compared to that of a baseline model
86
Table B.1 Average treatment effect on difference in ratings between
frequent and infrequent reviewers for Los Angeles (vs. Las
Vegas) zip codes using Generalized Synthetic Control
101
Table D.1 Estimates of the first stage multinomial logit model 108
Table D.2 Raw and standardized estimates of the second stage hazard
model
112
viii
List of Figures
Figure 1.1 Minimum wage increases in Los Angeles but not in Las Vegas
over the analysis period
19
Figure 1.2 The trend of mean star ratings by frequent and infrequent
reviewers for restaurants in Las Vegas and Los Angeles before
and after minimum wage increase in Los Angeles for high priced
restaurants
23
Figure 1.3 Average star rating by frequent and infrequent reviewers for
restaurants in the Michelin guide and the LA Times vs. other
restaurants in Los Angeles
28
Figure 1.4 The average number of expectation and negative disconfirmation
words per review of frequent and infrequent reviewers at
different price levels for restaurants in Los Angeles
31
Figure 1.5 Average star rating by frequent and infrequent reviewers at
different price levels for restaurants in Los Angeles
32
Figure 1.6 The proportion of negative (1- & 2-star) and positive (4- & 5-
star) ratings by frequent and infrequent reviewers at different
price levels for restaurants in Los Angeles
33
Figure 1.7 Trend of infrequent reviewers’ ratings for high price level in
treated and synthetic control zip codes
38
Figure 2.1 Examples of the three launch strategies 59
Figure 2.2 Baseline hazard of failure of the three strategies for selected
categories
79
Figure 2.3 Predicted probability of new product failure as marketing mix
variables in the second stage model increase for selected
categories
81
ix
Figure 2.4 Factors related to recommended strategies 84
Figure 2.5 Products launched using the niche strategy expand well beyond
the initial set of markets
90
Figure C.1 Fifty geographical markets in IRI academic data set 102
x
Abstract
Quality and market entry are two strategic factors that influence the market leadership of
firms. This dissertation examines how consumers report quality heterogeneously in online
reviews in response to price changes and how managers could use various new product entry
strategies for favorable market outcomes.
Chapter 1 examines how frequent and infrequent reviewers report the quality of
restaurants differently in online review platforms, how changes in price influence reviewers’
ratings of quality, and the potential impact reviewers’ heterogeneous reviewing behavior has on
businesses. I compiled an extensive data set of over 2 million reviews that involved a quasi-
experimental design between two similar cities (LA and Las Vegas), two time periods (separated
by a change in the minimum wage in LA), and different reviewers’ frequency. I use text analysis,
difference-in-difference-in-differences, and generalized synthetic control to identify the causal
impact of price and reviewing frequency on ratings. I find that frequent reviewers write reviews
that are significantly more lengthy, useful, multi-dimensional, non-extreme, authentic, and
consistent with restaurants’ expert reviewers’ than do infrequent reviewers. These results suggest
that frequent reviewers are more knowledgeable about restaurants than infrequent reviewers.
Second, infrequent reviewers give lower ratings to high-priced restaurants than frequent
reviewers. I show that these low ratings could substantially negatively impact high-priced
restaurants’ revenue. These results have important implications for review platforms, consumers,
and businesses listed on the platforms.
Chapter 2 studies the different new product launch strategies that consumer packaged
goods companies use, which strategies lead to longer survival, and what category-specific factors
could guide the choice of the best launch strategy. Launch strategies play a critical role in the
xi
long-term success of new products. Identifying the best strategy for long-term survival is
managerially relevant. In this paper, I test the ten-year survival of over 650 new products
launched in 18 product categories in the U.S. I use a two-stage model: 1) multinomial logit
model of strategy choice as a function of the type of firm and category and 2) hazard of survival
as a function of strategy choice plus controls. I find that firms widely use a strategy that has not
been analyzed in the literature – niche. Even though the niche strategy has the widest use, the
waterfall strategy has the longest survival. Further, I find a substantial mismatch between firms’
current choice of strategy and the best strategy for long-term survival. I further show that a
model incorporating launch strategy as an independent variable has superior predictive
performance in out-of-sample tests. This study has critical implications for new product
strategies.
1
Chapter 1: The Impact of Price and Reviewing Frequency on
Ratings of Restaurants
1.1 Introduction
Consumers constantly seek information about product quality to make purchase decisions
(Milgrom & Roberts, 1986). Previously, consumers sought such information from various
sources like ads, websites, product catalogs, recommendations of experts, and word-of-mouth
communication from other consumers. The growth of e-commerce and online reviews has vastly
increased the quantity and use of peer-to-peer information. A recent survey by Murphy (2020)
found that 93% of U.S. consumers search online for a local business and 87% of consumers read
online reviews to get information about those businesses. Notably, Yelp.com, a review platform
founded in 2004, is among the top 3 websites (apart from Google and Facebook) that consumers
use to find local business information. It had more than 178 million monthly users across its web
and mobile platforms in 2020
1
. In comparison, Consumer Reports, a non-profit organization
founded in 1936 that provides expert opinions on products and services, has only about 6 million
paid members. Thus, online reviews have emerged in recent decades as a fundamental and
premier source of information on product quality for consumers.
So, in the last two decades, the academic literature has extensively documented the
impact of online reviews and ratings on purchase decisions (Chevalier & Mayzlin, 2006;
Chintagunta et al., 2010; Floyd et al., 2014; Luca, 2011; Rosario et al., 2016). Some studies have
demonstrated how online reviews provide valuable information about key product attributes,
dimensions of product quality, and market structure in a product category (Lee & Bradlow, 2011;
1
https://www.reviewtrackers.com/blog/yelp-factsheet/
2
Netzer et al., 2012; Tirunillai & Tellis, 2014). More recently, some studies have focused on the
antecedents of online ratings (e.g., Nguyen et al. 2021; Schoenmueller et al. 2020). We
contribute to the literature that examines antecedents of online ratings by addressing two
essential gaps.
First, only three studies have examined the impact of price on online ratings. Price is a
widely researched variable in the marketing literature and plays an important role in consumers’
product evaluations. However, the documented impact of price on online ratings is contradictory.
Li and Hitt (2010) and Luca and Reshef (2021) show that higher prices lead to lower ratings,
whereas de Langhe et al. (2016) document that high prices lead to high ratings. In this context, a
deeper understanding of how price impacts online ratings is necessary.
Second, it is unclear if any factors moderate the price-rating relationship in online
reviews. One could potentially explain the contradictory effect of price on online ratings using a
moderator whose impact on the price-rating relationship has not been highlighted in the extant
literature. The previous literature has documented that consumers with different levels of product
knowledge differentially weight price information when they evaluate products (Rao & Monroe,
1988). Moreover, several studies have documented how reviewing frequency or reviewing
expertise
2
affects the distribution of ratings, motivating us to explore it as a potential moderator
of the price-rating relationship (Moe & Schweidel, 2012; Nguyen et al., 2021; Schoenmueller et
al., 2020; Wojnicki & Godes, 2017).
We address the above gaps by answering the following research questions: How does the
price of a product influence its online ratings? Does reviewing frequency of reviewers moderate
2
Reviewing expertise has been operationalized as reviewing frequency in the literature (see Nguyen et al. 2021).
This is primarily because several online review platforms label users as experts based on the number of reviews they
contribute. We use reviewing frequency as the construct in this paper but discuss later how frequent reviewers may
demonstrate expertise.
3
the price-rating relationship in online reviews? If yes, what is its moderating effect on the price-
rating relationship? This is the only study that addresses all three questions.
The theory on the price-quality relationship has shown that consumers use price as a
signal of quality (Ding et al., 2010; Milgrom & Roberts, 1986; Shapiro, 1983). Generally,
consumers have high expectations of quality when prices are high. However, the extent to which
consumers use price to judge quality varies with their product familiarity. Frequent reviewers are
likely to have more product-related experiences and hence more product familiarity than
infrequent reviewers (Alba & Hutchinson, 1987). So, frequent reviewers could use various
intrinsic product attributes along with price information to evaluate a product’s quality. But
infrequent reviewers may have limited ability to judge quality and rely more on their price-
driven expectations to evaluate products. This potential interaction between price and reviewing
frequency on ratings of restaurants may drive the effects we find.
We address our research questions using data from the review platform Yelp.com. The
data includes 2.2 million reviews of over 5000 restaurants in Los Angeles and Las Vegas cities
in the U.S. from about 900,000 reviewers. We leverage data from a quasi-experiment involving
two cities (Los Angeles and Las Vegas), two time periods (separated by a wage increase in Los
Angeles but not in Las Vegas), and frequent and infrequent reviewers. Using text analytics,
difference-in-difference-in-differences (hereafter called triple differences), and generalized
synthetic control, we identify the causal impact of price and reviewing frequency on ratings. We
also perform several robustness checks using alternative measures of reviewing frequency and a
restricted sample of restaurants that controls for restaurant entry and exit.
Three key findings emerge from our study. First, frequent reviewers write reviews that
are significantly more 1) lengthy, 2) useful, 3) multi-dimensional, 4) non-extreme, 5) authentic,
4
and 6) consistent with restaurant critics than infrequent reviewers do. Frequent reviewers also
rate restaurants in the Michelin guide and the LA Times higher than other restaurants in our
sample. These results suggest that 1) frequent reviewers are more knowledgeable than infrequent
reviewers and 2) infrequent reviewers evaluate restaurants more critically than frequent
reviewers, which is different from prior studies that show that frequent reviewers evaluate
products more critically than infrequent reviewers (Moe & Schweidel, 2012; Nguyen et al.,
2021; Schlosser, 2005; Zhang et al., 2016). Second, infrequent reviewers systematically give
lower ratings than frequent reviewers for high-priced restaurants. Thus, the finding that the price-
rating relationship is positive for frequent reviewers and negative for infrequent reviewers offers
a potential explanation for the contradictory findings on the effect of price on ratings. This result
also suggests that average Yelp ratings of all reviewers may be biased by infrequent reviewers.
Third, infrequent reviewers’ harsh rating behavior leads to a half-star drop in ratings of 11% of
high-priced restaurants in our sample. This drop is substantial because an extra half-star on Yelp
causes restaurants to sell out 49% more frequently, and a one-star increase in Yelp rating leads to
a 5-9 percent increase in revenue (M. Anderson & Magruder, 2012; Luca, 2011). Our findings
have several implications.
First, platforms such as Yelp and Amazon need to aggregate ratings by frequent vs.
infrequent reviewers, in addition to those by verified and non-verified reviewers. Aggregating
ratings by reviewers’ reviewing frequency may be more useful to consumers who read reviews
than a single rating that masks the heterogeneous reviewing behavior. This approach could
increase the confidence users and businesses have in the review platform.
Second, consumers should consider the type of reviewers who have posted reviews for a
product. The previous literature has documented that consumers tend to weight the average
5
rating more than other factors like the number of ratings (de Langhe et al., 2016; Watson et al.,
2018). Consumers who read reviews may find that their preferences align better with one type of
reviewer (frequent vs. infrequent) than the other. If so, they may benefit from weighting that
reviewer’s rating more than the other’s while making purchase decisions.
Third, businesses listed on review platforms need to consider reviewers’ reviewing
frequency while responding to their comments. For instance, frequent reviewers’ feedback may
be more informative than that of infrequent reviewers for managerial action. Management
responses to online reviews are becoming more relevant in customer satisfaction management
(Chevalier et al. 2018, Proserpio and Zervas 2017). Our results suggest that online reputation
management systems should also account for reviewing frequency.
We make four important contributions to the marketing literature. First, while a few
papers have studied the effect of price on ratings, ours is the only paper that finds that the
relationship is positive for frequent reviewers and negative for infrequent reviewers. Our theory
suggests that this asymmetry may be due to frequent reviewers being better able to assess the
quality of premium services than infrequent reviewers. Second, while the extant literature shows
that frequent reviewers are more critical than infrequent reviewers, we find that the reverse is
also true, especially at high price levels. Third, we contribute to the literature on price-perceived
quality by documenting conditions under which price is negatively related to ratings of quality
instead of the conventional positive relationship. Fourth, very few studies that investigate the
effect of price and reviewing frequency on online ratings have been published in top quantitative
marketing journals. Published studies have documented the effect of either price or reviewing
frequency but not both. Table 1.1 shows how our study is different from previous studies. The
rest of the paper covers the theory, research design, model, results, and discussion.
6
Table 1.1: Classification of the relevant literature on the effect of price or reviewing
frequency/expertise on online ratings
Study Effect of price
on ratings is
positive for
frequent
reviewers and
negative for
infrequent
reviewers
Infrequent
reviewers post
more negative
ratings in
rising price
environments
Effect of
price and
reviewing
frequency/
expertise
on ratings
Published in
Marketing
Science or
Management
Science
Current article Yes Yes Yes ?
Luca and Reshef 2021
No No
No Yes
Moe and Schweidel 2012 No No No Yes
Li and Hitt 2010 No No No No
De Langhe et al. 2016
No No
No No
Schlosser 2005 No No No No
Zhang, Zhang, and Yang
2016
No No No No
Wojnicki and Godes 2017 No No No No
Nguyen et al. 2020 No No No No
Schoenmueller, Netzer,
and Stahl 2020
No No No No
7
1.2 Theory
This section builds the theory for the effect of price and reviewing frequency on online
ratings of restaurants. It answers three questions. What are the linguistic differences between
frequent and infrequent reviewers’ reviews? Are frequent reviewers more knowledgeable than
infrequent reviewers? How does price and reviewing frequency interactively affect the ratings
reviewers post? We also present in this section two sets of hypotheses. The first three hypotheses
are process tests that aim to link our construct reviewing frequency to the theoretical concept of
consumer expertise. The next three are the main hypotheses that consider the impact of price and
reviewing frequency on ratings.
1.2.1 Quality and online ratings
Quality is one of the most critical factors that influence the market leadership of firms
(Tellis et al., 2009). It has a strong positive impact on firm profitability, customer satisfaction,
and stock market returns (Rust et al., 1995, 2002; Tellis & Johnson, 2007). While scholars from
various disciplines like marketing, economics, and management have examined quality from
different perspectives, it is widely accepted that quality is a multi-dimensional construct (Klein
& Leffler, 1981; Tirunillai & Tellis, 2014).
The multiple dimensions of quality imply that consumers find it difficult to evaluate
quality (Curry & Faulds, 1986; Kopalle & Hoffman, 1992). The development of a composite
measure of product quality involves choosing various product attributes, assigning weights to
these attributes, rating each attribute, and then generating a weighted average index of quality.
Consumers often lack the required information and hence the ability to evaluate the quality of
products (Tellis & Wernerfelt, 1987). To overcome this problem, they could gather information
8
about quality from expert sources like Consumer Reports magazine that test products and
provide quality ratings. Such reviews of product quality are valuable and positively impact
abnormal stock returns of firms whose products are reviewed (Aaker & Jacobson, 1994; Tellis &
Johnson, 2007). Reviewed quality posted online by peer consumers on review platforms has
become a very popular source of information about product quality for many consumers. Studies
suggest that consumers rely more on the average star rating of a product than other factors like
the number of reviews to infer its quality (de Langhe et al., 2016; Watson et al., 2018).
1.2.2 Reviewer heterogeneity
Due to the importance of online ratings in consumers’ purchase decisions, it is important
to understand how reviewers’ heterogenous reviewing behavior affects the ratings they post. We
explain below how frequent and infrequent reviewers write reviews and rate restaurants
differently.
According to Alba and Hutchinson (1987), consumers who have more product-related
experiences are likely to be more familiar with the product category. Generally, increased
product familiarity is associated with increased expertise. Frequent reviewers, who often visit
restaurants and describe their dining experience on review platforms, are likely to have more
expertise than infrequent reviewers. This expertise could reflect in frequent reviewers’ reviews in
three ways. First, experts can elaborate on given information. They can consider more product-
related facts than novices to make inferences and explain their evaluations. So, frequent
reviewers may write longer reviews than infrequent reviewers. Second, as frequent reviewers
consider more product attributes to infer product quality, they may mention a greater number of
attributes in their reviews than infrequent reviewers do. Third, as frequent reviewers can explain
9
their quality inferences in detail, consumers who read reviews may find frequent reviewers’
reviews more useful than infrequent reviewers’. Thus, we hypothesize:
H1: Frequent reviewers write more 1) lengthy, 2) multi-dimensional, and 3) useful reviews than
infrequent reviewers.
As frequent reviewers are likely more knowledgeable than infrequent reviewers, they
may have a better ability to judge the quality of restaurants. Frequent reviewers should be able to
differentiate between high and low-quality restaurants from their experience visiting and
reviewing numerous restaurants. Thus, their ratings of restaurants should closely correspond to
restaurant critics’ evaluations. So, we hypothesize:
H2: Frequent reviewers’ ratings of restaurants are more consistent with restaurant critics’
evaluations.
Frequent and infrequent reviewers may also differ in the linguistic features of the text
they write in addition to the length and multi-dimensional nature of reviews. Kronrod, Lee, and
Gordeliy (2021) propose that authentic and fictitious reviews are linguistically different. They
suggest that authentic reviews have more concrete language, use more past tense verbs, and more
low-frequency words. As frequent reviewers are likely more knowledgeable than infrequent
reviewers, we propose that frequent reviewers may write more authentic reviews than infrequent
reviewers. Hence we hypothesize:
H3: Frequent reviewers write more authentic reviews than infrequent reviewers.
10
1.2.3 Price, reviewing frequency, and ratings
According to the expectancy disconfirmation paradigm, the key determinants of
consumers’ evaluation of products are expectations, perceived performance, and disconfirmation
(Oliver, 2014; Yi, 1990). We explain below how price and reviewing frequency could
interactively affect these three factors, resulting in frequent and infrequent reviewers producing
very different ratings for the same restaurant.
Consumers use various intrinsic and extrinsic cues to form expectations of product
quality (Zeithaml, 1988). Intrinsic cues are derived from the actual physical product and are
directly related to the product’s quality. For example, nutritional information like the protein
content or sodium level in a meal could convey its quality to consumers. Extrinsic cues are
typically product-related attributes that are not part of the physical product itself. Examples are
price, level of advertising, brand name, country of origin, and store name (Teas & Agarwal,
2000).
The literature on the price-quality relationship has found that consumers use price as an
important signal of quality (Ding et al., 2010; Klein & Leffler, 1981; Milgrom & Roberts, 1986;
Shapiro, 1983). Compared to the effect of other extrinsic cues, the effect of prices on forming
expectations of quality seems to be very robust, as suggested by two major meta-analyses
conducted two decades apart (Rao & Monroe, 1989; Völckner & Hofmann, 2007). The
consistent finding from this stream of literature is that consumers perceive high quality for high-
priced products. Consumers may rely more on price as a signal of quality for experience goods
like restaurants, which is the focus of the current study, because they may be unable to evaluate
other attributes of quality before purchase (Zeithaml, 1988).
11
Past research suggests that novices are more likely than experts to use price as an
indicator of quality (Miyazaki et al., 2005; Rao & Monroe, 1988). Experts tend to rely less on
price as a signal of quality when the availability of intrinsic information is high (Chang & Wildt,
1996; Cordell, 1997). Thus, frequent reviewers, who may have more expertise than infrequent
reviewers, are more likely to use their rich intrinsic knowledge to form expectations of quality. It
is not clear from the theory whether expectations created by price and other extrinsic cues will be
stronger for infrequent than for frequent reviewers, whose expectations are created mainly by
intrinsic knowledge. The difference in the level of expectations of frequent and infrequent
reviewers is an empirical issue that we test using our data. Hence, we hypothesize:
H4: Reviewers have higher expectations of quality for high-priced than low-priced restaurants.
Frequent and infrequent reviewers could perceive the performance of a product
differently. Past research suggests that experts and novices process product-related information
differently (Hong & Sternthal, 2010; Johnson & Russo, 1984; Park & Lessig, 1981). Experts
possess highly developed cognitive structures compared to novices (Alba & Hutchinson, 1987).
Cognitive structure refers to factual knowledge that consumers have about products and the
organization of that knowledge (Brucks, 1986). As they have more product knowledge, frequent
reviewers are better equipped than infrequent reviewers to understand the meaning of product
information. Frequent reviewers are also more likely than infrequent reviewers to consider
numerous product attributes to assess product performance (Golder et al., 2012). For example,
frequent reviewers may judge a steak by examining its cut, color, juiciness, marbling, and texture
and, as a result, can assess the quality of the steak in greater detail than infrequent reviewers.
Infrequent reviewers may judge the same steak by considering if it is soft vs. hard because they
12
have limited information to evaluate subtle differences in quality. They may not be able to fully
appreciate multiple attributes of a product that contribute to its overall performance. The result is
a gap between expectations of quality and the actual experience with the product, which is called
disconfirmation. A positive disconfirmation occurs when the experience is better than
expectations, and a negative disconfirmation occurs when the experience is worse than
expectations (Kopalle & Lehmann, 2001; Oliver, 1980). For high prices, infrequent reviewers
end up with high expectations, but they have limited ability to judge quality. Hence, we
hypothesize:
H5: Infrequent reviewers experience more negative disconfirmation than frequent reviewers for
high-priced restaurants.
Disconfirmation is likely to affect the ratings consumers post. Ratings tend to be above
average for positive disconfirmation and below average for negative disconfirmation (E. W.
Anderson & Sullivan, 1993; Oliver, 1980). As explained previously, infrequent reviewers are
more likely to experience negative disconfirmation than frequent reviewers during their
evaluation of high-priced products, making them more likely than frequent reviewers to assign
poor ratings to high-priced products. Hence, we hypothesize:
H6: Infrequent reviewers assign lower ratings than frequent reviewers for high-priced
restaurants.
13
1.3 Research Design and Models
This section describes the data, process tests (H1 to H3), measures, and the various
models used to test our main hypotheses (H4 to H6).
1.3.1 Data
We scraped review data for 990 restaurants in Los Angeles city for the time period
October 2004 to September 2019 from the review platform Yelp.com. After dropping a few
restaurants without price information from this set, the final dataset contained 871,503 reviews
of 921 restaurants from 398,354 reviewers. The data provides three types of information: 1)
business, which includes the name, address, total number of reviews received, average star
rating, and various attributes for each restaurant, 2) user, which includes total reviews posted, an
average of star ratings posted, date of joining the platform, and various compliments received by
each user, and 3) review, which includes the star rating, review text, review date, and
compliments received for each review of restaurants contained in business. The descriptive
results and regression analysis that follow are based on this data. We supplemented this data with
three similar datasets for Las Vegas that Yelp made publicly available for academic research
3
.
We discuss our identification strategy and the logic underlying our choice of these cities in a
later subsection.
1.3.2 Process tests and measures
Following Schoenmueller, Netzer, and Stahl (2020), we calculate the frequency of
reviewing as the number of reviews written by a reviewer per month since joining Yelp. We can
identify when a reviewer joined the platform and the total number of reviews the individual has
3
https://www.yelp.com/dataset
14
posted from the user data. We classify reviewers in the upper and lower quartiles of reviewing
frequency as frequent and infrequent reviewers respectively.
We compare the differences between frequent and infrequent reviewers using the
approach of Nguyen et al. (2021) for our first process test hypothesis. We first preprocessed the
review text from the review data to remove all punctuations and stop words (e.g., “and,” “then,”
“on,” “in,” etc.) to obtain a set of unique words. We then retained only words that occur at least
five times in our review corpus to create a vocabulary of words. Thus, each review is broken
down into a set of words contained in this vocabulary. We used an algorithm in Python to count
the number of words in each review. We also applied part-of-speech tagging using the Natural
Language Toolkit module in Python to label the words in our dictionary as nouns, adjectives,
adverbs, etc. We then focused on the nouns as they will closely correspond to restaurant-specific
attributes. We created a list of 100 restaurant-specific nouns most frequently used in our data,
such as food, service, dinner, etc. (see Appendix A). We then used an algorithm to count the
number of unique nouns in this list used in each review. Thus, in each review, we obtained the
number of restaurant-specific attributes mentioned.
For the second process test, we check whether frequent or infrequent reviewers’ ratings
correspond to restaurants’ expert reviewers’ ratings. We consider two kinds of restaurants’
expert reviews for our analysis – Michelin guide and the LA Times. The guide awards Michelin
stars, which is the most coveted stamp of quality, to a select few restaurants in various cities. In
addition to giving the Michelin stars, the guide also recognizes high-quality restaurants by
including them in two separate lists called Bib Gourmand and The Plate Michelin
4
. The guide
uses anonymous restaurant inspectors to check the quality of restaurants. We identified 66
4
https://guide.michelin.com/us/en/california/to-the-stars-and-beyond
15
restaurants in our sample in Los Angeles that are listed in the Michelin guide. The LA Times
also publishes a list of 101 best restaurants in Los Angeles after the daily’s restaurant experts’
evaluation. We compared frequent reviewers’ ratings for restaurants in the Michelin guide and
the LA Times with 1) infrequent reviewers’ ratings for the same restaurants and 2) frequent
reviewers’ ratings for restaurants not featured in the guide or the LA Times.
For our final process test, we compare linguistic differences between frequent and
infrequent reviewers’ reviews using the approach of Kronrod, Lee, and Gordeliy (2021) to verify
which type of reviewer writes more authentic reviews. They propose that authentic reviews use
more past tense verbs, more low-frequency words, and have more concrete language. Using the
part-of-speech tagged reviews, we count the number of verbs and past tense verbs in each review
to obtain the proportion of past tense verbs used. We divided the total number of words, N, in
our corpus of reviews by the number of unique words, F, to obtain the average number of times a
unique word occurs in the corpus. A low-frequency word is any word that occurs less than this
average of N/F.
We obtained the measures to test our main hypotheses H4 and H5 from the review text.
These hypotheses relate to expectation formation and negative disconfirmation experienced by
consumers. We created two lists of words to measure the extent of expectation and negative
disconfirmation that reviewers express in the text they post (see Appendix A). The words in the
expectation list include expect, hype, hope, anticipate, disappoint, dismay, etc. The words in the
negative disconfirmation list are a subset of words in the expectation list with a negative valence
such as disappoint, dismay, dishearten, dissatisfy, etc. Note that words in the negative
disconfirmation list indirectly suggest expectation, as these words are likely used when a
reviewer’s expectations are not met. We used an algorithm to calculate the number of times
16
words in these lists appear in each review. We then compared the average occurrence per review
of words in the expectation and negative disconfirmation lists for frequent and infrequent
reviewers at different price levels.
1.3.3 Models
We use three types of models in our empirical analysis. First, we estimate a preliminary
regression model on Los Angeles data to test our three main hypotheses (H4 to H6). Second, we
leverage a quasi-experimental design to estimate a triple difference to test the causal effect of
price increase on ratings (H6). Third, we use the generalized synthetic control to test H6 even
more rigorously.
1.3.3.1 Preliminary regression model
We specify the preliminary regression model as follows. For review i of restaurant j in
time t by reviewer r, we model the dependent variable as follows:
Yijrt = b1PriceLevelijt + b2ReviewerFrequencyijrt + b3PriceLevelijt x ReviewerFrequencyijrt +
b4AgeYelpr + b5Eliteir + b6UserReviewCountr + b7BusinessReviewCountj + b8Cuisinej +
ZipCodej + tt + eijrt (1)
The dependent variable is expectation words per review, negative disconfirmation words per
review, or star ratings. The main independent variables are the restaurant's price level, the
reviewer's frequency (frequent or infrequent), and the interaction between them.
We next include user-related variables and restaurant-specific characteristics as control
variables in our model. We also add location fixed effects (ZipCode), cuisine fixed effects, and
17
year fixed effects (tt). Table 1.2 describes the variables we use for our regression analysis. Note
that we treat PriceLevel as a continuous variable for easier interpretation of results even though it
is a categorical variable in the data.
The preliminary regression model uses variation in price level within Los Angeles to
estimate its effect on frequent and infrequent reviewers’ ratings. However, the impact of price
level on ratings in this model is likely biased as we do not account for unobservable factors
correlated with prices that may affect ratings. We next explain a quasi-experimental design and
two models that exploit it to estimate the causal effect of prices on ratings.
1.3.3.2 Quasi-experimental design
The estimation of the causal effect of price on ratings requires exogenous variation in
prices such that the changes in price are not correlated with changes in ratings. Since the price
level of restaurants is a categorical variable in our data for Los Angeles and does not change over
time, we look for other exogenous price variation to obtain the causal effect of prices on ratings.
We find the desired variation in a quasi-experiment involving differential minimum wage
changes in Los Angeles and Las Vegas.
About three-fourth of the workers in the U.S. earning the minimum wage or less in 2019
were employed in service occupations, mostly in food preparation and serving related jobs
5
.
Because labor costs account run about 28%-35% of restaurant sales, minimum wages are likely
to impact restaurant prices
6
. Past research suggests that restaurant prices increase when minimum
wages increase (Aaronson, 2001; Allegretto & Reich, 2018; Lemos, 2008). Figure 1.1 shows that
5
https://www.bls.gov/opub/reports/minimum-wage/2019/home.htm
6
https://upserve.com/restaurant-insider/how-much-does-it-cost-to-open-a-restaurant/
18
the minimum wage had increased on five occasions in Los Angeles city from $8 per hour in
January 2010 to $13.25 per hour in July 2019. At the same time, the minimum wage in Las
Vegas was constant at $8.25 per hour. We consider the increase in the minimum wage in Los
Angeles as the treatment in our quasi-experimental setting, with Jan 2010 to June 2014 as the
control period and July 2014 to Sep 2019 as the treatment period. Thus, Los Angeles is the
treated city and Las Vegas the control city in our setting.
Table 1.2: Description of variables used in regression analysis
Variable Description
Expectation Number of expectation words in a review
Negative disconfirmation Number of negative disconfirmation words in a review
Rating Rating on a 5-point scale that a restaurant receives in a review
PriceLevel Categorical variable (four levels) indicating the cost of a meal
per person at a restaurant
1: under $10, 2: $11-30, 3: $31-60, 4: above $61
ReviewerFrequency Indicator for type of reviewer (1 for infrequent reviewer)
AgeYelp Number of months ago a reviewer joined Yelp from the month
of review collection
Elite Indicator variable for a reviewer who has received the Elite
badge from Yelp at least for one year in the data period
UserReviewCount Number of reviews a reviewer has posted on Yelp
BusinessReviewCount Number of reviews a restaurant has received on Yelp
Cuisine Dummies for top 10 most common cuisine types in the dataset
(63% of total reviews) – American (New), Korean, Italian,
Mexican, Japanese, Asian Fusion, Mediterranean, American
(Traditional), Chinese, Thai. Remaining restaurants were
classified under Others.
ZipCode Zip code of the location of the restaurant
19
Figure 1.1: Minimum wage increases in Los Angeles but not in Las Vegas over the analysis
period
We choose Las Vegas as the comparison city because it is very similar to Los Angeles in
several important respects, as shown in Table 1.3. First, both cities are major entertainment hubs
in the southwestern U.S. Second, the number of visitors to each city, who likely visit restaurants
and post reviews online, is very similar. Third, the percentage of household income spent on
food away from home is the same in both cities suggesting a similar spending pattern on
restaurant meals at a household level. Fourth, the number of full-service restaurants per capita,
though slightly higher in Los Angeles, is still comparable across the two cities. Fifth, expenditure
per full-service restaurant is also comparable across the two cities though it is slightly higher in
Las Vegas.
20
Table 1.3: Comparison of Los Angeles and Las Vegas cities
a
Annual visitors
b
(millions)
% of household
income spent on
food away from
home
c
(%)
Full-service
restaurants per
capita
d
(per
million persons)
Spend per full-
service
restaurant
e
($000)
Year LA Vegas LA Vegas LA Vegas LA Vegas
2011 40.4 38.9 - - 717 566 - -
2012 41.4 39.7 - - 716 577 - -
2013 42.2 39.6 - - 723 586 - -
2014 44.2 41.1 - - 717 578 - -
2015 45.6 42.3 4.5% 4.4% 793 645 $746 $851
2016 47.3 42.9 4.5% 4.4% 807 673 $778 $856
2017 48.3 42.2 4.6% 4.5% 822 682 $827 $915
2018 50.0 42.1 4.6% 4.5% 831 686 $846 $954
a
Some data are at the city-level and others at the county level as indicated below; Las Vegas is in Clark
County in Nevada state
b
Los Angeles Department of Convention and Tourism Development; Las Vegas Convention and Visitors
Authority; city-level data
c,e
Consumer Expenditure Estimates, SimplyAnalytics; county-level data
d
County Business Patterns Summary, U.S. Census Bureau; county-level data
1.3.3.3 Identification strategy
Our identification strategy is to compare frequent and infrequent reviewers’ ratings for
high-priced restaurants in Los Angeles with high-priced restaurants in Las Vegas before and
after the minimum wage changes. We argue that in the posttreatment period, price increases in
Los Angeles, driven by an increase in the minimum wage (Allegretto & Reich, 2018), will cause
the average ratings of infrequent reviewers for high-priced restaurants in Los Angeles to drop
significantly more than those of infrequent reviewers for similar restaurants in Las Vegas. We do
not expect this drop in ratings for low-priced restaurants and frequent reviewers.
21
A key assumption in our identification strategy is that the increase in the minimum wage
in Los Angeles is an exogenous shock on prices but not on the ratings of restaurants in the city.
But restaurants could reduce staffing, in addition to raising prices, to manage higher operating
costs resulting from the increase in the minimum wage. This reduction in manpower may
adversely affect the ratings of restaurants in the city. However, we believe that this reduction in
rating is unlikely in our setting for three reasons. First, past research on the impact of minimum
wage on employment does not conclusively prove that employment levels in restaurants drop
due to an increase in the minimum wage. While some studies find a positive effect of increasing
the minimum wage on employment levels (Card & Krueger, 1994; Dube et al., 2010), others
suggest a low to moderate negative effect (Aaronson et al., 2008; Neumark & Wascher, 2000).
So, it seems a drop in rating due to the reduction of employees is less likely. Second, large chain
restaurants like McDonald’s, Burger King, Subway, etc., are more likely to manage their costs
using automated kiosks or robots to replace employees
7
. This adjustment involves the potential
risk of damaging their reputation due to a perceived drop in service quality. However, our data
does not include these chain restaurants. We believe that independent restaurants in our sample
are more averse to risking their reputation by reducing the number of employees or using lower
quality ingredients. High-priced restaurants, whose ratings we expect to drop in response to the
increase in the minimum wage, are likely to value their reputation more than low-priced
restaurants do. Thus, they may be even more careful about cutting corners to manage costs.
Third, recent evidence shows that when minimum wages increase, consumers perceive
higher service quality in restaurants as the higher wages incentivize workers to improve service
(Puranam et al., 2021). This finding implies that even if the increase in minimum wage affects
7
https://www.businessinsider.com/minimum-wage-increases-spur-fast-food-chains-to-consider-automation-2018-1
22
ratings, its effect is likely positive. So, a drop in infrequent reviewers’ ratings in Los Angeles
will provide very strong evidence of the negative impact of price increase (driven by the increase
in the minimum wage) on ratings.
Figure 1.2 shows the trend of average star ratings by frequent and infrequent reviewers
for high price levels in both cities before and after the increase in the minimum wage in Los
Angeles. Frequent reviewers’ ratings do not drop much in Los Angeles after the increase in the
minimum wage in July 2014, which is the first instance of wage increase in the city in our data,
but they slightly increase in Las Vegas during the same period. However, infrequent reviewers’
ratings further decline in Los Angeles, continuing the declining trend in the control period while
increasing in Las Vegas. Thus, as expected, the drop in infrequent reviewers’ ratings is
considerable for high-priced restaurants in Los Angeles, where prices increased due to rising
minimum wages. Importantly, such a significant drop in rating does not occur for frequent
reviewers in either city in the treatment period.
23
Figure 1.2: The trend of mean star ratings by frequent and infrequent reviewers for restaurants in
Las Vegas and Los Angeles before and after minimum wage increase in Los Angeles for high
priced restaurants
1.3.3.4 Triple differences model
We estimate the causal effect of price increase on infrequent reviewers’ ratings using a
triple differences design. This design allows us to control for two potentially confounding trends
simultaneously: changes in ratings of restaurants across the two cities and changes in ratings of
restaurants across frequent and infrequent reviewers, both of which are not related to price
increases. After controlling for these confounders, any differential trends in frequent and
infrequent reviewers’ ratings for restaurants in the two cities are due to the price increase in Los
Angeles. We specify the model for review i of restaurant j in time t by reviewer r as follows:
24
Yijrt = b0 + b1Dijt + b2ReviewerFrequencyijrt + b3PriceLevelijt + b4Dijt x ReviewerFrequencyijrt +
b5Dijt x PriceLevelijt + b6PriceLevelijt x ReviewerFrequencyijrt + dDijt x PriceLevelijt x
ReviewerFrequencyijrt + b7AgeYelpr + b8Eliteir + b9UserReviewCountr +
b10BusinessReviewCountj + b11Cuisinej + ZipCodej + gj + tt + eijrt (2)
The dependent variable is the star rating in a review. D is a treatment indicator that equals 1 for
reviews of restaurants in Los Angeles in the treatment period (from July 2014) and 0 otherwise.
gj is a restaurant fixed effect. All other variables in Equation 2 have already been described in
Table 1.2. The triple differences estimate is d, which compares the following differences: 1) the
difference between the posttreatment changes in infrequent reviewers’ ratings for restaurants in
Los Angeles and changes in infrequent reviewers’ ratings for restaurants in Las Vegas over the
same period, and 2) a similar difference for frequent reviewers’ ratings. We expect a negative
sign for d which suggests that infrequent reviewers give lower ratings than frequent reviewers for
high-priced restaurants in Los Angeles relative to Las Vegas after prices increase.
1.3.3.5 Generalized Synthetic Control
The identifying assumption of the triple differences model in Equation 2 is that the
difference in frequent and infrequent reviewers’ ratings in Los Angeles trends similarly to the
difference in their ratings in Las Vegas (Olden & Moen, 2020). Figure 1.2 shows that this
assumption may not hold. So, an alternative is to use the synthetic control (Abadie et al., 2010;
Abadie & Gardeazabal, 2003). Athey and Imbens (2017) note that the synthetic control is
“arguably the most important innovation in the policy evaluation literature in the last 15 years”.
Several marketing studies have used this model recently (Guo et al., 2020; Pattabhiramaiah et al.,
25
2019; Puranam et al., 2021; Tirunillai & Tellis, 2017). This model is similar to the difference-in-
differences framework because it compares outcomes of treatment and control units before and
after treatment. But instead of using a single control unit, synthetic control creates a simulated
control unit using a weighted average of several control units such that the pretreatment
outcomes and covariates of the synthetic control unit are matched with those for the treatment
unit. Thus, this model satisfies the parallel trend assumption that is necessary for causal
inference.
Following Xu (2017), we use the generalized synthetic control, a variant of the synthetic
control. The generalized synthetic control unifies the synthetic control approach with interactive
fixed effects models that explicitly account for unobserved time-varying factors (Bai, 2009). The
generalized synthetic control has two main advantages. First, it accommodates multiple
treatment units. We perform the analysis at the zip code level such that we have several treated
and control zip codes in Los Angeles and Las Vegas, respectively. The model will create a
synthetic control unit for each treated zip code in Los Angeles by combining several control zip
codes in Las Vegas to minimize the pretreatment differences between the treated zip code and
the synthetic control. Second, unlike the synthetic control, the generalized synthetic control
generates easily interpretable uncertainty estimates around the treatment effect like standard
errors and confidence intervals.
In the synthetic control, we must perform placebo tests in which each control unit is
iteratively assigned the treatment condition to obtain a distribution of simulated treatment
effects. Then, we can assess if the “true” treatment effect is significantly larger than the set of
simulated treatment effects. The generalized synthetic control automates this process and
26
provides standard errors and confidence intervals for the treatment effect more efficiently than
the synthetic control (Xu, 2017).
We specify the generalized synthetic control for zip code level ratings as follows:
ratingit = ditDit + x¢itb + l¢ift + eit (3)
where ratingit is the average star rating for zip code i in time t (month). Dit is a treatment indicator
that equals 1 for the zip codes in Los Angeles in the posttreatment period. dit is the main
coefficient of interest and captures the heterogeneous treatment effect of the price increase on
ratings in a treated zip code. x¢it is a vector of observed covariates, and b is the corresponding
vector of parameters to be estimated. We include the number of restaurants, reviews, and
reviewers in zip code i in time t as the covariates. ft is an (r x 1) vector of unobserved common
factors, and li is an (r x1) vector of corresponding unknown factor loadings. These two terms
capture a wide range of unobserved heterogeneities, including unit and time fixed effects. The
model assumes that the number of factors is fixed and that the same set of factors affect both the
treated and control zip codes. We calculate the average treatment effect on the treated zip codes
as 1/Ntr ∑ d
!" !#$
, where Ntr is the number of treated zip codes and T is the set of treated zip codes.
We estimate the model separately on frequent and infrequent reviewers’ ratings for high- and
low-price levels in both Los Angeles and Las Vegas. We expect the treatment effect to be
negative on infrequent reviewers’ ratings for high price levels, suggesting that infrequent
reviewers assign lower ratings for high price levels after prices increase in Los Angeles relative
to Las Vegas.
The estimation proceeds in three steps. In the first step, an integrated fixed effects model
is estimated using data from only the control zip codes to obtain the latent factors that capture the
unobserved time-varying effects on ratings for t periods as well as the coefficients of the
27
observed covariates. In the second step, the generalized synthetic control estimates the factor
loadings for each treated zip code by minimizing the mean squared error of the predicted ratings
of the treated zip codes in the pretreatment period. In the final step, using the factors obtained
from the control zip codes (in the first step) and the factor loadings obtained from the treated zip
codes (in the second step), the generalized synthetic control estimator calculates the
posttreatment ratings for the treated zip codes. These estimated ratings serve as the
counterfactuals for the treated zip codes in the posttreatment period. The estimation process
consists of a cross-validation procedure that chooses the number of latent factors that minimizes
the mean squared prediction error. The details of the estimation algorithm are available in Xu
(2017).
1.4 Results
1.4.1 Process tests and model-free evidence
We present the process test (H1 to H3) results first and then the model-free evidence for
our main hypotheses, H4 to H6.
First, we find that frequent reviewers write longer reviews than infrequent reviewers
(AvgWordsfrequent = 69 vs. AvgWordsinfrequent = 41, t = 143.28, p < .001). Second, frequent
reviewers use a greater number of restaurant-specific attributes in their reviews than infrequent
reviewers (AvgNounsfrequent = 12 vs. AvgNounsinfrequent = 7, t = 158.78, p < .001). Third, frequent
reviewers’ reviews receive a higher number of “useful” votes from other users than infrequent
reviewers’ reviews (AvgUsefulfrequent = 1.83 vs. AvgUsefulinfrequent = 0.52, t = 88.46, p < .001).
These three results support our first process test hypothesis H1.
28
Figure 1.3A shows that frequent reviewers rate the restaurants listed in the Michelin
guide significantly higher than infrequent reviewers do (AvgRatingfrequent,Michelin = 4.11 vs.
AvgRatinginfrequent,Michelin = 3.85, t = 16.01, p < .001). Additionally, frequent reviewers rate these
restaurants higher than they do for the remaining restaurants not listed in the guide
(AvgRatingfrequent,Michelin = 4.11 vs. AvgRatingfrequent,Non-Michelin = 3.99, t = 21.89, p < .001).
Similarly, figure 1.3B shows that frequent reviewers rate the restaurants listed in the LA Times
significantly higher than infrequent reviewers do (AvgRatingfrequent,LATimes = 4.18 vs.
AvgRatinginfrequent,LATimes = 3.95, t = 17.46, p < .001). Additionally, frequent reviewers rate these
restaurants higher than they do for the remaining restaurants not listed in the guide
(AvgRatingfrequent,LATimes = 4.18 vs. AvgRatingfrequent,Non-LATimes = 3.98, t = 36.73, p < .001). These
results suggest that frequent reviewers’ ratings are more consistent with restaurants’ expert
reviewers’ ratings, supporting our second process test hypothesis H2.
Figure 1.3: Average star rating by frequent and infrequent reviewers for restaurants in the
Michelin guide and the LA Times vs. other restaurants in Los Angeles
(A) (B)
29
Finally, we find that frequent reviewers use a higher proportion of past tense verbs in
their reviews than infrequent reviewers (PastTenseVerbsfrequent = 28.1% vs.
PastTenseVerbsinfrequent = 26.9%, t = 7.5, p < .001). Additionally, frequent reviewers use a greater
proportion of low-frequency words in their reviews than infrequent reviewers
(LowFrequencyWordsfrequent = 13.6% vs. LowFrequencyWordsinfrequent = 12.7%, t = 5.6, p < .001).
These results indicate that frequent reviewers write more authentic reviews than infrequent
reviewers, supporting our second process test hypothesis H3.
The results of our process tests suggest that frequent reviewers are more knowledgeable
than infrequent reviewers. We now present model-free evidence for our main hypotheses H4 to
H6.
The price variable in our data is categorical with four levels. The number of dollar signs
($) displayed near a restaurant’s name indicates its price level. The dollar signs correspond to the
price of a meal per person: $ = under $10, $$ = $11-30, $$$ = $31-60, and $$$$ = over $61.
Yelp uses the price information that reviewers submit while posting a review to display the price
level of a restaurant. Table 1.4 shows the number of restaurants in Los Angeles in each price
level and the proportion of frequent and infrequent reviewers’ reviews posted for each price
level. The two lower price-level restaurants (level 1 and 2) comprise 85% of our sample of
restaurants. Since frequent reviewers write about 82% of their reviews for these low-priced
restaurants compared to infrequent reviewers who write 84% of their reviews for the same
restaurants, we do not expect a large difference in the profile of restaurants that the two groups of
reviewers visit.
30
Table 1.4: Number of restaurants and percentage of reviews by price level in
Los Angeles
Percentage of reviews in each price level
Price Level Number of restaurants Frequent
reviewers’
reviews
Infrequent
reviewers’
reviews
All reviews
1 98 10.0% 8.0% 9.5%
2 681 71.8% 76.0% 72.8%
3 113 14.2% 13.7% 14.1%
4 29 4.0% 2.3% 3.6%
At higher price levels (see Figure 1.4), both types of reviewers tend to use expectation
words more frequently in their reviews (Expectation wordsfrequent,P1 = .23 vs. Expectation
wordsfrequent,P4 = .38, t = -23.5, p < .001; Expectation wordsinfrequent,P1 = .14 vs. Expectation
wordsinfrequent,P4 = .29, t = -14.3, p < .001). This result supports our hypothesis H4. We also find
that frequent reviewers have a higher level of expectation than infrequent reviewers at all price
levels. Note that we did not hypothesize about the level of expectations of frequent reviewers
vis-à-vis infrequent reviewers at high vs. low price levels. Past literature suggests that novices
are more likely than experts to use price as a signal of quality. However, as we explained in the
theory section, price is not the only factor that creates expectations of quality. Even though
frequent reviewers rely less on price compared to infrequent reviewers, they are likely to use
their rich intrinsic knowledge to form overall expectations of quality that are higher than that of
infrequent reviewers, which could explain why frequent reviewers have higher expectations than
infrequent reviewers at all price levels.
31
Figure 1.4: The average number of expectation and negative disconfirmation words per review
of frequent and infrequent reviewers at different price levels for restaurants in Los Angeles
We also find that (see Figure 1.4) reviewers tend to use negative disconfirmation words
more frequently at higher price levels (Negative disconfirmation wordsfrequent,P1 = .06 vs.
Negative disconfirmation wordsfrequent,P4 = .13, t = -21.6, p < .001; Negative disconfirmation
wordsinfrequent,P1 = .06 vs. Negative disconfirmation wordsinfrequent,P4 = .14, t = -12.5, p < .001).
Importantly, infrequent reviewers use more negative disconfirmation words than frequent
reviewers at high price levels. This difference is significant at price level 3 (Negative
disconfirmation wordsfrequent,P3 = .11 vs. Negative disconfirmation wordsinfrequent,P3 = .13, t =
-5.44, p < .001) and marginally significant at price level 4 (Negative disconfirmation
wordsfrequent,P4 = .13 vs. Negative disconfirmation wordsinfrequent,P4 = .14, t = -1.76, p = .07).
32
Overall, this result supports our hypothesis H5 that infrequent reviewers experience more
negative disconfirmation than frequent reviewers for high price levels.
Figure 1.5 shows that the average ratings by frequent reviewers increase by almost 0.2
stars (AvgRatingfrequent,P1 = 4.02 vs. AvgRatingfrequent,P4 = 4.18, t = -15.4, p < .001) as price
increases from the lowest to highest level. At the same time, infrequent reviewers’ ratings
decrease by almost 0.6 stars (AvgRatinginfrequent,P1 = 4.25 vs. AvgRatinginfrequent,P4 = 3.70, t = 16.4,
p < .001). A test of difference of slopes of the two OLS lines of best fit indicated by the dotted
lines in the plot is statistically significant (Slopeinfrequent = -.2 vs. Slopefrequent = .05, F = 19.2, p
=.01). This result gives preliminary evidence for our hypothesis H6 that infrequent reviewers
assign lower ratings than frequent reviewers for high-priced restaurants.
Figure 1.5: Average star rating by frequent and infrequent reviewers at different price levels for
restaurants in Los Angeles
33
We also examine how the proportion of negative (1- & 2-star) and positive (4- & 5-star)
ratings by frequent and infrequent reviewers vary with price levels (see Figure 1.6). Extant
research shows that experts generally rate more negatively than novices (Schlosser, 2005; Zhang
et al., 2016). However, we find that infrequent reviewers rate restaurants more negatively than
frequent reviewers do at higher price levels. As price increases from level 1 to 4, the proportion
of negative ratings by infrequent reviewers significantly increases from 13.8% to 28.4% (c2=
311.1, p < .001; see Figure 1.6A). At the same time, the proportion of positive ratings
significantly decreases from 80.6% to 64% (c2= 322.9, p < .001; see Figure 1.6B). However, the
trends for frequent reviewers are in the opposite direction. Their proportion of negative ratings
slightly decreases from 9.6% to 8.5% (c2= 14.9, p = .0001; see Figure 1.6A) as price goes from
level 1 to 4, and the proportion of positive ratings significantly increases from 74.7% to 78.6%
(c2= 86.9, p < .001; see Figure 1.6B). These results suggest that frequent reviewers rate
restaurants more positively than infrequent reviewers do at higher price levels. Thus, we obtain
more evidence for our hypothesis H6.
Figure 1.6: The proportion of negative (1- & 2-star) and positive (4- & 5-star) ratings by
frequent and infrequent reviewers at different price levels for restaurants in Los Angeles
(A) (B)
34
1.4.2 Model results
Table 1.5 shows the results of the preliminary regression model in Equation 1 with the
three different dependent variables. In Model 1, where the dependent variable is expectation
words per review, the main effect of price level is positive, and that of infrequent reviewer
(comparison level is frequent reviewer) is negative. The interaction effect is not significant. This
suggests that at higher price levels, both frequent and infrequent reviewers have higher
expectations in line with our Hypothesis H4. In Model 2, where the dependent variable is
negative disconfirmation words per review, the main effect of price level is positive, and that of
infrequent reviewer is negative. Importantly, the interaction effect is significant and positive,
suggesting that infrequent reviewers experience higher negative disconfirmation than frequent
reviewers at high price levels. This result supports our Hypothesis H5. Finally, in Model 3,
where the dependent variable is the rating, the main effects of both price level and infrequent
reviewer are positive. However, the interaction effect is negative, suggesting that infrequent
reviewers assign lower ratings than frequent reviewers at high price levels in line with our
Hypothesis H6. Thus, the preliminary regression model supports all our hypotheses.
35
Table 1.5: OLS regression estimates for testing hypotheses H4, H5, and H6
Dependent variable:
Model 1
Expectation
words per
review
Model 2
Negative
Disconfirmation
words per review
Model 3
Rating
Estimate S.E. Estimate S.E. Estimate S.E.
Main independent variables:
Price Level
Infrequent
Price Level x Infrequent
.054***
–.029***
.005
.002
.007
.003
.021***
–.016***
.011***
.001
.004
.002
.066***
.527***
–.262***
.003
.015
.007
Control variables:
AgeYelp
NonElite
UserReviewCount
TotalReviewCount
Cuisine type dummies
Year fixed effects
Zip code dummies
.000***
–.01***
.000***
.000***
Yes
Yes
Yes
.000
.002
.000
.000
.000*
–.026***
.000***
.000
Yes
Yes
Yes
.000
.001
.000
.000
–.001***
–.066***
–.000***
.000***
Yes
Yes
Yes
.000
.005
.000
.000
*p < .05; ***p < .001
Table 1.6 shows the results from the triple differences (Equation 2). The triple interaction
term gives the triple differences estimate, which is the average treatment effect on the treated
restaurants in Los Angeles. The negative coefficient of the triple interaction term implies that
infrequent reviewers’ ratings for high-priced restaurants in Los Angeles decrease by 0.1 stars due
to the price increase. This result supports Hypothesis H6.
36
Table 1.6: Estimates of the triple differences model
Estimate t-statistic
Main independent variable:
D x Price Level x Infrequent
Other independent variables:
D
Price Level
Infrequent
D x Price Level
D x Infrequent
Price Level x Infrequent
–.076*
.098
–.079
.157***
–.022
.198**
–.129***
–2.42
.09
–.00
4.70
–.88
2.81
–7.59
Control variables:
AgeYelp
NonElite
UserReviewCount
TotalReviewCount
Cuisine type dummies
Restaurant fixed effects
Year fixed effects
Zip code dummies
–.001***
–.083***
–.000***
–.042***
Yes
Yes
Yes
Yes
–19.45
–20.66
–18.61
–.00
*p < .05; **p < .01; ***p < .001
Notes: The dependent variable is the star rating in review i of restaurant j in time t by reviewer r.
t-statistics are clustered at the restaurant level. N= 1,123,689; R
2
within = .004
Table 1.7 shows the estimated average treatment effect on the treated zip codes in Los
Angeles using the generalized synthetic control. We estimated the model for four price-review
frequency conditions on split samples (i.e., price – high vs. low & frequency – frequent vs.
infrequent). As expected, the treatment effect on infrequent reviewers’ ratings for high price
level is negative and largest in magnitude. We compare the coefficients (see Clogg et al. 1995 for
details on comparison of coefficients) for frequent and infrequent reviewers at high price levels
using the z-test. We find a significantly higher treatment effect for infrequent reviewers (z-
statistic = -2.805, p = .005; see Table 1.7). Additionally, the treatment effect for infrequent
37
reviewers is larger at high price levels than at low price levels (z-statistic = -1.707, p = .043). On
average, infrequent reviewers’ ratings drop by .36 stars in the posttreatment period for the high
price level in Los Angeles. Thus, we obtain causal evidence for the negative effect of price
increase on infrequent reviewers’ ratings in Los Angeles in line with our hypothesis H6. Figure
1.7 shows the trend of infrequent reviewers’ ratings for high price levels in the treated and
synthetic control zip codes before and after the price increase in Los Angeles. We see that ratings
across the treated and synthetic control groups are very similar in the pretreatment period but
diverge during the posttreatment period. Specifically, infrequent reviewers’ average rating for
the treated zip codes in Los Angeles decreases relative to the synthetic control in Las Vegas.
Table 1.7: Average treatment effect on ratings for Los Angeles (vs. Las Vegas) zip codes
using the Generalized Synthetic Control
DV: Average rating at the zip code level
Frequent
reviewers
Infrequent
reviewers
z-statistic p-value
High Price –.088 (.042) –.361*** (.087) –2.823 .002
Low Price –.050 (.025) –.195*** (.043) –2.883 .002
z-statistic –.788 –1.707
p-value .215 .043
*p < .05; **p < .01; ***p < .001
Notes: Standard errors are obtained from a placebo test and are bootstrapped for 1000 times. Controlling
for monthly number of reviews, restaurants, & reviewers at the zip code level, and two-way fixed effects.
The treatment effect is evaluated at the mean counterfactual. Z-test performed to compare coefficients in
each row and column.
# Treated observations: frequent, high price = 3229; infrequent, high price = 2676; frequent, low price =
6099; infrequent, low price = 5550
# Control observations: frequent, high price = 1958; infrequent, high price = 1310; frequent, low price =
5616; infrequent, low price = 4742
38
Figure 1.7: Trend of infrequent reviewers’ ratings for high price level in treated and synthetic
control zip codes
We estimate an alternative generalized synthetic control using the difference in ratings
between frequent and infrequent reviewers for high- and low-price levels separately to verify that
the treatment effect is significantly different for them. The estimation results support our finding
that infrequent reviewers give lower ratings than frequent reviewers for high price levels. The
details are in Appendix B.
To assess the economic significance of infrequent reviewers’ low ratings for high priced
restaurants, we compared the average ratings for those restaurants in Los Angeles (N=142) with
and without infrequent reviewers’ ratings. The average ratings for these restaurants drop by 0.07
stars when infrequent reviewers’ ratings are included (AvgRatingfrequent = 4.04 vs.
AvgRatingfrequent & infrequent = 3.97, t = 12.4, p < .001). While the drop in average rating does not
39
seem dramatic, it is meaningful at the individual restaurant level due to Yelp’s practice of
rounding off ratings to the nearest half-star. For example, if a 3.24-star restaurant crosses the
3.25-star threshold, Yelp displays 3.5 stars for the restaurant. When we exclude infrequent
reviewers’ ratings, the displayed rating of 21 restaurants increases by a half-star while that of 5
restaurants decreases by the same amount.
So, a net of 16 restaurants or 11% of high-priced restaurants in our sample lose a half-star
on Yelp when infrequent reviewers’ ratings are included. This drop is critical because Anderson
and Magruder (2012) show that an extra half-star causes restaurants to sell out 49% more
frequently. Also, Luca (2011) shows that a one-star increase in Yelp rating leads to a 5-9 percent
increase in revenue. Thus, infrequent reviewers who penalize high priced restaurants on online
platforms can have a significant negative impact on restaurant revenues and profits.
1.4.3 Robustness checks
This subsection presents robustness of our results using alternative measures of reviewing
frequency. Yelp recognizes actively contributing users as “Elite” reviewers via the Elite Squad
program. Although the exact criteria that Yelp uses to assign the Elite status are not very clear,
Yelp seems to assign it to frequent reviewers who post “well-written” reviews and high-quality
photos
8
. So, we consider Elite reviewers as frequent reviewers and others infrequent reviewers.
However, we note that using Elite status to identify frequent and infrequent reviewers involves a
stricter test of our hypothesis because reviewers must self-nominate every year to be considered
for the Elite status in that year. So, the pool of infrequent (Non-Elite) reviewers is likely to
contain several frequent reviewers who did not nominate themselves to be considered for the
8
https://www.yelp-support.com/article/What-is-Yelps-Elite-Squad?l=en_US
40
“Elite” status. We estimate the generalized synthetic control in Equation 2 on four split samples
of Elite and Non-Elite reviewers for high and low price levels. We find a significant negative
treatment effect of increased prices on Non-Elite reviewers’ ratings for high price levels but not
on Elite reviewers’ ratings, as shown in Table 1.8. The treatment effect is larger for high price
levels than low price levels (z-statistic = -1.845, p = .043).
Another way to identify frequent and infrequent reviewers is to consider the count of
reviews written by reviewers for a cuisine type instead of considering all types of restaurants. For
this, we consider reviews posted in the “American (New)” category of restaurants, which is the
most reviewed cuisine type in our dataset. For each reviewer, we obtain the number of reviews
they have posted in this category. Using a quartile split of the review count, we classify
reviewers in the top quartile as frequent and those in the bottom quartile as infrequent reviewers
in the American (New) cuisine. We then estimate the generalized synthetic control in Equation 2
on four split samples of frequent and infrequent reviewers’ ratings for high- and low-priced
restaurants in the American (New) cuisine. We again find a significant negative treatment effect
of increased prices on infrequent reviewers’ ratings for high price levels, as shown in Table 1.8.
We next check whether our results hold for ratings of restaurants that operated during the
entire period of our analysis. This test will ensure that the change in the mix of restaurants in the
two cities before or after the change in minimum wage does not affect our results. So, we retain
12% of the restaurants (N=748) and 35% of the reviews (N=761,249) from the original sample.
We then estimate the generalized synthetic control on frequent and infrequent reviewers’ ratings
for these high- and low-priced restaurants. In line with our previous results, we find a significant
negative treatment effect of increased prices only on infrequent reviewers’ ratings for high price
levels, as shown in Table 1.9.
41
Table 1.8: Average treatment effect on ratings for Los Angeles (vs. Las Vegas) zip codes for
different measures of reviewing frequency using the Generalized Synthetic Control
DV: Average rating at zip code level
Measure: Elite status Number of reviews posted for
American (New) cuisine
Elite Non-Elite Frequent
reviewers
Infrequent
reviewers
High Price –.069 (.038) –.227*** (.048) –.106 (.055) –.201* (.098)
Low Price –.023 (.022) –.128*** (.024) –.129** (.039) –.172** (.060)
*p < .05;**p < .01; ***p < .001
Notes: Standard errors are obtained from a placebo test and are bootstrapped for 1000 times. Controlling
for monthly number of reviews, restaurants, & reviewers at the zip code level, and two-way fixed effects.
The treatment effect is evaluated at the mean counterfactual.
Table 1.9: Average treatment effect on ratings for Los Angeles (vs. Las Vegas) zip codes for
restaurants operational throughout the analysis period using the Generalized Synthetic Control
DV: Average rating at zip code level
Frequent reviewers Infrequent reviewers
High Price –.035 (.043) –.258* (.113)
Low Price .087*** (.028) –.041 (.058)
*p < .05; **p < .01; ***p < .001
Notes: Standard errors are obtained from a placebo test and are bootstrapped for 1000
times. Controlling for monthly number of reviews, restaurants, & reviewers at the zip code
level, and two-way fixed effects. The treatment effect is evaluated at the mean
counterfactual.
42
1.5 Discussion
This section summarizes the main findings and contributions, addresses some questions,
and draws implications.
1.5.1 Summary
Online ratings have become a widely used proxy for product quality. While prior
literature has mainly focused on the impact of online ratings on sales and stock performance, the
impact of price and reviewing frequency on ratings is less known. This study addresses the
following research questions: How does the price of a product influence its online ratings? Does
reviewing frequency of reviewers moderate the price-rating relationship in online reviews? If
yes, what is its moderating effect on the price-rating relationship? We examine the effect of price
and reviewing frequency on online ratings using 2.2 million reviews of over 5000 restaurants in
Los Angeles and Las Vegas from about 900,000 reviewers. We leverage data from a quasi-
experiment involving differential minimum wage changes in Los Angeles and Las Vegas and use
triple differences and the generalized synthetic control estimator to identify the causal impact of
price on ratings.
The key results of the study are as follows. First, frequent reviewers write reviews that
are significantly more 1) lengthy, 2) useful, 3) multi-dimensional, 4) non-extreme, 5) authentic,
and 6) consistent with restaurant critics than infrequent reviewers do. Second, infrequent
reviewers systematically give lower ratings than frequent reviewers for high-priced restaurants.
We show that this drop in infrequent reviewers’ ratings at high price levels is driven by negative
disconfirmation that they experience because of their high expectations of quality and low ability
to judge quality.
43
This study makes four important contributions to the marketing literature. First, ours is
the only study that finds that the price-rating relationship is positive for frequent reviewers and
negative for infrequent reviewers. Our theory suggests that this asymmetry may be due to
frequent reviewers being better able to assess the quality of premium services than infrequent
reviewers. Second, while the extant literature shows that frequent reviewers are more critical
than infrequent reviewers, we find that the reverse is also true, especially at high price levels.
Third, we contribute to the literature on price-perceived quality by documenting conditions under
which price is negatively related to ratings of quality instead of the conventional positive
relationship. Fourth, very few studies that investigate the effect of price and reviewing frequency
on online ratings have been published in top quantitative marketing journals. Published studies
have documented the effect of either price or reviewing frequency but not both.
1.5.2 Questions
We next discuss some questions that readers may pose.
1. Are infrequent reviewers’ low ratings for high-priced restaurants due to fake reviews?
Yelp uses an algorithm to filter out reviews that it finds as suspicious or fake, and these reviews
were explicitly excluded from our analysis to address this concern. Around 13% of Yelp reviews
of our sample of restaurants are filtered, a number comparable to that in Luca and Zervas (2016),
who treat filtered reviews as fake reviews. Additionally, 75% of infrequent reviewers in our
sample have written at least three reviews on Yelp, suggesting that our sample of infrequent
reviewers’ reviews is not made up of reviewers who write one or two reviews that tend to be
fake.
44
2. Are the frequent reviewers signaling their expertise by giving high ratings to high
priced restaurants like the restaurants’ experts? The frequent reviewers could emulate
restaurants’ expert reviewers to signal that they are knowledgeable like the experts. For example,
they could give high ratings to Michelin-starred restaurants even if the dining experience does
not live up to their expectations. However, this is unlikely in our context because the Michelin
guide was absent from Los Angeles during our analysis period. Michelin guide had exited Los
Angeles in 2010 and then returned to the city in 2019.
3. Is income effect driving infrequent reviewers’ low ratings for high priced restaurants?
One could argue that infrequent reviewers are low-income consumers who slam the ratings of
high priced restaurants because they feel the pinch of paying high prices. If infrequent reviewers
were indeed low-income consumers, their incomes may have increased higher than frequent
reviewers’ due to rising minimum wages in Los Angeles. This positive income effect should
have mitigated the pain of paying to some extent. Thus, the income effect alone may not explain
the effect of rising prices on infrequent reviewers’ ratings. Additionally, we do not see such a
strong negative effect of prices on infrequent reviewers’ ratings for high priced restaurants in Las
Vegas.
1.5.3 Implications
Our findings have several implications for platforms, consumers, firms, and theory.
First, review platforms such as Yelp and Amazon need to aggregate ratings by frequent
vs. infrequent reviewers, in addition to those by verified and non-verified reviewers. Consumers
use average ratings as a simplifying heuristic to learn about the quality of products. However, we
have shown how infrequent reviewers systematically give low ratings to high priced restaurants.
45
A single average rating may mask this heterogenous reviewing behavior and misinform
consumers. Separately displaying average ratings by frequent and infrequent reviewers could
help platforms provide more useful information to consumers of reviews, increasing the
confidence that users and businesses have on the platform.
Second, consumers should consider the type of reviewers who have posted reviews.
Consumers who read reviews may find that their preferences align better with one type of
reviewer (frequent vs. infrequent) than the other. If so, they may benefit from weighting that
reviewer’s rating more than the other’s while making purchase decisions.
Third, businesses listed on review platforms should consider reviewers’ reviewing
frequency while responding to their comments. Management response to online reviews is a
crucial component of online reputation management. Our results suggest that online reputation
management systems should also account for reviewing frequency.
Fourth, our results can be best explained if online ratings are considered as an indicator of
quality and frequent reviewers as experts, as suggested by our theory. Thus, we contribute to
marketing theory by suggesting easy measures for two important constructs in marketing, quality
and consumer expertise, in the context of online reviews.
1.5.4 Limitations and future research
This study has several limitations that provide opportunities for future research.
First, we do not analyze the ratings for chain restaurants because our sample does not
include those data. Several previous studies related to online reviews have examined
heterogeneous effects on chain vs. independent restaurants or hotels (e.g., Hollenbeck 2018,
Luca 2011). Future research could examine whether the effect of price and reviewing frequency
46
on ratings will be different for chain and independent restaurants. However, in line with findings
in the extant literature, we suspect a lower effect on chain restaurants because they deliver a
highly standardized quality which consumers are aware of and will not heighten expectations.
Second, we examined reviews in the restaurant category as it is one of the most
frequently reviewed product categories. Also, restaurant food is an experience good that needs to
be consumed to judge its quality, making it suitable for answering our research questions. Future
research could examine whether the role of price and reviewing frequency varies by product
category.
Third, we examined how price and reviewing frequency interact to affect ratings in our
study. It would be worthwhile to investigate whether other factors like external ratings by
restaurant critics (e.g., Michelin stars) influence frequent and infrequent reviewers’ subsequent
ratings on online platforms differently.
Fourth, our data has limited price variation over time. We use the four price levels of our
sample of restaurants and the minimum wage shock to estimate the impact of prices on ratings.
However, we may have obtained stronger effects of price on ratings with better variation in
price.
As next steps, it would be valuable to build a predictive model using the text-based
attributes to predict reviews by frequent and infrequent reviewers with reasonable accuracy. We
could use a supervised learning model using labeled data to predict in a holdout sample of
reviews whether they have been written by frequent or infrequent reviewers.
47
Chapter 2: New Product Entry for Long Term Success: Waterfall,
Sprinkler, or Niche?
2.1 Introduction
Firms’ choice of market launch strategy is an important consideration for new product
success (Urban and Hauser 1993). The stakes are high during a new product launch as huge
investments are involved in this stage. Previous literature has described two types of new product
launch strategies – sprinkler and waterfall. In the sprinkler strategy, a firm launches its new
product simultaneously in several markets. In the waterfall strategy, a firm launches a new
product in very few markets initially. After achieving success in the initial markets, the firm
launches the product in other markets in a cascading fashion.
Most existing empirical studies on launch strategies have investigated the takeoff and
growth of new product categories (Stremersch and Tellis 2004; Tellis, Stremersch, and Yin
2003), cross-national innovation diffusion (Ganesh and Kumar 1996), and the timeliness of new
product rollout (Chryssochoidis and Wong 1998). Some of these studies suggest that the
sprinkler strategy is preferable while others recommend the waterfall strategy. However, some
important gaps exist in the literature.
First, no empirical study has explicitly considered launch strategy as a driver of new
product success. Previous studies just explore the suitability of the two strategies given the
nature of innovation diffusion of various categories. The only study that treats launch strategy as
a driver of new product success uses an analytical model (Kalish, Mahajan, and Muller 1995).
So, extant studies have not tested the direct role of launch strategy on new product performance.
Second, extant literature uses case studies, descriptive results, analytical models, or single
stage empirical models to discuss the two launch strategies. These methods do not account for
48
the factors that influence a firm’s choice of launch strategy. Therefore, past studies may not be
able to provide strong evidence for the role of launch strategy on new product success and the
mechanism by which the strategies influence new product performance.
Third, most previous studies use short-term sales, market share, or self-reported measures
as the success metric. Only two studies have examined long-term product performance, but they
do not investigate how launch strategies affect new product success (Stremersch and Tellis 2004;
Tellis, Stremersch, and Yin 2003). No previous study has examined the relationship between
launch strategy and long-term survival of new products.
Fourth, past studies have considered global launches of new consumer durables or
technology products only. To our knowledge, no study has examined new product launch
strategies in the consumer-packaged goods (CPG) industry. The innovation diffusion models
used in past studies may not directly apply to CPG products. CPG is a $360 billion per year
market in the US, making up 6% of the US GDP (PWC 2017).
Fifth, almost all launch strategy related studies have examined only a few products in a
limited number of categories. The sole study to employ a large sample merely documents the
strategies employed in various product categories but does not examine product performance. A
study that covers a large variety of product categories and a large sample of products is more
likely to yield robust generalizable results about the role of launch strategy on product success.
We intend to address the above gaps by addressing the following research questions:
What are the distinct strategies of new product launches in CPG markets? Which category and
firm-level factors affect the adoption of alternate strategies? Which launch strategy is associated
with the longest survival and why?
49
We empirically study these questions using rich data provided by IRI (Bronnenberg,
Kruger, and Mela 2008). The data covers 50 markets, 2000 stores, and 31 CPG categories in the
US. This weekly data at the universal product code (UPC) level helps us identify new product
launches at a granular level. We use this data to identify the various launch strategies that firms
use. Specifically, for a launch window of 12 weeks, we consider the number of markets a new
UPC is launched in and the time taken for the launch to identify different strategies. In line with
definitions in extant literature, when products reach markets sooner in the launch window, we
call the strategy sprinkler, and when products take longer, we call the strategy waterfall. We then
test the ten-year survival of 658 new products using a two-stage model. In the first stage, we use
a multinomial logit model of strategy choice as a function of firm-level variables. In the second
stage, we model the hazard of survival as a function of strategy choice and various controls
including marketing mix variables. We believe our analysis of numerous product categories
including food and non-food products will yield generalizable results.
Our study has four key findings. First, we identify a new launch strategy – niche – that is
the most used strategy in our data (44%) but not discussed in the previous literature. Second, in a
majority of product categories (50%), products launched with the waterfall strategy survive the
longest. This result is surprising because numerous studies in the past have argued that the
sprinkler strategy, which gives the widest early penetration and a large scale market entry, is
more successful (e.g., Chryssochoidis and Wong 1998; Libai, Muller, and Peres 2005;
Mascarenhas 1997; Szymanski, Troy, and Bharadwaj 1995). We also find that the waterfall
strategy is more suitable in categories with relatively higher inter-brand competition. This
finding refutes the position taken by Kalish, Mahajan, and Muller (1995) who suggest that the
sprinkler strategy is optimal when competition is high. Third, we find that there is a significant
50
mismatch between the strategies that our study recommends and those that firms actually use.
This could imply that firms use past conventions or expectations to choose a launch strategy that
is different from the ideal strategy. This result is potentially useful for managers to reassess their
choice of strategies. Fourth, our model predicts survival better than a baseline model that does
not specify launch strategy as an explicit variable. Thus, the launch strategy chosen could be a
harbinger of success for a new product. This is very useful for managers because they can predict
success before a product has been launched in the market (Anderson et al. 2015).
Our study makes five important contributions to the literature on new product launch
strategies. First, unlike previous studies that do not empirically model launch strategy as a driver
of new product success, we specifically consider launch strategy as our main independent
variable that influences new product success. Second, previous literature has mostly used case
studies, descriptive results, analytical models, or single stage empirical models to argue for one
strategy vs. the other. We use a two-stage model: (1) Multinomial logit model of strategy choice
as a function of the type of firm and category, and (2) Hazard of survival as a function of strategy
choice plus controls. Thus, we control for the endogenous selection of strategy while testing its
effect on product success. Third, most previous studies use short-term sales, market share, or
self-reported measures as the success metric. We model the long-term survival of a new product
at the store level for ten years using highly discrete data. Fourth, most of the past studies
examine the diffusion of consumer durables (e.g., TV sets, printers, etc.) or industrial technology
products (e.g., point-of-sale scanners). Extant literature has not studied launch strategies for CPG
products. Our study aims to fill this gap. Fifth, almost all other studies have analyzed the success
of very few new products in a handful of product categories. Our study covers 18 product
51
categories and 658 new products, using millions of product-store-week observations to test
survival. Table 2.1 shows the major ways in which our study is different from previous studies.
The remainder of the article is organized as follows. In the next section, we present the
theory. We then describe the method and report the results of our analysis. Finally, we discuss
the findings, implications, and limitations of our study.
2.2 Theory
Past literature has highlighted the pros and cons of both the sprinkler and waterfall
strategies for new product launches. The sprinkler strategy could lead to new product success
due to three main reasons.
First, simultaneous launch in many markets allows a firm to pre-empt its competitors
(Kalish, Mahajan, and Muller 1995). In highly competitive industries, a firm’s entry into many
markets could raise entry barriers to its competitors, who will then have to play catch up in those
markets. Particularly in the CPG industry, simultaneous entry into multiple markets helps a firm
secure shelf-space in numerous stores. According to the resource-based view of the firm,
resources that are valuable, rare, imperfectly imitable, and non-substitutable could offer firms a
competitive advantage (Barney 1991). Scarce retail shelf-space is a valuable and not easily
substitutable resource that is crucial for new product success (Kaufman, Jayachandran, and Rose
2006). By securing shelf-space in multiple markets via a sprinkler strategy, a firm is able to erect
a resource position barrier that competitors find difficult to overcome (Wernerfelt 1984). This
barrier to competition could increase the likelihood of new product success.
52
53
Second, a sprinkler strategy allows a firm to exploit economies of scale and experience in
R&D and manufacturing (Stremersch and Tellis 2004). By introducing a new product to as many
markets as possible, a firm can maximize revenues and spread costs over a large base of sales.
When a firm enjoys significant cost advantages due to scale economies, its competitors may have
to retaliate on a large scale to negate the advantage (Harrigan 1981). But competitors may not
immediately respond because they need to invest heavily (Porter 2008). Thus, a delay in
competitive threat could help the new products succeed.
Third, a sequential launch into many markets is likely to involve rollout delays.
Chryssochoidis and Wong (1998) find in their study that most of the sequential rollouts were
delayed while the simultaneous launches were mostly timely. New product delays have a
significant negative impact on product success, firm market value (Hendricks and Singhal 1997),
and firm profitability (Hendricks and Singhal 2008). Delays could also offer competitors a
chance to attack the latter markets, negating the potential advantages of entry barriers and scale
economies. In some extreme situations, particularly in the computer industry, not launching a
product previously announced could invite antitrust lawsuits (Bayus, Jain, and Rao 2001). A
simultaneous launch can avoid such pitfalls.
However, a sprinkler strategy involves risks as well. First, entering several markets
simultaneously requires a huge investment of resources (Dekimpe, Parker, and Sarvary 2000).
Apart from capital investments for manufacturing, significant marketing investments like ad
buys, product samples, distributor incentives, slotting allowances, and consumer promotions
need to be committed upfront across many markets. For example, in 1995, P&G launched its
Crest Gum Care toothpaste nationally in the US, backed by a $40 million budget (Sloan 1995).
But the product did not do well because consumers complained of a poor taste in the mouth. It
54
was declared a failure and eventually pulled from the market. When market entry costs are very
high, a waterfall strategy is more profitable (Libai, Muller, and Peres 2005). Second, minor but
early product or marketing errors could be fatal, because the entrant lacks the time to correct
them and the product could earn a bad reputation. For example, in August 2016, Samsung
launched its new mobile device Galaxy Note 7 in ten countries simultaneously. Within weeks,
the phones started exploding due to faulty batteries. In September 2016, Samsung initiated its
biggest ever recall of 2.5 million devices that cost them $6.5 billion (Jeong 2016). Ultimately,
Samsung discontinued the product and launched a new version a year later.
On the other hand, the waterfall strategy could lead to new product success due to several
reasons.
First, revenues and profits from initial markets can be reinvested to support the product
launch in latter markets. This relieves the pressure on firms’ cash flows and offers a low-risk
option of launching new products (Tellis, Stremersch, and Yin 2003). In the motion picture
industry, a new movie’s box office revenues in the initial weeks and markets are used to expand
to more screens and international markets later (Eliashberg, Elberse, and Leenders 2006). Such
sequential movie releases help studios take advantage of the success-breeds-success phenomenon
and succeed in the movie launch (Elberse and Eliashberg 2003). Netflix’s streaming services
were available only in the US in 2010. First, it expanded to Canada, then to Latin America and
Central America in 2011, followed by a few European countries during 2012-13, and eventually
reached about 50 countries by 2015. This sequential expansion allowed Netflix to invest
gradually in content deals and in the production of original local content (Brennan 2018).
Second, in the waterfall strategy, firms get the option to learn from any mistakes and
fine-tune their marketing strategy. Organizational learning theory suggests that firms learn from
55
their direct experience of operating in an environment. They capture these experiences through
routines that are recorded in the organizational memory to be retrieved and used later when
required (Levitt and March 1988). Routines that contribute to success are used more frequently
than those associated with failure. The outcome of a new product launch is uncertain at the outset
and firms would like to minimize the risk of failure. When firms launch in few initial markets via
a waterfall strategy, they tend to acquire market knowledge from their experience with retailers,
suppliers, and consumers in those early markets (Chang 1995; Johanson and Vahlne 1977).
Based on this knowledge, firms could change their product strategy or marketing mix elements
to improve new product performance in the latter markets. For example, P&G launched the odor
eliminating product Febreze in 1996 in three test markets: Phoenix, Salt Lake City, and Boise.
Initial retail sales were dismal, and the company was on the verge of pulling the product. But,
consumer feedback from these markets helped P&G reposition the product and it eventually
became a billion-dollar brand (Duhigg 2012).
Third, the later a product is launched in a market, the faster its rate of adoption will be
(Takada and Jain 1991). According to the diffusion of innovations theory (Rogers 2003), four
perceived characteristics of a new product affect its rate of adoption: 1) relative advantage 2)
compatibility with consumer needs 3) observability, and 4) trialability. In a waterfall strategy,
consumers in latter markets get the opportunity to learn about the product benefits and whether
the product serves their needs (Ganesh and Kumar 1996). They will also be able to observe the
product via mass communication media or mixing with friends and relatives in markets where it
has been launched first (Putsis et al. 1997). These factors could increase the likelihood of
adoption and long-term success.
56
Fourth, a waterfall strategy requires a lower level of investment than a sprinkler strategy.
Manufacturing, marketing, and selling expenses can be kept relatively low. By focusing on a few
initial markets, firms do not need to spread their investments too thin. Also, if the new product
fails in the initial markets, then further investments can be avoided (Stremersch and Tellis 2004).
However, a firm’s key risk in employing the waterfall strategy is its competitors’ opportunity to
react quickly to the new product launch. Riesenbeck and Freeling (1991) advocate avoiding the
conservative step-by-step approach in the waterfall strategy because a rapid competitive response
is likely. The competitor could beat the focal firm to the latter potential markets and take away
substantial market share. This risk is significant especially in the case of vertically differentiated
products. When a manufacturer of a vertically differentiated new product chooses a waterfall
strategy, it gives a potential imitator a time window to gather more information about the
product. The imitator could then launch a product that is superior in quality to beat the original
new product because consumers are likely to switch to the imitation product due to discernible
quality differences (Ethiraj and Zhu 2008). Such copying was confirmed in the verdict on a
lawsuit that Apple filed in 2011 against Samsung for copying Apple’s mobile phones, violating
Apple patents covering the iPhone’s design (Reuters 2018). Bohlmann, Golder, and Mitra (2002)
find that late entrants perform better than pioneers in product categories where vertical
differentiation is more important than horizontal differentiation. This is because the late entrants
are able to use better technology at a cheaper cost to offer higher quality products at lower prices
to consumers. However, in CPG categories, tastes may be more important than objective quality
attributes. Tastes could also vary significantly across consumer segments and geographical
regions. So, it is not clear whether competitors can outpace a firm that employs the waterfall
strategy in the CPG industry.
57
The preceding discussion shows that past literature does not resolve the issue of which of
the two strategies is suitable for new product success. Additionally, the mechanisms that work
for or against the two strategies have been discussed in the context of international launches of
consumer durables only and have not been tested in an empirical setting. Thus, ours is the first
study to empirically test the suitability of strategies for within-country launches of new CPG
products and to explore the mechanisms by which the strategies influence new product success.
2.3 Data and Method
2.3.1 Data
We use the IRI Academic dataset for our study (Bronnenberg, Kruger, and Mela 2008).
The dataset contains UPC and store-level weekly sales, price, promotion, feature, and display
data in 31 product categories from over 2000 stores in 50 geographical regions from 2001 to
2012. Figure C.1 in Appendix C shows the 50 markets on a map along with their names.
For the current study, we consider new products launched in 2002 in 18 categories and trace their
survival till 2012. We chose the categories in which the greatest number of new products were
introduced. The dataset also has information about the product attributes of each UPC, which we
use to identify the new products as explained in the next subsection.
2.3.2 Definition of new products
The first step is to identify new products from the dataset. A new UPC is one that appears
for the first time in the dataset in a year and is not present in previous years. We identify all
UPCs that occur in the data set for the first time in 2002 and not present in 2001 in any of the
markets. From this list of new UPCs, we remove all promotional and temporary UPCs. We then
58
use the product attributes of the remaining UPCs to arrive at the final list of new products in each
category. Each category has ten product characteristics like product type, type of package, pack
size, flavor, brand name, etc. A UPC is classified as a new product only if at least one of these
characteristics is new to the market compared to previous years.
2.3.3 Definition of launch strategy
We use the geographical patterns of market launches for new products during the first 12
weeks of their launch (referred to as launch window from hereon) to identify the launch strategy
for that product. The first sales transaction of a new product in any store in a market indicates its
launch in that market. We use two factors to define the launch strategy of a new product: (1) the
number of markets the product enters during the launch window and (2) the time taken to enter
those markets.
We identify three distinct launch strategies as illustrated in Figure 2.1. We calculate the
median time that products take to reach the number of markets in the launch window. If a
product takes less than the median amount of time, then its launch strategy is sprinkler. But, if it
takes the median amount of time or more, then its launch strategy is waterfall. Besides these two
previously known strategies, we uncover a new strategy that we call niche. In a niche strategy, a
new product reaches less than 20% of the total available markets during the launch window. So,
we classify all products that reached less than or equal to 10 markets during the launch window
under the niche strategy. In the data, across the 18 product categories, 28% of new products are
launched via the sprinkler strategy, 28% via the waterfall strategy, and 44% using the niche
strategy. Thus, the new strategy that we identify in this study is the most widely used launch
strategy in CPG markets.
59
Figure 2.1: Examples of the three launch strategies
Sprinkler Waterfall
Niche
2.3.4 Threats to identification
There are three potential issues that could arise in the identification of new products and
the definition of launch strategies. We discuss these here.
First, the focal outcome variable is the survival of new products. Since our unit of
analysis is a UPC, newly launched products may include various promotional and temporary
UPCs that are introduced with the express purpose of being on the market for only a relatively
short time. All such products, which included percent free packs, buy-one-get-one-free promo
packs, trial size packs, bonus packs, and cents off packs, were identified and removed from the
analysis sample.
Second, the definition of launch strategy is based on the launch window of 12 weeks. We
chose the window of 12 weeks because it allowed us to identify three distinct geographical
60
patterns of launches. We avoided longer windows to allow for enough time to analyze the
survival of new products. A similar window has been used as the initialization period in past
studies that investigate new product failure (Anderson et al. 2015).
Third, firms could choose a particular strategy based on their assessment of their own
strengths and weakness and the anticipated performance of the chosen strategy. For example,
firms that have a high market share or a large distribution footprint in a product category may be
more likely to choose a sprinkler strategy to maximize revenues by achieving the widest early
penetration. Thus, strategy choice could be endogenous to survival. We address this endogeneity
issue by incorporating a two-stage model in our empirical analysis. The first stage explicitly
models firms’ choice of launch strategy using firm- and brand-specific variables as explained in
the subsection below. Next, following the approach taken by Akerman, Gaarder, and Mogstad
(2015) and Zervas, Proserpio, and Byers (2017), we include the predictors of the choice of
strategy in the first stage model as control variables in our second stage model that estimate the
influence of launch strategy on the hazard of failure. Thus, we control for major factors that
influence the choice of strategy. In an alternative approach, we use a Heckman two-step
estimation (Heckman, 1979) to account for strategy selection bias. As the substantive results do
not change much, we report the results of that approach in Appendix D.
2.3.5 Empirical model
Our approach incorporates two sub-models, one for the firm’s choice of strategy for the
introduced product and the second for the time to failure. The probability that a firm chooses a
particular launch strategy is specified as a function of several firm-level and brand specific
61
variables using an unordered multinomial logit model. The probability that company j chooses
strategy m to launch its new product i for brand b is given by:
Pr(Yijb = m | wijb) = Pijbm = exp(dmwijb) / (1 + å
2
n=1 exp(dnwijb)) (1)
where Yijb refers to the chosen launch strategy, wijb refers to the set of firm and brand specific
variables that influence the choice of strategy, and d refers to the coefficients to be estimated.
The dependent variable takes on one of three values – sprinkler, waterfall, or niche. For
parameter identification, the sprinkler strategy is selected as the reference by setting its
coefficients to zero (Greene 2012). n = 1 and n = 2 refer to the waterfall and niche strategy
respectively. Thus, model coefficients can be interpreted as the extent to which the
corresponding variables contribute to the firm choosing waterfall or niche strategy relative to the
sprinkler strategy.
The wijb vector includes the following variables (a) CompanyShare – market share of
company j, (b) NumMarkets – number of markets that company j is present, (c) NumChains –
number of retailer chains that company j is present, (d) NumStores – number of stores that
company j is present, (e) NumBrands – number of brands that company j has, and (f)
BrandStrength – number of categories in which brand b is present, all of which are product-
category specific. For example, CompanyShare refers to the market share of company j in a
particular product category and NumMarkets refers to the number of markets in which the
company is present in that category. Also, all these variables are static, i.e., based on data
preceding the launch of product i. Specifically, we measure these variables at the beginning of
2002, the year when our sample of new products is launched.
62
Equation 1 explicitly captures the role of a parent company’s existing market presence
and brand strength in its choice of launch strategy, thus accounting for the firm’s expectations of
the success of the product. For instance, a firm present in many markets covering multiple chains
and stores may believe that choosing a sprinkler strategy will lead to success because the firm
can leverage its existing national distribution footprint. Similarly, a large firm with strong brands
may choose the sprinkler or waterfall strategy over the niche strategy relative to a small firm
because it expects new products of a strong brand to perform well (Reddy, Holak, and Bhat
1994; Sullivan 1992). On the other hand, a small firm may not have the resources required for a
national launch using a sprinkler strategy. Hence, it may choose a niche strategy. Thus, our first
stage model captures the role of firms’ expectations of success in their choice of launch strategy.
In the second stage, we model new product failure using a discrete-time hazard model
(Allison 1995). This accounts for time-varying explanatory variables and for censoring, i.e.
products that have not failed during the observation period (Chandrasekaran et al. 2013; Jindal
and McAlister 2015). The focal event, product failure in a store, occurs in a week when the
product has zero sales in that store. When we identify the week of zero sales, we ensure that
there are no sales of the product in subsequent weeks in that store. This way we avoid
erroneously identifying a stock out or temporary no purchase situation as a product failure. We
model product failure using the following hazard specification:
hijbsct = Pr(Tijbsc = t | Tijbsc ³ t, xijbsct) (2)
where hijbsct is the probability that new product i of brand b of company j fails in store s of chain
c during week t, given that failure has not yet occurred. Tijbsc is the week when the new product
63
fails and xijbsct is a vector of explanatory variables observed for product i in week t. We trace the
survival of a new product in each store in each week since launch. The dependent variable takes
value 1 in a week if the product fails and 0 otherwise. We specify a time-varying logit for our
hazard model as:
log (hibjsct/(1- hibjsct)) = h0ijsct + b1LaunchStrategyibj
+ b2RelativePriceibjsct +b3RelativeFeatureibjsct + b4RelativeDisplayibjsct
+ b5RelativePromoibjsct + b6Relative Salesibjsct + b7MarketsEntryibjt
+ b8StoresEntryibjt + b9CompanyShareij + b10NumMarketsij
+ b11NumChainsij + b12NumStoresij + b13NumBrandsij
+ b14BrandStrengthibj (3)
The focal predictor of our study is the launch strategy dummy for product i of brand b of
company j (LaunchStrategyibj). The other predictors include the baseline hazard function (h0ijsct),
time-varying marketing mix variables relative to other products in the category (RelativePriceijsct,
RelativeFeatureijsct, RelativeDisplayijsct, and RelativePromoijsct), time-varying sales relative to
category sales (RelativeSalesijsct), time-varying product-level variables that capture the number
of markets and stores the product has entered in each week (MarketsEntryijt and StoresEntryijt),
and all the strategy predictors that were part of the first stage model in Equation 1. We specify
the baseline hazard as follows:
h0ijsct = as + ac + g1Timeijsct + g2Timeijsct
2
(4)
64
where as is a store intercept term, ac is a chain intercept term, and Time is the time elapsed in
weeks since firm j launched product i in store s of chain c. We use the R packages nnet and
alpaca (Stammann, 2018; Venables & Ripley, 2002) to estimate Equations 1 and 3. Since we
estimate the equations separately for each of the 18 product categories, we do not include the
category subscript in our equations for simplicity. Table 2.2 describes all the variables used in
Equations 1 and 3.
2.4 Results
2.4.1 First stage model
Table 2.3 presents the coefficient estimates of the multinomial logit model in Equation 1.
Recall that Equation 1 models the probability of a company choosing one of the three strategies
for launching its new product. With sprinkler as our reference strategy, we have two sets of
coefficients for our independent variables. The first set corresponds to the waterfall strategy and
the second set to the niche strategy. A positive coefficient suggests that the variable increases the
probability of choosing the waterfall or niche strategy over the sprinkler strategy.
As previously stated, we estimated Equation 1 separately in each of the 18 product
categories. In most categories, we see that McFadden’s pseudo R
2
values are fairly high
indicating a good model fit (McFadden, 1974). In all but one category (household cleaners), the
model generates statistically significant coefficients for most of the independent variables. We
see that the relationship between the independent variables and the probability of strategy choice
varies across categories. Based on the relative frequency of a positive or negative relationship
between the independent variables and the dependent variable, we make the following
observations: all else being equal 1) firms with higher market share are likely to choose waterfall
65
over niche and sprinkler, 2) firms present in a larger number of markets and stores are likely to
choose waterfall or sprinkler over niche, 3) firms present in a larger number of chains and with a
larger number of brands are likely to choose niche over waterfall and sprinkler, and 4) stronger
brands are likely to choose waterfall or niche over sprinkler.
The relationships between strategy choice and the various predictors are mostly in line
with our expectations. For instance, we expected firms present in a larger number of markets and
stores to choose waterfall or sprinkler over niche. This would be consistent with a firm’s strategy
to leverage its existing distribution footprint to ensure a wide reach for its new products.
Regarding the number of brands, firms with several brands in the same product category may
prefer a niche strategy over a sprinkler or waterfall strategy to prevent the risk of cannibalization.
The only unexpected finding is that stronger brands prefer niche over sprinkler. This underscores
the importance of further investigating the niche strategy, which has not been analyzed in past
literature.
66
67
68
69
70
71
2.4.2 Hazard model
The second-stage hazard model examines which one of the three strategies is associated
with the longest survival of new products. Note that our hazard model has a logit specification
that models the log odds of failure of a new product in a store of a chain in a given week. Thus,
positive coefficients increase the probability of failure and negative coefficients decrease the
probability of failure. Also, recall that sprinkler is our reference category. So, coefficients for the
strategy variables have to be interpreted relative to the sprinkler strategy.
Table 2.4 contains the raw and standardized parameter estimates for Equation 3 for 18
product categories. Standardized estimates are useful to compare the relative effects of
independent variables when the variables are measured in different units. The R package that we
use for estimation does not return standardized estimates. So, we manually computed the
standardized estimate bk
*
of an independent variable k from its raw estimate bk using the
following formula (Allison, 1999; Hilbe, 2009):
bk
*
= (bk)(sk) / [p/Ö3] (5)
where sk is the standard deviation of the independent variable k.
72
73
74
75
76
77
78
From the estimates, we see that in half the categories, products launched with a waterfall
strategy survive the longest. In one-third of the categories, products launched via the niche
strategy survive the longest, while products launched with a sprinkler strategy survive the longest
in only one-sixth of the categories. Thus, contrary to past literature that recommends the
sprinkler strategy, we find that the waterfall strategy is appropriate for new product survival in a
majority of categories. We examine the effect of the waterfall and niche strategies on product
failure relative to the sprinkler strategy by exponentiating the respective coefficients of the
strategy variables in Table 2.4. For instance, in the beer category, the odds of new product failure
are 45% lower (odds ratio = exp (coefficient of waterfall strategy from Table 2.4) = exp (-.61) =
.55) for waterfall strategy compared to sprinkler strategy, holding all other variables constant.
Similarly, the odds of new product failure are 31% lower (odds ratio = exp (coefficient of niche
strategy from Table 2.4) = exp (-.36) = .69) for niche strategy compared to sprinkler strategy.
The relative effect of the three strategies on product failure can also be examined by
plotting the baseline hazard given in Equation 4. Figure 2.2 displays the baseline hazard for the
strategies for selected categories. The X-axis shows the number of weeks since product launch
and the Y-axis shows the baseline hazard. These plots help visually clarify the key result
obtained from our hazard model estimation. For instance, we can clearly see that the hazard of
failure is significantly lower for the waterfall strategy than that for the other two strategies in the
case of household cleaners, beer, and shampoo. Two additional observations can be made from
these plots. First, for several categories, the hazard of failure initially increases to a maximum
and then declines over time. Second, the time taken to reach maximum hazard is different for
different product categories. For example, the hazard peaks in about 1000 weeks for cold cereals,
750 weeks for milk, and between 250 and 400 weeks for various other categories.
79
Figure 2.2: Baseline hazard of failure of the three strategies for selected categories
80
The standardized estimates in Table 2.4 can be interpreted as follows: a one standard
deviation increase in independent variable k is associated with bk standard deviation change in
the logit of failure. We can rank the independent variables in terms of their effect on product
failure using the standardized estimates. For example, in the beer category, an increase of one
standard deviation of RelativeSales is associated with a decrease of .46 standard deviations in the
logit of failure. Similarly, a one standard deviation increase in RelativePrice is associated with an
increase of .12 standard deviations in the logit of failure. Thus, RelativeSales has a larger effect
than RelativePrice on product failure in the beer category. When we compare the relative effects
of all the variables across all categories, we find that variables related to existing market
presence like NumMarkets, NumStores, and NumChains have the highest effect on product
failure in the maximum number of categories. We also find that the launch strategy variables are
among the top five variables associated with product failure in 10 out of 18 categories. This is
important because launch strategy has not been tested as a driver of success hitherto in the
literature. Our findings suggest that it indeed plays a more important role than other variables
like product sales, price, and promotion that have been frequently linked to new product success.
While we primarily focus on the launch strategy variables, we also note that most of the
marketing mix variables in Table 2.4 have the expected sign for their coefficients. The
relationships between these variables and the probability of failure for selected categories are in
Figure 2.3. The X-axis shows the values of these variables and the Y-axis shows the predicted
probability of failure of product using Equation 3. As expected, when RelativeSales,
RelativeFeature, and RelativeDisplay increases, the probability of failure decreases.
RelativePromo, which is a measure of the extent of price promotion of the new product, is
negatively associated with the probability of failure. This is consistent with past literature that
81
Figure 2.3: Predicted probability of new product failure as marketing mix variables in the
second stage model increase for selected categories
82
documents the negative impact of price promotions on long-term performance (Pauwels et al.,
2004). The effect of RelativePrice is also negative in various categories, though it is weakly
positive in some categories.
2.4.3 Do companies use the strategy with the lowest hazard of failure?
We examined whether firms mostly use the strategy that our model recommends in the
respective product categories. The strategy with the lowest hazard of failure in our second stage
model is the recommended one for long term survival. Recall that waterfall is the recommended
strategy in most categories followed by niche and sprinkler. Table 2.5 shows the percentage of
products using the three strategies when each of the three strategies is recommended according to
our results.
Table 2.5: Recommended vs. actual strategies used – percentage of products
Recommended
Actual
Waterfall Sprinkler Niche
Waterfall 32% 32% 22%
Sprinkler 29% 39% 18%
Niche 39% 30% 60%
We see that there is a significant mismatch between the recommended strategy and the
actual strategies that firms use. For example, in categories where the waterfall strategy is
recommended, only 32% of products use it whereas 39% and 29% use niche and sprinkler
respectively. Similarly, in categories where the sprinkler strategy is recommended, only 39% of
83
products use it while 32% and 30% use waterfall and niche respectively. However, in categories
where the niche strategy is recommended, 60% of products use it while only 22% and 18%
deviate to employ waterfall and sprinkler respectively. In summary, we find that majority of
products in many categories do not use sprinkler or waterfall when those are the recommended
strategies. This analysis implies that many firms do not know beforehand which strategy is
suitable for long term success. So, the endogeneity arising from firms’ expectations of success
driving the choice of strategy may not be a serious concern in our context.
2.4.4 Factors associated with recommended strategies
We examined two category specific factors that are potentially related to the
recommended strategy – brand-level Herfindahl index and average price per SKU. Herfindahl
index is a measure of competitive intensity in a category. A lower value of the index indicates
numerous brands competing in a category whereas a higher value indicates a high concentration
of players in the category and lower competition. Herfindahl index is calculated by taking the
sum of squared market shares (unit sales) of all brands in the category. We then plotted the
recommended strategy in each category in a price vs. Herfindahl graph as shown in Figure 2.4.
We see that the sprinkler strategy is suitable when the average price per SKU and level of
competition are relatively low (relatively high values of Herfindahl Index) as in the case of milk,
yogurt, ketchup, etc. On the other hand, waterfall and niche strategies are suitable mostly when
the average price per SKU and level of competition are relatively high as in the case of diapers,
beer, coffee, frozen pizza, etc. This finding is contrary to the result in Kalish, Mahajan, and
Muller (1995) that suggests the sprinkler strategy is optimal when the level of competition is
high.
84
Figure 2.4: Factors related to recommended strategies
2.4.5 Predictive analysis
To demonstrate the benefit of considering launch strategy as a driver of new product
success, we compare the out-of-sample predictive validity of two competing models. We
compare our proposed model in Equation 3 with a similar model that excludes the launch
strategy variables. Our competing model is a suitable baseline model because it is similar to
several new product forecasting models which assume that higher product sales signal a higher
likelihood of long-term success. Additionally, such a model has been used as the baseline in a
recent study that examines harbingers of failure of new products (E. Anderson et al., 2015).
85
We split our sample of new products into training and test sets using a 70%-30% split.
We calibrate the two models on the training set and then predict new product failure in the test
set at the product-store-week level. We then consider three metrics that are commonly used in
the machine learning literature to assess predictive performance (Hastie et al., 2013) –
sensitivity, specificity, and area under the receiver operating characteristic (ROC) curve (area
under the curve, AUC). Sensitivity is a model’s ability to detect true positives (failure in a store
in a week in our scenario) while specificity is the ability to detect true negatives. These are
calculated using the following formulae:
Sensitivity =
%&'( *+,-.-/(,
01.'23 *+,-.-/(,
=
%&'( *+,-.-/(,
%&'( *+,-.-/(,4523,( 6(72.-/(,
(6)
Specificity =
%&'( 6(72.-/(,
01.'23 6(72.-/(,
=
%&'( 6(72.-/(,
%&'( 6(72.-/(,4523,( *+,-.-/(,
(7)
In our scenario, we have many more instances of 0s than 1s (failure in a store in a week).
So, we have to choose a cut-off value other than .5 to predict failure using our model. We
perform a grid search to find the cut-off value that maximizes sensitivity and specificity,
balancing the trade-off between the two. The ROC is a plot of true positive rate vs. false positive
rate (one minus specificity). An AUC of .5 for a model indicates that it performs no better than
chance in predicting the outcomes. In our scenario, a higher AUC for the proposed model than
for the baseline model justifies the importance of considering launch strategy as a predictor of
success.
86
87
Table 2.6 reports the three predictive performance metrics for the proposed and baseline
models for 18 categories. The model coefficients are not reported in the table for easy
comparison of the metrics. First, we find that the sensitivity, specificity, and AUC values for the
proposed model are significantly higher than what could have been predicted by chance alone.
The values for sensitivity range from 66.2% to 81.2%, for specificity from 67.8% to 82.5%, and
for AUC from .735 to .892 across various categories. Second, the proposed model outperforms
the baseline model in sensitivity and specificity in 15 and 17 categories respectively. Also, the
proposed model has better AUC values than the baseline model in 17 categories. Thus, the
predictive analysis confirms the importance of considering launch strategy as a key explanatory
variable when examining new product failure.
2.5 Discussion
This section summarizes the main findings and contributions, addresses some questions,
and draws implications.
2.5.1 Summary
The choice of launch strategy is critical for the long-term survival of new product
management decisions. Prior research has not addressed the types and long-term success of new
product launch strategies in the $360 billion CPG market. This study addresses the following
questions: What new product launch strategies do firms adopt in CPG markets? What factors
affect the choice of strategies? What factors affect the long-term survival of these strategies? We
test the ten-year survival of over 650 new products launched in 18 categories in the US using a
two-stage model: 1) Multinomial logit model of strategy choice as a function of the type of firm
and category and 2) Hazard of survival as a function of strategy choice plus controls.
88
The key results of the study are as follows. First, firms widely use a strategy that has not
been analyzed in the literature – niche – besides previously reported waterfall and sprinkler
strategies. Second, the niche strategy is most frequently used in CPG markets with 44% of
products in our sample using it. However, the waterfall strategy has the longest survival in the
majority of the categories. Third, a substantial mismatch exists between firms’ current choice of
strategies and best strategies for long-term survival. Fourth, our hazard model’s predicted long-
term survival is superior to that of baseline models in out-of-sample tests.
This study makes five important contributions to the literature. First, unlike previous
studies that do not empirically model launch strategy as a driver of new product success, ours is
the first study to consider launch strategy as the main independent variable that influences new
product success. Second, previous literature has mostly used case studies, descriptive results,
analytical models, or single stage empirical models to argue for one strategy vs. the other. We
use a two-stage model: (1) Multinomial logit model of strategy choice as a function of the type
of firm and category, and (2) Hazard of survival as a function of strategy choice plus controls.
Thus, we control for past practices or expectations of firms while testing the role of strategy on
product success. Third, most previous studies use short-term sales, market share, or self-reported
measures as the success metric. We model the long-term survival of a new product at the store
level for ten years using highly discrete data. Fourth, most of the past studies examine the
diffusion of consumer durables (e.g., TV sets, printers, freezers, etc.) or industrial technology
products (e.g., point-of-sale scanners). Our study is the first to focus on launch strategy for CPG
products. Fifth, almost all other studies have analyzed the success of very few new products in a
handful of product categories. Our study covers 18 product categories and 658 new products,
using millions of product-store-week observations to test survival.
89
2.5.2 Questions
Are our results biased because firms choose strategies that they believe will lead to new
product success? We use a two-stage model to address this issue. In the first stage, we model the
choice of strategy using firm and brand specific variables like market share and existing market
presence. Firms with a high market share in a category may expect their products to do well and
choose sprinkler or waterfall over niche so that they can capitalize on higher potential sales from
numerous markets. Likewise, firms that have a wide distribution footprint across many markets
and retail stores may prefer to leverage the distribution strength using a sprinkler or waterfall
strategy. These variables have fairly high explanatory power in our first stage model in most
product categories. By controlling for these variables in our second stage hazard model, we are
able to identify accurately the effect of launch strategy on new product survival.
Is niche really a different strategy or is it just a failed waterfall? To check this, we
compare the average number of markets new products launched using niche strategy enter during
the launch window with the maximum markets they enter in their lifetime. Figure 2.5 depicts this
comparison. Except in a few product categories like milk, hotdog, and toilet tissue, we find that
the products enter many more markets during their lifetime subsequent to launch. If niche was a
failed waterfall, products launched using niche are unlikely to scale up their reach significantly
after the launch window. So, we argue that niche is definitely a distinct launch strategy.
2.5.3 Implications
Our results have several implications. First, managers are probably better off choosing
launch strategies that ensure the long-term survival of new products. Our study suggests that the
appropriate strategies for long-term survival differ by product categories. Notably, we find that
90
firms deviate significantly from the appropriate strategies. Our study also offers guidelines to
choose the appropriate strategy. We find that sprinkler strategy is suitable when competition and
price level is low in a category whereas waterfall or niche is suitable when competition and price
level is high. Second, our results show that in many product categories the baseline hazard of
failure increases initially and then decreases over time. This implies that managers should
probably not pull the plug early for new products. If they persist with marketing activities like
feature and display that reduce the hazard of failure, they may potentially benefit from continued
sales over a longer period of time. Third, managers could possibly benefit from the predictive
Figure 2.5: Products launched using the niche strategy expand well beyond the initial set of
markets
91
ability of our model. Our model has a true positive rate of about 62% to 83% and a true negative
rate of about 60% to 81% depending on the category. This performance is better than that of a
baseline model that utilizes sales and other marketing variables by about 5% to 10%.
2.5.4 Limitations and future steps
This study has several limitations that provide opportunities for future research. First, we
do not consider the nature of innovation (major vs. minor) of the new product. It is very useful to
understand whether the appropriate strategies depend on the nature of product innovation.
Second, we do not consider the effect of advertising on new product survival as data is
unavailable. Third, our study may be subject to the Lucas critique according to which our
recommended strategies in the future may no longer be the same if firms start using those widely
(Lucas, 1976). There are three main requirements for Lucas critique to be relevant in our
situation: 1) firms are aware of the appropriate strategies for survival, 2) firms are motivated to
implement the recommended strategies, and 3) firms are able to implement the recommended
strategies (Van Heerde et al., 2005). On the first point, we found that firms mostly do not use the
recommended strategies. Our analysis in an earlier subsection suggests that firms do not seem to
know which strategies lead to long term survival. Adoption of the recommended strategies is
contingent on firms learning from their experience or on the dissemination of our research
findings, both of which may take time. As for the second point, firms definitely have the
incentive to adopt the recommended strategies. But related to the third point, not all firms may be
able to immediately change their strategies due to resource constraints. For example, smaller
firms may not be able to use the sprinkler or waterfall strategies as significant investments are
required. So, at least in the short to medium term, our findings may not be subject to the Lucas
92
critique. However, future research could examine whether firms change their use of strategies
over time by considering new product launches during a larger time span.
Two next steps can further improve this research. First, we could analyze an independent
sample of new products to verify that the current results hold. Presently, we identify new
products using various product attributes available in the data. We have recently obtained a
database of new products maintained by a market research company. This data also has
innovation ratings which we think are valuable for our research because the type of innovation
could also affect the choice of launch strategy. Second, we need to analyze more recent product
launches to verify whether our results hold in the current market scenario. Currently, we have
examined the survival of new products launched in 2002, and the market dynamics could have
changed over the years. So, we have obtained access to the Nielsen data from the Kilts Center at
the Chicago Booth School of Business which has sales data till 2019.
93
Bibliography
Aaker, D. A., & Jacobson, R. (1994). The financial information content of perceived quality.
Journal of Marketing Research, 31(2), 191–201.
Aaronson, D. (2001). Price pass-through and the minimum wage. Review of Economics and
Statistics, 83(1), 158–169.
Aaronson, D., French, E., & MacDonald, J. (2008). The minimum wage, restaurant prices, and
labor market structure. Journal of Human Resources, 43(3), 688–720.
Abadie, A., Diamond, A., & Hainmueller, A. J. (2010). Synthetic control methods for
comparative case studies: Estimating the effect of California’s tobacco control program.
Journal of the American Statistical Association, 105(490), 493–505.
Abadie, A., & Gardeazabal, J. (2003). The economic costs of conflict: A case study of the
Basque country. American Economic Review, 93(1), 113–132.
Akerman, A., Gaarder, I., & Mogstad, M. (2015). The skill complementarity of broadband
internet. Quarterly Journal of Economics, 130(4), 1781–1824.
Alba, J. W., & Hutchinson, J. W. (1987). Dimensions of consumer expertise. Journal of
Consumer Research, 13(4), 411–454.
Allegretto, S., & Reich, M. (2018). Are local minimum wages absorbed by price increases?
Estimates from internet-based restaurant menus. ILR Review, 71(1), 35–63.
Allison, P. D. (1995). Survival analysis using SAS: A Practical Guide. SAS Institute Inc.
Allison, P. D. (1999). Logistic regression using SAS: Theory and application. SAS Institute Inc.
Anderson, E., Lin, S., Simester, D., & Tucker, C. (2015). Harbingers of failure. Journal of
Marketing Research, 52(5), 580–592.
Anderson, E. W., & Sullivan, M. W. (1993). The antecedents and consequences of customer
satisfaction for firms. Marketing Science, 12(2), 125–143.
Anderson, M., & Magruder, J. (2012). Learning from the crowd: Regression discontinuity
estimates of the effects of an online review database. Economic Journal, 122(563), 957–
989.
Athey, S., & Imbens, G. W. (2017). The state of applied econometrics: Causality and policy
valuation. Journal of Economic Perspectives, 31(2), 3–32.
Bai, J. (2009). Panel data models with interactive fixed effects. Econometrica, 77(4), 1229–1279.
94
Bronnenberg, B. J., Kruger, M. W., & Mela, C. F. (2008). Database paper — The IRI marketing
data set. Marketing Science, 27(4), 745–748.
Brucks, M. (1986). A typology of consumer knowledge content. Advances in Consumer
Research, 13, 58–63.
Card, D., & Krueger, A. B. (1994). Minimum wages and employment: A case study of the fast-
food industry in New Jersey and Pennsylvania. American Economic Review, 84(4), 772–
793.
Chandrasekaran, D., Arts, J. W. C., Tellis, G. J., & Frambach, R. T. (2013). Pricing in the
international takeoff of new products. International Journal of Research in Marketing,
30(3), 249–264.
Chang, T. Z., & Wildt, A. R. (1996). Impact of product information on the use of price as a
quality cue. Psychology and Marketing, 13(1), 55–75.
Chevalier, J. A., & Mayzlin, D. (2006). The effect of word of mouth on sales: Online book
reviews. Journal of Marketing Research, 43(3), 345–354.
Chintagunta, P. K., Gopinath, S., & Venkataraman, S. (2010). The effects of online user reviews
on movie box office performance: Accounting for sequential rollout and aggregation across
local markets. Marketing Science, 29(5), 944–957.
Clogg, C. C., Petkova, E., & Haritou, A. (1995). Statistical methods for comparing regression
coefficients between models. American Journal of Sociology, 100(5), 1261–1293.
Cordell, V. V. (1997). Consumer knowledge measures as predictors in product evaluation.
Psychology and Marketing, 14(3), 241–260.
Curry, D. J., & Faulds, D. J. (1986). Indexing product quality: Issues, theory, and results. Journal
of Consumer Research, 13(1), 134–145.
de Langhe, B., Fernbach, P. M., & Lichtenstein, D. R. (2016). Navigating by the stars:
Investigating the actual and perceived validity of online user ratings. Journal of Consumer
Research, 42(6), 817–833.
Ding, M., Ross, W. T., & Rao, V. R. (2010). Price as an indicator of quality: Implications for
utility and demand functions. Journal of Retailing, 86(1), 69–84.
Dube, A., Lester, T. W., & Reich, M. (2010). Minimum wage effects across state borders:
Estimates using contiguous counties. Review of Economics and Statistics, 92(4), 945–964.
Dubin, J. A., & McFadden, D. L. (1984). An econometric analysis of residential electric
appliance holdings and consumption. Econometrica, 52(2), 345–362.
95
Floyd, K., Freling, R., Alhoqail, S., Cho, H. Y., & Freling, T. (2014). How online product
reviews affect retail sales: A meta-analysis. Journal of Retailing, 90(2), 217–232.
Germann, F., Ebbes, P., & Grewal, R. (2015). The chief marketing officer matters! Journal of
Marketing, 79(3), 1–22.
Golder, P. N., Mitra, D., & Moorman, C. (2012). What is quality? An integrative framework of
processes and states. Journal of Marketing, 76(4), 1–23.
Greene, W. H. . (2012). Econometric analysis, 7th ed. In Prentice Hall (7th ed.). Prentice Hall.
Guo, T., Sriram, S., & Manchanda, P. (2020). “Let the sunshine In”: The impact of industry
payment disclosure on physician prescription behavior. Marketing Science, 39(3), 516–539.
Hastie, T., Tibshirani, R., James, G., & Witten, D. (2013). An introduction to statistical learning.
Springer.
Heckman, J. (1979). Sample specification bias as a selection error. Econometrica, 47(1), 153–
161.
Hilbe, J. M. (2009). Logistic regression models. CRC Press LLC.
http://ebookcentral.proquest.com/lib/socal/detail.action?docID=1633158
Hollenbeck, B. (2018). Online reputation mechanisms and the decreasing value of chain
affiliation. Journal of Marketing Research, 55(5), 636–654.
Hong, J., & Sternthal, B. (2010). The effects of consumer prior knowledge and processing
strategies on judgments. Journal of Marketing Research, 47(2), 301–311.
Jindal, N., & McAlister, L. (2015). The impacts of advertising assets and R&D assets on
reducing bankruptcy risk. Marketing Science, 34(4), 555–572.
Johnson, E. J., & Russo, J. E. (1984). Product familiarity and learning new information. Journal
of Consumer Research, 11(1), 542–550.
Kalish, S., Mahajan, V., & Muller, E. (1995). Waterfall and sprinkler new-product strategies in
competitive global markets. International Journal of Research in Marketing, 12(2), 105–
119.
Klein, B., & Leffler, K. B. (1981). The role of market forces in assuring contractual
performance. The Journal of Political Economy, 89(4), 615–641.
Kopalle, P. K., & Hoffman, D. L. (1992). Generalizing the sensitivity conditions in an overall
index of product quality. Journal of Consumer Research, 18(4), 530–535.
96
Kopalle, P. K., & Lehmann, D. R. (2001). Strategic management of expectations: The role of
disconfirmation sensitivity and perfectionism. Journal of Marketing Research, 38(3), 386–
394.
Kronrod, A., Lee, J. K., & Gordeliy, I. (2021). Liar, liar, pants on fire! Combining experimental
design and natural language processing to distinguish between authentic and fictitious
product reviews.
Lee, T. Y., & Bradlow, E. T. (2011). Automated marketing research using online customer
reviews. Journal of Marketing Research, 48(5), 881–894.
Lemos, S. (2008). A survey of the effects of the minimum wage on prices. Journal of Economic
Surveys, 22(1), 187–212.
Li, X., & Hitt, L. M. (2010). Price effects in online product reviews: An analytical model and
empirical analysis. MIS Quarterly: Management Information Systems, 34(4).
Luca, M. (2011). Reviews, reputation, and revenue: The case of Yelp. com.
Luca, M., & Reshef, O. (2021). The effect of price on firm reputation. Management Science,
67(7), 4408–4419.
Luca, M., & Zervas, G. (2016). Fake it till you make it: Reputation, competition, and Yelp
review fraud. Management Science, 62(12), 3412–3427.
Lucas, R. (1976). Econometric policy evaluation: A critique. The Phillips Curve and Labor
Markets, Carnegie-Rochester Conference Series on Public Policy, 19–46.
McFadden, D. (1974). The measurement of urban travel demand. Journal of Public Economics,
3(4), 303–328.
Milgrom, P., & Roberts, J. (1986). Price and advertising signals of product quality. Journal of
Political Economy, 94(4), 796–821.
Miyazaki, A. D., Grewal, D., & Goodstein, R. C. (2005). The effect of multiple extrinsic cues on
quality perceptions: A matter of consistency. Journal of Consumer Research, 32(1), 146–
153.
Moe, W. W., & Schweidel, D. A. (2012). Online product opinions: Incidence, evaluation, and
evolution. Marketing Science, 31(3), 372–386.
Murphy, R. (2020). Local consumer review survey 2020. BrightLocal.
Netzer, O., Feldman, R., Goldenberg, J., & Fresko, M. (2012). Mine your own business: Market-
structure surveillance through text mining. Marketing Science, 31(3), 521–543.
97
Neumark, D., & Wascher, W. (2000). Minimum wages and employment: A case study of the
fast-food industry in New Jersey and Pennsylvania: Comment. American Economic Review,
90(5), 1362–1396.
Nguyen, P., Wang, X. (Shane), Li, X., & Cotte, J. (2021). Reviewing experts’ restraint from
extremes and its impact on service providers. Journal of Consumer Research, 47(5), 654–
674.
Olden, A., & Moen, J. (2020). The triple difference estimator. SSRN Electronic Journal.
Oliver, R. L. (1980). A cognitive model of the antecedents and consequences of satisfaction
decisions. Journal of Marketing Research, 17(4), 460–469.
Oliver, R. L. (2014). Satisfaction: A behavioral perspective on the consumer (2nd ed.).
Routledge.
Park, C. W., & Lessig, V. P. (1981). Familiarity and its impact on consumer decision biases and
heuristics. Journal of Consumer Research, 8(2), 223–230.
Pattabhiramaiah, A., Sriram, S., & Manchanda, P. (2019). Paywalls: Monetizing online content.
Journal of Marketing, 83(2).
Pauwels, K., Silva-Risso, J., Srinivasan, S., & Hanssens, D. M. (2004). New products, sales
promotions, and firm value: The case of the automobile industry. Journal of Marketing,
68(4), 142–156.
Puranam, D., Kadiyali, V., & Narayan, V. (2021). The impact of increase in minimum wages on
consumer perceptions of service: A transformer model of online restaurant reviews.
Marketing Science, 40(5), 985–1004.
Rao, A. R., & Monroe, K. B. (1988). The moderating effect of prior knowledge on cue utilization
in product evaluations. Journal of Consumer Research, 15(2), 253–264.
Rao, A. R., & Monroe, K. B. (1989). The effect of price, brand name, and store name on buyers’
perceptions of product quality: An integrative review. Journal of Marketing Research,
26(3), 351–357.
Reddy, S. K., Holak, S. L., & Bhat, S. (1994). To extend or not to extend: Success determinants
of line extensions. Journal of Marketing Research, 31(2), 243–262.
Rosario, A. B., Sotgiu, F., De Valck, K., & Bijmolt, T. H. A. (2016). The effect of electronic
word of mouth on sales: A meta-analytic review of platform, product, and metric factors.
Journal of Marketing Research, 53(3), 297–318.
Rust, R. T., Moorman, C., & Dickson, P. R. (2002). Getting return on quality: Revenue
expansion, cost reduction, or both? Journal of Marketing, 66(4), 7–24.
98
Rust, R. T., Zahorik, A. J., & Keiningham, T. L. (1995). Return on quality (ROQ): Making
service quality financially accountable. Journal of Marketing, 59(2), 58–70.
Schlosser, A. E. (2005). Posting versus lurking: Communicating in a multiple audience context.
Journal of Consumer Research, 32(2), 260–265.
Schoenmueller, V., Netzer, O., & Stahl, F. (2020). The polarity of online reviews: Prevalence,
drivers and implications. Journal of Marketing Research, 57(5), 853–877.
Shapiro, C. (1983). Premiums for high quality products as returns to reputation. Quarterly
Journal of Economics, 98(4), 659–679.
Stammann, A. (2018). Fast and feasible estimation of generalized linear models with high-
dimensional k-way fixed effects. ArXiv Preprint ArXiv: 1707.01815.
Sullivan, M. W. (1992). Brand extensions: When to use them. Management Science, 38(6), 793–
806.
Teas, R. K., & Agarwal, S. (2000). The effects of extrinsic product cues on consumers’
perceptions of quality, sacrifice, and value. Journal of the Academy of Marketing Science,
28(2), 278–290.
Tellis, G. J., & Johnson, J. (2007). The value of quality. Marketing Science, 26(6), 758–773.
Tellis, G. J., & Wernerfelt, B. (1987). Competitive price and quality under asymmetric
information. Marketing Science, 6(3), 240–253.
Tellis, G. J., Yin, E., & Niraj, R. (2009). Does quality win? Network effects versus quality in
high-tech markets. Journal of Marketing Research, 46(2), 135–149.
Tirunillai, S., & Tellis, G. J. (2014). Mining marketing meaning from online chatter: Strategic
brand analysis of big data using latent dirichlet allocation. Journal of Marketing Research,
51(4), 463–479.
Tirunillai, S., & Tellis, G. J. (2017). Does offline TV advertising affect online chatter? Quasi-
experimental analysis using synthetic control. Marketing Science, 36(6), 862–878.
Van Heerde, H. J., Dekimpe, M. G., & Putsis, W. P. (2005). Marketing models and the Lucas
critique. Journal of Marketing Research, 42(1), 15–21.
Venables, W. N., & Ripley, B. D. (2002). Modern applied statistics with S, 4th ed. (Fourth).
Springer.
Völckner, F., & Hofmann, J. (2007). The price-perceived quality relationship: A meta-analytic
review and assessment of its determinants. Marketing Letters, 18(3), 181–196.
99
Watson, J., Ghosh, A. P., & Trusov, M. (2018). Swayed by the numbers: The consequences of
displaying product review attributes. Journal of Marketing, 82(6), 109–131.
Wojnicki, A. C., & Godes, D. (2017). Signaling success: Word of mouth as self-enhancement.
Customer Needs and Solutions, 4(4), 68–82.
Xu, Y. (2017). Generalized synthetic control method: Causal inference with interactive fixed
effects models. Political Analysis, 25(1), 57–76.
Yi, Y. (1990). A critical review of consumer satisfaction. Review of Marketing, 4(1), 68–123.
Zeithaml, V. (1988). Consumer perceptions of price, quality, and value: A means-end model and
synthesis of evidence. Journal of Marketing, 52(3), 2–22.
Zervas, G., Proserpio, D., & Byers, J. W. (2017). The rise of the sharing economy: Estimating
the impact of Airbnb on the hotel industry. Journal of Marketing Research, 54(5), 687–705.
Zhang, Z., Zhang, Z., & Yang, Y. (2016). The power of expert identity: How website-recognized
expert reviews influence travelers’ online rating behavior. Tourism Management, 55, 15–24.
100
Appendix A: Lists of words used to measure expectation, disconfirmation, and
mentions of restaurant-specific attributes in text analysis
Most frequently used 100 restaurant-specific nouns
food, place, service, time, chicken, order, wait, menu, people, meat, night, sauce, experience,
dish, bar, dinner, rice, bit, cheese, salad, way, pork, meal, taste, side, staff, dishes, drinks, fries,
friends, flavor, ramen, beef, spot, area, lunch, friend, hour, pizza, minutes, bread, quality, soup,
day, server, price, line, times, burger, sandwich, atmosphere, top, fish, sushi, places, tacos,
dessert, street, drink, shrimp, noodles, garlic, beer, pasta, waiter, wine, check, plate, steak,
location, things, egg, tables, prices, ambience, home, bbq, party, broth, fun, bowl, waitress,
review, reservation, cream, end, course, vegan, water, valet, portions, birthday, years, flavors,
selection, brunch, family, group, items, style
Words indicating expectations
hype, heard, expect, expectation, expecting, expected, anticipate, anticipated, anticipating,
hyped, hope, hoped, disappoint, disappointment, disappointments, disappointing, disappointed,
disenchanted, dissatisfied, disillusioned, dismay, dismayed, disheartening, disheartened
Words indicating negative disconfirmation
disappoint, disappointment, disappointments, disappointing, disappointed, disenchanted,
dissatisfied, disillusioned, dismay, dismayed, disheartening, disheartened
101
Appendix B: Generalized Synthetic Control using difference in ratings
GSC model using difference in zip code level ratings as the dependent variable
We estimate the GSC model on the difference in zip code level ratings between experts and
novices to confirm the differential treatment effect on experts and novices. We specify the model
as follows:
ratingDifferenceit = ditDit + x¢itb + l¢ift + eit (B1)
where ratingDifference is the difference between experts’ and novices’ ratings for zip code i in
month t. The remaining terms are the same as in Equation 2 in the main manuscript. We estimate
Equation B1 separately for high and low price levels. We show the estimated average treatment
effect on the treated zip codes in Table B.1. The difference in ratings between experts and
novices widens much more in the treated zip codes in Los Angeles for high price levels than for
low price levels.
Table B.1: Average treatment effect on difference in ratings between frequent and infrequent
reviewers for Los Angeles (vs. Las Vegas) zip codes using Generalized Synthetic Control
DV: Difference in rating between frequent and infrequent reviewers at the
zip code level
Estimate S.E.
High price .879*** .108
Low price .322*** .052
***p < .001
102
Appendix C: The fifty geographic markets in the IRI data set
Figure C.1: Fifty geographical markets in IRI academic data set
103
Appendix D: Two stage model with Heckman Correction
In this section, we explain an alternative approach of estimating Equations 1 and 3
described in subsection 2.3.5 in Chapter 2. As in that subsection, we have two models. First, we
model the probability of a firm choosing a particular launch strategy as a function of several
firm-level and brand specific variables using an unordered multinomial logit model. The
probability that company j chooses strategy m to launch its new product i for brand b is given by:
Pr(Yijb = m | wijb) = Pijbm = exp(dmwijb) / (1 + å
2
n=1 exp(dnwijb)) (D1)
In addition to all the variables listed in the main manuscript, wijb also includes an
instrument called strategy prevalence in the category. An instrument in the first stage is
necessary to recover consistent estimates of the effect of launch strategy on product failure in the
second stage. Following the approach in Germann, Ebbes, and Grewal (2015), we define strategy
prevalence as the proportion of new products launched in a category using a particular strategy.
For a focal firm, we calculate strategy prevalence by excluding the products that it has launched.
Since the sprinkler strategy is the reference strategy in the model, we calculate two instruments
for prevalence of waterfall and niche strategies.
Valid instruments should be (1) relevant and (2) meet the exclusion restriction condition.
First, the proposed instrument is likely relevant because firms often choose their strategy based
on strategies their competitors employ in a category. We do not ex ante predict the direction of
this influence as firms could imitate their competitors by adopting the most prevalent strategy or
differentiate themselves by adopting the less prevalent strategies. Second, to meet the exclusion
restriction condition, the instrument should not be correlated with the omitted variables that
104
affect product success. The instrument strategy prevalence results from the action of multiple
firms in a category. It is difficult for peer firms to collectively observe a focal firm’s omitted
variables that affect new product success and then strategically act on those variables (Germann
et al., 2015). Thus, the instrument and firm-level omitted variables are likely uncorrelated,
satisfying the instrument’s exclusion restriction condition.
As in subsection 2.3.5, we model new product failure using a discrete-time hazard model
in the second stage (Allison 1995). The focal event, product failure in a store, occurs in a week
when the product has zero sales in that store. To control for potential endogeneity, we use the
strategy predictors as well as selection correction terms from the first stage model as controls in
the second stage model. The selection correction terms, also known as inverse Mills ratio (IMR),
are obtained from the predicted probabilities from the first stage equation. We calculate these
terms using the approach in Dubin and McFadden (1984) who use a multinomial logit in their
first stage. Specifically, the inverse Mills ratio are calculated using the following:
IMR = å
r
n¹k [ p ̂ n log p ̂ n /(1- p ̂ n)+ log p ̂ k] (D2)
where n refers to one of the three strategies and p ̂ refers to the predicted probabilities of choosing
one of the strategies estimated using Equation D1. Since sprinkler is the reference strategy, we
calculate two inverse Mills ratio terms for waterfall and niche strategies respectively and use
them as control variables in the second stage model.
We specify product failure using the following:
hijbsct = Pr(Tijbsc = t | Tijbsc ³ t, xijbsct) (D3)
105
where hijbsct is the probability that new product i of brand b of company j fails in store s of chain
c during week t, given that failure has not yet occurred. Tijbsc is the week when the new product
fails and xijbsct is a vector of explanatory variables observed for product i in week t. We trace the
survival of a new product in each store in each week since launch. The dependent variable takes
value 1 in a week if the product fails and 0 otherwise. We specify a time-varying logit for our
hazard model as:
log (hibjsct/(1- hibjsct)) = h0ijsct + b1LaunchStrategyibj
+ b2RelativePriceibjsct +b3RelativeFeatureibjsct + b4RelativeDisplayibjsct
+ b5RelativePromoibjsct + b6Relative Salesibjsct + b7MarketsEntryibjt
+ b8StoresEntryibjt + b9CompanyShareij + b10NumMarketsij
+ b11NumChainsij + b12NumStoresij + b13NumBrandsij
+ b14BrandStrengthibj + b15WaterfallIMRibj + b16NicheIMRibj (D4)
The baseline hazard is the same as in Equation 5 in subsection 2.3.5:
h0ijsct = as + ac + g1Timeijsct + g2Timeijsct
2
(D5)
106
D.1 Estimation results
We limit our discussion of results to those that are specific to the Heckman correction
approach to avoid repetition of findings listed in section 2.4.
Table D.1 presents the coefficient estimates of the first stage multinomial logit model in
Equation D1. All key findings regarding the relationship between the probability of choosing one
of the three strategies and the independent variables are similar to those in the main manuscript.
Importantly, the instruments that we use in Equation D1 are statistically significant in all but one
category, suggesting that they are valid instruments. Based on the estimates, we find that firms
are likely to choose strategies that are different from those their peers adopt. For instance, when
the prevalence of waterfall strategy increases, firms in most categories are more likely to choose
sprinkler or niche strategies. Similarly, when the prevalence of niche strategy increases, firms are
more likely to choose sprinkler or waterfall strategies.
Table D.2 contains the raw and standardized parameter estimates for Equation D4 for all
product categories. From the estimates, we see that in majority of the categories (56%), products
launched with a waterfall strategy survive the longest. Products launched using sprinkler and
niche strategies survive the longest in an equal number of remaining categories (22% each).
Thus, the substantive results of our alternative approach reinforce the finding in the main
manuscript that the waterfall strategy is appropriate for new product survival in most categories.
However, the appropriate strategy for survival changes in five categories – carbonated beverages,
frozen dinner, frozen pizza, paper towels, and salty snacks – when the current results are
compared with those obtained with the models in the main manuscript. We suspect this change is
related to the amount of variation in strategy choice explained by the independent variables in
the first stage model. The pseudo R
2
values of the first stage model indicates the percentage
107
variation in strategy choice that the independent variables explain. The pseudo R
2
values for
some of the above-mentioned categories (see Table D.1) - carbonated beverages (17%), frozen
dinner (37%), and paper towels (36%) – are relatively much less compared to most other
categories. When the first stage predictors explain a significant portion of strategy choice, the
selection correction terms used in the Heckman two-step approach possibly does not influence
the second stage results enough to change the substantive results that we obtained in the main
manuscript. But, when the first stage predictors explain a relatively small portion of strategy
choice, the selection correction terms that proxy for the unobserved variables that drive strategy
choice likely play a larger role in the second stage model. This potentially explains why we
obtain different results in some categories while using the two different modeling approaches.
108
109
110
111
112
113
114
115
116
117
Abstract (if available)
Abstract
Quality and market entry are two strategic factors that influence the market leadership of firms. This dissertation examines how consumers report quality heterogeneously in online reviews in response to price changes and how managers could use various new product entry strategies for favorable market outcomes.
Chapter 1 examines how frequent and infrequent reviewers report the quality of restaurants differently in online review platforms, how changes in price influence reviewers’ ratings of quality, and the potential impact reviewers’ heterogeneous reviewing behavior has on businesses. I compiled an extensive data set of over 2 million reviews that involved a quasi-experimental design between two similar cities (LA and Las Vegas), two time periods (separated by a change in the minimum wage in LA), and different reviewers’ frequency. I use text analysis, difference-in-difference-in-differences, and generalized synthetic control to identify the causal impact of price and reviewing frequency on ratings. I find that frequent reviewers write reviews that are significantly more lengthy, useful, multi-dimensional, non-extreme, authentic, and consistent with restaurants’ expert reviewers’ than do infrequent reviewers. These results suggest that frequent reviewers are more knowledgeable about restaurants than infrequent reviewers. Second, infrequent reviewers give lower ratings to high-priced restaurants than frequent reviewers. I show that these low ratings could substantially negatively impact high-priced restaurants’ revenue. These results have important implications for review platforms, consumers, and businesses listed on the platforms.
Chapter 2 studies the different new product launch strategies that consumer packaged goods companies use, which strategies lead to longer survival, and what category-specific factors could guide the choice of the best launch strategy. Launch strategies play a critical role in the long-term success of new products. Identifying the best strategy for long-term survival is managerially relevant. In this paper, I test the ten-year survival of over 650 new products launched in 18 product categories in the U.S. I use a two-stage model: 1) multinomial logit model of strategy choice as a function of the type of firm and category and 2) hazard of survival as a function of strategy choice plus controls. I find that firms widely use a strategy that has not been analyzed in the literature – niche. Even though the niche strategy has the widest use, the waterfall strategy has the longest survival. Further, I find a substantial mismatch between firms’ current choice of strategy and the best strategy for long-term survival. I further show that a model incorporating launch strategy as an independent variable has superior predictive performance in out-of-sample tests. This study has critical implications for new product strategies.
Linked assets
University of Southern California Dissertations and Theses
Conceptually similar
PDF
Marketing strategies with superior information on consumer preferences
PDF
Quality investment and advertising: an empirical analysis of the auto industry
PDF
Essays on the turning points of the product life cycle
PDF
Essays on consumer conversations in social media
PDF
Essays on consumer product evaluation and online shopping intermediaries
PDF
Essays on understanding consumer contribution behaviors in the context of crowdfunding
PDF
Essays on product trade-ins: Implications for consumer demand and retailer behavior
PDF
Essays in non-market strategy and entrepreneurship
PDF
Creation and influence of visual verbal communication: antecedents and consequences of photo-text similarity in consumer-generated communication
PDF
Essays on digital platforms
PDF
Essays on the luxury fashion market
PDF
Essays on dynamic control, queueing and pricing
PDF
Essays on information, incentives and operational strategies
PDF
Three essays on agent’s strategic behavior on online trading market
PDF
Essays on information design for online retailers and social networks
PDF
Competing across and within platforms: antecedents and consequences of market entries by mobile app developers
PDF
The motivational power of beauty: how aesthetically appealing products drive purchase effort in consumers
PDF
The effects of a customer's comparative processing with positive and negative information on product choice
PDF
Efficient policies and mechanisms for online platforms
PDF
Whose quality is it? Transnational TV flows and power in the global TV market
Asset Metadata
Creator
Nair, Sajeev Vijayakumari Krishnan
(author)
Core Title
Essays on the role of entry strategy and quality strategy in market and consumer response
School
Marshall School of Business
Degree
Doctor of Philosophy
Degree Program
Business Administration
Degree Conferral Date
2022-05
Publication Date
04/13/2022
Defense Date
02/28/2022
Publisher
University of Southern California
(original),
University of Southern California. Libraries
(digital)
Tag
consumer packaged goods,difference-in-difference,expertise,generalized synthetic control,hazard model,innovation,launch strategy,new products,OAI-PMH Harvest,online reviews,Price,quality,quasi-experiment,Survival,text analysis
Format
application/pdf
(imt)
Language
English
Contributor
Electronically uploaded by the author
(provenance)
Advisor
Tellis, Gerard (
committee chair
), Jia, Nan (
committee member
), Puranam, Dinesh (
committee member
), Siddarth, S. (
committee member
)
Creator Email
sajeevvk@gmail.com,svijayak@usc.edu
Permanent Link (DOI)
https://doi.org/10.25549/usctheses-oUC110939611
Unique identifier
UC110939611
Document Type
Dissertation
Format
application/pdf (imt)
Rights
Nair, Sajeev Vijayakumari Krishnan
Type
texts
Source
20220413-usctheses-batch-923
(batch),
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 author, as the original true and official version of the work, but does not grant the reader permission to use the work if the desired use is covered by copyright. It is the author, as rights holder, who must provide use permission if such use is covered by copyright. The original signature page accompanying the original submission of the work to the USC Libraries is retained by the USC Libraries and a copy of it may be obtained by authorized requesters contacting the repository e-mail address given.
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
Repository Email
cisadmin@lib.usc.edu
Tags
consumer packaged goods
difference-in-difference
expertise
generalized synthetic control
hazard model
innovation
launch strategy
new products
online reviews
quality
quasi-experiment
text analysis