Can Your Advertising Really Buy Earned Impressions? The Effect … · 2019-01-01 · Can Your Advertising Really Buy Earned Impressions? The Effect of Brand Advertising on Word of - [PDF Document] (2025)

1

Can Your Advertising Really Buy Earned Impressions?

The Effect of Brand Advertising on Word of Mouth

Mitchell J. Lovett

Simon Business School

University of Rochester

[emailprotected]

Renana Peres

School of Business Administration

Hebrew University of Jerusalem, Jerusalem, Israel 91905

[emailprotected]

Linli Xu

Carlson School of Management

University of Minnesota

[emailprotected]

November 2018

Acknowledgment: We thank the Keller Fay Group for the use of their data and their groundbreaking

efforts to collect, manage, and share the TalkTrack data. We gratefully acknowledge our research

assistants at the Hebrew University--Shira Aharoni, Linor Ashton, Aliza Busbib, Haneen Matar, and

Hillel Zehavi—and at the University of Rochester--Amanda Coffey, Ram Harish Gutta, and Catherine

Zeng. We thank Daria Dzyabura, Sarah Gelper, Barak Libai, and Ken Wilbur for their helpful comments.

We thank participants of our Marketing Science 2015, INFORMS Annual Meeting 2015, Marketing

Science 2017, Marketing Dynamics 2017, NYU 2017 Conference on Digital, Mobile Marketing and

Social Media Analytics sessions and at the Goethe University, University of Minnesota, and University of

Houston Seminar Series.

This study was supported by the Marketing Science Institute, Kmart International Center for Marketing

and Retailing at the Hebrew University of Jerusalem, the Israel Science Foundation, and the Carlson

School of Management Dean’s Small Grant.

2

Can Your Advertising Really Buy Earned Impressions?

The Effect of Brand Advertising on Word of Mouth

Abstract

Paid media expenditures could potentially increase earned media exposures such as social media

posts and other word-of-mouth (WOM) . However, academic research on the effect of

advertising on WOM is scarce and shows mixed results. We examine the relationship between

monthly Internet and TV advertising expenditures and WOM for 538 U.S. national brands across

16 categories over 6.5 years. We find that the average implied advertising elasticity on total

WOM is small: 0.019 for TV, and 0.014 for Internet. On the online WOM (measured by

harvesting brand chatter on blogs, user-forums and volume), we find average monthly effects of

0.008 for TV and 0.01 for Internet advertising. Even the categories that have the strongest

implied elasticities are only as large as 0.05. Despite this small average effect, we do find that

advertising in certain events may produce more desirable amounts of WOM. Specifically, using

a synthetic control approach, we find that being a Super Bowl advertiser causes a moderate

increase in total WOM that lasts one month. The effect on online WOM is larger, but lasts for

only three days. We discuss the implications of these findings for managing advertising and

WOM.

Keywords: word of mouth, advertising, brands, dynamic panel methods, paid media, earned

media, synthetic control methods.

3

1 Introduction

Paid advertising could potentially increase earned media exposures such as social media posts

and word of mouth (WOM, hereafter). Brand conversations commonly reference advertisements

with estimates of online buzz about movie trailers ranging from 9% (Gelper, Peres and

Eliashberg 2016) to 15% (Onishi and Manchanda 2012), and Keller and Fay (2009) estimate that

20% of all WOM references TV ads. Some industry reports claim that the impact of advertising

on WOM is considerable (Graham and Havlena, 2007; Nielsen 2016; Turner 2016), and that the

impact on total WOM (online and offline) can amplify the effect of paid media on sales by 15%

(WOMMA 2014). In some cases, this expectation to boost earned mentions is used to justify

buying high priced ad spots in programs like the Super Bowl (Siefert et al. 2009; Spotts, Purvis,

and Patnaik 2014).

In contrast, scholarly research that estimates the WOM impressions gained from advertising

is scarce. As illustrated in Figure 1, the current literature either focuses on the influence of

advertising on sales (Naik and Raman 2003; Sethuraman, Tellis, and Briesch 2011; Danaher and

Dagger 2013; Dinner, Van Heerde, and Neslin 2014), WOM on sales (Chevalier and Mayzlin

2006; Liu 2006; Duan, Gu, and Whinston 2008; Zhu and Zhang 2010), or their joint influence

on behaviors (Hogan, Lemon, and Libai 2004; Chen and Xie 2008; Moon, Bergey, and Iacobucci

2010; Stephen and Galak 2012; Onishi and Manchanda 2012; Gopinath, Chintagunta, and

Venkataraman 2013; Lovett and Staelin 2016). Research on how advertising induces WOM is

mostly conceptual (Gelb and Johnson 1995; Mangold, Miller, and Brockway 1999), or

theoretical (Smith and Swinyard 1982; Campbell, Mayzlin, and Shin 2017). Existing empirical

studies that measure the effect of advertising on WOM, are sparse, and focus on case studies for

a single company (Park, Roth, and Jacques 1988; Trusov, Bucklin and Pauwels 2009; Pauwels,

4

Aksehirli, and Lackman 2016) or specific product launches such as Onishi and Manchanda

(2012) and Bruce, Foutz and Kolsarici (2012) for movies, and Gopinath, Thomas, and

Krishnamurthi (2015) for mobile handsets, and Hewett et al. (2016) for US banks. Recently,

Tirunillai and Tellis (2017), studied how a TV advertising campaign for HP influenced the

information spread and content of the blogs and product reviews of the brand. Tirunillai and

Tellis (2012) studied the effect of online WOM for 15 firms from 6 markets but the main focus

was on firms’ stock market performance. All these studies focused on online social media and

did not incorporate offline WOM, although offline WOM is estimated to be 85% of WOM

conversations (Keller and Fay 2012). Further, the results from these studies are mixed, with

some positive effects (Onishi and Manchanda 2012; Tirunillai and Tellis (2012, 2017); Gopinath,

Thomas, and Krishnamurthi 2015), non-significant effects (Trusov, Bucklin, and Pauwels 2009;

Onishi and Manchanda 2012; Hewett et al., 2016; Pauwels, Aksehirli, and Lackman 2016), and

even negative effects (Feng and Papatla 2011).

-------Insert Figure 1 about here ------

The main goal of this paper is to evaluate the effect of advertising on WOM. A key

component of the analysis is information on the number of brand mentions: we distinguish two

separate measures using different data sources. The first measure (labeled as total WOM) is

drawn from the Keller Fay TalkTrack database. This dataset includes comprehensive information

about individuals’ online and offline conversations about brands. From this dataset, we include

information on 538 US national brands across 16 broad categories and over 6.5 years. The

5

second measure (labeled as online WOM) comes from data on online social media posts for the

same set of 538 brands over 5 years on Twitter, blogs, and user forums.

We use two distinct analysis approaches to evaluate the effect of advertising on WOM. Our

main analysis leverages monthly WOM and advertising expenditures on Internet, TV, and other

media (from Kantar Media’s Ad$pender database) to quantify the relationship between

advertising expenditures and WOM. We use panel regressions that include brand fixed effects,

category-quarter fixed effects and time effects (trends), while also controlling for past WOM and

news mentions. All variables have brand-level heterogeneous effects.

We find that at a monthly level, the relationship between advertising and WOM is

significant, but small. The average implied elasticity of total WOM is 0.019 for TV advertising

expenditures and 0.014 for Internet display advertising expenditures. The average implied

elasticities of online WOM are in similar ranges: 0.009 for TV advertising and 0.010 for Internet

display advertising. To put these numbers into some perspective, for the average monthly

spending on TV advertising in our sample, approximately 58 million ad exposures are generated.

Based on our estimates, a 10% increase in TV advertising expenditure is associated with about

69,000 additional impressions from total WOM.

We find significant heterogeneity across brands and categories in the estimated relationship

between advertising and WOM. For instance, categories with the largest implied elasticities to

TV advertising on total WOM are Sports and Hobbies, Telecommunications, and Media and

Entertainment. However, the average implied elasticity, even for these largest categories, is still

relatively small (e.g., average elasticity between 0.03 and 0.05). Similar conclusions can be

drawn for the online WOM.

6

We conduct a series of robustness tests and find our results are largely consistent across

these specifications. In some of these tests, we use instrumental variables (advertising costs and

political advertising expenditures) to obtain exogenous variation to estimate the advertising-

WOM relationship. Because our results suggest small effects, the main endogeneity concerns are

downward biases which could arise from WOM acting as an advertising substitute or

measurement errors in the advertising expenditure variables. The results from these robustness

tests are supportive of a limited role for these concerns. The only exception is the relationship

between TV advertising and online WOM posts, which is estimated in the Lasso-IV analysis to

have an implied elasticity of 0.12, an order of magnitude larger than the implied elasticity from

the main model.

Our second set of analyses uses a different approach to causal inference and studies the

effect of advertising on WOM where the effect is expected to be large—Super Bowl advertising.

We conduct an analysis using the generalized synthetic control technique (Abadie, Diamond, and

Hainmeller 2010; Bai 2013; Xu 2017), which constructs a difference-in-difference type estimator

by matching the treatment group to a control group synthesized from a weighted combination of

the non-treated brands. This causal inference technique aims to reduce the potential sources of

bias in order to assess from non-experimental data the causal impact of a treatment (in this case,

advertising on the Super Bowl) on the outcome (WOM).

From this second set of analyses, we find that being a Super Bowl advertiser increases

monthly total WOM by 16% in the month of the Super Bowl and by 22% in the week after the

Super Bowl. This increase suggests “free” impressions of the order of 10%-14% of the average

monthly ad impressions, a meaningful contribution because most evidence suggests the impact

of WOM engagements on consumer choices is larger than that of advertising exposures

7

(Sethuraman, Tellis, and Briesch 2011; You et al., 2015; Lovett and Staelin 2016). Although the

estimated gain in total WOM from being a Super Bowl Advertiser is meaningful, It is perhaps

still not as large as one might expect given the large cost of becoming a Super Bowl advertiser.

In contrast, we find that online social media posts respond even more than total WOM to being a

Super Bowl advertiser, including an average increase of 68% on the day of the Super Bowl.

These estimates for online WOM posts are large—similar to the Lasso-IV robustness test—and

suggest that perhaps online posts respond much more than total WOM to advertising.

Our findings portray a world in which typical advertising does not really buy lasting, broad-

based earned impressions, but might increase online posts in the short-term. Paid advertising

developed for TV and the Internet should not automatically be associated with meaningful

increases in WOM. If a brand has the goal of increasing WOM, and uses advertising as the

vehicle to do so, then care must be taken both to design the campaign for this goal (Van der Lans

et al., 2010) and to monitor that the design is successful. In particular, our results suggest that

monitoring needs to include more than counts of online posts, as such measures are neither

representative nor reflect total WOM. Our results also demonstrate that some campaigns for

some brands such as Super Bowl advertisements generate far higher WOM response, but that the

small average implied elasticity and low heterogeneity across brands and categories suggest that

these larger effects are relatively rare and are not obtained without a focused investment of

considerable resources.

2 Existing Theory and Evidence on the Advertising-WOM

Relationship

Marketing theory provides a foundation for both a positive and a negative advertising-WOM

relationship. On the positive side, engaging in WOM is driven by the need to share and receive

8

information, have social interactions, or express emotions (Lovett, Peres, and Shachar 2013;

Berger 2014). Advertising can trigger these needs and potentially stimulate a WOM conversation

about the brand. Four routes through which advertising might trigger these needs include

attracting attention (Batra, Aaker, and Myers 1995; Mitra and Lynch 1995; Berger 2014),

increasing social desirability and connectedness (Aaker and Biel 2013; Van der Lans and van

Bruggen 2010), stimulating information search (Smith and Swinyard 1982), and raising

emotional arousal (Holbrook and Batra 1987; Olney, Holbrook, and Batra 1991; Lovett, Peres,

and Shachar 2013, Berger and Milkman 2012).

However, advertising can also have a negative influence on WOM. Dichter (1966) argues

that advertising decreases involvement, and if involvement has a positive influence on WOM

(Sundaram, Mitra, and Webster 1998), advertising would cause a decrease in WOM. Feng and

Papatla (2011) claim that talking about an advertised brand may make an individual look less

unique, and may harm her self enhancement. Similarly, if advertising provides sufficient

information so that people have the information they need, they will tend to be less receptive to

WOM messages (Herr, Kardes, and Kim 1991), which diminishes the scope for WOM.

The overall balance between the positive and negative influences is not clear. Scholarly

empirical research on this issue is limited and the available results are mixed. Onishi and

Manchanda (2012) estimated the advertising elasticity of TV advertising exposures on blog

mentions for 12 movies in Japan, and found an elasticity of 0.12 for pre-release advertising, and

a non-significant effect for post-release advertising. Gopinath, Thomas and Krishnamurthi

(2015) studied the impact of the number of ads on online WOM for 5 models of mobile phones

and estimated elasticities of 0.19 for emotion advertising and 0.37 for "attribute" (i.e.

informational) advertising. Feng and Papatla (2011) used data on cars to show both positive and

9

negative effects of advertising on WOM. Using a model of goodwill for movies, Bruce, Foutz

and Kolsarici (2012) found that advertising has a positive impact on the effectiveness of WOM

on demand, but did not study the effect on WOM volume. Bollinger et al. (2013) found positive

interactions between both TV and online advertising and Facebook mentions in influencing

purchase for fast moving consumer goods, but did not study how one affects the other. Tirunillai

and Tellis (2017), studied how a TV advertising campaign for HP influenced the information

spread and content of the blogs and product reviews of the brand. They found a 10-day elasticity

of 0.15, and short-term elasticity of 0.08 on the volume of WOM. Tirunillai and Tellis (2012)

studied 15 firms from 6 markets and estimated the elasticity of online WOM on advertising

expenditures to be 0.09. Hewett et al. (2016) find that advertising spend by four banks do not

affect online Twitter posts, and Pauwels, Aksehirli and Lackman (2016) find that for one apparel

retailer the effects of advertising on electronic brand WOM are relatively large in the long-term,

but small in the short-term.

Thus, both marketing theory and scholarly empirical research offer mixed guidance about

the direction and size of the advertising-WOM relationship. Our focus is to quantify this

relationship using data that cuts across many industries and brands, spans a long time-period, and

captures a wide set of controls. Our setting is mostly large established brands with relatively

large advertising budgets. We next describe the main dataset used in our analysis.

3 Data

Our dataset contains information on 538 U.S. national brands from 16 product categories (the list

is drawn from that of Lovett, Peres, and Shachar 2013), and the complete list appears in Web

Appendix A. The categories include: beauty products, beverages, cars, children’s products,

10

clothing products, department stores, financial services, food and dining, health products, home

design and decoration, household products, media and entertainment, sports and hobbies,

technology products and stores, telecommunication, and travel services. The brands in the list

include products and services, corporate and product brands, premium and economy brands. For

each brand from January 2007 to June 2013, we have monthly information on advertising

expenditures, total number of word-of-mouth mentions, and brand mentions in the news. We also

have data on online WOM between July 2008 and June 2013. We elaborate on each data source

and provide some descriptions of the data below.

3.1 Advertising expenditure data

We collect monthly advertising expenditures from the Ad$pender database of Kantar Media. For

each brand, we have constructed three categories of advertising—TV advertising, Internet

advertising, and other advertising. For TV advertising, we have aggregated expenditures across

all available TV outlets (DMA-level as well as national and cable). For Internet advertising

expenditures, we include display advertising (the only Internet advertising available in

Ad$pender). Display advertising is appropriate since it is more often used as a branding tool,

whereas search advertising is more closely connected with encouraging purchases directly.

Hence, display advertising is more closely aligned with the goal of obtaining WOM. We focus

on these TV and Internet advertising expenditures for three reasons. First, for our brands, these

two types of expenditures make up approximately 70% of the total advertising expenditures

according to Ad$pender. Second, TV advertising is the largest category of spending and has been

suggested to be the most engaging channel (Drèze and Hussher 2003) and often cited as

generating WOM. Third, Internet advertising is touted as the fastest growing category of

spending among those available in Kantar and reflects the prominence of “new media.” That

11

said, we also collect the total advertising expenditures on other media, covering the range of

print media (e.g., newspapers, magazines), outdoor, and radio advertisements.

3.2 Word of mouth and news data

We use two sources of word of mouth data. The first relates to total WOM and comes from an

industry-standard measure of WOM that uses a representative sample in each week of self-

reported brand conversations. The second is more typical of social media listening data and

comes from queries into a large corpus of text from public online posts. In addition, we also

collect the number of news and press mentions for each brand.

3.2.1 Total WOM data from the TalkTrack panel:

Our primary word-of-mouth data is drawn from the TalkTrack dataset of the Keller-Fay Group.

The dataset is an industry standard for measuring WOM, and has been used in various marketing

academic studies (Berger 2014; Baker, Donthu, and Kumar 2016; see Lovett, Peres, and Shachar

(2013) for a detailed description). It contains the number of mentions for each brand every week

across a sample of respondents, who are recruited to self-report for a 24-hour period on all their

word of mouth conversations. During the day they record their brand conversations and list the

brands mentioned in the conversation. Note that a list of brands is not provided to respondents –

i.e., they can mention any brand. These conversations can happen both online and offline. The

inclusion of offline WOM is important, since it is estimated to be 85% of WOM conversations

(Keller and Fay 2012).

The sample includes 700 individuals per week, spread approximately equally across the

days of the week. This weekly sample is constructed to be representative of the U.S. population.

The company uses a scaling factor of 2.3 million to translate from the average daily sample

12

mentions to the daily number of mentions in the population. We aggregate the WOM data to the

monthly level to match with the monthly advertising data on all brands in our main analysis.

3.2.2 Online posts from social media:

The second source of word-of-mouth data we use is a dataset of social media posts extracted

using the Nielsen-McKinsey Insight user-generated content search engine. This search engine

has conducted daily searches through blogs, discussion groups, and microblogs, and the brand

specific information is retrieved using designated queries written for each brand (see Lovett,

Peres, and Shachar (2013) for a detailed description). This dataset covers the time period

between July 2008 and June 2013. We use this dataset to study the effects of advertising on

online posts. In addition, this dataset allows us to conduct the second part of our analysis

described in section 6 at a more granular level since the online posts data are available at a daily

level.

3.2.3 News and Press Mentions

WOM may be triggered by news media, which might also proxy for external events (e.g., the

launch of a new product, a change of logo, product failure or recall). Such events could both lead

the firm to advertise and consumers to speak about the brand, so that the WOM is caused by the

event not the advertising. To control for such unobserved events and news, we use the

LexisNexis news and press database to collect the monthly number of news and press mentions

for each brand.

3.3 Descriptive statistics

Table 1 presents category specific information about the advertising, media mentions, and WOM

mentions data. This table communicates the large variation across categories in the use of the

13

different types of advertising and in the number of media mentions. For example, the highest

spender on TV ads is AT&T, the highest spender on Internet display ads is TD AmeriTrade, and

the brand with the highest number of news mentions is Facebook. The average number of total

mentions for a brand in the sample is 15.8 (equivalent to 36 million mentions in the population),

the brand with the highest total WOM is Coca Cola, and the brand with the highest online WOM

is Google. In Web Appendix A, we present time series plots for four representative brands as

well as descriptive statistics and correlations for the data.

-------- Insert Table 1 about here ----

4 Model for Main Analyses

In our main analysis, we focus on relating advertising expenditures to WOM. Our empirical

strategy is to control for the most likely sources of alternative explanations and evaluate the

remaining relationship between advertising and WOM. Hence, causal inference requires a

conditional independence assumption. We are concerned about several important sources of

endogeneity due to unobserved variables that are potentially related to both advertising and

WOM and, as a result, could lead to a spurious relationship between the two. The chief concerns

and related controls that we include are 1) unobserved (to the econometrician) characteristics at

the brand level that influence the advertising levels and WOM, which we control using brand

fixed effects (and in one robustness test, first differences), 2) WOM inertia that is spuriously

correlated with time variation in advertising, controlled for by including two lags of WOM, 3)

unobserved product introductions and related PR events that lead the firm to advertise and also

generate WOM, which we control using media mentions of each brand, and 4) seasonality and

14

time varying quality of the brand that lead to both greater brand advertising and higher levels of

WOM. For seasonality and time-varying quality, we use a (3rd order) polynomial function of the

month of year to control for high-low season within a year, and category-quarter-year fixed

effects. We also introduce common time effects in a robustness test.

With these controls in mind, our empirical analysis proceeds as a log-log specification

(where we add one to all variables before the log transformation).1 Under the conditional

independence assumption, this specification imposes a constant elasticity for the effect of

advertising expenditures on WOM and implies diminishing returns to levels of advertising

expenditures. For a given brand j in month t, the empirical model is defined as

(1) log(𝑊𝑂𝑀)𝑗𝑡 = 𝛼𝑗 + 𝛼𝑐𝑞 + 𝛽1𝑗𝑙𝑜𝑔(𝐴𝑑𝑇𝑉)𝑗𝑡 + 𝛽2𝑗log(𝐴𝑑𝐼𝑛𝑡𝑒𝑟𝑛𝑒𝑡)𝑗𝑡

+𝛾1𝑗log(𝑊𝑂𝑀)𝑗𝑡−1 + 𝛾2𝑗log(𝑊𝑂𝑀)𝑗𝑡−2 + 𝑋𝑗𝑡𝛽0𝑗 + 휀𝑗𝑡

where 𝛼𝑗 are brand fixed effects, 𝛼𝑐𝑞 are category-quarter-year fixed effects, log(AdTV) and

log(AdInternet) relate to the focal variables, logged dollar expenditures for TV and Internet

display ads, and 𝑋𝑗𝑡 contains control variables that include the logged dollar expenditures for

other advertising (print, outdoor, and radio), logged count of news and press articles mentioning

the brand, and a polynomial (cubic) of month of year. The 𝛾1𝑗 , 𝛾2𝑗 , 𝛽0𝑗 , 𝛽1𝑗 , 𝛽2𝑗 are random

coefficients for, respectively, the effect of lagged word-of-mouth variables, 𝑋𝑗𝑡, and the two

focal advertising variables. Here, we focus on the short-term impact of advertising on WOM by

including the contemporary advertising only. In section 5.3 and Web Appendix B, we report the

1 We also test adding alternative constants (0.1 and 0.01). The relative magnitudes and statistical significance are

consistent with the reported results, and qualitative conclusions remain the same. The results are available from the

authors.

15

empirical results of the model with both contemporary and lagged advertising as well as the

estimated long-term cumulative effects of advertising on WOM.

In what follows, we focus on the average relationship between advertising and WOM across

brands. In one set of results we also allow observable heterogeneity in brand coefficients in the

form of category-level differences.2 For the models that include both random coefficients and

fixed effects we use proc mixed in SAS with REML. For the models without random coefficients

we use plm in R, which estimates the model using a fixed effects panel estimator, noting that in

both models our longer time-series implies negligible `Nickell bias’ in the lagged dependent

variables (Nickell 1981).3

5 Main Results

We organize the results from our main analysis into four sections. The first section presents our

results related to the magnitude of the average relationship between advertising and WOM and

interpreting this magnitude in the broader context of advertising. The second section explores

how much heterogeneity in the advertising-WOM relationship exists across brands and

categories. The third section presents cumulative effects of advertising from a model with

multiple lags of advertising, and the final section discusses other robustness tests including using

instrumental variables to obtain exogenous variation to estimate the advertising-WOM

relationship.

2 We note that we also examined whether the WOM effects varied by brand characteristics using the data provided

by Lovett, Peres, and Shachar (2013). The relationships we found suggested there were few significant relationships,

so few that the relationships could be arising due to random variation rather than actual significance. 3 In robustness checks, we also conduct several two-stage least squares analyses and Lasso-IV analyses to evaluate

the extent of remaining endogeneity bias after our controls. These are also done in R using the plm and hdm

packages.

16

5.1 The advertising-WOM relationship

The first set of columns in Table 2 (Total) presents the results from estimating Equation (1) on

the total TalkTrack WOM dataset. In this section, we focus on the parameters related to the

population mean. We find that the advertising variables indicate significant positive coefficients

for both TV (0.019, s.e. =0.0017) and Internet display ad expenditures (0.014, s.e. =0.0021). The

second set of columns in Table 2 (Online) describes the results for the dataset of online posts.

We see that the estimated coefficients are similar but smaller – The coefficient for TV

advertising is 0.009 (s.e. 0.001), and for Internet advertising it is 0.01 (s.e. 0.002). The difference

between the coefficients for TV and Internet advertising are not significant in both datasets. This

is consistent with the results of Draganska, Hartmann, and Stanglein (2014), who find that

advertising on TV and the Internet do not have significantly different effects on brand

performance metrics.

The control variables take the expected signs, are significant, and have reasonable

magnitudes. Based on the estimated effects for the lagged dependent variables, WOM has a low

level of average persistence in WOM shocks that diminishes rapidly between the first and second

lag, keeping in mind that these effects are net of the brand fixed effects. News mentions have a

much larger significant and positive estimate, but we caution against interpreting this effect as

arising due to news per se, since this variable could also control for new product introductions

which typically are covered in the news. The variance parameters for the heterogeneity across

brands are also all significant.4 We discuss the heterogeneity related to the brand advertising

variables in more detail in section 5.2.

4 We note that we include only heterogeneity in the linear month of year of the cubic function. This was necessary

due to stability problems with estimating such highly collinear terms as random effects. Heterogeneity in the count

17

------- Insert Table 2 about here ---------

How big are these estimated advertising effects on WOM? Since the analysis is done in log-

log space, the estimated coefficients on advertising are constant advertising elasticities under the

causal interpretation of the coefficient. The implied elasticity of total WOM to TV advertising

expenditures is 0.019 and to Internet advertising expenditures is 0.014. For online WOM, the

implied elasticity to TV advertising expenditures is 0.009 and to Internet advertising

expenditures is 0.01.

We offer some perspective on the magnitude of this relationship. First, the relationship is

quite modest even in absolute magnitudes.5 For instance, in our sample, the average number of

conversations about a brand in a month is 15.8. Given the sampling procedure of Keller-Fay,

they project that one brand mention in their sample equals 2.3 million mentions in the United

States. This suggests there are 36.4 million conversations about the average brand in our dataset

in a month. Our elasticity estimate implies that a 10% increase in TV advertising corresponds to

around 69,000 additional conversations in total WOM about the brand per month. For the large,

high WOM national brands that we study, this number of brand conversations is quite small.

Consider the average spending of $5.89 million on TV advertising. A 10% increase in spending

at 1 cent per advertising impression on average generates 58.9 million advertising impressions.

In this case, the additional WOM impressions associated with advertising is orders of magnitude

smaller than the advertising impressions, only one thousandth.

of news mentions was excluded because they turn out to be zero after controlling for the category-quarter fixed

effects. 5 Due to the nature of the online data being non-representative to the U.S. population, we do not attempt to convert

the point estimate to the number of WOM conversations.

18

Second, the translation to sales based on the estimated WOM elasticities in the literature are

quite small, too. For instance, You et al. (2015) in a meta-study of electronic WOM find an

overall elasticity of 0.236 on sales. At this elasticity for WOM on sales, the average impact of

advertising through WOM would be less than 0.004. Further, the 0.236 eWOM elasticity of You

et al. (2015) is relatively large compared to recent studies that find elasticities between 0.01 and

0.06 (Lovett and Staelin 2016; Seiler, Yao, and Wang 2017). With these lower elasticities, the

effect would be an order of magnitude smaller. Given that the meta-studies on the influence of

advertising on sales (e.g., Sethuraman, Tellis, and Briesch 2011) reveal average advertising

elasticities of 0.12, the implied impact of advertising on sales through WOM is only a very small

part of the overall advertising influence.

How do these elasticities relate to the elasticities reported in the specific cases studied in the

scholarly literature? As mentioned above, reported results are mixed, with some analyses

showing a positive effect (Onishi and Manchanda 2012; Tirunillai and Tellis (2012); Gopinath,

Thomas, and Krishnamurthi 2015; Pauwels, Aksehirli and Lackman 2016), some showing no

significant effect (Trusov, Bucklin, and Pauwels 2009; Onishi and Manchanda 2012; Hewett et

al., 2016), and some even showing negative effects (Feng and Papatla 2011). The comparison,

even in the cases of positive elasticities is not very direct. For example, Onishi and Manchanda

(2012) provide an estimated elasticity of 0.12 for daily advertising exposures on pre-release

WOM, where the WOM is blogs about 12 different movies in Japan. For five models of mobile

phones Gopinath, Thomas, and Krishnamurthi (2015) find elasticities between 0.19 and 0.37 for

monthly online WOM to the number of advertisements. Pauwels et al. (2016) finds long-term

brand electronic WOM elasticities of 0.085, 0.149, 0.205, and 0.237 for TV, print, radio, and

paid search ads for weekly data about one apparel retailer. We differ notably in two ways. First,

19

our measure is the response of total monthly WOM, which may smooth some of the daily

variation captured in Onishi and Manchanda (2012) and the weekly variation in Pauwels et al.

(2016). Second, our data covers over 500 brands, spans 6.5 years, and covers all types of WOM,

not just online. With these broader definitions and sample, it appears the estimated average

relationship between advertising and WOM is much smaller than what is currently reported in

the literature.6 Hence, in absolute terms and relative to the positive findings in the literature, we

find a weak average advertising-WOM relationship.

5.2 Does the average effect mask larger effects for some brands or categories?

We now turn to how much stronger the relationship is for some brands and categories than the

average coefficients we reported thus far. Brand level heterogeneity in the relationship between

advertising and WOM could lead some brands to have strong relationships and others to have

weak relationships, resulting in the small average coefficients described above. For instance, this

variation could arise from different customer bases, different brand characteristics, different

degrees of engagement with the brand, or different types or quality of advertising campaigns

between brands. Heterogeneity variances in both Total WOM and Online WOM reported in

Table 2 shows that the standard deviations for the heterogeneity in advertising coefficients are

roughly the same size as the coefficients themselves, indicating that brands differ meaningfully

in the relationship between WOM and advertising. However, the cross-brand variation does not

produce an order of magnitude shift in the point estimates. For example, for the TV ads effect, a

two standard deviation shift implies that a few brands have point estimates as large as 0.059 for

total WOM and 0.049 for online WOM. Although the max of these point estimates is larger than

6 Our small effect size appears similar in some respects to Du, Joo and Wilbur’s (forthcoming) small correlation

between advertising and brand image measures using weekly data.

20

the overall average, 0.059 and 0.049 are still less than half the size of the typical sales elasticity

to advertising. This suggests that even for the brands with the largest relationships between

advertising and WOM, the magnitudes are relatively modest.

To understand whether the relationships systematically differ across categories, we

incorporate category dummy variables and interact them with the variables in Equation (1).

Figure 2a presents the category level estimates with +/− one standard error bars for both TV and

Internet dollar spend for total WOM. As apparent, the beauty category has the smallest average

TV advertising-total WOM relationship (-0.003, but not significantly different from zero),

whereas the highest estimate is 0.046 for Sports and Hobbies, significantly larger than zero and

the coefficient for beauty. Also, on the high-end are Telecommunications, which includes mobile

handset sellers, and Media and Entertainment, which includes movies. These latter two

categories are ones that past research has found to have significant, positive effects of advertising

on WOM (mobile handsets and movies). Hence, the category variation we find is directionally

consistent with the categories that have been studied in the past being exceptionally large. For

Internet display advertising expenditures, we find that the Clothing category has one of the

weakest relationships, whereas Media and Entertainment has the highest.

Figure 2b shows the same estimates for online WOM. Sports and Hobbies have the

strongest relationship for the Internet display advertising expenditure, followed by Media and

Entertainment; whereas department stores have the weakest relationship. The highest estimate of

the TV advertising-online WOM relationship occurs in Media and Entertainment. Overall, we

find that the categories with the largest advertising elasticities are still relatively small for both

total WOM and online WOM.

------- Insert Figure 2 about here ---------

21

5.3 Does the average effect mask larger cumulative effects?

Because the estimates we report could cover up a larger cumulative effect, we also consider

models with lagged advertising expenditure variables. The details of this examination are

available in Web Appendix B. Our finding is that although some lags are considered significant,

the results do not meaningfully alter the conclusions reported here. Our estimates indicate that

the cumulative relationship of advertising on total WOM is 0.031 for TV advertising expenditure

and 0.020 for Internet display advertising. For online WOM the cumulative relationship is 0.017

for TV advertising and 0.013 for Internet display advertising. Interestingly, TV advertising

expenditures appear to have some longer-term effects, but Internet advertising expenditures do

not.

5.4 Robustness tests and instrumental variables analyses

In Web Appendix C, we provide details on a range of model tests that support the robustness of

the main results presented above.7 First, we drop or add different components to the model to

evaluate robustness to specification. We find that as long as either lagged WOM or brand fixed

effects are included in the model, the small advertising-WOM relationships described above

maintain. Importantly, the brand fixed effects are critical controls since without them the

relationship between WOM and advertising expenditures would appear to be stronger than it

actually is.

7 We also examined whether WOM that mentions advertising has a stronger relationship with advertising

expenditures. We find that it is not meaningfully different in magnitude. These results are in Web Appendix D.

22

Second, we evaluate instrumental variables specifications. Our main empirical strategy

leverages control variables to reduce potential endogeneity concerns related to seasonality,

unobserved brand effects, secular trends, and new product/service launches. The causal

interpretation of our results relies on a conditional independence assumption. The main concerns

in estimating advertising causal effects typically involve positive biases (e.g., brands advertise in

the high season of sales that might falsely be attributed to the advertising). We have attempted to

control for these concerns and show that our control variables do not overly influence the results.

Since failing to account for endogeneity of advertising is usually expected to produce larger

effect sizes, our small effect size suggests the typical concerns are not a major threat.

Two main arguments specific to our context could lead to a downward bias. The first is that

advertising and WOM could serve as substitutes. However, since the advertising for large

established brands tends to be planned well in advance, advertising cannot easily respond to

short-term fluctuations in WOM. Hence, we can narrow the substitutes concern to planning to

cut advertising when WOM is expected to be high and vice versa. For example, when the

product would be on consumers’ minds and talked about (e.g., summer for ice cream), the firm

would choose not to advertise. On the face, this argument appears counterfactual (i.e., ice cream

is advertised more in summer). Even so, our brand level seasonality and secular trend controls

are intended to address this kind of concern. The second main argument that could lead to

smaller effect sizes is measurement error in the advertising expenditure variables. Classical

error-in-variables arising from measurement problems is known in linear models to produce

attenuation bias, underestimating the effect size. We next consider models that can account for

these endogeneity concerns.

23

We examine whether our results are robust to an instrumental variables approach to obtaining

exogenous variation to estimate the advertising-WOM relationship. We find that the main

estimates do not shift meaningfully for total WOM, but do in one specification for online WOM

posts. We use many instruments--interactions of costs and political advertising with brand

identities. This collection of instruments is jointly weak in the standard linear instrumental

variables specification, so we use post-LassoIV to approximate the optimal instruments (Belloni

et al 2012). In this post-LassoIV approach, the estimated effects for total WOM are still

qualitatively the same. Those for online WOM posts, however, are much larger and significant

with an implied elasticity of 0.12. These larger effects could be consistent with online posts

acting as substitutes with advertising expenditures. However, even these post-LassoIV estimates

could still face a weak instruments problem, as we find that many of the first stage coefficients

have unexpected signs making the theoretical argument for the instruments less clear.

Together, these additional analyses presented in Web Appendix C provide support for the

robustness of our main results, and in particular the small average effect. However, because the

instruments could be relatively weak, it is difficult to establish that endogeneity is not biasing our

results toward zero as a result of measurement error or advertising and WOM acting as

substitutes. In order to further address these potential issues, we use an alternative approach in

section 6 that specifically avoids both of these concerns.

6 Advertising in the Super Bowl

In this section, we examine what is typically considered a situation where advertising is intended

to generate WOM and is believed to have very large effects—Super Bowl advertisements. While

the heterogeneity in categories and brands described in the section 5.2 suggests that persistent

24

differences do not lead to large magnitudes for the advertising-WOM relationship, it is possible

that some events, periods, or specific campaigns might do so. One leading possibility is that

certain campaigns or events are simply better at generating conversation than others. To evaluate

this potential, we examine one of the most often cited sources of incremental WOM impressions

from advertising—the Super Bowl.8 We collected information on which of the brands in our

dataset advertised in the Super Bowl in each of the years in our sample. We apply this data to

two different analyses. In the first we examine the main model results including main effects for

being a Super Bowl advertiser and interactions between this variable and the advertising

expenditure variables. In the second analysis we apply a causal inference technique, synthetic

controls, to evaluate the relationship between being a Super Bowl advertiser and WOM.

6.1 Regression analysis of Super Bowl advertising effects?

We add to our main analysis of equation (1) both a main effect of being a Super Bowl advertiser

in the month of the Super Bowl (February) and interaction terms between this variable and the

logged advertising spending variables. If the Super Bowl increases the effectiveness of

advertising spending on WOM impressions, we would expect the coefficients on the interaction

terms to be positive. The Super Bowl main effect and interaction effects do not have random

coefficients, because they are not separately identified from the brand fixed effects and the

brand-specific advertising random coefficients.

Table 3 presents the results. We find some interesting differences between total WOM and

online WOM. For total WOM, we find that none of the Super Bowl interaction terms is large or

significant. In fact, the term for TV advertising, which one would expect to be positive if Super

8 We also collected data on advertising awards including, for instance, the EFFIE, CANNES, and OGILVY awards.

Interacting the award winners in a year with their advertising expenditure produced no statistically significant

interactions nor systematic pattern.

25

Bowl advertising is more efficient, is actually negative and small (but insignificant). While this

finding suggests that advertising on the Super Bowl does not lead to stronger relationships

between advertising expenditures and total WOM, the main effect potentially tells a different

story. In particular, the main effect of being a Super Bowl advertiser for total WOM is positive,

large (0.27) and significant (t-stat = 2.31). This indicates that although Super Bowl advertising

expenditures are not more efficient per dollar than at other times, Super Bowl advertisers have on

average 27% higher total WOM in the month of the Super Bowl than in other periods. This large

effect size could suggest that advertising is more effective in the Super Bowl for creating total

WOM, but that the variation in advertising spending on Super Bowl ads is insufficient to

attribute that gain to advertising expenditures. Since such an increase could translate to a much

larger effect than what we find in the small average elasticity, this result seems to provide an

opportunity for advertising to play a larger role in creating total WOM than our previous findings

suggest.

In contrast, for online WOM, we find that the Super Bowl interaction term with TV

advertising expenditure is positive (0.036) and significant (t-stat=1.99), while the other

interactions are not significant. This implies that advertising in the Super Bowl does lead to a

stronger relationship between TV advertising and online WOM posts. Further, the online WOM

posts effect is consistent with the instrumental variables analysis for TV advertising that has a

larger coefficient.

Taken together with the large main effect of Super Bowl in total WOM, these findings are

consistent with both the popular press and practitioner literature arguing that Super Bowl

advertisements lead to a large increase in WOM impressions. If these Super Bowl effects are

causal, then advertising may generate meaningful levels of WOM in some campaigns or when

26

combined with specific events. In the next section, we examine this finding in more detail using

a recent causal inference technique that can provide further robustness of our findings.

6.2 A synthetic controls analysis of advertising in the Super Bowl on WOM

Unlike in the main analysis, where we observe multiple continuous advertising expenditure

variables, the analysis in this section focuses on whether being a Super Bowl TV advertiser

causes an increase in WOM. In this case, we have a discrete “treatment” variable, Super Bowl,

which takes a value of 1 for Super Bowl advertisers in the time period(s) when we test for an

effect, and 0 otherwise. In this section, we present evidence about the effect of this Super Bowl

treatment using a causal inference method to reduce potential bias.

To measure this causal effect, we would ideally like to calculate the difference between the

realized WOM for the Super Bowl advertisers as compared to the counterfactual case, the WOM

these brands would receive had they not advertised in the Super Bowl. Of course, by definition,

we do not (and cannot) observe the counterfactual case for the same brands, and instead seek a

way to generate the missing counterfactual WOM data. Ideally, we would run a field experiment

that randomizes the assignment of Super Bowl advertising slots to brands in order to justify

using the non-treated brands as the counterfactual measure. This is infeasible.9

To construct the prediction for this missing counterfactual data, we use a recently developed

technique, the Generalized Synthetic Control Method (GSCM) of Xu (2017). This method is a

parametric approach that generalizes to multiple treatment units the synthetic control method

developed by Abadie, Diamond, and Hainmueller (2010). The method was originally developed

9 We note that some recent papers have used other strategies that leverage geographic variation and the surprise of

who plays in the Super Bowl (Hartmann and Klapper 2018). We could not employ this approach due to national

level data.

27

for comparative case studies, and has been used and extended broadly including in economics

(Doudchenko and Imbens 2016), finance (Acemoglu et al., 2016), political science (Xu 2017),

and, recently, in marketing (e.g., Vidal-Berastain, Ellickson, and Lovett 2018).

The intuition behind these methods is to use the non-treated cases—so called “Donors”—to

create a “synthetic control” unit for each treatment unit. The synthetic control unit is developed

by using a weighted combination of the donor pool cases, where the weights are selected so that

they create a synthetic control that closely matches the pre-treatment data pattern of the outcome

variable (in our context, logged WOM) for the treated cases. The synthetic control’s post-

treatment pattern is then used as the counterfactual prediction for the treated cases. Because the

synthetic controls method uses the pre-treatment outcome variable, it naturally conditions on

both observables and unobservables. As the pre-treatment time-series increases in length, the

level of control increases. Thus, the synthetic control approach can account for unobserved

variables that might otherwise invalidate causal inference.

In the GSCM, a parametric model of the treatment effect and data generating process

follows the interactive fixed effects model (see Bai 2009) and is assumed to be

(2) 𝑌𝑖𝑡 = 𝛿𝑖𝑡𝐷𝑖𝑡 + 𝑥𝑖𝑡′ 𝛽 + 𝜆𝑖

′𝑓𝑡 + 휀𝑖𝑡 ,

where

𝐷𝑖𝑡: binary treatment variable for a brand i in a Super Bowl in period t

𝛿𝑖𝑡: The brand-time specific treatment effect

𝑥𝑖𝑡: Fixed effect for every brand/Super Bowl-year and period

𝛽: The vector of common coefficients on the control variables

𝑓𝑡: The unobserved time-varying vector of factors with length F

𝜆𝑖: The brand-specific length F vector of factor loadings

28

휀𝑖𝑡: stochastic error, assumed uncorrelated with the 𝐷𝑖𝑡, 𝑥𝑖𝑡,𝑓𝑡, and 𝜆𝑖

The method requires three further assumptions related to only allowing weak serial dependence

of the error terms, some (standard) regularity conditions, and that the error terms are cross-

sectionally independent and homoscedastic. Given these assumptions, the average treatment

effect on the treated, 𝐴𝑇𝑇𝑡, for the set of 𝑁𝑇𝑟 Super Bowl advertising brands, 𝒯, can be estimated

based on the differences between i’s observed outcome 𝑌𝑖𝑡,𝑖∈𝒯 and the synthetic control for i,

𝑌𝑖𝑡,𝑆𝐶.

(3) 𝐴𝑇𝑇𝑡 =1

𝑁𝑇𝑟∑ [𝑌𝑖𝑡,𝑖∈𝒯 − 𝑌𝑖𝑡,𝑆𝐶]𝑖∈𝒯

Estimation proceeds in three steps. First, we estimate the parameters β, the λ𝑖 vectors for all

donor pool cases, and the vector 𝑓𝑡. These are estimated using only the data from the pre-

treatment period for the donor pool. Second, the factor loadings, λ𝑖 for each of the treated units

are estimated using the pre-treatment outcomes for the treatment cases conditioning on the β and

𝑓𝑡 estimates. Third, the synthetic control for the treated counterfactuals, 𝑌𝑖𝑡,𝑆𝐶, are calculated

using the β and 𝑓𝑡 estimates from step one and the λ𝑖 estimates from step two. This then allows

calculating the 𝐴𝑇𝑇𝑡 for each period. The number of factors, F, is selected via a cross-validation

procedure in which some pre-treatment observations are held back and predicted. The three-step

procedure is used for each number of factors and the number of factors with the lowest mean

squared prediction error is chosen. Inference proceeds using a parametric bootstrap. See Xu

(2017) for details on the procedure and inference.

We implement the procedure using the available software package in R, gsynth. We

estimate the causal effects including two-way fixed effects (time and brand-year). Our standard

errors are clustered at the brand-year level and we use 16,000 samples for bootstrapping the

29

standard errors. We report analyses for both the Keller-Fay total WOM measure and the Nielsen-

McKinsey Insight (NMI)’s online WOM measure. The two datasets overlap from 2008 onward

and so we use this common period to make the analyses comparable. We note that for the Keller-

Fay measures the reported subsample and the full available time period have quite similar effect

sizes and significances.

We report the average treatment effects in Table 4 along with the number of factors used

and the number of pre-periods, post-periods, and total observations. In most cases, the number of

factors reported is the optimal number selected by the cross-validation technique. In the total

WOM cases, the optimal number of unobserved factors was zero suggesting no meaningful

remaining interactive fixed effects in the data. This indicates that the fixed unit and time effects

already control for the unobserved time-varying influences. This finding provides indirect

support for our conditional independence assumption used in the main analysis section. In these

cases with zero optimal factors, we also present solutions where we forced the model to have one

unobserved factor to ensure robustness against more factors.

We begin with the monthly data that most closely approximates our main analysis. We

include the last six months prior to the Super Bowl as pre-treatment periods and consider the

Super Bowl treatment beginning in February (time 0) and continues through March. We find a

significant and positive average causal effect of being a Super Bowl advertiser for the month of

and after the Super Bowl. The average ATT for the two months is 10.8% (s.e.=0.043,p-

value=0.026) with the best fitting number of factors (zero) and 10.3% (s.e.=0.050,p-

value=0.035) with one factor. The ATT for the month of the Super Bowl, February, is estimated

to be 15.9% (s.e.=0.054, p-value<.01) with the optimal zero factors and 15% with one factor

(s.e.=0.062, p-value=0.013), but this effect rapidly declines in later months. Already in March,

30

the effect is insignificant with the ATT estimated to be 6% (s.e. 0.054, p-value=0.246) with zero

factors. Panel A of Figure 3 shows the time-varying estimated ATT for each month of the data,

showing the only individually significant month is the month of the Super Bowl. Thus, the effect

on total WOM caused by being a Super Bowl advertiser is reasonably large, but only lasts

approximately one month.

One major concern with this analysis is that, if the Super Bowl advertiser effect is actually

shorter-lived than one month, monthly data could have an aggregation bias. To examine this, we

conduct the analysis on weekly total WOM measures, which is the finest periodicity the Keller-

Fay dataset allows.10 We use 16 weeks prior to the Super Bowl week as pre-treatment periods,

and a total of 4 weeks of treatment periods including the week of and three weeks after the Super

Bowl. Panel B of Figure 3 shows the weekly pattern of the effects. The week of the Super Bowl

has no increase in total WOM (0.1%, s.e.=0.056), which may not be too surprising since the

Super Bowl airs on the last day of the week. We find the first week following the Super Bowl

has a 22.1% increase (s.e.=0.057, p-value<.01) in total WOM, but that the following weeks have

lower effect sizes of 10.9% (s.e.=0.056, p-value<.061), 14.3% (s.e.=0.058,p-value=0.012), and

10.4% (s.e.=0.058,p-value=0.068) respectively for weeks 2-4. The average ATT across the first

four weeks is estimated to be 11.8% and significant (s.e.=0.033, p-value<.01). While the weekly

data indicate a higher peak of WOM effect in the week following the Super Bowl, the general

patterns do not suggest the monthly data dramatically understate the average effect. In particular,

the effect stays significant for the entire month (4 weeks). Overall, these results indicate that

10 Note that we have access to the Keller-Fay WOM at the weekly level, but not the Kantar advertising data, so that

we are unable to conduct our main panel regression analysis at the weekly level.

31

being a Super Bowl advertiser causes a sizable increase in total WOM of 16% in the first month

of and 22% in the first week after the Super Bowl.

These results speak to the potential aggregation bias in the total WOM data, one possible

source of measurement error. First, the point estimate for the peak weekly effect is less than 50%

larger than that of the monthly average. Second, the estimated ATT for February from the

monthly data has a 95% confidence interval of (5.2%, 26.6%). This interval actually covers the

maximum weekly estimated value, suggesting we cannot statistically distinguish them. These

results suggest that our small result in the main analysis that uses monthly data is unlikely to be

explained away by short-lived total WOM effects. In sum, although there might be aggregation

bias, it appears not to be large enough to overturn the main result for total WOM.

We conduct the same kind of analysis on the online WOM measure in order to examine

whether the Super Bowl effect is larger for online social media posts and whether the effect is

shorter-lived than that of the total WOM. Panels C and D of Figure 3 present the effect patterns.

In the monthly analysis, the measured ATT for the month of the Super Bowl is significant and

26.6% (s.e.=0.039,p-value<.01), and in the month following the Super Bowl, the effect size falls

to be insignificant at 4.9% (s.e.=0.044,p-value=0.240). Thus, the effect does appear to be larger

for online posts than total WOM, but lasts at most one month. Considering weekly data, the ATT

for the week of the Super Bowl is significant and 48.0% (s.e.=0.042,p-value<.01), and the three

weeks after the Super Bowl are all insignificant at 4.1% (s.e.=0.048), 1.8% (s.e.=0.048), and

2.3% (s.e.=0.051), respectively. This analysis suggests that the Super Bowl has a much larger but

shorter effect on counts of online posts than on representative, total WOM mentions.

Because the Nielsen-McKinsey Insight data come daily, we can perform the analysis at this

even more fine-grained level. We use 60 days prior to the Super Bowl as the pre-treatment

32

period. Panel E of Figure 3 indicates that the incremental posts concentrate heavily on the first

few days with significant causal estimates of 67.7% (s.e.=0.062,p-value<.01) for the day of the

Super Bowl, 62.8% (s.e.=0.058,p-value<.01) for the day after, 39.7% (s.e.=0.068,p-value<.01)

for the second day after, 25.2% (s.e.=0.081,p-value<.01) for the third, 12.3% (s.e.=0.084,p-

value=.179) and insignificant for the fourth, and dropping to below 10% and insignificant

thereafter. These causal effects on online posts for the first three days are much larger than the

effects on total WOM measured with a representative sample. This analysis also reaffirms the

concentration of incremental impressions close to the Super Bowl for online posts, which is

distinct from the more spread out effects for total WOM.

How should we interpret these results for the online posts from Nielsen-McKinsey Insight

compared to the total WOM from Keller-Fay? First, the effects for online posts are larger for a

short duration (few days for daily or one week for weekly). In contrast, the effect on the total

WOM persists for approximately the full month. These shorter-term, stronger effects in the

online data might explain why studies that focus entirely on online posts may find larger effects

of advertising on WOM. Second, the monthly periodicity does not appear to produce measurable

aggregation bias for total WOM, since the effect is relatively consistent over the whole month. In

contrast, aggregation bias appears likely to be more severe in the monthly data for online WOM

posts. Daily and weekly effects are much larger than the monthly effects and do not last the full

month. This implies that we should interpret with caution the small effect sizes found in the

monthly data of section 5.1-5.3 for online WOM posts. When combined with the LassoIV results

in Web Appendix C, this suggests that the lift in online posts might be larger than the average

reported, but this larger lift might be relatively short-lived.

33

It is important to keep in mind that the total WOM from Keller-Fay is measured with a

representative sample of the U.S. population and can be interpreted as impressions. In contrast,

the online posts have only a vague connection to impressions with some posts never seen by

anyone and others seen by many people. Moreover, these posts are not collected to be

representative. It is possible that the difference in effects for these two types of WOM could arise

from sampling differences in the data or that the individuals that post online are only a (selected)

subset of those that talk about brands. In either case, for generalizations to earned impressions

that advertising creates, the Keller-Fay data has a stronger foundation.

7 Discussion

Can firms buy earned media impressions with paid media? We conducted an empirical analysis

to evaluate the relationship between advertising expenditures and WOM conversations about

brands. Our dataset contains information on 538 U.S national brands across 16 categories over a

period of 6.5 years and covers both online and total (including online and offline) WOM

mentions. Our main analyses control for news mentions, time lagged WOM, seasonality, secular

trends, brand fixed effects, category-quarter fixed effects, and random coefficients, and checks

robustness against model misspecification. In a second set of analyses, we apply a causal

inference technique, generalized synthetic controls (Xu 2017), on Super Bowl advertisers to

evaluate the possible impact of large, WOM-focused advertising campaigns on WOM. Together,

these analyses present a compelling story. Our main findings include:

1. The relationship between advertising and WOM is positive and significant, but small.

Assuming causality, the average implied elasticity on total WOM is 0.019 for TV advertising,

and 0.014 for Internet advertising and on online WOM posts is 0.009 for TV advertising and

34

0.010 for Internet advertising. Projecting from our sample to the entire US population, for an

average brand in our dataset this implies that a 10% increase in TV advertising leads to 69,000

additional total WOM conversations about the brand per month. This amounts to approximately

0.1% of the paid advertising exposures for the same advertising spend.

2. Cross-brand and cross-category heterogeneity in the advertising-WOM relationship is

significant. The categories with the largest implied elasticities of total WOM to TV advertising

are Sports and Hobbies, Media and Entertainment, and Telecommunications. However, even for

these categories, the average implied elasticity is relatively small, with values between 0.03 and

0.05. Similarly, the “best” brands are estimated to have average elasticities of only around 0.05.

This implies the brands with the most effective brand advertising for total WOM would be

associated with increases in WOM conversations that are less than 1% of the increase in

advertising exposures. Online WOM posts also have small elasticities among the most

responsive categories and brands.

3. Certain events and campaigns are able to achieve higher impact on total WOM. Our

synthetic controls analysis of the Super Bowl advertisers indicates that total WOM mentions

increase 16% in the month of the Super Bowl and 22% in the week after the Super Bowl. This

implies an increase of 10-13% of the average advertising impressions.

4. The Super Bowl advertiser impact on online posts, harvested from the Internet (instead of

using a representative sample) is even higher, but much shorter-lived. The effect of being a

Super Bowl advertiser is a 27% increase in online posts for the month of, 48% for the week of,

and 68% for the day of the Super Bowl.

These results imply that the advertising-WOM relationship is small on average, but that a

larger effect is possible for WOM in certain campaigns both total and online. The effects for

35

online WOM posts may be relatively large, but also short-lived. As a result, one should be

cautious about generalizing the impact of advertising on WOM based only on online post data

collected by crawling the web. Generally, the online posts may signal larger effects than one

should expect for total WOM mentions, where the bulk of brand conversation happens.

What are the managerial implications of our findings? Our findings suggest “there is no free

lunch” when it comes to WOM. Mass TV and Internet display advertising expenditures do not

automatically imply large gains in WOM. More precisely, across 538 brands and many

campaigns per brand over the 6.5-year observation window, high advertising expenditures on

average are not associated with a large increase in total WOM. Similarly, based on our analysis,

no single brand or category appears to generate large average effects. We do find, for Super

Bowl advertisers, where expenditures are very large and the event is a social phenomenon with

the advertisements playing a relatively central role in media attention about the event, the causal

effect on total WOM can be larger, though still modest. However, such successful WOM

campaigns must be relatively rare to still find the average advertising effects to be so small.

Does the small average effect we find imply that investing in advertising to generate WOM is

foolish? Not necessarily. First, our results suggest that online posts might be more responsive to

advertising. Second, if marketers seek to enhance WOM through advertising, they will likely

need to go beyond the typical advertising campaigns contained in our dataset. Our analysis

reveals that managers are unlikely to generate meaningful increases in total WOM unless they

obtain deeper knowledge of which expenditures and campaigns generate WOM on which

channels. We suggest that for managers to pursue the goal of generating WOM from advertising,

they need to be able to track WOM carefully and use methods that can assess the effectiveness of

advertising in generating WOM at a relatively fine-grained level (e.g., campaign or creative).

36

Importantly, because of the disconnect between online posts and total WOM, it is critical to

evaluate total WOM in order to understand whether the more easily tracked online posts translate

into meaningful changes in total mentions.

References

Aaker, David A. and Alexander Biel (2013), Brand equity & advertising: advertising's role in

building strong brands, Psychology Press.

Abadie, Alberto, Alexis Diamond, and Jens Hainmeller (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.

Acemoglu, Daron, Simon Johnson, Amir Kermani, and James Kwak (2016), “The Value of

Connections in Turbulent Times: Evidence from the United States,” Journal of Finanial

Economics, 121, 368-391.

Angrist, Joshua and Jörn Steffen Pischke (2008), Mostly Harmless Econometrics: An

Empiricist’s Companion, Princeton University Press.

Bai, Jushan (2009), “Panel Data Models with Interactive Fixed Effects,” Econometrica, 77(4),

1229-1279.

Bai, Jushan (2013), “Notes and Comments: Fixed Effects Dynamic Panel Models, A Factor

Analytical Method,” Econometrica, 81(1), 285-314.

Baker, Andrew M., Naveen Donthu, and V. Kumar (2016), "Investigating how word-of-mouth

conversations about brands influence purchase and retransmission intentions," Journal of

Marketing Research 53 (2), 225-239.

Batra, Rajeev, David A. Aaker, and John G. Myers (1995), Advertising Management, 5th Edition,

Prentice Hall.

Belloni, Alexandre, Daniel Chen, Victor Chernozhukov, and Christian Hansen (2012), "Sparse

models and methods for optimal instruments with an application to eminent domain,"

Econometrica, 80 (6), 2369-2429.

Berger, Jonah (2014), "Word of mouth and interpersonal communication: A review and

directions for future research," Journal of Consumer Psychology, 24 (4), 586-607.

37

Berger, Jonah and Katherine L. Milkman (2012), "What makes online content viral?" Journal of

marketing research, 49 (2), 192-205.

Bollinger, Bryan, Michael Cohen, and Lai Jiang (2013), "Measuring Asymmetric Persistence and

Interaction Effects of Media Exposures across Platforms," working paper.

Bruce, Norris I., Natasha Zhang Foutz, and Ceren Kolsarici (2012), "Dynamic effectiveness of

advertising and word of mouth in sequential distribution of new products," Journal of

Marketing Research, 49 (4), 469-486.

Campbell, Arthur, Dina Mayzlin, and Jiwoong Shin (2017), "Managing Buzz", The RAND

Journal of Economics, 48 (1), 203-229.

Chen, Yubo and Jinhong Xie (2008), "Online consumer review: Word-of-mouth as a new

element of marketing communication mix," Management Science, 54 (3), 477-491.

Chevalier, Judith A. and Dina Mayzlin (2006), "The effect of word of mouth on sales: Online

book reviews," Journal of marketing research, 43 (3), 345-354.

Danaher, Peter J., and Tracey S. Dagger (2013), "Comparing the relative effectiveness of

advertising channels: A case study of a multimedia blitz campaign," Journal of

Marketing Research, 50 (4), 517-534.

Dichter, Ernest (1966), "How Word-of-mouth Advertising Works," Harvard Business Review,

16 (November-December), 147-166.

Dinner, Isaac M., Harald J. Van Heerde, and Scott A. Neslin (2014), "Driving online and offline

sales: The cross-channel effects of traditional, online display, and paid search

advertising," Journal of Marketing Research, 51 (5), 527-545.

Doudchenko, Nikolay and Guido Imbens (2016), “Balancing, Regression, Difference-in-

Difference and Synthetic Control Methods: A Synthesis,” working Paper.

Draganska, Michaela, Wes Hartmann, and Gena Stanglein (2014), "Internet vs. Television

advertising: A brand-building comparison," Journal of Marketing Research, 51 (5), 578-

590.

Drèze, Xavier and François‐Xavier Hussherr (2003), "Internet advertising: Is anybody

watching?" Journal of Interactive Marketing, 17 (4), 8-23.

Du, Rex, Mingyu Joo, and Kenneth C. Wilbur (forthcoming), “Advertising and Brand Attitudes:

Evidence from 575 Brands over Five Years,” Quantitative Marketing and Economics.

38

Duan, Wenjing, Bin Gu, and Andrew B. Whinston (2008) "The dynamics of online word-of-

mouth and product sales: An empirical investigation of the movie industry," Journal of

Retailing, 84 (2), 233-242.

Feng, Jie and Purushottam Papatla (2011), "Advertising: stimulant or suppressant of online word

of mouth?" Journal of Interactive Marketing, 25 (2), 75-84.

Furrier, John (2013) “Innovative Advertising Is About Integrating TV with Social - Kia Motors

Uses Social Media to Reach the New Audience,” Forbes, Feb-25-2013.

Gelb, Betsy and Madeline Johnson (1995), "Word-of-mouth communication: Causes and

consequences," Marketing Health Services, 15 (3), 54.

Gelper, Sarah, Renana Peres, and Jehoshua Eliashberg (2016), "Talk Bursts: The Role of Spikes

in Pre-release Word-of-Mouth Dynamics," working paper.

Gopinath, Shyam, Pradeep K. Chintagunta, and Sriram Venkataraman (2013), "Blogs,

advertising, and local-market movie box office performance," Management Science, 59

(12), 2635-2654.

Gopinath, Shyam, Jacquelyn S. Thomas, and Lakshman Krishnamurthi (2014), "Investigating the

relationship between the content of online word of mouth, advertising, and brand

performance," Marketing Science, 33 (2), 241-258.

Graham, Jeffrey, and William Havlena (2007), "Finding the “missing link”: Advertising's impact

on word of mouth, web searches, and site visits," Journal of Advertising Research, 47 (4),

427-435.

Hartmann, W. and D. Klapper (2018), “Super Bowl Ads,” Marketing Science, 37(1), 78-96.

Herr, Paul M., Frank R. Kardes, and John Kim (1991), "Effects of word-of-mouth and product-

attribute information on persuasion: An accessibility-diagnosticity perspective," Journal

of Consumer Research, 17 (4), 454-462.

Hewett, Kelly, William Rand, Roland T. Rust, and Harald J. van Heerde (2016), "Brand buzz in

the echoverse," Journal of Marketing, 80 (3), 1-24.

Hogan, John E., Katherine N. Lemon, and Barak Libai (2004), "Quantifying the ripple: Word-of-

mouth and advertising effectiveness," Journal of Advertising Research, 44 (3), 271-280.

Holbrook, Morris B. and Rajeev Batra (1987), "Assessing the role of emotions as mediators of

consumer responses to advertising," Journal of Consumer Research, 14 (3), 404-420.

Keller, Ed and Brad Fay (2009), "The role of advertising in word of mouth," Journal of

Advertising Research, 49 (2), 154-158.

39

Keller, Ed and Brad Fay (2012), The Face-To-Face Book: Why Real Relationships Rule In a

Digital Marketplace, Simon and Schuster.

Liu, Yong (2006), "Word of mouth for movies: Its dynamics and impact on box office

revenue." Journal of Marketing 70 (3), 74-89.

Lovett, Mitchell J., Renana Peres, and Ron Shachar (2013), “On brands and word-of-mouth,”

Journal Marketing Research, 50 (4), 427-444.

Lovett, Mitchell J. and Richard E. Staelin (2016), “The Role of Paid, Earned, and Owned Media

in Building Entertainment Brands: Reminding, Informing, and Enhancing Enjoyment,”

Marketing Science, 35(1), 142-157.

Mangold W. Glynn, Fred Miller, and Gary R. Brockway (1999), “Word-of-mouth

communication in the service marketplace,” Journal of Services Marketing, 13(1), 73-89.

Mitra, Anusree and John G. Lynch Jr. (1995),"Toward a reconciliation of market power and

information theories of advertising effects on price elasticity," Journal of Consumer

Research, 21 (4), 644-659.

Moon, Sangkil, Paul K. Bergey, and Dawn Iacobucci (2010), "Dynamic effects among movie

ratings, movie revenues, and viewer satisfaction," Journal of Marketing, 74 (1), 108-121.

Naik, Prasad A. and Kalyan Raman (2003), "Understanding the impact of synergy in multimedia

communications," Journal of Marketing Research, 40 (4), 375-388.

Nickell, Stephen (1981), "Biases in dynamic models with fixed effects," Econometrica, 49 (6),

1417-1426.

Nielsen (2016), "Stirring up buzz: How TV ads are driving earned media for brands,"

http://www.nielsen.com/us/en/insights/news/2016/stirring-up-buzz-how-tv-ads-are-driving-earned-media-for-brands.html

Olney, Thomas J., Morris B. Holbrook, and Rajeev Batra (1991), "Consumer responses to

advertising: The effects of ad content, emotions, and attitude toward the ad on viewing

time," Journal of Consumer Research, 17 (4), 440-453.

Onishi, Hiroshi and Puneet Manchanda (2012), "Marketing activity, blogging and

sales," International Journal of Research in Marketing, 29 (3), 221-234.

Park, C. Whan, Martin S. Roth, and Philip F. Jacques (1988), "Evaluating the effects of

advertising and sales promotion campaigns," Industrial Marketing Management, 17 (2),

129-140.

40

Pauwels, Koen, Zeynep Aksehirli, and Andrew Lackman (2016), "Like the ad or the brand?

Marketing stimulates different electronic word-of-mouth content to drive online and

offline performance," International Journal of Research in Marketing, 33 (3), 639-655.

Seiler, Stephan, Song Yao, and Wenbo Wang (2017), “Does Online Word-of-Mouth Increase

Demand (and How?): Evidence from a Natural Experiment,” Marketing Science, 36(6),

838-861

Sethuraman, Raj, Gerard J. Tellis, and Richard A. Briesch (2011), "How well does advertising

work? Generalizations from meta-analysis of brand advertising elasticities," Journal of

Marketing Research, 48 (3), 457-471.

Siefert, Caleb J., Ravi Kothuri, Devra B. Jacobs, Brian Levine, Joseph Plummer, and Carl D.

Marci (2009), "Winning the Super “Buzz” Bowl," Journal of Advertising Research, 49

(3), 293-303.

Sinkinson, Michael and Amanda Starc (2017), "Ask your doctor? Direct-to-consumer advertising

of pharmaceuticals," working paper.

Smith, Robert E., and William R. Swinyard (1982), "Information response models: An integrated

approach," The Journal of Marketing, 46 (1), 81-93.

Spotts, Harlan E., Scott C. Purvis, and Sandeep Patnaik (2014),"How digital conversations

reinforce super bowl advertising," Journal of Advertising Research, 54 (4), 454-468.

Stephen, Andrew T. and Jeff Galak (2012), "The effects of traditional and social earned media

on sales: A study of a microlending marketplace," Journal of Marketing Research, 49 (5),

624-639.

Stock, James and Motohiro Yogo (2005), "Testing for weak instruments in linear IV regression,"

in chapter 5 in Identification and inference for econometric models: Essays in honor of

Thomas Rothenberg.

Sundaram, D.S. Kaushik Mitra, and Cynthia Webster (1998), "Word-Of-Mouth

Communications: A Motivational Analysis", in Advances in Consumer Research, Vol. 25,

Joseph W. Alba, and J. Wesley Hutchinson, eds. Provo, UT: Assoc. for Consumer Research,

527-531.

Tirunillai, Seshadri, and Gerard J. Tellis (2012), "Does chatter really matter? Dynamics of user-

generated content and stock performance," Marketing Science, 31 (2), 198-215.

Tirunillai, Seshadri, and Gerard J. Tellis (2017), "Does offline TV advertising affect online

chatter? Quasi-experimental analysis using synthetic control," Marketing Science 36 (6), 862-

878.

41

Trusov, Michael, Randolph E. Bucklin, and Koen Pauwels (2009), "Effects of word-of-mouth

versus traditional marketing: findings from an internet social networking site," Journal of

Marketing, 73 (5), 90-102.

Turner (2016), "Television advertising is a key driver of social media engagement for brands," http://www.4cinsights.com/wp-content/uploads/2016/03/4C_Turner_Research_TV_Drives_Social_Brand_Engagement.pdf

Van der Lans, Ralf and Gerrit van Bruggen (2010), "Viral marketing: What is it, and what are

the components of viral success," in The Connected Customer: The Changing Nature of

Consumer and Business Markets, New York, 257-281. Eds: Stefan Wuyts, Marnik G.

Dekimpe, Els Gijsbrechts, Rik Pieters.

Van der Lans, Ralf, Gerrit van Bruggen, Jehoshua Eliashberg, and Berend Wierenga (2010), "A

viral branching model for predicting the spread of electronic word of mouth," Marketing

Science, 29 (2), 348-365.

Vidal-Berastain, Javier, Paul Ellickson, and Mitchell Lovett (2017), “Understanding the Impact

of Entry of Alternative Retailer Types on Consumer Grocery Shopping Habits,” working

paper.

WOMMA (2014), Return on Word of Mouth. https://womma.org/wp-content/uploads/2015/09/STUDY-WOMMA-

Return-on-WOM-Executive-Summary.pdf

Xu, Yiqing (2017), “Generalized Synthetic Control Method: Causal Inference with Interactive

Fixed Effects Models,” Political Analysis, 25, 57-76.

You, Ya, Gautham G. Vadakkepatt, and Amit M. Joshi (2015), "A meta-analysis of electronic

word-of-mouth elasticity," Journal of Marketing, 79 (2), 19-39.

Zhu, Feng and Xiaoquan Zhang (2010), "Impact of online consumer reviews on sales: The

moderating role of product and consumer characteristics," Journal of Marketing, 74 (2),

133-148.

1

Table 1: Monthly spending on advertising (in thousand dollars) on TV and Internet, and number of news and press mentions (in

thousands) per category.

*NFL=National Football League, MLB= Major Baseball League, NHL = National Hockey League

Category Avg TV $K/mo

Brand with max spending TV

Avg Internet $K/mo

Brand with max spending Internet

Avg news K/mo

Brand with max news

Avg Total WOM K/mo

Brand with max Total WOM

Avg Online WOM K/mo

Brand with max Online WOM

Beauty products

374.7 L’oreal 18.7 L’oreal 0.9 Chanel 0.774 Dove

167.29 Chanel

Beverages 365.0 Pepsi 13.6 Pepsi 1.4 Coca-Cola 1.789 Coca Cola 457.86 Coca-Cola

Cars 1196.1 Ford 112.3 Chevrolet 16.2 GM 1.931 Ford 2256.84 Ford

Children's products

230.5 Mattel 11.2 Lego 0.9 Mattel 0.810 Toys R Us

260.11 Lego

Clothing products

225.5 Lowes 14.4 Lowes 4.1 GAP 1.031 Nike

627.57 Nike

Department stores

699.5 Walmart 67.7 Target 5.8 Walmart 3.235 Walmart

1404.20 Amazon

Financial services

505.8 Geico 210.6 TD Ameritrade

24.7 Bank of America

0.830 Bank of America

287.80 Mastercard

Food and dining 570.3 General Mills

23.2 General Mills

2.4 Banquet 0.948 McDonalds

444.92 McDonalds

Health 946.0 Johnson & Johnson

52.4 Johnson & Johnson

5.0 Pfizer 0.837 Tylenol

344.22 Tylenol

Home design 464.5 Home Depot

35.3 Home Depot 3.6 Home Depot

1.324 Home Depot

433.26 Ikea

Household Products

337.9 Clorox 16.1 Clorox 0.9 P&G 0.846

Tide

86.61 Black & Decker

Media and entertainment

222.6 Time Warner

56.5 Netflix 31.0 Facebook 0.653 American Idol

4234.32 Facebook

Sports and hobbies

36.6 NFL * 14.4 MLB * 188.0 NHL* 1.413 NFL*

3883.39 NFL*

Technology 294.4 Apple 54.7 Microsoft 6.7 Apple 2.052 Apple 2810.69 Apple

Telecom 1074.9 AT&T 107.4 AT&T 28.2 iPhone 2.778 AT&T 3528.49 iPhone

Travel services 163.5 Southwest Airlines

44.4 Expedia 4.7 Holiday Inn

0.791 Southwest Airlines

161.09 American Express

1

Table 2: Main Model with Dependent Variable Ln(WOM).

Variables

Total WOM Online WOM

Population Means Population Means

Estimate Standard

Error Estimate

Standard Error

Ln (Advertising $ TV)+ 0.019 0.0017 ** 0.009 0.001 **

Ln (Advertising $ Internet) + 0.014 0.0021 ** 0.010 0.002 **

Ln (Advertising $ Other) + 0.013 0.0018 ** 0.004 0.002 **

Ln (No of news mentions) 0.103 0.0049 ** 0.138 0.009 **

Ln (WOM(t-1)) 0.167 0.0087 ** 0.429 0.009 **

Ln (WOM(t-2)) 0.075 0.0064 ** 0.039 0.007 **

Brand Fixed Effects? Yes

Yes

Brand Random Coefficients? Yes

Yes

Time Effects? Category-Year-Quarter fixed effects and cubic functions of month of year

Category-Year-Quarter fixed effects and cubic functions of month of year

Heterogeneity Variances Heterogeneity Variances

Estimate

Standard Error

Estimate Standard

Error

Ln (Advertising $ TV) 0.0004 0.0001 ** 0.0002 0.0000 **

Ln (Advertising $ Internet) 0.0008 0.0001 ** 0.0004 0.0001 **

Ln (Advertising $ Other) 0.0003 0.0001 ** 0.0002 0.0001 **

Ln (No of news mentions)

0.0272 0.0026 **

Ln (WOM(t-1)) 0.0250 0.0021 ** 0.0162 0.0016 **

Ln (WOM(t-2)) 0.0078 0.0010 ** 0.0054 0.0008 **

Sample size 40,888 21,689 +

Spending is the log of $1,000’s of dollars plus one per brand per month. * indicates p-value<.05; ** indicates p-value<.01.

2

Table 3. Main Model Results with Super Bowl Interactions

Variables

Total WOM Online WOM

Population Means Population Means

Estimate Standard

Error Estimate

Standard Error

Ln (Advertising $ TV) 0.019 0.002 ** 0.008 0.001 **

Ln (Advertising $ Internet) 0.014 0.002 ** 0.010 0.002 **

Ln (Advertising $ Other) 0.013 0.002 ** 0.004 0.002 **

Super Bowl 0.271 0.117 * 0.065 0.034

Ln(Advertising $ TV)*SuperBowl -0.028 0.017 0.036 0.018 *

Ln (Advertising $ Internet)*SuperBowl 0.004 0.017 0.000 0.018

Ln (Advertising $ Other)(SuperBowl 0.000 0.017 0.027 0.026

Ln (No of news mentions) 0.103 0.005 ** 0.137 0.009 **

Ln (WOM(t-1)) 0.167 0.009 ** 0.428 0.009 **

Ln (WOM(t-2)) 0.075 0.006 ** 0.040 0.007 **

Brand Fixed Effects? Yes Yes

Brand Random Coefficients? Yes

Yes

Time Effects? Category-Year-Quarter fixed effects and cubic functions of month of year

Category-Year-Quarter fixed effects and cubic functions of month of year

Heterogeneity Variances Heterogeneity Variances

Estimate Standard

Error Estimate

Standard Error

Ln (Advertising $ TV) 0.000407 0.000079 ** 0.0002 0.0000 **

Ln (Advertising $ Internet) 0.000807 0.000139 ** 0.0004 0.0001 **

Ln (Advertising $ Other) 0.000328 0.000084 ** 0.0002 0.0001 **

Super Bowl

Ln(Advertising $ TV)*SuperBowl

Ln (Advertising $ Internet)*SuperBowl

Ln (Advertising $ Other)(SuperBowl

Ln (No of news mentions) 0.0271 0.0026 **

Ln (WOM(t-1)) 0.02495 0.002119 ** 0.0162 0.0016 **

Ln (WOM(t-2)) 0.00783 0.001007 ** 0.0054 0.0008 **

Sample size 40,888 21,689

3

Table 4: Average Treatment Effect on the Treated (ATT) for total WOM and for online WOM,

in various time resolutions.

Type of WOM Data Data Frequency

Effect size (ATT.avg)

Standard Error p.value #Factors

#Pre Periods

#Treatment Periods

Overall WOM on a representative sample week 0.1181 0.0335 0.0005 0 16 4

Overall WOM on a representative sample week 0.1047 0.0444 0.0113 1 16 4

Overall WOM on a representative sample month 0.1076 0.0431 0.0124 0 6 2

Overall WOM on a representative sample month 0.1029 0.0505 0.0354 1 6 2

Online Posts week 0.1405 0.0383 0.0003 3 16 4

Online Posts month 0.1574 0.0370 0.0000 1 6 2

Online Posts day 0.1511 0.0875 0.1789 10 60 31

Online Posts day 0.2660 0.0638 0.0000 9 60 8

4

Figure 1: Overview of the literature of advertising and WOM

5

Figure 2a: Effect of advertising on total WOM by product category

Figure 2b. Effect of advertising on Online WOM per product category

-0.020-0.0100.0000.0100.0200.0300.0400.0500.0600.070

Category Estimates TotalTV Internet

-0.040

-0.020

0.000

0.020

0.040

0.060

0.080

Category Estimates OnlineTV Internet

6

Figure 3: Time-varying estimated average treatment effect on the treated (ATT) for total WOM

and online WOM, in various time resolutions.

-0.2

-0.15

-0.1

-0.05

0.05

0.1

0.15

0.2

0.25

0.3

0.35

-8 -6 -4 -2 0 2 4

Ave

rage

est

imat

ed

eff

ect

Period

A. Total WOM monthly

-0.3

-0.2

-0.1

0.1

0.2

0.3

0.4

0.5

-20 -15 -10 -5 0 5 10

Ave

rage

est

imat

ed

eff

ect

Period

B. Total WOM weekly

-0.2

-0.1

0.1

0.2

0.3

0.4

-8 -6 -4 -2 0 2 4

Ave

rage

est

imat

ed

eff

ect

Period

C. Online WOM monthly

-0.4

-0.3

-0.2

-0.1

0.1

0.2

0.3

0.4

0.5

0.6

0.7

-20 -15 -10 -5 0 5 10

Ave

rage

est

imat

ed

eff

ect

Period

D. Online WOM weekly

-0.4

-0.2

0.2

0.4

0.6

0.8

1

-80 -60 -40 -20 0 20 40Ave

rage

est

imat

ed

eff

ect

Period

E. Online WOM daily

ATT

CI.lower

CI.upper

7

Web Appendix A: Data Description

Our brand list contains 538 brands from 16 product categories. The brand list is described in

Table A1. Figure A1 presents time series plots for four representative brands in the dataset. Our

main analysis uses these time series for the full set of brands to evaluate the relationship between

advertising and WOM. These data patterns do not present a clear pattern indicating a strong

advertising-WOM relationship.

----------- Insert Table A1 about here ---------

----------- Insert Figure A1 about here ---------

Table A2 presents descriptive information about the main variables in the study. We have

41,964 brand-month observations. The table indicates that the majority of advertising spending is

on TV advertising. Internet display advertising is, on average, 10% of TV advertising. Brands

greatly differ in their advertising spending. On average, a brand receives 205 news mentions per

month. Some brands (e.g., Windex and Zest) do not receive any mentions in some months, and

the most mentioned brand (Facebook) receives the highest mentions (18,696) in May of 2013. As

for WOM, the average number of monthly total mentions for a brand is 15.8 in our sample,

which translates to an estimated 36.4 million average monthly conversations in the U.S.

population. Table A3 presents similar descriptive information about the main variables, but in

the log scales we use in estimation, along with the correlations across these key variables.

----------- Insert Table A2 about here ---------

----------- Insert Table A3 about here ---------

8

Figure A1: Illustration of the time series data - monthly advertising expenditure on TV and

Internet, number of news mentions, Total and online WOM, for four brands.

9

Table A1: Brand List

Beauty Products Sephora Minute Maid Infiniti Clothing products

Always St. Ives Monster Energy Drink Jaguar Adidas

Arm And Hammer Suave Mountain Dew Jeep Aeropostale

Aveeno Tampax Nestea Jiffy Lube American Eagle

Avon Tresemme Ocean Spray Kia Armani

Axe Zest Patron Tequila Land Rover Banana Republic

Bath & Body Works Beverages Pepsi Lexus Bloomingdales

Caress A And W Root Beer Poland Spring Lincoln Chicos

Chanel Absolut Propel Fitness Water Mazda Coach

Charmin Anheuser Busch Red Bull Mercedes Benz Converse

Clairol Aquafina Sam Adams Mercury Eddie Bauer

Clinique Bacardi Sierra Mist Mini Cooper Foot Locker

Colgate Budweiser Smirnoff Mitsubishi Gap

Covergirl Canada Dry Snapple Nissan Gucci

Crest Captain Morgan Sobe Pep Boys H&M

Degree Coca-Cola Sprite Pontiac Hot Topic

Dial Soap Coors Sunkist Porsche Jcrew

Dove (Personal Care) Coors Light Sunny Delight Saab Kohls

Estee Lauder Corona Tropicana Subaru Lane Bryant

Garnier Fructis Crystal Light Welch Suzuki Levis

Gillette Dasani Water Cars Toyota Louis Vuitton

Head & Shoulders Diet Mountain Dew Acura Volkswagen Lowes

Herbal Essence Diet Pepsi Audi Volvo Marshalls

Irish Spring Dr Pepper Autozone Yamaha New Balance

Ivory Fanta BMW Children's Products Nike

Jergens Fresca Buick Carters Nordstrom

Kotex Gatorade Cadillac Enfamil Old Navy

Lancome Grey Goose Chevrolet Fisher Price Pac Sun

Listerine Guinness Chrysler Gerber Payless

Loreal Heineken Dodge Leapfrog Polo

Mary Kay Jack Daniels ExxonMobil Lego Prada

Maybelline Jose Cuervo Firestone Little Tikes Ralph Lauren

Neutrogena Juicy Juice Ford Luvs Reebok

Nivea Koolaid GM Mattel TJ Maxx

Old Spice Lipton GMC Oshkosh Tommy Hilfiger

Pantene Maxwell House Good Year Tires Pampers Under Armour

Playtex Michelob Harley Davidson Playskool Wilson

Revlon Mikes Hard Lemonade Honda Toys R Us

Scott Tissue Miller Brewing Hyundai

Secret Miller Lite Infiniti

10

Department Stores Td Ameritrade Marie Callender Velveeta Household Products

Barnes & Noble Trowe Price Mcdonalds Wegmans Cascade

BJs USAA Nabisco Whole Foods Cheer

Borders Vanguard Nestle Winn Dixie Clorox

Costco Visa Olive Garden Yoplait Dawn

Kmart Wachovia Oreos Health Downy

Meijer Wells Fargo Oscar Mayer Advil Febreze

Office Depot Food And Dining Outback Steakhouse Aetna Gain

Sams Club Albertsons Panera Aleve Hoover

Sears Applebees Papa Johns Band Aid Kitchen Aid

Staples Arbys Perdue Chicken Bayer Lysol

Target Banquet PF Chang Benadryl Mr Clean

Walmart Butterball Pillsbury Blue Cross/Blue Shield P&G

Financial Services Campbell Pizza Hut Cigna Palmolive

AIG Cracker Barrel Popeyes Claritin Pine Sol

Allstate Dannon Prego CVS Pledge

American Express Del Monte Publix Excedrin Purex

Bank Of America Dennys Quaker Oats GNC Swiffer

BB&T Bank Digiorno Quiznos Johnson & Johnson Tide

Capital One Dole Ragu Kaiser Permanente Windex

Charles Schwab Dominos Pizza Ralphs Grocery Lipitor Media & Entertainment

Citibank Doritos Red Lobster Merck 24tvshow

Discover Card Dunkin Donuts Red Robin Pfizer ABC

Dow Jones Frito Lay Romanos Macaroni Grill Prilosec Amazing Race

Edward Jones General Mills Ruby Tuesday Rite Aid American Idol

Etrade Giant Eagle Safeway Tylenol America's Next Top Model

Fidelity Investments Giant Food Sara Lee Walgreens BBC

Fifth Third Bank Healthy Choice Shaw's Supermarket Home Design Bet

Geico Heinz Slim Fast GE Big Brother

H&R Block Hershey Snickers Home Depot Blockbuster

HSBC Hot/Lean Pockets Sonic Ikea Cartoon Network

Ing Direct Ihop Starbucks Kenmore CBS

Mastercard Jack In The Box Stouffers La-Z-Boy CNBC

Merrill Lynch Jello Subway Maytag CNN

Metlife Kelloggs Swansons Pier 1 Imports Comedy Central

Morgan Stanley KFC Taco Bell Whirlpool CSI

Prudential Kraft Texas Roadhouse Dancing With The Stars

Regions Bank Kroger TGI Fridays Deal Or No Deal

Smith Barney Lays Chips Tostitos Desperate Housewives

Suntrust Long John Silvers Tyson DirectTV

11

Discovery Channel PBS YMCA Wii Fit Travel Services

E! People Magazine Technology Xbox Alamo

Ebay Prison Break Acer Xbox 360 Alaska Air

ESPN Scrubs Apple Zune American Airlines

Everybody Loves Raymond Shrek (Movie) Best Buy Telecommunications Amtrak

Facebook Simpsons Bose AOL Best Western

Family Guy Sirius Brother AT&T British Airways

Food Network Smallville Canon Blackberry Budget Car Rental

Fox South Park Circuit City Boost Mobile Carnival Cruise Lines

Fox News Spongebob Squarepants Compaq Charter Communications Comfort Inn

Friends Survivor Dell Cox Continental Airlines

Fringe (TV Show) The Office Fuji Dish Network Days Inn

General Hospital Time Warner Garmin Iphone Delta Airlines

Google TNT Gateway Computer Motorola Enterprise Car Rental

Greys Anatomy Two And A Half Men Halo (Video Game) Nokia Expedia

Hallmark Ugly Betty HP Qwest Frontier Airlines

Harry Potter VH1 iPod Road Runner Hampton Inn

HBO Wall Street Journal iTunes TMobile Hertz

Heroes (TV Show) Wheel Of Fortune Kodak Virgin Mobile Holiday Inn

House (TV Show) Yahoo Lexmark Vonage Hyatt

Incredible Hulk (Movie) You Tube LG Jet Blue

Indiana Jones (Movie) Sports and Hobbies Microsoft Marriott

Iron Man (Movie) Atlanta Braves Nikon Orbitz

Jeopardy Boston Celtics Nintendo Priceline.Com

Law And Order Boston Red Sox Palm/Treo Royal Caribbean Cruises

Lifetime Television Curves Panasonic Sheraton Hotels

Lost La Lakers Pioneer Southwest Airlines

Money Magazine MLB Playstation 3 Travelocity

MSN Nascar Radio Shack United Airlines

MSNBC NBA RCA US Air

Mtv New England Patriots Samsung

Myspace.Com NFL Sandisk

NBC NHL Sanyo

Ncis NY Giants Sharp

Netflix NY Mets Sony Playstation

Nickelodeon NY Yankees Super Mario Brothers (Video Game)

NY Times Pittsburgh Steelers Tivo

Oprah WWE Wii

12

Table A2: Summary statistics of main variables

Table A3: Summary statistics and correlations of main variables

Variable/per brand per month Descriptive Statistics

average std.dev min max

Advertising Expenditures

K$ TV spending 5890.46 12,743.14 0 153,886.6

K$ Internet spending 665.34 2002.29 0 47,928.3

K$ Other ad spending 2828.46 6124.37 0 105,786.9

News Mentions

News mentions 205.31 778.21 0 18696

WOM

WOM total mentions 15.81 31.11 0 394

WOM online mentions (K posts) 35.9 191.2 0 6,264.3

Variable/per brand per month Descriptive Statistics Correlations

average std.dev min max 1 2 3 4 5 6

Advertising Expenditures

1 Ln (K$ TV spending) 5.407 3.825 0 11.94 1 0.56 0.56 0.06 0.42 0.11

2 Ln (K$ Internet spending) 3.777 2.781 0 10.78 0.56 1 0.60 0.31 0.39 0.30

3 Ln (K$ Other ad spending) 5.532 3.171 0 11.57 0.56 0.60 1 0.23 0.33 0.16

News Mentions

4 Ln (News mentions) 3.322 1.909 0 9.84 0.06 0.31 0.23 1 0.26 0.55

WOM

5 Ln (WOM total mentions) 2.142 1.088 0 5.98 0.42 0.39 0.33 0.26 1 0.39

6 Ln (WOM online mentions) 8.511 2.011 0 15.65 0.10 0.30 0.16 0.55 0.40 1

13

Web Appendix B: Model with lagged advertising

In section 4, we presented the model of our main analysis to discover the relationship between

advertising and WOM. In this appendix, we consider a model that allows the effects of

advertising to carry over to the future months. This helps us understand the long-term

relationship between advertising expenditure and WOM. For a given brand j and month t, the

model is defined as

(B1) log(𝑊𝑂𝑀)𝑗𝑡 = 𝛼𝑗 + 𝛼𝑐𝑞 +∑ 𝛽1𝑗,𝜏𝑙𝑜𝑔(𝐴𝑑𝑇𝑉)𝑗𝑡−𝜏𝐿𝜏=0 + ∑ 𝛽2𝑗,𝜏log(𝐴𝑑𝐼𝑛𝑡𝑒𝑟𝑛𝑒𝑡)𝑗𝑡−𝜏

𝐿𝜏=0

+𝛾1𝑗log(𝑊𝑂𝑀)𝑗𝑡−1 + 𝛾2𝑗log(𝑊𝑂𝑀)𝑗𝑡−2 + 𝑋𝑗𝑡𝛽0𝑗 + 휀𝑗𝑡

where the number of lags L included in the empirical analysis is chosen based on model fit

statistics (e.g., AIC, BIC) and the likelihood ratio test.

Table B1 reports that the optimal choice is L=1 for total WOM and L=3 for online WOM.

Note that all control variables in equation (B1) remain the same as equation (1) except that the

log dollars in other media advertising expenditure in 𝑋𝑗𝑡 here include both contemporary and

lagged terms, i.e., ∑ 𝛽0𝑗,𝜏𝑙𝑜𝑔(𝐴𝑑𝑂𝑡ℎ𝑒𝑟)𝑗𝑡−𝜏𝐿𝜏=0 .

----------- Insert Table B1 about here ---------

We present the results of the above model for both total WOM and online WOM in Table

B2. The contemporary effects of advertising expenditure are all positive, significant and remain

similar sizes as the results reported in Table 2. For total WOM, there is a positive and significant

relationship between TV advertising expenditure and total WOM in the next month (1 lag),

whereas the relationship does not have significant lags for Internet display advertising. Given

these estimates, we calculated the cumulative relationship of advertising on WOM: 0.031 for TV

14

advertising expenditure and 0.020 for Internet display advertising. The second set of columns in

Table B2 suggests that TV advertising expenditures have a longer-term relationship with online

WOM, lasting three months, but again none of the lagged Internet display advertising

coefficients are statistically significant. The cumulative effects on online WOM are 0.017 for TV

advertising and 0.013 for Internet display advertising. These findings imply that Internet display

advertising expenditures only have a short-term impact on total WOM and online social media

posts; while TV advertising expenditures seem to have a longer-lasting impact. However, the

total effects for both total and online WOM for both TV and Internet display advertising

expenditures are still small.

----------- Insert Table B2 about here ---------

Table B1: Model Selection for Lagged Advertising Models

Total WOM Online WOM

2 lags 1 lag 4 lags 3 lags 2 lags 1 lag

-2 Log Likelihood 53256.5 53259.8 4376.1 4386.6 4415.9 4459.3

AIC (Smaller is Better) 55150.5 55145.8 4868.1 4866.6 4883.9 4921.3

AICC (Smaller is Better) 55195.4 55190.3 4874.4 4872.5 4889.6 4926.9

BIC (Smaller is Better) 59211.1 59189.2 5918.7 5891.6 5883.3 5907.9

15

Table B2: Lagged Advertising Model.

Variables

Total WOM Online WOM

Population Means Population Means

Estimate StdErr Estimate StdErr

Ln (Advertising $ TV) 0.016 0.002 ** 0.009 0.001 **

Ln (Advertising $ TV)(t-1) 0.008 0.002 ** 0.000 0.001

Ln (Advertising $ TV)(t-2) -0.004 0.001 **

Ln (Advertising $ TV)(t-3) 0.003 0.001 *

Ln (Advertising $ Internet) 0.012 0.002 ** 0.011 0.002 **

Ln (Advertising $ Internet)(t-1) 0.003 0.002 -0.004 0.002

Ln (Advertising $ Internet)(t-2) 0.000 0.002

Ln (Advertising $ Internet)(t-3) -0.001 0.002

Ln (Advertising $ Other) 0.010 0.002 ** 0.004 0.002 *

Ln (Advertising $ Other)(t-1) 0.005 0.002 ** 0.003 0.001 *

Ln (Advertising $ Other)(t-2) -0.006 0.002 **

Ln (Advertising $ Other)(t-3) 0.001 0.002

Ln (No of news mentions) 0.102 0.005 ** 0.133 0.009 **

Ln (WOM(t-1)) 0.162 0.009 ** 0.444 0.009 **

Ln (WOM(t-2)) 0.074 0.006 ** 0.059 0.008 **

Brand Fixed Effects + Rand Coeff. Yes Yes

Time Effects? Category-Year-Quarter, cubic functions of month of year

Category-Year-Quarter, cubic functions of month of year

Heterogeneity Variances Heterogeneity Variances

Estimate StdErr Estimate StdErr

Ln (Advertising $ TV) 0.0004 0.0001 ** 0.0002 0.0000 **

Ln (Advertising $ TV)(t-1) 0.0002 0.0001 ** 0.0001 0.0000 *

Ln (Advertising $ TV)(t-2) 0.0000

Ln (Advertising $ TV)(t-3) 0.0000 0.0000

Ln (Advertising $ Internet) 0.0008 0.0001 ** 0.0003 0.0001 **

Ln (Advertising $ Internet)(t-1) 0.0002 0.0001 *

Ln (Advertising $ Internet)(t-2) 0.0000

Ln (Advertising $ Internet)(t-3) 0.0003 0.0001 ** 0.0002 0.0001 **

Ln (Advertising $ Other) 0.0000 0.0001 ** 0.0001 0.0001 *

Ln (Advertising $ Other)(t-1) 0.0000

Ln (Advertising $ Other)(t-2) 0.0002 0.0001 **

Ln (Advertising $ Other)(t-3) 0.0001 0.0001 *

Ln (No of news mentions) 0.0261 0.0025 **

Ln (WOM(t-1)) 0.0244 0.0021 ** 0.0159 0.0017 **

Ln (WOM(t-2)) 0.0080 0.0010 ** 0.0071 0.0010 **

Sample size 40,350 20,102

16

Web Appendix C: Robustness checks

In this section, we provide evidence on the robustness of our main analysis results to different

model specifications and to potential remaining endogeneity concerns. To illustrate the

robustness of our results presented in Table 2, Tables C1 and C2 present six model specifications

that delete or adjust variables or model components from the model of Equation (1) for total

WOM and online WOM, respectively. In Model 1a, we only include the advertising variables,

the news mentions, and the time and seasonality controls (i.e., no lagged dependent variable, no

brand fixed effects, no brand random coefficients). This model with very limited controls

produces implied elasticities that are larger (0.083 for TV and 0.064 for Internet for the total and

0.057 and 0.074 for the online), noting that the TV elasticity on online WOM is an order of

magnitude larger than of Table 2. However, without the additional controls, these estimates are

likely to be spurious. Model 1b adds to Model 1a the two lags of Ln(WOM). In this model, the

estimated elasticities are already quite small (0.017 for TV and 0.009 for Internet for total WOM

and 0.005 and 0.004 for online WOM). Model 1c adds brand fixed effects to the model and we

find that the implied elasticities actually grow slightly (0.018 for TV and 0.016 for Internet for

total WOM, and 0.009 and 0.01 for online WOM). Model 1d deletes the News Mentions variable

from the main model of Table 2. Again, we find the remaining coefficients are quite similar in

size and significance. In Model 1e, we replace the fixed effects with first differences. In model

1f, we include time effects for each month in the data instead of cubic trends of month of year. In

model 1g, we drop the lagged Ln(WOM) and include the brand fixed effects. Looking across

these specifications, the implied advertising elasticities appear to be consistently small and of the

relative magnitude reported in Table 2, whenever reasonable controls are included. Further, the

main controls that are important are brand fixed effects (or first differences) and the lagged

Ln(WOM).

17

--------- Insert Table C1 about here ---------

--------- Insert Table C2 about here ---------

Although, as noted, we include controls for the main endogeneity concerns, one might

remain concerned that brand managers anticipate some specific shocks to WOM and also plan in

advance their advertising around those anticipated shocks or that there is measurement error in

our advertising expenditure measures. To examine whether our results are biased due to any such

remaining endogeneity between, for example, TV advertising and the unobserved term in the

regression, we apply a two-stage least squares analysis. This analysis is applied to the model

without random coefficients. As instruments, we use average national advertising costs per

advertising unit obtained from Kantar Media’s Ad$pender data. The argument for validity of the

instrument comes from a supply-side argument that advertisers respond to advertising costs, and

the exclusion here is that no single brand sets the price of advertising that prevails in the market.

We were able to obtain these per unit costs for TV, magazines, and newspapers. For Internet

display advertising, we include total political Internet display advertising expenditures. Here, we

follow the argument made by Sinkinson and Starc (2017) that political advertising can crowd out

commercial advertisers. We interact these cost and political advertising variables with the brand

indicators, producing 2152=538*4 instruments.

We found that the Cragg-Donald statistic for this full set of instruments (3.24 with 3

endogenous variables) failed to achieve the minimum thresholds suggested by Stock and Yogo

(2005), suggesting these are weak instruments when all are included. Further, the first stage

regression coefficients were counterintuitive for the total political Internet advertising variables,

for example, with many brands having higher Internet display advertising expenditures when

18

political advertising expenditures were larger. These results suggest that we should interpret this

analysis using the full set of instruments with caution since it could produce biased estimates due

to weak instruments that are not operating as theoretically predicted.

In principle, the indication of weakness and biasing could arise because we include many

potentially weak instruments that may not be helpful (Angrist and Pischke 2009). To examine

this, we also estimated the model via two-stage least squares where we use a post-LASSO

technique in the first stage to select the optimal instruments for each of the three endogenous

variables (Belloni et al., 2012). If a subset of all instruments is strong, then this analysis would

select the optimal set of instruments. The post-LASSO procedure “deselects” between 56% and

66% of the instruments depending on the dependent variable so that the union of these Lasso

selections, the set used in the IV estimates, includes 67% of all instruments.11 Recall that our

arguments for these instruments imply negative overall effects for the variables. Of the

coefficients on the instruments that are retained in the first stage, for TV, Internet, and other

advertising expenditures, 65%, 59%, and 65%, respectively, are negatively signed. This suggests

that a meaningful proportion of the coefficients take positive signs that are counter to the

theoretical direction of effect for the instrument’s exogeneity argument. This doesn’t necessarily

invalidate the use of the instruments, since the positive coefficients could arise from

approximation error and correlations with other variables. However, to be cautious, we treat

these analyses as robustness tests and present as our main results the ones requiring a conditional

independence assumption.

11 Following Angrist and Pischke (2009), we also explored results from the LIML estimator. The estimated values

were quite similar to the LASSO-IV estimates. We also examined the instruments without the brand interactions and

this similarly produced weak instrument results and unreasonable first stage estimates.

19

Table C3 presents the focal coefficient results of the two different instrumental variables

analyses for the total and online WOM. For total WOM (models IV1 & IV2), the results for the

full set of IVs are significant, but insignificant for the LassoIV analysis. The estimates for

advertising expenditures still have effect sizes that are quite small with the largest being 0.026

for Internet display advertising. Although this estimate is larger than our main estimates reported

in Table 2, the standard errors clearly cover the previous main results, and the more uncertain

LassoIV results have implied 95% confidence intervals that do not allow values above 0.10 for

the advertising variables. This suggests that the potential endogeneity bias should not be too

severe and barely increases the point estimates at all.

For the online WOM posts (models IV3 & IV4), the results are very similar to the main

results when all IVs are included. The TV and Internet advertising expenditure variables are

significant and positive, but quite small and the other advertising expenditures are small and

insignificant. The LassoIV, however, has point estimates that are much larger with a significant

result for TV advertising expenditures. In this case, the advertising effects appear to be larger

when accounting for the IV analysis. When considering the full set of our results, these results

are consistent with online WOM response to advertising being larger than total WOM and

potentially meaningfully large even in the average effects, not just Super Bowl advertising.

--------- Insert Table C3 about here ----------

20

Table C1: Nested Models with dependent variable, Ln(Total WOM). Monthly data. N=40,888.

Model1a Model1b Model1c Model1d

Variable Estimate SE Estimate SE Estimate SE Estimate SE

Intercept 0.786 0.107 ** 0.256 0.0627 **

Ln (Advertising $ TV)+ 0.083 0.002 ** 0.017 0.0010 ** 0.018 0.0012 ** 0.020 0.0012 **

Ln (Advertising $ Internet)+ 0.064 0.002 ** 0.009 0.0014 ** 0.016 0.0016 ** 0.018 0.0017 **

Ln (Advertising $ Other)+ 0.008 0.002 ** 0.002 0.0012 0.014 0.0015 ** 0.017 0.0015 **

Ln (#of news mentions) 0.182 0.003 ** 0.033 0.0020 ** 0.129 0.0049 **

Ln (WOM(t-1)) 0.482 0.0046 ** 0.253 0.0049 ** 0.266 0.0049 **

Ln (WOM(t-2)) 0.365 0.0046 ** 0.151 0.0048 ** 0.159 0.0048 **

Brand Fixed Effects? No No Yes Yes

Brand Random Coeff.? No No No No

Time Effects? Category-Year-Quarter fixed effects and cubic functions

of month of year

Category-Year-Quarter fixed effects and cubic functions of

month of year

Category-Year-Quarter fixed effects and cubic

functions of month of year

Category-Year-Quarter fixed effects and cubic functions

of month of year

Model1e Model1f Model1g

Variable Estimate SE Estimate SE Estimate SE

Intercept

Ln (Advertising $ TV)+ 0.017 0.0010 ** 0.018 0.0012 ** 0.022 0.0013 **

Ln (Advertising $ Internet)+ 0.010 0.0014 ** 0.016 0.0016 ** 0.022 0.0018 **

Ln (Advertising $ Other)+ 0.002 0.0012 0.014 0.0015 ** 0.019 0.0016 **

Ln (#of news mentions) 0.033 0.0020 ** 0.128 0.0049 ** 0.180 0.0052 **

Ln (WOM(t-1)) 0.481 0.0046 ** 0.256 0.0049 **

Ln (WOM(t-2)) 0.365 0.0046 ** 0.149 0.0048 **

Brand Fixed Effects? First Differences Yes Yes

Brand Random Coeff.? No No No

Time Effects? Category-Year-Quarter fixed effects and cubic functions

of month of year

Category-Year-Quarter fixed effects and month of year

fixed effects

Category-Year-Quarter fixed effects and cubic

functions of month of year

+Spending is the log of $1,000’s of dollars plus one per brand per month. * indicates p-value<.05; ** indicates p-value<.01.

21

Table C2: Nested Models with dependent variable, Ln(Online WOM). Monthly data. N=21,689. +

Spending is the log of $1,000’s of dollars plus one per brand per month. * indicates p-value<.05; ** indicates p-value<.01.

Model1a Model1b Model1c Model1d

Variable Estimate SE Estimate SE Estimate SE Estimate SE Estimate

Intercept 5.221 0.206 ** 0.435 0.0569 **

Ln (Advertising $ TV)+ 0.057 0.003 ** 0.005 0.0009 ** 0.009 0.0012 ** 0.011 0.0012 **

Ln (Advertising $ Internet)+ 0.074 0.005 ** 0.004 0.0013 ** 0.010 0.0016 ** 0.012 0.0016 **

Ln (Advertising $ Other)+ 0.038 0.004 ** 0.001 0.0011 0.003 0.0014 * 0.005 0.0014 **

Ln (#of news mentions) 0.427 0.006 ** 0.028 0.0019 ** 0.114 0.0048 **

Ln (WOM(t-1)) 0.758 0.0066 ** 0.563 0.0067 ** 0.576 0.0068 **

Ln (WOM(t-2)) 0.199 0.0065 ** 0.048 0.0065 ** 0.049 0.0066 **

Brand Fixed Effects? No No Yes Yes

Brand Random Coeff.? No No No No

Time Effects? Category-Year-Quarter fixed effects and cubic

functions of month of year

Category-Year-Quarter fixed effects and cubic

functions of month of year

Category-Year-Quarter fixed effects and cubic functions of

month of year

Category-Year-Quarter fixed effects and cubic functions of month of

year

Model1e Model1f Model1g

Variable Estimate SE Estimate SE Estimate SE

Intercept

Ln (Advertising $ TV)+ 0.006 0.0009 ** 0.010 0.0012 ** 0.015 0.0016 **

Ln (Advertising $ Internet)+ 0.005 0.0013 ** 0.009 0.0015 ** 0.017 0.0021 **

Ln (Advertising $ Other)+ 0.000 0.0011 0.004 0.0014 ** 0.007 0.0019 **

Ln (#of news mentions) 0.030 0.0020 ** 0.113 0.0047 ** 0.177 0.0064 **

Ln (WOM(t-1)) 0.743 0.0068 ** 0.602 0.0067 **

Ln (WOM(t-2)) 0.211 0.0068 ** 0.014 0.0065 *

Brand Fixed Effects? First Differences Yes Yes

Brand Random Coeff.? No No No

Time Effects? Category-Year-Quarter fixed effects and cubic

functions of month of year

Category-Year-Quarter fixed effects and cubic

functions of month of year

Category-Year-Quarter fixed effects and cubic functions of

month of year

22

Table C3: Instrumental variables analysis for the model estimated in Table 2 (Eq 1)

Total WOM IV1: 2SLS-All Instruments IV2: 2SLS - LASSO-IV

Variable Estimate Std. Error

Estimate Std. Error

Intercept

Ln (Advertising $ TV) 0.020 0.0027 ** 0.021 0.0334

Ln (Advertising $ Internet) 0.026 0.0040 ** 0.017 0.0425

Ln (Advertising $ Other) 0.009 0.0036 * 0.000 0.0308

Controls? News mentions; Two lags of WOM News mentions; Two lags of WOM

Brand Fixed Effects? Yes Yes

Brand Random Coefficients No No

Time Effects? Category-Year-Quarter fixed effects

and cubic functions of month of year

Category-Year-Quarter fixed effects and cubic functions of month of year

Sample size 40,888 40,888

Online WOM IV3: 2SLS-All Instruments IV4: 2SLS - LASSO-IV

Variable Estimate Std. Error Estimate Std. Error

Intercept

Ln (Advertising $ TV) 0.008 0.0024 ** 0.124 0.0414 ** Ln (Advertising $ Internet) 0.010 0.0035 ** 0.050 0.0453

Ln (Advertising $ Other) 0.006 0.0031 0.046 0.0398

Controls? News mentions; Two lags of WOM News mentions; Two lags of WOM

Brand Fixed Effects? Yes Yes

Brand Random Coefficients No No

Time Effects? Category-Year-Quarter fixed effects

and cubic functions of month of year

Category-Year-Quarter fixed effects and cubic functions of month of year

Sample size 21,689 21,689 +Spending is the log of $1,000’s of dollars plus one per brand per month. * indicates p-value<.05; ** indicates p-value<.01.

23

Web Appendix D: WOM Mentioning Advertisements

In this appendix, we provide an analysis of WOM that specifically mentions advertisements. The

Keller-Fay TalkTrack dataset includes information about whether a brand mention references

advertising. Out of all the brand mentions a respondent provides, 10 are randomly selected for

the respondent to provide additional information about the conversation surrounding the brand

mention. Specifically, respondents were asked to indicate whether the conversation included a

reference to media or marketing about the brand. The exact question, “Did anyone in the

conversation refer to something about the brand from any of these sources?” used a multi-select

format allowing up to two answers. The response categories include TV advertisements and

Internet advertisements as options. We use this item to count the number of cases in which the

brand conversation referred to an ad.

Figure D1 presents the percentage of brand conversations that reference ads for each brand

(which we refer to as ad-WOM), including WOM with references to TV (top panel) and Internet

(bottom panel) ads. First, the distribution suggests that a meaningful proportion of all brand

conversations contain references to advertising. Unsurprisingly, far more of the conversations

contain mentions of TV ads than Internet ads. For most brands, TV ads are referenced between

6% and 14% of the time, whereas Internet ads are only referenced between 2% and 6% of all

conversations. Both distributions are skewed right, so that there are some brands for which

advertising is referenced quite frequently during conversations.

-------- Insert Figure D1 about here -------

24

In our main analysis, we pooled all WOM together, which could cover up a stronger

relationship between advertising expenditures and the number of brand conversations that

reference ads. To test whether this is the case, we estimate the same model but use as the

dependent variable (and lagged dependent variables) the WOM that references either TV ads or

Internet ads (ad-WOM).

The results of the two analyses are presented in Table D1. The TV ad-WOM analysis

reveals that advertising coefficients have a similar magnitude and significance as those presented

in the main analysis in Table 2. The relationship between the advertising variables and the

Internet ad-WOM is estimated to be smaller than that found for the total WOM measures. Recall

that the construction of the ad-WOM measure differs from that of the total WOM so that a direct

comparison of the estimates is not possible. Yet we can conclude that these results provide no

evidence that the advertising-WOM relationship is meaningfully stronger when considering only

WOM that discusses advertising. Hence, although many brand conversations talk about the ads,

the advertising did not necessarily “cause” the conversation about the ads. Instead, the

advertising becomes part of the existing conversations that would have happened anyway.

------- Insert Table D1 about here --------

25

Table D1: Models with ad-WOM as dependent variable. Monthly data. N=40,888.

DV: TV ad-WOM DV: Internet ad-WOM

Description Estimate Standard Error

Estimate Standard Error

Ln (Advertising $ TV) + 0.011 0.0014 ** 0.003 0.0010 **

Ln (Advertising $ Internet) + 0.010 0.0018 ** 0.003 0.0012 *

Ln (Advertising $ Other) + 0.007 0.0018 ** 0.004 0.0012 **

Ln (#of news mentions) 0.057 0.0074 ** 0.027 0.0044 **

Ln (DV(t-1)) 0.033 0.0070 ** 0.001 0.0066

Ln (DV(t-2)) 0.000 0.0058

-0.011 0.0063 **

Month of Year -0.079 0.0785

0.031 0.0610

(Month of Year)2 0.057 0.1372

-0.050 0.1066

(Month of Year)3 0.000 0.0697

0.032 0.0542

Year 0.626 0.1725 ** 0.264 0.1337 *

Year2 -0.500 0.4877 -0.066 0.3776 Year3 -0.090 0.4107 -0.159 0.3181

Brand Fixed Effects? Yes Yes

Brand Random Coefficients? Yes Yes +Spending is the log of $1,000’s of dollars plus one per brand per month. The heterogeneity variances are similar in size and significance to

those in Table 2. * indicates p-value<.05; ** indicates p-value<.01.

26

Figure D1: Percentage of WOM conversations mentioning TV advertising (top panel) or Internet

advertising (bottom panel).