Informational and Non-Informational Advertising Content

∗

Current version: October 2019

Yi-Lin Tsai

†

Elisabeth Honka

‡

Abstract

We investigate the relationship between both advertising content and quantity and several

stages of consumers’ decision-making, namely, unaided and aided awareness, considera-

tion, and purchase. Understanding how the amount and content of advertisements aﬀect

consumers’ decision-making is crucial for companies to eﬀectively and eﬃciently use their

advertising budgets. Spanning a time period from 2010 to 2016, we combine a unique

data set on TV advertising content and quantities with individual-level data containing

information on purchases, consideration and awareness sets, demographic variables, and

perceived prices. Our results reveal that advertising quantity signiﬁcantly increases con-

sumer (unaided and aided) awareness, but has no eﬀect on conditional consideration and

conditional purchase. However, when measuring the separate eﬀects of diﬀerent types of

advertising content, we ﬁnd a more nuanced set of results: advertising only containing

non-informational content increases unaided awareness, while advertising only containing

informational content increases aided awareness and purchase conditional on consider-

ation. And lastly, since many companies spend a signiﬁcant portion of their budgets

on advertising with both informational and non-informational content, we investigate

whether this type of advertising signiﬁcantly aﬀects certain groups of consumers and ﬁnd

it to indeed be the case. We conclude that the eﬀects of advertising content vary along

consumers’ purchase funnel.

Keywords: Advertising, Advertising Content, Purchase Funnel, Auto Insurance Industry

JEL Classiﬁcation: D83, G22, M37

∗

We thank J.D. Power and Associates for supplying us with data. We would like to especially thank Brett

Hollenback, Stephen Spiller, and Stephan Seiler. We thank Pradeep Chintagunta, Xiaojing Dong, G¨unter

Hitsch, Sylvia Hristakeva, Robert Kent, Jura Liaukonyte, Carl Mela, David Muir, Thomas Otter, Brad Shapiro,

Matt Shum, Jeremy Tobacman, Ken Wilbur, and Nathan Wilson, seminar participants at Duke University,

INSEAD, Rotterdam School of Economics, University of Michigan, and Tulane University, our discussants

Pedro Gardete, Tatsuo Tanaka, and Zsolt Sandor, and the participants of the 2017 International Industrial

Organization Conference (IIOC), the 8th Workshop on Search and Switching Costs, the 2017 Boulder Summer

Conference on Consumer Financial Decision Making, the 2017 Marketing Science/INFORMS conference, the

2018 UT Dallas FORMS conference, the 2018 Marketing Dynamics conference, and the 11th Workshop on the

Economics of Advertising and Marketing for their comments. All errors are our own.

†

University of Delaware, yilin[email protected].

‡

University of California Los Angeles, elisabeth.honk[email protected].

1 Introduction

Companies spend millions, and in some cases even billions, of dollars to advertise their products

to consumers.

Given the large amounts spent on advertising, the question of how well the

money is spent, i.e., how eﬀective advertising spending is in inﬂuencing consumers’ purchase

behavior, is obviously a very important one to marketing managers. Numerous studies in

both marketing and economics have provided answers to this question for various products

and industries.

However, another crucial, but much less investigated question is whether

the content of ads matters, i.e., whether the eﬀects of advertising vary across diﬀerent types of

content, or whether it only matters that companies advertise? Little is known about the answer

to this important question.

Marketing managers (often together with an advertising agency) must decide what to com-

municate to consumers in their ads. On the one hand, product-related content might be more

useful to consumers during the purchase process as it informs them about relevant product

attributes. On the other hand, given the abundance of ads consumers encounter every day, ads

must catch consumers’ attention and be memorable to be eﬀective. It is an empirical question

which type of advertising content is eﬀective with consumers. Furthermore, the eﬀects of dif-

ferent types of advertising content might vary depending on consumers’ stage in the purchase

process. For example, early on, attention-catching and memorable content might be more im-

portant, while later, closer to the actual purchase, product-related information might be more

relevant. It is important for marketing managers to know how diﬀerent types of advertising

content aﬀect each stage of consumers’ purchase process so that they can choose the advertising

content that is most eﬀective in achieving their marketing goal.

Much of the diﬃculty in quantifying the eﬀects of advertising content stems from a lack of

data on advertising content especially for non-digital advertising, e.g., TV advertising or direct

For example, P&G spent $8.3 billion on advertising its Pampers brand in 2016

(http://www.businessrevieweurope.eu/marketing/856/Top-20-companies-with-the-biggest-advertising-budget).

For example: Ackerberg (2001), Ackerberg (2003), Narayanan, Manchanda, and Chintagunta (2005), Ching

and Ishihara (2012), Chan, Narasimhan, and Xie (2013), Lovett and Staelin (2016), Hastings, Horta¸csu, and

Syverson (2017), Shapiro (2018), Shapiro, Hitsch, and Tuchman (2019).

mail. Previous research on non-digital advertising content has therefore mostly investigated

a small number of easily identiﬁable or manipulatable aspects of content such as mention

of a competitor, photo of an attractive woman, or a call for action (Bertrand et al. 2010,

Liaukonyte, Teixeira, and Wilbur 2015, Anderson et al. 2016). In this paper, we set out to

conduct a systematic and comprehensive empirical analysis of the eﬀects of the main aspects

of advertising content on consumers’ purchase process.

Because of the unavailability of canned data on advertising content, we acquired data on TV

advertising quantities and the corresponding creatives, i.e., ﬁles containing the TV commercials,

for the U.S. auto insurance industry between 2010 - 2016 from Kantar. We then hired a team of

research assistants who watched the ads and recorded the content. More speciﬁcally, research

assistants recorded the presence or absence of content on prices, non-price product features,

brand name focus,

and content with emotional appeal. We classify price-related and (non-

price) product feature-related content as “informational,” i.e., transmitting information about

the product and its characteristics to consumers, and brand name focused and emotionally

appealing content as “non-informational,” i.e., not containing information about the product

and its characteristics.

We then use these data on advertising quantities and content to investigate how both fac-

tors inﬂuence each stage of consumers’ purchase process, i.e., awareness, consideration, and

purchase.

Our data on consumers’ purchase process come from J.D. Power and Associates’

annual screener surveys and annual “Insurance Shopping Studies” conducted between 2010 and

2016. These surveys provide us with information on shoppers’ and non-shoppers’ unaided and

aided awareness sets, consideration sets, and purchase decisions.

Additionally, we also know

the survey and shopping months, have location and demographic information, information on

By deﬁnition, the brand name is mentioned/shown in all ads. The content type “brand name focus” does

not capture the simple mention of the brand name in an ad. It captures the focus on the brand name in an ad

– e.g. frequent repetitions/showings of the brand name in a TV ad, multiple/large prints of the brand name in

a magazine ad, ads with no product-related information.

In this paper, we use the terms “consider,” “search,” and “shop” interchangeably.

In this paper, we view shopping, i.e. requesting at least one price quote from an insurance company that is

not the consumer’s current insurance provider, as a prerequisite for an (active) purchase decision that involves

deciding whether to switch auto insurers.

the identity of the previous insurance provider, and categorical information on insurance pre-

mia.

We quantify the eﬀects of both advertising quantity and advertising content on consumers’

(unaided and aided) awareness, consideration, and purchase using a set of linear probability

models. We account for the endogeneity of the advertising decision using the regression dis-

continuity approach suggested by Shapiro (2018). Our results reveal that advertising intensity

aﬀects consumer (unaided and aided) awareness, but has insigniﬁcant eﬀects on conditional

consideration and conditional purchase. These ﬁndings are consistent with prior literature

(e.g., Honka, Horta¸csu, and Vitorino 2017). However, when measuring the separate eﬀects of

diﬀerent types of advertising content, i.e., ads with only informational content, ads with only

non-informational content, and ads with both informational and non-informational content,

we ﬁnd a more nuanced set of results: advertising only containing non-informational content

increases unaided awareness, while advertising only containing informational content increases

aided awareness and purchase conditional on consideration. We do not ﬁnd signiﬁcant eﬀects

of advertising containing both informational and non-informational content. Next, since many

companies spend a signiﬁcant portion of their budgets on advertising with both informational

and non-informational content, we investigate whether this type of advertising signiﬁcantly

aﬀects certain groups of consumers. We ﬁnd it to increase consumers’ unaided (and, in some

cases, aided awareness) for high risk consumers, consumers who recently moved, and consumers

with a change in their family or policy circumstances.

The contribution of this paper is two-fold: First, we provide new insights on advertising

content. Because of very limited data availability, systematic, large-scale research especially on

non-digital advertising content is scarce. We overcome this challenge by creating our own data

set containing information on the main aspects of advertising content for the auto insurance

industry for a time period of 7 years. Our data describe the diﬀerent strategies used by brands

Some information is (logically) unavailable for non-shoppers. For example, since non-shoppers do not shop,

they do not form consideration sets and do not make an (active) purchase decision, but remain passively insured

with the same insurance company. Further, since non-shoppers do not shop, we have no information on their

shopping month and (quoted) insurance premia.

in relation to advertising content. And second, we quantify the eﬀects of advertising on each

stage of the consumers’ purchase process. Understanding how both the amount and content

of advertisements aﬀect each stage of consumers’ decision-making is crucial for companies to

eﬀectively and eﬃciently use their advertising budgets. We show that advertising primarily

aﬀects consumer awareness and that diﬀerent types of advertising aﬀect diﬀerent stages of

the purchase process. Our results contribute to managers’ and researchers’ knowledge of how

advertising inﬂuences consumers.

The remainder of this paper is organized as follows: In the next section, we review the

relevant literature. In Section 3, we discuss advertising endogeneity and how we address this

issue. We describe our data in the following section. In Section 5, we introduce our models

and estimation approach. In the following two sections, we discuss our results and present

robustness checks. In Section 8, we examine limitations of our work and opportunities for

future research. And, ﬁnally, we conclude in Section 9.

2 Relevant Literature

Our paper is related to three streams of literature on advertising, consumers’ limited informa-

tion, and demand for ﬁnancial services. In the following, we review the relevant literature and

delineate the positioning of our research vis-`a-vis the ﬁndings from extant research.

Empirical researchers have long tried to determine the role(s) advertising plays in consumers’

decision-making. Most work has focused on ﬁnding empirical evidence for the informative or

persuasive view of advertising ﬁrst developed by Chamberlin (1933) (e.g., Ackerberg 2001,

Ackerberg 2003, Narayanan, Manchanda, and Chintagunta 2005, Ching and Ishihara 2012,

Chan, Narasimhan, and Xie 2013, Lovett and Staelin 2016).

Our focus is on ﬁnancial services

and, more speciﬁcally, on auto insurance. There is little academic research that investigates

the precise way through which advertising aﬀects consumer demand for ﬁnancial products.

There is little empirical work on the complementary (Stigler and Becker 1977, Becker and Murphy 1993)

and signaling views (Nelson 1970, Nelson 1974) of advertising. Recent exceptions are Tuchman, Nair, and

Gardete (2018) for complementarity and Sahni and Nair (2018) for signaling.

Gurun, Matvos, and Seru (2016) and Hastings, Horta¸csu, and Syverson (2017) explore the

eﬀects of advertising in the subprime mortgage and social security markets, respectively, but

neither of these studies can distinguish whether advertising aﬀects awareness and/or consider-

ation/purchase because of data limitations.

Most closely related to our paper is Honka, Horta¸csu, and Vitorino (2017) who investi-

gate the role of advertising in the retail banking industry. However, our paper diﬀers from

theirs in several respects: ﬁrst, the questions both papers can and do answer are diﬀerent.

Honka, Horta¸csu, and Vitorino (2017) ﬁnd that advertising plays a primarily informative role

by informing consumers about the existence of banks. They use their results to quantify branch-

advertising substitutability and to analyze the competitive eﬀects of advertising in the retail

banking industry. While we also study whether advertising aﬀects awareness and/or consider-

ation/purchase in the auto insurance industry, we focus on investigating how diﬀerent types of

advertising content aﬀect diﬀerent stages of consumers’ purchase process. Further, we study

whether advertising content has heterogeneous eﬀects across diﬀerent consumer groups. And

second, to answer the respective research questions, the empirical approaches are diﬀerent.

While Honka, Horta¸csu, and Vitorino (2017) develop a structural model and address the issue

of advertising endogeneity using the control function approach, we use reduced-form modeling

and the regression discontinuity approach to address advertising endogeneity.

The majority of the empirical literature on advertising content investigates the eﬀects of

speciﬁc informational cues on consumers’ purchase decisions (e.g., Bertrand et al. 2010, Li-

aukonyte, Teixeira, and Wilbur 2015, Tucker 2015, Anderson et al. 2016, Sahni, Wheeler, and

Chintagunta 2018). For example, Bertrand et al. (2010) conduct a direct mail ﬁeld experiment

and ﬁnd that showing fewer sample loans or including of a photo of an attractive woman in-

creases the demand for loans. They conclude that advertising content persuades by appealing

“peripherally” to intuition rather than to reason. Liaukonyte, Teixeira, and Wilbur (2015)

The data are also diﬀerent. While Honka, Horta¸csu, and Vitorino (2017) only have data on consumer

(aided) awareness, consideration, and purchase for one year and only data on advertising quantities, our data

used in this paper spans a time period of seven years, also includes unaided awareness, and we not only have

information on advertising quantities, but also on advertising content.

are closest to our paper in that they also investigate four content pieces, albeit diﬀerent ones

(action-focus, information-focus, emotion-focus, imagery focus).

Zooming in on ﬁnancial services, there are a handful of papers that investigate how diﬀerent

types of advertising content (together with advertising quantity) aﬀect consumer demand in

this industry. Using data from Sweden, Cronqvist (2006) ﬁnds that only a small fraction of

advertisements for funds is informational in the sense that the ads contain information on rel-

evant product characteristics. Nevertheless, he ﬁnds that advertising aﬀects investors’ choices

even though it provides little information. Agarwal and Ambrose (2010) use data on home eq-

uity credit choices from direct mail and walk-in customers and ﬁnd non-informational content

to inﬂuence consumer choices. Gurun, Matvos, and Seru (2016) analyze consumers’ borrow-

ing behavior in the context of subprime mortgages. They ﬁnd that initial/introductory rates

are frequently and prominently advertised, while reset rates and other characteristics of mort-

gages or lenders are rarely advertised. Further, Gurun, Matvos, and Seru (2016) show that

expensiveness and advertising intensity of a lender within a market are positively correlated

and conclude that their results are consistent with the persuasive view of advertising. And

lastly, Mullainathan, Schwartzstein, and Shleifer (2008) investigate whether predictions from

their theoretical model of the role of advertising in the mutual funds industry are consistent

with empirical patterns. They analyze the content of ads from two business magazines and ﬁnd

that the inclusion of past returns data is used to frame mutual fund investing as grabbing an

opportunity rather than as hiring advice. The results from these four papers are broadly con-

sistent with a persuasive role of advertising, i.e., advertising mostly not containing information

on product characteristics, but nevertheless aﬀecting consumer choice. However, what these

four papers implicitly assume is that consumers have full information in the sense that they

know that all these ﬁnancial institutions operate in the marketplace. While we use data on

advertising content and quantity as do the previous papers, what distinguishes our paper from

theirs is that we have information on consumers’ awareness and consideration sets allowing us

to relax the full information assumption made by previous literature.

3 Advertising Endogeneity

Advertising endogeneity is a well-known concern when estimating the eﬀects of advertising on

demand. Its cause are omitted variables or, more speciﬁcally, time-varying local “events” which

we do not observe in the data, but which might be correlated with brands’ advertising decisions.

Such time-varying local “events” include sponsorships of local sports teams or festivals and

changes in the focal brand’s or competitive brands’ local agent network (i.e., openings and

closings of agencies). They can also include changes in the communicated content. For example,

sponsorships usually only show the brand name and a brand might place more price-related

ads due to the opening of a competitor’s new local oﬃce.

We apply the regression discontinuity approach to address advertising endogeneity (Shapiro

2018). In the following, we brieﬂy describe the main idea of the regression discontinuity ap-

proach using advertising quantity. Our measure of advertising quantity is the logarithm of

total, i.e., national and DMA-level, TV advertising spending per household by brand b in DMA

d and month t, log (1 + A

bdt

), and serves as a GRP approximation (Shapiro 2018). We then

discuss the appropriateness of the regression discontinuity approach for advertising content.

Our three measures of advertising content, i.e., ad types, are total TV advertising spending

per household by brand b in DMA d and month t on (i) ads with only informational content,

log



1 + A

bdt



, (ii) ads with only non-informational content, log



1 + A

bdt



, and (iii) ads with

both informational and non-informational content, log



1 + A

f,nf

bdt



Figure 1 shows an example of two DMAs in Texas – Austin and San Antonio (in blue and

red, respectively). Note that the border between these two DMAs does not – as they do not for

most DMAs – coincide with state borders. Rather, historically, DMAs were centered around

a large city or a metropolitan area. The border strategy to deal with advertising endogeneity

considers the six counties directly adjacent to the DMA border belonging to the Austin DMA

(in dark blue) and the six counties directly adjacent to the DMA border belonging to the San

Antonio DMA (in dark red) as two treatment groups in every month. While consumers living

on diﬀerent sides of the DMA border are similar, they are being treated with diﬀerent amounts

of advertising. The advertising eﬀect can be identiﬁed by comparing how consumers living in

the two groups of border counties react diﬀerently to diﬀerences in advertising quantities.

=========================

Insert Figure 1 about here

=========================

Next, we discuss how the regression discontinuity approach can be applied to our data on

advertising content. We assume that, within a brand-DMA-month combination, the amount

of spending on each ad type can increase, decrease or stay constant independently of changes

in spending on any other ad type. For example, a brand can spend more on ads with only

non-informational content in a DMA-month (e.g., by sponsoring a local music festival that is

advertised on TV) without changing spending on the other two ad types in the same DMA-

month. An aspect of this assumption is that brands must have access to all three ad types at all

times even if they choose not to employ them. Further, the regression discontinuity approach

would not work with a percentage speciﬁcation as the percentage of one ad type increases when

the percentage of another ad type decreases. Given our assumption and the operationalization

of the advertising content variables, the intuition for the identiﬁcation of the advertising content

eﬀects is the same as the intuition for the identiﬁcation of the advertising quantity eﬀect using

the regression discontinuity approach: consumers living on diﬀerent sides of a DMA border are

similar, but being treated with diﬀerent amounts of a speciﬁc advertising type. Therefore the

eﬀect of that speciﬁc advertising type can be identiﬁed by comparing how consumers living in

the two groups of border counties react diﬀerently to diﬀerences in spending on that advertising

ad type.

Two types of ﬁxed eﬀects are crucial for the implementation of the regression discontinuity

approach for both advertising quantity and advertising content: brand-border-DMA and brand-

border-month ﬁxed eﬀects. The former control for persistent diﬀerences across diﬀerent border

regions and the latter capture unobserved border-region-speciﬁc trends. We include both types

of ﬁxed eﬀects in all empirical models.

Geographic variation, i.e., across-DMA-border variation, in log (1 + A

bdt

), log



1 + A

bdt



log



1 + A

bdt



, and log



1 + A

f,nf

bdt



within a brand and month is crucial for identifying the

eﬀects of advertising using the regression discontinuity approach (Shapiro 2018), i.e., the ef-

fects of advertising are identiﬁed by variation in DMA-level advertising (and not by national

advertising). To put it diﬀerently, we need discontinuities in all four advertising measures at

DMA borders to be able to identify the eﬀects of advertising. Furthermore, since our data

span a time period of seven years, we also need variation in the discontinuities in the four

advertising measures over time, i.e., across months. We present decriptive evidence that our

data contain such variation in Section 4.2.1 for advertising quantity and in Section 4.2.2 for

advertising content.

4 Data

We combine data from two sources to investigate the relationship between advertising and each

stage of consumers’ purchase process. Our data on advertising come from Kantar. Kantar tracks

TV advertising expenditures (in dollars and units) at the national and Designated Market Area

(DMA) level. We have monthly data from 2010 to 2016. Additionally, Kantar supplied us with

the creatives, i.e. the ﬁles containing the TV commercials.

Our second data come from J.D. Power and Associates who generously shared data from

their annual screener surveys and annual “Insurance Shopping Studies” covering consumer be-

havior from 2010 to 2016. The data sets contain individual-level information on consumers’

awareness and consideration sets, the identity of the purchased option, the identity of the pre-

vious insurance provider, location and demographic information, survey and shopping months,

perceived categorical price information for shoppers, and representativeness weights.

9, 10

Unaided awareness: “When you are thinking of auto and home insurance, which companies come to mind?”

Aided awareness: “Please review the list below and select ALL the insurance companies that you recognize.”

Consideration: “From which of the following insurance companies did you receive a quote?” Purchase: “Which

company is your current auto insurer?” Previous insurer: “Which company was your auto insurer prior to [pipe

in current insurer]?” Price information: “Which auto insurer oﬀered you the lowest price/premium?”

J.D. Power and Associates calculates representativeness weights for each annual survey, i.e., each screener

survey and each Insurance Shopping Study, to ensure that the data and results are representative. We use

those representativeness weights for consumers within each survey. Next, for each survey year, J.D. Power

and Associates reports the percentage of shoppers (versus non-shoppers) in the market. For each year, we

4.1 Data Processing

4.1.1 Advertising Content

We have all ads, i.e., “creatives,” placed by auto insurance companies on TV between 2010

and 2016 (both in Spanish and English), i.e. 2,965 unique creatives across 21 auto insurance

brands.

To code the content of these creatives, we hired a team of 25 student research assis-

tants during an 18-months time period.

These research assistants were trained to code

whether a creative (i) talked about prices/rates/discounts, (ii) conveyed (non-price) product

feature information, (iii) focused on the brand name, (iv) had emotional appeal (i.e., humor-

ous/funny/entertaining and/or fear-inducing). We developed these four content types based on

previous literature (e.g., Resnik and Stern 1977, Stern, Krugman, and Resnik 1981) and taking

the characteristics of the auto insurance industry into account. A detailed description of each

content type with examples is shown in Web Appendix A. Note that creatives can contain more

than one piece of content, e.g., price-related and emotionally appealing content. The training

of the research assistants was conducted as follows: ﬁrst, research assistants were screened for

language skills (English and Spanish) and basic knowledge of the auto insurance market before

employment. Then research assistants received a document containing a written description of

each content type and a set of 20 creatives that they coded on their own. Afterwards, they

met with one of the authors to discuss their coding decisions and to resolve other uncertainties.

re-weight all observations of non-shoppers (from that year’s screener survey) and all observations of shoppers

(from that year’s Insurance Shopping Study) so that they reﬂect that percentage of non-shoppers and shoppers

reported by J.D. Power and Associates. And lastly, given diﬀerences in sample sizes across years, we re-weight

all observations from a year so that the data from each year have the same weight. We report all descriptive

statistics and estimation results using these modiﬁed representativeness weights.

Alternatively, one could rely on machine learning algorithms to code advertising content. We decided

not to do that for several reasons: e.g., creatives are deposited in diﬀerent formats (e.g. gif, swf, and ﬂm

for an Internet display advertisement), diﬀerent resolutions, and in diﬀerent languages (English and Spanish).

Based on our observation, well-trained research assistants are more eﬀective in identifying content types and

picking up non-informational cues such as humor. To predict whether a human will ﬁnd a photograph or video

funny/entertaining remains a challenging machine learning task. Computational humor is sometimes considered

to be an “AI-complete” problem (Binsted et al. 2006). Although there has been some progress in identifying

visual humor (e.g. Chandrasekaran et al. 2016), we decide to use trained research assistants because auto

insurance advertisements often are nuanced and speciﬁc to the social context. For example, a cartoon in which

a car is destroyed by a superhero ﬁght (such as in a TV commercial by Mercury) could be considered funny,

but a car damaged completely in real life is typically horrifying.

After this meeting, research assistants started coding creatives.

Each creative was independently coded by at least three research assistants and we use

majority coding across the three research assistants for each creative in the analyses. Fleiss’

kappa is a measure of inter-rater agreement in coding. Figure 2 shows a histogram of Fleiss’

kappas for the coded creatives across the three research assistants. The average value is .68

with a median of .70 indicating substantial agreement.

=========================

Insert Figure 2 about here

=========================

4.1.2 Screener Surveys and Insurance Shopping Studies

The screener surveys conducted by J.D. Power and Associates between 2010 and 2016 pro-

vide information on a large number of non-shoppers (on average, 15,000 individuals annually).

Non-shoppers are consumers who were not actively shopping for auto insurance during a par-

ticular year. From these screener surveys, we have information on non-shoppers’ unaided and

aided awareness sets and their current insurance provider.

Each year, J.D. Power and Asso-

ciates also conduct “Insurance Shopping Studies” surveying about 10,000 individuals annually.

From these Insurance Shopping Studies, we have information on shoppers’ unaided and aided

awareness sets, consideration sets, and purchase decisions. Additionally, we also have location

and demographic information for all consumers, i.e. shoppers and non-shoppers, survey and

shopping months, and information on the identity of the previous insurance provider. And

lastly, for shoppers, we also have categorical information on insurance premia. It is important

to note that both the screener surveys and the Insurance Shopping Studies contain repeated

cross-sections of consumers and not a panel of consumers.

The original data contain information on 360,182 individuals with valid FIPS codes (108,942

shoppers and 251,240 non-shoppers). Unfortunately, detailed location information (beyond the

Since non-shoppers do not shop, they do not form consideration sets and do not make an (active) purchase

decision, but remain passively insured with the same insurance company.

state) was not available for respondents from the 2011 Insurance Shopping Study and the 2014

screener survey so these respondents were dropped. In our empirical analysis, we focus on

respondents living in counties at the borders of the top 130 DMAs excluding the Bakersﬁeld,

CA, and San Diego, CA, DMAs. We excluded the Bakersﬁeld, CA, and San Diego, CA, DMAs

because, in both cases, the whole DMA only contains one county and therefore a border could

not be deﬁned. We further focus on the top 21 brands (measured by revenue) that were

consistently part of the surveys from 2010 to 2016 and held a joint market share of about 85%.

This focus implies that respondents who purchased auto insurance from a brand that is not part

of the top 21 brands were removed. These data cleaning steps left us with 161,096 consumers

(50,181 shoppers and 110,915 non-shoppers).

Next, we dropped respondents who (i) indicated to be younger than 18 years or older than

75 years, (ii) reported an annual income of over $1,000,000, (iii) stated to own more than four

cars, (iv) reported paying a premium of more than $4,000 for a 6-months policy, (v) were

not a decision-maker regarding the auto insurance purchase, (vi) inconsistently reported their

location, and (vii) did not provide demographic information. These data cleaning steps left us

with 98,438 consumers (30,855 shoppers and 67,583 non-shoppers). And lastly, for the Insurance

Shopping Studies, to avoid any memory issues consumers might develop over time, we restrict

our data to consumers who completed the survey two or fewer months after shopping for car

insurance. This step left us with our ﬁnal sample of 77,309 respondents (67,583 non-shoppers

and 9,726 shoppers) located in 1,263 diﬀerent counties. These 1,263 counties belong to 250

diﬀerent border regions, i.e., a cluster of geographically adjacent counties spanning across both

sides of a DMA border. We reweigh the individuals in our ﬁnal sample using representativeness

weights. The reweighted ﬁnal sample contains 77,309 consumers (25,104 shoppers and 52,205

non-shoppers).

4.2 Data Description

4.2.1 Advertising Quantity

Insurance companies can place TV advertisements nationally and locally, i.e., at the DMA-

level.

On average, a single insurance brand spends about $6 million monthly on national

TV advertising placing around 2, 000 ads. There is large variation in national TV advertising

spending ranging from $0 (Auto Owners, Erie, GMAC) to $36 million (Geico) per month. All 21

brands together spend about $123 million monthly on national TV advertising placing around

43, 000 ads. At the DMA-level, a single insurance brand spends, on average, around $2, 400

per DMA and per month placing about 12 TV ads. Across all DMAs, all 21 brands spend, on

average, around $11 million per month on DMA-level advertising placing about 53, 120 TV ads.

89% of brands’ TV advertising spending is utilized on national advertising and the remaining

11% are spent on DMA-level advertising.

While national TV advertising has increased during

the observation period, DMA-level TV advertising has decreased.

Focusing on the top 130 DMAs excluding the Bakersﬁeld, CA, and San Diego, CA, DMAs,

average total monthly advertising spending per household was $0.06 with brands such as Erie,

GMAC, and Safeco spending $0 and brands such as Progressive and Geico spending $0.20 and

$0.34, respectively, per household (see column (i) in Table 1). Following Shapiro (2018), we

use the logarithm of total advertising spending per household, log (1 + A

bdt

), as our measure

of advertising intensity in the empirical analyses. As robustness checks of our advertising

measure, we also use total advertising units and DMA-level ad expenditure per household in $

as measures of advertising. Descriptive statistics for these two variables are shown in columns

(ii) and (iii) in Table 1.

=========================

Insert Table 1 about here

=========================

Across the diﬀerent media channels, auto insurance brands spend 80% of their total advertising budget on

TV, 7% on Internet (display advertising), 6% on the radio, 7% on print, and less than 1% on outdoor advertising.

In terms of units, 67% of brands’ TV advertising units are utilized on national advertising and the remaining

27% are spent on DMA-level advertising.

A potential concern with the regression discontinuity approach is whether the amount of

variation left in the advertising variable after the inclusion of all ﬁxed eﬀects is suﬃcient

to identify the eﬀect of advertising. To evaluate the amount of remaining variation, we

regress log (1 + A

bdt

) on the full set of ﬁxed eﬀects, i.e., brand-border-DMA, brand-border-

month, brand-state-month, survey, brand-demographics-year, and brand-online-demographics-

year ﬁxed eﬀects (see Section 5 for details on ﬁxed eﬀects), and plot the residuals from this

regression in Figure 3. The average residual equals, as expected, (approximately) zero with a

standard deviation of 0.0041. For comparison, the mean of log (1 + A

bdt

) is 0.0449. Following

Shapiro, Hitsch, and Tuchman (2019), we calculate a “coeﬃcient of variation” as the ratio

of the standard deviation of the residuals and the mean of log (1 + A

bdt

). This coeﬃcient of

variation equals 0.09, i.e., the standard deviation of the residuals is about 9% the average size

of log (1 + A

bdt

=========================

Insert Figure 3 about here

=========================

As discussed in Section 3, geographic variation in advertising across the two sides of a border

within a brand and month is crucial for identifying the eﬀects of advertising using the regression

discontinuity approach (Shapiro 2018). Since geographic variation in our advertising measure

stems from variation in DMA-level advertising (and not national advertising), we focus here on

DMA-level advertising per household. Furthermore, since our data span a time period of seven

years, we also need variation in the discontinuity in advertising over time, i.e., across months.

We show that our data contain both types of variation using two approaches: we ﬁrst show

evidence of discontinuities in the advertising data and that these discontinuities vary across

geographies (borders) and time (months). In a second step, we then regress our advertising

measure on the full set of ﬁxed eﬀects and show that there are discontinuities in residuals that

vary across borders and months.

We display the detailed results in column (i) in Table 2. Here, we brieﬂy summarize our

ﬁndings: ﬁrst, the mean absolute diﬀerence in DMA-level advertising spending per household for

each brand-border-month observation is 0.0077 with a standard deviation of 0.0497. Given that

brands, on average, spend 0.0046 on DMA-level advertising per household (see Table 1), these

discontinuities can be substantial. We then show the proportion of brand-border combinations

that show variation in the advertising discontinuity over time (i.e., across 84 months) and the

proportion of brand-month combinations that show variation in the advertising discontinuity

across geographies (i.e., across 250 borders). The proportions with variation are large: 74%

and 72%.

=========================

Insert Table 2 about here

=========================

And second, we regress DMA-level advertising spending per household on the full set of

ﬁxed eﬀects. We then calculate the residuals from this regression and the absolute diﬀerences

in residuals for each brand-border-month combinations. As shown in column (i) in Table 2,

the mean absolute diﬀerence in residuals in 0.0030 with a standard deviation of 0.0086. Lastly,

we also observe a large amount of variation across time and geographies (63% and 83%). We

conclude that our data contain suﬃcient variation to identify the eﬀects of advertising intensity

using the regression discontinuity approach.

4.2.2 Advertising Content

Table 3 depicts the percentages of TV ads (weighted by spending) for each brand that contain

a speciﬁc mix of content. On average, 14% of ads only contain informational content, 34% of

ads only contain non-informational content, and the remaining 52% contain both informational

and non-informational content. There is a large amount of variation in advertising content

across brands. The percentage of ads with only informational content in column (i) in Table

3 varies from 0% (e.g., 21st Century, American Family, Auto Owners) to 100% (Erie), the

percentage of ads with only non-informational content in column (ii) ranges from 0% (e.g.,

21st Century, Erie) to 95-100% (Auto Owners, Safeco, USAA), and the percentage of ads that

include both informational and non-informational content in column (iii) varies from 0-1% (e.g.,

Auto Owners, Safeco, USAA) to 98-100% (21st Century, Geico).

=========================

Insert Table 3 about here

=========================

Brands also communicate diﬀerent combinations of the three advertising content types

shown in columns (i) to (iii): some brands mostly focus on one type – non-informational

only for Auto Owners, Safeco, and USAA, informational only for Erie, and both informational

and non-informational for 21st Century and Geico. Other brands employ a rather balanced

combination of all three types. For example, 41% of Allstate’s ads are only informational, 27%

of Allstate’s ads are only non-informational, and the remaining 31% are both informational and

non-informational. A similar pattern can be observed for Nationwide and State Farm. And

lastly, some brands such as American Family or Travelers employ a combination of two types:

non-informational only content and informational and non-informational content.

Using the same approach as for our advertising quantity measure in Section 4.2.1, we evalu-

ate whether the amount of variation in the three advertising content variables after the in-

clusion of all ﬁxed eﬀects is suﬃcient to identify the eﬀects of advertising. We do so by

regressing log



1 + A

bdt



, log



1 + A

bdt



, and log



1 + A

f,nf

bdt



on the full set of ﬁxed eﬀects

and plotting the residuals from this regression in Figure 4. The average residuals equal, as

expected, (approximately) zero with standard deviations of 0.0011, 0.0049, and 0.0025 for

log



1 + A

bdt



, log



1 + A

bdt



, and log



1 + A

f,nf

bdt



, respectively. For comparison, the average

values of log



1 + A

bdt



, log



1 + A

bdt



, and log



1 + A

f,nf

bdt



are 0.0039, 0.0076, and 0.0331,

respectively. Calculating the coeﬃcient of variation of each ad type, we get 0.28, 0.64, and 0.07

for ads with only informational content, only non-informational content, and both informational

and non-informational content, respectively.

=========================

Insert Figure 4 about here

=========================

Next, we show empirical evidence for (i) variation in advertising content across geographies

within a brand and month and for (ii) variation in these discontinuities across time – the same

two types of variation as for advertising quantity. We use the same set of approaches as for the

advertising quantity data in Section 4.2.1 and the results are shown in columns (ii) - (iv) in Table

2. We ﬁrst display the mean absolute diﬀerences and standard deviations in log



1 + A

DM A,f

bdt



log



1 + A

DM A,nf

bdt



, and log



1 + A

DM A,f,nf

bdt



. Given that brands, on average, spend 0.0009,

0.0025, and 0.0111 on DMA-level advertising per household on only informational ads, only

non-informational ads, and ads containing both informational and non-informational content,

respectively, these discontinuities can be substantial. Further, we ﬁnd variation in all three

advertising measures across time and geographies. And second, we display the same statistics

for the absolute diﬀerences in advertising residuals in the lower half of Table 2. We conclude

that our data contain suﬃcient variation to identify the eﬀects of diﬀerent advertising types

using the regression discontinuity approach.

4.2.3 Consumer Shopping Behavior

We ﬁrst discuss consumer characteristics and then consumer shopping behavior. In Table 4,

we compare descriptive statistics for all consumers (column (i)), non-shoppers (column (ii)),

and shoppers (column (iii)) separately.

Among all consumers, about 80% of respondents are

between 25 and 65 years old, 42% are male, and 58% are married. 59% of respondents have a

college degree and 25% of respondents have an annual income of more than $100k. Comparing

the two subgroups of shoppers and non-shoppers (columns (ii) and (iii) in Table 4), we ﬁnd

shoppers to be more likely male, married, Black, and Hispanic than non-shoppers. For shoppers

(only), we have additional information on insurance-related variables: 44% of shoppers were

also shopping for homeowner’s or renter’s insurance and 8% of shoppers indicated having a poor

credit history. Further, 3% and 4% of shoppers reported having had two or more accidents and

Recall that we only include respondents living in border counties of the top 130 DMAs excluding the

Bakersﬁeld, CA, and San Diego, CA, DMAs in our ﬁnal sample. Descriptive statistics comparing the original

and ﬁnal data sets are shown in Web Appendix B. The distributions of demographic and insurance-related

variables are largely similar.

tickets, respectively, during the last three years.

=========================

Insert Table 4 about here

=========================

Next, we discuss consumer shopping behavior. In our data, 32% of consumers are shoppers

and the remaining 68% of consumers are non-shoppers.

This proportion of shoppers is con-

sistent with proportions reported by other sources: 46% of consumers reported having shopped

for auto insurance during the past 12 months according to a 2015 comScore survey,

25% of

consumers reported having shopped for auto insurance during the past 12 months according

to a 2017 Princeton Research Survey Associates International survey, and 33% of consumers

reported having shopped for auto insurance during the past 12 months according to the 2012

McKinsey Auto Insurance Customer Insights Research report.

Among shoppers, 48% of consumers switch their auto insurance provider after the shopping

occasion under study and the remaining 52% of consumers remain with their previous insurance

provider.

Projecting to the whole population, we ﬁnd that 15% of all consumers switch their

auto insurance provider in a year. The 2012 McKinsey Auto Insurance Customer Insights

Research similarly report found about 1/3 of shoppers or 13% of the total population to switch

insurance providers.

The average number of auto insurance brands consumers are aware of is 4.21 for unaided

awareness and 12.49 for aided awareness. As expected, non-shoppers are aware of fewer brands

than shoppers: 3.82 vs. 5.01 (diﬀerence statistically signiﬁcant at p < 0.001) for unaided

awareness and 12.37 vs. 12.67 (diﬀerence statistically insigniﬁcant) for aided awareness. We

next turn to the brands that consumers are aware of (see Table 5). The probability that a

consumer is aware of any brand is 20% (unaided) and 59% (aided) (columns (i) to (ii) in Table

The percentage of shoppers increased from 29% in 2010 - 2012 to 37% in 2014 - 2016.

https:www.comscore.comInsightsPress-Releases201511comScore-Releases-2015-US-Online-Auto-

Insurance-Shopping-Report;

https:www.huﬃngtonpost.comentrypaying-too-much-for-auto-insurance-many-consumers us 58c2dbede4b070e55af9ee2b;

https:www.mckinsey.com mediamckinseydotcomclient serviceFinancial%20ServicesLatest%20thinkingInsurance

Winning share and customer loyalty in auto insurance.ashx

The percentage of switchers among shoppers increased from 45% in 2010 - 2012 to 49% in 2014 - 2016.

5). Across all consumers, the brand-speciﬁc awareness probabilities range from 1% (GMAC) to

71% (State Farm) for unaided awareness and 10% (Auto Owners, Erie) to 96% (State Farm) for

aided awareness. Further, we compare the brand-speciﬁc awareness probabilities for shoppers

and non-shoppers (columns (iii) to (vi) in Table 5). Not surprisingly, shoppers have, on average,

a higher probability of being aware of any brand than non-shoppers. We provide more details

on consumers awareness and consideration in Web Appendix C.

=========================

Insert Table 5 about here

=========================

Table 6 contains consideration and purchase shares for all brands as well as conversion rates

for consideration, i.e., conditional on being aware of a brand the proportion of consumers that

consider the brand, and for purchase, i.e., conditional on considering a brand the proportion of

consumers that choose the brand.

Conditional on unaided awareness, the conversion rates to

consideration vary from 47% (Farmers) to 93% (GMAC) with an average of 64%. Conditional

on aided awareness, the conversion rates to consideration vary from 10% (Metlife) to 43%

(Geico) with an average of 26%. And lastly, conditional on consideration, the conversion rates

to purchase range from 15% (Geico) to 63% (Erie) with an average of 28%. To summarize,

there is substantial variation within a purchase stage and across purchase stages both in shares

and in conversion rates across brands.

=========================

Insert Table 6 about here

=========================

5 Model and Estimation

Using a set of linear probability models, we estimate the eﬀect of adverting quantity on con-

sumers’ awareness and conditional consideration as follows: let Y

ibt

be the binary dependent

Note that these shares were calculated based on the available information in our data, i.e., the number

if individuals who bought an insurance policy from a brand. In constrast, many outlets and websites publish

market shares based on each brand’s revenue.

variable of interest, i.e., (unaided or aided) awareness or consideration for consumer i, brand b,

and month t. Then

ibt

= β log (1 + A

bd,t−1

) + %

+ ϕ

+ ν

bdr

+ η

brt

+ ζ

bst

+ τ

+ 

ibt

(1)

where log (1 + A

bd,t−1

) captures advertising intensity by brand b in DMA d in month t −

1. Moreover, we include a large set of ﬁxed eﬀects: ﬁrst, we include brand-demographic-

group-year ﬁxed eﬀects, %

While we do not observe other potentially targeted oﬄine

marketing activities such as direct mail, as long as the targeting is based on demographics,

our brand-demographic-group-year ﬁxed eﬀects control for it. Second, we include online-brand-

demographic-group-year-ﬁxed eﬀects, ϕ

. The dummy variable I

is individual-speciﬁc and

indicates whether a consumer spends more than the median amount of hours per week online

(14 hours). While we do not observe online search advertising in our data, these ﬁxed eﬀects

control for the amount of exposure to targeted online advertising as long as the targeting is

based on demographics.

Third, brand-border-DMA ﬁxed eﬀects, ν

bdr

, capture persistent diﬀerences across diﬀerent

border regions. Fourth, brand-border-month ﬁxed eﬀects, η

brt

, capture unobserved border-

region-speciﬁc trends. The two last mentioned types of ﬁxed eﬀects are essential for the re-

gression discontinuity approach (Shapiro 2018). Fifth, brand-state-month ﬁxed eﬀects, ζ

bst

capture changes at the state level such as changes in insurance rates, i.e. premium levels, or

regulations over time. And sixth, we also include survey ﬁxed eﬀects, τ

, to control for any

systematic diﬀerences across surveys. 

ibt

is a standard normally distributed error term and

θ =



β, %

, ϕ

, ν

bdr

, η

brt

, ζ

bst

, τ



is the vector of parameters to be estimated.

Note that we condition on consumers’ awareness sets when estimating the eﬀects of adver-

tising on consumers’ consideration decisions, i.e., we only include the set of brands for each

consumer that the consumer is aware of. We do so once using consumers’ unaided and once

The demographic groups for which we estimate ﬁxed eﬀects are as follows: age < 25 years, age between 25

and 45 years, age between 45 and 65 years, male, shopped for homeowner insurance, more than one accident in

last 3 years, has college degree, income of more than $100k.

consumers’ aided awareness sets. Similarly, in the following model describing the eﬀects of

advertising quantity on purchases, we condition on each individual consumer’s consideration

set.

We quantify the eﬀects of advertising quantity on consumers’ purchase decision as follows:

Let Y

ibt

= 1 if consumer i purchases an insurance policy from brand b in month t and Y

ibt

= 0

otherwise. Then

ibt

= β log (1 + A

bd,t−1

) + δ

ibt

+ δ

ibt

+ %

+ ϕ

+ ν

bdr

+ η

brt

+ ζ

bst

+ τ

+ 

ibt

(2)

where I

ibt

captures state dependence and is operationalized as a dummy variable indicating

whether brand b chosen in time period t is the same brand that consumer i chose in the

previous policy period. The variable I

ibt

is also a dummy variable that indicates whether brand

b oﬀered the lowest premium for consumer i in time period t among the brands consumer i

considered and is a self-reported variable. Thus, while the brand-state-year ﬁxed eﬀects ζ

bst

capture average premium changes across all consumers in a state (for a company and year),

the dummy variable I

ibt

is speciﬁc to each consumer and his consideration set. The remaining

variables are deﬁned as in Equation (1). 

ibt

is a standard normally distributed error term and

θ =



β, δ

, δ

, %

, ϕ

, ν

bdr

, η

brt

, ζ

bst

, τ



is the vector of parameters to be estimated.

In the next two equations, we describe how we jointly study the eﬀects of advertising quan-

tity and advertising content on consumers’ awareness, consideration, and purchase. Equation

(3) is our speciﬁcation for (unaided and aided) awareness and conditional consideration and

Equation (4) depicts our conditional purchase model. Here, again, we condition on consumers’

awareness sets when modeling consideration and on consumers’ consideration sets when mod-

eling purchase.

ibt

=β

log



1 + A

bd,t−1



+ β

log



1 + A

f,nf

bd,t−1



+ β

log



1 + A

bd,t−1



+ %

+ ϕ

+ ν

bdr

+ η

brt

+ ζ

bst

+ τ

+ 

ibt

(3)

ibt

=β

log



1 + A

bd,t−1



+ β

log



1 + A

f,nf

bd,t−1



+ β

log



1 + A

bd,t−1



+ δ

ibt

+ δ

ibt

+ %

+ ϕ

+ ν

bdr

+ η

brt

+ ζ

bst

+ τ

+ 

ibt

(4)

where log



1 + A

bd,t−1



, log



1 + A

f,nf

bd,t−1



, and log



1 + A

bd,t−1



capture the logarithms of

total spending in dollar per household on ads with only informational content, ads with both

informational and non-informational content, and ads with only non-informational content,

respectively. The remaining variables are deﬁned the same way as in Equations (1) and (2).

A potential concern are unobserved, individual- and brand-speciﬁc variables that are corre-

lated across the three purchase stages (awareness, consideration, purchase). Not accounting for

such variables can lead to selection issues and biased estimates in the conditional consideration

and conditional choice stages of the advertising quantity and advertising content regressions

(but is not a concern for the awareness stage – see Maddala 1983 or Wachtel and Otter 2013).

Since we model three stages of consumers’ purchase process, selection could potentially oc-

cur twice: as consumers move from awareness to consideration and as consumers move from

consideration to purchase. However, in our speciﬁc empirical content of the auto insurance

industry, selection can only happen once when consumers move from awareness to considera-

tion/shopping, i.e., decide whether to actively shop or to passively remain insured with their

previous insurance company. The reason is that having auto insurance is mandatory and thus

all consumers who shop also have to buy an insurance policy, i.e., no purchase is not an option.

While we do not account for the potential existence of such unobserved, individual- and brand-

speciﬁc variables in our modeling and estimation decisions, we discuss why selection does not

appear to be a concern in our speciﬁc empirical setting in Section 6 and also present robustness

checks in Web Appendix F.

6 Results and Discussion

6.1 Advertising Intensity

The results for advertising intensity are shown in Table 7: the top half of the table shows

the model estimates using the border strategy and the bottom half of the table shows the

model estimates without the border strategy (for the same sample of consumers). Note that all

standard errors are clustered at the DMA-level and that we report the “eﬀective” number of

observations in all tables, i.e., the number of observations remaining after observations collinear

with included ﬁxed eﬀects have been dropped.

Using the border strategy, we ﬁnd advertis-

ing intensity, i.e., total advertising spending per household, to signiﬁcantly aﬀect consumers’

unaided and aided awareness (see columns (i) and (ii) in Table 7).

To understand the magni-

tudes of these advertising eﬀects, we calculate average advertising elasticities. Using the border

strategy, the average advertising elasticities for unaided and aided awareness are 0.05 and 0.01,

respectively.

=========================

Insert Table 7 about here

=========================

We now describe the eﬀects of advertising intensity on the other stages of the purchase

process. Recall that to do so we use data on shoppers only since non-shoppers do not shop and

thus do not form consideration sets. In columns (iii) and (iv) in Table 7, we show the results

for consideration conditional on unaided and aided awareness, respectively. In both cases,

advertising does not have signiﬁcant eﬀects on consideration. Turning to the purchase stage,

we ﬁnd total advertising spending per household to have an insigniﬁcant eﬀect on conditional

purchase (column (v) in Table 7). Compared to the awareness and conditional consideration

Our results – in terms of which coeﬃcients are signiﬁcant and which ones are not – are robust to alternative

clusterings at the individual and at the DMA-border level.

Note that we also investigated alternative model speciﬁcations (for all models estimated in this paper) in

which we additionally also control for competitors’ advertising. The results (for advertising quantity and content)

are qualitatively and quantitatively similar for the focal brand. Therefore we present the most parsimonuous

model speciﬁcation in this paper. The results for the alternative model speciﬁcations are shown in Tables F-5

and F-6 in Web Appendix F.

regressions, we include two additional variables in the conditional purchase regression (column

(v) in Table 7): a dummy variable indicating whether a brand is a consumer’s previous insurance

provider and a dummy variable indicating whether a brand oﬀered a consumer the lowest

premium. The parameter estimates for both variables are signiﬁcant and have the expected

signs: consumers are more likely to purchase an insurance policy from their previous insurance

provider and are also more likely to pick the insurance brand that oﬀers them the lowest

premium.

Lastly, in column (vi) in Table 7, we compare our results to those from an unconditional

purchase model, sometimes also called a full information model, i.e., a model in which consumers

are aware of and consider all brands in the market for purchase. We ﬁnd advertising to have a

small, but signiﬁcant positive eﬀect on purchase (elasticity: 0.02). The advertising elasticity is

in line with those found in Shapiro, Hitsch, and Tuchman (2019) for a large number of products.

However, this result stands in contrast to the result from the conditional purchase model in

column (v) in Table 7 in which advertising does not have a signiﬁcant eﬀect. Further, under full

information, the eﬀect of the previous insurer dummy is overestimated and the eﬀect of the best

price dummy is underestimated. Thus, similar to previous literature (e.g., Goeree 2008, Pires

2016), we ﬁnd that not accounting for consumers’ limited information leads to quantitatively

and qualitatively diﬀerent results.

To recap, we ﬁnd that advertising intensity positively aﬀects consumer purchase behavior.

However, it does so not by directly aﬀecting consumer purchase decisions, but rather indirectly

by aﬀecting consumers’ (unaided and aided) awareness. These ﬁndings are consistent with

those in Honka, Horta¸csu, and Vitorino (2017) in the context of retail banks. Further, these

results are also consistent with advertising professionals’ recent demands for companies to focus

on consumer awareness rather than consumer engagement.

See, e.g., http://www.adweek.com/brand-marketing/advertisers-need-to-stop-chasing-engagement-and-

get-back-to-focusing-on-awareness/?utm content=buﬀer42f1f&utm medium=social&utm source=facebook.com&

utm campaign=buﬀer

6.2 Advertising Content

We now discuss our results on the eﬀects of advertising content. Recall that we include three

advertising content variables and operationalize them as total spending per household on (a) ads

with only informational content, (b) ads with only non-informational content, and (c) ads with

both informational and non-informational content. Columns (i) and (ii) in Table 8 show the

parameter estimates for unaided and aided awareness. Our results reveal that advertising with

only non-informational content has a signiﬁcant positive eﬀect on unaided awareness (elasticity:

0.02), while advertising with only informational content has a positive signiﬁcant eﬀect on aided

awareness (elasticity: 0.01).

=========================

Insert Table 8 about here

=========================

It is well-known that unaided and aided awareness do not refer to the same concept. Unaided

awareness or (brand) recall captures situations in which a consumer must independently produce

previously acquired information, while aided awareness or (brand) recognition describes cases

in which a consumer is given possible choices and must indicate which one was previously

seen (Lynch and Srull 1982, Alba, Hutchinson, and Lynch 1991). Therefore diﬀerent types of

advertising content might aﬀect recall but not recognition and vice versa.

We ﬁrst discuss the eﬀect of non-informational content on brand recall (unaided awareness).

The inﬂuence and importance of memory on brand recall and through brand recall subsequently

on brand choice has long been acknowledged (Alba, Hutchinson, and Lynch 1991). For example,

Nedungadi (1990) highlights the role of memory in brand retrieval and thus consumer awareness

of a brand. An important question is why non-informational content leads to better brand recall.

Recall that non-informational ads are brand name focused and/or emotionally appealing ads.

Brand name focused ads are creatives that either dominantly and/or frequently show the brand

name or contain no information on car insurance but that mention the brand name (e.g.,

TV program sponsorships, public service messages). It has long been known that repetition

enhances memory (e.g., Ebbinghaus 1885, Hintzman 1976). As the term says, emotionally

appealing creatives appeal to consumers’ emotions, often contain a story or an unexpected

event, and allow for imagination and inspiration (Zwaan and Radvansky 1998, Heath and Heath

2008). Emotional messages make people care and develop empathy for the main character.

Consequently, emotional events and messages are better remembered and recalled (Reisberg

and Hertel 2004). And lastly, funny and humorous ads often contain an element of surprise that

catches and increases the consumer’s attention (Pieters, Warlop, and Wedel 2002). Through

the increase in attention, unexpected events then turn into memories.

Our results show that informational content enhances consumers’ recognition of brands

(aided awareness). Previous research has shown that familiarity, i.e., the number of product-

related experiences and exposures, increases recognition (Alba, Hutchinson, and Lynch 1991).

Further, the accessibility of product attributes/positioning aﬀects which brands are recognized

as members of a product category (McCloskey and Glucksberg 1979, Alba, Hutchinson, and

Lynch 1991). Recall that informational advertising content contains descriptions of price and

non-price product attributes. Thus it increases the accessibility of these product attributes and

accessibility, in turn, increases brand recognition.

In columns (iii) and (iv) in Table 8, we show the results for consideration conditional on

unaided and aided awareness, respectively, and the results for conditional purchase in column

(v). No advertising type has a signiﬁcant eﬀect on conditional consideration. However, we ﬁnd

a signiﬁcant eﬀect of only informational advertising on conditional purchase. At this stage in

the purchase process, information on product attributes that are important to the consumer

(e.g., accident forgiveness or roadside assistance) might directly inﬂuence his purchase decision.

Further, in the conditional purchase regression (column (v) in Table 8), the coeﬃcient estimates

for the previous insurer and lowest price dummies are very similar to those in Table 7 (column

(v)) where we showed results from modeling the eﬀects of advertising intensity.

Similar to the analyses for advertising intensity, we also estimate a full information model,

i.e., unconditional purchase model, for advertising content. The results are shown in column

(vi) in Table 8. We ﬁnd a small and signiﬁcant eﬀect of only non-informational advertising

(elasticity: 0.04). This result is consistent with Cronqvist (2006) and Bertrand et al. (2010) who

ﬁnd non-informational content to inﬂuence consumers’ decision-making for ﬁnancial services.

Considering a wider range of industries and contexts, Sahni, Wheeler, and Chintagunta (2018)

and Lee, Hosanagar, and Nair (2018) also found that non-informational content can aﬀect email

openings, sales, and customer engagement.

As discussed in Section 5, unobserved, individual- and brand-speciﬁc variables that are

correlated across the three purchase stages can lead to selection issues and biased estimates in

the conditional consideration and conditional choice stages (Maddala 1983, Wachtel and Otter

2013). This is a potential concern for both the advertising quantity and advertising content

regressions. To present our approach to this issue and ﬁndings in a concise manner, we pool the

discussion on selection as related to the advertising quantity and advertising content analyses

here.

Positively (negatively) correlated, unobserved, individual- and brand-speciﬁc variables can

lead to an overestimation (underestimation) of the eﬀect of advertising in the conditional con-

sideration and conditional choice stages. Given that we ﬁnd the eﬀects of advertising quantity

in the conditional consideration and conditional purchase stages to be insigniﬁcant (see columns

(iii) - (v) in Table 7), overestimation is not a concern. For advertising content, our results show

insigniﬁcant eﬀects of all three ad types for conditional consideration and a signiﬁcant eﬀect of

ads with only informational content on conditional purchase (see columns (iii) - (v) in Table

8) so overestimation is somewhat of a concern. Underestimation of the eﬀects of advertis-

ing quantity and advertising content in the conditional consideration and conditional purchase

regressions is a potential concern for the results in both Table 7 and Table 8.

We address selection concerns with two robustness checks. The results for conditional

consideration and conditional purchase are shown in Tables F-3 and F-4 in Web Appendix F.

Recommendations by family and friends are an example of an individual- and brand-speciﬁc

variable. In our data, we observe whether a consumer received a recommendation, but not

the valence of the recommendation. Therefore, in the ﬁrst robustness check whose results are

shown in Table F-3, we drop all consumers who report having received a recommendation.

Furthermore, we also observe in our data whether there is a brand a consumer would never

consider. In a second robustness check whose results are shown in Table F-4, we control for

the presence of such a brand. In both robustness checks, we ﬁnd our advertising coeﬃcient

estimates for conditional consideration and conditional purchase in the advertising quantity

and advertising content regressions to be similar to those from our main model speciﬁcation.

We therefore conclude that selection is not a concern in our speciﬁc empirical application.

Taking a step back, our results also speak to the concepts of informative and persuasive ad-

vertising (Bagwell 2007). Informative advertising informs consumers about product existence

and (price and non-price) product features, while persuasive advertising alters consumers’ tastes

and creates spurious product diﬀerentiation (Bagwell 2007). Any type of advertising that con-

veys the existence of a product to consumers, i.e. makes consumers aware of a product, has an

informative eﬀect. To put it diﬀerently, the eﬀect of both informational and non-informational

advertising content on consumer awareness is informative. We ﬁnd this informative eﬀect in

our empirical application.

To understand whether advertising has an informative and/or persuasive eﬀect in the con-

sideration and purchase stages of the purchase process, the content of ads has to be observed. If

non-informational ad content were to aﬀect conditional consideration and/or conditional pur-

chase, the eﬀect of advertising could be interpreted as persuasive. If informational ad content

were to aﬀect conditional consideration and/or purchase, the eﬀect of advertising could be in-

terpreted as informative. In our empirical application, we only ﬁnd informational ad content to

aﬀect conditional purchase. Thus, overall, our results are consistent with an informative eﬀect

of advertising in the auto insurance industry.

To summarize, we ﬁnd advertising only containing non-informational content to increase

unaided awareness, while advertising only containing informational content increases aided

awareness and purchase conditional on consideration. We do not ﬁnd signiﬁcant eﬀects of

advertising containing both informational and non-informational content. This last ﬁnding

prompts the question why companies spend a large portion of their advertising budgets on

advertising that contains both informational and non-informational content if that type of

advertising is not eﬀective. We provide an explanation in the following section.

6.3 For Whom is Advertising with Both Informational and Non-

Informational Content Eﬀective?

Our main set of results in the previous section shows that advertising with both informational

and non-informational content has insigniﬁcant eﬀects consumers’ awareness, consideration,

and purchase. Yet, brands spend about 52% of their advertising budgets on advertising with

that mix of content. If advertising with both informational and non-informational content is

not eﬀective, it raises the question why brands spend so much money on it. One potential

explanation is that advertising with both informational and non-informational content has

signiﬁcant eﬀects for certain groups of consumers. In this section, we investigate this potential

explanation by looking at the heterogeneous eﬀects of advertising for three groups of consumers:

high risk consumers, consumers with a change in their family or policy circumstances, and

consumers who recently moved.

We deﬁne high risk consumers as consumers who had accidents or tickets during the last

three years, who have a poor credit history or who are younger than 25 years. Consumers

have to satisfy one of these criteria to be classiﬁed as high risk consumers. In our data, 8.5%

of individuals are high risk consumers. Consumers with a change in their circumstances are

individuals who either added/dropped a vehicle to their policy, added/dropped a driver to their

policy or had a change in their family (e.g., marriage, divorce) during the last 12 months. They

represent 6.0% of consumers in our data. And lastly, consumers who moved during the last

year constitute 4.4% of shoppers (1.4% of consumers) in our ﬁnal sample.

The results for advertising quantity are shown in Table E-1 in Web Appendix E. Overall, our

results are consistent with those from the main speciﬁcation: advertising intensity has signiﬁ-

cant eﬀects on awareness, but insigniﬁcant eﬀects on conditional consideration and conditional

purchase for all groups of consumers. For all three groups of consumers advertising intensity

has signiﬁcantly larger eﬀects on unaided awareness than for the remainder of the population.

Here, we focus on the results for advertising content which are shown in Table 9. Overall,

our results are consistent with those from the main speciﬁcation: advertising only contain-

ing non-informational content increases unaided awareness, while advertising only containing

informational content increases aided awareness and purchase conditional on consideration.

However, we also ﬁnd advertising with both informational and non-informational content to

have signiﬁcant eﬀects on unaided and aided awareness for the three groups of consumers. For

example, advertising with both informational and non-informational content has signiﬁcant ef-

fects for unaided and aided awareness of high risk consumers and consumers with a change in

circumstances. It also signiﬁcantly aﬀects unaided awareness of consumers who recently moved.

=========================

Insert Table 9 about here

=========================

7 Robustness Checks

We evaluate the robustness of our results using several checks. First, we evaluate the robustness

of our results with respect to an alternative operationalization of the advertising quantity

variable. Here, we re-estimate our models using the logarithm of total advertising units as our

measure of advertising intensity. The results are shown in the top half of Table F-1 in Web

Appendix F. While the results for awareness are directionally robust, the advertising coeﬃcient

for unaided awareness is insigniﬁcant. Second, we investigate the robustness of our results

with respect to our use of total advertising spending. Here, we re-estimate our models using

the logarithm of DMA-level advertising spending per household as our measure of advertising

intensity. The results are shown in the lower half of Table F-1 in Web Appendix F. Overall,

our results are qualitatively robust to this alternative operationalization.

Third, we evaluate the robustness of our results with respect to the degree of inter-rater

agreement in coding. To do so, we re-estimate our models for advertising content only using

ads with Fleiss’ kappa of at least 0.4 (we drop all ads with lower kappas from our data; 5%

of all ads). The results are shown in the upper half of Table F-2 in Web Appendix F and

conﬁrm that our results also hold for this subsample of ads. And lastly, we investigate the

robustness of our results with respect to our use of total advertising in the ad content regressions.

Here, we re-estimate our models using the logarithm of DMA-leve advertising spending per

household on only informational ads, only non-informational ads, and both informational and

non-informational ads. The results are shown in the lower half of Table F-2 in Web Appendix

F and conﬁrm that our ﬁndings are robust to this alternative operationalization.

8 Limitations and Future Research

There are several limitations to our paper and opportunities for future research. First, while we

observe the shopping month, we do not observe the shopping date. This is a limitation of our

data. Therefore we quantify the eﬀects of advertising in month t − 1 on outcome variables in

month t. However, consumers are also likely inﬂuenced by advertising during month t in the days

prior to the shopping date. Thus our estimates on the eﬀects of advertising should be interpreted

as lower bounds. Second, we quantify the eﬀects of only informational, only non-informational,

and both informational and non-informational ads. However, ads containing both informational

and non-informational content could potentially be divided in at least three subgroups: ads

containing more informational than non-informational content, ads containing equal amounts

of informational and non-informational content, and ads containing less informational than

non-informational content. We leave such a more detailed examination of the eﬀects of ads

with both informational and non-informational content for future research.

Third, our results are based on data from one industry. While our results are broadly

consistent with those found by previous literature for other ﬁnancial services (when comparisons

can be made), it is likely that the generalizability of our results decreases the further one

moves away from the ﬁnancial services sector. This represents a limitation of our data. And

lastly, we have information on four content pieces: price and non-price product features, brand

name focus, and emotional appeal which we aggregate to informational and non-informational

content. Exploring how the eﬀects vary across the four content pieces is left for future research.

9 Conclusion

Understanding how advertising inﬂuences consumers’ decision-making is crucial for companies.

Marketing managers must not only decide how much to spend on advertising, but also what

to communicate to consumers in their advertisements. In this paper, we quantify both the

eﬀects of TV advertising quantities and TV advertising content in the context of the U.S. auto

insurance industry. We ﬁnd advertising content to matter – a lot. This ﬁndings should give

pause to marketing managers and increase their focus on employing advertising content that is

eﬀective in achieving their marketing goals.

To summarize, our results show that advertising quantity primarily aﬀects consumer aware-

ness and has no discernible eﬀects on conditional consideration and conditional purchase. How-

ever, when measuring the separate eﬀects of diﬀerent types of advertising content, i.e., ads

with only informational content, ads with only non-informational content, and ads with both

informational and non-informational content, we ﬁnd a more nuanced set of results: advertising

only containing non-informational content increases unaided awareness, while advertising only

containing informational content increases aided awareness and purchase conditional on con-

sideration. We do not ﬁnd signiﬁcant eﬀects of advertising containing both informational and

non-informational content. Since many companies spend a signiﬁcant portion of their budgets

on advertising with both informational and non-informational content, we investigate whether

this type of advertising signiﬁcantly aﬀects certain groups of consumers. We ﬁnd it to increase

consumers’ unaided (and, in some cases, aided awareness) for high risk consumers, consumers

with a change in circumstances, and consumers who recently moved.

References

Ackerberg, D. (2001), “Empirically Distinguishing Informative and Prestige Eﬀects of Advertising,”

RAND Journal of Economics, 32 (2), 316–333.

— (2003), “Advertising, Learning, and Consumer Choice in Experience Good Markets: An Empirical

Examination,” International Economic Review, 44 (3), 1007–1040.

Agarwal, S. and B. Ambrose (2010), “The Eﬀect of Advertising on Home Equity Credit Choices,”

Working Paper, Pennsylvania State University.

Alba, J., J. Hutchinson, and J. Lynch (1991), “Memory and Decision Making,” in H. Kassarjian and

T. Robertson (Editors), “Handbook of Consumer Behavior,” NJ: Prentice Hall, 1–49.

Anderson, S., F. Ciliberto, J. Liaukonyte, and R. Renault (2016), “Push-me Pull-you: Comparative

Advertising in the OTC Analgesics Industry,” RAND Journal of Economics, 47 (4), 1029–1056.

Bagwell, K. (2007), “The Economic Analysis of Advertising,” in Mark Armstrong and Robert H.

Porter (Editors), “The Handbook of Industrial Organization,” New York: North Holland, volume 3,

1701–1844.

Becker, G. and K. Murphy (1993), “A Simple Theory of Advertising as a Good or Bad,” Quarterly

Journal of Economics, 108, 942–964.

Bertrand, M., D. Karlan, S. Mullainathan, E. Shaﬁr, and J. Zinman (2010), “What’s Advertising Con-

tent Worth? Evidence from a Consumer Credit Marketing Field Experiment,” Quarterly Journal

of Economics, 125 (1), 263–306.

Binsted, K., A. Nijholt, O. Stock, C. Strapparava, G. Ritchie, R. Manurung, H. Pain, A. Waller, and

D. O’Mara (2006), “Computational Humor,” IEEE Intelligent Systems, 21 (2), 59–69.

Chamberlin, E. (1933), Cambridge, Massachusetts: Harvard University Press.

Chan, T., Ch. Narasimhan, and Y. Xie (2013), “Treatment Eﬀectiveness and Side Eﬀects: A Model

of Physician Learning,” Management Science, 59 (6), 1309–1325.

Chandrasekaran, A., A. Vijayakumar, S. Antol, M. Bansal, D. Batra, C. Zitnick, and D. Parikh (2016),

“We Are Humor Beings: Understanding and Predicting Visual Humor,” Proceedings of the IEEE

Conference on Computer Vision and Pattern Recognition, 4603–4612.

Ching, A. and M. Ishihara (2012), “Measuring the Informative and Persuasive Roles of Detailing on

Prescribing Decisions,” Management Science, 58 (7), 1374–1387.

Cronqvist, H. (2006), “Advertising and Portfolio Choice,” Working Paper, Ohio State University.

Ebbinghaus, H. (1885), Memory, Reprint: Martino Fine Books, CT: Eastford.

Goeree, M. (2008), “Limited Information and Advertising in the US Personal Computer Industry,”

Econometrica, 76 (5), 1017–1074.

Gurun, U., G. Matvos, and A. Seru (2016), “Advertising Expensive Mortgages,” Journal of Finance,

71, 2371–2416.

Hastings, J., A. Horta¸csu, and Ch. Syverson (2017), “Advertising and Competition in Privatized Social

Security: The Case of Mexico,” Econometrica, 85 (6), 1723–1761.

Heath, Ch. and D. Heath (2008), Made to Stick, Random House, NY: New York, 2nd edition.

Hintzman, D. (1976), “Repetition and Memory,” Psychology of Learning and Motivation, 10, 47–91.

Honka, E., A. Horta¸csu, and M. Vitorino (2017), “Advertising, Consumer Awareness, and Choice:

Evidence from the U.S. Banking Industry,” RAND Journal of Economics, 48 (3), 611–646.

Lee, D., K. Hosanagar, and H. Nair (2018), “Advertising Content and Consumer Engagement on

Social Media: Evidence from Facebook,” Management Science, 64 (11), 5105–5131.

Liaukonyte, J., T. Teixeira, and K. Wilbur (2015), “Television Advertising and Online Shopping,”

Marketing Science, 34 (3), 311–330.

Lovett, M. and R. Staelin (2016), “The Role of Paid and Earned Media in Building an Entertainment

Brand: Reminding, Informing, and Enhancing Enjoyment,” Marketing Science, 35, 142–157.

Lynch, J. and T. Srull (1982), “Memory and Attentional Factors in Consumer Choice: Concepts and

Research Methods,” Journal of Consumer Research, 9 (1), 18–37.

Maddala, G. (1983), Limited-Dependent and Qualitative Variables in Econometrics, NY: Cambridge

University Press, 1st edition edition.

McCloskey, M. and S. Glucksberg (1979), “Decision Processes in Verifying Category Membership

Statements: Implications for Models of Semantic Memory,” Cognitive Psychology, 11 (1), 1–37.

Mullainathan, S., J. Schwartzstein, and A. Shleifer (2008), “Coarse Thinking and Persuasion,” Quar-

terly Journal of Economics, 123 (2), 577–619.

Narayanan, S., P. Manchanda, and P. Chintagunta (2005), “Temporal Diﬀerences in the Role of

Marketing Communication in New Product Categories,” Journal of Marketing Research, 42 (3),

278–290.

Nedungadi, P. (1990), “Recall and Consumer Consideration Sets: Inﬂuencing Choice without Altering

Brand Evaluations,” Journal of Consumer Research, 17 (3), 263–276.

Nelson, P. (1970), “Information and Consumer Behavior,” Journal of Political Economy, 78 (2), 311–

329.

— (1974), “Advertising as Information,” Journal of Political Economy, 82 (4), 729–754.

Pieters, R., L. Warlop, and M. Wedel (2002), “Repetition and Memory,” Management Science, 48 (6),

765–781.

Pires, T. (2016), “Costly Search and Consideration Sets in Storable Goods Markets,” Quantitative

Marketing and Economics, 14 (3), 157–193.

Reisberg, D. and P. Hertel (2004), Memory and Emotion, Oxford University Press, Oxford.

Resnik, Alan and Bruce Stern (1977), “An Analysis of Information Content in Television Advertising,”

Journal of Marketing, 41 (1), 50–53.

Sahni, N. and H. Nair (2018), “Does Advertising Serve as a Signal? Evidence from Field Experiments

in Mobile Search,” Review of Economic Studies, forthcoming.

Sahni, N., Ch. Wheeler, and P. Chintagunta (2018), “Personalization in Email Marketing: The Role

of Non-Informative Advertising Content,” Marketing Science, 37 (2), 236–258.

Shapiro, B. (2018), “Positive Spillovers and Free Riding in Advertising of Prescription Pharmaceuti-

cals: The Case of Antidepressants,” Journal of Political Economy, 126 (1), 381–437.

Shapiro, B., G. Hitsch, and A. Tuchman (2019), “Generalizable and Robust TV Advertising Eﬀects,”

Working Paper, University of Chicago.

Stern, Bruce, Dean Krugman, and Alan Resnik (1981), “Magazine Advertising: An Analysis of Its

Information Content,” Journal of Advertising Research, 21 (2), 39–44.

Stigler, G. and G. Becker (1977), “De Gustibus Non Est Disputandum,” American Economic Review,

67, 76–90.

Tuchman, A., H. Nair, and P. Gardete (2018), “Television Ad-Skipping, Consumption Complementar-

ities and the Consumer Demand for Advertising,” Quantitative Marketing and Economics, 16 (2),

111–174.

Tucker, C. (2015), “The Reach and Persuasiveness of Viral Video Ads,” Marketing Science, 34 (2),

281–296.

Wachtel, S. and Th. Otter (2013), “Successive Sample Selection and Its Relevance for Management

Decisions,” Marketing Science, 32 (1), 170–185.

Zwaan, R. and G. Radvansky (1998), “Situation Models in Language Comprehension and Memory,”

Psychological Bulletin, 2 (123), 162–185.

Tables and Figures

Figure 1: Example of Border Strategy: Austin and San Antonio DMAs

0 10 20 30 40

Percent

−.25 0 .25 .5 .75 1

Fleiss’ Kappa

Figure 2: Histogram of Fleiss’ Kappas

(Median = 0.70 and Mean = 0.68)

0 10 20 30 40

Percent

−.002 −.001 0 .001 .002

Figure 3: Histogram of Residuals from Regression of Log(1+Total Advertising

Spending per Household) on Full Set of Fixed Eﬀects

(Truncated at -0.002 and 0.002)

(a) Only Informational Ads

0 10 20 30 40 50 60 70

Percent

−.002 −.001 0 .001 .002

(b) Only Non-Informational Ads

0 10 20 30 40 50 60 70

Percent

−.002 −.001 0 .001 .002

Informational Ads

0 10 20 30 40 50 60 70

Percent

−.002 −.001 0 .001 .002

Figure 4: Histogram of Residuals from Regression of Log(1+Total Ad Content

Spending per Household) on Full Set of Fixed Eﬀects

(Truncated at -0.002 and 0.002)

(i) (ii) (iii)

Total Ad Total Ad Units Local Ad

Expenditure Expenditure

Brand Per Household in $ Per Household in $

21st Century 0.0141 2,029 0.0004

AAA 0.0095 43 0.0095

Allstate 0.1230 4,703 0.0021

American Family 0.0041 13 0.0041

Amica 0.0112 244 0.0094

Auto Owners 0.0002 1 0.0002

Erie 0.0001 0 0.0001

Esurance 0.0956 3,528 0.0005

Farmers 0.0096 487 0.0014

Geico 0.3303 10,275 0.0142

GMAC 0.0000 0 0.0000

Hartford 0.0090 253 0.0015

Liberty Mutual 0.1117 5,041 0.0016

Mercury 0.0008 50 0.0003

Metlife 0.0008 26 0.0000

Nationwide 0.0553 1,374 0.0023

Progressive 0.2018 9,337 0.0182

Safeco 0.0000 1 0.0000

State Farm 0.1241 5,013 0.0042

Travelers 0.0068 258 0.0010

USAA 0.0388 742 0.0017

Average 0.0546 2,068 0.0035

Table 1: Monthly TV Advertising Quantities

(i) (ii) (iii) (iv)

log



1 + A

DM A

bdt



log



1 + A

DM A,f

bdt



log



1 + A

DM A,nf

bdt



log



1 + A

DM A,f,nf

bdt



Advertising Data

Mean Absolute Diﬀerence 0.0077 0.0004 0.0008 0.0027

Standard Deviation of Absolute Diﬀerence 0.0497 0.0057 0.0070 0.0135

Proportion of Brand-Border Combinations

with Variation Across Time 74% 39% 59% 57%

Proportion of Brand-Month Combinations

with Variation Across Geographies 72% 27% 48% 51%

Advertising Residuals

Mean Absolute Diﬀerence 0.0030 0.0005 0.0012 0.0019

Standard Deviation of Absolute Diﬀerence 0.0086 0.0041 0.0059 0.0078

Proportion of Brand-Border Combinations

with Variation Across Time 63% 28% 34% 49%

Proportion of Brand-Month Combinations

with Variation Across Geographies 83% 45% 51% 65%

Table 2: Discontinuities in DMA-Level Advertising

(i) (ii) (iii)

Both Informational

Brand Informational Only Non-Informational Only and Non-Informational

21st Century 0.00 0.00 100.00

AAA 0.04 19.86 80.10

Allstate 41.03 27.49 31.48

American Family 0.00 76.07 23.93

Amica Mutual 1.38 10.90 87.73

Auto Owners 0.00 99.33 0.67

Erie 100.00 0.00 0.00

Esurance 18.34 9.68 71.98

Farmers 1.31 69.83 28.86

Geico 0.30 0.94 98.76

Hartford 1.31 15.70 82.98

Liberty Mutual 13.89 1.94 84.18

Mercury 0.62 12.96 86.42

Metlife 57.16 0.19 42.64

Nationwide 21.26 19.67 58.91

Progressive 1.15 10.89 87.96

Safeco 0.00 100.00 0.00

State Farm 21.81 42.28 35.91

Travelers 0.00 54.32 45.68

USAA 0.00 99.73 0.27

Average 13.98 33.59 52.42

Table 3: Percentage of Ads (Weighted by Spending) Containing Informational

Content Only, Non-Informational Content Only, and Both Informational and Non-

Informational Content

(i) (ii) (iii)

Demographics All Consumers Non-Shoppers Shoppers

Age ≤ 25 Years 0.0455 0.0395 0.0580

26 Years < Age ≤ 45 Years 0.3598 0.3548 0.3702

46 Years < Age ≤ 65 Years 0.4440 0.4515 0.4284

Age > 65 Years 0.1507 0.1542 0.1434

Male 0.4170 0.4038 0.4445

Black 0.0432 0.0375 0.0550

Hispanic 0.0309 0.0212 0.0511

Asian 0.0633 0.0714 0.0466

Married 0.5759 0.5708 0.5863

College Degree 0.5870 0.6239 0.5103

Income Greater than $100k 0.2468 0.2524 0.2351

Lived in Urban Area 0.1510

Someone under 25 Years Insured under the Policy 0.1357

Shopped for Homeowner’s Insurance 0.3410

Shopped for Renter’s Insurance 0.0978

Shopped for Life Insurance 0.0386

Shopped for Personal Umbrella Insurance 0.0568

Two or More Accident(s) in the Last 3 Years 0.0288

Two or More Ticket(s) in the Last 3 Years 0.0385

Poor Credit History 0.0758

Same Insurer as in Previous Year 0.5201

Table 4: Descriptive Statistics

(i) (ii) (iii) (iv) (v) (vi)

All Consumers Non-Shoppers Shoppers

Brand Unaided Aided Unaided Aided Unaided Aided

21stCentury 0.0765 0.5108 0.0505 0.5057 0.1306 0.5198

AAA 0.2652 0.8543 0.2733 0.8779 0.2483 0.8125

Allstate 0.6403 0.9503 0.6167 0.9482 0.6894 0.9541

American Family 0.0730 0.3337 0.0635 0.3333 0.0926 0.3344

Amica 0.0263 0.1408 0.0134 0.0843 0.0533 0.2409

Auto Owners 0.0214 0.1006 0.0145 0.0864 0.0358 0.1257

Erie 0.0284 0.0978 0.0223 0.0836 0.0411 0.1230

Esurance 0.1017 0.6334 0.0588 0.6061 0.1907 0.6817

Farmers 0.4051 0.8760 0.4184 0.8922 0.3775 0.8473

Geico 0.6099 0.9505 0.5635 0.9463 0.7062 0.9579

GMAC 0.0106 0.2321 0.0037 0.2312 0.0248 0.2338

Hartford 0.0900 0.6655 0.0638 0.6614 0.1445 0.6726

Liberty Mutual 0.1374 0.7963 0.0928 0.7830 0.2300 0.8199

Mercury 0.0553 0.2883 0.0593 0.3436 0.0471 0.1905

Metlife 0.0486 0.7629 0.0347 0.7600 0.0774 0.7680

Nationwide 0.1979 0.8352 0.1661 0.8255 0.2639 0.8522

Progressive 0.4631 0.9181 0.3934 0.9076 0.6081 0.9367

Safeco 0.0450 0.3594 0.0308 0.3519 0.0744 0.3725

State Farm 0.7132 0.9640 0.7152 0.9662 0.7092 0.9603

Travelers 0.1015 0.7310 0.0740 0.7191 0.1586 0.7522

USAA 0.1020 0.4851 0.0968 0.4632 0.1127 0.5238

Average 0.2006 0.5946 0.1822 0.5894 0.2389 0.6038

Table 5: Awareness Probabilities

Considered Chosen Aware Aware Considered

(unaided) (aided)

Brand → Considered → Considered → Chosen

21st Century 0.1378 0.0424 0.9005 0.2524 0.3075

AAA 0.2343 0.0955 0.8110 0.2854 0.4076

Allstate 0.3342 0.0522 0.4980 0.3509 0.1563

American Family 0.2005 0.0803 0.7071 0.2447 0.4007

Amica 0.0478 0.0192 0.7464 0.1920 0.4022

Auto Owners 0.0677 0.0424 0.8489 0.3068 0.6257

Erie 0.1609 0.1024 0.8062 0.3720 0.6366

Esurance 0.1759 0.0524 0.8535 0.2559 0.2979

Farmers 0.2259 0.0539 0.4942 0.2544 0.2388

Geico 0.4072 0.0614 0.6101 0.4282 0.1508

GMAC 0.0344 0.0191 0.9329 0.1386 0.5563

Hartford 0.1281 0.0494 0.7557 0.1861 0.3853

Liberty Mutual 0.1644 0.0481 0.7078 0.1997 0.2927

Mercury 0.0728 0.0356 0.8210 0.2478 0.4890

Metlife 0.0742 0.0360 0.7632 0.0958 0.4848

Nationwide 0.1527 0.0449 0.5418 0.1771 0.2941

Progressive 0.3892 0.0695 0.6716 0.4165 0.1786

Safeco 0.0864 0.0501 0.8589 0.2192 0.5798

State Farm 0.3538 0.0541 0.5157 0.3699 0.1529

Travelers 0.1178 0.0469 0.6944 0.1546 0.3983

USAA 0.0912 0.0469 0.7640 0.1691 0.5141

Average 0.1778 0.0500 0.6426 0.2627 0.2811

Table 6: Consideration, Purchase, and Conversion Probabilities (Shoppers Only)

(i) (ii) (iii) (iv) (v) (vi)

Unaided Awareness Aided Awareness Consideration Consideration Choice Full Information

Conditional on Unaided Awareness Aided Awareness Consideration

Border Strategy

Advertising Spending per Household in $ * 0.2544

0.2152

-0.1220 0.0186 0.2199 0.0779

(0.0712) (0.0998) (0.4503) (0.3961) (0.4891) (0.0321)

Same Insurer as in Previous Year 0.1947

0.8756

(Yes = 1) (0.0244) (0.0295)

Insurer Provided the Best Price 0.7170

0.4456

(Yes = 1) (0.0243) (0.0234)

Brand-Demographics-Year FEs yes yes yes yes yes yes

Online-Brand-Demographics-Year FEs yes yes yes yes yes yes

Brand-State-Month FEs yes yes yes yes yes yes

Brand-Border-DMA FEs yes yes yes yes yes yes

Brand-Border-Month FEs yes yes yes yes yes yes

Survey FEs yes yes yes yes yes yes

Eﬀective Number of Observations

†

1,754,151 1,154,013 31,383 100,914 13,558 1,511,581

Without Border Strategy

Advertising Spending per Household in $ * 0.1500

0.0965 -0.5846 0.0682 0.2625 0.0418

(0.0690) (0.0670) (0.3314) (0.1960) (0.2302) (0.0203)

Same Insurer as in Previous Year 0.2085

0.8520

(Yes = 1) (0.0132) (0.0332)

Insurer Provided the Best Price 0.7144

0.4541

(Yes = 1) (0.0171) (0.0239)

Brand-Demographics-Year FEs yes yes yes yes yes yes

Online-Brand-Demographics-Year FEs yes yes yes yes yes yes

Brand-State-Month FEs yes yes yes yes yes yes

Brand-DMA FEs yes yes yes yes yes yes

Survey FEs yes yes yes yes yes yes

Eﬀective Number of Observations

†

1,778,112 1,178,583 49,299 132,536 27,043 1,532,965

Table 7: Results for Advertising Quantity

a: <.001, b: <.01, c: < .05.

Standard errors in parentheses (clustered at the DMA level).

* Measured on a logarithmic scale.

†

After dropping observations due to collinearity with ﬁxed eﬀects.

(i) (ii) (iii) (iv) (v) (vi)

Unaided Awareness Aided Awareness Consideration Consideration Choice Full Information

Conditional on Unaided Awareness Aided Awareness Consideration

Border Strategy

Spending per Household in $ on Ads with...

... Informational Content Only * -0.0111 0.5804

-1.4005 -0.1008 2.6961

0.0087

(0.2125) (0.2720) (0.9471) (0.6710) (1.1324) (0.1705)

... Non-Informational Content Only * 0.1728

0.0097 -0.5459 -0.6025 -0.2891 0.1974

(0.0551) (0.0418) (0.4456) (0.3273) (0.3863) (0.0848)

... Both Informational and Non-Informational Content * 0.2005 0.0086 0.6391 0.3968 -0.1305 0.0033

(0.2049) (0.1163) (0.6999) (0.5880) (0.6849) (0.0806)

Same Insurer as in Previous Year 0.1950

0.8525

(Yes = 1) (0.0244) (0.0335)

Insurer Provided the Best Price 0.7166

0.4564

(Yes = 1) (0.0244) (0.0266)

Brand-Demographics-Year FEs yes yes yes yes yes yes

Online-Brand-Demographics-Year FEs yes yes yes yes yes yes

Brand-State-Month FEs yes yes yes yes yes yes

Brand-Border-DMA FEs yes yes yes yes yes yes

Brand-Border-Month FEs yes yes yes yes yes yes

Survey FEs yes yes yes yes yes yes

Eﬀective Number of Observations

†

1,754,151 1,154,013 31,383 100,914 13,558 1,517,379

Without Border Strategy

Spending per Household in $ on...

... Informational Content Only * 0.6898

0.5866

-0.8119 -0.5249 0.1628 -0.0000

(0.2039) (0.1822) (0.6878) (0.5160) (0.5826) (0.1690)

... Non-Informational Content Only * 0.0533 -0.0157 -0.5946 -0.1749 0.0284 0.1924

(0.0656) (0.0321) (0.6051) (0.4640) (0.3211) (0.0847)

... Both Informational and Non-Informational Content * 0.2044 -0.0906 -0.8636 0.0653 0.4177 -0.0079

(0.1345) (0.1031) (0.4848) (0.2819) (0.4283) (0.0800)

Same Insurer as in Previous Year 0.2085

0.8524

(Yes = 1) (0.0133) (0.0335)

Insurer Provided the Best Price 0.7144

0.4564

(Yes = 1) (0.0171) (0.0265)

Brand-Demographics-Year FEs yes yes yes yes yes yes

Online-Brand-Demographics-Year FEs yes yes yes yes yes yes

Brand-State-Month FEs yes yes yes yes yes yes

Brand-DMA FEs yes yes yes yes yes yes

Survey FEs yes yes yes yes yes yes

Eﬀective Number of Observations

†

1,778,112 1,178,583 49,299 132,536 27,043 1,517,581

Table 8: Results for Advertising Content

a: <.001, b: <.01, c: < .05.

Standard errors in parentheses (clustered at the DMA level).

* Measured on a logarithmic scale.

†

After dropping observations due to collinearity with ﬁxed eﬀects.

(i) (ii) (iii) (iv) (v)

Unaided Awareness Aided Awareness Consideration Consideration Choice

Conditional on Unaided Awareness Aided Awareness Consideration

High Risk Consumers

Spending per Household in $ on Ads with...

... Informational Content Only * -0.0577 0.5862

-1.1798 -0.0583 2.7136

(0.2072) (0.268) (1.0090) (0.6705) (1.1282)

... Non-Informational Content Only * 0.1655

0.0104 -0.4693 -0.5591 -0.2875

(0.0562) (0.0418) (0.4270) (0.3214) (0.3950)

... Both Informational and 0.1932 0.0062 0.6040 0.3740 -0.1179

Non-Informational Content * (0.2052) (0.1162) (0.7139) (0.5852) (0.6690)

High Risk Consumer ×

Spending per Household in $ on...

... Informational Content Only * 0.4163 -0.1814 -1.5684 -0.3171 -0.2357

(0.2882) (0.2624) (1.1000) (0.5133) (0.4742)

... Non-Informational Content Only * 0.3609

-0.0598 -0.7580 -0.3470 -0.5136

(0.1591) (0.1482) (0.5350) (0.3752) (0.4662)

... Both Informational and 0.3039

0.0965

-0.0421 0.1292 0.1797

Non-Informational Content * (0.0657) (0.0452) (0.1730) (0.1316) (0.0973)

Same Insurer as in Previous Year 0.1949

(Yes = 1) (0.0246)

Insurer Provided the Best Price 0.7164

(Yes = 1) (0.0243)

Brand-Demographics-Year FEs yes yes yes yes yes

Online-Brand-Demographics-Year FEs yes yes yes yes yes

Brand-State-Month FEs yes yes yes yes yes

Brand-Border-DMA FEs yes yes yes yes yes

Brand-Border-Month FEs yes yes yes yes yes

Survey FEs yes yes yes yes yes

Eﬀective Number of Observations

†

1,754,151 1,154,013 31,383 100,914 13,558

Moved Last Year

Spending per Household in $ on...

... Informational Content Only * -0.0656 0.5261 -1.3788 -0.1343 2.6957

(0.2131) (0.2685) (0.9369) (0.6834) (1.1402)

... Non-Informational Content Only * 0.1714

0.0081 -0.5484 -0.6262 -0.2637

(0.0579) (0.0417) (0.4851) (0.3449) (0.4053)

... Both Informational and 0.2070 0.0124 0.6242 0.3921 -0.1337

Non-Informational Content * (0.2064) (0.1168) (0.6980) (0.5922) (0.6922)

Moved Last Year ×

Spending per Household in $ on...

... Informational Content Only * 0.7469 0.5945 -0.7917 -0.3636 0.2175

(0.6220) (0.5395) (1.0081) (0.7754) (0.6631)

... Non-Informational Content Only * 0.1145 0.2246 0.0037 0.1293 -0.1873

(0.2880) (0.3160) (0.7551) (0.4751) (0.3430)

... Both Informational and 0.2776

0.0457 -0.0072 0.2218 0.0351

Non-Informational Content * (0.1009) (0.0731) (0.2564) (0.1538) (0.1470)

Same Insurer as in Previous Year 0.1950

(Yes = 1) (0.0244)

Insurer Provided the Best Price 0.7165

(Yes = 1) (0.0244)

Brand-Demographics-Year FEs yes yes yes yes yes

Online-Brand-Demographics-Year FEs yes yes yes yes yes

Brand-State-Month FEs yes yes yes yes yes

Brand-Border-DMA FEs yes yes yes yes yes

Brand-Border-Month FEs yes yes yes yes yes

Survey FEs yes yes yes yes yes

Eﬀective Number of Observations

†

1,754,151 1,154,013 31,383 100,914 13,558

Table 9: Results for Advertising Content – Eﬀect Heterogeneity

a: <.001, b: <.01, c: < .05.

Standard errors in parentheses (clustered at the DMA level).

* Measured on a logarithmic scale.

†

After dropping observations due to collinearity with ﬁxed eﬀects.

(i) (ii) (iii) (iv) (v)

Unaided Awareness Aided Awareness Consideration Consideration Choice

Conditional on Unaided Awareness Aided Awareness Consideration

Change in Circumstances

Spending per Household in $ on Ads with...

... Informational Content Only * -0.0644 0.4595 -0.7992 0.0553 2.1624

(0.2106) (0.2711) (1.0147) (0.6642) (1.2083)

... Non-Informational Content Only * 0.1658

0.0054 -0.5963 -0.7051

-0.3186

(0.0569) (0.0417) (0.4807) (0.3401) (0.3871)

... Both Informational and 0.1868 -0.0071 0.5389 0.3735 -0.0545

Non-Informational Content * (0.2039) (0.1159) (0.7033) (0.5786) (0.6813)

Change in Circumstances ×

Spending per Household in $ on Ads with...

... Informational Content Only * 0.3285 0.6986

-1.1178 -0.5810 1.2889

(0.2677) (0.2215) (0.7157) (0.3401) (0.7190)

... Non-Informational Content Only * 0.4949

0.3099

0.5235 0.5224 0.0762

(0.1331) (0.1107) (0.5059) (0.2800) (0.2097)

... Both Informational and 0.1681

0.1027

0.1162 0.0550 -0.1031

Non-Informational Content * (0.0449) (0.0272) (0.0836) (0.0609) (0.0566)

Same Insurer as in Previous Year 0.1944

(Yes = 1) (0.0243)

Insurer Provided the Best Price 0.7174

(Yes = 1) (0.0246)

Brand-Demographics-Year FEs yes yes yes yes yes

Online-Brand-Demographics-Year FEs yes yes yes yes yes

Brand-State-Month FEs yes yes yes yes yes

Brand-Border-DMA FEs yes yes yes yes yes

Brand-Border-Month FEs yes yes yes yes yes

Survey FEs yes yes yes yes yes

Eﬀective Number of Observations

†

1,754,151 1,154,013 31,383 100,914 13,558

Table 9: Results for Advertising Content – Eﬀect Heterogeneity (Continued)

a: <.001, b: <.01, c: < .05.

Standard errors in parentheses (clustered at the DMA level).

* Measured on a logarithmic scale.

†

After dropping observations due to collinearity with ﬁxed eﬀects.

Web Appendix A: Advertising Content Classiﬁcations

Coders were provided with the following set of instructions to classify advertising content into

categories:

• Price / Rate / Discount

Does the ad mention any price-related information? For example: competitive rate, low-

cost, premium, discount, budget, saving you money, save 28% on your premium.

• (Non-Price) Product Features

Does the ad talk about (new) non-price characteristics of the insurance product? For

example: accident forgiveness, safe driver discount, getting competitive quotes on its

website.

• Brand Name Focus

Is the focus of the ad the name of the insurance brand? Ads with brand name focus can

usually be found in “holiday wishes”, “thank you to the customers”, or it could be that

mentioning the company name is the biggest part of the ad.

• Humor / Fun / Entertainment

Is the ad funny/entertaining? For example: it’s a cartoon or cars do weird things or dam-

age from a superhero ﬁght is covered. Note: about humor/entertainment and fear/safety

concerns of an ad. Please pick one (dominant) emotion for the ad coding. For example,

Allstate has ads with a guy causing crazy mischief – that falls under humor as these ads

are intended to be fun/enjoyed by consumers. Fear/safety concerns means that the ad is

trying to instill serious (not funny) concerns into the consumer.

• Fear / Need for Safety

Does the ad try to induce fear or a need for (ﬁnancial) safety? For example, fear of an

accident / what will happen after an accident, fear of losing money / car, need to make

family (ﬁnancially) safe in case of disaster.

Web Appendix B: Descriptive Statistics for Original Data

In this appendix, we compare descriptive statistics from the ﬁnal (border) samples to those

from the original data sets. Table B-1 shows the statistics for all consumers (columns (i) and

(ii)), non-shoppers (columns (iii) and (iv)), and shoppers (columns (v) and (vi)). Comparing

all consumers who belong to the original and ﬁnal samples, we ﬁnd few diﬀerences among the

demographics. Consumers from the ﬁnal sample who live in border regions (column (i)) are

more likely to be between 45 and 65 years old, more likely to be married, and less likely to have

a college degree compared to consumers from the original sample (column (ii)). The diﬀerences

for the other demographic variables are all smaller than two percentage points.

=========================

Insert Table B-1 about here

=========================

Comparing non-shoppers from the ﬁnal and original data samples (columns (iii) and (iv)),

we ﬁnd non-shoppers from the ﬁnal sample to be less likely younger than 45 years, more likely

to be between 45 and 65 years, more likely to be married, and less likely to have a college degree

compared to non-shoppers from the original sample. Among the remaining demographics, the

diﬀerences are all smaller than two percentage points.

Comparing shoppers from the original and ﬁnal samples (columns (v) and (vi) in Table

B-1), we ﬁnd shoppers from the ﬁnal sample to be less likely to have a college degree and to be

less likely to have an income above $100k compared to shoppers from the original sample. The

diﬀerences for the other demographic variables are all smaller than 2 percentage points. For

shoppers only, we have access to more information on insurance-related consumer characteristics

such as other insurance products the consumer shopped for or past accidents and tickets. Among

these variables, we ﬁnd – not surprisingly – that shoppers from the ﬁnal sample are less likely

to live in an urban area than shoppers from the original sample. Three other diﬀerences are

that shoppers from the ﬁnal sample are less likely to also have been shopping for homeowner

insurance, more likely to have a poor credit history, and less likely to stay with their previous

insurance provider compared to shoppers from the original sample. The diﬀerences for the

other insurance-related variables are all smaller than two percentage points.

We conclude that the distributions of demographic and insurance-related variables for all

consumers and the subgroups of shoppers and non-shoppers in the original data are largely

similar to those in the ﬁnal data.

(i) (ii) (iii) (iv) (v) (vi)

All Consumers Non-Shoppers Shoppers

Demographics Final Original Final Original Final Original

Age ≤ 25 Years 0.0455 0.0490 0.0395 0.0421 0.0580 0.0632

25 Years < Age ≤ 45 Years 0.3598 0.3869 0.3548 0.3791 0.3702 0.4030

45 Years < Age ≤ 65 Years 0.4440 0.4299 0.4515 0.4391 0.4284 0.4106

Age > 65 Years 0.1507 0.1343 0.1542 0.1397 0.1434 0.1232

Male 0.4170 0.4131 0.4038 0.4016 0.4445 0.4370

Black 0.0432 0.0531 0.0375 0.0465 0.0550 0.0667

Hispanic 0.0309 0.0299 0.0212 0.0186 0.0511 0.0532

Asian 0.0633 0.0628 0.0714 0.0656 0.0466 0.0568

Married 0.5759 0.5518 0.5708 0.5455 0.5863 0.5650

College Degree 0.5870 0.6260 0.6239 0.6624 0.5103 0.5505

Income Greater than $100k 0.2468 0.2565 0.2524 0.2613 0.2351 0.2465

Lived in Urban Area 0.1510 0.2061

Someone under 25 Years Insured under the Policy 0.1357 0.1387

Shopped for Homeowner Insurance 0.3410 0.3546

Shopped for Renters Insurance 0.0978 0.1102

Shopped for Life Insurance 0.0386 0.0343

Shopped for Personal Umbrella Insurance 0.0568 0.0611

Two or More Accident(s) in the Last 3 Years 0.0288 0.0336

Two or More Ticket(s) in the Last 3 Years 0.0385 0.0406

Poor Credit History 0.0758 0.0647

Same Insurer as in Previous Year 0.5201 0.5121

Table B-1: Descriptive Statistics

Web Appendix C: Details on Consumer Shopping Behav-

ior

In Figure C-1, we show the distributions of awareness set sizes for all consumers, shoppers,

and non-shoppers separately. The left panel shows the distributions of unaided awareness set

sizes and the right panel shows the distributions of aided awareness set sizes. Across the three

groups of consumers, the distributions have similar shapes. However, the right tail of the

unaided awareness set size distribution for shoppers has more mass than the right tail of the

unaided awareness set size distribution for non-shoppers pointing to shoppers being aware of

more brands.

=========================

Insert Figure C-1 about here

=========================

Figure C-2 visualizes the relationship between unaided and aided awareness set sizes for all

consumers. The Pearson correlation coeﬃcient between consumers’ unaided and aided aware-

ness set sizes is 0.44 (p < .001), i.e. there is a relatively consistent though far from perfect

relationship between unaided and aided awareness set sizes. Between 2010 and 2016, average

unaided awareness set sizes increased from 3.73 to 4.09 (increase of 10%), while aided awareness

set sizes decreased from 13.22 to 12.36 (decrease of 7%). This latter decrease is likely due to

several companies either being acquired by another auto insurance company and stopping to

sell insurance under their old brand name (e.g. 21st Century) or re-branding (GMAC is now

National General).

=========================

Insert Figure C-2 about here

=========================

In the following, we focus on shoppers and their consideration and purchase decisions. The

top half of Figure C-3 visualizes the relationship between unaided awareness and consideration

set sizes and the bottom half of Figure C-3 visualizes the relationship between aided awareness

and consideration set sizes. Recall that shoppers are, on average, aware of 5.01 brands (un-

aided) and 12.67 brands (aided). On average, they consider 3.23 brands which includes their

previous insurance provider.

The Pearson correlation coeﬃcient between shoppers’ unaided

awareness and consideration set sizes is 0.73 (p < .001) and the Pearson correlation coeﬃcient

between shoppers’ aided awareness and consideration set sizes is 0.22 (p < .001), i.e. the rela-

tionship between shoppers’ unaided awareness and consideration set sizes is much closer than

the relationship between shopppers’ aided awareness and consideration set sizes.

=========================

Insert Figure C-3 about here

=========================

Unaided Awareness Aided Awareness

0 10 20 30

Percent

0 2 4 6 8 10 12 14 16 18 20 22

All Consumers

0 10 20 30

Percent

0 2 4 6 8 10 12 14 16 18 20 22

All Consumers

0 10 20 30

Percent

0 2 4 6 8 10 12 14 16 18 20 22

Non−Shoppers

0 10 20 30

Percent

0 2 4 6 8 10 12 14 16 18 20 22

Non−Shoppers

0 10 20 30

Percent

0 2 4 6 8 10 12 14 16 18 20 22

Shoppers

0 10 20 30

Percent

0 2 4 6 8 10 12 14 16 18 20 22

Shoppers

Figure C-1: Awareness Set Sizes

Average consideration set sizes increased from 3.14 in 2010 - 2012 to 3.25 in 2014 - 2016.

0 5 10 15 20

Size of Aided Awareness Set

0510152025

Percent

0 5 10 15 20

Size of Aided Awareness Set

0 5 10 15 20

Size of Unaided Awareness Set

0510152025

Percent

0 5 10 15 20

Size of Unaided Awareness Set

Figure C-2: Aided and Unaided Awareness Set Sizes

Unaided Awareness and Consideration

0 5 10 15 20

Size of Unaided Awareness Set

051015202530

Percent

0 5 10 15 20

Size of Unaided Awareness Set

0 5 10 15 20

Size of Consideration Set

051015202530

Percent

0 5 10 15 20

Size of Consideration Set

Aided Awareness and Consideration

0 5 10 15 20

Size of Aided Awareness Set

051015202530

Percent

0 5 10 15 20

Size of Aided Awareness Set

0 5 10 15 20

Size of Consideration Set

051015202530

Percent

0 5 10 15 20

Size of Consideration Set

Figure C-3: Awareness and Consideration Set Sizes (Shoppers Only)

Web Appendix D: Results without Border Strategy – All

Consumers Living in Top 130 DMAs

(i) (ii) (iii) (iv) (v)

Unaided Awareness Aided Awareness Consideration Choice

Conditional on Unaided Awareness Aided Awareness Consideration

Advertising Quantity

Advertising Spending per Household in $ * 0.1944

0.0784 -0.0791 0.2056 0.1829

(0.0384) (0.0487) (0.2084) (0.1304) (0.1729)

Same Insurer as in Previous Year 0.1935

(Yes = 1) (0.0076)

Insurer Provided the Best Price 0.7385

(Yes = 1) (0.0089)

Brand-Demographics-Year FEs yes yes yes yes yes

Online-Brand-Demographics-Year FEs yes yes yes yes yes

Brand-DMA FEs yes yes yes yes yes

Brand-State-Month FEs yes yes yes yes yes

Survey FEs yes yes yes yes yes

Eﬀective Number of Observations

†

3,579,114 2,360,862 94,338 252,425 50,767

Advertising Content

Spending per Household in $ on...

... Informational Content * 0.3578

0.2643 -0.7620 -0.3214 -0.3313

(0.1284) (0.1441) (0.3873) (0.2968) (0.4954)

... Non-Informational Content * 0.0641 0.0137 -0.0193 0.1194 0.5222

(0.0360) (0.0336) (0.2820) (0.2208) (0.3020)

... Both Informational and 0.2733

-0.1416 -0.2150 0.3347

-0.0730

Non-Informational Content * (0.0757) (0.0740) (0.2917) (0.1605) (0.2196)

Same Insurer as in Previous Year 0.1922

(Yes = 1) (0.0076)

Insurer Provided the Best Price 0.7389

(Yes = 1) (0.0091)

Brand-Demographics-Year FEs yes yes yes yes yes

Online-Brand-Demographics-Year FEs yes yes yes yes yes

Brand-DMA FEs yes yes yes yes yes

Brand-State-Month FEs yes yes yes yes yes

Survey FEs yes yes yes yes yes

Eﬀective Number of Observations

†

3,492,825 2,288,223 92,752 248,557 49,871

Table D-1: Results withour Border Strategy – All Consumers Living in Top 130

DMAs

a: <.001, b: <.01, c: < .05.

Standard errors in parentheses (clustered at the DMA level).

* Measured on a logarithmic scale.

†

After dropping observations due to collinearity with ﬁxed eﬀects.

Web Appendix E: Eﬀect Heterogeneity – Advertising Quan-

tity

(i) (ii) (iii) (iv) (v)

Unaided Awareness Aided Awareness Consideration Choice

Conditional on Unaided Awareness Aided Awareness Consideration

High Risk Consumers

Advertising Spending per Household $ * 0.2378

0.2155

-0.1040 0.0128 0.2148

(0.0708) (0.0998) (0.4546) (0.3980) (0.4847)

High Risk Consumers ×

Advertising Spending per Household $ * 0.2440

-0.0102 -0.1779 0.0390 0.1228

(0.0618) (0.0426) (0.1432) (0.0896) (0.0797)

Same Insurer as in Previous Year 0.1945

(Yes = 1) (0.0245)

Insurer Provided the Best Price 0.7173

(Yes = 1) (0.0244)

Brand-Demographics-Year FEs yes yes yes yes yes

Online-Brand-Demographics-Year FEs yes yes yes yes yes

Brand-Border-DMA FEs yes yes yes yes yes

Brand-Border-Month FEs yes yes yes yes yes

Brand-State-Month FEs yes yes yes yes yes

Survey FEs yes yes yes yes yes

Eﬀective Number of Observations

†

1,754,151 1,154,013 31,383 100,914 13,558

Moved Last Year

Advertising Spending per Household $ * 0.2513

0.2147

-0.1033 0.0103 0.2159

(0.0710) (0.1002) (0.4491) (0.3973) (0.4915)

Moved Last Year ×

Advertising Spending per Household $ * 0.2973

-0.0577 -0.1534 0.1202 0.0433

(0.0989) (0.0865) (0.1406) (0.1373) (0.1124)

Same Insurer as in Previous Year 0.1946

(Yes = 1) (0.0244)

Insurer Provided the Best Price 0.7169

(Yes = 1) (0.0243)

Brand-Demographics-Year FEs yes yes yes yes yes

Online-Brand-Demographics-Year FEs yes yes yes yes yes

Brand-Border-DMA FEs yes yes yes yes yes

Brand-Border-Month FEs yes yes yes yes yes

Brand-State-Month FEs yes yes yes yes yes

Survey FEs yes yes yes yes yes

Eﬀective Number of Observations

†

1,754,151 1,154,013 31,383 100,914 13,558

Change in Circumstances

Advertising Spending per Household $ * 0.2432

0.2046

-0.1812 -0.0053 0.2451

(0.0698) (0.0972) (0.4500) (0.3920) (0.4788)

Change in Circumstances ×

Advertising Spending per Household $ * 0.2406

0.1764

0.1377

0.1009 -0.0616

(0.0481) (0.0344) (0.0674) (0.0538) (0.0397)

Same Insurer as in Previous Year 0.1945

(Yes = 1) (0.0244)

Insurer Provided the Best Price 0.7172

(Yes = 1) (0.0243)

Brand-Demographics-Year FEs yes yes yes yes yes

Online-Brand-Demographics-Year FEs yes yes yes yes yes

Brand-Border-DMA FEs yes yes yes yes yes

Brand-Border-Month FEs yes yes yes yes yes

Brand-State-Month FEs yes yes yes yes yes

Survey FEs yes yes yes yes yes

Eﬀective Number of Observations

†

1,754,151 1,154,013 31,383 100,914 13,558

Table E-1: Eﬀect Heterogeneity – Advertising Quantity

a: <.001, b: <.01, c: < .05.

Standard errors in parentheses (clustered at the DMA level).

* Measured on a logarithmic scale.

†

After dropping observations due to collinearity with ﬁxed eﬀects.

Web Appendix F: Robustness Checks

(i) (ii) (iii) (iv) (v)

Unaided Awareness Aided Awareness Consideration Choice

Conditional on Unaided Awareness Aided Awareness Consideration

Total Advertising Units

Advertising Units * 0.0012 0.0083

-0.0089 -0.0113 -0.0268

(0.0020) (0.0027) (0.0190) (0.0086) (0.0291)

Same Insurer as in Previous Year 0.1948

(Yes = 1) (0.0244)

Insurer Provided the Best Price 0.7167

(Yes = 1) (0.0244)

Brand-Demographics-Year FEs yes yes yes yes yes

Online-Brand-Demographics-Year FEs yes yes yes yes yes

Brand-Border-DMA FEs yes yes yes yes yes

Brand-Border-Month FEs yes yes yes yes yes

Brand-State-Month FEs yes yes yes yes yes

Survey FEs yes yes yes yes yes

Eﬀective Number of Observations

†

1,754,151 1,154,013 31,383 100,914 13,558

DMA-Level Advertising Spending Only

Advertising Spending per Household $ * 0.1549

0.0833 -0.0899 0.0189 0.1367

(0.0754) (0.0564) (0.3932) (0.3203) (0.3833)

Same Insurer as in Previous Year 0.1947

(Yes = 1) (0.0244)

Insurer Provided the Best Price 0.7170

(Yes = 1) (0.0243)

Brand-Demographics-Year FEs yes yes yes yes yes

Online-Brand-Demographics-Year FEs yes yes yes yes yes

Brand-Border-DMA FEs yes yes yes yes yes

Brand-Border-Month FEs yes yes yes yes yes

Brand-State-Month FEs yes yes yes yes yes

Survey FEs yes yes yes yes yes

Eﬀective Number of Observations

†

1,754,151 1,154,013 31,383 100,914 13,558

Table F-1: Robustness Checks – Results for Advertising Quantity

a: <.001, b: <.01, c: < .05.

Standard errors in parentheses (clustered at the DMA level).

* Measured on a logarithmic scale.

†

After dropping observations due to collinearity with ﬁxed eﬀects.

(i) (ii) (iii) (iv) (v)

Unaided Awareness Aided Awareness Consideration Choice

Conditional on Unaided Awareness Aided Awareness Consideration

High Kappa Content Only

Spending per Household in $ on...

... Informational Content Only * -0.0443 0.9623

-1.3610 -0.0237 2.6779

(0.2434) (0.2956) (0.9373) (0.6943) (1.1343)

... Non-Information Content Only * 0.1622

0.0050 -0.1788 -0.6010 -0.4671

(0.0568) (0.0418) (0.4976) (0.3543) (0.4288)

... Both Informational and Non-Informational Content * 0.1871 0.0173 0.5569 0.4322 -0.1594

(0.2093) (0.1232) (0.6839) (0.5873) (0.7023)

Same Insurer as in Previous Year 0.1951

(Yes = 1) (0.0243)

Insurer Provided the Best Price 0.7166

(Yes = 1) (0.0244)

Brand-Demographics-Year FEs yes yes yes yes yes

Online-Brand-Demographics-Year FEs yes yes yes yes yes

Brand-Border-DMA FEs yes yes yes yes yes

Brand-Border-Month FEs yes yes yes yes yes

Brand-State-Month FEs yes yes yes yes yes

Survey FEs yes yes yes yes yes

Eﬀective Number of Observations

†

1,754,151 1,154,013 31,383 100,914 13,558

DMA-Level Advertising Only

Spending per Household in $ on...

... Informational Content * 0.0009 0.5854

-1.3626 -0.0946 2.6768

(0.2108) (0.2629) (0.9468) (0.6643) (1.1241)

... Non-Informational Content * 0.1682

0.0106 -0.5160 -0.5756 -0.2618

(0.0568) (0.0402) (0.4239) (0.3053) (0.3633)

... Both Informational and Non-Informational Content * 0.1467 -0.0290 0.5034 0.3807 -0.0815

(0.2036) (0.0939) (0.5437) (0.4733) (0.5533)

Same Insurer as in Previous Year 0.1950

(Yes = 1) (0.0244)

Insurer Provided the Best Price 0.7166

(Yes = 1) (0.0244)

Brand-Demographics-Year FEs yes yes yes yes yes

Online-Brand-Demographics-Year FEs yes yes yes yes yes

Brand-Border-DMA FEs yes yes yes yes yes

Brand-Border-Month FEs yes yes yes yes yes

Brand-State-Month FEs yes yes yes yes yes

Survey FEs yes yes yes yes yes

Eﬀective Number of Observations

†

1,754,151 1,154,013 31,383 100,914 13,558

Table F-2: Robustness Checks – Results for Advertising Content

a: <.001, b: <.01, c: < .05.

Standard errors in parentheses (clustered at the DMA level).

* Measured on a logarithmic scale.

†

After dropping observations due to collinearity with ﬁxed eﬀects.

(iii) (iv) (v)

Consideration Choice

Conditional on Unaided Awareness Aided Awareness Consideration

Recommendations

Advertising Spending per Household $ * -0.3028 -0.1773 0.2478

(0.5077) (0.4105) (0.4518)

Same Insurer as in Previous Year 0.2063

(Yes = 1) (0.0255)

Insurer Provided the Best Price 0.7128

(Yes = 1) (0.0258)

Brand-Demographics-Year FEs yes yes yes

Online-Brand-Demographics-Year FEs yes yes yes

Brand-Border-DMA FEs yes yes yes

Brand-Border-Month FEs yes yes yes

Brand-State-Month FEs yes yes yes

Survey FEs yes yes yes

Eﬀective Number of Observations

†

27,875 90,455 11,863

Recommendations

Spending per Household in $ on...

... Informational Content Only * -1.5258 -0.5535 1.6897

(1.1214) (0.8389) (1.1838)

... Non-Information Content Only * -0.8302 -0.5262 -0.1702

(0.5494) (0.4578) (0.4073)

... Both Informational and Non-Informational Content * 0.7913 0.4349 0.0543

(0.6547) (0.5303) (0.6336)

Same Insurer as in Previous Year 0.2122

(Yes = 1) (0.0257)

Insurer Provided the Best Price 0.7077

(Yes = 1) (0.0241)

Brand-Demographics-Year FEs yes yes yes

Online-Brand-Demographics-Year FEs yes yes yes

Brand-Border-DMA FEs yes yes yes

Brand-Border-Month FEs yes yes yes

Brand-State-Month FEs yes yes yes

Survey FEs yes yes yes

Eﬀective Number of Observations

†

27,875 90,455 11,863

Table F-3: Selection Robustness Checks 1 – Recommendations

a: <.001, b: <.01, c: < .05.

Standard errors in parentheses (clustered at the DMA level).

* Measured on a logarithmic scale.

†

After dropping observations due to collinearity with ﬁxed eﬀects.

(iii) (iv) (v)

Consideration Choice

Conditional on Unaided Awareness Aided Awareness Consideration

Would Never Consider

Advertising Spending per Household $ * -0.1166 0.0051 0.2199

(0.4305) (0.3844) (0.4891)

Would Never Consider -0.6099

-0.3069

(Yes = 1) (0.0216) (0.0085)

Same Insurer as in Previous Year 0.1947

(Yes = 1) (0.0244)

Insurer Provided the Best Price 0.7170

(Yes = 1) (0.0243)

Brand-Demographics-Year FEs yes yes yes

Online-Brand-Demographics-Year FEs yes yes yes

Brand-Border-DMA FEs yes yes yes

Brand-Border-Month FEs yes yes yes

Brand-State-Month FEs yes yes yes

Survey FEs yes yes yes

Eﬀective Number of Observations

†

30,736 97,628 13,558

Would Never Consider

Spending per Household in $ on...

... Informational Content * -1.1911 -0.0312 2.6961

(0.8466) (0.7167) (1.1324)

... Non-Informational Content * -0.2788 -0.5174 -0.2891

(0.4769) (0.3412) (0.3863)

... Both Informational and Non-Informational Content * 0.3572 0.3271 -0.1305

(0.6865) (0.5873) (0.6849)

Would Never Consider -0.6095

-0.3068

(Yes = 1) (0.0216) (0.0085)

Same Insurer as in Previous Year 0.1950

(Yes = 1) (0.0244)

Insurer Provided the Best Price 0.7166

(Yes = 1) (0.0244)

Brand-Demographics-Year FEs yes yes yes

Online-Brand-Demographics-Year FEs yes yes yes

Brand-Border-DMA FEs yes yes yes

Brand-Border-Month FEs yes yes yes

Brand-State-Month FEs yes yes yes

Survey FEs yes yes yes

Eﬀective Number of Observations

†

30,736 97,628 13,558

Table F-4: Selection Robustness Checks 2 – Would Never Consider

a: <.001, b: <.01, c: < .05.

Standard errors in parentheses (clustered at the DMA level).

* Measured on a logarithmic scale.

†

After dropping observations due to collinearity with ﬁxed eﬀects.

(i) (ii) (iii) (iv) (v)

Unaided Awareness Aided Awareness Consideration Choice

Conditional on Unaided Awareness Aided Awareness Consideration

Advertising Quantity

Advertising Spending per Household in $ * 0.2546

0.2158

-0.1158 0.0195 0.2203

(0.0713) (0.1003) (0.4501) (0.3922) (0.4900)

Sum of Competitors’ Advertising Spending 0.0373 0.0450 0.2814 0.0165 -0.0213

per Household in $ * (0.0193) (0.0506) (0.3767) (0.1843) (0.2608)

Same Insurer as in Previous Year 0.1947

(Yes = 1) (0.0244)

Insurer Provided the Best Price 0.7171

(Yes = 1) (0.0243)

Brand-Demographics-Year FEs yes yes yes yes yes

Online-Brand-Demographics-Year FEs yes yes yes yes yes

Brand-Border-DMA FEs yes yes yes yes yes

Brand-Border-Month FEs yes yes yes yes yes

Brand-State-Month FEs yes yes yes yes yes

Survey FEs yes yes yes yes yes

Eﬀective Number of Observations

†

1,754,151 1,154,013 31,383 100,914 13,558

Advertising Content

Spending per Household in $ on...

... Informational Content * -0.0107 0.5697

-1.5198 -0.1198 2.8692

(0.2129) (0.2749) (0.9815) (0.6812) (1.0546)

... Non-Informational Content * 0.1728

0.0089 -0.5378 -0.5987 -0.2356

(0.0585) (0.0420) (0.4388) (0.3331) (0.4190)

... Both Informational and Non-Informational Content * 0.2012 0.0203 0.6484 0.3989 -0.1952

(0.2057) (0.1194) (0.7034) (0.5789) (0.6703)

Sum of Competitors’ Advertising Spending per Household in $ on...

... Informational Content * 0.0711 0.6653

-0.2559 -0.0944 2.6006

(0.0595) (0.2030) (1.2882) (0.5331) (0.7979)

... Non-Informational Content * 0.0259 0.0268 0.1364 0.0297 0.4313

(0.0140) (0.0214) (0.3235) (0.1181) (0.2346)

... Both Informational and Non-Informational Content * 0.0006 -0.0230 0.4773 0.1176 -0.6790c

(0.0281) (0.1137) (0.4370) (0.1914) (0.3199)

Same Insurer as in Previous Year 0.1945

(Yes = 1) (0.0244)

Insurer Provided the Best Price 0.7171

(Yes = 1) (0.0244)

Brand-Demographics-Year FEs yes yes yes yes yes

Online-Brand-Demographics-Year FEs yes yes yes yes yes

Brand-Border-DMA FEs yes yes yes yes yes

Brand-Border-Month FEs yes yes yes yes yes

Brand-State-Month FEs yes yes yes yes yes

Survey FEs yes yes yes yes yes

Eﬀective Number of Observations

†

1,754,151 1,154,013 31,383 100,914 13,558

Table F-5: Competitive Advertising Robustness Check 1

a: <.001, b: <.01, c: < .05.

Standard errors in parentheses (clustered at the DMA level).

* Measured on a logarithmic scale.

†

After dropping observations due to collinearity with ﬁxed eﬀects.

(i) (ii) (iii) (iv) (v)

Unaided Awareness Aided Awareness Consideration Choice

Conditional on Unaided Awareness Aided Awareness Consideration

Advertising Quantity

Advertising Spending per Household in $ * 0.2539

0.2135

-0.1175 0.0191 0.2223

(0.0713) (0.0998) (0.4493) (0.3915) (0.4915)

Sum of Top 4 Brands’ Advertising Spending per Household in $ * 0.0226 -0.0001 0.1353 -0.0050 -0.0021

(0.0129) (0.0293) (0.3844) (0.1935) (0.2701)

Sum of Small Competitors’ Advertising Spending per Household in $ * 0.0472 0.2410

0.2559 0.0262 -0.1151

(0.0278) (0.0934) (0.5120) (0.1575) (0.2534)

Same Insurer as in Previous Year 0.1948

(Yes = 1) (0.0244)

Insurer Provided the Best Price 0.7171

(Yes = 1) (0.0244)

Brand-Demographics-Year FEs yes yes yes yes yes

Online-Brand-Demographics-Year FEs yes yes yes yes yes

Brand-Border-DMA FEs yes yes yes yes yes

Brand-Border-Month FEs yes yes yes yes yes

Brand-State-Month FEs yes yes yes yes yes

Survey FEs yes yes yes yes yes

Eﬀective Number of Observations

†

1,754,151 1,154,013 31,383 100,914 13,558

Advertising Content

Spending per Household in $ on...

... Informational Content * 0.0000 0.5640

-1.5541 -0.1311 2.8418

(0.2116) (0.2697) (0.9118) (0.6813) (1.0502)

... Non-Informational Content * 0.1729

0.0090 -0.6066 -0.5956 -0.2394

(0.0587) (0.0419) (0.4521) (0.3347) (0.4219)

... Both Informational and Non-Informational Content * 0.1997 0.0185 0.6398 0.4003 -0.2236

(0.2052) (0.1228) (0.7296) (0.5784) (0.6761)

Sum of Top 4 Brands’ Advertising Spending per Household in $ on...

... Informational Content * 0.0319 0.6607

0.1434 -0.1452 2.7291

(0.1028) (0.2040) (1.0019) (0.4977) (0.8509)

... Non-Informational Content * 0.0132 0.0084 0.5178

-0.0067 0.1617

(0.0081) (0.0151) (0.2447) (0.1510) (0.1713)

... Both Informational and Non-Informational Content * 0.0340 -0.0817 -0.4119 0.1220 -0.7016

(0.0286) (0.1040) (0.4836) (0.1960) (0.4115)

Sum of Small Competitors’ Advertising Spending per Household in $ on...

... Informational Content * 0.1466 0.5746 -1.1543 -0.0168 2.2014

(0.1022) (0.3721) (1.9118) (0.6327) (1.1751)

... Non-Informational Content * 0.1247

0.1870

-0.6155 0.0777 0.8652

(0.0377) (0.0864) (0.4997) (0.1838) (0.2783)

... Both Informational and Non-Informational Content * -0.0357 0.0529 0.8065

0.1203 -0.4924

(0.0421) (0.1349) (0.3960) (0.1675) (0.2923)

Same Insurer as in Previous Year 0.1943

(Yes = 1) (0.0244)

Insurer Provided the Best Price 0.7170

(Yes = 1) (0.0244)

Brand-Demographics-Year FEs yes yes yes yes yes

Online-Brand-Demographics-Year FEs yes yes yes yes yes

Brand-Border-DMA FEs yes yes yes yes yes

Brand-Border-Month FEs yes yes yes yes yes

Brand-State-Month FEs yes yes yes yes yes

Survey FEs yes yes yes yes yes

Eﬀective Number of Observations

†

1,754,151 1,154,013 31,383 100,914 13,558

Table F-6: Competitive Advertising Robustness Check 2

a: <.001, b: <.01, c: < .05.

Standard errors in parentheses (clustered at the DMA level).

* Measured on a logarithmic scale.

†

After dropping observations due to collinearity with ﬁxed eﬀects.