Super Returns to Super Bowl Ads?

∗

Seth Stephens-Davidowitz

[email protected]

Hal Varian

hal@ischool.berkeley.edu

Michael D. Smith

mds@cmu.edu

December 10, 2016

Abstract1

This paper uses a natural experiment—the Super Bowl—to study the2

causal eﬀect of advertising on demand for movies. Identiﬁcation of3

the causal eﬀect rests on two points: 1) Super Bowl ads are purchased4

before advertisers know which teams will play; 2) home cities of the5

teams that are playing will have proportionally more viewers than6

viewers in other cities. We ﬁnd that the movies in our sample expe-7

rience on average incremental opening weekend ticket sales of about8

$8.4 million from a $3 million Super Bowl advertisement.9

∗

This paper beneﬁted greatly from discussions with Randall Lewis, David Reiley, Bo

Cowgill, Lawrence Katz, and Lawrence Summers. The referees and editors were particu-

larly helpful. We also thank participants at IO Fest at Berkeley and the NBER Summer

Institute for helpful comments.

1

1 Introduction10

The United States spends roughly 2 percent of its GDP on advertising (Galbi11

[2008]). Not surprisingly, whether, when, and why advertising increases prod-12

uct demand is of considerable interest to economists and marketers. However,13

empirically measuring the impact of advertising is notoriously diﬃcult. Prod-14

ucts that are heavily advertised tend to sell more, but this in itself does not15

prove causation (Sherman and Tollison [1971], Comanor and Wilson [1971]).16

A particular product often sees an increase in sales after increasing its ad ex-17

penditures, but here too the causation could run the other way (Heyse and18

Wei [1985], Ackerberg [2003]). For example, ﬂower companies increase ad ex-19

penditures in the weeks leading up to Valentine’s Day and see increased sales20

around Valentine’s Day. But it is not easy to determine the causal impact21

of that ad expenditure since many of the same factors that aﬀect consumer22

demand may also aﬀect advertising purchase decisions (Schmalensee [1978],23

Lee et al. [1996]).24

Testing for causal eﬀects requires an exogenous shock to ad exposures.25

The gold standard, as usual, is a randomized experiment. For this reason,26

ﬁeld experiments have become increasingly popular among economists and27

marketers studying advertising (Simester et al. [2009], Bertrand et al. [2010],28

Lewis and Rao [2012]). However, these experiments tend to be expensive29

and require access to proprietary data. Moreover, they tend to have low30

power, often do not produce statistically signiﬁcant eﬀects, and have not led31

2

to consensus on advertising eﬀectiveness (Hu et al. [2007], Lewis and Reiley32

[2008], Lewis and Rao [2012]).33

Further, ﬁeld experiments tend to involve a particular subset of ads: those34

that a ﬁrm is uncertain enough about to agree to conduct an experiment.35

These ads may be quite diﬀerent from ads that are routinely purchased by36

ﬁrms. By contrast the diﬀerential viewership associated with the the Super37

Bowl and other sports events yields natural experiments that can be used to38

estimate advertising eﬀectiveness.39

Two weeks prior to the Super Bowl, the NFC and AFC Championship40

games are played. Controlling for the point spread, the winners of these41

games are essentially random. On average, the Super Bowl will be watched42

by an additional eight percentage points, or roughly 20 percent, more house-43

holds, in the home cities of the teams that play in the game compared to44

other cities. There is a similar increase in viewership for the host city of the45

Super Bowl. We refer to these boosts in viewership as the “home-city” and46

“host-city” eﬀects respectively.47

Super Bowl ads are typically sold out several weeks or months before48

these Championship games, so ﬁrms have to decide whether to purchase ads49

before knowing who will be featured in the Super Bowl. Hence the outcomes50

of the Championship Games are essentially random shocks to the number51

of viewers of Super Bowl ads in the home cities of the winning teams. The52

increased sales of advertised products in cities of qualifying teams, compared53

to sales in home cities of near-qualifying teams, can thus be attributed to54

3

advertisements.55

There are three attractive features to studying movies advertised in the56

Super Bowl. First, movie advertisements are common for Super Bowls, with57

an average of about 7 per game in our sample. Second, diﬀerent movies58

advertise each year. Third, Super Bowl ad expenditure represents a large59

fraction of a movie’s expected revenue. For a Pepsi ad to be proﬁtable, it60

only needs to move sales by a very small amount. As Lewis and Rao [2012]61

show, in their Super Bowl Impossibility Theorem, for products like Pepsi,62

it can be virtually impossible to detect even proﬁtable eﬀects. The cost of63

Super Bowl ads, on the other hand, can represent a meaningful fraction of a64

movie’s revenue.65

There are however, two notable disadvantages to studying movies. First,66

city-speciﬁc, movie sales data are costly to obtain. Nonetheless, we were67

able to acquire this data for a limited sample of movies and cities. However,68

we also have an additional proxy for movie demand—Google searches after69

the Super Bowl. Miao and Ma [2015] and Panaligan and Chen [2013] have70

illustrated that Google searches are predictive of opening week revenue, and71

Google searches have the advantage of being available for the full sample of72

cities.73

The second disadvantage of studying movies is that movies do not have74

a standard measure of expected demand prior to the broadcast of the Super75

Bowl ads. Here too Google searches can be helpful in that they can serve as76

a proxy for pre-existing interest in the movie and help improve the prediction77

4

of the outcome (box oﬃce or searches) when the movie opens.78

Wesley Hartmann and Daniel Kapper proposed the idea of using the79

Super Bowl as a natural experiment at a presentation at the June 7-9, 201280

Marketing Science conference. They subsequently circulated a June 201281

working paper examining the impact of the Super Bowl ads on beer and soft82

drink sales. The most recent version of their working paper is Hartmann and83

Klapper [2015].84

We independently came up with a similar idea in February of 2013. We85

focused on Super Bowl movie ads and thought of “fans” as an instrumental86

variable for ad exposures. Our initial analysis used Google queries for movie87

titles as the response variable, but eventually we were able to acquire movie88

revenue data by DMA. Earlier versions of Hartmann and Klapper [2015] and89

this paper were presented at the same session at the 2014 summer NBER90

meeting in Cambridge.91

Both papers ﬁnd a substantial eﬀect of advertising on purchases in quite92

diﬀerent markets. Beer and soft drinks involve substantial repeat purchases93

and have familiar brands. Movies are typically purchased only once and94

each is unique. Given these quite diﬀerent characteristics, it is comforting95

that both papers ﬁnd an economically and statistically signiﬁcant impact of96

advertising on sales.97

In a related paper, Ho et al. [2009] build an econometric model of ex-98

hibitors’ decisions to show a movie, and consumers’ decisions to view a movie99

during its opening weekend. The ﬁrst stage equation models the probability100

5

of placing a Super Bowl ad for a movie as a function of the movie’s budget,101

genre, rating, and distributor, whether the movie is released on a holiday102

week, and the timing of the ad relative to the movies release. Using this103

estimate, the authors construct expected expenditure on the Super Bowl104

ad. In the second stage regressions, they use the predicted expenditure as105

an explanatory variable for exhibitor decisions to show the movie, and for106

consumers’ decisions to view the movies during the opening weekend.107

Our model diﬀers from the approach in Ho et al. [2009] in that we do not108

model the studios’ decisions to purchase ads. It is possible (though in our109

opinion not likely) that astute theater chains recognize that the home cities110

of the Super Bowl teams will be exposed to more ads and thus be more likely111

to want to see the advertised movies. If this is so, then our model is about112

the joint impact of advertising on both consumer and exhibitor decisions.113

However, in our view the primary response is likely consumer decisions since114

exhibitors typically have to construct their distribution schedules months in115

advance. We expand on this point in Section 5.116

Our work is also related to Yelkur et al. [2004] who analyze the eﬀective-117

ness of Super Bowl advertising by comparing box oﬃce revenue for movies118

that were advertised on the Super Bowl to a set of popular movies that did119

not have Super Bowl advertisements. The authors ﬁnd that, controlling for120

budget size and release date, movies with Super Bowl advertisements had121

nearly 40 percent higher gross theatrical revenue than other non-promoted122

movies. Of course, the movies that were selected to be advertised were likely123

6

chosen for some reason, so there could potentially be bias in this estimate124

due to confounding variables.125

Overall, with our method we ﬁnd strong evidence of large eﬀects of ad-126

vertising on movie demand. Our results suggest that a 100 ratings point127

increase due to additional Super Bowl ad impressions increases opening week-128

end movie revenue by 50–70 percent. For the average movie in our sample,129

this translates into an incremental return of at least $8.4 million in opening130

weekend ticket sales associated with a $3 million Super Bowl advertisement.131

We believe that researchers can use this methodology for other types of132

advertising. Sports events such as the World Series, basketball playoﬀs, col-133

lege bowls, the Olympics, and the World Cup create many large, essentially134

random shocks to viewership of ads shown during these events that can serve135

as natural experiments to measure ad impact.136

2 Empirical speciﬁcation137

We use the following notation.

t = date where outcome is measured (opening week) (1)

s = date when ads are seen (Super Bowl) (2)

y

mct

= outcome for movie m in city c at time t (3)

x

cms

= adviews for movie m in city c at time s (4)

z

cms

= fans of team from city c exposed to ad for movie m at time s (5)

7

The variable outcome is the measure of ad performance, which in the ini-138

tial speciﬁcation is Google searches immediately prior to the opening week-139

end. Later we use opening weekend revenue for a subset of the movies ad-140

vertised as our ad performance measure.141

The adviews are the Nielsen ratings for the relevant Super Bowl. Nielsen142

ratings correspond to the percent of households watching the Super Bowl in143

an average half hour.144

The fans variable in the initial speciﬁcation consists of 3 dummy variables145

indicating whether the home team of the city in question is the AFC partici-146

pant in the Super Bowl, whether the home team of the city in question is the147

NFC participant in the Super Bowl, and whether the city in question hosts148

the Super Bowl. Later on we investigate some reﬁnements to this measure.149

Our model speciﬁcation is then a classic instrumental variable model.

1

150

y

cmt

= α

0

+ α

1

x

cms

+ 

cmt

(6)

x

cms

= β

0

+ β

1

z

cms

+ δ

cms

(7)

Equation (6) says that the outcome, y

cmt

, depends on prior ad exposure,151

x

cms

. We would not expect that estimating this single equation by ordinary152

least squares would produce a good estimate of the causal eﬀect of advertis-153

ing, since x

cms

could be correlated with 

cmt

.154

1

We also include city and movie ﬁxed eﬀects along with an index of Google searches

prior to the Super Bowl as control variables in our regressions.

8

There are a variety of ways that x

cms

could be correlated with 

cmt

. For155

example, suppose that in some years, some cities are particularly interested156

in entertainment. These cities might watch the Super Bowl more than usual157

and attend movies more than usual. Or suppose diﬀerent types of movies158

appealed to diﬀerent geographic audiences. In this case, the teams that159

compete in the Super Bowl could aﬀect the choice of movie advertised.160

Another potential issue is measurement error. The city-level Nielsen rat-161

ings are based on a relatively small number of households. We would expect162

measurement error associated with the ratings numbers would attenuate the163

estimated eﬀect of ad viewership on outcomes toward zero.164

In order to estimate the causal impact of ad views on outcomes, we need165

an instrument—a variable that perturbs ad views exogenously.166

Equation (7) contains such instruments, namely the home-city and the167

host-city eﬀects we described earlier. We know from prior experience, and168

will verify in Section 4.1, that this instrument is a strong predictor of ad169

views. Furthermore, this instrument should be independent of 

cmt

since170

advertising expenditures typically are chosen well before it is known which171

teams will play in the Super Bowl. We present additional arguments for172

identiﬁcation in Section 5.173

9

3 Data174

3.1 Ad views175

We measured ad views using Nielsen ratings for the 2004-2014 Super Bowls,176

for 56 designated media markets (cities) from Street & Smith’s Sports Busi-177

ness Daily Global Journal. Total local ad spend, which we use in Section 5.4,178

is taken from Kantar Media. This data is only available starting in 2009.179

3.2 Movies180

We looked at a sample of 70 movies that were advertised in the Super Bowl181

and were released within 6 months after the game date. The average gap182

between the Super Bowl and the movie release was about 66 days and the183

median was 54 days. The gap varied quite a bit, with a standard deviation of184

about 50 days. Roughly speaking, the median date of release was mid-March,185

but there is substantial variation in the release date.186

We obtained the list of movies that advertised for the Super Bowl from187

the USA Today’s AdMeter, which lists commercials and viewer ratings for188

all commercials after every Super Bowl. Release dates, distributor, budget,189

and national sales by week for every movie were found at the-numbers.com.190

Data on movie opening weekend sales is from Rentrak.191

10

3.3 Fans and Host City192

As indicated above, the simplest proxy for fans of a team in a city is just193

a dummy variable that equals 1 if the team plays in the home city and 0194

otherwise. We split the fans into AFC fans and NFC fans. We also add the195

host city in some speciﬁcations. Though the host city is known in advance,196

we argue in Section 5 that it represents such a small part of the total boost197

in viewership that it is unlikely to have a meaningful impact on advertiser198

choices. The advantage of including the host city is we get more power.199

However, the quantitative results are similar with and without host city,200

suggesting advertisers do not select ads considering which city is hosting the201

game.202

To test the sensitivity of our results to alternate speciﬁcations, in Sec-203

tion 6.1 we reﬁne the deﬁnition of fans using Google searches, and in Sec-204

tion 5.3 we adjusted the fans measure using Vegas odds in the playoﬀs so as205

to reﬂect the estimated fans at the time of the playoﬀs.206

3.4 Searches207

Movie titles frequently contain common words, making it diﬃcult to use208

simple text matching to identify queries related to movies. For example, the209

word [wolverine] could refer to an animal, a university mascot, a brand of210

boots, or a Marvel comics character.211

We address this problem by using the Google entity identiﬁer associated212

11

with the movies in our sample. Google’s entity identiﬁer attempts to disam-213

biguate diﬀerent uses of a word by using contextual information associated214

with the search. So if a user searched for other animals in the session where215

a search for [wolverine] occurred, that user is likely looking for information216

about the animal. On the other hand, if a user included movie related terms217

along with a search for [wolverine] it is likely that they were using the word218

as short-hand for the movie X-Men Origins: Wolverine.219

With the Google entity identiﬁer, we generate a control variable in our220

regressions based on the Google Trends index prior to the Super Bowl for221

each city and movie in our data. The Google Trends index for the week222

preceding the opening weekend was used as an outcome variable in the initial223

speciﬁcation. We interpret this index as a measure of “interest” in a movie.224

The Google Trends data has the advantage of being complete—available for225

all movies in the sample—and non-proprietary.

2

By contrast, the Rentrak226

data on opening weekend revenue is available only for a subset of movies and227

is proprietary and cannot be freely redistributed.228

In addition to the Google Trends index of searches on the movie prior to229

the Super Bowl, we also use city and movie ﬁxed eﬀects.230

We also conﬁrm that a movie’s opening weekend box oﬃce sales can be231

well-predicted by a few key features. In particular, we regress box oﬃce sales232

per capita on searches prior to the Super Bowl, the type of movie (comedy,233

2

The number of queries in a given city must be larger than an unspeciﬁed privacy

threshold to show up in the index, so there are a few smaller cities that report zero

searches on movie entities prior to the Super Bowl. We drop these cities from the analysis.

12

●

−4 −3 −2 −1 0

−5 −4 −3 −2 −1 0 1

Linear regression

box.office.reg

box.office.actual

●

−4 −3 −2 −1 0

−5 −4 −3 −2 −1 0 1

Random forest

box.office.rf

box.office.actual

Figure 1: This shows how opening box oﬃce sales per capita compare to a

prediction using a few features: Google searches prior to the Super Bowl, the

type of movie (comedy, adventure, etc.), the distributor, the rating, and the

DMA. We use both an ordinary linear regression and a random forest model.

adventure, etc.), the distributor, the rating, and the DMA. We use both an234

ordinary linear regression and a random forest model. The OLS prediction235

has an R

2

of 0.51 and the random forest has an R

2

of 0.87. Figure 3.4 shows236

how these two predictions compare to the actual box oﬃce sales.237

13

4 Results for Google Searches238

4.1 First stage239

This section examines the eﬀects of advertising on Google searches in the240

week prior to opening weekend. Table 1 shows that the Nielsen ratings in a241

given city are strongly related to whether that city is a home city for one of242

the teams playing or the host city for the game.243

Table 1: First Stage: Super Bowl Ratings and Fans of Teams

Nielsen Ratings

(1) (2)

City of AFC Championship Game Winner 0.077***

(0.009)

City of NFC Championship Game Winner 0.076***

(0.008)

Super Bowl Host City 0.063***

(0.008)

Constant 0.455*** 0.451***

(0.004) (0.003)

Adjusted R-squared 0.66 0.75

Observations 616 616

* p < 0.1; ** p < 0.05; *** p < 0.01

Notes: Robust standard errors clustered at the city-year level are shown in parentheses.

City and year ﬁxed eﬀects are included in all speciﬁcations. Nielsen ratings correspond to

the percent of households watching the Super Bowl in an average half hour. Home city is

a dummy variable that takes the value 1 if a team plays in a city; 0 otherwise. The Green

Bay Packers’ Home city is Milwaukee, since we do not have ratings data on Green Bay.

Data sources are discussed in more detail in Section 3.

Column (1) of Table 1 shows the R

2

for the regression that only uses244

city and movie ﬁxed eﬀects. Column (2) shows what happens to R

2

when245

14

Search index

Time

0 5 10 15 20 25 30

200 400 600 800 1000

Advertised

Not advertised

Figure 2: Nationwide searches for movies advertised during the Super Bowl

and similar movies that were not advertised during the Super Bowl.

we include dummy variables for the teams that are playing and the host246

city. The R

2

moves from 66 percent to 75 percent, indicating that these247

instruments signiﬁcantly improve the prediction of Nielsen ratings.248

As the regression shows, about 8 percentage points more households will249

watch the Super Bowl in the home city of qualifying teams. This is about a250

20 percent increase in ratings compared to the sample average.251

4.2 Second stage252

Figure 2 shows the nationwide queries on movie titles advertised in the Super253

Bowl.254

It is clear that movies advertised in the Super Bowl see a signiﬁcant bump255

15

in searches. We also contrast these searches with national search volume for256

a set of placebo movies that had similar qualities to the advertising movies257

but did not advertise in the Super Bowl. We discuss how we select these258

movies in Section 6.3.259

While it is clear there is an increase in interest in advertising movies260

immediately after the ads are shown, it is not apparent how much of that261

initial interest translates into box oﬃce revenue. That question is what our262

model is designed to answer.263

The regression results in Table 2 use an ordinary least squares regression264

in Column (1) to show that, for movies that advertised in the Super Bowl,265

Google searches on release week are notably higher in cities with higher Super266

Bowl ratings than in other cities. Note that Google Trends numbers for the267

search volume in a particular geo are measured relative to the total number268

of searches in that geo. Hence the Trends numbers are already normalized269

for population size.270

Column (2) uses both home and host cities as instruments and ﬁnds about271

twice as large an eﬀect as the OLS estimate. Our baseline model uses host272

cities as an instrument but Table 5 shows the estimated eﬀect is similar if273

we use only home cities.274

16

Table 2: Eﬀects of Advertising

log(Google Searches on Release Week) log(Box Oﬃce PC)

(1) (2) (3) (4)

Nielsen Ratings 0.314 0.762** 0.484** 0.771**

(0.243) (0.318) (0.225) (0.362)

log(pre Search) 0.068*** 0.069*** 0.035*** 0.035***

(0.018) (0.017) (0.013) (0.012)

Adj. R-squared 0.89 0.89 0.96 0.96

Observations 3,080 3,080 1,088 1,088

Speciﬁcation OLS 2SLS OLS 2SLS

* p < 0.1; ** p < 0.05; *** p < 0.01

Notes: Robust standard errors clustered at the city-year level are shown in parentheses.

City and year ﬁxed eﬀects are included in all speciﬁcations. Super Bowl ratings are Nielsen

ratings, corresponding to percent of households watching the Super Bowl in an average half

hour. Instruments include dummy variables for the home and host cities. Data sources

are discussed in more detail in Section 3.

5 Identiﬁcation275

In this section we consider arguments questioning the validity of the fans276

instrument and present rebuttals to these arguments.277

We note that there could be a potential problem in our estimates above if278

East Coast and Midwest football fans liked diﬀerent kinds of movies. In such279

a case, the movie that a studio chooses to advertise could, in principle, depend280

on which teams play in the Super Bowl. In our view, this is conceivable, but281

not likely.282

The reason this shouldn’t impact our estimates is that the decisions about283

which movies to promote and how much to spend on promotion are made at284

a national level. This means that variations in attendance will be determined285

17

primarily by local tastes. The only role that advertiser decisions might make286

is in determining which movies to advertise nationwide. This will typically287

not depend on which teams end up playing since the choice of which movies288

to advertise 1) is made well in advance and 2) has a tiny impact on the total289

size of the audience, as we establish below.290

5.1 Ad decisions are made in advance291

The decision to show an ad in the Super Bowl is typically made far in advance292

of the actual game, when advertisers would have little idea which teams would293

play. (They would know the host city, which we deal with shortly.) Table 3294

presents a list of press reports about the status of Super Bowl ad sales.295

(We report short URLs for reasons of space; complete URLs are provided296

in a spreadsheet in the Appendix.) Of course most advertisers do not wait297

until the last minute to purchase ads. According to our discussions with298

ﬁlm industry executives, the decision about which movies to advertise in the299

Super Bowl are decided well in advance of the game. Generally studios only300

have a few choices of movies that will be released in an appropriate time301

frame, and a great deal of care goes into planning and executing marketing302

for the hoped-for blockbusters.303

18

Table 3: Ad sales for Super Bowl.

Year Snippet Date Source

2003 fewer than 10 spots available Jan 06 2003 superbowl.ads.com

2004 –NA– –NA– –NA–

2005 said Thursday all 59 slots had been sold Feb 02 2005 money.cnn.com

2006 80% sold Dec 18 2005 www.mediapost.com

2007 ﬁrst half sold out Jan 03 2007 money.cnn.com

2008 90% sold out by ﬁrst week in Nov Nov 07 2007 money.cnn.com

2009 much was sold out by September Jan 09 2008 money.cnn.com

2010 had ﬁnished selling commercial time Feb 01 2010 articles.latimes.com

2011 3 months before Oct 29 2010 adage.com

2012 has sold out Jan 02 2012 www.bloomberg.com

2013 advertisers need to announce 5 months out Sep 03 2013 www.usatoday.com

Notes: The columns show the Super Bowl year and extracts from news articles that

appeared on the indicated date from the indicated source. The full URL for these snippets

is available in the online Appendix.

5.2 Home-city and host-city eﬀects are small304

It may well be that studios would advertise diﬀerent movies in diﬀerent ge-305

ographies if they were able to do so, but in this case there is a single nation-306

wide audience and advertisers must choose one movie for the entire audience.307

This restriction makes it implausible that the host cities and home cities of308

the teams playing in the Super Bowl would have any impact on advertising309

decisions since the aggregate audience for the ad is not very sensitive to which310

teams actually play and where they play.311

To see this, we constructed an estimate of what would have happened to312

viewership if the teams that lost the championship games instead won those313

games and competed in the Super Bowl.314

Consider for example, Pittsburgh’s 2005 loss. This meant that 161,000315

19

fewer households watched the ad in Pittsburgh than would have watched had316

Pittsburgh won. However, compared to the total viewership for the Super317

Bowl that year of 86 million this is only 0.2 percent, a tiny factor in an318

advertiser’s decision.319

The largest city in our sample is New York, but even in this case, the320

impact of the counterfactual is only 1.2 percent. Nationwide the average321

absolute diﬀerence in viewers across all DMAs and years was 0.4 percent of322

national viewership. Would the choice of ad to be shown in the Super Bowl323

depend on a 20 percent boost in viewership for 0.4 percent of the population?324

We believe that this eﬀect is insigniﬁcant from an economic viewpoint and325

unlikely to aﬀect studio decisions.326

A similar argument applies to the host cities which are known in advance.327

However, the population of the host cities comprise only 1.6 percent on av-328

erage of the DMAs in our sample. It seems implausible that choosing which329

movie ad to show nationwide in the Super Bowl would be inﬂuenced by a 0.2330

percent boost in viewership (1.6 percent of the population times a 15 percent331

boost).332

5.3 Expected fans333

Even though advertisers do not know with certainty who will play in the Su-334

per Bowl game, they can form judgments about who will play. Our contacts335

in the movie business tell us that decisions on which movies to advertise are336

made far in advance of the playoﬀs, and they would be highly unlikely to337

20

substitute at the last minute based on which teams were playing due to the338

major investments they have made in planning, publicity, and production of339

the movie ad. Furthermore, as we have seen, the eﬀect on viewership of the340

movie ad is tiny.341

Nevertheless, let us take this critique seriously and see how plausible it is.342

Consider the Vegas odds for the AFC and NFC Championship games.

3

We343

converted these odds to probabilities using the method described in Stern344

[1986] and calculated the expected fans for each city, where the expectation345

is made using the Vegas odds just prior to the championship game. We346

then used the expected fans as control variables in the regressions described347

earlier.348

We did not use the host city as an instrument since we thought that if349

advertisers were so sophisticated that they considered expected fans in their350

decisions, they would certainly take into account the host city in those deci-351

sions, which would make the host city an invalid instrument. The expected352

fans speciﬁcation made no essential diﬀerence in the results.353

Let us summarize the argument. In our baseline speciﬁcation, the instru-354

ment is whether a city’s team qualiﬁed for the Super Bowl. If advertisers355

were highly sophisticated and picked advertisements based on which teams356

were performing well up to the point they chose to advertise this could be357

a biased instrument. By controlling for the probability a team makes it to358

3

These are available at http://www.vegasinsider.com/nfl/afc-championship/

history/.

21

the Super Bowl at the time of the Championship games, we ensure that our359

instrument is “as good as random.”360

5.4 Impact of outcome on subsequent ad spend361

If advertisers choose their subsequent ad spend on a movie based on the362

associated Super Bowl ratings, our instrument would not be valid. To check363

this possibility we ran a regression to see if local ad spend was associated364

with home and host cities. Our data on local ad spend is from Kantar Media,365

and these data were only available to us starting in 2009.366

The results of this regression are shown in Column 2 of Table 4. These367

estimates should be compared to those in Column 1 which is the ﬁrst-stage368

regression from Table 1 but restricted to data from 2009 onward. Both de-369

pendent variables, the Nielsen ratings and ad spend per capita, are expressed370

in logs. Hence, the regression coeﬃcients can be interpreted as percentage371

response. The impact of host and home cities on Nielsen ratings is large and372

statistically signiﬁcant while the corresponding coeﬃcients for local ad spend373

are small and statistically insigniﬁcant.374

6 Variations on the baseline model375

Here we consider a few variations on the baseline model.376

22

Table 4: Local ad spend compared to Nielsen ratings

log(Nielsen Ratings) log(Ad Spend PC+1)

(1) (2)

City of AFC Championship Game Winner 0.107*** -0.003

(0.021) (0.008)

City of NFC Championship Game Winner 0.133*** -0.011

(0.020) (0.008)

Super Bowl Host City 0.093*** 0.005

(0.020) (0.008)

Adjusted R-squared 0.80 0.58

Observations 336 336

Fixed Eﬀects City and Year City and Year

* p < 0.1; ** p < 0.05; *** p < 0.01

Notes: These regressions include only observations for 2009 onward due to data availability.

For each year and city, we add up local television spending across all Super Bowl movies.

This gives one observation for each year and city, making the data directly comparable to

Nielsen ratings data. There are a small number of zeros in local ad spend for a few small

cities and niche movies, which is why we took log of adspend + 1. Note that these cities

may well have seen some movie ads through national advertising campaigns.

23

4

8

12

16

value

New.York.Giants

10

20

30

value

San.Francisco.49ers

Figure 3: Heat map of estimated fan density for New York Giants and San

Francisco 49ers using method described in text.

6.1 Other deﬁnitions of fans377

In our baseline model we use dummy variables for the home cities of the two378

participating teams. However, some major cities do not have an NFL team,379

but football fans in those cities may identify with teams from other cities.380

We use Google entity search data from Google Trends in each NFL city for381

each NFL team to measure the local interest in that team. See Figure 3382

which shows the distribution of searches for the New York Giants and the383

San Francisco 49ers. The geographic pattern suggests that this is a plausible384

measure for the fan distribution. Our results using this deﬁnition of fans are385

shown in Column (2) of Table 5.386

24

Table 5: Variations on baseline model for opening week searches

log(Google Searches on Release Week)

(1) (2) (3) (4) (5)

Nielsen ratings 0.762** 0.684** 0.687* 0.705 0.721**

(0.318) (0.333) (0.355) (0.620) (0.360)

log(Pre search) 0.069*** 0.069*** 0.069*** 0.069*** 0.078***

(0.017) (0.017) (0.017) (0.017) (0.017)

Adj. R

2

0.89 0.89 0.89 0.89 0.92

Observations 3,080 3,080 3,080 3,080 3,080

Speciﬁcation +Host Trends −Host +Vegas Weighted

* p < 0.1; ** p < 0.05; *** p < 0.01

Notes: Robust standard errors clustered at the city-year level are shown in parentheses.

City and movie ﬁxed eﬀects are included in all speciﬁcations. Nielsen ratings correspond

to the percent of households watching the Super Bowl in an average half hour. Column

(1) uses home and host cities as instruments, Column (2) uses the Google Trends data

to measure fans, Column (3) omits the host variable, Column (4) uses the expected fans

measure based on Vegas odds, Column (5) uses the original speciﬁcation with population

weighting. Data sources are discussed in more detail in Section 3.

25

6.2 Opening weekend box oﬃce revenue387

As mentioned above, we have two measures of outcome: Google searches on388

the movie title and opening weekend revenue.389

The movie sales data we have is only available for a subset of cities. In390

particular, we only have data for movies that advertised in the Super Bowl391

and cities that were the home cities for teams that qualiﬁed for a Super Bowl392

or were the runners-up.393

Despite the smaller sample, there is evidence of a signiﬁcant positive eﬀect394

of Super Bowl ratings on movie sales as shown in Table 2, Columns (3) and395

(4). Note, though, that the eﬀect on ticket sales is smaller than the eﬀect on396

Google searches. This is true even if we use only the sub-sample of cities for397

which we have box oﬃce data. Table 6 reports regressions using the alternate398

deﬁnition of fans.399

6.3 Placebo analysis400

It is conceivable that Super Bowl ratings could inﬂuence subsequent movie401

attendance for all movies. We consider this possibility highly implausible,402

but decided to check it anyway.403

One could look at city-by-city movie attendance following the Super Bowl,404

but a better test is to look at movies that were similar to those advertised405

in the Super Bowl. Accordingly, we constructed a placebo set of movies. If406

watching the Super Bowl is correlated with subsequent overall movie atten-407

26

Table 6: Variations on baseline model for opening week box oﬃce

log(Box Oﬃce PC)

(1) (2) (3) (4) (5)

Nielsen Ratings 0.771** 0.705** 0.507 1.401*** 0.444

(0.362) (0.342) (0.352) (0.527) (0.283)

log(Pre Search) 0.035*** 0.035*** 0.035*** 0.038*** 0.055***

(0.012) (0.012) (0.012) (0.012) (0.016)

Adj R

2

0.96 0.96 0.96 0.96 0.97

Observations 1,088 1,088 1,088 1,088 1,088

Speciﬁcation +Host Trends −Host +Vegas Weighted

* p < 0.1; ** p < 0.05; *** p < 0.01

Notes: Robust standard errors clustered at the city-year level are shown in parentheses.

City and movie ﬁxed eﬀects are included in all speciﬁcations. Nielsen ratings correspond

to the percent of households watching the Super Bowl in an average half hour. Column

(1) uses home and host cities as instruments, Column (2) uses Google Trends data to

measure fans, Column (3) omits the host variable, Column (4) uses the expected fans

measure based on Vegas odds, Column (5) uses the original speciﬁcation with population

weighting. Data sources are discussed in more detail in Section 3.

27

dance, we would expect to see it aﬀect both those movies that were advertised408

and similar movies that weren’t advertised.409

Speciﬁcally, we used nearest-neighbor matching based on the movie bud-410

get, movie category (comedy, action, etc.), distributor, critic ratings, and411

year and month of release. We used the matchit R package which is speciﬁ-412

cally designed for this purpose and described in detail in Ho et al. [2007a,b].413

We provide lists of the advertised and matched movies in the online appendix.414

In our view, these two lists appear to be similar.415

The results are shown in Table 7 for our baseline speciﬁcation and a few of416

the variations considered above. What is noteworthy is that the coeﬃcient417

on Nielsen ratings is insigniﬁcant for all speciﬁcations. Of course, the418

movies advertised in the Super Bowl were chosen for that distinction and our419

matching is far from perfect, so this analysis cannot be considered deﬁnitive420

evidence. Nevertheless, it is suggestive.421

We can test to see whether the estimated coeﬃcient on ad views (Nielsen422

ratings) is diﬀerent for the advertised and placebo movies. To do this we423

combine the two datasets and add an interaction term for Nielsen ratings424

and the advertised movies. This is denoted by Nielsen × Super Ad in425

Table 8. The interaction eﬀect is signiﬁcant at the 10 percent level in our426

baseline speciﬁcation (Column 2) and at the 5 percent level when we use the427

Google Trends measure for fans (Column 3).

4

428

4

Another question is whether placebo movies do worse than they would have if the

Super Bowl ads had not run. That is, does advertising for Super Bowl movies cause

substitution away from placebo movies? The relevant coeﬃcient to test this is the ﬁrst

one in Table 8, Nielsen Super Bowl Ratings. Unfortunately, we get diﬀerent answers

28

Table 7: Eﬀects of Advertising: Placebo movies

log(Google Searches on Release Week)

(1) (2) (3) (4)

Nielsen Ratings -0.091 -0.373 0.059 0.198

(0.374) (0.387) (0.444) (0.876)

log(Pre-Super Search) 0.083*** 0.083*** 0.084*** 0.084***

(0.019) (0.018) (0.019) (0.019)

Adjusted R-squared 0.87 0.87 0.87 0.87

Observations 2,747 2,747 2,747 2,747

Speciﬁcation 2SLS 2SLS (Trends fans) 2SLS (-Host) 2SLS (+Vegas)

* p < 0.1; ** p < 0.05; *** p < 0.01

Notes: Column (1) shows the baseline IV estimates from Table 2 using the placebo data.

Columns (2)-(4) illustrate variations on the baseline model that we consider elsewhere in

the paper, such as other deﬁnition of fans (Section 6.1), excluding the host city as an

instrument, and using Vegas odds to compute expected fans (Section 5.3). The notes from

Table 2 apply here as well.

Table 8: Placebo and advertised movies

log(Google Searches on Release Week)

(1) (2) (3) (4)

Nielsen Super Bowl Ratings -0.483** -0.091 -0.373 0.059

(0.238) (0.374) (0.387) (0.444)

Nielsen X Super Ad 0.797*** 0.853* 1.057** 0.628

(0.305) (0.448) (0.461) (0.483)

log(Pre-Super Search) 0.083*** 0.083*** 0.083*** 0.084***

(0.019) (0.019) (0.018) (0.019)

Adjusted R-squared 0.88 0.88 0.88 0.88

Observations 5,827 5,827 5,827 5,827

Speciﬁcation OLS 2SLS 2SLS (Trends fans) 2SLS (-Host)

* p < 0.1; ** p < 0.05; *** p < 0.01

Notes: City and movie ﬁxed eﬀects are used in all speciﬁcations. See the notes to the

previous table for deﬁnitions. Coeﬃcients for other speciﬁcations are available in the

online appendix.

29

6.4 Interpretation429

The results suggest that an increase of 100 ratings points raises weekend430

ticket sales for a movie advertised on the Super Bowl by at least 50 per-431

cent. Note that 100 ratings points means a switch from 0 percent of people432

watching to 100 percent of people watching. In other words, it measures the433

diﬀerence from a hypothetical situation in which everybody watched the ad434

to a hypothetical situation in which nobody watched the ad.435

Since the Super Bowl averages about 42 ratings points overall, this implies436

that a Super Bowl ad increases release-week ticket sales by about 21 percent.437

In other words, the coeﬃcient suggests there are 21 percent more ticket sales438

when 42 percent of the country watched the Super Bowl than there would439

have been if nobody watched the Super Bowl. The average movie in our440

sample took in $40 million on the opening weekend. Thus the incremental441

ticket revenue from the Super Bowl ad were roughly $8.4 million on average.442

Since a Super Bowl ad cost is about $3 million, this means an overall return443

of 2.8 to 1.444

According to industry practice, the studio typically pays for the entire445

marketing costs and receives 40-50 percent of the domestic box oﬃce revenue.446

(The exact numbers are closely guarded secrets, but see Danzig and Hughes447

[2014] for some estimates.) Hence the return to the studio from the Super448

Bowl ad is about 1.4 to 1, or a 40 percent ROI.

5

449

depending on the speciﬁcation. It is usually negative – suggesting there is substitution –

but only statistically signiﬁcant in one out of four main speciﬁcations.

5

Hartmann and Klapper [2015] estimate a 153 percent ROI for Super Bowl beer ads,

30

We want to emphasize four caveats in interpreting these results.450

First, this back of the envelope calculation ignores future revenue streams451

such as ticket sales after the opening weekend and other revenue through452

home movie purchases, TV licensing, and so on. Some of this additional453

revenue may be attributable to the Super Bowl ad impressions, though we454

have no easy way to measure this.455

However, a causal relationship between increased movie attendance and456

increased home entertainment sales is consistent with Choi et al. [2015] who457

use opening-weekend snowstorms as an instrument and ﬁnd that a 10 percent458

rise in theatrical attendance causes an 8 percent increase in DVDs/Blu-ray459

sales when they are released. Cable licensing deals are also directly tied to460

box oﬃce success so that any increase in box oﬃce revenue will positively461

impact revenue from this channel.462

We also do not know how the incremental revenue is divided among the463

various parties—how much goes to the studios, producers, writers, stars,464

and so on. Similarly, we don’t know exactly how the costs of the Super Bowl465

ad are divided among the various parties. However, as indicated above, it466

appears that studios are the primary decision makers with respect to Super467

Bowl ads and bear most of the marketing costs.468

Second, in calculating the return to advertising, we are assuming that the469

incremental viewers of the Super Bowl have the same response to ads as those470

who would watch the Super Bowl anyway. It is possible that the committed471

but caution that this is a likely an overestimate.

31

fans pay more attention to the game and less to ads. Or perhaps they are472

much more engaged with the entire experience and so pay more attention473

to ads than the incremental viewers. It is also possible that the incremental474

fans have substantially diﬀerent tastes in movies than the fans you would get475

simply by purchasing more ad slots. We provide some evidence on this in476

Section 7.477

Third, we don’t know how these results extend to other settings, as the478

Super Bowl has unique qualities. There are other similar events such as the479

World Series, basketball playoﬀs, the Summer and Winter Olympics, and so480

on. These natural experiments are not quite as clean-cut as the Super Bowl,481

but are certainly worthy of future study.482

Fourth, one might ask why the estimated return is so high. First, it483

is important to understand that our results pertain to returns on movies484

that the studio has chosen to advertise on the Super Bowl. The return on485

advertising movies with mediocre prospects could be much lower. Second,486

once the network has set a market-clearing price, we would expect that the487

marginal ad would earn a normal, risk-adjusted rate of return. However,488

the average ad would typically earn a return higher than the marginal ad.489

One might then ask “if the return to the movie ad is so high, why don’t the490

studios advertise more movies?” The answer to this question may be that491

they only have a few movies for which a Super Bowl ad makes economic492

sense. Movie theaters can only show a limited number of movies at any one493

time, and the conventional wisdom in the industry is that if two blockbusters494

32

are released on the same weekend, the revenues of both movies will suﬀer.495

As a result, studios typically try to stagger the release of blockbusters, so at496

any one time there are only a few movies that would warrant Super Bowl497

treatment. Whatever the explanation, we typically see only 6-8 movie ads498

per Super Bowl and this number does not vary much from year to year.499

Finally, we want to clarify how these results ﬁt with the Super Bowl500

Impossibility Theorem (Lewis and Rao [2012]). They argue that it is nearly501

impossible for a ﬁrm to test the eﬀects of an individual ad campaign, even if it502

randomly assigned DMAs during a Super Bowl. How, then, can we ﬁnd such503

highly statistically signiﬁcant results? The answer is that the Super Bowl504

Impossibility Theorem refers to the question of measuring the eﬀectiveness505

of a single campaign. But here, we study the average eﬀect of 70 campaigns.506

The noise level is too high to say anything about the eﬀects of a particular507

advertisement, but the average performance of all movies in our sample can508

be estimated reasonably precisely.509

7 Heterogeneous treatment eﬀects510

We have shown that the incremental ad exposures due to the home-team511

eﬀect have a causal impact on both Google queries and opening weekend512

revenue. This suggest that increased ad expenditure would also have an513

incremental impact on these outcomes. However, the incremental ad views514

from the home-team eﬀect may well be diﬀerent than the incremental ad515

33

views from simply spending more money on advertising.516

We can oﬀer some suggestive evidence on this point. We ran a Google517

Consumer Survey and asked the 2,568 respondents whether they watched518

the Super Bowl on TV in 2013, 2014 or both years. The question of interest519

was whether those who watched both years were diﬀerent than those who520

watched only one year. The dimensions on which the respondents could diﬀer521

were inferred age, gender, and income.

6

522

We found that those who watched the Super Bowl in both years, rather523

than a single year, tended to be older, more male, and live in wealthier areas.524

However, most of these eﬀects tended to be statistically insigniﬁcant, with525

the exception of gender. We suspect that there is some diﬀerence between the526

incremental viewers from the home-city eﬀect and the incremental viewers527

that would be reached by increased ad spend.528

Nevertheless, we believe that our estimates can be useful in estimating529

the response to ad spend. Suppose that a movie advertiser targeted its ads to530

reﬂect the audience composition of the incremental Super Bowl viewers. This531

targeting could be informed by a more sophisticated version of our survey.532

That advertiser might well expect a response to its ad spend along the lines533

of that described in Section 6.2. So those estimates of the impact of spend534

on box oﬃce should be a lower bound on what ad eﬀectiveness would be if535

ad targeting could be fully optimized.536

6

Inferred age and gender are based on web site visits and inferred income is based on

the IP address of the respondent and Census data.

34

We also can test whether there are diﬀerential eﬀects based on when a537

movie is released. Are ads less eﬀective for movies released well after the538

Super Bowl? We divided our sample into movies with release dates more539

than 70 days out and those with release dates less than 70 days out. We540

recreated the regressions in Table 2. Somewhat surprisingly, we did not see541

a diﬀerence in the eﬀects of ads on box oﬃce sales in these two groups.

7

542

8 Discussion543

We use a natural experiment—the Super Bowl—to study the causal eﬀect544

of advertising on movie demand. Our identiﬁcation strategy relies on the545

fact that Super Bowl ads are purchased before advertisers know which teams546

will play in the Super Bowl and that cities where there are many fans of the547

qualifying teams have substantially larger viewership than other cities do.548

Within this setting we study 70 movies that were advertised during the549

2004-2014 Super Bowls. We compare product purchase patterns for adver-550

tised movies in cities with fans from the qualifying teams to cities with fans551

of near-qualifying teams. We ﬁnd a substantial increase in opening weekend552

revenue due to Super Bowl advertisements. On average, the movies in our553

sample experience an incremental increase of $8.4 million in opening weekend554

7

In general, we don’t have suﬃcient power to break down the treatment eﬀects. There

are several other interesting questions, such as whether there are diﬀerential eﬀects for

movies with more competition, but we have to leave these questions for further research.

It may be possible to investigate such issues after we accumulate a few more years of Super

Bowl data.

35

box oﬃce revenue from a $3 million Super Bowl advertisement.555

We suggest that our methodology can be generalized to a variety of sports556

settings where the nature of qualifying creates a large random shock to ad557

viewership in a particular area, and that this methodology has notable ad-558

vantages compared to the more common approach of using ﬁeld experiments559

to determine the causal impact of advertising. The best identiﬁcation comes560

from sporting events such as the Super Bowl in which the teams that will play561

are unknown at the time companies purchase advertising spot. However, even562

if the home cities are known it seems to us unlikely that advertisers would563

take this information into account when choosing its ad expenditure. So the564

methodology could well be applicable for a broader set of media broadcasts565

with diﬀerential appeal across geographies.566

36

References567

Daniel A. Ackerberg. Advertising, Learning, and Consumer Choice in Ex-568

perience Good Markets: an Empirical Examination. International Eco-569

nomic Review, 44(3):1007–1040, August 2003. ISSN 0020-6598. doi:570

10.1111/1468-2354.t01-2-00098. URL http://doi.wiley.com/10.1111/571

1468-2354.t01-2-00098.572

Marianne Bertrand, Dean Karlan, Sendhil Mullainathan, Eldar Shaﬁr,573

and Jonathan Zinman. What’s Advertising Content Worth? Ev-574

idence from a Consumer Credit Marketing Field Experiment.575

The Quarterly Journal of Economics, 125(1):44, 2010. URL576

http://www.ingentaconnect.com.ezp-prod1.hul.harvard.edu/577

content/oup/qje/2010/00000125/00000001/art00007.578

Patrick Choi, Peter Boatwright, and Michael D. Smith. The perfect storm:579

Using snowstorms to analyze the eﬀect of theatrical attendance on de-580

mand for subsequently released DVDs. Technical report, Carnegie Mellon581

University, 2015. URL http://ssrn.com/abstract=2639303.582

William S Comanor and Thomas A Wilson. On Advertising and583

Proﬁtability. The Review of Economics and Statistics, 53(4):408–10,584

1971. URL http://econpapers.repec.org/RePEc:tpr:restat:v:53:585

y:1971:i:4:p:408-10.586

Scott Danzig and Mark Hughes. Breakdown of587

movie costs, 2014. URL http://www.quora.com/588

What-is-the-breakdown-of-costs-associated-with-making-a-high-budget-Hollywood-film.589

Douglas Galbi. U.S. Annual Advertising Spending Since 1919, 2008. URL590

http://www.galbithink.org/ad-spending.htm.591

Wesley R. Hartmann and Daniel Klapper. Super Bowl ads. Technical report,592

Stanford Graduate School of Business, 2015. URL http://faculty-gsb.593

stanford.edu/hartmann/SuperBowl.pdf. First version presented at IN-594

FORMS Marketing Science, June 7-9, 2012.595

Joseph F. Heyse and William W. S. Wei. Modelling the Advertising-Sales596

Relationship through Use of Multiple Time Series Techniques. Journal597

of Forecasting, 4(2):165–181, 1985. ISSN 02776693. doi: 10.1002/for.598

3980040206. URL http://doi.wiley.com/10.1002/for.3980040206.599

37

Daniel Ho, Kosuke Imai, Gary King, and Elizabeth Stuart. Matching as600

nonparametric preprocessing for reducing model dependence in parametric601

causal inference. Political Analysis, 2007a. URL http://gking.harvard.602

edu/files/abs/matchp-abs.shtml.603

Daniel Ho, Kosuke Imai, Gary King, and Elizabeth Stuart. Matchit: Non-604

parametric preprocessing for parametric causal inference. Journal of Statis-605

tical Software, 15(3):199–236, 2007b. URL http://gking.harvard.edu/606

matchit/.607

Jason Ho, Tirtha Dhar, and Charles Weinberg. Playoﬀ payoﬀ: Super Bowl608

advertising for movies. International Journal of Research in Marketing,609

26:168–179, 2009. URL www.elsevier.com/locate/ijresmar.610

Ye Hu, Leonard M. Lodish, and Abba A. Krieger. An Analysis of Real World611

TV Advertising Tests: a 15-Year Update. Journal of Advertising Research,612

47(3):341–353, 2007.613

Junsoo Lee, B. S. Shin, and In Chung. Causality Between Advertising and614

Sales: New Evidence from Cointegration. Applied Economics Letters, 3(5):615

299–301, 1996. ISSN 1350-4851. URL http://econpapers.repec.org/616

RePEc:taf:apeclt:v:3:y:1996:i:5:p:299-301.617

Randall A. Lewis and Justin M. Rao. On the Near Impossibility of Measuring618

Advertising Eﬀectiveness. Working Paper, 2012.619

Randall Aaron Lewis and David H. Reiley. Does Retail Advertising Work?620

Measuring the Eﬀects of Advertising on Sales Via a Controlled Experiment621

on Yahoo! SSRN Electronic Journal, June 2008. ISSN 1556-5068. doi:622

10.2139/ssrn.1865943. URL http://papers.ssrn.com.ezp-prod1.hul.623

harvard.edu/abstract=1865943.624

Rui Miao and Yueyue Ma. The dynamic impact of websearch volume on625

product — an empirical study based on box oﬃce revenues. WHICEB 2015626

Proceedings, 2015. URL http://aisel.aisnet.org/whiceb2015/14.627

Reggie Panaligan and Andrea Chen. Quantifying movie magic with google628

search. Google Think Insights, 2013. URL http://ssl.gstatic.com/629

think/docs/quantifying-movie-magic_research-studies.pdf.630

38

Richard Schmalensee. A Model of Advertising and Product Quality. Journal631

of Political Economy, 86(3):485–503, 1978. URL http://ideas.repec.632

org/a/ucp/jpolec/v86y1978i3p485-503.html.633

Roger Sherman and Robert D. Tollison. Advertising and Prof-634

itability. The Review of Economics and Statistics, 53(4):397–407,635

1971. URL http://econpapers.repec.org/RePEc:tpr:restat:v:53:636

y:1971:i:4:p:397-407.637

Duncan Simester, Yu Hu, Erik Brynjolfsson, and Eric T. Anderson. Dynam-638

ics of Retail Advertising: Evidence from a Field Experiment. Economic639

Inquiry, 47(3):482–499, 2009. ISSN 0095-2583. URL http://econpapers.640

repec.org/RePEc:bla:ecinqu:v:47:y:2009:i:3:p:482-499.641

Hal Stern. The probability of winning a football game as a function of the642

pointspread. Technical report, Department of Statistics, Stanford Univer-643

sity, 1986.644

Rama Yelkur, CHuck Tomkovick, and Patty Traczyk. Super Bowl advertising645

eﬀectiveness: Hollywood ﬁnds the games golden. Journal of Advertising646

Research, 44(1):143–159, 2004. URL http://journals.cambridge.org/647

abstract_S0021849904040085.648

39