You are on page 1of 39

Advertising Effects in Presidential Elections

June 15, 2011

Abstract

We estimate advertising effects in the context of presidential elections. This setting overcomes many data challenges in previous advertising studies, while arguably providing one of the most interesting empirical settings to study advertising’s effects. The four-year presidential election cycle facilitates measurement in two ways. First, the gap between elections depreciates past advertising stocks such that large advertising investments are concentrated within relatively short periods. Second, the lack of political advertising between elections allows lagged advertising prices to serve as instruments that are safely independent of candidates’ current advertising choices. To further aid estimation, the winner-take-all nature of the electoral college generates broad variation in advertising levels across states. We analyze the data using an aggregate discrete choice approach with extensive fixed effects at the party-market level to control for unobservable cross-sectional factors that might be correlated with advertising, outcomes, and instruments. The results indicate significant positive effects of advertising exposures for the 2000 and 2004 general elections. Advertising elasticities are smaller than are typical for branded goods, yet significant enough to shift election outcomes. For example, if advertising were set to zero and all other factors held constant, three states’ electoral votes would have changed parties in 2000, leading to a different president.

Keywords: advertising, instrumental variables, presidential elections, politics.

1

Introduction

Advertising is ubiquitous. In 2008, firms spent roughly $65 billion on television advertising alone on products ranging from branded goods to political candidates.1 The prevalence of advertising suggests it must be influential. Consequently, the study of advertising often turns to understanding what it affects and why: economists and marketers debate whether advertising is informative or persuasive; marketers also assess its effects on intermediate measures such as brand recall; political scientists wonder if negative advertisements depress voter turnout.2 Nevertheless, conclusive evidence on the efficacy of advertising is still quite elusive. Most papers lack any source of exogenous variation, and those studies with experimental variation have had trouble detecting robust effects.3 We focus on the question of effectiveness in the context of presidential elections. The potential to alter the choice of president of the United States is perhaps one of the widest-reaching decisions on which advertising focuses. Most advertising occurs at the brand level, seeking to influence individual consumers or households. An election aggregates the decisions of many into a single outcome affecting the inhabitants of that country and countless others. Concerns over advertising’s ability to affect choice are therefore substantially greater in this setting, leading to debates about whether to impose fundraising and spending limits in elections. Despite these concerns, the evidence on advertising effects in elections is inconclusive (Gordon et al. 2010). Some studies suggest only a positive turnout effect
Source: TNS Media Intelligence. For work analyzing whether advertising is informative versus persuasive, see, for example, Nelson (1974), Anand and Shachar (2000), Ackerberg (2001), Narayanan and Manchanda (2009), Clark, Doraszelski, and Draganska (2009). Draganska and Klapper (2010) incorporate the effects of brand recall through advertising on demand, and Kanetkar, Weinberg, and Weiss (1992) and Mela, Gupta, and Lehmann (1997) measure the effects of advertising on price sensitivity. Ansolabehere et al., (1999), Wattenberg and Brians (1999), and Freedman and Goldstein (1999) investigate the effects of negative advertisements on voter turnout. 3 Lodish et al. (1995) conduct a meta-analysis of split cable television experiments and do not find conclusive positive effects of advertising. In the context of internet advertising, the experimental variation alone in Riley and Lewis (2009) was unable to find significant positive advertising effects.
2 1

1

(Shachar, 2009). Others suggest the effect is entirely persuasive (Huber and Arceneaux, 2007; Lovett and Peress, 2010). Some debate whether negative ads decrease turnout or not (Wattenberg and Brians, 1999; Ansolabehere et al., 1999). Presidential elections provide a data-rich setting well suited to identify the causal effects of advertising. Two challenges in estimating the effects of advertising are econometric endogeneity and disentangling the effects of past and present advertising. First, as with most empirical questions, the existence of unobservables creates an endogeneity problem in isolating the causal effect of advertising. Potential instruments are variables that enter the decision process of advertisers, but not that of the targets to be influenced. One potential variable is the price of advertising, which is excluded from demand just as costs are excluded when seeking instruments for price.4 A problem with using contemporaneous prices is that advertisers may not be price takers. The impact of demand for political advertising on advertising prices was noted during the 2010 midterm elections in a number of markets.5 To address this issue, we take advantage of the fact that there are no elections during odd years, and thus use the prior year’s advertising prices as cost instruments that are net of political campaign effects. The second challenge is that advertising effects are typically spread over long horizons and multiple choice occasions. This fact may help explain why the few studies that have causally identified positive advertising effects have primarily examined new products (e.g. Ackerberg, 2001, Lodish et al., 1995, Eastlack and Rao, 1989). Much of this literature, motivated in part by Nerlove and Arrow (1962), focuses on measuring latent advertising stocks that depreciate and are reinvested over long horizons (Naik, Kalyan, and Srinivasan, 1998; Dub´, Hitsch, and e
Dub´ and Manchanda (2005) and Doganoglu and Klapper (2006) both use advertising costs as instruments. e Associated Press, “Sick of Campaign Ad Avalanche? TV Stations Aren’t,” October 30, 2010. Accessed at http://finance.yahoo.com/news/Sick-of-campaign-ad-avalanche-apf-3274707232.html.
5 4

2

The electoral college system distorts advertising incentives across geographic areas such that advertising varies from zero in some markets to significant per-capita levels in battleground states. 2011).500 advertising exposures and resulting vote shares. Rutz and Bucklin. political advertising concentrates both the choices and spending into a well-defined short window of time. Although advertising capital stocks may depreciate a little during the post-primary period. but the absence of advertising between elections suggests a substantial depreciation of any advertising stocks. by pooling candidate-share observations across counties and two elections. The estimates show a robust positive advertising effect across a number of specifications. 2005. The fixed effect shifts inference to how within-market changes in advertising prices between two elections indirectly affect within-market changes in vote shares. Focusing on within-market variation removes the worry that unobservables in the candidate choice equation might be cross-sectionally correlated with the advertising price instrument. Such a correlation could exist because major metropolitan areas have higher advertising prices and tend to lean Democrat.Manchanda. Spending in presidential general elections is concentrated in the post-primary period between Labor Day and Election Day. providing us with rich variation to estimate advertising’s efficacy. We use advertising data from the 2000 and 2004 presidential elections to measure the effect of advertising on county-level voting decisions. we are able to observe over 9. Advertising elasticities are smaller than estimates typically found in consumer packaged goods 3 . we can avoid specifying a stock of advertising goodwill or awareness that might depend heavily on initial conditions. Choices are fully concentrated on Election Day. Another concern might be the persistence of advertising from past elections. To measure the advertising effect as cleanly as possible. we include an extensive set of fixed effects at the market-party level. Furthermore. Fortunately. We estimate an aggregate discrete choice model in which the candidate and his advertising influence voters’ decisions.

we consider how the absence of advertising would have affected state outcomes. First. Lacking information on exposure rates. and Hillygus. Gerber. and Shanmugam (2007) and Rekkas (2007). 1988. and most critically. Our paper contributes in two ways to the literature on measuring the effects of political advertising. for which candidate to vote. Huber and Arceneaux 2007). our model combines a voter’s decision to turn out to vote. Due to the difficulty of identifying reasonable instruments. and if so.categories. more recent work in political science employs field or natural experiments to estimate advertising’s causal effect (Gerber et al. We find that some states in both the 2000 and 2004 elections shift sides. but they still serve as a useful benchmark for evaluating an election outcome’s sensitivity to advertising. the authors use the total number of ads aired. 6 4 . The point of this exercise is merely to highlight that advertising’s causal effects are great enough to shift the election outcome and that advertising can be disproportionate between candidates. our particular instrumental variables strategy allows us to address the endogeneity of advertising. with the 2000 shift being sufficient for Gore to overtake Bush in electoral votes. To provide a better metric for the role and importance of advertising. Second. 1998. Iyer. whereas most prior work considers these decisions separately (Ashworth and Clinton 2006. and hence electoral votes. Rekkas (2007) studies the effects Degan and Merlo (2009) study participation and voting in an uncertain voter model but do not consider the effects of advertising.6 Perhaps the two papers closest to our own are Che. 2005). The former estimates an individual-level nested logit model using a combination of voter surveys and market-level advertising quantity data. These results should not be interpreted as a strict prediction since our demand specification must hold many factors fixed. 2011). which has long been recognized as a challenge in the political science literature (Green and Krasno.

in all but a few cases. We measure voting outcomes at the county level. hereafter BLP) style model. Section 4 presents the estimates and the zero-advertising analysis. which. Electoral votes are measured at the state level. Section 3 describes the aggregate discrete choice demand model. 2 Data This section details our data sources and approach to constructing our instruments. Our strategy of focusing on within-market changes alleviates such a concern about our analysis. but candidates set advertising quantities at the media market level (DMA). only includes one media market.7 2. Shachar (2009) also analyzes whether candidate marketing mix variables affect voting outcomes.of overall campaign spending on Parliamentary elections in Canada using a Berry. Both papers consider only a single election year. The data vary in the geographic unit at which they are measured. but restricts advertising to effect turnout and not candidate choice and does not account for potential unobservable shocks. Section 5 concludes. and Pakes (1995. only five belong to multiple DMAs.596 counties. which can span multiple states. We use zip-code-level population data to weight the advertising proportionally according to the share of the population in a given state. The remainder of the paper is structured as follows. 7 5 . Levinsohn.1 Advertising The primary advertising data come from the Campaign Media Analysis Group (CMAG) for the 2000 and 2004 presidential elections. such that identifying advertising’s effects rests on cross-market variation that might be confounded with market-party unobservables. The next section describes the advertising and election outcome data we use. CMAG monitors political advertising activity on all Of the 1. and were made available through the University of Wisconsin Advertising Project.

The standard practice in analyzing advertising effects.g. See Freedman and Goldstein (1999) for more details on the creation of the CMAG data set. Although we do not directly observe GRPs in our data. which give a measure of the number of exposures per capita. the national party. however. a measure which cannot account for the number of viewers that observe the ad. Since our identification strategy focuses on cross-election changes in outcomes. Independent. Republican. etc. and 807. The data provide a complete record of every advertisement broadcast in each of the country’s top designated media markets (DMAs). representing 78% of the country’s population. Democrat. We further concentrate on all advertisements appearing after Labor Day. One key variable we observe is an estimate of the ad’s cost that CMAG calculates. Television ads are the largest component of media spending for political campaigns according to AdWeek (2009). the length in seconds. we restrict all subsequent analysis to the 75 largest DMAs. Total spent on television advertising by all candidates was $168 million in 2000 and $564 million in 2004. The data contain a large number of individual presidential ads: 247.national television and cable networks. the candidate supported (e. The GRP measure therefore captures the actual “quantity” relevant for impact. we reconstruct the GRPs based 6 ..643 for the 75 largest DMAs in 2000. or “hybrid/coordinated”).). is to use Gross Rating Points (GRPs). We use the candidate supported to assign it to the appropriate candidate and include all advertisements regardless of the sponsoring group.g. GRPs fundamentally differ from the number of ads. we observe all the dates and times which it aired. For each ad..296 for the 100 largest DMAs in 2004. and the sponsoring group (e. As a measure of advertising effectiveness. independent groups. and assigns each advertisement to support the proper candidate. the candidate. which marks the conclusion of the primaries and the beginning of earnest competition in the general election.

8 7 . Second. a market research firm that specializes in estimating media costs.8 This second issue is unlikely to be present in our data because CMAG reconstructed the cost from actual GRPs they observed and their own estimates of the CPP. We obtain the total GRPs by summing across ads: GRPtmj = a Costatdmj /CP Ptdm where a indexes an ad. Candidates might pay a higher-than-average advertising price in the week prior to Election Day. then match each advertisement with the corresponding CPP according to the three categories just mentioned.on the advertisement’s cost and the price per GRP. Our GRP measure contains two potential sources of measurement error. We focus on the 18-and-over demographic to align with voting age. so the true CPP likely differed from our data. We use this measure of GRPs as our advertising variable later in our model. t the 3rd quarter of the election year. First. These data provide quarterly forecasts of CPP by market. We obtain the CPP data from SQAD. the price a particular candidate received could have been above or below the market CPP. We therefore believe our GRP measure suffers only Even with a unforecasted CPP. the CPP we observe is a forecast made by SQAD. so the costs we observe do not include such candidate specific CPPs. provided that this particular measurement error is purely random. which is commonly referred to as the Cost-Per-Point (CPP). the price represents an average taken over the entire quarter. divided by the cost per GRP in that market and daypart. The equation above simply represents the total advertising expenditure by a candidate in a particular market and daypart. and j the candidate. then it will simply be absorbed into the unobservable shock terms we include in the model. The ratio of these two quantities should roughly equal the number of GRPs a candidate purchased in the two month period leading up to election day. d denotes the daypart. However. population subcategory. and time slot (daypart). m the media market.

and from $143 to $139 for Democrats. the average price for Republicans dropped from $151 to $144 per point in 2004.from forecast error in measurement. The Republicans chose not to advertise at all in 20 markets in 2000 and 32 markets in 2004. the advertising prices for the most common daypart (early news) increased by about five percent. The structure of the electoral college creates particular incentives that drive much of the observed pattern of advertising. Dividing the total expenditures by the GRPs. we are fortunate to observe rich variation in total advertising expenditures and GRPs between 2000 and 2004. non-battleground states. First. consistent with the growing importance of advertising in political elections. receive little to no advertising from either candidate because any such intervention would not be expected to alter the state’s election outcome. but the cost is prohibitive. Table 2 shows that both total ad quantities and total ad expenditures significantly increased from 2000 to 2004. the changes in GRPs across candidates are highly correlated. The figure also exhibits significant variation both between elections and across DMAs. However.10 Second. Conversely. We explore candidates’ advertising allocation decisions in detail in Gordon and Hartmann (2010). For Democrats. We have explored obtaining the GRP data directly from CMAG and Nielsen. The support of the advertising distribution ranges from zero to about six million dollars in 2000 and about nine million dollars in 2004. where a given party expects to win by a handsome margin. given that our estimation strategy focuses on within-DMA variation. perhaps by selecting cheaper dayparts. the numbers are 28 in 2000 and 25 in 2004. Figure 1 plots the change in GRPs from 2000 to 2004 for the Republican and Democratic candidates. 10 9 8 . Two important points are worth noting.9 Table 2 displays summary statistics for the major party candidates’ advertising in 2000 and 2004. so-called battleground states receive a disproportionate share of a candidate’s advertising due to the expected narrow margin of victory. This observation indicates that candidates may have found more economical means of exposures in 2004. As expected. In particular. we observe significant variation in advertising across the markets within a given election year. indicating the candidates tended to increase and decrease spending in the same DMAs.

11 9 . we use the prior year’s advertising price (1999 for 2000 and 2003 for 2004). ignoring price endogeneity leads the researcher to underestimate price sensitivity because the unobservables are positively correlated with prices. one might be concerned that the measurement errors in the lagged CPP estimates Using lagged prices instead of current also reduces the potential concern of having the instrument enter the denominator of the variable for which it is instrumenting in the GRP approximation in the preceding section. BLP). it is possible that candidates purchased enough advertising in a given market to have an effect on the equilibrium price of advertising in that market. Candidates are both unlikely to advertise in markets where they have little chance of winning and are unlikely to advertise in markets where they strongly expect to win. such that candidates would no longer act as price-takers. one obvious potential variable affecting advertising allocations but unlikely to affect voters’ preferences is the price of advertising.. whether unobserved demand shocks are substantially higher or lower in the presence of more advertising is unclear.2 Instruments A central issue in our empirical application is to address the endogeneity of candidate advertising. However. the direction of the bias is ambiguous. when market advertising prices were free of political factors. Although we do not model the candidates’ decisions here. Therefore.g. We consider a candidate’s decision process to find suitable instruments. There are two potential concerns that arise from such an instrument choice. To avoid this concern.11 Second. In standard differentiated product choice contexts (e.2. in the context of political candidate choice. We naturally expect the advertising variation depicted in Figure 1 to reflect some knowledge the candidates observe. making it hard to sign the direction of the endogeneity bias. This would invalidate the instrument since the price would not be exogenous to a candidate’s advertising decisions. which informs their advertising decisions. but that we do not. First.

with only a few markets experiencing declines.due to SQAD’s methodology could be systematically related to current CPP estimates.55. Table 3 provides summary statistics of the change in the CPM between 1999 and 2003 in each daypart. whereas the maximum is between 10 . yet Gore. To convert the instrument to a per-capita basis. and the importance of each daypart CPM varies across DMAs. Most CPMs increased over this period. the CPM in a market varies over the dayparts because the cost of reaching a thousand viewers varies over the day. Our motivation for using CPM is similar to our decision to use GRPs as our endogenous advertising variable since CPMs more accurately account for exposures per capita. for instance. For example. we use the cost per thousand impressions (CPM) instead of per point. We do not expect such a systematic bias to exist because SQAD updates its advertising price predictions each quarter to account for realized prices in the past quarters. which seems unlikely given the nature of the firm’s business. Although 30 percent of Bush’s GRPs in Spokane were in Early News. If the measurement errors were correlated. each daypart CPM is potentially relevant for advertising decisions. As with the CPP. then SQAD would be making a systematic mistake in the same direction. We use the CPM in each of the eight dayparts as instruments. Daypart CPMs are correlated—though not perfectly—with one another. We observe significant variation across candidates in when they choose to advertise during the day and in the cost of advertising across elections. The Early News and Daytime slots are the most common across DMAs. less than 15 percent in Milwaukee were in Early News. bought fewer GRPs in Kansas City during the Early News than in Prime Access or Late Fringe. Given this mix. Figures 2 and 3 show how each candidate spread his GRPs across dayparts in ten DMAs in 2000. the minimum correlation is between Daytime and Prime Time at 0. and demonstrates substantial variation in the daypart CPMs over time.

and the results suggest that congressional campaign spending have little effect on voting outcomes. First. Ansolabehere et al. Recognizing this issue. A wealthier candidate should be able to spend more on advertising. Given the motivation for our particular instrumental variables. Gerber (1998) uses a combination of instruments. including a measure of the challenger’s personal wealth and the state voting-age population. This instrument seems valid in a Senate context but does not transfer to a presidential election setting. Most of this analysis focuses on congressional races to take advantage of a larger number of independent observations. instrumental variable techniques gained early traction in work that sought to measure the effects of aggregate candidate campaign spending on voting outcomes (Jacobson 1978). Work in political science uses both instrumental variables and field/natural experiments to deal with the endogeneity of candidate choice variables. (1999) consider the effects of negative ads on voter turnout using GRPs as instruments. Figure 4 illustrates how the Early News CPM varied within DMAs between 2000 and 2004. A candidate has more citizens from which to raise funds in more populous states. More recent work tends to focus on specific candidate activities such as television advertising or direct mailings. although campaign spending levels are much lower than a presidential campaign.93. but the GRPs are a choice variable the candidate potentially determines in response to an econometric unobservable in the choice equation.the Late Fringe and Prime Time at 0. Green and Krasno (1988) use lagged incumbent spending in Senate elections to instrument for current incumbent spending. Differencing eliminates any fixed candidate or local influences. 12 11 . but must assume challenger spending is exogenous. although one concern might be that candidate wealth is correlated with an unobserved quality attribute that affects voters’ decisions. we now compare our approach to the extant literature.12 Levitt (1994) addresses candidate unobservables by examining congressional races in which the same two candidates faced one another in multiple elections.

Together. the Republicans had higher average shares per county. exposing voters in the non-battleground state to higher levels advertising than the candidate intended. these voting outcomes reveal that the Democrats tend to do better in See the web page maintained by Michael McDonald at http://elections.htm for more information on measures of voter turnout.596 counties. Gerber et al. 2. For each of the 1. The VAP estimates serve as our market-size parameters and allow us to calculate a measure of voter turnout at the county level. They link the advertising data to survey data from the National Annenberg Election Surveys (NAES) that reports individuals’ campaign interest and intentions to vote. The Democrats had a higher number of average votes per county in both years than did Republicans.polidata.org. Huber and Arceneaux (2007) take advantage of the fact that some media markets overlap battleground and non-battleground states.3 Votes The county-level vote data are available from www. and found evidence supporting the notion that advertising influences voters’ choice of candidate but not whether to turn out to vote. finding that televised ads had strong but short-lived effects on voting preferences. is only available at the state level. we observe the number of votes cast for all possible candidates and the size of the voting-age population (VAP). On the other hand. 13 12 . a more accurate measure for calculating turnout that removes non-citizens and criminals. (2011) conducted a randomized field experiment in the 2006 gubernatorial election in Texas to examine the effect of advertising on voters’ stated attitudes and intentions.A second approach that has recently become popular is to exploit natural experiments or to conduct field experiments to generate exogenous sources of variation.13 Table 1 summarizes the total votes and vote shares of each candidate by county and election year.edu/voter_turnout. The voting-eligible population (VEP).gmu.

we take the voting outcomes in these excluded counties as fixed. and collective third-party candidates. In 2004. and David Cobb (Green Party). However. In 2000. who ran on the Green Party ticket in 2000 and as an independent in 2004. Democrat. 2. and many spent little on advertising. In our estimation.74 percent of the popular vote. the other candidates were Michael Badnarik (Libertarian Party). accounting for some within-market preference shifts is useful. (2) variables that affect voter turnout but not candidate choice (the occurrence of a senate or gubernatorial election and local weather patterns). In 2000. Estimating the model without aggregating the smaller candidates into a single option is possible. Whether excluding such Republican-leaning counties would bias our parameter estimates is unclear. Michael Peroutka (Constitution Party). we prefer to aggregate them so we can focus on measuring the effectiveness of advertising for the Republican. Thus. We include three categories of variables that attempt to absorb some of the remaining within-market variation: (1) variables that measure local political preferences (market-level party affiliation and distance to the candidate’s home state). and John Hagelin (Natural Law Party).5 percent). summing the votes and GRPs across these candidates. In computing our counterfactual in which we set advertising to zero. our data omit a greater number of Republican votes.larger counties. and both times did not win any electoral votes.4 Other Explanatory Variables By focusing on within-market variation. fixed effects absorb the systematic variation across geographies.38 percent in 2004. some of the other candidates included Harry Browne (Libertarian Party). By focusing on counties within the top 75 DMAs. Nader received 2. the majority of voters were probably unaware of many of these candidates. 14 13 . the other non-major party candidates’ vote shares were very small (below 0. but only 0. such that we can estimate the advertising effect as cleanly as possible. and (3) The only third party candidate to run in both elections who had any public visibility was Ralph Nader.14 With the exception of Ralph Nader in the 2000 election. we group all candidates not belonging to one of the two major parties into a single third-party candidate option. Howard Phillips (Constitution Party). That said.

373 observations for 2000 and 81. Gomez. resulting in 58. In addition. a hotly contested senate seat or governor race could generate some spillover effects. In each year. First. whereas the greatest increase was 7 points. Between 2000 and 2004. we merged the six national cross-sectional surveys into a single data set. but we also want to include variables that primarily affect a voter’s decision to turn out for the election.4 percentage points while the percentage of registered Democrats increased 1. we include countylevel estimates of rain and snowfall on Election Day from the National Climatic Data Center’s “Summary of the Day” database (obtained through EarthInfo. we want to include some demographic type variables which exhibit some variation between the two elections. and Krause (2007) show that weather can play a significant role in affecting voter turnout in presidential elections. the percentage of Republicans increased about 2. we include as a control the distance (in miles) between a given DMA and the candidate’s home state. Some county’s preferences may shift to the left or right depending on their match with the incumbent party or even the local political climate. Finally. we include separate variables to indicate whether a senate or gubernatorial election also occurred that year. The party affiliation variables above are designed to capture some variation in preferences across parties.demographic and economic variables (population age demographics and local unemployment). whereas Democratic shares of registered voters dropped by at most 5. Republican shares varied between 10 positive and negative percentage points at the extremes.422 observations for 2004. We therefore use the National Annenberg Election Surveys to include a measure of the percentage of voters in a media market who identify with a political party. Hansford.3 points on average across all DMAs. Second. and hence candidates. within a market. We chose to include census data on the percentage of the county’s 14 . Inc).5 percentage points in a DMA. Although presidential elections are much more likely to drive turnout.

as in BLP. 3 Modeling Voter Preferences We specify an aggregate discrete choice model of demand for political candidates. Due to the lingering effects of the baby boom and migration patterns. To capture variation in the local economic conditions. Atc is a vector of advertising levels across candidates. and εitcj captures idiosyncratic variation in utility across voters. Xtc is a vector of county-level observables that may affect voters’ decisions to turn out for the election. 3. αi is a vector of parameters that captures the marginal utility of advertising. candidates. One key distinction between our model’s specification and that found in other differentiated products settings is that we allow the advertising of other candidates to effect the utility from choosing a given candidate. We allow for unobserved heterogeneity. and include a rich set of fixed effects and other control variables to explain the variation in county-level voting patterns. and periods. γmj represents market-party fixed effects. (1) where βitj is a voter-specific taste for a candidate from party j in election t. the price of other products does not generally enter the utility for a given product.1 Voter Utility A voter’s utility for candidate j in election t is: uitcj = βitj + αi Atc + φ Xtc + γmj + ξtcj + εitcj .population that is between the ages of 45 and 64 and the percentage 65 or older. If a voter 15 . we obtain percentage unemployment data at the county level from the Bureau of Labor Statistics (BLS). In differentiated product contexts. these percentages change even within the four-year time span.

Note that by permitting Atc to be a vector over multiple candidates’ advertising levels. To capture heterogeneity in voter preferences.2. we allow the party-election-specific intercepts 16 . The market-party fixed effects γmj help fit the mean utility level for a party in a specific market. she selects the outside good and receives a (normalized) utility of uitc0 = εitc0 . we include only a subset of all candidates’ advertising levels in Atc . corresponding to time-specific deviations from the unobserved market-party mean utility. ξtcj is a time-county-candidate-specific demand shock that is perfectly observable to voters when casting their votes. We make this change to allow for the possibility that another candidate’s advertising may affect a voters’ decision to vote for a given candidate. The fixed effects also control for the endogeneity of advertising through any unobserved characteristics at the market-party level. and we discuss the identification strategy for allowing multiple advertising variables in the utility function in Section 3. We use the variables in Xtc to control for observable factors at the county level that may change over time which would not otherwise be captured in our fixed effects. We do require an instrument to address the remaining unexplained variation. In our estimation.does not turn out for the election. we relax the common assumption in aggregate discrete choice models that only variables indexed to alternative j enter the choice-specific utility of j. removing any concern of a correlation between advertising and market-specific party preferences without relying on an instrument. but is unobservable to the researcher. Candidates have beliefs about the demand shocks ξtcj that induce endogeneity in candidates’ advertising strategies.

α exp{βitj + αi Atc + φ Xtc + γmj + ξtcj } dF (β. θ) = β. it allows for a correlation between the intercept for candidates j and k that might occur. Specifically. demand side covariates and instruments across time and many markets.J} The model of voter choice above does not consider whether a voter acts strategically based on whether she expects her vote to be pivotal in deciding the election outcome. Conlin and Moro 2008). we integrate over the idiosyncratic shocks to obtain the following vote shares: ˆ stcj (Atc .i.... Allowing for off-diagonal terms in Σ is important to remove the property of independence from irrelevant alternatives (IIA) common to logit demand models..g. α).d. or decides not to vote. Each voter either selects the candidate who gives her the highest utility.and the marginal utility of advertising to vary across individuals. advertising levels. for instance. We assume   ¯tj  βitj   β    ∼ N (  . Coate. (3) exp{βitk + αi Atc + φ Xtc + γmk + ξtck } k∈{0. 3. 15 The model assumes voters act sincerely in casting their votes.2 Identification We observe variation in vote shares. Σ) αi α ¯   (2) where Σ is the full covariance matrix of voter tastes.15 Assuming {εitcj }j are i. There are two distinct factors in our identification which we now discuss. the effect vanishes in larger elections (Feddersen and Pesendorfer 1996). 17 . ξtc .. Identification of the parameters follows from standard arguments when estimating random coefficient models of demand using aggregate market shares data. Xtc . Although voters’ expectations of being pivotal can play a role in small elections (e.. when individuals who vote for either candidate are more similar than those individuals less likely to turn out at all.

XZ would enter the optimal advertising equation and form a relevant interaction for the reduced-form of the model. we require another instrument when the opposing candidates’ advertising is included.16 Yet. Although the sign of the bias on the advertising coefficients is ambiguous (because both high and low demand shocks favor little advertising by a candidate). it is clear from the specification of our model that ω and X will be interacted on the supply side. Since advertising cost is common to both candidates. suggesting such an interaction might not be a viable instrument. Furthermore. ω. In the typical additively linear simultaneous equations model. We now discuss the instrumenting strategy in this context. Under the more realistic specification we outline below. suggesting it is not a viable instrument for advertising. we use market-party fixed effects which narrow the identification to explaining within market variation in a party’s performance based on within market variation in explanatory variables and instruments. This accounts for unobserved market-specific factors which may be correlated with both candidate shares and instruments. Second. the significance of the interaction in a first-stage advertising equation can be tested. consider the candidate’s objective as Such an additively linear specification might be: Vjm = αAm + φXm + ξjm and Ajm = dVm + ωZm + wjm . we consider interactions between the observed advertising cost. To find an excluded variable that differentially affects candidates’ advertising incentives. while most aggregate demand models use cross-market variation for identification. allowing for multiple advertising variables to enter the utility for a particular candidate creates multiple endogenous variables. To see how ω and X interact on the supply-side. the reduced form would not involve an interaction between ω and X.First. The XZ term neither enters the advertising equation nor would ever be interacted in the reduced-form of the model. 16 18 . and X. we still require a proper instrumenting strategy to deal with all the sources of endogeneity.

At−j . θ)] − m=1 ωtm Atmj where Atj = (At1j . Xt . the non-linearity in the solution to the candidate’s optimization problem permits this interaction to serve as a valid instrument. Thus. . Since X and Atmj are not additively separable in the expression for ∂Eξ /∂Atmj . we formulate the estimation objective function as a Mathematical Program with Equilibrium Constraints (MPEC). Xt . Xt . 3. then solution for a candidate’s choice of Atmj must necessarily interact X and ωtm . ξ t . AtM j ) is a vector advertising choice variables over the M = 75 markets. The first-order condition with respect to advertising in a given market is FOC: Rtj ∂Eξ [dtj (Atj . At−j . . To consider the BLP version of the model. following the work of Su and Judd (2010). . A∗ ) set a system of JM such first-order conditions to tj t−j be simultaneously zero. At2j .specified in Gordon and Hartmann (2010). Fox. and ωtm is the observed marginal cost of advertising. ξ t . Thus. whereby the model reduces to a simple aggregate logit model we can estimate via two-stage least squares. and Su (2009). θ) = Rtj Eξ [dtj (Atj . ξ t . .3 Estimation Our first step in estimation treats individuals as homogeneous. who show how to estimate the aggregate e demand model in BLP (1995) by formulating the GMM objective function as an MPEC 19 . θ)] − ωtm . where the parameter Rtj converts the probability of winning into monetary terms. the candidate’s objective function is to maximize the expected probability of winning the election less the total cost of advertising. At−j . dtj (·) is an indicator for whether candidate j wins the election. ∂Atmj The optimal advertising levels (A∗ . The candidate seeks to maximize: M πtj (Atj . We use the approach in Dub´.

chicagobooth. The first two specifications include party and party-year fixed effects. We extend their model to include market-party fixed effects and a full covariance matrix in taste heterogeneity. Without any controls and only a few fixed We altered the Matlab coded posted at http://faculty.edu/jean-pierre. and the remaining specifications add DMA-party fixed effects. although this estimate does not account for the endogeneity of advertising. As we previously indicated. Advertising has a strong positive effect. The next specification in column (2) uses 2SLS to instrument for the advertising levels using the one-year lagged CPMs and interactions with demographics. We begin with an OLS regression in column (1) in which the dependent variable is the difference in the log shares of a candidate and the outside no-vote option. which is equivalent to a homogeneous aggregate logit specification. 4. We also report advertising elasticities and the results of a counterfactual in which we set all advertising to zero and recompute the predicted election outcome.1 Parameter Estimates Table 4 presents the estimates from various specifications. We cluster standard errors at the DMA level to account for the fact that advertising and some of the instruments are the same across counties within a DMA. the sign of the endogeneity bias is ambiguous because advertising occurs when candidates are close in a market.dube/vita/ MPEC to estimate the voter model. 17 20 . significant at the one percent level.problem. The advertising coefficient increases by 60 percent and is significant at the five percent level.17 4 Results This section begins by presenting parameter estimates from a variety of specifications. such that strong positive or negative unobservables both tend toward zero advertising.

Column (5) starts by including the county-level percentage of self-identified Republicans and Democrats interacted with the three candidate intercepts. correcting for endogeneity seems to increase the measured effectiveness of advertising. to have a different influence on the vote share of the Republican candidate than on the Democratic candidate. the unemployment variable is negative and significant. although they use a data set with a much longer time horizon. the advertising coefficient barely changes and only one of the interaction terms is significant. for example. We find that the presence of a gubernatorial election seems to indicate lower overall turnout and senate elections have no significant effect. Column (3) estimates an OLS regression including the DMA-party fixed effects to facilitate comparison with column (1). Hansford.05 but it remains highly significant. despite controlling for cross-sectional unobservables via the DMA-party fixed effects. To illustrate. leading to a reasonable decrease in the advertising coefficient to a value of 0. However. When we introduce the instruments in the 2SLS specification in column (4). The additional fixed effects reduce the advertising coefficient by a third to 0. Interestingly. still highly significant. The next column in Table 4 introduces DMA-party fixed effects.4. Neither of the weather variables are significant. and Krause (2007). in contrast to the findings in Gomez. The interaction permits the percentage of Republicans in a county.053. and these variables should pick up local variation in political preferences.075 and is significant at the one percent level. the advertising coefficient rises again by 50 percent to 0. correcting for the remaining within-county endogeneity still leads to an increase in the advertising coefficient. Next we introduce a series of additional covariates that we explained in Section 2. These fixed effects account for potential correlation between the advertising instruments and cross-sectional variation in voter preferences. Thus. suggesting 21 . The party identification candidate-intercept interaction variables are unchanged. In column (6) we include the rest of our control variables.effects.

The reason is that the DMA-party fixed effects absorb the cross-sectional variation that otherwise helps identify the unobserved heterogeneity. The estimated own-candidate advertising coefficient increases to 0.18 The important point to note is that across all of the specifications. and then conditional on voting. we only include the Democratic advertising in the Republican choice utility and the Republican advertising in the Democratic utility—the third-party candidate’s advertising enters its utility alone and does not affect the other choice options. and the heterogeneity parameters were insignificant. suggesting that an opponent’s advertising reduces the current candidate’s votes. Without the fixed effects. we find significant parameter heterogeneity.that higher unemployment leads to lower voter turnout. The estimated nesting parameter was not significantly different from zero. We also estimated a BLP version of the model that allows for continuous unobserved parameter heterogeneity.083 and is strongly significant. fixed effects. the opposing candidate’s advertising coefficient is negative. despite including numerous control variables. 18 22 . the estimated advertising coefficients were generally the same as those found in the 2SLS specifications. so a nested model does not appear to be helpful. and correcting for the endogeneity of advertising.05. With the fixed effects. Given the low vote shares and advertising levels for the collective third-party option. for which candidate to vote. we rely on non-IIA substitution patterns within a DMA between elections to identify the heterogeneity and correlation in tastes for the candidates. Our last specification in column (7) allows other candidates’ advertising levels to enter the utility for a particular candidate. the coefficient is insignificant. the own-candidate advertising coefficient robustly remains positive and above 0. Consistent with our intuition. Implicitly. The nested structure allows us a voter first to choose whether to vote or not. so drawing a firm conclusion about the effects of opponent’s advertising is not possible. However. This limited variation leads us to find no significant unobserved We also estimated a version of the BLP model using a nested logit with only the own candidate’s advertising.

Ackerberg (2001) finds an elasticity of 0. Although we might at first suspect political ads to be of greater importance. In the case of new products. 4. the estimated elasticities are roughly 0. Few studies report advertising elasticities.050 for Democrats. Ch.027 for Democrats.025 for the Republicans and 0.15. experience an intense advertising campaign at launch. seeing that people are more wedded to a political candidate than to a yogurt product is unsurprising.heterogeneity once sufficient controls are in place. 8). With only the own advertising. These sensitivities are actually rather small own advertising elasticities if we compare them to consumer packaged goods. Kadiyali. Vilcassim. Parsons. The third-party candidate is about twice as sensitive to changes in the Republican’s advertising than to the Democrat’s advertising.1. Gerber (2004) compares the estimated effects of campaign spending (as opposed to specifically television 23 .03 using GRP data for a personal care product. Comparing our estimates of advertising elasticity to previous work in the political science literature is difficult. For instance.2 Elasticity Estimates Table 5 presents the elasticity estimates from specifications (6) and (7) using all the control variables to compare the results when only one advertising variable is included to when the major opposing party’s advertising is also included.045 for Republican and 0. Including the opponent’s advertising raises the elasticities to about 0. which. and Schultz (2001. and Chintagunta (1999) estimate an advertising elasticity of about 0. and an order of magnitude smaller for the third-party candidate. where advertising elasticities are roughly 0. and many use stated preference and intentions data as dependent variables instead of actual voting outcomes. The effectiveness of advertising is likely to vary significantly depending on the product category. similar to elections. as reported in Hanssens.

please refer to pages 969 and 975 of Huber and Arceneaux (2007). (2011) ran a field experiment in which different DMAs randomly receives varying levels of television and radio advertisements. (2011) permit some comparison because both use GRPs of television advertising and focus on large. The studies by Huber and Arceneaux (2007) and Gerber et al.000 GRPs yields on average 2 percent more votes in 2000 and 1. respectively.7 percent more votes in 2004. we calculate the percentage change in votes for a candidate given an increase of 1.76 percent. An extra 1. We exclude from this calculation those counties that received zero GRPs.000 GRPs increases the probability a voter supports Bush by 1. but 19 For details.advertising) across several prominent studies and shows their predictions vary by more than an order of magnitude depending on the estimation technique. To help make our elasticity estimates comparable with these results. Huber and Arceneaux (2007) use cross-sectional survey data to show that increasing Bush’s advertising by 1. the same increase in GRPs generates an extra 2. 24 . For Bush. Using a regression and actual voting data. the authors find a robust effect that 1. both studies address the endogeneity of advertising using a natural experiment in the 2000 presidential election and a field experiment in the 2006 Texas gubernatorial election. For the Democratic candidates.72 percent. an increase of 1.19 Gerber et al.2 percent vote share. Importantly. A polling firm then conducted telephone surveys to voters in each market to gather information on candidate preferences and voting intentions.000 GRPs represents a non-trivial amount of advertising dollars. the authors estimate that 1. Restricting attention to non-battleground states. while increasing Gore’s advertising by the same amount increases his support by 3. Combining the survey data and the experiment.000 GRPs increases Bush’s proportion of the two-party vote by 4 percent. recent elections.000 GRPs.000 television GRPs increases respondents’ intentions to vote for that candidate by about 5 percent.

without television advertising. we consider the power of advertising as an influential variable in the competitive interactions between candidates. Thus. It is also unclear whether the effectiveness of advertising will be the same across elections for different offices. For example. holding all other factors fixed. The column “Switched States” lists states that switched to the candidate listed in that row. To assess the potential of advertising to shift election outcomes. Table 6 below indicates how the electoral votes would have changed under this scenario. 4. Note that we do not consider this exercise to be an actual prediction of the election outcome if advertising were banned because many other variables would endogenously respond. the exercise still gives us some idea of the rough preferences of voters were it not for the influence of advertising and of the sensitivity of the election outcome. Although the counterfactual is unable to address these issues. The estimates in Table 4 demonstrate that advertising can influence choice. but also that one candidate’s advertising might partially cancel out the effects of another candidate’s advertising. This comparison comes with several caveats because our study differs from the other two. or perhaps candidates would even be forced to change some of their policy stances. we therefore consider how the electoral votes would have changed if all advertising were set to zero.the predicted response could potentially have a significant effect in a state’s voting outcome. 25 . our estimates seem roughly consistent with two other careful studies that exploit (quasi-) experimental variation to isolate the causal effects of advertising.3 Zero Advertising Counterfactual In this section. candidates might use other means to communicate with voters about their stances on policy issues. especially in our use of instrumental variables to measure advertising efficacy with actual voting data.

26 . tipping the election outcome in Gore’s favor. Aside from the political implications. demonstrate the sensitivity of the election outcome to changes in advertising. Bush and Gore each spent under $3 million on advertising across the two states that switched. an amount that represents less than 5% of each candidate’s advertising budget. Bush wins in either specification. With zero advertising in 2004. These results. the occurrence of elections only in even years allows lagged advertising prices to serve as instruments since they are independent of any political election influence. with no change in the upper panel and gains New Hampshire (NH) and Wisconsin (WI) in the lower panel. this new scenario represents a net loss of 18 electoral votes to Gore. The use of advertising prices as instruments is well-motivated. analyzing advertising in elections frees the researcher from many of the dynamic concerns in which advertising is invested over long periods of time. Finally. Furthermore. Including the opposing candidate’s advertising leads Gore to also win New Hampshire (NH) and gain an extra four electoral votes. 5 Conclusions This paper documents a robust positive effect of advertising in the case of general elections for the US president. whereas the results for branded goods often find no effects with experimental variation.With zero advertising in 2000. in particular those for the 2000 election. In 2000. We believe this estimation strategy allows us to find robust positive effects of advertising. based on the underlying structure of the candidate’s decision process. For Bush. and effects persist across multiple choice occasions. the application is well suited to estimating the effect of advertising in general. The findings illustrate that advertising is capable of shifting the electoral votes of multiple states and consequently the outcome of elections. Bush would have won Oregon (OR) and Gore would have picked up Florida (FL).

and to analyze them from a causal perspective requires an instrument that shifts the relative balance of positive and negative advertisements across DMAs. Ansolabehere et al. and how this might alter their decision to vote for a given candidate. 27 .g. We made this choice because numerous papers in political science specifically examine positive versus negative advertising (e. Third. we do not model voters’ expectations about the potential outcome of the election.. This is beyond the scope of available data because one would need to observe a sample with substantial variation in such positions.Our analysis comes with several caveats that might be interesting to pursue as future research. Expectations data have well-known challenges and would require additional structure that we felt would impose too stringent parametric restrictions while trying to focus on the causal relationship in the data.. First. we aggregate all of a candidate’s advertising into a single variable and do not separately consider the effects of positive or negative advertisements. our model takes the policy positions of candidates fixed and does not consider any interaction between the advertising content and their positions. Second. We hope our illustration of the value of fixed effects and instruments motivated by advertiser objective functions influences future empirical studies of these political applications as well as other questions in advertising more broadly. 1999).

P. Manchanda (2005). (2001). Simon (1999). J.” RAND Journal of Economics. and J. 3. S. [6] Che. [4] Ashworth. A. Klapper (2006). “Choice Set Heterogeneity and the Role of Advertising: An Analysis with Micro and Macro Data. “Goodwill and Dynamic Advertising Strategies.” Marketing Science. M. P. “Negative Advertising and Voter Choice.” Harvard Business School. Conlin. and R. [11] Draganska. T. 02-099. D. [3] Ansolabehere. T. 841-890. 92(3-4).. M. and A. U. Doraszelski. [2] Anand. A.” Working paper. “The Swing Voter’s Curse. 408-424. M. D.” Journal of Public Economics. Pesendorfer (1996). 4(1). [8] Clark. 32(2). G.” Quantitative Marketing and Economics. 27-41. Rao (1989). University of Pennsylvania. and P. [15] Eastlack. [7] Doganoglu. O. 107-144. H. B. Berkeley. J.. J.” Working paper. “Modeling Response to Advertising and Pricing Changes for V-8 Cocktail Vegetable Juice. Stanford University GSB. 28 . [10] Degan. Su (2009). 86(3).” American Economic Review. 207-236. University of Chicago.. [12] Dub´.” Quantitative Marketing and Economics. Pakes (1995). and D. Clinton (2006). and Draganska. “Risk Aversion and Apparently Persuasive Advertising. 63(4).References [1] Ackerberg. and A. [16] Feddersen.” Working paper. 2. “An Empirical Model of Advertising e Dynamics. No.. Manchanda (2005). 5-29. and R. University of California. (2009). and P. R. [13] Dub´. S. S.. “Empirically Distinguishing Informative and Prestige Effects of Advertising. Fox. J. 316-333. “Automobile Prices in Market Equilibrium.. and W. 81-95.D. “Replicating Experiments Using Aggregate and Survey Data: The Case of Negative Advertising and Turn Out. “The Effect of Advertising on Brand Awareness and Perceived Quality: An Empirical Investigation Using Panel Data.” Quantitative Marketing and Economics. Moro (2008).. 245-259. D. and A. Shanmugam (2007).” Working paper. and A. [5] Berry. (2010). “Differences in Dynamic Brand Competition Across e Markets: An Empirical Analysis.” Marketing Science. 5(3). Shachar (2002). “Does Advertising Exposure Affect Turn Out?” Quarterly Journal of Political Science. C. “The Performance of Pivotal-Voter Models in Small-Scale Elections: Evidence from Texas Liquor Referenda. Iyer. P. S. “A Structural Model of turn out and Voting in Multiple Elections. J. S. [14] Dub´.” Econometrica. “Improving the Numerical Performance of BLP Static e and Dynamic Discrete Choice Random Coefficients Demand Estimation. Working Paper Series. [9] Coate. 7(2). Iyengar. L. and Klapper. 582-596. B. G. T. Hitsch. J. Levinsohn. Merlo (2009). 24(1).” The American Political Science Review. 93(4) 901-909.

” Journal of Econometrics. and K. A. “Structural Equilibrium Analysis of Political Advertising. S. Columbia Business School. 957-977. L. J. Shaw (2011). P. [19] Gerber. 50-68. 119-29. Kluwer Academic Publishers. A. Turn Out. Behavior. and L. Krause (2007). and S. G. A. Goldstein (1999). L. “Measuring Media Exposure and the Effects of Negative Campaign Ads. 92(2). “How Large and Long-lasting Are the Persuasive Effects of Televised Campaign Ads? Results from a Randomized Field Experiment. G. Chintagunta (1999). [20] Gerber. “Estimating the Effect of Campaign Spending on Senate Election Outcomes Using Instrumental Variables.” Journal of Marketing Research. M.” American Journal of Political Science. “The Effects of Campaign Spending in Congressional Elections. and Voting in U. Parsons. Urminsky (2010). S. [24] Gordon. 2nd Edition. and W. G. (2001). G. Presidential Elections. D. P. 72. and J. D. [33] Krishnamurthi. Weiss (1992). 22. C. Cambridge. Hansford. “Salvation for the Spendthrift Incumbent: Reestimating the Effects of Campaign Spending in House Elections.” Marketing Science. 135-150.” Journal of Politics. and K. [27] Hillygus. R. “The Effect of Advertising on Consumer Price Sensitivity. C. 359-71.. V. P. M. [22] Gomez.R. “Does Campaign Spending Work?: Field Experiments Provide Evidence and Suggest New Theory. O. “Price Sensitivity and Television Advertising Exposures: Some Empirical Findings. Lovett. B. [31] Kadiyali.” American Journal of Political Science. L.. Soberman. 401. (1998). Market Response Models: Econometric and Time Series Analysis.” American Political Science Review. [23] Gordon. B. P. A.” Journal of Politics. (2005). Hagen. K. D.. [21] Gerber. (2004). L. [18] Garrett. Weinberg and D. Raj (1985). 884-907. P. “The Dynamics of turn out Intention in Election 2000. R. “Identifying the Persuasive Effects of Presidential Advertising. Rao.” working paper.. 541-574. Order Code RS22644. “The Republicans Should Pray for Rain: Weather. and D. 32(4). and Schultz. 1189-1208. 105(1). D. N.” working paper. [30] Johnston. [32] Kanetkar. H. “Coordinated Party Expenditures in Federal Elections: An Overview. R. R.. 51(4). Green.. A. [26] Hanssens. 339-363. 11(4). G. Vilcassim. and K. (1978). 89.S.” American Political Science Review.. and Policy Implications. 469-491. Whitaker (2007). Arceneaux (2007). The Library of Congress. M. S. B. [29] Jacobson. Krasno (1988). V. “Marketing and Politics: Models. S. Hartmann (2010). S. R. Peress. Arceneaux. Moorthy. B.” American Political Science Review. 47(5). P. [25] Green. Sen. and G.[17] Freedman. “Product line extensions and competitive market interactions: An empirical analysis. 69(3).. 29 . R.” American Journal of Political Science. R. 67(1). T. Shachar. (2004).” American Behavioral Scientist.. Jamieson. T. D. Gimpel. 649-663.” Congressional Research Service. [28] Huber. 43(4). UK: Cambridge University Press. J. The 2000 Presidential Election and the Foundations of Party Politics.

Lehmann (1997). Marketing Science. University of Rochester. Bernhardt. [48] Su. (1994). [49] Wattenberg. Advertising Experiments. “How T. T.” Journal of Marketing Research. K.” Working paper. E. M. S. 102(4). Kalmenson. S. Chicago: AMA. G. S. 48(1). “Constrained Optimization Approaches to Estimation of Structural Models. “From Generic to Branded: A Model of Spillover Dynamics in Paid Search Advertising. C. Advertising Works: A Meta-Analysis of 389 Real World Split Cable T. 46(6). Mantrala.P. 125-139. 32(2). “The Long-Term Impact of Promotion and Advertising on Consumer Brand Choice. “Negative Campaign Advertising: Demobilizer or Mobilizer.” Economica. Marketing 1776-1976 and Beyond: 1976 Educators’ Proceedings. [37] Mela.” Journal of Marketing Research. L.” in Kenneth L. D. 798-815. R. J. House. 214-235.” Public Choice.. Gupta and D. S. 424-441. Bucklin (2011). and A. B. M and M. Winer (1976). Lubetkin. L. M. and K. F. 357-377. S. “The Impact of Campaign Spending on Votes in Multiparty Elections. Richardson.” Journal of Marketing Research. 777-798.” Journal of Political Economy. (Ed. Abraham. P. 30 . “Optimal Advertising Policy under Dynamic Conditions.V. (2009). (1974).” Working paper. 29(114). “Targeting Political Advertising on Television. [47] Staelin. and C.” Journal of Political Economy. [39] Narayanan. Peress (2010). [36] Lovett. (2007). 28(3). 140. “An Unobservable Model for Determining the Effect of Advertising on Consumer Purchases. M.” Journal of Marketing Research. “Heterogeneous Learning and the Targeting of Marketing Communication for New Products”. P.” Review of Economics and Statistics. C. “Planning Media Schedules in the Presence of Dynamic Advertising Quality. 891-899. Manchanda (2009). R. and Randolph E. Stevens (1995). Brians (1999).” Working paper. 82. [40] Nelson. David and Randall A Lewis (2009). 129-142. R. 93(4). 729-753. [41] Nerlove. 87-102. [35] Lodish. University of Chicago..L. Sawyer (1998).” Marketing Science.. Arrow (1962). [43] Riley. “The Political Participation Puzzle and Marketing. 89(3). [45] Shachar. 573-585. 34(2). [44] Rutz. Oliver J. “Using Repeat Challengers to Estimate the Effect of Campaign Spending on Election Outcomes in the U. and P. B. [38] Naik. Marc and Kenneth J.V. Livelsberger. [46] Stratmann.).” American Political Science Review. Yahoo! Research. 17(3). [42] Rekkas.S. “Advertising as Information. “Does Retail Advertising Work? Measuring the Effects of Advertising on Sales via a Controlled Experiment on Yahoo. “How Prices Matter in Politics: The Returns to Campaign Advertising. and M. (2009). 248-261. M.[34] Levitt. and R. Judd (2010).

285 95 1.057 0.081 0.047 0.890 0 17.285 2.386.410 0 9.944 20.675 75 116 879.856.809 10.221 0 12 0 8.726 681.349.607 0 322.472 63.Table 1: Summary Statistics of County-Level Voting by Candidate and Election Year Obs Mean Std Dev Min Max 2000 Election Votes Bush 1596 23.225 89.907.179 1.520 Share Bush 1596 0.185.445 0 5.229 51.755 1.009 44.782 5.039 0.569 Table 2: Market-Level Advertising by Candidate and Election Year Obs 2000 Election GRPs Bush GRPs Gore GRPs Other Expenditures Bush Expenditures Gore Expenditures Other 2004 Election GRPs Bush GRPs Kerry GRPs Other Expenditures Bush Expenditures Kerry Expenditures Other 75 75 75 75 75 75 75 75 75 75 75 75 Mean Change Std Dev Min Max 5.505 0.520 0 14.207.439 9.593 4.056 0.569 Votes Gore 1596 25.823 5.064 0.089 0.341 Share Kerry 1596 0.123.940 0 423 0 6.879 0 35.268 0 15.710.666 0.315 216 1.863.979 0 46.678 7.165 31 .698 Share Bush 1596 0.807 3 3 1.842 1.630 0.214 2004 Election Votes Bush 1596 29.941.526 1.168 Votes Kerry 1596 29.930 78.086 0.726 12.433 1.142 77 1.324 2.889 210 871.736 0.076.218.297 Share Gore 1596 0.072.

56 1.26 5.07 1.21 0.Table 3: Summary Statistics of the change in the Lag CPMs from 1999 to 2003 Daypart Early Morning Daytime Early Fringe Early News Prime Access Prime Time Late News Late Fringe Obs 75 75 75 75 75 75 75 75 Mean Std Dev 1.30 0.82 3.02 0.11 10.85 -2.06 -3.04 1.72 1.98 Min Max -1.21 32 .24 4.88 2.10 3.50 3.38 -1.75 9.06 -2.40 -1.05 2.18 1.88 1.80 2.75 -1.48 11.74 -2.47 6.65 1.98 2.

108** (0.111 (6.045) -0.036 (0.276 (0.585) 1. “Gubernatorial Election” for the existence of a Gubernatorial race in the same state.565 (3.053*** (0.071*** (0.022 (0.070) (3) OLS 0.077*** (0.473*** (0.921) 0.122 (6. ‘**’ significance at α = 0.023 (0.075*** (0.055 (0.077*** (0.546) -0. Robust standard errors clustered at the DMA level appear in parentheses.032) -0.720) 1.101*** (0.358) -0.008) 0. 33 . “Senate Election” indicates the inclusion of fixed effects for the existence of a Senate race.028 (0.357*** (0.059 (0.039 (0.920) 0.500) -0.039) 5.847 (3.549) 2. “% Republican x Democrats”.003 (0.004) Y Y Y 0.350** (0. represents the % of voters who identify as Republican interacted with the Democratic candidate’s intercept.086 (6.664) 0. Party-Year indicates the inclusion of party-year dummy variables.028 0.019) (6) 2SLS 0.032) -0.267 (0.008) 0.910) 0.021) (2) 2SLS 0.704) 1.040) -0.083** (0.911) 0.566) 2.573 (0.128 (0.081 (0.356*** (0.058 (0.535) 0. for example.513** (0.540) -0.020) (5) 2SLS 0.048) 0.004) Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Y Notes: Obs = 9.569 (3.008) (4) 2SLS 0.048) 0.050*** (0.950) 0.039) 5.019) (7) 2SLS 0.612) -0. 576.621) 2.05 and ‘***’ significance at α = 0.045 -0.1083** (0.160** (0.110) -0.035) 0.502) -0.050 (0.Table 4: Parameter Estimates (1) OLS Candidate’s Ads Opponent’s Ads % Republican x Republican % Republican x Democrats % Republican x 3rd Party % Democrat x Republicans % Democrat x Democrats % Democrat x 3rd Party Senate Election Gubernatorial Election Rain (inches) Snow (inches) % 45 ≤ Age < 64 % 65 < Age % Unemployment Distance*100 Fixed Effects Party Year-Party DMA-Party -0.572 (0. DMA-Party indicates fixed effects for each DMA and Party.054 (0.003 (0.126 (0.332) -0.540) 0.01.

01 level.0028 -0.0001 -0. using either the specification in (6) or (7) in Table 4. The right-most column indicates which states switched hands.0269 -0.0114 0. whereas in the lower panel the same change implies a decrease of 0.0088 -0.0245 -0.0088% decrease in the market share of the Republican candidate. For example. Table 6: Zero Advertising Counterfactual Election Year Own Ads 2000 2000 2004 2004 Both Ads 2000 2000 2004 2004 Electoral Votes Observed Zero Ad Switched States (Electoral Votes) 271 267 286 252 271 267 286 252 253 285 286 252 249 289 300 238 OR(7) FL(25) — — OR(7) FL(25). WI(10) — Candidate Bush Gore Bush Kerry Bush Gore Bush Kerry Notes: Electoral college results from setting advertising to zero in both elections. using the own advertising specification in the 2000 election.0119 -0.0063 -0.0001 0. in the upper panel.0305 0. in Table 4.0501 -0.0305% for the Republican. respectively.0001 -0.0088 0. For example.0114 -0. and the “Zero Ad” column presents the predicted number after setting advertising to zero everywhere. 34 . NH(4) NH(4). a 1% increase in Democrat advertising implies a 0.0448 -0.0044 Notes: All estimates are significant at the α = 0.0363 -0. Bush picks up Oregon but loses Florida to Gore.0001 0. The “Observed” column presents the actual number of electoral votes received by each candidate. Upper and lower panels report estimates from specifications (6) and (7).Table 5: Elasticity Estimates Republican Democrat 3rd Party Own Ads Republican Democrat 3rd Party Both Ads Republican Democrat 3rd Party 0.

000 ‐10.000 0 10.000 30.000 30.000 ‐20.000 40.000 ‐10.000 Demo ocratic GRP Change from 2000 to 2004 20.000 20.000 Republican GRPs Change from 2000 to 2004 35 .Figure 1: 2000 to 2004 within-DMA Changes in GRPs 40.000 0 ‐20.000 10.

7 0.5 prime_time prime_access 0.9 0.1 0 36 .Figure 2: Daypart Mix for Democrats in 2000: Top 10 DMAs in GRPs 1 0.6 late_fringe late_news 0.3 daytime early_morning 0.2 0.8 0.4 early_news early_fringe 0.

6 late_fringe late_news prime_time 0.7 0.3 early_morning 0.2 0.8 0.9 0.5 prime_access early_news 0.1 0 GREEN BAY MIAMI DETROIT DES MOINES MILWAUKEE SPOKANE HARRISBURG WILKES  BARRE LITTLE ROCK PORTLAND 37 .4 early_fringe daytime 0.Figure 3: Daypart Mix for Republicans in 2000: Top 10 DMAs in GRPs 1 0.

2004 25 20 15 CPM in 2 2004 CPM 10 45 degree 5 0 0 5 10 15 20 25 CPM in 2000 38 .Figure 4: Early News CPM by DMA: 2000 vs.