The Air War versus the Ground Game:
An Analysis of Multi-Channel Marketing in U.S. Presidential Elections

Lingling Zhang1, Doug J. Chung2 1
1

Robert H. Smith School of Business, University of Maryland, College Park, MD 20742,
lingzhang@rhsmith.umd.edu;

2

Harvard Business School, Harvard University, Boston, MA 02163, dchung@hbs.edu

Abstract
This study jointly examines the effects of television advertising and field operations in U.S.
presidential elections, with the former referred to as the “air war” and the latter as the “ground game.”
Specifically, the study focuses on how different campaign activities—personal selling in the form of field
operations and mass media advertising by the candidates and by outside sources—vary in their
effectiveness with voters who have different political predispositions. The voting choice model takes into
consideration voter heterogeneity and analyzes comprehensive data that include voting outcomes,
detailed campaign activities, and voters’ party affiliation for three presidential elections (2004-2012). The
results reveal that different campaign activities have heterogeneous effects depending on voters’ party
affiliation. Field operations and political advertising from outside groups are more effective with partisans,
while a candidate’s advertising is more effective with non-partisans. These findings can help strategists
better allocate resources across and within channels to design an effective political marketing campaign.

Keywords: multi-channel marketing, ground campaigning, advertising, personal selling, political
campaigns, consumer heterogeneity.

1

The authors thank the senior editor, associate editor, and two anonymous reviewers for their feedback during the review
process. The authors also thank the following individuals for their comments and suggestions: Sunil Gupta, Vineet Kumar,
Roland Rust, and the seminar participants at Harvard University, the Korean Advanced Institute of Science and
Technology, Temple University, Texas Christian University, Yonsei University, the University of Maryland, the University
of Texas at Dallas, and the University of Wisconsin, in addition to the conference participants at the 2016 INFORMS
Marketing Science Conference.

 1 Introduction
Multi-channel marketing has increasingly become a critical competitive strategy for a firm’s success. In
particular, mass-media advertising and personal selling are among the most important instruments at a
firm’s disposal. Advertising has the advantage of reaching a large audience via well-scripted
communications. Its importance goes without saying, given that global advertising spending was around
$563 billion in 2019.1 Personal selling, on the other hand, occurs at a more micro and personal level and
takes the form of in-person visits, distribution of fliers, and telemarketing, to name just a few. Like
advertising, personal selling is of great importance to many businesses. In the United States alone, the
total spending on a company’s sales force is about three times the total spending on advertising (Zoltners
et al. 2006), with approximately 10% of the nation's labor force directly involved in sales or sales-related
activities.2 As advertising and personal selling are foremost in the minds of marketers, it is essential to
understand the relative effect of these channels, as well as the dynamics between them.
This study examines the effects of multi-channel marketing—mass media advertising and personal
selling—in the context of U.S. presidential elections. Choosing the right product (the “president”) every
four years is, perhaps, among the most critical decisions that U.S. consumers (the “voters”) face. The
amount of marketing behind each campaign is massive: the Democrats and Republicans spent roughly
$2 billion on presidential campaigns in the 2012 election alone, making it one of the most expensive in
U.S. history. 3 In recent years, presidential campaigns have increasingly employed a multi-channel
strategy. One notable phenomenon is the large-scale personal selling effort in the form of ground
campaigning, also known as field operations or grassroots campaigning. Former President Barack H.
Obama deployed unprecedented field operations in 2008, such that many attributed his election success
to these on-the-ground efforts: “. . . Obama's effective organization [of the field teams] . . . could be a
harbinger for how successful elections are won in battlegrounds in years to come.” 4 However, this
speculation remains untested. Were Obama’s ground campaigns as pivotal as the press stated? If so,
which type of voters were most affected by different types of campaigns? In terms of voting behavior,
people are likely to have heterogeneous preferences at different stages of their decision-making process
(DMP). Some may have already established a political predisposition towards a certain candidate and
need only a nudge to turn out to cast a ballot, whereas others may remain undecided and continue to
1
2
3
4

Statistica, https://www.statista.com/statistics/236943/global-advertising-spending, accessed on April 2, 2020.
U.S. Bureau of Labor Statistics, May 2019, https://www.bls.gov/news.release/pdf/ocwage.pdf, accessed on April 2, 2020.
The New York Times, “The Money Race”, 2012.
Sherry Allison, “Ground Game Licked G.O.P.” The Denver Post, November 5, 2008.

1

 explore their preferences. Hence, how do field operations and advertising persuade and/or mobilize
individuals with different political predispositions?
Researchers have long been interested in the effect of campaigning in political elections. As better
data and empirical models have become available, there has been an increase in empirical research
assessing the effect of campaign advertising (Gordon and Hartmann 2013, 2016; Huber and Arceneaux
2007; Klein and Ahluwalia 2005; Krasno and Green 2008; Lovett and Shachar 2011; Lovett and Peress
2015; Shaw 1999; Wang, Lewis, and Schweidel 2018) and field operations (Darr and Levendusky 2014;
Masket 2009). While these studies have generated insights into political campaigns, two issues remain.
First, extant research has focused largely on just one campaign activity, which does not allow for a
comparison of multiple marketing instruments within the same context (e.g., U.S. presidential elections).
Second, and perhaps more importantly, existing research has yet to provide sufficient insights into the
heterogeneous effect of campaign activities. With more technologies to better target individual voters,
the question of how different campaign activities work in different voter segments has become
increasingly important. Hence, this study seeks to generate more-granular insights into the heterogeneous
effects of multiple political campaign activities—field operations and candidate and outside advertising—
on different types of voters.
In U.S. presidential elections, campaign staff and volunteers carry out on-the-ground campaigning,
which includes conducting door-to-door canvassing, making phone calls, sending mailings, and delivering
door hangers with candidate signs. Although these outreach activities may serve the purposes of both
persuading voters and inducing them to turn out, ground campaigns tend to focus more on get-out-thevote (GOTV) efforts (e.g., Gerber and Green 2000a, 2000b; Green et al. 2003; Nickerson et al. 2006;
Wielhouwer 2003). Since these “retailing” efforts involve personal contact, they give the candidate an
opportunity to directly nudge voters towards turning out to vote. In contrast, the classic “wholesale”
campaigning method of television advertising sends broad messages to express a candidate’s policies or
to attack his or her rivals. Research on campaign advertising has found that it has a limited effect on
turnout (Krasno and Green 2008) but a positive effect on establishing preferences and persuading voters
to see candidates more or less favorably (Huber and Arceneaux 2007; Lovett and Peress 2015).
Establishing preferences and deciding to turn out likely occur at different stages of an electorate’s
DMP. Naturally, turnout—the choice between going and not going to the polls—is determined at a later
stage of the DMP, after voters have established their candidate preferences. Voters who have already
formed a political predisposition for a particular candidate would be more likely to identify with more2

 targeted turnout messages from that candidate, whereas voters who have not yet decided on their
preferred candidate would be more receptive to broad messages that provide information about the
candidates. Following these arguments, one would expect that different presidential campaign
activities—field operations and television advertising—would have varying effects on different types of
voters, depending on where they are in their voting DMP. Field campaigning in modern elections focuses
on partisan turnout operations, and, thus, this approach is expected to have a stronger effect among
voters with an established political predisposition. In contrast, television advertising, being less targeted
but more informational in nature, is expected to be better received by those who are yet undecided and,
thus, might still change their minds.
Campaign advertising in the extant literature typically corresponds to advertisements sponsored by
the candidates and their party committees5 (Huber and Arceneaux 2007); however, recent presidential
elections have seen rapid growth in advertisements sponsored by outside political groups, known as
Political Action Committees (PACs or Super PACs). 6 Existing research has found that the two types of
campaign advertisements have different average effects in U.S. senatorial elections (Wang et al. 2018).
We further hypothesize that they could have heterogeneous effects on different voter segments.
This study’s theoretical arguments originate from the well-established Receive-Accept-Sample (RAS)
model in political science (Zaller 1992). According to the RAS model, voters form political preferences
through a two-stage process: receive and accept. The model states that people with higher political
awareness (e.g., those with an established predisposition) are more likely to receive campaign information
but are less likely to accept messages that are inconsistent with their predisposition. In contrast, people
with moderate political awareness (typically, those who are unaffiliated with any major party) are less
likely to receive political messages but are more open to persuasion and, thus, are more likely to accept
and be influenced by such information. This process holds for traditional candidate advertising (Huber
and Arceneaux 2007), but may apply less to PAC advertising because the content of the latter tends to

5

Hereafter, we refer to advertising sponsored by a candidate and his or her affiliated party committee as candidate
advertising.
6
For ease of exposition, we refer to outside political groups as PACs. Unlike traditional PACs, Super PACs can raise an
unlimited amount of money for an election, but the spending has to be independent of any candidate or party. Hereafter,
we use the terms PAC advertising and outside advertising interchangeably to refer to advertising that is not sponsored
by the candidate or his or her affiliated party.

3

 be far more negative and have an attacking tone rather than a promoting tone.7 Moderately aware people
would be less likely to receive and accept the PAC messages because they tend to have a lower tolerance
for negativity than do those with higher political awareness (Fridkin and Kenney 2011). Therefore, the
net effect of PAC advertising on moderate voters remains an empirical question.
Based on existing theoretical and empirical findings, voter segments are hypothesized to moderate
the effects of campaign activities. To examine this, we compile a unique and comprehensive dataset
integrating multiple sources. The ground-campaigning data include detailed records of field operations
from the Democratic and Republican candidates. The data on television advertising include ad
impressions at the designated-market-area (DMA) level for advertisements made by the candidates and
their party committees, as well as those by outside political groups. For voter segments, this study
focuses on voters’ party affiliation: those who have affiliated themselves with a party tend to have a
stronger political predisposition and awareness (Holbrook and McClurg 2005; Iyengar and Hahn 2009).
Thus, the partisan versus unaffiliated segmentation matches the voter segmentation proposed in the
RAS theoretical framework.
To model individual voting preferences, we specify a random-coefficient aggregate discrete-choice
model, which estimates the interaction between campaign effects and voter characteristics using
aggregated data. However, estimating the true effects is nontrivial because campaign resources are
strategically allocated, causing an endogeneity problem. To address this challenge, the model incorporates
an extensive set of county-party fixed effects and uses advertising cost shifters as instrumental variables.
The results show that campaign effects significantly depend on voters' affiliated partisanship. In
particular, field operations are more effective with partisans than with non-partisans, while candidate
advertising is more effective with non-partisans. PAC advertising, which typically consists of negative
and attacking messages, is more effective with partisan voters. It is novel to see that the two types of
advertising have different effects on different voter segments.
This study makes three contributions to the literature on political campaigns. First, we collect
detailed and comprehensive data on candidate advertising, outside advertising, and ground campaigning.

7

The Wisconsin Advertising Project (WAP) coded the content of the creatives for the 2004 and 2008 presidential elections
for selected DMAs (Goldstein and Rivlin 2008). The data pattern confirms that PAC advertisements in those two elections
were far more negative (i.e., using an attacking tone rather than a promoting tone) than the candidate and party
advertisements. For example, in 2008, 33.1% of the advertisements from PACs used a negative tone compared to 2.7%
from candidates and party committees.

4

 By jointly examining these campaign activities, the study speaks to the relative importance of various
marketing instruments within a specific event (the U.S. presidential election). Second, the study identifies
and distinguishes the effect of outside advertising from that of candidate advertising. This study is among
the first few to systematically examine political advertising sponsored by outside interest groups in U.S.
presidential elections, a highly controversial subject among the general public. 8 Third, and most
importantly, the study makes inferences about the heterogeneous campaign effects to determine whom
(partisans versus non-partisans) to target with specific campaign activities (field operations versus
advertising). The results provide guidance on targeted campaign resource allocation across different voter
segments.
The remainder of the paper is organized as follows. Section 2 describes the campaign activities and
the data used for empirical analysis. Section 3 specifies the model and discusses identification. Section 4
presents the parameter estimates and the counterfactual results, and Section 5 concludes.

2 Data
The data come from multiple sources and include voting outcomes and various campaign activities for
the 2004-2012 U.S. presidential elections. Compared with those used in other studies, the data for this
study are more comprehensive in at least four aspects. First, the collection of multiple campaign activities
encompasses a more complete record of advertising and ground campaigning than seen in previous studies.
Knowing where and to what extent candidates chose to campaign enables us to assess the effects of
various campaign activities after controlling for each other. Second, the unit of analysis in this study is
at the county level, which is as granular as it can be to reliably obtain voting outcomes. In addition, the
measure of campaign activities is also at the most granular level possible. Hence, the disaggregated data
allow us to take a finer look at campaign allocation, curtailing potential aggregation bias. Third, the
data include registered party affiliation at the county level, making it possible to examine how campaign
effects differ according to voter partisanship. Lastly, the data span a period of three elections, and, thus,
the results are not confined to a particular combination of candidates.

8

Public interest in PAC advertising remained high in the 2016 presidential election, when more than 100 organizations
contributed to election advertising (https://www.opensecrets.org/pres16/outside-groups). The role of PAC advertising
also received considerable attention and discussion; e.g., https://www.bloomberg.com/politics/graphics/2016presidential-campaign-fundraising/, https://www.harvardmagazine.com/2016/07/does-money-matter.

5

 2.1 Election Votes
The outcome variable of this study is the number of votes cast for the presidential candidates in each
county. These data come from the CQ Press Voting and Elections Collection, a database that tracks
major U.S. political elections.
A county constitutes a “market” in which residents choose up to one “product” (candidate).9 A
county's “market size” equals the total number of resident citizens aged 18 and older, typically known
as the voting-age population (VAP).10 The county-level age-specific population counts come from the
U.S. census database. The “market share” or “vote share” of each candidate is the percentage of votes
that he or she receives out of the county’s VAP. Table 1 lists the county-level summary statistics for
vote outcomes.
< Table 1>
2.2 Ground Campaigning
To an average voter, presidential elections are likely most visible on the ground level through personal
selling activities. Candidates establish field operations to organize ground-level voter outreach (hereafter,
ground campaigning and field operations are used interchangeably). As field offices are the backbone of
outreach activities, the number of field offices deployed in each county is the measure used to quantify
the scale of a candidate's field operations. The 2004 and 2008 field office data come from the “Democracy
in Action” project at George Washington University, and the 2012 data come from Newsweek’s Daily
Beast website.11
Table 2 displays the summary statistics for field operations. Across all elections, the Democratic
candidates had an indisputable lead in establishing field operations: the ratios of the Democratic and
Republican field offices were 3.51, 3.53, and 2.69 in 2004, 2008, and 2012, respectively. Furthermore,
even between the Democratic candidates, field offices were more prominent in the Obama campaign than
in the John Kerry campaign: the latter set up field offices in 237 (8%) counties and the former in 624
(20%) and 439 (14%) counties in 2008 and 2012, respectively. 12

9

Hereafter, we use “market” and “county” interchangeably. There are a total of 3,144 counties and county equivalents in
the United States. We exclude Alaska and Louisiana from the analysis due to data issues.
10
Perhaps a better measure of the market size of a county is the voting-eligible population (VEP), which equals the VAP
minus ineligible felons. This metric, however, is available only at the state level.
11
We use the Geographic Information System (GIS) software to map the office addresses onto the corresponding county.
12
Field operations remained a major campaign activity in the 2016 presidential election. Clinton had more than twice as
many field offices as Trump nationwide (489 versus 207) (https://fivethirtyeight.com/features/trump-clinton-field-offices).
However, the ratio between the parties was 2.36, which is smaller than the ratio of 2.69 in the 2012 election.

6

 < Table 2 >
It merits mentioning that the number of field offices in each county is a proxy for voters’ exposure
to field operations. This metric, which has become available only in recent elections, provides a more
objective measure of field operations at a granular geographical level than alternative survey-based
measures, which are prone to recall errors and non-response bias. Due to this advantage, several recent
studies have used the number of field offices (Darr and Levendusky 2014; Masket 2009). However, this
measure is not without limitations. For example, a field office may serve multiple purposes, including
coordinating voter contacts, organizing fundraising events, or even laying the groundwork to raise voter
support for future party candidates. Some of these activities may not be directly related to winning the
current election (Darr and Levendusky 2014). However, the primary goal for a field office during a
presidential election should center on the “Race to 270” electorate votes. Therefore, the number of field
offices should still indicate the degree to which a candidate uses ground campaigning to gain votes.
Nevertheless, a question remains. To what extent does the number of field offices reflect the level of
voter exposure to ground campaigning? One way to assess the validity of this metric is to correlate it
with the number of voter contacts made by the ground campaign personnel. The American National
Election Studies (ANES), a large-scale survey on voting and political participation, provides survey
responses regarding voter contacts. In the 2004, 2008, and 2012 ANES time-series surveys, respondents
were asked whether they had been contacted by a party about the campaign and, if yes, by which party.
The correlation between voter contacts (in the ANES responses) and the number of field offices was 0.76
for the Democrats and 0.73 for the Republicans.13 Hence, despite its limitations, the number of field
offices provides the best available proxy for field operations in the scope of this study (nationwide at the
county level across the three elections). At the same time, we acknowledge the limitations and believe
that future research can benefit from improving the measure for field operations.
2.3 Television Advertising
Three types of advertising sponsors exist in U.S. presidential elections: (1) the candidates; (2) their party
committees—namely, the Democratic National Committee (DNC) and the Republican National
Committee (RNC); and (3) outside political groups. Because a candidate and his or her party committee
often coordinate advertising efforts (Wielhouwer 2003), the combined amounts of advertising from these

13

ANES responses were available only at the census-region level instead of at the county level and, thus, cannot be used
as a proxy for field operations in this study.

7

 two sources represent the candidate’s own advertising amount—hereafter, referred to as candidate
advertising.
Outside political groups (i.e., PACs), the third type of advertising sponsor, also finance independent
television advertising spots to support their preferred candidates and to attack their rivals. Although
they have played a role in presidential elections for decades, PACs have taken on much greater
prominence in recent elections, partly due to the 2002 campaign finance reform law that set stricter
restrictions on fundraising and spending for campaigns. As a result, outside political groups stepped in
to fill the gap. Especially in the 2012 election, a relatively new kind of organization, Super PACs, emerged
as a major advertiser. Super PACs, consisting of several independent PACs, support a candidate with
unlimited, and often anonymous, donations from unions, companies, or individuals. Due to the large
number of PACs that advertise in presidential elections, it is challenging to track all of their
advertisements. The advertising data used in this study include the top PAC advertisement spenders,
which represent more than 90% of the total ad spending by PACs.
In this study, the advertising measure uses gross rating points (GRPs), which quantify advertising
impressions as a percentage of the target audience reached. Using GRPs is a better measure of ad
exposure than dollar spending because the cost of advertising varies significantly across markets. For
example, the same amount of ad dollars would yield less exposure in Boston than in Kansas City. In
contrast, GRPs provide a measure of audience reach that is independent of the advertising cost.
The advertising data come from Nielsen Media Research. Nielsen divides the U.S. media market into
210 designated market areas (DMA): residents from the same DMA receive the same television
advertising. Therefore, the advertising measures are at the DMA level. It is noteworthy that the outcome
variable of interest (vote share) is at the county level, with each county belonging to one and only one
DMA. Hence, the implicit assumption is that the percent of the audience reached in a county equals the
percent of the audience reached in the DMA to which the county belongs.

It is fair to assume that the cumulative amount of advertising affects a voter’s preference on Election
Day, and, thus, the cumulative GRPs of each DMA since September 1 of that year form the advertising
measure for analysis. Table 3 presents the summary statistics for candidate advertising and PAC
advertising, respectively. The Democratic candidates had more candidate advertising than the
Republicans by 20%, 50%, and 40% in the 2004, 2008, and 2012 elections, respectively. Interestingly, the
PACs, which had less advertising than the candidates in 2004 and 2008, reached a much greater scale in
the 2012 election. In particular, PAC advertisements supporting Mitt Romney were responsible for
8

 roughly 46% of the total advertising, and there were seven times more PAC advertisements for Romney
than for Barack Obama. Although the Obama campaign had more candidate advertising than the
Romney campaign, the PACs filled the gap. In the end, the pro-Romney advertisements outnumbered
the pro-Obama ones by 25%.
< Table 3 >
2.4 Additional Variables
The data used in this study also include a large set of county-level contextual factors that may influence
a county’s voter preferences. First, the percentage of African-American residents is included to capture
the racial composition of a county. In addition, the data include a large set of variables that measure
demographics and the socio-economic conditions of a county: distribution of age, educational attainment,
population density, number of new establishments, number of newly created employment positions,
average wage, median household income, unemployment rate, Gini index, median house value, highschool dropout rate, and poverty rate. These variables come from the U.S. Bureau of Labor Statistics
and the U.S. Census Bureau databases. Furthermore, the data include the percentage of registered
partisan voters for the Democratic and the Republican parties, respectively, which is a key variable to
examine voter heterogeneity in response to various campaign activities (see Section 3.2 for details). This
variable comes from a proprietary database tracking U.S. elections. 14 Lastly, the data also contain
variables that measure the local political climate regarding a particular party (Campbell 1992): whether
it is the home state for the presidential candidate; whether the governor is from the same party; whether
the governor is an incumbent (if from the same party); and how far each county is from the campaign
headquarters. These variables vary by party within the same county.
2.5 Model-free Evidence
This section presents the model-free evidence on the campaign effect and campaign resource allocation.
2.5.1 Campaign Effects
To visually examine the campaign effects, each campaign activity is plotted against the voting outcome.
To control for the large cross-sectional variation across counties, the plots show the changes in the vote
share and campaign activities from one election to the next. Figure 1a depicts the association between

14

Because not all states require voters to declare party affiliation during registration, partisan information is available for
27 states in 2004 and 2008 and 28 states in 2012. Data come from a repository tracking of U.S. elections
(http://uselectionatlas.org/), where partisan numbers come from various official websites, such as those of the Secretary
of State and the Office of Elections.

9

 the vote share and ground campaigning. The vertical axis of the figure corresponds to the change in vote
share—i.e., scj ,t  1   scj ,t , where the vote share in county c for party j during election t is computed as
the vote counts for that party divided by the county's VAP. The horizontal axis is the difference in the
number of field offices (in logarithm), and each dot corresponds to a county-party combination. Figure
1a also includes the best-fitting fractional polynomial curve with a 95% confidence interval. The pattern
exhibits an overall positive relation, indicating that a candidate’s vote share in a county increases with
additional field offices. The positive trend tails off and turns downward at the right end, driven largely
by a few outlier counties in which the competition was intense, and the candidates added five or more
field offices. However, because of the sparse data support, one should be cautious when interpreting the
relation at the extremes.
Similarly, Figure 1b and Figure 1c show the changes in a county’s vote share against the changes in
candidate advertising and PAC advertising, respectively. The horizontal axis now corresponds to the
changes in advertising GRPs (in logarithm) for each county-party combination. The slopes for advertising
are flatter than those for field operations but still reveal a positive trend: a candidate's vote share goes
up with an increase in advertising. The trend for PAC advertising is less obvious than that for candidate
advertising, suggesting a potentially weaker effect for outside advertising.
< Figure 1 >
2.5.2 Voter Heterogeneity
A key research question of this study is: Do different voters respond to campaigns differently? To obtain
insights into this question, the association between campaigns and vote outcomes is further related to
voter partisanship, which is an essential characteristic signaling voters' political predisposition (Campbell
1992). Thus, counties are divided into two groups: high-partisan Democratic (Republican) counties in
which the percentage of registered Democratic (Republican) voters is above the mean; and low-partisan
Democratic (Republican) counties in which the corresponding percentage of partisans is below the mean.
Figure 2a depicts the relation between the vote share and field operations, with the lines fitted for low(dashed line) and high- (solid line) partisan counties, respectively. Again, for illustration, the figure shows
the scatter plot and the best-fitting fractional polynomial curve with its 95% confidence interval. While
both lines exhibit a positive trend, the solid line has a steeper slope, suggesting that ground campaigning
seems to have a stronger effect in counties with a higher percentage of partisan voters.
Interestingly, similar plots for advertising show different patterns for candidate advertising (Figure
2b) and PAC advertising (Figure 2c). Similar to the pattern shown for ground campaigning, the effect
10

 of PAC advertising is stronger in counties with a higher percentage of partisans (solid line) than in those
with a lower percentage (dashed line). The opposite pattern exists for candidate advertising: the slope is
steeper in counties with a lower percentage of partisans than in those with a higher percentage, which
seems to suggest that candidate advertising is more effective with non-partisans than with partisans.
< Figure 2 >
2.5.3 Campaign Allocation
Figures 1 and 2 provide suggestive evidence of the overall campaign effects and their heterogeneity by
voter partisanship, but such correlational observations are simply descriptive and not causal. The
observed correlation would differ from the true effect if candidates strategically allocated campaign
activities. There are at least two primary reasons that campaign deployment could be endogenous: (1)
the need to campaign where it is necessary; and (2) the need to campaign where the effects are stronger.
Regarding the first reason, the nature of U.S. presidential elections is such that candidates are
incentivized to campaign where they are not necessarily winning (or losing) but could have a chance—
i.e., the so-called “battleground” states. In contrast, states in which they are likely to win or to lose
would have only modest importance—i.e., there is a non-monotonic relationship between competition
intensity and campaign allocation. The second reason for endogeneity is that candidates may also allocate
resources where they expect them to be more effective. If they have information on which voters are
more responsive, the campaign managers could consider engaging in more campaign activities to take
advantage of the potentially higher returns on investment. Hence, we examine whether and how
candidates strategically allocate campaign activities (see Section 3.3 for the empirical strategy to address
these endogeneity concerns).
First, we examine the amount of campaigning based on competition intensity. Figure 3a shows the
Democratic field offices versus the Republican field offices for battleground and non-battleground states,
respectively.15 Because states vary in voter size, the number of field offices is normalized by the VAP

15

Typically, a battleground state (also known as a swing state or a contested state) refers to a state in which no party is
predicted to have an obvious winning margin, based on polls and election results from previous years. This study uses the
list of contested states defined by Real Clear Politics. In 2004, the contested states included Arkansas, Colorado, Florida,
Iowa, Maine, Michigan, Minnesota, Missouri, Nevada, New Hampshire, New Jersey, New Mexico, Ohio, Oregon,
Pennsylvania, West Virginia, and Wisconsin. In 2008, the identified battleground states were Colorado, Florida, Georgia,
Indiana, Iowa, Michigan, Minnesota, Mississippi, Missouri, Montana, Nevada, New Hampshire, New Jersey, New Mexico,
North Carolina, Ohio, Oregon, Pennsylvania, South Dakota, Virginia, and Wisconsin. In 2012, the battleground states
were Arizona, Colorado, Florida, Iowa, Michigan, Minnesota, Missouri, Nevada, New Hampshire, North Carolina, Ohio,
Oregon, Pennsylvania, Virginia, and Wisconsin.

11

 population for each state. The figure shows that the Democratic candidates had more field offices than
the Republicans and that the battleground states had more field offices than the non-battleground states.
Similarly, Figures 3b and 3c illustrate the VAP-normalized GRPs for candidate advertising and PAC
advertising, respectively. Candidates advertised more intensively in battleground than in nonbattleground states. However, PAC advertising exhibited a less obvious association with competition
intensity. PACs advertised somewhat more evenly across battleground and non-battleground states,
perhaps because PAC advertising comes from multiple organizations, which likely have varying political
objectives and strategic focuses. Nevertheless, the difference between PAC and candidate advertising
empirically confirms that there was little coordination between the two types of sponsors. 16
< Figure 3 >
Next, we examine the extent to which campaign managers allocate resources according to the
expected effectiveness. This task is nontrivial, as it is unknown to researchers how managers form
expectations about campaign effectiveness. However, Section 2.5.2 provides suggestive evidence that
campaign effects seem to vary by voter partisanship. If the observed pattern is consistent with the
campaign managers’ underlying decision process, one would expect the counties with relatively higher
partisan support to have more field offices and PAC advertising, but less candidate advertising, after
controlling for competition intensity. To test this hypothesis, counties are again split into low and high
groups, depending on whether the percentage of partisan voters is below or above the mean. The upper
half of Table 4 presents the summary statistics for the campaign activities in each county by the level
of partisan support and battleground status. The bottom half of Table 4 shows the results of a regression
analysis to test the statistical significance of competition intensity and partisan support with regards to
the three campaign activities. Not surprisingly, far more campaign activities are present in battleground
states than in non-battleground states. More interestingly, the levels of field operations and PAC
advertising are, indeed, significantly higher in counties with more partisan voters than in those with less
partisan support. However, for candidate advertising, the county’s level of partisan support is statistically
insignificant. These patterns are directionally consistent with the conjecture that field operations and
PAC advertising are more effective with partisan voters, whereas non-partisans tend to respond more to
candidate advertising. Overall, the analysis here provides correlational support that candidates allocate
campaign activities by competition intensity and that the effectiveness of various campaign activities

16

According to the Federal Election Commission (FEC) advisories, Super PACs cannot legally coordinate with candidates
or political parties directly; nor can they donate funds to candidates.

12

 may differ by the level of local partisanship. Thus, to estimate the true campaign effects, it is important
to control for voter partisanship and the unobserved competition intensity. Given these considerations,
the next section describes the modeling approach and presents the identification strategy.
< Table 4>

3 Model
3.1 Model of Voter Preference
Individual i from county c has latent voting utility that he or she associates with a candidate from party
j during election t, denoted as uicjt . An individual faces three voting options: the Democratic candidate,
the Republican candidate, and the outside option (i.e., voting for an independent candidate or choosing
not to vote). Individual i chooses the option that yields the highest utility, and the market share for the
three options results from aggregating individual choices. The utility of individual i takes the form of

uicjt   (i (Gcjt , Acjt )   Bi   IXcjt   Ycj   %Ycjt   Gjt   Ficjt .

(1)

The term (i (Gcjt , Acjt ) captures how individual i's goodwill towards party-candidate j is affected by
the candidate's ground campaigning, Gcjt , and television advertising, Acjt . The effect of campaigns may
differ by individual, denoted by the subscript i, because of heterogeneous tastes (see Section 3.2. for the
detailed specification of the campaign effect).
The term Bi captures the remaining individual-specific heterogeneity in voting preferences. It can
be understood as the mean voting utility for i that is not explained by the voter’s exposure to campaigns.
This term is further decomposed into three parts: (1) the grand mean across individuals, B1 ; (2) the
deviation from the mean that is attributed to observable individual characteristics, such as the voter’s
partisan affiliation, Pijt ; and (3) the individual’s departure from the mean related to all other
unobservable individual characteristics, TOi , where Oi comes from a standard normal distribution. The
unobservable characteristics include factors such as the individual’s financial conditions, which could
shape the voting preference but are missing from the data. The three terms enter utility linearly such
that Bi   B1   B2Pijt   TOi .
The term IXcjt captures how observed county-election-specific characteristics affect voting utility.
Examples of such variables include the county's racial composition, political climate, and socio-economic
conditions, such as the median household income and the unemployment rate, all of which may influence

13

 voters’ preferences for candidates. The interaction of these contextual variables with a party indicator
captures the differential effects across parties.
Utility is also a function of unobservable characteristics that are specific to a county-party-election
combination. This is further decomposed into three parts: Ycj , %Ycjt , and Gjt . The unobserved Ycj refers
to the mean utility toward the candidate from party j across all residents in the same county. People
from the same county likely exhibit similar political preferences due to exposure to the same socioeconomic stimuli. The fixed effect Ycj absorbs the cross-sectional variation in party preferences across
counties, and, thus, the estimates result from the cross-election variation within counties.
The component %Ycjt is the county-party-election-specific deviation from the mean utility Ycj , which
quantifies the hard-to-measure utility shifts over time. For example, when Hurricane Sandy hit the
northeastern part of the United States in 2012, then-President Barack Obama promptly committed to
relief operations and was praised for his crisis leadership, causing a positive boost in his support, especially
in the areas impacted by the hurricane. Such unobserved factors would not be reflected in Ycj but would
be captured by %Ycjt . The county-party-election-specific deviation in candidate favorability, though
unobserved by the researcher, impacts voters’ preference for candidates. Thus, the potential correlation
between the campaign variables and the unobserved candidate favorability may bias the estimate for the
parameters in (i (Gcjt , Acjt ) in an ordinary least squares (OLS) regression (see Section 3.3 for the empirical
strategy to address this issue).
The component Gjt captures the common shift in voters’ overall preferences across elections and is
allowed to differ between parties; for example, there could have been unobserved differences when the
Republican candidate changed from George W. Bush to John McCain and when the Democratic
candidate changed from John Kerry to Barack Obama. Finally, Ficjt is the idiosyncratic utility shock
that is assumed to be independently and identically distributed (i.i.d.) Type I extreme value across
individuals, counties, candidates, and elections.
3.2 Specification of Campaign Effect
The campaign effect, (i (Gcjt , Acjt ) , is a function of party-candidate j’s ground campaigning, Gcjt , and
mass-media advertising activities, Acjt . Campaign advertising has two primary types: advertisements
made by the candidates and their respective parties, Acjtp ; and advertisements sponsored by outside PAC
o
groups, Acjt
. All three types of campaign measures enter the model in the log form to capture a potentially

14

 diminishing return. Furthermore, all campaign activities can have heterogeneous effects across
individuals, captured by the subscript i. Hence, the combined campaign effect takes the following form:
o
(i (Gcjt , Acjt )   CiGcjt   Hi Acjtp   Qi Acjt
.

(2)

Parameter Ci captures voter i's taste for field operations and consists of two components: (1) the
mean effect of field operations, C1 ; and (2) the deviation from the mean that could be attributed to
observable individual characteristics, C2Pijt . 17 The parameters Hi and Qi consist of two components
similarly, such that
Ci
Hi

  C1   C2Pijt
  H1   H2Pijt ,

Qi

  Q1   Q2Pijt

(3)

where Pijt represents a voter’s party affiliation, a key factor affecting the political preference for
candidates. This individual partisan variable can take on three values—affiliated with the Democratic
Party, the Republican Party, or neither—and, thus, follows a multinomial distribution, where the
empirical means of the categories correspond to the observed percentages of registered partisan voters
for each county. For example, if a county had 30% registered Democrats and 35% Republicans, the
simulated individual partisanship would have roughly 30% labeled Democratic, 35% labeled Republican,
and the remaining 35% labeled neither.18 Note that the individual’s partisanship may also affect the
overall voting preference, which is captured in Bi .
The estimation procedure follows the standard random-coefficient model as defined in Berry,
Levinsohn, and Pakes (1995). This is typically referred to as the “BLP” model and has been used in
various marketing applications (e.g., Chung 2013; Sudhir 2001).
3.3 Identification and Estimation
The BLP model specification allows us to examine the individual heterogeneity in campaign effectiveness,
even though voters’ choices are observed at the aggregated county level. The parameters are estimated

17

The results of an alternative model, which includes unobserved individual heterogeneity for each campaign effect, show
negligible unobserved heterogeneity (i.e., parameter values were small and insignificant).
18
States differ in terms of how independent or unaffiliated voters can participate in a state primary. In states with closed
primaries, only voters affiliated with a party can vote in the party’s nominating contests. However, in a number of other
states, unaffiliated voters can take part in a party’s primary. However, voters affiliated with the other party cannot cross
over to vote. Therefore, the meaning of “unaffiliated” voters could be different in different states. The inclusion of countyparty fixed effects can help address this cross-sectional difference since the requirement is at the state level and constant
over time. We thank an anonymous reviewer for raising this point.

15

 using the generalized method of moments (GMM), where the moment conditions involve a vector of
instrumental variables, which is orthogonal to the unobserved county-party-election shock %Ycjt (Hansen
1982).
Inferences about the campaign effects requires carefully addressing the endogeneity concern (i.e.,

%Ycjt is correlated with the level of campaigning). As discussed in Section 2.5.3, at least two reasons
could explain how campaign allocations are strategically determined: the need to campaign where it is
necessary and the need to campaign where it is effective. The data used in this study indicate that
competition intensity relates to the scale of campaigning (Shachar 2009). Furthermore, the data provide
suggestive evidence that campaign allocations vary by local partisan support. When campaigns are
deployed where they are better received, the OLS estimate would be greater than the true campaign
effect (i.e., upward bias). When campaigns are planned according to the intensity of competition, the
sign of the estimation bias becomes ambiguous. Campaign efforts are likely to decrease if the preference
for the candidate is either too strong or too weak (Gordon and Hartmann 2013); thus, the OLS estimates
would be biased downward for the former and upward for the latter. Combined, the sign of the
endogeneity bias becomes unclear. To address the potentially different campaign effects related to
partisanship, we control for voters’ partisan status and explicitly model how the effects vary by partisan
segment. To address the endogeneity concern related to unobserved competition intensity, we estimate
the effects for field operations and advertising using different identification strategies because the two
types of campaign activities employ different geo-based allocation strategies.
Regarding the identification strategy of field operations, compared to advertising, which is
determined at the DMA level, grassroots field operations are targeted at a more granular and local level
(Huber and Arceneaux 2007; Wielhouwer 2003). For example, the DNC stated in its Campaign Training
Manual that “[grassroots campaigns] are designed to reach a very special target group of favorable voters
either individually, in select precincts or in most cases a combination of both” (Wielhouwer 2003). In
other words, the level of field operations closely relates to the local political predisposition and voter
characteristics. Thus, it is important that the model includes the county-party-level fixed effects, which
help address the targeting of field operations related to unobserved county-level preferences for
candidates.
Even with the fixed effects, it is possible that %Ycjt , the unobserved time-varying shock to voter
preference, may still be correlated with field operations—for example, if the campaign managers detect
a change in the local political favorability (relative to the previous election) and adjust the number of
16

 field offices accordingly. Wielhouwer (2003) identifies several groups of variables that may influence field
office allocations. The most important factor is voters’ registered partisanship, as managers can use voter
registration to estimate the number of contacts to make. The second group of variables includes voters’
age, education, income, and African-American ethnicity. For example, field operations tend to target
older and better-educated residents, and Democrats and Republicans target African American voters
differently. This study includes these county-level influencers, so that their effects are eliminated from

%Ycjt . According to Wielhouwer (2003), the next variable is residential connectedness—that is, whether
voters live in a more socially connected (e.g., urban) or less socially connected (e.g., rural) area. It is
reasonable to assume that the urban versus rural status does not change substantially over eight years
(the data collection period of this study), so the county-party fixed effects largely absorb this effect. The
changes in county-level population density from one election to the next capture the remaining crosselection variation in social connectedness. Lastly, a vector of party-county-level variables is included to
further control for the local political climate: whether the candidates have home-state advantage; whether
the governor is from the same party; whether the governor is an incumbent (if from the same party);
and the distance of each county to the campaign headquarters. Given these considerations, the
combination of granular (county-party) fixed effects and the many county-level control variables should
address the endogeneity concerns related to field operations. Nevertheless, we acknowledge that if some
unobserved contextual changes over elections (that this study does not control for) affect the number of
field offices, the model estimates would produce biased results. As with advertising, the direction of the
bias is ex ante unclear because campaign strategists may decide to have fewer offices when their
candidate’s favorability is very low or when the focal county is considered secure for voting outcomes
(i.e., favorability is very high).
Next, regarding the identification strategy of advertising, the county-party fixed effects would not
directly address the endogeneity problem related to advertising because advertising is determined
strategically at the DMA level. Thus, the identification strategy relies on the instrumental variable
approach, using cost-shifting variables. The instruments follow those of Gordon and Hartmann (2013)—
i.e., the third-quarter DMA-level ad prices in the year before each election. The argument for the validity
of these instruments is that price changes affect advertising costs and, hence, shift the amount of
advertising. However, the cause of the price fluctuation in the previous year, instead of in the election
year, is assumed to be outside the system (i.e., independent of %Ycjt ).

17

 The ad-price data come from the Kantar Media SRDS TV and Cable Source, in which prices exist
for four dayparts: prime access, prime, late news, and late fringe. Although they are invariant across
candidate advertising and PAC advertising, the ad costs can instrument both types of advertising
through the difference in airtime for each type. Data from the University of Wisconsin Advertising
Project (Goldstein and Rivlin 2008) illustrate that candidates’ advertisements aired more during the
prime and prime access dayparts than did PAC advertisements, while the latter appeared more frequently
during the late news dayparts (see Figure A1 in the online Appendix). Therefore, the costs for different
dayparts have varying effects on the two types of advertising, providing the variation needed for
identification. Section 4.1 presents the detailed diagnostic statistics for the instrumental variables.
The instruments by themselves, however, would not provide the between-party variation because
they are constant across parties. The interactions between the party indicator and each of the cost
shifters would provide the between-party variation in the first-stage fitted values for the endogenous
variables (Gordon and Hartmann 2013). Furthermore, the interactions between the ad costs and the
county-level percentage of partisans for each party would help identify the observed heterogeneity
parameters associated with individual partisanship. This set of instruments, adding the partisan-related
variation in the first-stage fitted values, helps leverage the covariance between advertising and vote share
that is associated with voter partisanship concentration. Thus, the final matrix of instruments contains
the lagged ad prices, the interactions with the party variable, their respective interactions with countylevel percentage partisans, and all of the exogenous variables in Equation (1), including the county-party
fixed effects.
Individual partisanship is coded using the “simple contrast” method: 0.5 for a partisan voter for each
party and -0.5 for an unaffiliated voter. Using simple contrast coding instead of conventional dummy
variable coding, the main campaign effect (mean coefficients) in the linear component of the model
becomes the average campaign effect for a typical voter (Irwin and McClelland 2001). For the states
with open primaries, voters do not declare party affiliation, and, thus, such data are unavailable. 19 For
those states, the model includes only the linear components to capture the aggregated average campaign
effects, and the instruments include only the ad costs and their interactions with the party indicator.

19

In states with open primaries, voters do not reveal party affiliation during registration. Thus, whether or not partisanship
data are available depends on a state’s type of primary, which is determined by historical factors. Hence, the availability
of partisanship is likely exogenous to the competitive level of each election. A robustness check shows that states with
and without the partisan variable are not statistically different in terms of the level of the three types of campaign
activities.

18

 Note that the coefficients regarding observed heterogeneity are identified from the states with
partisanship information. Heuristically, the variation in vote share for counties with different partisan
density but the same campaign activities helps identify the observed heterogeneity parameters related
to campaign effects. For example, suppose that two counties observe the same increase in the number of
Democratic field offices from one election to the next. If the increase in the vote share is higher for a
county with a higher percentage of registered Democrats, the coefficient on the partisan variable is
identified as positively interacting with the effect of field operations. The same logic applies to how
partisans respond to advertising. Unobserved heterogeneity is identified through different substitution
patterns in the data (Nevo 2000).

4 Results
4.1 Parameter Estimates
Table 5 presents the results of four model specifications.20 The first two specifications are based on an
OLS regression with and without the county-party fixed effects, respectively. The third specification
incorporates the instrumental variables (IVs) for advertising, and the fourth specification allows for
heterogeneous campaign effects across individual voters.
The addition of the county-party fixed effects increases the model R-squared from 0.601 in column
1 to 0.955 in column 2. Furthermore, the estimates of campaign effects decline significantly with the
inclusion of the fixed effects, confirming the importance of addressing the heterogeneous political
preferences across counties. The Durbin–Wu–Hausman (DWH) test (Durbin, 1954; Wu, 1973; Hausman,
1978) is also performed to examine the model specification of treating county-party effects as fixed effects.
If the unobserved county-party heterogeneity is correlated with the explanatory variables, the fixedeffect model is consistent; otherwise, the more efficient random-effect model is preferred (Greene 2018).
The results show that the random-effect model is rejected and that the fixed-effect specification is
preferred ( D2 (51)   2,157.6 , p<0.001). Therefore, the subsequent analyses always include the fixed effects.
< Table 5>
Column 3 of Table 5 presents the estimates based on the instrumental variables. First, the estimates
for field operations remain qualitatively similar to those in column 2. The similarity of the IV estimator
with the fixed-effect estimator is expected, as the instrumental variables of DMA-level cost shifters should

20

Table 5 presents abridged results for simplicity of exposition. Table A2 in the Online Appendix shows the full results,
including parameter estimates for the control variables.

19

 not co-vary with the number of field operations in a county. Next, with the use of instrumental variables,
the estimates for both types of advertising are larger than the fixed-effect estimates. The larger IV
estimates (i.e., larger than the corresponding OLS estimates) suggest that the unobserved demand shocks
are negatively correlated with the changes in advertising: candidates and PACs tend to increase
advertising in the contested areas in which the favorability is low, thus causing the OLS estimates to be
attenuated.
To examine whether the advertising variables should be treated as endogenous, the DWH test is
performed by comparing the more efficient OLS estimator with the less efficient, but consistent, IV
estimator (Hausman 1978). The null hypothesis of the test is that the preferred model is the more efficient
OLS estimator. After including the fixed effects and the socio-economic variables, the DWH test results
are significant for candidate advertising ( D2 (1)   54.284 , p<0.001) and PAC advertising ( D2 (1)   25.945 ,
p<0.001), indicating that the estimators without the IVs are inconsistent for both advertising variables.
The diagnostic statistics on the relevance of our instrumental variables are as follows. The first-stage
regression equations are specified following Angrist and Pischke (2008), and Shea’s partial R-squared is
computed to account for the potential collinearity among multiple endogenous variables and the
instruments. For each endogenous variable, Shea’s partial R-squared directly measures the correlation
between the variation orthogonal to the other endogenous variables, and the component of its projection
on the instruments orthogonal to the projection of the other endogenous variables on the instruments
(Shea 1997). Shea’s partial R-squared reveals the relevance of the instruments in uniquely identifying
each endogenous variable. Note that this is after controlling for the county-party fixed effects and all the
control variables. Shea’s partial R-squared is 5.2% for candidate advertising and 5.3% for PAC
advertising, indicating that the instruments have exploratory power to identify both types of advertising.
The Stock and Yogo (2005) test is further conducted to test weak instruments. The minimum eigenvalue
statistic from the test is 18.22, which is greater than the critical value of 10.96 if one is willing to tolerate
10% bias of the 2SLS estimator. Thus, the null hypothesis of weak instruments is rejected for a small
bias of 10%. See Table A1 in the Online Appendix for the results of the first-stage diagnostic statistics.
The final specification in Table 5 incorporates voter heterogeneity. In particular, the model examines
ways in which the effect of campaign activities varies with voter partisanship. The first column under
specification (4) lists the parameter estimates for an average voter. The column (labeled “Partisanship”)
under (4) is the estimated difference between partisan and non-partisan voters (i.e., the interactions
between the campaign effect and voter partisanship).
20

 The advertising elasticities are computed based on the parameter estimates (0.034 for candidate
advertising and 0.026 for PAC advertising). For candidate advertising, on average, a 1% increase results
in a 0.023% increase in the vote share for the Republicans and a 0.026% increase for the Democrats. The
elasticity for PAC advertising is 0.017% and 0.020% for the Republicans and Democrats, respectively.
To benchmark these estimates with existing elasticities, Gordon and Hartmann (2013) estimate the ad
elasticities (with candidate and PAC ads combined) to be 0.033% for the Republicans and 0.036% for
the Democrats in the 2000 and 2004 presidential elections. 21 Using data from the 2010 and 2012
senatorial elections, Wang et al. (2018) estimate the elasticity of negative candidate advertising to be
0.013% for the Republicans and 0.015% for the Democrats, whereas the overall ad elasticity (including
positive and negative advertising) is not significant. Thus, the ad elasticities estimated in this study fall
within the middle of existing estimates.
The parameter estimate for field operations is 0.078. It is challenging to benchmark the fieldoperation elasticity, as empirical studies on this topic are scarce. In one exception, Darr and Levendusky
(2014) estimate the magnitude of the effect for the Democrats and identify a 1.04% increase in countylevel vote share with the presence of a field office. In this study, the percentage change in vote share in
response to an additional Democratic field office is 1.25% in 2008 and 1.31% in 2012. 22 The finding is
qualitatively consistent with Darr and Levendusky’s (2014) and, thus, suggests that the estimated
magnitude for field operation elasticity is robust to different model specifications (i.e., controlling for
more county-level characteristics and taking the county-level partisan support into consideration).
Not surprisingly, voter partisanship has a strong main effect. Independent from the campaign effects,
partisans tend to favor the candidates from their party more than do non-partisans. Second, the
interaction estimates with campaign activities are all significant, indicating that partisans and nonpartisans react differently to campaign activities. Interestingly, partisans do not always react more
positively to campaign messages, and the nature of the interaction depends on the type of campaigning.
Field operations are more influential with partisan voters, whereas candidate advertising is less effective

21

Another robustness analysis that combined the candidate advertising and PAC advertising, as done in Gordon and
Hartmann (2013), shows that the advertising elasticity is 0.024 for the Republicans and 0.028 for the Democrats. Note
that Gordon and Hartmann (2013) reconstructed the advertising GRPs based on price, whereas this study uses the
actual GRPs.
22
The addition of one field office can be understood as a proxy for the average number of voter contacts associated with
a typical field office. The fact that the number of field offices is highly correlated with the number of self-reported voter
contacts (see Section 2.2) suggests that there may be a narrow distribution for the number of voter contacts behind each
field office. We thank an anonymous reviewer for this comment.

21

 with partisans than with non-partisans. PAC advertising, surprisingly, is different from candidate
advertising and has a stronger effect on partisans.
Table 6 summarizes the campaign-effect estimates by partisanship. Field operations have a positive
effect on partisan voters but a limited effect on non-partisans. Candidate advertising has a positive effect
on both partisan and non-partisan voters, whereas the effect is larger for non-partisans. In contrast, the
opposite pattern exists for third-party-sponsored PAC advertising, with a larger effect on partisans.
< Table 6 >
Why would the effectiveness of various campaign activities differ by voter partisanship? We first
discuss the difference in the mechanism between field operations and advertising in general, followed by
the difference between candidate advertising and PAC advertising.
The interaction between voters’ predispositions and DMP status may explain the difference between
field operations and candidate advertising. When a voter encounters on-the-ground canvassing efforts by
a campaign, the messages typically involve why the voter should turn out to support the focal candidate
(e.g., Gerber and Green 2000a; Green et al. 2003; Nickerson et al. 2006; Wielhouwer 2003). The voter’s
decision of whether to vote for his or her preferred candidate on Election Day occurs after the voter has
decided on the preferred candidate from the pool of candidates. Compared with unaffiliated voters,
partisan voters are likely to have established favorable opinions about their party’s candidate, and, thus,
they are more likely to be at a later stage in their DMP. As a result, the partisan turnout messages
delivered during field campaigning will be better received by partisan voters than by unaffiliated voters.
In contrast, existing research suggests that advertising has a limited effect on turnout (Krasno and
Green 2008) but a positive effect on influencing voters’ preferred candidate choices (Huber and Arceneaux
2007; Lovett and Peress 2015) at an earlier stage of voters’ DMP. For a typical voter’s process of receiving
persuasive messages and responding to them, the RAS model (Zaller 1992) postulates that voters form
preferences through a two-stage process: receive and accept. Table 7 lays out the framework.
< Table 7>
In large part, there are three unique electorates in the population: those who are highly aware; those
who are moderately aware; and those who are not aware at all and do not care about politics. Awareness
here does not narrowly refer to being aware of a particular advertising message; rather, it means
awareness of politics in a broader sense (i.e., whether or not people care about politics). Partisan voters,
as the name implies, are people who care deeply enough about politics to have registered with a particular
party. Thus, it is reasonable to assume that partisans have a higher level of political awareness than an
22

 average voter (the general public). The majority of independent voters, on the other hand, are people
who are moderately aware of politics. They care enough to cast their votes on Election Day but do not
identify with a particular party. Of course, some independent voters are highly aware of politics;
nevertheless, partisan voters as a group are relatively more aware. Finally, some people simply do not
care about politics. These people are not likely to vote in any election (about half of the entire U.S.
electorate) and, thus, are not the focus of campaigns.
When a message (information) intended to persuade arrives (through advertising), highly and
moderately aware people will “receive” the message—highly-aware people even more so—whereas
unaware people will not. For example, when seeing an advertisement on television, both highly and
moderately aware people will understand its political nature and consume the message, whereas unaware
people will simply ignore it. However, after receiving the message, highly aware people will not accept it
if the message goes against their political predisposition (e.g., the message attacks their party’s candidate).
Even if the message is consistent with their political predisposition, it still will have limited influence
because they are likely to have already established their preferences for the candidate, and there is no
room to increase their favorability. In contrast, moderately aware people, although less likely to receive
the message, would be more open to accepting and being influenced by the information in the message.
This mechanism helps explain the differentiating effect of candidate advertising on voter types.
Regarding PAC advertisements, besides an obvious difference in sponsorship, they are also known
to differ from candidate advertisements in their tone. PAC advertising is typically known for its
negativity and attacking tone. Because moderately aware people (non-partisans) have a low tolerance
for negativity (Fridkin and Kenney 2011), they are less likely to receive negative or combative messages.
For example, independent voters may be put off by attack advertisements and, thus, may simply ignore
them. Therefore, although unaffiliated voters have more room to be persuaded (similar to candidate
advertising), the negativity of PAC advertisements may create a backlash and cause non-partisan voters
to not receive the message. This leads to a weakened effect on non-partisan voters. Partisan voters, on
the other hand, are more tolerant of a negative tone and, thus, allow the PAC advertisements to still
have an effect (Finkel and Geer 1998).

23

 4.2 Robustness Analyses
Table 8 presents the results from alternative model specifications to test the robustness of the results. 23
Model 1 includes only field operations, while Model 2 includes only candidate and PAC advertising.
Models 3 and 4 retain field operations and only one type of advertising: candidate for Model 3 and PAC
for Model 4. Model 5 replaces the two original types of advertising with the combined total of candidate
and PAC advertising.
Overall, the alternative models yield results that are qualitatively consistent with those of the study’s
main model: all three campaign activities have a significant average effect, and the effect varies with a
voter’s partisanship. Field operations and PAC advertising interact positively with the voter’s
partisanship, while candidate advertising interacts negatively with partisanship. When excluding one or
both advertising variables (Models 1, 3, 4, and 5), the average effect of field operations is slightly
overestimated, whereas the IV estimates for advertising are more robust to the removal of field operations
(Model 2).
Furthermore, in Model 5, in which candidate advertising and PAC advertising are combined into
one advertising variable, the coefficient estimate for an average voter is 0.021 (p < 0.01), whereas the
interaction between partisanship and combined advertising is -0.008 (versus -0.019 for candidate
advertising and 0.018 for PAC advertising in the main model). The different partisan effect of advertising
between Model 5 and the main model emphasizes the importance of modeling the separate effect for
candidate advertising and PAC advertising, respectively.
< Table 8>
4.3 Counterfactual Analyses
This section provides results of counterfactual simulations using estimates from the main model to answer
the “what if” question: What would the election results have been if the candidates had allocated
campaign resources differently?
The first counterfactual analysis quantifies the magnitude of the effect for each campaign activity.
It computes the would-have-been vote results had the corresponding activity been eliminated from the
campaign. To do so, we (1) set the focal activity to zero and leave the other campaign activities intact;
(2) compute the predicted county-level votes for the Democrats and the Republicans, respectively; and
(3) predict the winner in each state. Before discussing the results, it is worth pointing out that this

23

We thank an anonymous reviewer for suggesting these robustness checks.

24

 counterfactual exercise is not an equilibrium analysis. The simulation does not take into consideration
the competitive response or the re-allocation of resources, in the sense that when a party removes or
alters one activity from the campaign mix, the party’s and its rival’s other activities are kept intact. A
full equilibrium analysis would require a supply-side model that solves the new equilibrium level for all
the remaining activities, given a fixed campaign budget. Such an analysis should consider the costs and
competitive response, but doing so is beyond the scope of the current study. Nevertheless, the analysis
here still helps compare the relative effect of the various campaign activities.
The results in Table 9 highlight the importance of field operations for the Democrats. No field
operations by either side would have resulted in the Republicans winning Pennsylvania and Wisconsin
in 2004, Florida, Indiana, and North Carolina in 2008, and Florida in 2012. Although the final national
results would have remained unchanged, the absence of field operations would have reduced the winning
margin for the Democrats. In the absence of candidate advertising, the Republicans would have won
New Hampshire and Wisconsin in 2004 and Indiana and North Carolina in 2008, whereas in the 2012
election, the changes in the popular votes would not have led to different electoral votes for the two
parties. The effect of field operations on electoral outcomes is more substantial than that of candidate
advertising, which is consistent with the findings of existing research in a more general setting.
Specifically, meta-analyses of various studies find that the elasticity of personal selling expenditures is
about 0.34 (Albers et al. 2010), whereas that of advertising expenditures is 0.22 (Assmus et al. 1984).
Furthermore, jointly examining a particular context of Navy enlistment, Carroll et al. (1985) find that
the elasticity of a field sales force was large and significant, whereas advertising was less so.
The next question relates to PAC-sponsored advertisements. Eliminating outside advertising barely
moved the needle on election results in 2004 and 2008, which is unsurprising given that the amount of
PAC advertising in these elections was small and comparable between the parties. However, in the 2012
election, for which the amount of outside advertising was more sizable, eliminating PAC advertising
could have helped the Democrats win Georgia, Indiana, and North Carolina. In other words, there is
some evidence that the high PAC advertising spending helped the Republicans in 2012. Given that the
amount of outside advertising has grown substantially in recent presidential elections, this counterfactual
analysis suggests that PAC advertising could play an important role in shaping election results and, thus,
is a factor that cannot be ignored in the presidential campaign mix.

25

 Currently, Super PACs are legally prohibited from directly coordinating their advertising efforts with
candidates.24 The next counterfactual analysis is to understand the effect of this policy (row 5 in Table
9). Specifically, PACs are made to donate their advertising money to the candidates they favor. Therefore,
PAC advertising becomes zero, and candidate advertising increases. The relative prices paid by the
candidate and the PACs, respectively, determine the amount of the increase. 25 For the 2004 and 2008
elections, when the level of PAC advertising was relatively low, donating PAC advertising expenditures
to campaign advertising yielded the same state-level voting outcomes as eliminating PAC advertising
from the campaign mix. In 2012, however, with the PAC spending donated to the candidates, the
Democrats would have won Indiana and North Carolina. Given that the Republicans received far more
PAC advertising, it may seem surprising at the first glance that donating PAC ad spending benefited
the Democrats more than the Republicans. The reason is, perhaps, the net effect of two factors. First,
PAC advertising is, overall, more expensive than candidate advertising (Moshary 2020). When PACs
donate advertising funds to the candidates, they are trading more-expensive ads for less-expensive ones,
which translates into more exposure. Second, however, the incremental advertising exposure from the
candidate campaigns has diminishing returns due to saturation and fatigue, for example. The
combination of these forces benefits the Democrats more than the Republicans.
< Table 9 >
The importance of field operations for the Democrats warrants a counterfactual analysis that
computes the “value” of a field office. In other words, how much additional spending would have been
needed to obtain the same election results if the Democrats had eliminated their field operations? The
focus is on the three states in which the Democrats would have lost to the Republicans with no field
operations in the 2008 election. For each of these states, we compute the optimal additional candidate
advertising required by the Democrats to still win the state (again, in the absence of field operations).
The values are searched through constrained optimization: the additional Democratic advertising should

24

The detailed regulations are in the FEC report: https://www.fec.gov/resources/cms-content/documents/ colagui.pdf.
The following steps are conducted for this analysis. First, for each party in each DMA, we use the total amount of
advertising spending and GRPs to compute the average advertising prices paid by the candidate and the PACs,
respectively. Then, with the relative prices and the total amount of PAC advertising spending, we compute the
counterfactual amount (in GRPs) had the money been donated to the candidate.
25

26

 be the least costly, given that the Democrats would still win the state. 26 Table 10 shows the results. Note
that the advertising amount and cost are at the DMA level to mimic the actual ad allocation process as
much as possible. Again, this is not a full equilibrium analysis, as the analysis assumes that the
Republicans would not adjust their campaign efforts upon observing the changes in the Democrats’
campaigns.
The results indicate that field offices were “valuable,” in the sense that in the absence of field
operations, the Obama campaign would have needed about $3.0 million worth of additional television
advertising to win Florida, $4.0 million to win Indiana, and $9.3 million to win North Carolina. These
figures could help assign a dollar amount to an Obama field office: an office was, on average, worth about
$0.05 million (of television advertising) in Florida, $0.20 million in Indiana, and $0.19 million in North
Carolina. It is interesting to zoom in on the comparison between Florida and North Carolina. In 2008,
the Obama campaign deployed a comparable scale of field operations and TV advertising in these two
states. However, the above analysis shows that a field office would have been more valuable in North
Carolina, as it would have been more expensive for the Obama campaign to try to replace field operations
with TV advertising. What could have accounted for this difference is that the percentage of Democratic
voters is higher in North Carolina than in Florida. Because field operations are more effective on partisans,
while candidate advertising is more effective on non-partisans, for a state with higher partisan support,
replacing field operations with TV advertising would be costlier and less efficient.
< Table 10 >

5 Conclusion
This study examines the effects of marketing through personal selling and mass-media advertising in the
2004-2012 U.S. presidential elections. By linking campaign activities to county-level vote results, the
study offers a comprehensive view of the effects of various campaign activities. Distinct from extant
research, this study simultaneously considers on-the-ground campaign efforts and television advertising
and distinguishes between the effects of candidate and PAC advertising. Most importantly, the study
examines the heterogeneous effects of campaign activities on two types of voters: partisans versus non-

26

The detailed procedure of this counterfactual exercise is as follows: (1) we set the field offices to zero; (2) we set the
choice variable as the additional DMA-level candidate advertising GRPs for the Democrats; (3) the DMA-level GRPs are
mapped to the county-level; (4) we estimate the county-level new vote counts based on the new campaign variables for
the Democrats (assuming that the Republicans do not readjust their campaigns); and (5) the state-level voting results are
realized. The values in (2) are searched by minimizing the additional total advertising spending under the constraint that
the Democrats should still win the state in step (5).

27

 partisans. The results generate insights into the effectiveness of each campaign activity by voter segment.
Field operations and PAC advertising are more effective with partisan voters, while candidate advertising
is more effective with unaffiliated voters. This study’s findings can provide guidance to campaign
strategists on resource allocation.
Counterfactual election results in several hypothetical scenarios show that ground campaigning
played an important role for the Democratic Party in several battleground states during the 2008 and
2012 elections. In addition, replacing field operations with candidate advertising in these states (to
achieve winning status) would have been quite costly.
Some of the results merit further discussion. First, it is interesting to note that candidate advertising
is more effective with non-partisan voters, whereas PAC advertising is less effective with this segment,
likely because of the difference in advertisement content. PAC advertisements are predominately negative
and tend to attack rivals rather than promote the preferred candidate. Such a strong negative tone may
turn off unaffiliated voters, who are less tolerant of negative and attack messages in political
advertisements (Fridkin and Kenney 2011). Furthermore, the finding that candidate advertising is more
effective with non-partisans also seems to suggest that there is a stronger effect on persuasion than on
turnout (Krasno and Green 2008; Huber and Arceneaux 2007; Lovett and Peress 2015).
Second, the definition of field operations is worth closer examination. Field offices differ in terms of
their intended target population, the efficiency of their internal organization (e.g., the data available on
voters and techniques used to target and persuade them 27), and even the primary focus of the office (i.e.,
get-out-the-vote versus fundraising). Therefore, the effect of field operations on voting could be
heterogeneous across offices. Despite the importance of office heterogeneity, detailed data on how field
teams operate are challenging to obtain, especially at a granular level. We acknowledge this data
limitation and believe that future research can benefit from better data to more precisely examine the
operations of ground campaigning.
Another issue worth discussing is the mechanism of field operations. Field campaigning in this study
is largely related to partisan outreach efforts organized by the candidate’s field offices, in the form of
phone calls and canvassing contacts that were meant to deliver partisan messages favoring the focal
candidate. The study finds that these field operations are effective for partisans but have limited effects
on non-partisan voters. However, there are also non-partisan-based mobilization campaigns, during which

27

In the same vein, there may be reasons to suspect that field offices, even from the same party, may have a heterogeneous
effect due to operational differences. We thank an anonymous reviewer for this comment.

28

 volunteers encourage general turnout without explicitly advocating a specific candidate. For these nonpartisan grassroots activities, Gerber and Green (2000a) find a larger effect with the independents than
with partisans. These different results seem to suggest an interaction between voter segment and the
content of canvassing messages. Since this study is restricted to partisan canvassing only, we are unable
to make inferences on non-partisan field campaigns. This study aims to raise interest in modern partisan
field operations by documenting the scale and magnitude of the effect. We hope that future research can
incorporate non-partisan field operations and validate this study’s findings using alternative identification
strategies.

29

 References
Albers S, Mantrala MK, Sridhar S (2010) Personal selling elasticities: A meta-analysis. Journal of
Marketing Research 47(5): 840–853.
Angrist JD, Pischke J (2008) Mostly harmless econometrics: An empiricist’s companion. Princeton,
NJ: Princeton University Press.
Assmus G, Farley JU, Lehmann DR (1984) How advertising affects sales: Meta-analysis of
econometric results. Journal of Marketing Research 21(1): 65-74.
Berry S, Levinsohn J, Pakes A (1995) Automobile prices in market equilibrium. Econometrica 63(4):
841–890.
Campbell JE (1992) Forecasting the presidential vote in the states. American Journal of Political
Science 36(2): 386–407.
Carroll VP, Rao AG, Lee HL, Shapiro A, Bayus BL (1985) The Navy enlistment marketing
experiment. Marketing Science 4(4): 352–374.
Chung DJ (2013) The dynamic advertising effect of collegiate athletics. Marketing Science 32(5):
679–698.
Darr JP, Levendusky MS (2014) Relying on the ground game: The placement and effect of campaign
field offices. American Politics Research 42(3): 529–548.
Durbin, J (1954) Errors in variables. Revue de l'institut International de Statistique 22(1/3): 23-32.
Finkel SE, Geer JG (1998) A spot check: Casting doubt on the demobilizing effect of attack
advertising. American Journal of Political Science 42(2): 573–595.
Fridkin KL, Kenney P (2011) Variability in citizens’ reactions to different types of negative
campaigns. American Journal of Political Science 55(2): 307-325.
Gerber AS, Green DP (2000a) The effect of a nonpartisan get‐out‐the‐vote drive: An experimental
study of leafletting. Journal of Politics 62(3): 846-857.
Gerber AS, Green DP (2000b) The effects of canvassing, telephone calls, and direct mail on voter
turnout: A field experiment. The American Political Science Review 94(3): 653–663.
Gordon BR, Hartmann WR (2013) Advertising effects in presidential elections. Marketing Science
32(1): 19–35.
Gordon BR, Hartmann WR (2016) Advertising competition in presidential elections . Quantitative
Marketing and Economics 14(1): 1–40.
Goldstein K, Rivlin J (2008) Presidential Advertising Combined File. Madison, WI: The University
of Wisconsin Advertising Project, The Department of Political Science at The University of
Wisconsin-Madison.
Green DP, Gerber AS, Nickerson DW (2003) Getting out the vote in local elections: Results from
six door-to-door canvassing experiments. The Journal of Politics 65(4): 1083-1096.
Greene, WH (2018) Econometric Analysis (8th Edition). Pearson, New York University, NY.
Hansen LP (1982) Large sample properties of generalized method of moments estimators.
Econometrica 50(4): 1029–1054.
Hausman, J (1978) Specification tests in econometrics. Econometrica 46(6): 1251–1271.

30

 Holbrook TM, McClurg SD (2005) The mobilization of core supporters: Campaigns, turnout, and
electoral composition in United States presidential elections. American Journal of Political
Science 49(4): 689-703.
Huber GA, Arceneaux K (2007) Identifying the persuasive effects of presidential advertising.
American Journal of Political Science 51(4): 957-977.
Irwin JR, McClelland GH (2001) Misleading heuristics and moderated multiple regression
models. Journal of Marketing Research 38(1): 100-109.
Iyengar S, Hahn KS (2009) Red media, blue media: Evidence of ideological selectivity in media
use. Journal of Communication 59(1): 19-39.
Klein JG, Ahluwalia R (2005) Negativity in the evaluation of political candidates. Journal of
Marketing 69(1): 131-142.
Krasno JS, Green DP (2008) Do televised presidential ads increase voter turnout? Evidence from a
natural experiment. The Journal of Politics 70(1): 245-261.
Lovett MJ, Peress M (2015) Targeting political advertising on television. Quarterly Journal of
Political Science 10(3): 391–432.
Lovett MJ, Shachar R (2011) The seeds of negativity: Knowledge and money. Marketing Science
30(3): 430-446.
Moshary S (2020) Price discrimination in political advertising: evidence from the 2012 US
presidential election. Forthcoming at RAND Journal of Economics.
Masket SE (2009) Did Obama’s ground game matter? The influence of local field offices during the
2008 presidential election. Public Opinion Quarterly 73(5): 1023–1039.
Nevo A (2000) A practitioner's guide to estimation of random‐coefficients logit models of
demand. Journal of economics & management strategy 9(4): 513-548.
Nickerson DW, Friedrichs RD, King DC (2006). Partisan mobilization campaigns in the field:
Results from a statewide turnout experiment in Michigan. Political Research Quarterly 59(1):
85-97.
Shachar R (2009) The political participation puzzle and marketing. Journal of Marketing Research
XLVI: 798–815.
Shaw DR (1999) The effect of TV ads and candidate appearances on statewide presidential votes,
1988-96. The American Political Science Review 93(2): 345–361.
Shea J (1997) Instrument relevance in multivariate linear models: A simple measure. Review of
Economics and Statistics 79(2): 348-352.
Stock, JH, Yogo M. (2005) Testing for weak instruments in linear IV regression. In Identification
and Inference for Econometric Models: Essays in Honor of Thomas Rothenberg, ed. D. W. K.
Andrews and J. H. Stock, 80–108. New York: Cambridge University Press.
Sudhir K (2001) Competitive pricing behavior in the auto market: A structural analysis. Marketing
Science 20(1): 42–60.
Wang Y, Lewis M, Schweidel DA (2018) A border strategy analysis of ad source and message tone
in Senatorial campaigns. Marketing Science 37(3):333-355.
Wielhouwer PW (2003) In search of Lincoln's perfect list: Targeting in grassroots
campaigns. American Politics Research 31(6): 632-669.
31

 Wu, DM (1973) Alternative tests of independence between stochastic regressors and disturbances.
Econometrica 41(4): 733-750.
Zaller JR (1992). The Nature and Origins of Mass Opinion. Cambridge university press.
Zoltners AA, Sinha P, Lorimer SE (2006) The Complete Guide to Sales Force Incentive
Compensation: How to Design and Implement Plans That Work. New York: AMACOM.

32

 Table 1: Summary Statistics for Vote Outcomes by County
Year
Vote Share
2004
2008
2012

Party

N

Democrat
Republican
Democrat
Republican
Democrat
Republican

Turnout Rate
2004
2008
2012

Mean

SD

Min

Max

3,111
3,111
3,106
3,106
3,108
3,108

0.22
0.35
0.25
0.33
0.22
0.33

0.08
0.09
0.10
0.09
0.10
0.10

0.04
0.05
0.04
0.04
0.02
0.02

0.58
0.76
0.64
0.78
0.70
0.73

3,111
3,106
3,108

0.57
0.58
0.55

0.09
0.09
0.09

0.19
0.16
0.15

0.98
0.90
0.99

Note: The turnout rate is the sum of votes for the Democratic and the Republican candidates divided by the number of
resident citizens aged 18 and above. The vote share is the total votes cast for a candidate, divided by the number of
resident citizens aged 18 and above.

Table 2: Summary Statistics for Field Operations by County
N
Number of Field Offices
2004 Democrat
Republican
2008 Democrat
Republican
2012 Democrat
Republican
Presence of Field Offices
2004 Democrat
Republican
2008 Democrat
Republican
2012 Democrat
Republican

Mean

SD

Min

Max

Total

3,111
3,111
3,106
3,106
3,108
3,108

0.10
0.03
0.28
0.08
0.24
0.09

0.45
0.23
0.75
0.46
0.93
0.37

0
0
0
0
0
0

12
5
11
17
21
6

313
89
874
247
750
278

3,111
3,111
3,106
3,106
3,108
3,108

0.08
0.02
0.20
0.06
0.14
0.07

0.27
0.14
0.40
0.24
0.35
0.26

0
0
0
0
0
0

1
1
1
1
1
1

237
65
624
192
439
222

Note: The unit of observation is a county. The presence of field operations denotes whether there is at least one field office
in the county.

33

 Table 3: Summary Statistics for Television Advertising by DMA
N
Candidate Advertising (GRPs)
2004 Democrat
Republican
2008 Democrat
Republican
2012 Democrat
Republican
Party Advertising (GRPs)
2004 Democrat
Republican
2008 Democrat
Republican
2012 Democrat
Republican
PAC Advertising (GRPs)
2004 Democrat
Republican
2008 Democrat
Republican
2012 Democrat
Republican

Mean

SD

Min

Max

Total

206
206
206
206
206
206

1,420.4
1,885.9
3,809.9
2,075.3
2,232.2
1,320.7

2,374.1
2,764.0
3,797.8
2,422.7
4,143.7
2,576.5

0
0
255
77
0
0

8,933
8,440
13,838
8,452
15,779
9,535

292,611
388,501
784,848
427,517
459,827
272,055

206
206
206
206
206
206

1,942.0
868.0
1,766.5
1,553.5
1,063.8
1,069.4

2,211.0
1,271.1
2,324.9
1,890.5
1,604.2
1,774.5

0
0
0
0
0
0

7,561
5,858
12,277
11,035
7,270
11,044

400,054
178,814
363,905
320,013
219,144
220,291

206
206
206
206
206
206

255.9
394.4
159.4
217.2
254.3
2,030.9

505.0
866.2
407.3
435.7
694.6
2,714.9

0
0
0
0
0
67

2,248
4,023
2,513
2,188
3,840
12,137

52,726
81,250
32,830
44,736
52,378
418,356

Note: Gross rating points (GRPs) correspond to the percentage of the target population reached in each DMA.

Table 4: Campaign Activities by County-level Partisan Support
Voter Partisan Level

Field Operations
High
Low

Battleground States

0.26 (0.44)

0.24 (0.43)

Non-battleground
States

0.04 (0.20)

0.03 (0.17)

Battleground
High Partisan
State Fixed Effects
Election Fixed Effects

Field Operations
Est (SE)
p
0.067
<0.001
(0.016)
0.030
<0.001
(0.007)
Included
Included

Candidate Advertising
High
Low
72,359.8
67,912.9
(189,133.3)
(168,686.9)
26,726.3
22,562.3
(94,848.9)
(99,165.3)
Candidate Advertising
Est (SE)
p
1.837
<0.001
(0.147)
-0.044
0.521
(0.069)
Included
Included

PAC Advertising
High
Low
12,695.8
10,560.0
(58,770.6)
(39,382.4)
7,901.8
2,901.6
(61,869.8)
(24,526.0)
PAC Advertising
Est (SE)
p
0.873
<0.001
(0.180)
0.206
0.015
(0.084
Included
Included

Note: The unit of observation is a county-party combination. Counties are classified into high versus low partisan levels depending on whether
the county-level percent of partisan voters is above or below the mean. In battleground states, there are 1,468 and 1,606 counties with high
and low partisan levels, respectively. In non-battleground states, there are 2,518 and 2,380 counties with high and low partisan levels,
respectively. The advertising variables are normalized for 100,000 age-eligible voters. For field operations, the percentage of counties that had
at least one field office is reported. The bottom portion of the table presents the linear regression results for the battleground indicator and
high partisan indicator after controlling for the state and election fixed effects. The dependent variable is each of the log-transformed campaign
variables.

34

 Table 5: Parameter Estimates
(1)

Field Operations
Candidate Advertising
PAC Advertising
Year 2008
Year 2012
Year 2008 X Democrat
Year 2012 X Democrat
Intercept

(2)

(3)

(4)

Est
(SE)
0.252***
(0.012)
0.014***
(0.001)
0.020***
(0.001)
-0.033***
(0.012)
-0.187***
(0.017)
-0.053***
(0.016)
-0.183***
(0.022)
-1.303***
(0.343)

Est
(SE)
0.072***
(0.006)
0.006***
(0.0004)
0.008***
(0.001)
-0.059***
(0.007)
-0.183***
(0.013)
0.123***
(0.010)
0.009
(0.018)
0.776
(0.629)

Est
(SE)
0.058***
(0.008)
0.034***
(0.004)
0.018***
(0.004)
-0.112***
(0.012)
-0.192***
(0.024)
0.114***
(0.012)
0.037
(0.032)
1.415**
(0.600)

Est
(SE)
0.078***
(0.007)
0.034***
(0.005)
0.026***
(0.004)
-0.121***
(0.013)
-0.298***
(0.027)
0.119***
(0.014)
0.126***
(0.036)
1.314**
(0.648)

Included
No
No
18,266
0.601

Included
Yes
No
18,266
0.955

Included
Yes
Yes
18,266
0.943

Included
Yes
Yes
18,266

Unobserved Heterogeneity (Sigma)
Control Variables
County-Party Fixed Effects
Instrumental Variables
Sample Size
R-squared

Partisanship
(SE)
0.135***
(0.018)
-0.019***
(0.002)
0.018***
(0.002)

0.251***
(0.059)
0.097
(0.476)

7,588

*** p<0.01; ** p<0.05; * p<0.10
Note: Column (1) reports the estimates of an OLS model without fixed effects, and column (2) reports estimates with
fixed effects. Column (3) reports the estimates with the use of instrumental variables. In specifications (1) through (3),
the dependent variable is the log ratio of the candidate vote share over the outside option. Column (4) reports the
estimates of the model, including voter heterogeneity in campaign effects in addition to the use of instrumental variables.
The column labeled “Partisanship” lists the estimated difference in campaign effects between partisan and non-partisan
voters. The intercept estimate under the “Partisanship” column corresponds to the main effect of partisanship. The table
reports abridged results for illustrative purposes. Table A2 in the Online Appendix reports the full results, including the
parameter estimates for the control variables: home state advantage for presidential candidates; home state advantage for
vice-presidential candidates; whether the governor is from the same party; whether the same-party governor is an
incumbent; the distance to campaign headquarters; percent of African Americans; population density; age group indicators
(18~24, 45~64, 65 and above); educational attainment indicators (less than high school, high school and associate, graduate
and above); median household income; unemployment rate; Gini index; median house value; dropout rate; poverty rate;
number of new employment positions; number of new establishments; weekly wage; and the interaction between the party
indicator and each of the control variables.

35

 Table 6: Campaign Effects by Voter Partisanship
Field Operations
Candidate Advertising
PAC Advertising

Average
0.078***
(0.007)
0.034***
(0.005)
0.026***
(0.004)

Non-partisan
0.011
(0.012)
0.043***
(0.005)
0.017***
(0.004)

Partisan
0.146***
(0.011)
0.025***
(0.005)
0.035***
(0.004)

Note: The effects are computed based on the parameter estimates from model specification 4 in Table 5.
*** p<0.01; ** p<0.05; * p<0.10

Table 7: RAS Model Framework
Type of electorate
Highly aware (partisan)
Moderately aware (unaffiliated)
Unaware (do not care about politics)

Receive message
Always
Occasionally
Never

36

Accept message
Not open to persuasion
Open to persuasion
Open to persuasion

 Table 8: Robustness Check: Alternative Models
Linear Coefficients
Field Operations ( E1 )

Model 1
Est
SE

Model 2
Est

SE

Model 3
SE

Est

Est

Model 4
SE

Model 5
SE

Est

0.093***

0.005

--

--

0.099***

0.005

0.111***

0.006

0.114***

0.005

Candidate Advertising ( J1 )

--

--

0.033***

0.005

0.028***

0.004

--

--

--

--

PAC Advertising ( S 1 )

--

--

0.028***

0.004

--

--

0.017***

0.003

--

--

Combined Advertising

--

--

--

--

--

--

--

--

0.021***

0.005

0.326

0.234

0.353***

0.058

0.349***

0.056

0.381***

0.058

0.339***

0.059

0.142***

0.015

--

--

0.121***

0.016

0.217***

0.016

0.163***

0.015

--

--

-0.024***

0.002

-0.013***

0.002

--

--

--

--

--

--

0.020***

0.002

--

--

0.005***

0.001

--

--

--

--

--

--

--

--

--

--

-0.008***

0.002

0.385**

0.160

0.097

0.475

0.090

0.463

0.096

0.424

0.099

0.411

Random Coefficients
Main Effect of Partisan ( D 2 )
Field Operations X Partisan ( E 2 )
Candidate Advertising X Partisan ( J 2 )
PAC Advertising X Partisan ( S 2 )
Combined Advertising X Partisan
Sigma
For Main Effect ( V )

Controls
County-Party Fixed Effects

Included
Included

Included
Included

Included
Included

Included
Included

Included
Included

Note: The alternative models differ from the full BLP model with regard to the inclusion of campaign variables. The remaining model specifications and model assumptions
are kept the same.
*** p<0.01; ** p<0.05; * p<0.10

37

 Table 9: Predicted Total Electoral Votes for Counterfactual Analyses

Actual
No field operations
No candidate advertising
No PAC advertising
PAC advertising rolled over
to candidate advertising

2004
Democratic Republican
252
277
221
308
238
291
253
276
253

2008
Democratic Republican
365
164
217
312
339
190
354
175

276

354

175

2012
Democratic Republican
332
198
227
303
332
198
374
156
358

172

Note: For the counterfactual scenario “no field operations,” each candidate was assigned with no field offices, keeping the
other campaigning activities unchanged. Similar steps were taken for the other counterfactual conditions, except for the
last counterfactual exercise, where the PACs donated their advertising money to the candidate, so that PAC advertising
became additional candidate advertising (respective of the relative advertising prices). The reported numbers are the total
electoral votes won by each party. The predicted winner of each election is in bold.

Table 10: Value of Democratic Field Offices in the 2008 Election
State
Florida
Indiana
North Carolina

Actual Field

Offices
58
18
49

Actual Ads
($M)

Additional Ads
($M)

$ 7.5
$10.2
$ 8.6

$2.99
$3.96
$9.34

Value of Office
($M)
$0.05
$0.20
$0.19

Note: The table reports the minimum additional advertising spending that would have allowed the Democrats to win the
state without field operations. The included states are those that the Democrats would have lost to the Republicans in
the absence of field operations.

38

 Figure 1: Vote Share versus
Campaigning

Figure 2: Effect of Campaigning for Low- and
High-Partisan Counties

(a)

(a)

.3

.3

.15

.15

0

0

-.15

-.15

-.3

-.3
-3

-2

-1

0

1

2

Change in Field Operations (logarithm)

3

High % Partisans
Low % Partisans
-3

-2

-1

.1

.15

.05

0

0

-.15

-.05

-.3

-.1
-5

0

5

Change in Candidate Advertising (logarithm)

10

-5

3

5

10

(c)
.1

.15

.05

0

0

-.15

-.05

-.3

-.1
0

0

Change in Candidate Advertising (logarithm)

(c)

-5

2

High % Partisans
Low % Partisans
-10

.3

-10

1

(b)

(b)
.3

-10

0

Change in Field Operations (logarithm)

5

Change in PAC Advertising (logarithm)

10

Note: Each dot corresponds to a county-party combination.
The line is the best-fitting fractional polynomial curve with
its 95% confidence interval shaded in gray. One unit was
added to the campaign variables before taking the logarithm.

High % Partisans
Low % Partisans
-10

-5

0

5

Change in PAC Advertising (logarithm)

10

Note: Each dot corresponds to a county-party combination. The solid
and dashed lines are the best-fitting fractional polynomial curve with its
95% confidence interval shaded in gray for counties with above and
below the average percentage of partisans, respectively. To better exhibit
the patterns, (b) and (c) focus on the middle range of the vertical axis,
which excludes only 1% of the data.

39

 Figure 3: Campaign Allocation by States

Note: The plot depicts the campaign allocations between the Democrats and the Republicans, split by battleground and
non-battleground states. Panel (a) is for field operations; (b) is for candidate and party advertising; and (c) is for PAC
advertising. Campaign variables are normalized for 100,000 age-eligible voters. Advertising variables are on the log scale.
The diagonal reference line is y=x.

40