Bias in Cable News: Real Effects and Polarization

NBER WORKING PAPER SERIES

BIAS IN CABLE NEWS:

REAL EFFECTS AND POLARIZATION

Gregory J. Martin

Ali Yurukoglu

Working Paper 20798

http://www.nber.org/papers/w20798

NATIONAL BUREAU OF ECONOMIC RESEARCH

1050 Massachusetts Avenue

Cambridge, MA 02138

December 2014

We thank Tom Clark, Greg Crawford, Ruben Enikopolov, Matthew Gentzkow, Ben Golub, Marit

Hinnosaar, Kei Kawai, Robin Lee, Claire Lim, Paul Oyer, Ariel Pakes, Jesse Shapiro, Michael Sinkinson,

Gaurav Sood, Daniel Stone, and seminar and workshop participants at Boston College, Boston University,

the BFI Media and Communications Conference, Emory, Harvard, NYU Stern, Stanford GSB, the

Wallis Political Economy Conference, and the Workshop on Media Economics for comments and

suggestions, and Carlos Sanchez-Martinez for excellent research assistance. The views expressed herein

are those of the authors and do not necessarily reflect the views of the National Bureau of Economic

Research.

NBER working papers are circulated for discussion and comment purposes. They have not been peer-

reviewed or been subject to the review by the NBER Board of Directors that accompanies official

NBER publications.

exceed two paragraphs, may be quoted without explicit permission provided that full credit, including

Bias in Cable News: Real Effects and Polarization

Gregory J. Martin and Ali Yurukoglu

NBER Working Paper No. 20798

December 2014

JEL No. D72,D83,L82

ABSTRACT

We jointly measure the persuasive effects of slanted news and tastes for like-minded news. The key

ingredient is using channel positions as exogenous shifters of cable news viewership. Local cable positions

affect viewership by cable subscribers. They do not correlate with viewership by local satellite subscribers,

who are observably similar to cable subscribers. We estimate a model of voters who select into watching

slanted news, and whose ideologies evolve as a result. We estimate that Fox News increases the likelihood

of voting Republican by 0.9 points among viewers induced into watching four additional minutes per

week by differential channel positions.

Gregory J. Martin

Tarbutton Hall

1555 Dickey Dr.

Atlanta, GA 30322

[email protected]

Ali Yurukoglu

Graduate School of Business

Stanford University

Stanford, CA 94305

and NBER

[email protected]

1 Introduction

Political observers across the ideological spectrum routinely make allegations of media

bias and its detrimental eﬀect on society. Two of the three 24-hour cable news channels,

the Fox News Channel and MSNBC, are frequent targets of such allegations. In this

paper, we address two questions about cable news. First, how much does consuming

slanted news, like the Fox News Channel, alter the propensity of an individual to

vote Republican in Presidential elections, if at all? Second, how intense are consumer

preferences for cable news that is slanted towards their own ideology?

The answers to these questions are key inputs for designing optimal public pol-

icy towards the media sector. If consumers simply prefer news that resonates with

their pre-existing ideology, as in Mullainathan and Shleifer (2005) and Gentzkow and

Shapiro (2010), then the news media sector is similar to any other consumer product,

and should be treated as such by public policy. However, if consuming news with a

slant alters the consumer’s ideology, then public policy towards the news media sector

becomes more complex.

In particular, if news consumption alters ideology, and con-

sumers have a taste for like-minded news, then the existence of slanted news could lead

to a polarizing feedback loop: an “echo chamber” where partisans can reinforce and

strengthen their initial biases.

Furthermore, an interested party could potentially in-

ﬂuence the political process by owning or controlling media outlets.

Such concerns led

the Federal Communications Commission (FCC) to condition approval of the merger

of Comcast Corporation and NBC Universal in 2010 on the requirement that Comcast

take steps to promote independent news services.

Diﬀerentiating the taste mechanism from the inﬂuence mechanism is diﬃcult in ob-

servational data. The analyst observes a positive correlation between the propensity to

Gentzkow and Shapiro (2008) detail the complexities in designing optimal regulatory policy for media

markets.

Gentzkow and Shapiro (2011) indicate that current media consumption tends to be balanced across

ideologically slanted sources. This paper identiﬁes trends suggesting that the “echo chamber” scenario may

be increasing in relevance.

Existing evidence from Gentzkow and Shapiro (2010) shows that owner partisanship is not an important

determinant of newspaper slant. The sample size is too small to test this hypothesis in the cable news case.

The condition required that Comcast move “independent” news channels such as Bloomberg Television

into “news neighborhoods.” This eﬀectively required Comcast to move Bloomberg next to channels such as

MSNBC and CNN in their channel lineups.

vote Republican and hours spent watching Fox News. Were Fox News viewers already

predisposed to vote Republican, and the observed correlation driven by preference for

watching like-minded news? Or were some fraction of those viewers persuaded to vote

Republican as a consequence of watching Fox News?

The essential ingredient in our analysis is the use of the channel positions of news

channels in cable and satellite television lineups as instrumental variables. Variation

in channel positions causes some viewers to watch more or less of these channels. We

use the corresponding induced variation in time watched to estimate whether or not

watching slanted news changes voting behavior. We estimate that watching the Fox

News Channel (at its current ideological positioning) for four additional minutes per

week

increases the probability of intending to vote for the Republican presidential

candidate by 0.9 percentage points for voters induced into watching by variation in

channel position. Watching MSNBC (at its current ideological positioning) for four

additional minutes per week increases the probability of intending to vote for the

Democratic presidential candidate by about 0.7 percentage points for voters induced

into watching by variation in channel position.

As with any instrumental variables design, it is critical that the channel positions

for Fox News and MSNBC are exogenous, and not chosen to accord with local political

tastes. In Section 2, we describe turbulence in the cable industry in the years 1994-2000

that induced as good as random variation in channel positions across locations. We

then directly test and conﬁrm the validity of the instrument by demonstrating that

the local cable channel positions of Fox News and MSNBC correlate strongly with the

channels’ viewership among cable subscribers but do not correlate with viewership by

satellite television subscribers in the same zip code.

Satellite viewership provides a useful placebo test because the two satellite providers

each use a single nationwide channel position lineup; thus, satellite subscribers’ viewing

decisions cannot be directly inﬂuenced by the local cable operator’s choice of lineup.

However, satellite subscribers’ observable characteristics and viewing tastes are highly

positively correlated with the characteristics and tastes of cable subscribers in the same

zip code. If the same correlation holds for unobservable political ideology, and if cable

Approximately four minutes per week is the additional time spent watching Fox News associated with

moving from a cable system with Fox News channel position at the 75

percentile of the distribution to one

with Fox News channel position at the 25

percentile.

companies endogenously choose channel positions to suit local political leanings, then

the channels’ positions on cable systems should predict viewing by satellite subscribers

in the same zip code. Our data reject this hypothesis. Across the set of news channels,

the coeﬃcient estimates of cable channel positions on satellite viewing are close to zero.

Statistically, we can (1) not reject that eﬀects of the cable positions on satellite viewing

are zero, and (2) reject that the eﬀects of cable positions on satellite viewing are equal

to the eﬀects of cable positions on cable viewing.

Our approach to quantifying the second object of interest, the preference for like-

minded news, follows Gentzkow and Shapiro (2010), who estimate this quantity in the

context of US daily newspapers. We ﬁrst place the cable news channels on the ideo-

logical spectrum by quantifying how similar the language employed by the channels is

to the language employed by individual members of Congress. This method provides

a measure of ideological slant for each channel in each year. We measure the relation-

ship between changes in the slant measure over time and the characteristics of viewers

of these channels. A key source of variation in this exercise is MSNBC’s change in

business strategy towards oﬀering more explicitly liberal content around 2006. Our

ideology estimates pick up this format switch - MSNBC closely tracks CNN in the

early 2000s, but then moves left following the format switch in 2006. We estimate

Fox News’ ideology to the right of CNN throughout the sample period, although the

distance between the two has been widening in the most recent years.

We conduct the analysis of these two questions in a uniﬁed framework. We model

consumer-voters who choose how much time to spend watching the cable news chan-

nels; whether to subscribe to cable, satellite or nothing at all; and for whom to vote

in presidential elections. Consumers’ allocation of time to television channels is gov-

erned by their preferences for the channels (which are a function of their ideology,

the channels’ ideologies, and their demographics), and the availability of the channels

(whether the cable operator carries them and, if so, the positions they occupy on the

channel lineup). Consumers’ ideologies evolve from their initial position depending on

how much time they allocate to watching channels of diﬀerent ideologies. This process

culminates in a presidential election in which consumers choose for whom to vote.

We estimate the parameters of the model by simulated indirect inference. The cri-

terion function is the distance between two-stage least squares estimates of intention

to vote on demographics and hours watched of each channel, using channel positions

as instrumental variables, in the actual data and in data simulated from the model. In

addition to matching the second stage regression coeﬃcients, we also match the ﬁrst

stage (viewership equation) regression coeﬃcients and the “mis-speciﬁed” OLS regres-

sion coeﬃcients. We use data covering 1998 to 2008 from multiple sources including

(1) high quality channel lineup data that provides channel positions and availability by

zip code, provider, and year, (2) individual level viewership data on hours watched by

channel and year together with demographics and zip code, (3) individual level survey

data on intent to vote Republican in presidential elections together with demographics

and zip code, (4) county-level presidential vote shares, (5) broadcast transcripts of Fox

News, CNN, and MSNBC by year, and (6) the Congressional record by year.

We use the estimated model to quantitatively assess the degree of ideological polar-

ization induced by cable news and separately the eﬀect of the entry of Fox News prior

to the year 2000 election. We ﬁnd that cable news does increase polarization among

the viewing public. Furthermore, the increase in polarization depends critically on the

existence of both a persuasive eﬀect and a taste for like-minded news. We estimate

that removing Fox News from cable television during the 2000 election cycle would

have reduced the average county’s Republican vote share by 1.6 percentage points.

This paper contributes to the empirical literature on the causes and eﬀects of the

news media, particularly regarding political outcomes.

The closest papers to this

study are by DellaVigna and Kaplan (2007) and Gentzkow and Shapiro (2010).

DellaVigna and Kaplan (2007) seek to study the eﬀects of Fox News by comparing

vote shares in locations with and without cable access to Fox News by November

2000, as partially measured by the Cable and Television Factbook data set. Our

contribution to this strand of the literature is to introduce a new and more credibly

exogenous source of variation, channel positions, to measure the eﬀects of Fox News as

well other cable news like MSNBC, using correct data.

The use of channel positions

A number of papers have demonstrated that media usage or availability aﬀects behavior. Amongst

others, Chiang and Knight (2011) ﬁnd positive eﬀects of unexpected newspaper endorsements on vote shares

for the endorsed candidate, Gentzkow (2006) ﬁnds decreased voter turnout from television access, Gerber

et al. (2009) ﬁnd positive eﬀects of newspaper exposure, regardless of slant, on Democratic vote shares in

the 2005 Virginia gubernatorial elections. Enikolopov et al. (2011) ﬁnd that viewing an independent news

channel in Russia increased vote shares for the opposition parties and decreased overall turnout in 1999.

Lim et al. (2014) ﬁnd that media coverage can aﬀect criminal sentencing decisions for judges.

In Appendix A, we document that the data set used in DellaVigna and Kaplan (2007) has severe mis-

measurement of Fox News availability. Nearly 40% of the “control group,” the locations that they consider

as instrumental variables could be useful for studying the eﬀects of media consumption

in other contexts. In terms of results, we estimate a Fox News eﬀect that is statistically

positive and quantitatively large whereas the DellaVigna and Kaplan (2007) analysis,

updated to use the correct channel availability data, is inconclusive. We also ﬁnd a

large MSNBC eﬀect in 2008.

Our approach follows Gentzkow and Shapiro (2010) in several dimensions, includ-

ing the use of text analysis to measure media outlet slant. Like Gentzkow and Shapiro

(2010), we treat that measure as a characteristic over which consumers have heteroge-

neous tastes when choosing media consumption levels. Our contribution is to model

media consumption together with voting to separately measure tastes for like-minded

news and the inﬂuence of slanted media consumption on consumer ideology. The inﬂu-

ence eﬀect also interacts with the existence of tastes for like-minded news. Consumers

for whom both eﬀects are present can be induced into feedback loops where they con-

sume slanted media, their ideologies then evolve in the direction of the slant, then their

taste for that slanted media increases, and so on in a loop.

2 Institutional Overview

During our study period of 1998-2008, most households had three options for television

service: subscribe to a cable (that is, a wire-based transmission) package, subscribe

to a satellite television package, or subscribe to neither and receive only over-the-air

broadcast signals.

In 2000, the vast majority of cable or satellite subscribers were

cable subscribers, but by 2008, satellite providers had a market share of about 30%.

Diﬀerent locations have diﬀerent cable providers such as Comcast, Time Warner Cable,

Cox, Cablevision, or Charter. The set of channels, or content, in a cable package varies

across providers and within providers across locations. A typical set of cable packages

as not having cable access to Fox News in 2000, did in fact have cable access to Fox News. 25% of the

control group had Fox News availability since 1998. Their data set simply had not been updated to reﬂect

Fox News’s arrival in those locations. We detail how their results change upon correcting the measurement

error in Appendix A.

Some households, for example households in remote rural areas, did not have a cable option. Some

households which did not have a direct line of sight due to physical obstructions like tall buildings, trees,

or steep slopes, did not have a satellite option. And some households, mostly in urban areas, had two

wire-based cable operators.

would have one Basic package which retransmits the over-the-air signals, an Expanded

Basic package which includes the top 40 to 80 cable channels such as ESPN, USA, TNT,

CNN, Nickelodeon, MTV, Comedy Central, and similar, and a digital package which

oﬀers more niche content like the DIY Channel or the Tennis Channel. Throughout the

period, there were two nationwide satellite providers: DirecTV and the Dish Network.

Each satellite provider oﬀers the same channel lineup and packages in all locations.

Cable content is produced by media conglomerates such as Viacom, News Corpora-

tion, ABC-Disney, or NBC Universal. The cable and satellite providers contract with

these ﬁrms to oﬀer their content to subscribers. This bilateral contracting is the focus

of Crawford and Yurukoglu (2012), which provides more detail about the industry’s

structure. There was some vertical integration during our sample period: News Cor-

poration had a controlling interest in DirecTV, and Time Warner and Time Warner

Cable were integrated.

The foci of this study are the cable news channels CNN, the Fox News Channel,

and MSNBC. CNN began broadcasting in 1980 as one of the earliest cable channels of

any genre, and pioneered the 24 hour news channel format. It was acquired by Time

Warner in 1996. CNN does not have any explicit ideological orientation.

The Fox

News Channel and MSNBC both entered the market in the mid 1990’s. Launched

by the News Corporation in late 1996, Fox News Channel’s business strategy was to

provide news with a more conservative slant. This strategy and the perception of such

a slant continues today. Fox News has become one of the most highly rated cable

channels across all genres. It is a cultural force in the U.S. synonymous with media

bias and the mixing of news and entertainment. MSNBC began as a joint venture

between NBC and Microsoft. At the outset, MSNBC did not have any explicit slant.

MSNBC changed its business strategy in the mid-2000’s to provide news with a more

liberal slant (Sanneh, 2013).

The channel lineup, or the numerical ordering of channels, varies by local cable

system. In most cases the ﬁrst ten to twenty channel positions are allocated to the

over-the-air broadcast aﬃliates. For example, NBC4 occupies channel position 4 in

Washington D.C. area cable systems. After the over-the-air channels, the cable chan-

In our analyses we make no assumption that CNN’s content is neutral or moderate, and treat it symmet-

rically with MSNBC and Fox. We apply the same text-based measures to estimate its ideology, and ﬂexibly

estimate its eﬀect on viewers’ political preferences.

nels begin. We assert in this paper that the ordering of a channel in the lineup can

have signiﬁcant eﬀects on the viewership of news channels.

The obvious empirical concern is that a channel might be placed in lower positions

in localities with high tastes for the channel. We later examine that concern with a

placebo test of whether satellite subscriber viewership is correlated with local cable

channel positions. Describing the process by which channel positions were determined

provides additional support for the claim that channel positions are valid instruments.

The 1994-2001 period during which Fox News and MSNBC were rolling out was

a tumultuous time for the cable television industry. This period saw many systems

upgrade from older analog to newer digital equipment, expanding the number of chan-

nels cable operators were able to oﬀer their subscribers. Coincident with this technical

advance, a wave of new channels (including the two cable news channels that are the

focus of this paper) entered cable lineups alongside ﬁrst-generation channels like CNN,

ESPN, MTV, TBS, TNT and the USA Network.

However, the timing of the advances in content and technology were not coordi-

nated: some systems invested in upgrades early, before the wave of new channel entry,

and some later. Meanwhile, bilateral deals for content distribution were being struck

between the numerous new channels and cable system operators, of which in this period

before the early-2000s wave of consolidation there were many. As a result, the channel

positioning that Fox or MSNBC ended up with on a given local system depended on

the speciﬁc timing of that system’s negotiations with multiple new channels as well as

its decision of when to upgrade. Channels were often allocated positions sequentially,

in the order in which they were added to a system.

Combined with the key principle

in lineup design of limiting the changes in channel positions as much as possible so as

to not confuse existing customers, these chaotic historical factors generated persistent

cross-system variation in the positioning of Fox News and MSNBC.

Figure 1 plots the growth in subscribers for a group of peer channels during this

The signiﬁcant relationship between channel position and viewership holds for all genres, not just news.

In Appendix C, we show that channel positions correlate with the best available position in the year

before a channel was added.

Some systems have shuﬄed positions over time as channels went out of business, as channel capacity

expanded and as new channels came online. Some local managers pursued a strategy of moving channels

with similar content or in the same genre together into “neighborhoods,” when possible. In general, however,

the ordering of cable channels is highly persistent from year to year.

time period. The top line shows ESPN, which was available on virtually every cable

system. The other channels in the graph all experienced substantial growth during this

time period. Idiosyncracies in the timing of contracts and system upgrades created

variation in channel positions for a given channel across locations. In some cases, if

Fox News was being added to a system facing capacity constraints, its channel position

was determined by the position of the channel it was replacing. On systems owned by

the multiple-system operator TCI in 1996, Fox News was reported to have replaced one

of as many as twelve diﬀerent channels depending on the location (Dempsey (1996)).

0 20 40 60 80

National Subscribers (M)

1994 1996 1998 2000 2002

Year

Animal Planet BET Bravo

Cartoon Network Comedy Central E! Entertainment TV

ESPN ESPN 2 FX

Food Network Fox News Channel History Channel

MSNBC SyFy, Sci−Fi truTV, Court TV

Figure 1: Number of subscribers for a group of peer channels by year for the period 1994-

2001. National subscriber numbers according to SNL Kagan data.

3 Data

We use seven categories of data sets: (1) Nielsen FOCUS data on cable channel lineups

by zip code by year, (2) the National Annenberg Election Survey (NAES) and the

Cooperative Congressional Election Study (CCES, Ansolabehere (2011)) on individual

demographics, zip code, and intent to vote Republican in 2000, 2004, and 2008 U.S.

Presidential elections, (3) Mediamark and Simmons individual survey data on hours

spent watching cable news by channel, individual demographics, and zip code, (4)

County level presidential election vote share data compiled by Congressional Quarterly,

(5) U.S. Census demographics by zip code, (6) Broadcast transcripts of cable news from

Lexis-Nexis, and (7) the Congressional Record. We now describe each data set and

exposit several empirical relationships that are central to our results. Most of our

analysis focuses on the years 2000 to 2008, but some data sets cover through 2011.

3.1 Cable Lineups: Nielsen FOCUS

The Nielsen FOCUS database consists of yearly observations on cable systems. The

key variables in this data set are, for each system and year, the availability of CNN,

Fox News, and MSNBC, the channel positions of CNN, Fox News, and MSNBC, when

available, and the zip codes served by the system. In Figure 2, we document the

availability of each of these news channels by year. CNN was already near-universal

by 1998. Fox News and MSNBC became widespread over the sample period. Table

1 presents the mean and standard deviation of channel position for each of the three

news channel by year. CNN is generally lower than Fox, which is generally lower than

MSNBC. The pairwise correlations in positions of the channels are around .3, with

variation depending on the channel pair and the year. As seen in Figure 4, Fox News

Channel is in a lower position than CNN in 11% of the observations when both are

available. MSNBC is in a lower position than CNN in 7% of the observations, and in a

lower position than Fox News Channel in 23% of the observations. Table 2 shows the

corresponding Republican vote shares and cable news viewership levels conditional on

lowest ordered channel. The average Republican vote share is higher and the average

hours of Fox News watched are higher when Fox News is in the lowest position.

There are two important facts about this data set. First, the Nielsen FOCUS

database contains the universe of cable systems. Second, all entries are updated on a

regular basis. This feature is diﬀerent from the Cable and Television Factbook used in

previous studies. We detail this important diﬀerence in Appendix A.

Figure 2: Availability of cable news channels by year. The solid lines represent the fraction

of cable subscribers for whom the news channel was carried on their system. The dashed

lines represent the fraction of cable systems which carry the news channels. By 2002, nearly

all cable subscribers had access to Fox News and MSNBC.

3.2 Individual Voting Data: NAES and CCES

The National Annenberg Election Study (NAES) is a large-scale phone survey con-

ducted each presidential election cycle which asks individual respondents a range of

political preference questions, along with demographic identiﬁers. We use data from

the 2000, 2004, and 2008 election cycles, including the conﬁdential zip-code ﬁeld. The

key variables are demographic variables such as race, age, and income; zip code; stated

ideology; and actual or intent to vote in the current presidential election. The NAES

surveys were conducted on a rolling basis over the course of each election, with most

Year CNN FNC MSNBC

Mean SD Mean SD Mean SD

2000 17.34 10.49 35.40 12.94 39.00 12.76

2004 21.83 12.83 38.22 14.31 41.66 13.53

2008 24.67 14.40 37.48 14.42 42.87 18.18

All 21.14 13.06 37.28 14.17 41.59 15.68

Table 1: Mean and standard deviation of channel position by news channel by year in election

years.

0 .01 .02 .03 .04

Density

0 25 50 75 100 125 150

Channel Position

CNN

Fox News

MSNBC

Figure 3: Kernel Density Estimate of Distribution of Channel Positions by Channel

Figure 4: Fraction of systems with certain ordinal conﬁgurations.

CNN Lowest FNC Lowest MSNBC Lowest All

Republican Vote Share 0.5023 0.5082 0.4900 0.5017

Fox Hours Watched 1.0556 1.1440 1.0608 1.0699

CNN Hours Watched 1.2078 1.2657 1.2471 1.2222

MSNBC Hours Watched 0.5219 0.5279 0.5690 0.5294

Table 2: Mean Republican presidential vote shares and cable news hours watched by lowest

ordered cable news channel.

respondents contacted before election day but some after. We combine actual vote

(from respondents contacted after election day) together with intent to vote (from

those contacted before) into a single variable.

The 2004 and 2008 NAES surveys also asked respondents to report their “most

watched” cable news source, if any. We use this variable in estimating OLS regressions

of vote intention on channel viewership.

These data are summarized in Table 3. For 2008, we add data from the Cooperative

Congressional Election Study (CCES) on the same variables that we use from the

NAES.

3.3 Individual Viewership Data: Mediamark and Sim-

mons

Mediamark and Simmons are two commercial data vendors who survey individuals on

their usage of diﬀerent brands, including media usage. We use Mediamark for 2000

to 2007, and Simmons for 2008 to 2011. The key variables for our study are year, zip

code, individual demographics, whether the respondent subscribes to cable, satellite,

or neither, and the reported number of hours watched per week of CNN, Fox News

Channel, and MSNBC. These data are also summarized in Table 3.

3.4 County Level Vote Shares and Census

We use county level presidential vote shares for the Presidential elections in 2000, 2004,

and 2008 obtained from the Voting and Elections Collection Database maintained by

Congressional Quarterly. We also use zip code level demographic statistics from the

2000 and 2010 US Census. We use these data to construct county-level distributions

of household income, age, race, education, and initial ideology, from which we draw a

set of simulated consumer-voters for the model of section 5. For zip codes which span

multiple counties, we split the zip code across the relevant counties in proportion to

the county size.

3.5 Broadcast Transcripts and Congressional Record

To quantify the slant of each news channel in each year, we follow Gentzkow and

Shapiro (2010)

in comparing the language that the channels use to language that

Congresspeople use. We modify their statistical procedure as well as create scores for

each channel for each year. This procedure does not recognize irony, satire, sub-text,

nor tone, and thus likely underestimates the true dispersion in slant as the slanted

outlets sometimes employ the language of the other side of political spectrum for pur-

poses of mockery or derision.

We obtained broadcast transcripts for CNN, Fox News

The idea is similar in spirit to Groseclose and Milyo (2005)

This is one reason why we exclude Comedy Central, which features two prominent slanted cable news

programs, The Daily Show with Jon Stewart and The Colbert Report, from the analysis. Their slant relies

heavily on satire and is not as reasonably quantiﬁed based on phrase usage. As a separate matter, Comedy

Central has other highly viewed shows which are not explicitly political such as South Park, and our data

N Mean SD

NAES/CCES

Male 122243 0.460 0.498

White 122243 0.841 0.365

Black 122243 0.082 0.275

Hispanic 122243 0.067 0.251

Age 122243 47.801 16.071

College Graduate 122243 0.369 0.482

Household Income ($000s) 122243 65.312 50.796

Intent to Vote: Republican 122243 0.498 0.500

Self Reported Ideology 119524 3.185 1.042

Mediamark/Simmons

Male 209352 0.505 53.130

White 209352 0.812 15.738

Black 209352 0.106 0.390

Hispanic 209352 0.109 0.308

Age 209352 46.314 0.312

College Graduate 209352 0.328 0.475

Household Income ($000s) 209352 70.744 0.303

Cable Subscriber 209352 0.656 0.261

DirecTV Subscriber 209352 0.102 0.469

Dish Network Subscriber 209352 0.074 0.500

Watch any CNN 209352 0.360 0.480

CNN Hours per Week 209352 1.223 2.816

Watch any Fox News Channel 209352 0.291 0.454

Fox News Channel Hours per Week 209352 1.070 2.847

Watch any MSNBC 209352 0.203 0.402

MSNBC Hours per Week 209352 0.529 1.726

Table 3: Summary Statistics for individual level NAES/CCES and Mediamark/Simmons

data. Ideology ranges from Very Liberal (1) to Very Conservative (5). The variables except-

ing age, household income, and the hours per week are dummy variables.

Channel, and MSNBC from the Lexis-Nexis database for the sample period 1998-2012

by downloading all transcripts per year for each identiﬁable cable news program from

each of the three channels.

Taking the Congressional Record for each year, the ﬁrst steps are to stem the

words, remove stop words, and then count the frequency of usage of two word phrases

by each Congress person. Each Congress person has a measure of their ideology, the

DW-NOMINATE score from McCarty et al. (1997), which places them on the interval

[−1, 1] with more positive being more conservative. The second step is to correlate

phrase usage with the DW-Nominate score. There are many more two word phrases

than Congresspeople, and an ordinary least squares criterion is therefore useless because

there are more variables than observations. For each year, we run an Elastic Net (Zou

and Hastie, 2005) regression of DW Nominate score of frequency of phrase usage where

an observation is a Congressperson. The Elastic Net regression is a variable selection

algorithm that combines the LASSO and the Ridge Regression regularization penalties.

In Table 4, we follow Gentzkow and Shapiro (2010) in showing a subset of the most

indicative partisan phrases selected by the Elastic Net regression for 2000, 2004, and

2008.

We use the estimated coeﬃcients to predict the DW-Nominate score for each cable

news channel in each year. We then apply a three period moving average smoothing

ﬁlter. The results of this procedure are in Figure 5. Fox News is consistently more

conservative than the other two channels. MSNBC switches to being more liberal in

the mid-2000’s. The estimates also reveal increased polarization of cable news over

time. The text based measures produce estimated ideologies for the channels that are

more moderate than the median members of each party. In the modelling to come, we

allow for consumers to perceive these news channels to be more or less ideologically

diﬀerentiated, in proportion to these estimates. Indeed, our estimates for this scale

factor put Fox News Channel very close to the median Republican voter.

are aggregated to the channel level.

Figure 5: Estimated Ideology by Channel-Year

2000 Party 2004 Party 2008 Party

60 minut R 17 month D 11 countri R

administr fail R administr continu D allow vote R

american coupl R administr fail D american resourc R

bank credit D administr refus D approv rate R

beneﬁt wealthiest D administr republican D bring skyrocket R

big bank D administr want D bush chenei D

break wealthi D american without D bush took D

break wealthiest D bid contract D busi come D

bush tax D billion iraq D busi todai D

busi come R bipartisan commiss D call abort R

caught nap R bush budget D can produc R

child tax R busi come R capit gain R

continu everi D civil justic R control two D

coupl pai R compani hmo D cost energi R

cut wealthi D corpor proﬁt D dai spent R

cut wealthiest D cost energi R death tax R

eight billion R cost war D deep sea R

elderli peopl D cut wealthi D deep water R

elimin death R cut wealthiest D democrat bill R

feder bureaucraci R date time R democrat colleagu R

follow morn R don nickl R develop resourc R

gun hand D econom advis D entitl reform R

hard earn R evil empir R explor oil R

huge tax D fall far D kill littl R

increas domest R far short D liquid fuel R

labor right D gdp growth R major parti R

largest tax R govern regul R minimum wage D

larg tax D govern spend R never express R

line vote D hold line R new nuclear R

live poverti D increas medicar D new reﬁneri R

modern school D invas iraq D nuclear plant R

monei washington R job administr D pelosi said R

name peopl D largest deﬁcit D plan bring R

need prescript D lawrenc v R process law R

pai social R liabil cost R produc american R

per child R major want D product american R

pm todai R marriag licens R properti without R

presid busi R marriag will R record proﬁt D

reduc tax R mass grave R reﬁneri capac R

reproduct health D medicar premium D remind realli R

republican bill D million manufactur D safe wai R

republican friend D presid aristid D sign petit R

republican propos D presid econom D soon on R

right organ D print report R sue opec R

riski scheme R protect tradit R suppli energi R

sensibl gun D reserv us R tax american R

seven million R revenu feder R tax burden R

sinc columbin D reward compani D tax hike R

state arizona R sector job D tax oil R

still republican D servic author R thing common R

tax death R ship job D took oﬃc D

tax hike R social justic D trillion barrel R

tax just R time administr D trillion cubic R

unidentiﬁ male D trillion surplu D unfund liabil R

us later R univers health D v wade R

wealthiest american D war cost D warrantless surveil D

wealthiest peopl D wealthiest american D wast spend R

work condit D woman right D without due R

year administr D yet todai R

Table 4: Top decile of partisan indicative phrases selected by Elastic Net for years 2000,

2004, and 2008. Word variants are stemmed to their roots.

4 Regression Analysis

This section provides the regression results on the relationship between channels po-

sitions, watching Fox News and MSNBC, and voting Republican. These results serve

as the basis for the model estimation and simulation in the following sections. The re-

sults in this section do not depend on the behavioral model estimation that we specify

in Section 5. This section can be read and evaluated as a stand-alone instrumental

variables regression analysis.

The ﬁrst stage regressions are channel viewership against channel positions, chan-

nel availability, year eﬀects, and demographics. The second stage regression is intent

to vote Republican against predicted viewership, channel availability, year eﬀects, and

demographics.

We present the placebo ﬁrst stage regression of channel viewership

by satellite subscribers against cable channel positions, cable channel availability, year

eﬀects, and demographics.

We also present the OLS regression of intent to vote

Republican against channel viewership, channel availability, year eﬀects, and demo-

graphics because the OLS regression is also relevant for the model estimation.

4.1 Viewership on Channel Positions: First Stage with

Individual Level Data

The ﬁrst stage describes how hours watched varies with channel position. We use

channel-year ﬁxed eﬀects to ask whether a given channel has more viewership when it

is in a lower position. The idea is that lower channel positions induce more viewership

for channels such as Fox News and MSNBC because the higher watched content tends

to be in lower channel positions for historical reasons.

The most obvious mechanism

to generate such an eﬀect is a costly search model. Consider a viewer who just ﬁnished

watching a television program, and begins to search for a new program. Their search

will begin from the channel they were watching, which is likely to be in a low position.

The variation in viewership attributable to diﬀerential availability of Fox News and MSNBC is thus not

part of the estimation strategy.

This serves to argue the validity of channel positions as instrumental variables because if channel positions

were tailored to local tastes, they should correlate with viewership of satellite subscribers in the same zip

code.

In addition to the broadcast networks ABC, CBS, Fox, and NBC, the lower channel positions are occupied

by the earliest entrants into cable (eg ESPN, MTV, TNT, USA) which also have high viewership.

They will move away from that channel, thereby making it more likely they stop nearer

to that channel than further away.

The obvious worry is that channel positions are

tailored to local tastes so that channels which will be watched more often are easier to

ﬁnd. We address this concern with a placebo test that shows that local cable positions

do not correlate with viewership by satellite subscribers in the same zip code.

There are two regressions for each channel to model explicitly the mass of consumers

who watch zero hours of a given news channel, which is a salient feature of the data.

The median individual hours watched for each news channel is zero. One regression

for each channel is a linear probability model for whether one watches the channel at

all, or not. The second regression for each channel is for how long one watches the

channel, conditional on watching at all.

= δ

+ a

+ η

+ θ

c,CNN

CNN

+ θ

c,F NC

F NC

+ θ

c,MSN BC

MSN BC

+ e

ict

+ ˜a

+ ˜η

c,CNN

CNN

c,F NC

F NC

c,MSN BC

MSN BC

+ ˜e

ict

where p

is the logged channel position of j in individual i’s zip code in year t, δ

and

are channel-year ﬁxed eﬀects, x

are individual level demographics, and a

and

˜a

are ﬁxed eﬀects for availability of the cable news channels to individual i. We choose

a log functional form for the position eﬀects on the basis of the empirical relationship

between position and viewership. Figure 6 shows the relationship in the full set of

96 channels measured by MediaMark. We ﬁrst regress hours watched on demographic

attributes of the respondents plus ﬁxed eﬀects for each channel and each year in the

dataset, and plot the residual hours watched against ordinal channel position. The

resulting relationship is negative, with a slope that steadily diminishes as channel

position increases.

To simplify the presentation, we present in Table 5 ﬁrst stage estimates for re-

gressions of hours watched on channel positions without modeling the mass of zero

viewership.

Own channel position has the expected sign and is strongly statistically

Bias to the top of a list or default option in search is documented in eye tracking studies for yellow pages

(Lohse (1997)) and survey response (Galesic et al. (2008)). There is a theoretical literature in economics

modelling such behavior (see Rubinstein and Salant (2006), Horan (2010), Masatlioglu and Nakajima (2013),

and the literature on status-quo bias more generally.)

For the 2SLS estimates, we model the ﬁrst stage as the combination of a linear probability model of

●

−0.25

0.00

0.25

0.50

0.75

0 50 100 150

Ordinal Channel Position

Residual Ratings (Hours)

Figure 6: Residual hours watched (after removing individual demographic eﬀects plus chan-

nel and year ﬁxed eﬀects) of all 96 channels in the MediaMark dataset, plotted against the

channel’s ordinal position in the lineup.

(1) (2) (3)

VARIABLES Fox News Hours CNN Hours MSNBC Hours

log(Fox News Position) -0.122*** -0.0309 0.0499***

(0.0212) (0.0215) (0.0134)

log(CNN Position) -0.0106 -0.111*** -0.0148*

(0.0136) (0.0138) (0.00859)

log(MSNBC Position) 0.128*** 0.0721*** -0.0969***

(0.0230) (0.0233) (0.0145)

HH Income 0.0116*** 0.00677*** 0.00535***

(0.00134) (0.00137) (0.000847)

HH Income

-0.000110*** -4.86e-05*** -4.23e-05***

(1.41e-05) (1.43e-05) (8.87e-06)

HH Income

2.86e-07*** 1.22e-07*** 1.08e-07***

(4.04e-08) (4.10e-08) (2.54e-08)

Age -0.00282 -0.0137*** 0.00229

(0.00357) (0.00363) (0.00225)

Age

0.000340*** 0.000531*** 8.20e-05***

(3.76e-05) (3.82e-05) (2.37e-05)

White 0.175*** -0.329*** -0.0981***

(0.0336) (0.0341) (0.0212)

Black 0.268*** 0.0934** 0.0231

(0.0416) (0.0423) (0.0262)

Hispanic -0.256*** -0.184*** -0.0997***

(0.0301) (0.0306) (0.0190)

College -0.124*** 0.287*** 0.113***

(0.0186) (0.0189) (0.0117)

Male 0.245*** 0.192*** 0.120***

(0.0165) (0.0168) (0.0104)

log(N Channels) -0.0155 -0.0815** 0.0328

(0.0327) (0.0332) (0.0206)

Observations 137,312 137,312 137,312

R-squared 0.037 0.040 0.013

Year FE Yes Yes Yes

F-test for Positions 16.33 23.61 18.14

Prob > F 1.32e-10 0 0

Standard errors in parentheses

*** p<0.01, ** p<0.05, * p<0.1

Table 5: Three OLS Regressions of hours watched by channel on cable channel availability,

local cable channel positions interacted with availability, demographics, and year dummy

variables for the sample of cable subscribers. The F statistic is for the joint test that all

the coeﬃcients on the channel positions interacted with availability are equal to zero.

(1) (2) (3)

VARIABLES Fox News Hours CNN Hours MSNBC Hours

log(Fox News Position) -0.0150 0.00411 0.0146

(0.0442) (0.0381) (0.0255)

log(CNN Position) -0.00219 0.0147 -0.00924

(0.0300) (0.0258) (0.0173)

log(MSNBC Position) 0.0220 -0.000588 -0.0201

(0.0480) (0.0413) (0.0276)

HH Income 0.00662** -0.000931 0.00233

(0.00264) (0.00227) (0.00152)

HH Income

-4.42e-05* 2.59e-05 -1.39e-05

(2.67e-05) (2.30e-05) (1.54e-05)

HH Income

9.92e-08 -8.23e-08 3.66e-08

(7.52e-08) (6.47e-08) (4.33e-08)

Age -0.0145* -0.00497 0.00737*

(0.00753) (0.00648) (0.00433)

Age

0.000581*** 0.000367*** 4.13e-05

(8.01e-05) (6.90e-05) (4.61e-05)

White 0.176*** -0.440*** -0.0566

(0.0601) (0.0518) (0.0346)

Black 0.138* 0.0452 0.126***

(0.0790) (0.0680) (0.0455)

Hispanic -0.353*** -0.144*** -0.0950***

(0.0559) (0.0481) (0.0322)

College -0.0138 0.312*** 0.0834***

(0.0391) (0.0337) (0.0225)

Male 0.210*** 0.197*** 0.116***

(0.0334) (0.0288) (0.0192)

log(N Channels) 0.0478 0.0196 0.0181

(0.0560) (0.0482) (0.0323)

Observations 36,735 36,735 36,735

R-squared 0.046 0.033 0.012

Year FE Yes Yes Yes

F-test for Positions 0.0863 0.124 0.331

Prob > F 0.968 0.946 0.803

Standard errors in parentheses

*** p<0.01, ** p<0.05, * p<0.1

Table 6: Placebo Regressions: Three OLS Regressions of hours watched by channel on cable

channel availability, local cable channel positions interacted with local cable availability,

demographics, and year dummy variables for the sample of satellite subscribers. The F

statistic is for the joint test that all the coeﬃcients on the channel positions interacted with

availability are equal to zero.

signiﬁcantly diﬀerent from zero for own hours watched. The magnitudes imply that

an improvement of the Fox News channel position from the 75th percentile to 25th

percentile would increase viewership by 4 minutes. The cross-channel position eﬀects

are generally positive, reﬂecting substitution consistent with our viewership model.

The ﬁrst stage F statistics on the instrumental variables range from 16.3 for the Fox

News Channel regression to 23.6 in the MSNBC regression.

As a placebo test for whether channel positions respond to unobserved character-

istics of the local population, we repeat the same regressions on the sample of satellite

subscribers but using the channel positions on the local cable system. If the channel

positions on the local cable system are chosen in response to unobservable local charac-

teristics, then these positions should predict satellite viewership. The results in Table 6

indicate that channel positions on the cable system do not predict viewership on satel-

lite. None of the coeﬃcients on cable channel position are statistically distinguishable

from zero for satellite subscribers. These coeﬃcients are always much closer to zero

than in the sample of cable subscribers. Finally, we can reject equality of nearly all of

them with the corresponding cable coeﬃcients with high degrees of conﬁdence.

Demographics predict viewership in similar manners for both cable and satellite

subscribers. Across locations, satellite subscriber characteristics correlate strongly with

cable subscriber characteristics. Table 7 shows the regression coeﬃcients of mean

satellite subscriber characteristics on mean cable subscriber characteristics in the same

cable system territory, nearly all of which are positive and large. Since the means

of these characteristics are measured with sampling error - as they are constructed

watching the channel, and a linear regression of how much one watches conditional on watching. The 2SLS

coeﬃcients have the same estimated magnitudes using either ﬁrst stage model. The second stage Fox News is

only signiﬁcant at the 90% level and the MSNBC eﬀect drops just below conventional conﬁdence levels when

we do not model the mass on zero explicitly, though the signs and magnitudes of the eﬀects are the same.

The loss of precision is not surprising given that a linear model for viewership will grossly under-predict the

number of viewers who watch zero hours.

CNN’s channel position is negative, though imprecise, in the Fox News and MSNBC regressions. A

negative coeﬃcient would suggest a complementarity between CNN and these two channels, whereas our

model does not admit such a complementarity. We note this as a weakness of the analysis. Modelling the

complementarity in the model we use is diﬃcult computationally as the ﬁrst stage optimization would have

to be solved numerically rather than with an analytical solution that our current demand system admits.

In the Appendix, we carry out Chow tests assuming the demographic eﬀects on viewership are the same

between cable and satellite subscribers, and formally reject that the cable position eﬀect on cable subscribers

is equal to the eﬀect on satellite subscribers.

from the television viewership survey samples - the OLS coeﬃcients are attenuated.

In the table, we address this measurement error problem in two ways.

First, we

progressively restrict the regression to markets with more and more survey respondents

as these markets will have less sampling error. Second, we instrument for the mean

cable characteristic with lead and lagged mean cable characteristic. Survey respondents

are sampled independently from year to year. Consistent with measurement error, the

coeﬃcients genereally tend upwards to one when we restrict to system-years with more

respondents. Furthermore, the IV coeﬃcients are generally very close to one.

In the same vein, we can look directly at viewership patterns. Satellite viewers

watch 1.2 fewer minutes per week of Fox News Channel on average relative to cable

viewers (on an overall mean of 90 minutes). At the bottom of Table 7, we regress

predicted mean viewership of satellite subscribers (predicted from demographics) on

that of cable subscribers. We also regress the cable system territory mean residual

viewership of satellite subscribers (net of demographics) on the mean residual viewer-

ship of cable subscribers. Across the board, cable and satellite subscribers within the

same cable system territory display strong correlations of demographics and viewing

behavior.

What is important is that satellite subscribers’ ideologies are correlated with cable

subscribers’ ideologies. Given that all observable characteristics correlate positively,

and that demographics explain little of whether or not an individual is a satellite sub-

scriber, we ﬁnd such a scenario to be plausible. Our principal identiﬁcation assumption

is thus that cable channel positions are not chosen to reﬂect unobservable attributes

that are unique to cable subscribers’ (to the extent that they diﬀer from satellite sub-

scribers’) political leanings in a locality. The placebo tests provide empirical evidence

in favor of this assumption.

4.2 Voting on Projected Viewership: 2SLS with Individ-

ual Level Data

Let y

be an indicator for whether individual i intends to vote for the Republican

Presidential candidate in the election of year t. Let h

be the reported hours watched

One could also dis-attenuate the coeﬃcients as the variance induced by sampling is known. This exercise

is complicated because each cable system-year has diﬀerent sampling variance.

Characteristic N>0 N>10 N>50 N>100 IV

Black 0.581*** 0.708*** 0.783*** 0.912*** 0.996***

(0.0129) (0.0148) (0.0279) (0.0571) (0.0388)

College 0.398*** 0.540*** 0.705*** 0.716*** 0.917***

(0.0165) (0.0202) (0.0412) (0.0714) (0.0779)

HH Income 0.498*** 0.612*** 0.820*** 0.886*** 0.973***

(0.0144) (0.0166) (0.0309) (0.0607) (0.0637)

Age 0.261*** 0.358*** 0.395*** 0.490*** 0.791***

(0.0165) (0.0212) (0.0458) (0.0764) (0.0998)

Hispanic 0.538*** 0.665*** 0.778*** 0.843*** 0.838***

(0.0138) (0.0159) (0.0234) (0.0345) (0.0304)

Party ID R 0.105*** 0.289*** 0.629*** 0.888*** 1.552***

(0.0286) (0.0503) (0.106) (0.172) (0.437)

Party ID D 0.118*** 0.228*** 0.630*** 1.174*** 2.947*

(0.0282) (0.0506) (0.117) (0.211) (1.690)

Predicted Fox News 0.766*** 0.804*** 0.912*** 0.924*** 0.965***

(0.0130) (0.0137) (0.0215) (0.0274) (0.0453)

Predicted CNN Viewing 0.506*** 0.556*** 0.633*** 0.720*** 0.831***

(0.0177) (0.0198) (0.0361) (0.0523) (0.146)

Predicted MSNBC Viewing 0.766*** 0.791*** 0.831*** 0.810*** 0.828***

(0.0132) (0.0141) (0.0259) (0.0397) (0.0761)

Fox News Residual 0.111*** 0.168*** 0.453*** 0.376*** 0.612**

(0.0237) (0.0282) (0.0624) (0.0862) (0.304)

CNN Residual 0.180*** 0.190*** 0.272*** 0.340*** 0.645***

(0.0191) (0.0201) (0.0452) (0.0656) (0.196)

MSNBC Residual 0.101*** 0.107*** 0.424*** 0.642*** 0.588**

(0.0210) (0.0221) (0.0613) (0.106) (0.246)

Note: The ﬁrst column of coeﬃcients uses all cable system territory-years. These coeﬃcients

are attenuated because the mean cable is constructed from samples of survey respondents

which can be as few as 2 per cable system territory-year. The second column of coeﬃcients

restricts to those with more than ten surveyed respondents. The third column of coeﬃcients

restricts to those with more than ﬁfty survey respondents. The fourth column of coeﬃcients

restricts to those with more than 100 survey respondents. The ﬁnal column of coeﬃcients

are uses lead and lagged means of cable subscribers as instrumental variables as respondents

are sampled independently from year to year.

Table 7: Regression coeﬃcients of demographic characteristics and cable news viewership of

satellite subscribers on the characteristics of cable subscribers in the same cable territory-year

in MediaMark / Simmons viewership data.

per week of channel j where j ∈ c, f, m with c corresponding to CNN, f to Fox News,

and m to MSNBC. We are interested in the coeﬃcients in the population relationship:

= γ

+ a

+ αx

+ β

+ 

(1)

where γ are election ﬁxed eﬀects and α are election-speciﬁc coeﬃcients on demo-

graphics x

. The assumption that  is uncorrelated with the vector of hours watched

is untenable if consumers have a preference for like-minded news. We use channel po-

sitions as instrumental variables for these three endogenous variables, and estimate by

2SLS. Results from the second stage regressions are in Table 8. We compute standard

errors by bootstrap as deemed appropriate in two-sample IV settings by Inoue and

Solon (2010). Our estimates imply that being induced to watch an additional hour per

week of Fox News by the channel position instruments would lead to an approximately

14-point increase in the probability of voting Republican in presidential elections for

those induced into watching by the instrument.

In terms of magnitudes, the estimated eﬀect of one hour of Fox News is just under

one-half of the eﬀect of being black on voting Republican, and about equal to the

diﬀerence in dummy coeﬃcients for residence in Ohio versus residence in Massachusetts.

We point out a few caveats regarding this seemingly large magnitude. First, the implied

eﬀect on an election of removing Fox News will be signiﬁcantly smaller, as the majority

of people do not watch any Fox News, and among those who do, the majority are

already going to vote Republican.

Second, our survey-based vote intention measure

may include individuals who are both unlikely to vote and lack an established political

ideology, and thus are more open to persuasion by slanted news.

When we restrict

our second stage to registered voters, the Fox News Channel coeﬃcient falls to 0.107.

Enikolopov et al. (2011) examined both individual level survey data and actual vote

The typical change in viewership induced by the instrument is signiﬁcantly less than one hour per week.

Given our ﬁrst-stage estimates and the distribution of the instrument presented in Figure 3, a one-standard

deviation increase in channel position induces a roughly 3-minute-per-week increase in Fox News viewing.

In Section 7, we estimate a 1.3% change in the 2000 presidential election’s Republican vote share resulting

from the existence of Fox News. This estimate accounts for selective viewership.

An extensive literature on turnout in political science, e.g. Leighley and Nagler (2013), has found non-

voters to be much more likely to report independent partisan aﬃliation and “moderate” ideological leanings

than are regular voters.

(1)

VARIABLES Intent to Vote R

Fox News Channel Hours 0.137***

(0.0108)

CNN Hours 0.0221

(0.0197)

MSNBC Hours -0.101***

(0.0221)

HH Income 0.00273***

(0.000286)

HH Income

-2.22e-05***

(3.04e-06)

HH Income

5.65e-08***

(9.06e-09)

Age 0.00377***

(0.000620)

Age

-7.47e-05***

(9.25e-06)

White 0.0745***

(0.00740)

Black -0.357***

(0.00755)

Hispanic -0.0571***

(0.00606)

College -0.0636***

(0.00556)

Male 0.0516***

(0.00339)

log(N Channels) -0.0975***

(0.00463)

Observations 122,243

R-squared 0.089

Year Eﬀects Yes

Channel Availability Yes

Standard errors in parentheses

*** p<0.01, ** p<0.05, * p<0.1

Table 8: Second Stage regressions of NAES and CCES intent to vote Republican vote share

on predicted hours watched from the ﬁrst stage.

shares and found signiﬁcantly larger eﬀects of media consumption using the individual

level survey data. Finally, as the behavioral model makes clear, we are estimating a

single coeﬃcient in a world of heterogenous treatment eﬀects. The IV estimates reﬂect

an average treatment eﬀect on individuals induced into watching by the instrument.

These viewers will lack strong political orientations, and thus be more malleable.

There is additional validation when we interact the instrumented time watched

with election year dummy variables. The Fox News eﬀect increases over time. The

MSNBC eﬀect is only present in 2008, after MSNBC made the switch to liberal slanted

programming. These interaction terms are both consistent with the ideology estimates

in Figure 5. First, the estimated ideology of Fox News moves further to the right over

the sample period. Second, MSNBC’s persuasive eﬀect only dips to the left after the

2004 election. Sorting out the dynamics of persuasion implied by these results are an

interesting angle for future research.

4.3 Voting on Viewership: OLS with Individual Level

Data

We also present the estimated coeﬃcients in an OLS regression of intent to vote Re-

publican on demographics and indicators for most watched cable news channel in the

NAES data. This is the OLS analog to the second stage above with two slight diﬀer-

ences. First, it uses only indicators for most watched cable news channel rather than

hours watched. Second, the sample is restricted to NAES data. The reason for these

diﬀerences is because the OLS regression is only possible on the NAES data where

hours watched is not available. In the 2SLS, we estimate the ﬁrst stage on Mediamark

and Simmons data,

and use the estimated coeﬃcients to predict hours watched on

the NAES and CCES data which has intent to vote.

The purpose of presenting the mis-speciﬁed OLS regression is because we use it

in the model estimation later to estimate the taste for like-minded news. Imagine a

researcher who purports to estimate the inﬂuence of Fox News on voting Republican by

regressing intent to vote Republican on an indicator for whether Fox News is the most

watched news channel. Such an estimate would lack credibility because it mixes any

inﬂuence eﬀect with the selection into viewership of Fox News by those who intend to

Mediamark and Simmons data do not include Intent to vote.

(1)

VARIABLES Intent to Vote R

Fox News Channel Hours * Year 2000 0.100***

(0.0304)

Fox News Channel Hours * Year 2004 0.122***

(0.0204)

Fox News Channel Hours * Year 2008 0.141***

(0.0154)

CNN Hours * Year 2000 0.0148

(0.0253)

CNN Hours * Year 2004 -0.0105

(0.0258)

CNN Hours * Year 2008 0.0337

(0.0205)

MSNBC Hours * Year 2000 0.0296

(0.0416)

MSNBC Hours * Year 2004 -0.0144

(0.0470)

MSNBC Hours * Year 2008 -0.113***

(0.0226)

HH Income 0.00246***

(0.000450)

HH Income

-1.99e-05***

(4.10e-06)

HH Income

5.03e-08***

(1.14e-08)

Age 0.00328***

(0.000636)

Age

-6.87e-05***

(1.02e-05)

White 0.0754***

(0.00841)

Black -0.358***

(0.00864)

Hispanic -0.0574***

(0.00671)

College -0.0681***

(0.00646)

Male 0.0497***

(0.00512)

log(N Channels) -0.0964***

(0.00468)

Observations 122,243

R-squared 0.090

Year Eﬀects Yes

Channel Availability Yes

Standard errors in parentheses

*** p<0.01, ** p<0.05, * p<0.1

Table 9: Second Stage regressions of NAES and CCES intent to vote Republican vote share

on predicted hours watched from the ﬁrst stage interacted with election year eﬀects.

vote Republican because of tastes for like-minded news. Our approach is to use channel

positions to estimate the inﬂuence eﬀect, while jointly matching the OLS coeﬃcient to

estimate the tastes for like minded news.

Table 10 conﬁrms a positive correlation between intending to vote Republican and

indicating that one’s most watched cable news channel is Fox News, and corresponding

negative coeﬃcients for indicating MSNBC and CNN.

5 Model

The model for a given election cycle has two stages. In the ﬁrst stage, the consumer-

voters choose a television package and how much time to watch the cable news channels.

In the second stage, the consumer-voters vote in the Presidential election. Between

the ﬁrst and second stage, the consumer-voters’ ideologies evolve as a function of the

ideologies of and time spent watching the news channels.

5.1 Voter Ideology and Presidential Vote Decision

Our consumer-voters have a latent unidimensional political ideology which determines

their vote choice in presidential elections. We denote the left-right ideology of consumer-

voter i by r

We specify voters’ initial ideologies as a function of their county of residence. Specif-

ically, we estimate a logit model of vote choice with county dummies as explanatory

variables, which matches county level vote shares from the previous election cycle.

The county-level intercepts from this model then determine the simulated consumers’

initial ideologies, along with an iid logit error term:

ij0

= δ

+ 

(2)

Where δ

is the estimated county intercept for county j, consumer i’s county of

residence. From this starting point, ideology may evolve if the consumer watches cable

news, according to a process described in detail later in this section.

E.g., simulated voters in the 2008 election cycle have their ideologies initialized using coeﬃcients that

match county-level vote shares in the 2004 election.

(1)

VARIABLES Intent to Vote R

Most Watched CNN -0.0869***

(0.00633)

Most Watched Fox News Channel 0.329***

(0.00654)

Most Watched MSNBC -0.0963***

(0.00844)

HH Income 0.00427***

(0.000408)

HH Income

-3.11e-05***

(4.62e-06)

HH Income

7.21e-08***

(1.38e-08)

Age 0.00333***

(0.000759)

Age

-3.47e-05***

(7.38e-06)

White 0.103***

(0.00876)

Black -0.285***

(0.0111)

Hispanic -0.0909***

(0.00908)

College -0.0556***

(0.00465)

Male 0.0447***

(0.00423)

log(N Channels) -0.104***

(0.00786)

Observations 44,472

R-squared 0.230

Year Eﬀects Yes

Channel Availability Yes

Standard errors in parentheses

*** p<0.01, ** p<0.05, * p<0.1

Table 10: OLS linear probability regression of intent to vote Republican on demographics

and indicators for most watched cable news channel.

At election time, each voter votes for the party whose candidate’s announced po-

sition is closest to her own. This behavior is consistent with voting given a utility

function over the ideology of the winning candidate that is single-peaked with maxi-

mum at r

. As in all such spatial models, only the cutpoint between the candidate

positions, and not the absolute values of the positions, determine the voting decision.

We can, therefore, describe each presidential election using a single parameter P

, for

t ∈ {2000, 2004, 2008}. All voters to the left of the cutpoint (with r

< P

) vote for the

Democratic candidate in the election, and those to the right vote for the Republican.

5.2 Viewership and Subscription

The viewership time allocation and subscription portion of the model follows Crawford

and Yurukoglu (2012). Given access to the news channels C

in package j in year

t, consumer-voter i allocates their time amongst watching those channels and other

activities to maximize:

c∈C

ict

log(1 + T

ijc

) (3)

where γ

ict

is consumer-voter i’s preference parameter for news channel c in year

t. We choose the normalization that the outside option (doing anything other than

watching cable news) has γ

i0t

= 1 for all i, t, and parameterize the remaining vector of

ict

= χ

◦ ν

ict

∼ Bernoulli(α

0ct

+ Π

+ ζ

pos

ict

− η((a + br

) − r

)

) (4)

ict

∼ Exp(α

+ Π

+ ζpos

ict

) (5)

ict

determines whether consumer-voter i has a non-zero preference for channel c.

It is a random function of demographics d

according to parameters Π

, a channel-

year speciﬁc ﬁxed eﬀect α

0ct

, the position of the channel in the lineup according to

We use this formulation because most consumers watch zero or one news channel.

, and the distance of consumer-voter i’s one dimensional political ideology r

from

the channel’s estimated ideology r

according to η. This last term represents taste

for like-minded news and follows a similar parameterization to Gentzkow and Shapiro

(2010). The parameters a and b scale the text based ideology measures to allow for

consumers to perceive slant as a linear function of the text based slant measure. If

η is positive, then increasing the ideological distance between consumer-voter i and

channel c reduces the probability i watches c.

If the consumer-voter has a non-zero preference, the intensity of her preference is

drawn from an exponential distribution whose distributional parameter depends on

, a channel-year speciﬁc ﬁxed eﬀect, demographics d

according to parameters Π,

and the position of the channel in the lineup according to ζ. The exponential shape

assumption mixed with a mass at zero is inspired by the raw data on hours watched

which features a mass at zero and right-skewed and monotonically decreasing density.

The constrained maximization problem deﬁned by (3) has an analytic solution that

can be determined as follows. Deﬁne ρ

ict

as the Lagrange multipliers associated with

the non-negativity constraints on T

ict

. By complementary slackness, if ρ

ict

> 0 then

ict

= 0. From the ﬁrst order condition, ρ

ict

= λ

− γ

ict

where λ

is the Lagrange

multiplier on consumer i’s budget constraint. Therefore, T

ict

can be zero if and only if

ict

< λ

For all the channels with γ

ict

> 0, λ

= γ

ict

/(1+ T

ict

). Additionally, each consumer

faces a time-budget constraint,

ict

= B, where B is the total time available (in our

scaling, the number of hours in a week: 168). This gives a system of equations with

solution:

1 +

B + C

where c

are the indices of the channels that i watches a positive amount, and C

is the total number of such channels. Given this result, the iterative solution is to

replace the γ

ict

’s below the cutoﬀ (1 +

)/(B + C) with zero. If there were any

ict

’s below this threshold, we now have a new cutoﬀ deﬁned by the remaining positive

ict

’s, and we repeat the process again. There are at most C steps of this until we hit

the ﬁnal set of positive γ

ict

’s, at which point we compute the times watched as:

ict

= (T + C

)

ict

− 1(γ

ict

> 0)

The indirect utility from solving this problem enters into the consumer-voter’s de-

cision whether to subscribe to cable, satellite, or no television package at all. The

conditional indirect utility from subscribing to package j is

= v

∗

+ δ

+ 

where δ

is the mean utility of package j, 

is an idiosyncratic logit error term

and j corresponds to cable or satellite. We also allow consumers to subscribe to no

package at all. This choice is associated with a normalized δ

= 0 and, since we

assume consumers cannot watch cable news if they do not subscribe, the only choice is

to spend their entire time budget in non-cable-news activity. This yields corresponding

= log(1 + B) + 

5.3 Ideological Inﬂuence

After watching cable news, consumer-voter i’s one-dimensional political ideology evolves

as a function of how much time i spends watching the news channels and the ideology

of the news channels.

We assume that i is attracted towards the ideologies of the

news channels he watches, the more so the more time i spends watching. Speciﬁcally

i,t−1

+ ρ

ic,t−1

(a + br

c,t−1

)

1 + ρ

ic,t−1

(6)

where r

i,t−1

is i’s ideology prior to watching, r

is i’s new ideology, and ρ is a

parameter to be estimated which controls the magnitude of news channels’ inﬂuence

on viewers’ ideology. One interpretation of ρ is as a (per-hour) rate at which viewers

receive ideological signals while watching cable news. If voters treat signals from slanted

outlets as true draws on the state of the world, and further, if they do not account for

the lack of independence between repeated signals from the same source as in the model

The ideology measure is the same function of the text based slant measure that enters the viewership

decision problem.

of DeMarzo et al. (2003),

then equation (6) arises as the inverse-variance-weighted

average of signals observed by viewer i in period t.

The functional form here implies that a consumer-voter’s attraction is governed by

the same parameter (ρ), whether coming from the left or the right. This parameter

doesn’t depend on how far away the consumer-voter starts from the channel. It rules

out that a voter might watch a slanted channel, become disgusted, and move in the

opposite direction of the channel as in Arceneaux et al. (2012). Furthermore, consumer-

voters are naive about the inﬂuence eﬀect when choosing time watched.

6 Estimation, Results, and Empirical Identiﬁ-

cation

We estimate the parameters of the model by indirect inference (Smith (1990); Gourier-

oux et al. (1993)). This implies choosing the estimates of the model’s parameters to

match estimates of an auxiliary model. The auxiliary model consists of eight linear

regressions that fall into four categories, and a set of unconditional moments: (1) a

regression of individual level time spent watching each cable news channel on demo-

graphics and channel positions, conditional on watching the channel (three regressions),

(2) a linear probability regression of individual level watching any positive amount of

each cable news channel on demographics and channel positions (three regressions),

and (3) a linear probability regression of individual level intent to vote Republican on

demographics and predicted time spent watching from (1) and (2), and (4) an OLS

regression of intent to vote Republican on reporting whether Fox News, MSNBC, or

CNN is an individual’s most watched cable news channel. (1)-(3) correspond to a two-

stage least squares estimate of the eﬀects of watching the cable news channels on voting

Republican using channel positions as instrumental variables. (4) corresponds to an

OLS regression of intent to vote Republican on viewership. Finally, we also match (5)

the actual vote shares in each presidential election, the year by year hours watched for

each channel, and the year by year fraction of non-zero viewership for each channel.

Gentzkow and Shapiro (2006) explore media consumption and endogenous slant with fully Bayesian

consumers.

For this interpretation to hold over a series of periods, we require that at the beginning of each period

the consumer gets an ideology shock which returns the variance of his ideology to 1.

We choose the model’s parameters so that estimating (1)-(5) on data simulated from

the model produce coeﬃcient estimates with minimum distance to those in the data.

We weight the distance metric in proportion to the inverse of the standard errors in

the estimated relationships in the real data, with the exception that the IV and OLS

hours watched coeﬃcients are weighted more heavily.

We also impose two constraints in our estimation procedure to accomodate the lin-

ear probability structure of equation 4. First, all the channel-year ﬁxed eﬀects in the

Bernoulli parameter (α

0ct

) are constrained to lie between zero and one. Second, the

demographic eﬀects in the Bernoulli parameter (Π

) are constrained such that their

linear combination lies between -1 and 1 for all demographic types.

Note that these

constraints apply only to the ﬁxed eﬀects and the demographic parameters, and thus

do not prevent the model from setting the probability of a given individual watch-

ing a channel to be exactly zero, for example if the individual is very ideologically

distant from the channel or has the channel in very high position on her cable sys-

tem. We impose them to prevent the optimizer from moving into territory where the

channel-position and ideological eﬀects have no inﬂuence on large numbers of simulated

individuals.

6.1 Empirical Identiﬁcation

Empirical identiﬁcation of the model’s parameters is relatively straightforward. In

terms of the notion of parameter estimated sensitivity, formalized in Gentzkow and

Shapiro (2013), ρ, the parameter which determines the degree of inﬂuence, is sensitive

to the coeﬃcients on projected time in the second stage regression. η, the parameter

governing the degree of tastes for like-minded news, is sensitive to coeﬃcients on which

channel is reported as most watched in the OLS regression relative to the coeﬃcients

on projected hours watched in the second stage regressions. One intuition for these

estimates comes from considering the OLS regression of intent to vote Republican on

Fox News Channel hours watched. The coeﬃcient estimates on hours watched of Fox

We begin with the inverse regression standard errors as our weights vector, but scale up the OLS and

IV hours watched weights by a relative factor of 100.

Our demographic variables are all either binary (such as our race and education dummies) or continuous

but bounded in the data (such as income and age). Hence, it is possible to deﬁne a demographic proﬁle for

each channel with minimum and maximum viewership.

News Channel would not be a credible measure of the eﬀects of consuming media

because the estimate would conﬂate tastes for like-minded news with any inﬂuence

eﬀect. However, if one knew the level of the inﬂuence eﬀect, then this estimate would

be informative about the tastes for like-minded news. Our approach is to measure the

inﬂuence eﬀect by using channel positions as instrumental variables, and choose the

level of tastes for like minded news to explain the OLS coeﬃcient conditional on the

inﬂuence eﬀect.

ζ, the parameters determining the strength of channel positions in the time al-

location problem, are sensitive to the ﬁrst stage coeﬃcients on channel positions. A

similar straightforward relationship applies to the demographic factors inﬂuencing time

watched and the coeﬃcients on demographics in the ﬁrst stage regressions.

, the parameters characterizing the three presidential elections in our sample

period, are sensitive to the year-eﬀects in the voting equations. These parameters

allow the model to capture national trends in party preference.

Finally, a and b, the parameters scaling our text-based ideology measures, are

sensitive to both the OLS regression coeﬃcients on which channel is reported as most

watched as well as the coeﬃcients on projected hours watched in the second stage

regressions. Separate identiﬁcation of these parameters from ρ and η is possible because

there are three channels and thus six moments to work with. The asymmetries in the

channels’ estimated eﬀects relative to their text-based ideological positioning provide

variation to distinguish the scaling parameters from ρ and η. To make this concrete,

consider the OLS estimates for Fox News and MSNBC. The Fox News coeﬃcient is

more positive than the MSNBC coeﬃcient is negative. Increasing η intensiﬁes the

magnitude of both OLS coeﬃcients generated from the model in similar proportions.

Increasing b at a ﬁxed η increases the magnitude of the Fox News coeﬃcient at a faster

rate than the MSNBC coeﬃcient, because the text-based Fox News ideology is more

conservative than the text-based MSNBC ideology is liberal.

To test this intuition, we veriﬁed that the derivative of the model’s Fox News OLS coeﬃcient with respect

to η is on the same order as the same derivative for the MSNBC OLS coeﬃcient, while the derivative of the

model’s Fox News coeﬃcient with respect to b is larger than that derivative for the MSNBC OLS coeﬃcient.

6.2 Model Estimates

Table 11 shows the main parameter estimates from the model.

We estimate positive

values for both ρ, the inﬂuence parameter, and η, the taste for like-minded news,

implying a positive feedback process where voters watch slanted news, are inﬂuenced

to move closer to the news’ channel’s ideology, and subsequently have even stronger

preference for that channel, due to the decreased ideological distance.

Parameter Estimate Bootstrapped Standard Error

Slant Preference (η) 0.163 0.0109

Ideological Inﬂuence (ρ) 0.096 0.0080

Position Eﬀect - Ratings -0.002 0.0002

Position Eﬀect - Viewership -0.085 0.0030

2000 R/D Threshold -0.184 0.0130

2004 R/D Threshold 0.055 0.0127

2008 R/D Threshold 0.106 0.0167

Channel Ideology Intercept (a) -0.246 0.0179

Channel Ideology Slope (b) 5.378 0.2441

Table 11: Key parameter estimates.

The magnitude of the estimate of the taste for like minded news parameter η

implies that an ideological distance of one unit between viewer and channel reduces

that viewer’s probability of watching by about 16%. For reference, at our estimated

scaling parameters, the ideological distance between Fox News and MSNBC in 2008

is close to one unit. The magnitude of ρ implies that a voter watching an hour per

week of a news channel for a year would be inﬂuenced to a new ideological position

just under 10% of the distance to the channel’s ideology. Estimates of the channel

position parameters, consistent with the data, imply that increasing channel position

decreases both the probability of watching any of a channel, as well as the number of

hours watched conditional on watching any. The eﬀect on the probability of watching

any - row 4 in the table - implies that doubling the channel position decreases the

probability of a typical voter watching a channel by about 8.5%.

The full set of parameters additionally contains channel-year ﬁxed eﬀects and demographic terms, sepa-

rately for the amount watched and the probability of watching any. These are omitted here for brevity. The

estimated model’s ﬁt on regression coeﬃcients is available in the Appendix.

The channel position eﬀect on the number of hours watched is harder to interpret

directly, as the hours-watched model is nonlinear and hence eﬀects of changing these

quantities depend on the values of all the other covariates. Tables 12 and 13 therefore

show some interpretable quantities generated by the model for viewers with various

demographic and ideological proﬁles.

Channel Position Elasticity

Age Income ($000s) Ethnicity College Gender Ideology CNN FOX MSNBC

65 25 White No Man Centrist 17.5 16.6 16.0

65 25 White No Man Median Republican 0.0 16.1 0.0

65 25 White No Man Median Democrat 16.4 13.4 15.9

30 85 Black Yes Man Centrist 15.8 11.7 13.6

30 85 Black Yes Man Median Republican 12.8 11.2 0.0

30 85 Black Yes Man Median Democrat 14.6 8.2 13.5

65 85 Hispanic No Man Centrist 20.8 16.8 16.3

65 85 Hispanic No Man Median Republican 18.0 16.3 0.0

65 85 Hispanic No Man Median Democrat 19.7 13.5 16.2

30 25 White Yes Woman Centrist 0.0 8.8 12.5

30 25 White Yes Woman Median Republican 0.0 7.4 0.0

30 25 White Yes Woman Median Democrat 0.0 0.0 12.4

65 25 Black No Woman Centrist 18.2 16.9 15.3

65 25 Black No Woman Median Republican 0.0 16.5 0.0

65 25 Black No Woman Median Democrat 17.1 13.7 15.3

30 85 Hispanic Yes Woman Centrist 14.6 9.0 12.8

30 85 Hispanic Yes Woman Median Republican 0.0 8.5 0.0

30 85 Hispanic Yes Woman Median Democrat 13.4 0.0 12.8

Table 12: Change in expected ratings (minutes watched per week) following a move from

channel position 50 to channel position 30, for selected demographic and ideological proﬁles.

Table 12 shows computed elasticities of viewers’ expected minutes watched with

respect to channel position. We compute the change in ratings (measured in minutes

per week) resulting from a channel’s moving from position 50 to position 30 in the

lineup.

All are weakly positive, as expected, although some are exactly zero because

the average viewer of the given demographic and ideological proﬁle does not watch

any of the channel, regardless of position. Viewers’ demographics and initial ideologies

have an important inﬂuence on their sensitivity to channel position, with viewer-types

who initially watch more of a channel showing larger changes in minutes.

Table 13 shows a diﬀerent look at the relationship of viewer preference for channels

to demographics and channel position. For the same ideological and demographic pro-

Positions 50 and 30 correspond approximately to the 75th and 25th percentile position, respectively, of

Fox News in 2008. We show the same channel position diﬀerence for all three channels to make the ﬁgures

comparable to each other.

Age Income ($000s) Ethnicity College Gender Ideology Channel Rank-Order FNC ⇔

MSNBC

65 25 White No Man Centrist FNC, CNN, MSNBC -2.25

65 25 White No Man Median Republican FNC, CNN, MSNBC -4.12

65 25 White No Man Median Democrat MSNBC, CNN, FNC 1.08

30 85 Black Yes Man Centrist CNN, MSNBC, FNC 0.47

30 85 Black Yes Man Median Republican FNC, CNN, MSNBC -2.98

30 85 Black Yes Man Median Democrat CNN, MSNBC, FNC 4.75

65 85 Hispanic No Man Centrist CNN, FNC, MSNBC -1.75

65 85 Hispanic No Man Median Republican FNC, CNN, MSNBC -4.71

65 85 Hispanic No Man Median Democrat CNN, MSNBC, FNC 1.44

30 25 White Yes Woman Centrist MSNBC, FNC, CNN 1.07

30 25 White Yes Woman Median Republican FNC, CNN, MSNBC -0.18

30 25 White Yes Woman Median Democrat MSNBC, CNN, FNC 1.83

65 25 Black No Woman Centrist FNC, CNN, MSNBC -2.59

65 25 Black No Woman Median Republican FNC, CNN, MSNBC -4.29

65 25 Black No Woman Median Democrat MSNBC, CNN, FNC 0.63

30 85 Hispanic Yes Woman Centrist CNN, MSNBC, FNC 1.63

30 85 Hispanic Yes Woman Median Republican FNC, CNN, MSNBC -1.22

30 85 Hispanic Yes Woman Median Democrat MSNBC, CNN, FNC 3.79

Table 13: Preference orderings of channels, and change in log channel position needed to ﬂip

FNC / MSNBC preference order, for selected demographic proﬁles.

ﬁles as in the previous table, Table 13 lists that type of viewer’s preference ordering

over the three cable channels, on a hypothetical system where all three are available

and positioned at their median position in the data in 2008. The last column shows the

diﬀerence in log channel position that would be required to ﬂip that type of viewer’s

preference between Fox and MSNBC. Demographic eﬀects play a large role in deter-

mining the most preferred channel. Within demographic proﬁles, ideology can drive

diﬀerences in preferences: all Republican types prefer Fox News to MSNBC, and vice

versa for Democrats.

In both our raw data and in the simulations, slanted media are consumed by agents

who do not necessarily share the same ideology as the media outlet. This result is

consistent with the analysis in Gentzkow and Shapiro (2011) who ﬁnd that much

of Fox News Channel’s audience is composed of people who do not self-identify as

conservative, and related, that self-identiﬁed conservatives watch other cable news

besides Fox News.

Our model estimates match these facts. Furthermore, such a

lack of ideological segregation is a necessary precursor in our model for cable news

consumption to change voter intentions.

Their results apply more broadly showing that individuals across the political spectrum tend to consume

media that is ideologically diverse.

Election Voter Ideology 1 Hour CNN 1 Hour FNC 1 Hour MSNBC

2000

Centrist -0.016 0.014 0.004

Median Republican -0.078 -0.054 -0.062

Median Democrat 0.048 0.073 0.065

2004

Centrist -0.003 0.022 0.005

Median Republican -0.069 -0.048 -0.062

Median Democrat 0.058 0.080 0.065

2008

Centrist -0.011 0.035 -0.036

Median Republican -0.076 -0.039 -0.097

Median Democrat 0.051 0.092 0.031

Table 14: Eﬀects of watching an additional 1 hour per week on the probability of voting

Republican.

Election Voter Ideology CNN FNC MSNBC

2000

Median Republican 0.41 0.46 0.32

Median Democrat 0.62 0.35 0.32

2004

Median Republican 0.31 0.53 0.19

Median Democrat 0.41 0.36 0.20

2008

Median Republican 0.00 0.31 0.00

Median Democrat 0.12 0.00 0.20

Table 15: Probability of cable subscribers with access to the channel watching each channel

in each election year, for diﬀerent ideological types.

We ﬁnd that the perception of slant for the channels is a multiple of about ﬁve

times the text based slant measure. The text based slant measures place Fox News and

MSNBC in 2008 closer to the center than the median Republican or median Democratic

congressman, respectively. The scaled ideology estimates place Fox News close to the

median Republican voter.

Table 14 shows the change in the probability of voting Republican with respect

to watching one hour per week of each of the cable channels, again for viewers with

diﬀerent initial ideological types. For initially centrist voters, watching additional CNN

has a small negative inﬂuence on the probability of voting Republican, by between 0

and 1.5 percentage points depending on the election. The eﬀect of MSNBC is very

close to zero in 2000 and 2004, but becomes substantially negative (at 3.6 percentage

points) in 2008 after MSNBC’s format switch. The eﬀect of Fox is positive, ranging

Election All voters Only attached voters

FNC (D to R) MSNBC (R to D) FNC (D to R) MSNBC (R to D)

2000 13% 12% 2% 3%

2004 34% 9% 17% 1%

2008 50% 30% 29% 23%

Table 16: Persuasion rates of Fox News and MSNBC. “All voters” counts as a Democrat

any voter initially to the left of the election cutoﬀ, and counts as a Republican any voter

initially to the right. “Only attached voters” includes only voters in the leftmost 25% and

rightmost 25% of the voter ideology distribution. Percentages are conditional on watching

the channel.

from 1.4 point in 2000 to 3.5 points in 2008. The largest elasticity magnitudes are

on individuals from the opposite ideology of the channel, with Fox generally better at

inﬂuencing Democrats than MSNBC is at inﬂuencing Republicans. This last feature

is consistent with the regression result that the IV eﬀect of Fox is greater than the

corresponding eﬀect for MSNBC.

Table 15 shows the eﬀects of viewers’ initial ideology on the probability that they

choose to watch each channel. The noteworthy pattern is that both Fox and MSNBC

appear to become more ideologically distinctive over time. In 2000, the diﬀerence be-

tween the probability of a typical Republican and that of a typical Democrat watching

Fox is about 11 percentage points; by 2008 that gap widens to over 30 points. For

MSNBC the diﬀerence is essentially zero in 2000 and 2004, but widens to 20 points by

2008.

Finally, Table 16 shows the estimated persuasion rates of the channels at converting

votes from one party to the other. The numerator here is the number of, for example,

Fox News viewers who are initially Democrats but by the end of an election cycle

change to supporting the Republican party. The denominator is the number of Fox

News viewers who are initially Democrats.

Again, Fox is more eﬀective at converting

viewers than is MSNBC. The persuasion rates are high, and increasing over time as

the channels polarize. However, due to the taste for like-minded news, the eﬀect on

the total number of votes converted is ambiguous; as shown in Table 15, there are far

fewer Democrats deciding to watch Fox News in 2008 than in 2000.

As our model has no inherent notion of partisanship, only an ideological cutpoint between the parties,

in Table 16 we consider two deﬁnitions of what constitutes a Democrat or Republican partisan.

7 Polarization Dynamics

In this section, we perform two exercises to quantify the eﬀects of slanted cable news on

election outcomes. First, we simulate the evolution of ideology for groups of voters over

time to document the incidence of slanted cable news on these agents’ political ideolo-

gies. Second, we estimate the eﬀect of the entry of Fox News on the 2000 presidential

election.

A positive ρ, implying that watching slanted news aﬀects ideology, and a positive

η, implying a taste for like-minded news, together create the potential for a polarizing

feedback loop. Consider forcing a centrist voter to watch only the Fox News Channel.

The more that individual watches the Fox News Channel, the more they drift to the

right; the more they drift to the right, the more they are attracted to watching Fox

News, and so on. These two eﬀects reinforce each other, in a positive feedback process

related to theoretical models from the literature on network formation (Holme and

Newman, 2006). In this section, we quantify the rate at which such polarization can

occur, given our model’s estimates. We employ the polarization index developed in

Esteban and Ray (1994) and Duclos et al. (2004).

Evolution of viewer ideology Figure 7 shows the results of a simulation of

viewing behavior given the model parameters estimated in the previous section. A

sample of 10,000 hypothetical viewers in an average cable system

in a county with

average demographic characteristics are initialized are assigned ideologies from the

initial ideology distribution for 2008, conditional on their simulated demographics.

In each year, they choose whether and how much to watch of each channel, given

individual-speciﬁc preferences. Their ideologies then adjust towards the ideology of

the channels they view in accordance with equation 6. This process repeats over the

next ten years.

The resulting distribution of ideologies changes according to two contrasting pat-

terns. Cable viewers who are initially centrist quickly polarize according to their chosen

channel, leading to the emergence of new poles at the ideologies of CNN, FNC, and

MSNBC respectively. The peaks of the distribution at those points become succes-

For purposes of this simulation, all viewers are given access to all three cable channels, at the channels’

mean positions in the data.

sively sharper in each year. However, for initially extreme cable viewers - those to the

right of FOX or to the left of MSNBC - watching cable news exerts a centralizing force.

Correspondingly, mass in the far tails of the ideology distribution declines.

Because of this combination of drawing in viewers from the extremes, and driving

apart those initially centrist, the aggregate eﬀect of cable news on ideological polar-

ization is ambiguous and not easily discerned from plots like that in Figure 7. We

therefore turn to the work of Esteban and Ray (1994); Duclos et al. (2004), who pro-

vide a measure of polarization derived from axioms. Employing their measure on

simulated individuals, we ﬁnd that the existence of cable news increases polarization.

Interestingly, this increase in polarization is dependent on the existence of a taste for

like-minded news; without such a taste, cable news actually reduces polarization. The

ﬁrst row in Table 17 presents this polarization measure at the initial ideology distri-

bution for the presidential election cycle in question. The second row presents the

polarization measure after four years of watching cable news. The third row presents

the same measure, but shutting down the taste for like-minded news.

Without a

taste for congruent news, viewers from across the political spectrum would be exposed

to and persuaded by news from the other side.

2000 2004 2008

Initial 0.439 0.439 0.440

Post-Exposure 0.455 0.503 0.470

Post-Exposure (no slant preference) 0.412 0.420 0.396

Table 17: Esteban and Ray polarization measure, before and after exposure to cable news.

Fox Entry in 2000 Next, we estimated the eﬀect of the entry of Fox News into the

cable news market beginning in late 1996 on the 2000 presidential election. Using our

estimated model parameters, we simulated two conditions. First, a base case where Fox

was available to cable subscribers in the 1997-2000 period according to the observed

rollout pattern. Second, a scenario where Fox was available exclusively to satellite

subscribers and not on any local cable system. We computed aggregate viewer welfare

and aggregate vote outcomes under each scenario.

I.e., we set the parameter η equal to zero, rather than the estimated value.

Year 1 Year 2 Year 3

Year 4 Year 5 Year 6

0.0

0.2

0.4

0.6

0.0

0.2

0.4

0.6

−2 0 2 −2 0 2 −2 0 2

Ideology

density

Figure 7: Density plots of viewers’ ideology over time. Voters are initially drawn from the

unconditional ideology distribution in 2008. The remaining 5 plots show the change in the

ideology distribution over time, in years 2 through 6.

Cycle Intention Intention (voters) Welfare Substitution

2000 -0.013 -0.009 -0.38 0.87

2004 -0.057 -0.040 -0.34 0.84

2008 -0.059 -0.042 -0.40 0.80

Table 18: Eﬀects of elimination of the Fox News Channel from cable lineups. Column 2 is the

change in the Republican share of the presidential vote intention; column 3 rescales these to

approximate the eﬀects on registered voters only; column 4 is the average fractional change

in welfare from cable TV of Fox viewers; column 5 is the fraction of initial Fox viewers who

switch to CNN or MSNBC after Fox is eliminated.

Table 18 shows the eﬀects of eliminating Fox from cable lineups in the 1997-2000

period and subsequent election cycles. County-level Republican vote shares on aver-

age fall by 1.3 percentage points under the no-Fox scenario relative to the baseline.

This prediction is roughly an order of magnitude higher than the previous estimate

of DellaVigna and Kaplan (2007).

We note that because our simulation model pre-

dicts vote intention, and not actual votes,

these estimates overstate the eﬀect on

actual election outcomes. As a rough estimate of the importance of this compositional

diﬀerence between the NAES sample and the population of voters, we scale our coun-

terfactual vote changes down in proportion to the reduction in the 2SLS eﬀect seen

when we restrict our regression sample to registered voters (discussed in section 4.2).

Welfare from cable news of those cable subscribers who watched some Fox under the

baseline scenario but could not in the no-Fox case falls as well, by about 40% overall.

The welfare loss is mitigated to some degree by the availability of substitutes in CNN

and MSNBC; large majorities of former Fox viewers switch to watching one of the other

two channels when Fox availablility is eliminated.

We also repeat this no-Fox counterfactual exercise in the two subsequent election

cycles. In subsequent cycles, the implied Fox News eﬀect increases due to two forces.

First and most importantly, overall Fox News viewership approximately doubles during

the period 2000 to 2008, meaning nearly twice as many viewers are exposed to Fox News

One can attribute the diﬀerences to measurement error in the DellaVigna and Kaplan (2007) study

detailed in Appendix A of this paper. Furthermore, their study considers only a subset of geographies for

which they could obtain voting data whereas our estimates cover the universe of US cable systems.

The underlying data come from the NAES, which surveyed a random sample of American adults, not

all of whom are likely voters or even registered to vote.

in later cycles. Second, Fox News moves to the right, increasing its persuasive eﬀect

enough to outweigh any loss in viewership due to the ideological drift. Finally, as Fox

becomes more ideologically distinctive over time, the proportion of its viewers who ﬁnd

MSNBC or CNN to be acceptable substitutes falls as well.

MSNBC Format Switch Finally, we estimated the eﬀects of MSNBC’s format

switch to providing more explicitly liberal coverage in 2005. We simulated a condition

where MSNBC’s ideology matched that of CNN, and compared to our base case. Table

19 shows that the estimated eﬀect in the 2008 election cycle of this switch is to increase

the Republican share of presidential vote intention by over 4 percentage points, an eﬀect

somewhat smaller but comparable in magnitude to the estimated eﬀect of eliminating

Fox News. This change induces welfare losses of MSNBC subscribers, who prefer the

base case ideology; however, preferences are not as strong as those of Fox viewers.

For comparison purposes, we also run this scenario for the earlier two election cycles,

showing that prior to 2005, MSNBC was a net conservative force.

Cycle Intention Intention (voters) Welfare

2000 -0.016 -0.011 -0.25

2004 -0.019 -0.013 -0.16

2008 0.043 0.030 -0.20

Table 19: Eﬀects of setting MSNBC’s ideology to match that of CNN. Column 2 is the

change in the Republican share of the presidential vote intention; column 3 rescales these to

approximate the eﬀects on registered voters only; column 4 is the average percentage change

in welfare from cable TV of MSNBC viewers.

8 Conclusion

This paper provides estimates of both the inﬂuence of slanted news on political views

and the taste for like-minded news in the context of cable television news in the U.S. The

key ingredient in the analysis is the use of channel positions as instrumental variables to

estimate a model of viewership, voting, and ideology evolution. We show instrumental

variables estimates that watching the Fox News Channel increases the probability of

voting Republican, and watching MSNBC after 2004 decreases the probability of voting

Republican in presidential elections.

We estimate a model of consumer-viewer-voters who choose cable subscriptions, al-

locate time to watching news channels, and vote in elections. The tastes for news chan-

nels are partly determined by the closeness of the news channels’ estimated ideology

to the individuals. Individual ideology evolves towards the estimated ideologies of the

news channels that a consumer watches. We use the estimated model to characterize

the degree of polarization that one can attribute to slanted cable news consumption,

and measure eﬀects on elections. Our estimates imply large eﬀects of Fox News on

presidential elections. Furthermore, we estimate that cable news does increase polar-

ization, and that this increase depends on both a persuasive eﬀect of cable news and

the existence of tastes for like-minded news.

Future research could go in a number of directions. The use of channel positions

as instrumental variable could be useful in other studies of how media consumption

aﬀects behavior. One could also use channel position variation to study the cable

news channels in more detail by examining speciﬁc programs, e.g. “The O’Reilly

Factor,” and speciﬁc issues like abortion, gay marriage, or government spending. In a

diﬀerent direction, studying the causes and consequences of the divergence in estimated

ideologies seems fruitful.

It would also be useful to test, reﬁne, or expand the speciﬁc

model we employ for belief updating after media consumption. For example, one could

allow for a joint distribution of inﬂuence parameters and tastes for like-minded news

in the population.

This includes improving these text based procedures to allow for sentiment analysis or other partisan

indicators.

References

Ansolabehere, Stephen, “CCES, Common Content, 2008,” 2011.

, Maxwell Palmer, and Amanda Lee, “Precinct-Level Election Data,” 2014.

Arceneaux, Kevin, Martin Johnson, and Chad Murphy, “Polarized political

communication, oppositional media hostility, and selective exposure,” The Journal

of Politics, 2012, 74 (01), 174–186.

Black, Eric and Tom Hamburger, “Major event, minor coverage: GOP debate

won’t make networks,” Star Tribune: Newspaper of the Twin Cities, December 1999.

Chiang, Chun-Fang and Brian Knight, “Media bias and inﬂuence: Evidence from

newspaper endorsements,” The Review of Economic Studies, 2011, 78 (3), 795.

Colman, Price, “TCI will carry Fox news channel,” Broadcasting & Cable, July 1996.

Crawford, Gregory S and Ali Yurukoglu, “The welfare eﬀects of bundling in

multichannel television markets,” The American Economic Review, 2012, 102 (2),

643–685.

DellaVigna, Stefano and Ethan Kaplan, “The Fox News eﬀect: Media bias and

voting,” The Quarterly Journal of Economics, 2007, 122 (3), 1187–1234.

DeMarzo, Peter M, Dimitri Vayanos, and Jeﬀrey Zwiebel, “Persuasion bias,

social inﬂuence, and unidimensional opinions,” The Quarterly Journal of Economics,

2003, 118 (3), 909–968.

Dempsey, John, “Spurned cable webs revile TCI,” Variety, August 1996.

Duclos, Jean-Yves, Joan Esteban, and Debraj Ray, “Polarization: concepts,

measurement, estimation,” Econometrica, 2004, 72 (6), 1737–1772.

Enikolopov, Ruben, Maria Petrova, and Ekaterina Zhuravskaya, “Media and

political persuasion: Evidence from Russia,” The American Economic Review, 2011,

101 (7), 3253–3285.

Esteban, Joan-Maria and Debraj Ray, “On the measurement of polarization,”

Econometrica: Journal of the Econometric Society, 1994, pp. 819–851.

Galesic, Mirta, Roger Tourangeau, Mick P Couper, and Frederick G Con-

rad, “Eye-Tracking data new insights on response order eﬀects and other cognitive

shortcuts in survey responding,” Public Opinion Quarterly, 2008, 72 (5), 892–913.

Gentzkow, Matthew, “Television and voter turnout,” The Quarterly Journal of

Economics, 2006, 121 (3), 931–972.

and Jesse M. Shapiro, “Media Bias and Reputation,” Journal of Political Econ-

omy, 2006, 114 (2), 280–316.

and , “Competition and Truth in the Market for News,” The Journal of Economic

Perspectives, 2008, pp. 133–154.

and , “What drives media slant? Evidence from US daily newspapers,” Econo-

metrica, 2010, 78 (1), 35–71.

and , “Ideological segregation online and oﬄine,” The Quarterly Journal of

Economics, 2011, 126 (4), 1799–1839.

and , “Measuring the Sensitivity of Parameter Estimates to Sample Statistics,”

2013. Working Paper.

Gerber, Alan S., Dean Karlan, and Daniel Bergan, “Does the Media Matter?

A Field Experiment Measuring the Eﬀect of Newspapers on Voting Behavior and

Political Opinions,” American Economic Journal: Applied Economics, 2009, 1 (2),

35–52.

Gourieroux, Christian, Alain Monfort, and Eric Renault, “Indirect inference,”

Journal of Applied Econometrics, 1993, 8 (S1), S85–S118.

Groseclose, Tim and Jeﬀrey Milyo, “A measure of media bias,” The Quarterly

Journal of Economics, 2005, pp. 1191–1237.

Holme, Petter and Mark E.J. Newman, “Nonequilibrium phase transition in the

coevolution of networks and opinions,” Physical Review E, 2006, 74 (5), 056108.

Horan, Sean, “Sequential search and choice from lists,” Unpublished paper., 2010.

Inoue, Atsushi and Gary Solon, “Two-sample instrumental variables estimators,”

The Review of Economics and Statistics, 2010, 92 (3), 557–561.

Leighley, Jan E. and Jonathan Nagler, Who Votes Now?: Demographics, Issues,

Inequality, and Turnout in the United States, Princeton University Press, 2013.

Lim, Claire S.H., James M. Snyder Jr., and David Stromberg, “The Judge,

the Politician, and the Press: Newspaper Coverage and Criminal Sentencing Across

Electoral Systems,” Working paper., 2014.

Lohse, Gerald L, “Consumer eye movement patterns on yellow pages advertising,”

Journal of Advertising, 1997, 26 (1), 61–73.

Masatlioglu, Yusufcan and Daisuke Nakajima, “Choice by iterative search,”

Theoretical Economics, 2013, 8 (3), 701–728.

McCarty, Nolan M, Keith T Poole, and Howard Rosenthal, “Income redistri-

bution and the realignment of American politics,” 1997.

Mullainathan, Sendhil and Andrei Shleifer, “The market for news,” American

Economic Review, 2005, pp. 1031–1053.

Rubinstein, Ariel and Yuval Salant, “A model of choice from lists,” Theoretical

Economics, 2006, 1 (1), 3–17.

Sanneh, Kelefa, “Twenty-Four Hour Party People: MSNBC tries to ﬁgure out what

liberals really want.,” The New Yorker, September 2013.

Smith, Anthony Alan, “Three essays on the solution and estimation of dynamic

macroeconomic models.” PhD dissertation, Duke University 1990.

Zou, Hui and Trevor Hastie, “Regularization and variable selection via the elastic

net,” Journal of the Royal Statistical Society: Series B (Statistical Methodology),

2005, 67 (2), 301–320.

A DellaVigna and Kaplan (2007)

DellaVigna and Kaplan (2007) (henceforth DVK) compare changes in presidential vote

shares in towns which had access to the Fox News Channel by the year 2000 compared

to towns that did not conditional on a rich set of co-variates. The ﬁrst-order problem

in DVK’s data is severe mis-measurement due to non-updated entries. Speciﬁcally,

37% of control group observations, the towns which the DVK data indicate did not

have cable access to the Fox News Channel in the year 2000, actually did have access

to Fox News, but were not properly updated in that data source. In fact, about 25%

of these towns already had cable access to the Fox News Channel in 1998. When one

re-runs their speciﬁcations with the Nielsen FOCUS data,

one draws new inferences

from the estimated coeﬃcients in DVK’s two preferred speciﬁcations.

In the speciﬁcation with county-level ﬁxed eﬀects, the coeﬃcient on having access

to Fox News drops from 0.00694 to 0.00215, and is no longer statistically signiﬁcant. In

the speciﬁcation with district-level ﬁxed eﬀects, the coeﬃcient on having access to Fox

News remains roughly the same. However, this speciﬁcation now performs poorly on

the placebo test that DVK used to argue that the estimate is not driven by selection of

towns into having access to the Fox News Channel. The estimated coeﬃcient of cable

availability of Fox News in 2000 on the change in Republican vote share from 1992 to

1996 is nearly the same as the estimated coeﬃcient for the change between 1996 and

2000.

A.1 The Data Problem

The data source in DVK is the Warren’s Cable and Television Factbook (henceforth

Factbook). The Factbook updates only a minority of cable systems every year. The

extent of non-updating has been documented by Crawford and Yurukoglu (2012). We

reproduce the relevant years from their Appendix table below in Table 20. Updating

is especially poor around DVK sample year. Between 1999 and 2000, only 22% of

observations were updated. Between 1998 and 1999, only 37% of observations were

These data are discussed in Section 3. We detail in the next subsection why the Factbook data are not

suitable for evaluating the eﬀects of Fox News in 2000 while the Nielsen FOCUS data are.

The former is not statistically distinguishable from the latter nor from zero. The placebo estimate

(0.00296) is closer to the actual estimate (0.0037) than it is to zero.

updated. Since Fox News was expanding across the country rapidly during these years,

this infrequent updating is consequential: many towns in the Factbook were listed as

not having cable access to Fox News, when in fact they did but the Factbook simply

wasn’t updated yet. Nearly all systems in the Nielsen FOCUS data are updated every

year.

Table 20: Data Quality of Factbook

Year Variable Number of Bundles Fraction of Bundles

1998 Total bundles 15,743 100.0%

Full information 10,872 69.0%

Updated 4,714 30.0%

Full information and updated 3,461 22.0%

1999 Total bundles 15,497 100.0%

Full information 10,444 67.0%

Updated 5,663 37.0%

Full information and updated 3,595 23.0%

2000 Total bundles 15,453 100.0%

Full information 10,312 67.0%

Updated 3,358 22.0%

Full information and updated 2,478 16.0%

2001 Total bundles 15,391 100.0%

Full information 9,793 64.0%

Updated 4,173 27.0%

Full information and updated 2,663 17.0%

2002 Total bundles 15,287 100.0%

Full information 7,776 51.0%

Updated 5,086 33.0%

Full information and updated 1,484 10.0%

1997-2007 Total bundles 166,619 100.0%

Full information 91,100 55.0%

Updated 62,299 37.0%

Full information and updated 31,493 19.0%

Notes: This table is a reproduction from Crawford and Yurukoglu (2012) indicating the degree of non-updating

in Factbook data.

One can verify that the Nielsen FOCUS data are accurate, while the non-updated

Factbook data are not. First, one can cross check the data sources against newspa-

pers from the time period. Black and Hamburger (1999) is a newspaper article from

December 2, 1999 stating that “Fox News Channel is channel 21B to subscribers in

Minneapolis.” According to the Factbook data used in DVK, Minneapolis did not have

access to the Fox News Channel by November 2000. The Nielsen FOCUS data indi-

cate that Minneapolis did have access to Fox News Channel in 1999, and also correctly

indicates the channel number of 21B. Second, we investigated the systems with the

largest discrepancy: those where Nielsen FOCUS indicated had Fox News availability

in 1998 while the Factbook indicated no availability by 2000. 353 of these systems

were operated by Tele-Communcations Inc. (TCI) in 1998. Press reports from the

time period indicate that Fox News would be available to over 90% of TCI customers

by 1998 (Colman (1996)).

Finally, the number of subscribers for Fox News implied by the Factbook data

conﬂict with the amount of viewership Fox News had in 2000, including the viewership

data used in DVK. According to DVK, “About half of the Fox News audience, therefore,

watches Fox News in ways other than via cable, possibly via satellite. This ﬁnding could

also be due to measurement error in our measure of availability via cable.” According

to their data, 17% of households were watching Fox News in 2000. Therefore, 8.5% of

all households must have been simultaneously satellite subscribers and watching Fox

News. However, the market share of satellite in the year 2000 was 11.4%

Therefore,

a vast majority of satellite subscribers must have been watching Fox News in 2000 to

be consistent with the Factbook availability measures. Our Mediamark data indicate

that the fraction of satellite subscribers watching Fox News in 2000 is only 19%.

To correct this issue, we matched the voting and demographic data in DVK to

Nielsen FOCUS. The identiﬁcation numbers in the Factbook and Nielsen FOCUS do

not match. We employed a matching procedure based on community names and ﬁrm

names, using manual inspection when matches weren’t obvious. We were able to re-

liably match 8,286 observations out of 9,256 to Nielsen FOCUS. Tables 21 and 22

compare the availability of Fox News according to the two data sources.

About 40 percent of the control group in DVK is mis-classiﬁed as not having cable

access to Fox News. About 25 percent already had access in 1998 and hadn’t been

updated for at least two years in the Factbook.

The cable market share was 70.2% implying a 81.6% total market share. Thus, about 14% of cable or

satellite subscribers were satellite subscribers.

Their viewership data and our Mediamark data agree on the aggregate 17% number. Our Mediamark

data indicate the conditional probability of watching Fox News conditional on satellite is only marginally

higher at 19%.

Factbook Fox News

(Year 2000)

0 1 Total

Nielsen Fox News

(Year 2000)

0 3,632 58 3,690

1 3,076 1,520 4,596

Total 6,708 1,578 8,286

Table 21: Year 2000: Nielsen Fox News Availability and Factbook non-updated Fox News

Availability.

Factbook Fox News

(Year 2000)

0 1 Total

Nielsen Fox News

(Year 1998)

0 4,837 358 5,195

1 1,871 1,220 3,091

Total 6,708 1,578 8,286

Table 22: Nielsen Fox News Availability in 1998 and Factbook non-updated Fox News Avail-

ability in 2000.

A.2 Estimates with Nielsen Data

We now re-run the two “benchmark” speciﬁcations from DVK: the county level ﬁxed

eﬀects regression and the US House district level ﬁxed eﬀects regression. These corre-

spond to equation (2) in DVK. Table 23 compares the resulting estimates.

The estimate in the county level ﬁxed eﬀects regression drops from a statistically

signiﬁcant 0.00694 (Column 7) to 0.00215 (Column 9), and becomes statistically in-

distinguishable from no eﬀect of Fox News on change in Republican vote share. The

diﬀerence cannot be attributed to not matching all of DVK’s observations. Their es-

timated eﬀect is stronger when using their Fox variable, but only on the subset of

matching observations (Column 8).

The estimate in the Congressional district ﬁxed eﬀects regression remains stable

with the Nielsen data. However, this speciﬁcation delivers new results in the placebo

test DVK use to argue that their speciﬁcations are estimating the causal eﬀects of Fox

News rather than selection of towns trending Republican into carrying Fox News. Table

24 compares the placebo regression estimates using the original data and the corrected

data. Using the correct data, the placebo regression indicates that availability of Fox

Republican two-party vote share change between 2000 and 1996 pres. elections

(1) (2) (3) (4) (5) (6) (7) (8) (9)

Factbook Fox 0.00798*** 0.00869*** 0.00421*** 0.00473*** 0.00694*** 0.00741***

(0.00257) (0.00270) (0.00154) (0.00163) (0.00150) (0.00158)

Nielsen Fox 0.00786*** 0.00377*** 0.00215

(0.00171) (0.00117) (0.00131)

Observations 9,256 8,286 8,286 9,256 8,286 8,286 9,256 8,286 8,286

R-squared 0.557 0.559 0.561 0.753 0.755 0.755 0.812 0.815 0.814

Data Set Factbook Factbook Nielsen Factbook Factbook Nielsen Factbook Factbook Nielsen

Sample Full Matched Matched Full Matched Matched Full Matched Matched

FE OLS OLS OLS District District District County County County

Robust standard errors in parentheses, clustered by cable ﬁrm

*** p<0.01, ** p<0.05, * p<0.1

Table 23: OLS, District FE, and County FE speciﬁcations from DVK and with corrected Fox News availability

data.

News might very well be positively correlated with the change in Republican vote

share between 1992 and 1996 in the district ﬁxed eﬀect speciﬁcation. The coeﬃcient’s

precision can not rule out a zero eﬀect, but the district ﬁxed eﬀects regression should

be interpreted in light of the placebo results.

B Comparison of Regression Coeﬃcients in Real

and Simulated Data

This section reports the ﬁt of the indirect inference estimation routine. Tables 25, 26,

and 27 report the ﬁt for the ﬁrst and second stages, respectively.

C More on Channel Positions

The combination of the placebo test of whether cable channel positions predict view-

ership of news channels on satellite together with the institutional narrative of the

period 1992-2000 is the most convincing argument for the validity of channel positions

as instrumental variables for the eﬀect of watching cable news on voting Republican.

In this section, we provide additional evidence in support of the validity for the in-

strumental variables assumption. First, we show that cable channels whose viewership

is composed of similar demographics have uncorrelated channel positions. Second, we

show that Fox News and MSNBC channel positions are highly correlated with the best

available position on the system at the time they were added. Third, we carry out

the formal test for whether local cable position eﬀects are equal for cable and satellite

subscribers.

C.1 Channel Positions for Channels with Similar Demo-

graphics

If channel positions are tailored to local tastes, one would expect channels whose view-

ership is composed of similar demographics would have correlated channel positions.

We examine this possibility in the data. The results are in Table 28. First, the Fox

News Channel position is positively correlated with the MSNBC position despite their

Republican two-party vote share change between 1996 and 1992 pres. elections

(1) (2) (3) (4) (5) (6)

Factbook Fox 0.00539 0.00459 -0.00237 -0.00271

(0.00503) (0.00507) (0.00313) (0.00325)

Nielsen Fox 0.00702** 0.00296

(0.00337) (0.00205)

Observations 4,006 3,637 3,637 4,006 3,637 3,637

R-squared 0.327 0.337 0.341 0.620 0.625 0.626

Data Set Factbook Factbook Nielsen Factbook Factbook Nielsen

Sample Full Matched Matched Full Matched Matched

Speciﬁcation OLS OLS OLS District FE District FE District FE

Robust standard errors in parentheses, clustered by cable ﬁrm

*** p<0.01, ** p<0.05, * p<0.1

Table 24: OLS and District FE Placebo speciﬁcations from DVK and with corrected Fox News availability

data.

diﬀerences in slant. This is perhaps because they are both news channels, and some

system managers prefer to group same genre channels together even if they cater to

diﬀerent segments of the population. Compare Columns (2) and (3) to see which other

channel positions correlate with Fox News and MSNBC positions. There is one chan-

nel, Bravo, which suggests endogenous positioning. During the sample period, Bravo

reformatted as a channel catering to younger and more urban viewers, including some

of the ﬁrst programming to include gay characters in lead roles. Indeed, the Bravo

position is negatively correlated with the Fox Position, and positively correlated with

the MSNBC position. On the other hand, most of the other coeﬃcients tell a diﬀerent

story. Comedy Central, which airs the liberal slanted Daily Show and Colbert Report,

is negatively correlated with MSNBC. Country Music Television (CMT) is positively

correlated with MSNBC. The Trinity Broadcasting Network (TBN) which is explicitly

religious is negatively correlated with Fox News. Column (4) shows that Speed Chan-

nel, which airs coverage of NASCAR and other motorsports, is positively correlated

with Bravo.

C.2 Best Available Channel Position

We demonstrate one example of this historical inﬂuence in Table 29. We regress the

ordinal positions of Fox News and MSNBC on the system’s best available ordinal

position in 1997, along with a control for the overall size of the system - its total

number of channels.

The best available position in 1998 is a strong predictor of

the current position, even though the positioning data here extends through 2008. A

system’s channel conﬁguration prior to the addition of Fox or MSNBC exerts a lasting

inﬂuence on the positioning of Fox and MSNBC today.

Our lineup data begins in 1998, and hence we restrict the sample for this regression to cable systems

that did not have Fox/MSNBC in 1998. “Best available” is deﬁned as the lowest open slot (unoccupied by

an existing channel) in the region of the lineup dedicated to cable (i.e. non-network and non-local-access)

channels. We deﬁne the cable region by locating the positions of CNN, ESPN, TNT, and The Discovery

Channel, and consider any open slot above at least one of those channels to be available.

C.3 Chow Test for Positions

This subsection formally carries out the Chow tests for whether the local cable position

aﬀects satellite subscribers in the same manner it aﬀects cable subscribers. This diﬀer-

ence between this regression and Tables 5 and 6 is that here we restrict the demographic

coeﬃcients to be equal for both sets of subscribers.

D Other Outcome Variables

In this section, we explore the eﬀect of cable news on other measures of political par-

tisanship aside from the surveyed Presidential vote variable. We look at self-reported

ideology on a 1 to 5 scale (Very conservative, Conservative, Moderate, Liberal, and

Very liberal) as well as precinct level actual vote tallies from the 2008 election.

D.1 Self-reported Ideology

Table 31 presents the results for self reported ideology. The Fox News eﬀect corresponds

to one-quarter of a standard deviation of self-reported ideology for an extra hour per

week. The MSNBC eﬀect is statistically imprecise. Interestingly, there is a liberal

CNN eﬀect of about one-sixth of a standard deviation for an extra hour of CNN per

week.

D.2 2008 Precinct Level Vote Shares

We regress two party vote share at the precinct level on cable news channel positions

using the data set compiled by Ansolabehere et al. (2014). The advantage of these

data is that they capture actual realized vote totals. However, because we do not have

precinct level viewership, we can not directly estimate the viewership eﬀect. Across a

variety of speciﬁcations, higher positions for Fox News predict a lower Republican vote

share, and higher positions of MSNBC predict a higher Republican vote share.

CNN-Positive FNC-Positive MSNBC-Positive

Regressor Real Simulated Real Simulated Real Simulated

CNN Position -0.087 -0.172 -0.005 0.032 -0.046 -0.013

FOX Position -0.012 0.004 -0.043 -0.261 0.041 0.030

MSN Position 0.034 -0.052 0.038 0.015 -0.001 -0.175

2000 3.837 3.520 2.791 2.281 2.844 3.210

2001 4.515 4.970 3.163 2.918 3.095 3.367

2002 4.492 4.404 3.470 3.092 2.989 3.305

2003 4.179 4.157 3.417 3.054 2.916 3.399

2004 4.333 4.098 3.699 3.178 2.960 3.724

2005 4.301 4.175 3.612 3.942 2.927 3.433

2006 4.113 3.966 3.437 3.205 2.908 3.128

2007 4.167 4.039 3.478 3.334 2.843 3.073

2008 6.446 7.141 6.897 3.864 5.878 6.033

FOX Only 0.125 -0.056 0.490 1.087 -0.450 -0.122

MSN Only -0.101 0.115 -0.548 0.088 0.096 0.774

Both Available -0.022 0.061 -0.004 1.023 -0.224 0.681

Income -0.899 -0.693 -0.244 -0.300 -0.746 -0.454

Income

0.348 0.256 -0.076 0.015 0.299 0.096

Income

-0.019 -0.014 0.048 0.016 -0.018 0.024

Age -0.020 -0.027 -0.016 -0.022 -0.011 -0.044

Age

0.001 0.001 0.001 0.001 0.000 0.001

White -0.323 -0.256 0.391 0.081 -0.139 0.059

Black 0.151 0.206 0.191 0.078 0.232 -0.227

Hispanic -0.025 0.099 -0.152 0.018 0.084 -0.082

College Graduate 0.021 -0.024 -0.179 -0.106 -0.140 -0.066

Man 0.010 0.090 0.042 0.047 -0.049 -0.075

Log Number of Channels -0.135 -0.084 -0.163 -0.089 -0.015 -0.082

Table 25: Comparison of regression coeﬃcients in real data and simulations. Dependent

variable is hours watched of each channel, conditional on watching any.

CNN-Zero FNC-Zero MSNBC-Zero

Regressor Real Simulated Real Simulated Real Simulated

CNN Position -0.013 -0.030 -0.001 -0.002 -0.001 0.001

FOX Position 0.000 -0.005 -0.018 -0.037 0.009 0.001

MSN Position -0.005 -0.005 0.011 -0.004 -0.033 -0.042

2000 -0.004 0.013 -0.235 -0.245 -0.131 -0.089

2001 0.041 0.010 -0.179 -0.246 -0.103 -0.082

2002 0.042 0.049 -0.138 -0.215 -0.123 -0.109

2003 0.016 0.027 -0.121 -0.202 -0.094 -0.084

2004 0.005 0.027 -0.091 -0.164 -0.096 -0.082

2005 0.013 0.023 -0.081 -0.214 -0.076 -0.070

2006 -0.004 0.010 -0.095 -0.194 -0.099 -0.081

2007 -0.016 0.007 -0.096 -0.181 -0.095 -0.072

2008 -0.157 -0.137 -0.225 -0.302 -0.198 -0.191

FOX Only 0.045 0.030 0.124 0.373 -0.022 0.003

MSN Only 0.083 0.027 -0.018 0.009 0.205 0.340

Both Available 0.068 0.051 0.077 0.390 0.160 0.333

Income 0.348 0.358 0.262 0.180 0.209 0.179

Income

-0.129 -0.082 -0.107 -0.034 -0.077 -0.054

Income

0.013 0.005 0.011 0.000 0.008 0.002

Age 0.003 0.005 0.004 0.004 0.003 0.002

Age

0.000 0.000 0.000 0.000 0.000 0.000

White -0.025 -0.019 0.030 0.007 -0.001 -0.023

Black 0.020 0.017 0.057 0.061 0.007 0.013

Hispanic -0.041 -0.035 -0.042 -0.033 -0.030 -0.031

College Graduate 0.063 0.064 -0.019 -0.012 0.040 0.043

Man 0.043 0.047 0.048 0.047 0.038 0.036

Log Number of Channels -0.001 -0.003 0.008 -0.012 0.004 -0.008

Table 26: Comparison of regression coeﬃcients in real data and simulations. Dependent

variable is an indicator for watching any of the channel.

Vote Intention - IV Vote Intention - OLS

Regressor Real Simulated Real Simulated

CNN Time 0.009 -0.011

FOX Time 0.138 0.125

MSN Time -0.103 -0.076

CNN Most-Watched -0.087 -0.087

FOX Most-Watched 0.329 0.331

MSN Most-Watched -0.095 -0.099

2000 0.741 0.699

2004 0.693 0.722 0.648 0.823

2008 0.682 0.772 0.649 0.854

FOX Only 0.009 -0.007 0.060 0.065

MSN Only 0.066 0.069 0.085 0.033

Both Available 0.037 0.018 0.060 0.060

Income 0.382 0.018 0.430 0.023

Income

-0.341 -0.014 -0.314 -0.006

Income

0.091 0.002 0.073 0.000

Age 0.003 0.002 0.003 0.000

Age

0.000 0.000 0.000 0.000

White 0.074 0.028 0.103 0.038

Black -0.354 -0.031 -0.286 -0.012

Hispanic -0.060 -0.018 -0.092 -0.020

College Graduate -0.060 0.004 -0.056 -0.011

Man 0.053 -0.006 0.044 0.004

Log Number of Channels -0.095 -0.064 -0.102 -0.087

Table 27: Comparison of regression coeﬃcients in real data and simulations. Dependent

variable is Republican vote intention.

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

VARIABLES Fox News Position Fox News Position MSNBC Position SPEED Position

ESPN Position 0.0377* 0.0761***

(0.0210) (0.0217)

TNT Position -0.0281 0.0391

(0.0228) (0.0279)

USA Position 9.66e-06 -0.141***

(0.0216) (0.0255)

Bravo Position -0.0449*** 0.0823*** 0.311***

(0.0112) (0.0153) (0.0271)

Comedy Central Position 0.00983 -0.0483**

(0.0194) (0.0206)

Trinity Broadcasting Network Position -0.0142** -0.00180

(0.00579) (0.00627)

CMT Position 0.00308 0.0528*** 0.0120

(0.00928) (0.0107) (0.0184)

SPEED Position -0.00318 0.00922

(0.00593) (0.00662)

N Channels 0.0184*** 0.0267*** -0.00975** 0.244***

(0.00221) (0.00434) (0.00480) (0.00835)

MSNBC Position 0.324***

(0.0161)

Observations 47,524 26,173 24,591 40,464

R-squared 0.154 0.022 0.031 0.186

Year Eﬀects Yes Yes Yes Yes

Robust standard errors in parentheses

*** p<0.01, ** p<0.05, * p<0.1

Table 28: Channel positions regressed on other channel positions.

Coeﬃcient MSNBC Fox

(Intercept) 33.8 30.7

(0.573) (0.432)

Number of Channels 0.032 0.032

(0.003) (0.002)

Best Available 0.181 0.148

(0.014) (0.012)

0.066 0.077

N 29,337 38,328

Table 29: Ordinal channel position vs. best available ordinal channel position, among sys-

tems where the channel (MSNBC or Fox News) was added in 1998 or later. Standard errors

clustered by cable system.

(1) (2) (3)

VARIABLES Fox News Hours CNN Hours MSNBC Hours

FNC Cable Position x Cable -0.122*** -0.0316 0.0500***

(0.0214) (0.0210) (0.0132)

FNC Cable Position x Satellite -0.00617 0.0328 0.00273

(0.0457) (0.0449) (0.0282)

CNN Cable Position x Cable -0.0126 -0.112*** -0.0139

(0.0137) (0.0135) (0.00849)

CNN Cable Position x Satellite 0.00759 0.0137 -0.00864

(0.0311) (0.0306) (0.0192)

MSNBC Cable Position x Cable 0.129*** 0.0718*** -0.0973***

(0.0232) (0.0228) (0.0143)

MSNBC Cable Position x Satellite 0.0511 0.0135 0.0111

(0.0500) (0.0491) (0.0309)

Observations 168,862 168,862 168,862

R-squared 0.038 0.041 0.013

F Stat for Own Position x Cable = Own Position x Satellite 5.320 14.14 10.18

P Value for Own Position x Cable = Own Position x Satellite 0.0211 0.000170 0.00142

Standard errors in parentheses

*** p<0.01, ** p<0.05, * p<0.1

Table 30: Chow Tests for whether cable position aﬀects cable subscribers and satellite sub-

scribers equally.

(1)

VARIABLES Intent to Vote R

Fox News Channel Hours 0.233***

(0.0234)

CNN Hours -0.174***

(0.0428)

MSNBC Hours -0.0300

(0.0480)

HH Income 0.00369***

(0.000622)

HH Income

-2.65e-05***

(6.59e-06)

HH Income

6.21e-08***

(1.96e-08)

Age 0.0157***

(0.00135)

Age

-0.000104***

(2.01e-05)

White 0.0236

(0.0161)

Black -0.127***

(0.0164)

Hispanic -0.00404

(0.0131)

College -0.142***

(0.0120)

Male 0.147***

(0.00734)

log(N Channels) -0.208***

(0.0101)

Observations 119,524

R-squared 0.038

Year Eﬀects Yes

Channel Availability Yes

Standard errors in parentheses

*** p<0.01, ** p<0.05, * p<0.1

Table 31: Second Stage regressions of NAES and CCES self-reported ideology on predicted

hours watched from the ﬁrst stage.

(1) (2) (3) (4) (5) (6)

VARIABLES R Vote Share R Vote Share R Vote Share R Vote Share R Vote Share R Vote Share

log(Fox News Position) -0.0100 -0.0135 -0.0106* -0.0115* -0.00650* -0.00756**

(0.00841) (0.00831) (0.00613) (0.00613) (0.00348) (0.00347)

log(MSNBC Position) 0.0229** 0.0202** 0.00653 0.00388 -0.000393 0.000428

(0.00986) (0.00968) (0.00775) (0.00750) (0.00427) (0.00440)

log(CNN Position) 0.00578 0.00352 -0.00279 -0.00268 0.00178 0.00202

(0.00643) (0.00635) (0.00466) (0.00451) (0.00265) (0.00270)

Observations 23,107 23,107 21,874 21,874 21,874 21,874

R-squared 0.129 0.154 0.478 0.493 0.710 0.717

Channel Availability Yes Yes Yes Yes Yes Yes

Demographics No No Yes Yes Yes Yes

Number of Channels Log Log Log Log Log Log

Number of OTA Broadcast Log Full Set Log Full Set Log Full Set

State FE No No No No Yes Yes

Standard errors clustered by cable system

*** p<0.01, ** p<0.05, * p<0.1

Table 32: OTA stands for “Over the Air” broadcast stations likes ABC, CBS, Fox, and NBC.