“Understanding Echo Chambers and Filter Bubbles: The Impact of

RESEARCH ARTICLE

UNDERSTANDING ECHO CHAMBERS AND FILTER BUBBLES:

THE IMPACT OF SOCIAL MEDIA ON DIVERSIFICATION AND

PARTISAN SHIFTS IN NEWS CONSUMPTION

Brent Kitchens, Steven L. Johnson, and Peter Gray

University of Virginia McIntire School of Commerce,

P.O. Box 400173, Charlottesville, VA 22904 U.S.A.

{bkitchens@virginia.edu} {[email protected]} {pet[email protected]}

Echo chambers and filter bubbles are potent metaphors that encapsulate widespread public fear that the use

of social media may limit the information that users encounter or consume online. Specifically, the concern

is that social media algorithms combined with tendencies to interact with like-minded others both limits users’

exposure to diverse viewpoints and encourages the adoption of more extreme ideological positions. Yet empi-

rical evidence about how social media shapes information consumption is inconclusive. We articulate how

characteristics of platform algorithms and users’ online social networks may combine to shape user behavior.

We bring greater conceptual clarity to this phenomenon by expanding beyond discussion of a binary presence

or absence of echo chambers and filter bubbles to a richer set of outcomes incorporating changes in both

diversity and slant of users’ information sources. Using a data set with over four years of web browsing

history for a representative panel of nearly 200,000 U.S. adults, we analyzed how individuals’ social media

usage was associated with changes in the information sources they chose to consume. We find differentiated

impacts on news consumption by platform. Increased use of Facebook was associated with increased informa-

tion source diversity and a shift toward more partisan sites in news consumption; increased use of Reddit with

increased diversity and a shift toward more moderate sites; and increased use of Twitter with little to no

change in either. Our results demonstrate the value of adopting a nuanced multidimensional view of how

social media use may shape information consumption.

Keywords: Echo chamber, filter bubble, diversity, polarization, slant, news, personalization

Introduction

Echo chambers and filter bubbles are potent metaphors that

encapsulate widespread public fear that the use of social

media may limit the information that users encounter or con-

sume online, thus failing to promote a shared experience of

free-flowing information. Specifically, the concern is that

social media algorithms combine with tendencies to interact

with like-minded others to create an environment that pre-

dominantly exposes users to congenial, opinion-reinforcing

content to the exclusion of more diverse, opinion-challenging

content. This intuitive understanding of echo chambers and

filter bubbles is well-accepted. Yet empirical evidence about

how social media and other digital platforms shape informa-

tion consumption is inconclusive. Despite decades of interest

in this phenomenon, researchers’ ability to analyze the preva-

lence or formation of echo chambers and filter bubbles is

stymied by a lack of consensus regarding their conceptuali-

zation and measurement.

There is growing concern that social media and other infor-

mation discovery platforms promote information-limiting

environments by shielding users from opinion-challenging

information, thereby encouraging users to adopt more extreme

Shaila Miranda was the accepting senior editor for this paper. Ofir Turel

served as the associate editor.

DOI: 10.25300/MISQ/2020/16371 MIS Quarterly Vol. 44 No. 4 pp. 1619-1649/December 2020 1619

Kitchens et al./Echo Chambers and Filter Bubbles

ideological positions. Indeed, there are aspects of popular

social media platforms that may foster information-limiting

environments. For instance, as with offline relationships,

people in online social networks interact most frequently with

like-minded others. In a study of 10.1 million U.S. Facebook

users with self-reported ideological affiliation, Bakshy et al.

(2015) found that more than 80% of these Facebook friend-

ships shared the same party affiliation. Accordingly, the news

and information sources that individuals discover through

their social relationships may also reflect a lack of ideological

diversity. Further, even within this already constrained choice

set, selective exposure theory predicts that individuals prefer

to consume opinion-reinforcing news sources over opinion-

challenging ones (Frey 1986). Garrett (2009) found support

for this tendency in a field experiment with 727 online news

readers: individuals expressed interest in reading online news

stories they perceived to be supportive of their existing

opinion and expressed disinterest in consuming opinion-

challenging stories. Finally, researchers have long expressed

concern about the potential for algorithmic filtering to reduce

the diversity of information sources that individuals are

exposed to, engage with, or consume (Van Alstyne and Bryn-

jolfsson 1996, 2005). Personalization technology is sensitive

to personal preferences; once a user engages with opinion-

reinforcing content, algorithmic filtering may constrain further

exposure to a narrower, more closely aligned range of content

(Pariser 2011; Stroud 2010). This, in turn, may foster the

adoption of more extreme opinions (Festinger 1964; Hart et

al. 2009).

Thus, there is a clear potential for the use of social media to

be associated with a narrowing of information diversity and a

partisan shift in the slant of news consumed by their users.

For example, Lawrence et al. (2010) found readers of political

blogs to be more ideologically segregated and more ideologi-

cally extreme than nonreaders. Likewise, Wojcieszak and

Mutz (2009) found that participants in online politics-related

groups were less likely to be exposed to political information

they disagree with (and more likely to be exposed to infor-

mation they agree with) than participants in many other

categories of online groups (e.g., hobby or leisure related).

Nonetheless, the prevalence and magnitude of information-

limiting environments may be overstated. Other studies have

shown social media platforms to be information-expanding.

Social media helps users discover new information sources,

thereby potentially expanding the diversity of viewpoints,

opinions, and information to which users are exposed. For

example, Flaxman et al. (2016) “uncover evidence for both

sides of the debate, while also finding that the magnitude of

the effects [are] relatively modest” (p. 298). Gentzkow and

Shapiro (2011) performed a large-scale analysis of web

browsing habits of U.S. adults, concluding that “ideological

segregation on the Internet is low in absolute terms” and that

the Internet exposes individuals to a broader range of view-

points than “face-to-face interactions with neighbors, co-

workers, or family members” (p. 1799). In summary, despite

a great deal of interest in the idea of echo chambers and filter

bubbles, empirical evidence to date is inconclusive as to

whether social media is information-limiting or information-

expanding.

We identify six key reasons contributing to why researchers

have reached conflicting conclusions about the existence of

echo chambers and filter bubbles, as well as the processes that

may lead to their formation. First, a fundamental issue in this

body of research is the lack of conceptual clarity, with vague

and conflicting definitions of constructs, processes, and

outcomes. Second, imprecise conceptualization has only

compounded issues of inconsistent measurement and

incommensurate research designs in prior studies, precluding

the systematic integration of findings. Most importantly, a

variety of online platforms have been studied either indepen-

dently or lumped together in aggregate, making it difficult to

ascertain the relationship between social factors, technology

features, and information-limiting environments. Third, empi-

rical results are also difficult to compare when they measure

disparate, nonequivalent outcomes, often measuring content

exposure, sharing, or generation, while consumption may be

more directly salient to the concerns surrounding information-

limiting environments. Fourth, as Shore et al. (2018) note,

one “likely reason for the conflicted nature of the literature is

that earlier work has generally focused too narrowly on

unrepresentative or incomplete data sets,” such as “focusing

on highly active users” (p. 850). Fifth, inferring the impact of

individuals’ technology use from analysis of population-level

distributions may suffer from either an ecological fallacy or an

aggregation bias (Freedman 1999). Finally, social media plat-

forms frequently adjust their algorithmic filters and rarely

disclose when those changes occur. The dynamic nature of

the phenomenon combined with a lack of transparency by

platform providers may limit the generalizability of empirical

results over time.

To further understand this much-debated phenomenon, we

propose the concept of information-limiting environments as

encapsulating the primary concerns regarding echo chambers

and filter bubbles—namely, that social network homophily

and algorithmic filtering constrain the information sources

that individuals choose to consume, shielding them from

opinion-challenging information and encouraging them to

adopt more extreme viewpoints. By identifying diversity and

partisan slant as distinct characteristics of information con-

sumption, we articulate how social media can shape informa-

tion consumption in ways that move beyond the simple

presence or absence of echo chambers and filter bubbles. We

use broadly representative data to investigate how three

1620 MIS Quarterly Vol. 44 No. 4/December 2020

Kitchens et al./Echo Chambers and Filter Bubbles

popular social media platforms vary in their impact on content

consumption by individual users, and thereby may or may not

contribute to underlying outcomes associated with

information-limiting environments. Through an analysis of

over four years of web browsing history for a representative

panel of nearly 200,000 U.S. adults, we demonstrate differ-

entiated impacts by platform and discuss these differences in

terms of the variation in platform features at the time.

Increased use of Facebook was associated with increased

information source diversity and a partisan shift in news site

visits. Increased use of Reddit was also associated with

increased diversity, but a moderating shift. Increased use of

Twitter had little to no association with changes in either

diversity or slant. Our results demonstrate the value of

adopting a more nuanced and multidimensional conceptua-

lization of how social media shapes information consumption.

Understanding Echo Chambers

and Filter Bubbles

Given the extensive range of approaches to conceptualizing,

measuring, and identifying echo chambers and filter bubbles,

they can be difficult to apply as precisely definable constructs

for academic research. To better understand these two terms,

it is useful to consider the genesis of each. The concerns arti-

culated by Cass Sunstein, a primary voice warning of echo

chambers (Sunstein 2001, 2017), originate from earlier

findings that group polarization—the individual and collective

adoption of more extreme views—occurs in intensive deliber-

ations by small groups of homogenous individuals on topics

of high relevance and concern (Sunstein 2002). He noted that

“widespread error and social fragmentation are likely to result

when like-minded people, insulated from others, move in

extreme directions simply because of limited argument pools

and parochial influences” (Sunstein 2002, p. 186). Extrapo-

lating from findings regarding small group dynamics, Sunstein

(2001) predicted that algorithmic filtering would also lead to

group polarization on a larger scale. Likewise, Eli Pariser’s

influential book The Filter Bubble (2011) predicted that

individualized personalization by algorithmic filtering would

lead to intellectual isolation and social fragmentation. His

thesis is succinctly captured in the book’s subtitle: What the

Internet Is Hiding from You. Both authors drew caution from

a scenario that Negroponte (1996) positively presented as the

“daily me”: information carefully selected to match individ-

ual preferences. But in contrast to Negroponte, both Sunstein

and Pariser made equally dire predictions that the use of

information discovery platforms would lead to information-

limiting environments with negative individual and societal

impacts.

Although both Sunstein and Pariser were concerned that

social media and other information discovery platforms would

shape what information individuals choose to consume—and,

ultimately, an individual’s viewpoints and opinions—they

differed on how this might happen. Pariser identified person-

alization technology as the primary mechanism, voicing

concern that it strengthens individual preferences for seeking

out opinion-reinforcing information to the exclusion of

opinion-challenging information (Frey 1986; Garrett 2009).

Pariser stressed the negative impacts of individual isolation in

creating epistemic bubbles where personal viewpoints persist,

unchallenged and untested. Sunstein stressed the potential of

technology to reinforce fragmentation at a larger scale, one

where people are not individually isolated, but instead form

groups in which individuals with similar ideological predilec-

tions interact exclusively with each other. Sunstein argued

that online interactions can reinforce ideological segregation

and thereby facilitate limited information pools that strengthen

preexisting biases, promote groupthink, and encourage adop-

tion of even more extreme viewpoints. In this way, Sunstein

was concerned with interactions among homogenous groups

that share a common social identity. Alternatively, Pariser

was concerned with individual isolation and a lack of shared

information in opinion formation. From Pariser’s perspective,

an individual isolated in their own personalized information

bubble still may suffer from the negative impacts of limited

information, even if this isolation makes them immune from

the social pressures that reinforce group solidarity and

engender polarizing groupthink. Despite these differences in

perspectives, both Pariser and Sunstein argued that the

remedy for information-limiting environments is individual

consumption of ideologically diverse content that encom-

passes opinion-challenging viewpoints.

These conceptualizations of echo chambers and filter bubbles

are reactionary, in that they portray not the creation of an

observable outcome, but rather the absence of an idealized

one. This ideal is not well defined, however, other than the

general normative assertion that individuals should be

exposed to and consume opinion-challenging information.

The quantity of opinion-challenging information that an

individual should consume and the degree to which it should

challenge their opinions is left ambiguous, rendering it all but

impossible to determine if someone is actually within an echo

chamber or filter bubble. Further, it is doubtful whether this

ideal of a well-informed public with rigorously examined

opinions has ever existed. For example, before the creation

of Internet-enabled personalization technology, news and

information was disseminated largely via newspapers, maga-

zines, radio, and TV broadcasts that catered to ideologically

diverse audiences. Even then, “a number of studies …

indicate[d] that persuasive mass communication functions far

more frequently as an agent of reinforcement than as an agent

MIS Quarterly Vol. 44 No. 4/December 2020 1621

Kitchens et al./Echo Chambers and Filter Bubbles

of change” (Klapper 1960, p. 15). With commonly vague and

reactionary conceptualizations, echo chambers and filter

bubbles are certainly powerful metaphors, but are ill-defined

for use as rigorous academic constructs.

Echo chambers and filter bubbles are also often depicted as

something that an individual (or group) either is in, or is not—

a simplistic binary that is unnecessarily reductionist.

Focusing solely on whether or not individuals consume

opinion-challenging information sources neglects more

nuanced ways in which technology use may shape the con-

sumption of information sources. For example, even though

other impacts are quite plausible, Bakshy et al. (2015)

considered only the consumption of ideologically discordant

content (e.g., a liberal consuming conservative news or vice

versa) in concluding that algorithmic ranking has limited

impact on news consumption.

Furthermore, this emphasis on consumption of opinion-

challenging and ideologically discordant information may be

misplaced. Groupthink, extreme viewpoints, and intellectual

isolation can persist even when opinion-challenging informa-

tion is readily available. Individuals and groups frequently

adopt extreme opinions despite being confronted with chal-

lenges to their viewpoints. For example, studies have found

that exposure to divergent viewpoints may fail to have a

moderating effect and can even catalyze a polarizing backlash

that hardens existing ideological positions (Bail et al. 2018;

Stroud 2010). Further, some individuals—particularly highly

politically-engaged ones—may purposely seek out ideolo-

gically discordant information sources to gain awareness of

opposing viewpoints while remaining highly antagonistic to

them (Shore et al. 2018), a condition referred to as affective

polarization (see Iyengar et al. 2019).

In summary, discourse regarding echo chambers and filter

bubbles has often been both reactionary and reductionist.

Consumption of ideologically discordant content that includes

opinion-challenging viewpoints is neither a necessary nor a

sufficient remedy for the fundamental ills associated with

echo chambers and filter bubbles. Thus, we conclude that

although Sunstein and Pariser identified key mechanisms that

may contribute to information-limiting environments, a richer

conceptualization of potential outcomes is needed to under-

stand how social media platforms can impact the content that

users choose to consume.

Constituent Characteristics of Echo

Chambers and Filter Bubbles

Although there is no consensus definition for echo chambers

or filter bubbles, in considering the range of descriptions we

identify two constituent characteristics that stand out. The

first is a lack of information diversity due to restriction of

information sources. In echo chambers, “individuals are

exposed only to information from like-minded individuals”

(Bakshy et al. 2015, p. 1130), that “confirms their previously-

held opinions” (Shore et al. 2018, p. 850), and “is devoid of

other viewpoints” (Garrett 2009, p. 279). Filter bubbles are

a “unique universe of information for each of us” (Pariser

2011, p. 9), “devoid of attitude-challenging content” (Bakshy

et al. 2015, p. 1130), where “individuals only see posts that

they agree with” (Lazer 2015, p. 1090). Reduced information

diversity exaggerates confirmation bias—the individual and

collective tendency to seek out information that supports

preexisting beliefs (Nickerson 1998). It also facilitates ideo-

logical groupthink—a collective manifestation of closed-

mindedness and an overestimation of the value of collective

beliefs that are reinforced by pressure towards uniformity

(Janis 1982). Narrowing of information sources is prob-

lematic, as “exposure to differing political views increases

people’s knowledge of rationales for political perspectives

other than their own and also fosters political tolerance”

(Mutz and Martin 2001, p. 140).

Second, both echo chambers and filter bubbles are commonly

characterized by ideological segregation (the tendency of

individuals to associate with others who share their view-

points) and by partisan polarization (the adoption of more

extreme views). Echo chambers are associated with “frag-

mentation of users into ideologically narrow groups” (Shore

et al. 2018, p. 850), with “political fragmentation and social

polarization” (Garrett 2009, p. 278), and with “segregation by

interest or opinion [that] will … increase political polari-

zation” (Dubois and Blank 2018, pp. 1-2) and “foster social

extremism” (Barberá 2015, p. 86). Similarly, filter bubbles

are a “centrifugal force pulling us apart” (Pariser 2011, p. 10),

“in which algorithms inadvertently amplify ideological segre-

gation” (Flaxman et al. 2016, p. 299). In this way, the

increasing ability to interact online is viewed not as a unifying

force but, rather, one that may tear apart the fabric of society

as individuals adopt more extreme views.

In summary, we argue that while echo chambers and filter

bubbles are potent, flexible metaphors that have broadly

captured the public’s imagination and serve as a distillation of

widespread fears, their vague and disparate conceptualizations

make them difficult to study. To make these attractive meta-

phors concrete, we propose that research into information-

limiting environments requires a more nuanced focus on the

separate characteristics of information source diversity and

information source slant. Information source diversity

reflects separation, variety, and disparity among information

sources an individual consumes (see Harrison and Klein

2007). A change in the consumption of information sources

1622 MIS Quarterly Vol. 44 No. 4/December 2020

Kitchens et al./Echo Chambers and Filter Bubbles

may lead to a broadening increase or a narrowing decrease in

information source diversity. Information source slant

reflects the dominant ideological perspective provided by

information sources an individual consumes. A change in

consumption of information sources may lead to a partisan

shift towards more extreme information sources or a moder-

ating shift toward more centrist ones. Our objective is to

examine how social media use may be associated with

changes in the diversity and slant of the information sources

an individual chooses to consume.

Categorizing Changes in

Information Consumption

Information Source Diversity

Information source diversity is an area of interest across aca-

demic disciplines, including management, communications,

and political science. Receiving information from diverse

sources can help individuals better understand the world

around them, develop more robust opinions, and make better

decisions (Jehn et al. 1999; Mutz and Martin 2001; Van

Alstyne and Brynjolfsson 1996). Information source diversity

captures the idea that the value of information sources derives

from providing nonredundant information and alternative

viewpoints. There are many ways to conceptualize and mea-

sure diversity (see Harrison and Klein 2007; Page 2010).

Most approaches involve summarizing how elements of

interest, such as unique information sources, are allocated

along a continuum or across a set of categories (McDonald

and Dimmick 2003). We identify constructs related to the

diversity of information sources an individual chooses to con-

sume, following Harrison and Klein’s (2007) dimensions of

separation, variety, and disparity. Together, these constructs

provide a holistic view of information source diversity (see

Table 1 for a summary). We also consider a precursor to

these measures: information source quantity (the number of

unique information sources an individual consumes). Al-

though this construct does not reflect how information sources

may vary in the kinds of information they provide, as the

quantity of information sources increases, so does the poten-

tial for nonredundant information.

Separation describes the degree of diversity along a single

lateral dimension (Harrison and Klein 2007)—such as the

dimension of political ideology. Differentiation by ideolo-

gical slant is particularly evident in the contemporary U.S.

media landscape (Jurkowitz et al. 2020) where there is

increasing separation of information sources by dominant

ideological perspective (Flaxman et al. 2016). This speciali-

zation allows reliable categorization of information sources

based on the predominant viewpoint of typical content

(Gentzkow and Shapiro 2011; Shore et al. 2018). Information

source dispersion reflects ideological separation among a set

of information sources an individual chooses to consume

(Shore et al. 2018). When an individual consumes informa-

tion sources from a narrow ideological range, they are less

likely to be exposed to a diversity of viewpoints and opinions

than when they read sites encompassing a broader range.

However, dominant ideological perspectives may be insuf-

ficient for understanding differences among information

sources. For example, some news and information sources are

not overtly ideological, and even among those with similar

ideological slant, there are other bases for differentiation.

Drawing on the concept of interorganizational competition

and niche overlap theory (Burt 1992; Sohn 2001), we con-

ceptualize information source variety as reflecting how likely

a set of information sources is to provide nonredundant

information along an assortment of dimensions. Similar to the

idea of brokerage across structural holes in a communication

network (Burt 1992), when an individual consumes multiple

information sources with minimally overlapping audiences,

they are likely to be exposed to more nonredundant infor-

mation and divergent viewpoints than if they consumed

multiple information sources with substantial audience over-

lap. Thus, increased information source variety reflects the

consumption of a broader range of knowledge, expertise, and

unique information.

Finally, disparity complements separation and variety as an

additional form of diversity. Disparity is frequently used to

assess income inequality, but may also be applied to the distri-

bution of attention (e.g., Li et al. 2019). In considering

consumption of multiple information sources, an increasing

disparity of the time spent per source reflects a concentration

of attention that effectively reduces diversity. For example,

if an individual regularly visits five information sources but

spends the preponderance of their time with only two of them,

they are less likely to consume nonredundant information than

when they give all five equal attention. Thus, information

source parity reflects equality of attention given to different

information sources.

Information Source Slant

In addition to considering changes in information source

diversity, we focus on the complementary, yet distinct, con-

cept of changes in information source slant—where slant

represents the dominant ideological perspective of the infor-

mation sources an individual consumes. This concept relates

to the conclusion that interaction among like-minded

individuals can lead to more extreme (less centrist) opinions

(Sunstein 2009). Such individual-level change is a form of

MIS Quarterly Vol. 44 No. 4/December 2020 1623

Kitchens et al./Echo Chambers and Filter Bubbles

Table 1. Dimensions of Information Consumption

Dimension Construct Description Categorization

Diversity

Information source

quantity

Quantity of unique information sources

consumed

None

Information source

dispersion

Separation of information sources along a

single ideological dimension

Slant (continuous)

Information source

variety

Variety of information among information

sources

Information source audience

overlaps (continuous)

Information source parity

Equality of time spent consuming

information sources

Time (continuous)

Slant Information source slant

Dominant ideological perspective among

information sources

Slant (continuous)

Diversity

and slant

Cross-cutting content

consumption

Proportion of information sources with

opposing viewpoints

Slant (discrete categories)

polarization that reflects a process of shifting or strengthening

“one’s original position” (Stroud 2010, p. 557) “further in the

direction of … original views” (Mutz 2006, p. 227). Yet, a

partisan shift is but one possibility; the slant of news sources

consumed could also shift in a centrist direction. We define

a shift in slant in terms of differences in an individual’s

consumption choices between points in time. In comparing

the information sources consumed at various times, a change

in the dominant ideological position towards a more centrist

perspective represents a moderating shift and a change

towards the extreme represents a partisan shift.

The concept of polarization (and its less frequently articulated

converse, moderation) is multifaceted. We identify three

alternative conceptualizations of polarization that encompass

antecedents or consequences of individual information source

consumption. First, social polarization is reflected in par-

ticular patterns of interactions and affiliations. For example,

the preference of individuals to maintain social relationships

with like-minded others manifests as modularity in social

networks (Baldassarri and Bearman 2007). This homophilous

tendency, which has been observed in Facebook friendships

(Bakshy et al. 2015) and in interactions among Twitter users

(Bright 2018), may impact individuals’ consumption

behaviors. Second, attitudinal polarization represents seg-

mentation of beliefs and attitudes (Baldassarri and Bearman

2007; Boxell et al. 2017). For example, individuals are more

trusting of and more likely to consume information sources

that align with their ideological views (Adamic and Glance

2005; Jurkowitz et al. 2020). At the group or population

level, this may be observed as audience fragmentation in

information sources that individuals engage with and consume

(Jacobson et al. 2016; Lawrence et al. 2010). Third, affective

polarization is a phenomenon of animosity between

individuals in opposing political parties that stems from parti-

sanship as a social identity (Iyengar et al. 2019). Individuals

with high affective polarization may frequently engage with

or consume opinion-challenging content, but remain un-

changed in or even strengthen their views as a result. This is

consistent with users who “read at least some information

from both sides of the political spectrum, but only tweet out

information consistent with their own side” (Shore et al. 2018,

p. 852).

In summary, polarization is an individual-, group-, and

population-level phenomenon encompassing static states and

dynamic processes (Mutz 2006). In the discussion of echo

chambers and filter bubbles, polarization is a frequently raised

theme, yet often without precise conceptualization or distinc-

tion among its many forms. Social, attitudinal, and affective

polarization are all potential antecedents or consequences of

the content a user is exposed to, engages with, and consumes.

To better understand how social media platforms may influ-

ence information consumption, we focus on the individual-

level behavioral outcome of partisan or moderating shifts in

slant of information a user consumes across periods of varied

platform use.

Cross-Cutting Content Consumption

Cross-cutting content consumption measures the extent to

which an individual consumes information sources that

provide ideologically discordant information (Bakshy et al.

2015). Assuming that an individual’s attitudes and behaviors

are grounded in and reflect a dominant ideological viewpoint,

an information source is considered cross-cutting when there

is sufficient ideological difference between the slant of the

information source and the individual’s dominant viewpoint

(Lawrence et al. 2010). When cross-cutting consumption is

high, an individual is regularly consuming information

sources that are opinion-challenging. A low amount of cross-

1624 MIS Quarterly Vol. 44 No. 4/December 2020

Kitchens et al./Echo Chambers and Filter Bubbles

Change in

Information

Source

Diversity

Change in Information Source Slant

Moderating Shift

H2B: more centrist No Change

Partisan Shift

H2A: more extreme

Narrowing

H1A: less

diverse

No change

Broadening

H2B: more

diverse

In each scenario, the gray bubble represents the original mean slant (left-right position) and diversity (size) of news consumed by an example

user. This user’s original consumption is liberal, as it is left of the center as denoted by the vertical hash along the axis. The white bubble

represents the changed consumption of the end user under a treatment condition (e.g., in a period of higher social media use).

Figure 1. Categorizing Changes in Information Consumption

cutting consumption is consistent with being in an

information-limiting environment. However, a focus on

cross-cutting content can distract researchers from under-

standing more modest shifts in consumption, such as a highly

conservative user reading more moderately conservative

news.

Decreasing consumption of cross-cutting information sources

results in both narrowed information diversity and a partisan

shift. Alternatively, increasing cross-cutting content con-

sumption concurrently results in both a broadening of infor-

mation diversity as well as a moderating shift in information

source slant. Because the cross-cutting consumption construct

conflates both diversity and slant of information sources, we

argue that it is a poor measure of either and that it fails to

capture the nuance provided by their separate measurement.

Categorizing Information Source Consumption

Considering changes in the diversity and slant of information

sources as two separate dimensions provides a richer cate-

gorization of the potential impacts of social media platforms

on information source consumption. As depicted in Figure 1,

three possible changes in information source diversity

(narrowing, no change, broadening) can combine with three

possible changes in information source slant (moderating

shift, no change, partisan shift) for nine potential scenarios

(see Figure 1).

Explicitly considering changes in both dimensions is essential

for moving beyond a simplistic conclusion that information-

limiting environments either do or do not exist and leads to a

more nuanced understanding of how social media use could

be associated with a wide range of outcomes. Of the scen-

arios shown in Figure 1, any of those involving a narrowing

decrease in information source diversity (the top row), as well

as those involving a partisan shift (the right-hand column),

could be considered as information-limiting environments; yet

each has unique characteristics and implications. Research

into how technology shapes information consumption would,

therefore, be enhanced by independently considering changes

in both information source diversity and slant when devel-

oping theory, designing studies, and selecting measures.

MIS Quarterly Vol. 44 No. 4/December 2020 1625

Kitchens et al./Echo Chambers and Filter Bubbles

Information Source Consumption

To frame our investigation of how social media platforms

shape the information sources an individual chooses to con-

sume, we provide a general model of how social network

homophily, algorithmic filtering, and individual behavioral

responses can lead to variation in information source con-

sumption. We then detail how this model relates to three

popular social media platforms. Finally, we hypothesize how,

based on these factors, social media use may be associated

with changes in information diversity and slant in information

source consumption.

As depicted in Figure 2, social network homophily (path A)

and algorithmic filtering (path B) are two key determinants of

the information sources a social media to which a user is

exposed. Individuals’ social networks vary widely in size, the

density of shared ties, and the frequency of interactions.

Likewise, social media users vary in their social networks,

including who they are connected with on a platform. The

preponderance of information sources shared on social media

is generated by platform users. Yet, there is wide variability

among platforms in how user-generated content is organized,

the extent to which algorithmic filtering is used in prioritizing

which content is displayed, and the factors considered in that

prioritization. Once a platform presents a user with a preview

of potential information sources, the user has the opportunity

(path C) to engage on that platform and/or to “click-through”

to the external source itself. Additionally, an individual’s

response to content exposure can serve as an input for subse-

quent algorithmic filtering (path E) and influence who that

user affiliates with in the future (path D).

Next, we build on Figure 2 to discuss how social media plat-

forms vary in the prioritization of information sources that

users are exposed to. As noted in Table 2, Facebook is an

example of a platform with high levels of social network

homophily and extensive algorithmic filtering. Facebook

determines what content to show users based on an estimated

likelihood of engagement (Bakshy et al. 2015; Vaidhyanathan

2018). In deciding what to present, Facebook chooses from

among recent posts made by others in a user’s social network

(e.g., by Facebook “friends”). Facebook considers an indi-

vidual’s past engagement history as well as the overall

popularity of content. Thus, algorithmic filtering on Face-

book is highly personalized: even if two users have an

identical set of Facebook friends, what each user sees may

vary considerably based on their prior engagement with

similar content.

Twitter also has a high level of social network homophily.

Twitter users build their online social network by identifying

other Twitter accounts to “follow.” For the first 10 years of

its existence (from 2006 to mid-2016), Twitter presented a

primary timeline of Tweets posted by others in a user’s social

network (those whom they “follow”) in reverse chronological

order. During this period, algorithmic filtering was minimal:

two users who followed the same Twitter accounts would see

the same content, in the same order. Recently, Twitter has

implemented more extensive algorithmic filtering (similar to

Facebook). However, the data analyzed in this study was col-

lected prior to this implementation, and therefore provides a

contrast to Facebook.

Finally, Reddit is an example of a platform that is primarily

interest-based, with content shared by users through a hier-

archy of topic-based communities (a.k.a. subreddits). Users

self-identify their interests by joining these communities. Users

may directly communicate with one another, but these

connections are not used to filter content. Indeed, Reddit

performs minimal algorithmic filtering, merely prioritizing

content based on popularity determined by up- and down-

votes. A Reddit user can choose to sort and filter content

based on recency or popularity, but members of the same

communities see the same content by default (Jürgens and

Stark 2017).

In summary, because social media platforms embody a wide

variety of features and uses, it is not reasonable to make

blanket statements regarding how their use may produce

information-limiting environments. It is more useful to discuss

how—through the interplay of social networks, algorithmic

filtering, and individual choices—the mechanisms in Figure 2

may be associated with changes in diversity and slant in the

information sources an individual chooses to consume.

Because of conflicting theoretical claims and mixed empirical

evidence to date, we divide studies into rival camps and

articulate competing hypotheses (e.g., Gray and Cooper 2010).

While only one of each pair of hypotheses can survive a given

empirical test, the reality that there are likely many different

boundary conditions—such as differentiated effects by

technology platform—suggests the need for a systematic

program of research over time to identify the conditions under

which each effect may dominate (Burton-Jones et al. 2017).

Narrowing Diversity

The argument that social media use is associated with nar-

rowing diversity in information source consumption begins

with the observation that social networks frequently exhibit

ideological homophily (e.g., Bakshy et al. 2015; Himelboim et

al. 2013). Individuals are more likely to affiliate with others

who share similar experiences, perspectives, and opinions. To

the extent that individuals are part of ideologically segregated

social networks, the content posted by others

1626 MIS Quarterly Vol. 44 No. 4/December 2020

Kitchens et al./Echo Chambers and Filter Bubbles

Figure 2. Social Media Use and Information Source Consumption

Table 2. Determinants of Prioritization of Information Sources by Social Media Platform

Facebook Twitter Reddit

Social Network Homophily Platform social graph High High Low

Algorithmic Filtering

Engagement history High Low

†

Low

Content popularity High Low High

Topic-based Low Low High

†

Low for chronological Twitter feed before mid-2016, high for curated feed initiated in mid-2016.

with whom they interact on social media may also be ideolo-

gically constrained (path A in Figure 2). For example, a study

of news links in partisan Facebook pages showed that com-

menters in both liberal- and conservative-oriented groups

posted links to a small number of information sources, with

little overlap between the two groups (Jacobson et al. 2016).

To the extent that a social media platform algorithmically

filters or prioritizes content based on users’ prior engagement

(path B), users may also be presented with a narrower range

of content (Bozdag and van den Hoven 2015; Pariser 2011).

When this constrains information sources available to engage

with (path C) it can also create a feedback loop (path D, E)

that further constrains variation in subsequent exposure

(Prawesh and Padmanabhan 2011). Algorithmic filtering and

prioritization based on general popularity can also lead to a

narrowing of information sources in a rich-get-richer dynamic

(Welch et al. 2011). Further, among information sources to

which users are exposed, individuals’ preference to engage

with opinion-confirming sources rather than opinion-

challenging ones—referred to as selective exposure theory

(e.g., Garrett 2009)—may reduce engagement with and con-

sumption of diverse information sources (paths C). Indeed,

Schmidt et al. (2017) found “that the more active a user is, the

more the user tends to focus on a small number of news

sources” (p. 4). Bakshy et al. (2015) concluded that restricted

information diversity occurs in part due to algorithmic filter

effects on exposure (path B), but more so because of indi-

vidual choices in engagement and consumption (paths C, D,

E). Thus, we propose:

H1a: Social media use is associated with a nar-

rowing in the diversity of information

source consumption.

Broadening Diversity

A contrasting argument holds that the social media use may be

associated with a broadening of information diversity in infor-

mation source consumption. Although the argument follows

similar logic and pathways as the narrowing effect (H1a), it

arises from different assumptions about ideological segrega-

tion in online social networks, associated content exposure,

and individual preferences in content engagement and

consumption.

There is some evidence that ideological segregation may be

less prevalent in online relationships than it is in offline ones.

As such, online social networks may even broaden the ideo-

MIS Quarterly Vol. 44 No. 4/December 2020 1627

Kitchens et al./Echo Chambers and Filter Bubbles

logical range of social interactions (path A). For example,

Goel et al. (2010) found that people’s ties on Facebook often

included others who held different ideological views—for

example, through familial, professional, and hobby-related

associations. These diverse social interactions can increase

individuals’ incidental exposure to ideologically diverse infor-

mation sources (path A) (Fletcher and Nielsen 2016).

Platforms that filter and curate information sources relative to

topics of interest (path B) must cater to a variety of users.

Heterogeneity of user interests provides incentives for these

algorithms to increase the diversity of information sources.

This works against the rich-get-richer popularity bias men-

tioned previously and leads to increased variety (Fortunato et

al. 2006). This rationale applies to generic filtering algo-

rithms, but not necessarily to personalized filtering (path G,

D). Personalized filtering algorithms are essentially sophisti-

cated recommender systems, a domain in which the novelty-

accuracy tradeoff has long been a well-known and researched

issue—intuitively, “it would be almost always correct, but

useless, to recommend bananas, bread, milk, and eggs” to

grocery shoppers (Herlocker et al. 2004, p. 14). Under-

standing the value of novel information, recommender

systems are commonly designed with diversity, serendipity,

coverage, and related objectives in mind, all of which would

increase the breadth of information sources to which users are

exposed (Kaminskas and Bridge 2016).

Finally, as a platform for information discovery, a funda-

mental purpose of social media is to provide access to

information sources otherwise unknown to users. Because

individuals have a limited capacity for search and exploration,

it has long been recognized that intermediate tools are needed

in order to overcome the difficulty of finding information

(Bakos 1997). Although there are a variety of use cases for

social media, the value of these platforms is broadly derived

from their ability to connect users with content that is infor-

mative, engaging, or entertaining. It stands to reason, there-

fore, that platform usage may be associated with discovery of

novel information sources that enhance an individual’s

information source diversity. This assertion is supported by

the finding that news site visits from social media platforms

were more diverse than direct visits (Flaxman et al. 2016). In

summary, we propose:

H1b: Social media use is associated with a

broadening in the diversity of information

source consumption.

Partisan Shift

The argument that social media use is associated with a

partisan shift in information source consumption is difficult to

disentangle from the argument for the narrowing of infor-

mation diversity (H1a above). It is therefore unsurprising that

the consumption of cross-cutting content, which shifts both

information diversity and slant, is viewed as an antidote to

information-limiting environments. Indeed, researchers often

discuss a narrowing of diversity and a partisan shift in infor-

mation sources as occurring together. For example, the large-

scale study of Facebook usage by Schmidt et al. (2017) con-

cluded that “users [tend] to limit their exposure to a few sites”

and “there is major segregation and growing polarization in

online news consumption” (p. 4). Moreover, an important

argument for a partisan shift follows a logic similar to that for

narrowing of diversity: exposure to opinion-reinforcing

information is high due to ideological segregation (path A),

individuals prefer opinion-conforming information (path C),

and algorithms reinforce these preferences (paths B, E).

Bakshy et al. (2015, p. 2) found the “factor decrease in the

likelihood that an individual clicks on a cross-cutting article

relative to the proportion available in News Feed to be 17%

for conservatives and 6% for liberals” (path C). This reduc-

tion in cross-cutting content consumption contributes to both

a narrowing of information diversity as well as a partisan

shift.

We identify three distinct arguments for how partisan shifts

may occur independently of narrowing. First, individuals who

primarily interact with like-minded others online (path A) are

likely to experience the same group solidarity and in-group

identification that occur offline, with a polarizing effect on

ideological beliefs (e.g., Schkade et al. 2007; Sunstein 2009).

Second, a Twitter experiment by Bail et al. (2018) demon-

strated that exposure to opposing viewpoints resulted in a

partisan shift in engagement (path C), albeit only for conser-

vatives. However, Bail et al.’s experimental conditions were

extreme and may not represent either typical use of social

media or typical patterns of ideological content exposure.

Third, social media platforms may include partisan content

created and promoted by zealots or malicious actors at-

tempting to convince others or actively sow discord through

Facebook groups, Twitter hashtags, and Reddit subreddits

(e.g., Allcott and Gentzkow 2017; Lazer et al. 2018). In

summary, we propose:

H2a: Social media use is associated with a

partisan shift in the slant of information

source consumption.

Moderating Shift

However, other research suggests the opposite effect. Just as

narrowing of information diversity and a partisan shift are

often complementary, so too is the opposite combination of

1628 MIS Quarterly Vol. 44 No. 4/December 2020

Kitchens et al./Echo Chambers and Filter Bubbles

broadening diversity and a moderating shift. The studies in

support of broadening of information diversity (in H1b above)

often assume that broadening occurs through exposure to

opinion-challenging, cross-cutting content.

To the extent that individuals’ online social networks are

diverse, users may experience increased exposure to moder-

ating news sources (path A). Based on experimental manipu-

lation of social endorsements and online news consumption

choices, Messing and Westwood (2014) concluded that “the

mere presence of social endorsements reduced partisan

selectivity to levels indistinguishable from chance” (p. 1056).

Thus, when an individual observes others in their social net-

work engaging with content—a form of social endorsement—

they are more likely to engage with it themselves, regardless

of perceived slant. Selective exposure to opinion-confirming

content appears to be even lower when mediated by online

platforms than through more direct interaction with others,

causing users to encounter information with dissimilar poli-

tical views (Mutz and Martin 2001, p. 98).

Filtering and prioritization algorithms (path B) may also show

users more mainstream, moderate content because it is typi-

cally more popular and commercially viable (Cooper 2003).

Particularly if a platform is ad-supported, there may be incen-

tive to avoid polarizing content in order to appease advertisers

(Gabszewicz et al. 2002). For similar reasons, platforms may

also choose to present a balance of ideological perspectives

(Fletcher and Neilsen 2018b).

There is also direct empirical evidence consistent with a

moderating effect of social media use. For example, a

population-level analysis of technology use found that the

demographic groups most likely to use social media were also

the least likely to be ideologically segregated (Boxell et al.

2017). Shore et al. (2018) offer empirical evidence of a

potential moderating effect of platform usage, finding that

Twitter users typically share news links that are more centrist

than the news links to which they are exposed (thereby

creating more opportunity for others to encounter centrist

news through). Thus, we propose:

H2b: Social media use is associated with a

moderating shift in the slant of information

source consumption.

Data and Methods

Research Setting

To investigate how social media use is associated with infor-

mation source consumption, we obtained data from Comscore,

a leading provider of digital audience measurement services.

They recruit and compensate a demographically represen-

tative sample of active U.S. Internet users who install an

apparatus that automatically records and reports granular

Internet usage. This apparatus enables Comscore to gather

accurate data about the timing and duration of all web pages

visited by each panelist, including use of social media and

visits to news sites. Companies widely trust this data to accu-

rately reflect the reach and effectiveness of marketing efforts,

similar to Neilsen ratings for television. Although they

usually report audience data in aggregate format, we obtained

de-identified user-level clickstream browsing data for over 3

million U.S. adults who were panelists between January 2012

and June 2016.

Panel membership rotates, with existing users leaving and new

users joining the panel at variable intervals. To understand

within-person behavioral trends over time, we limited our

analysis to users who remained on the panel for at least 365

consecutive days. Although social desirability bias could be

a concern as users are aware of the monitoring they agreed to,

we empirically established that panel users continued to

browse all manner of sites and conclude that it is unlikely that

users systematically altered their browsing behaviors.

A limitation of our data set is that it only includes browsing

behavior originating from desktop and laptop computers

(PCs); the data set does not include mobile device usage.

While it would be optimal to have a view of online usage that

encompasses mobile devices, this data set can still provide

robust insights into the association between social media use

and news site visits. An analysis of similar Comscore data by

Mitchell and Jurkowitz (2014) found that online news con-

sumption patterns were similar for PCs and mobile devices.

Also, we conducted an additional survey to better understand

mobile and PC device visits to social media and news sites.

We randomly surveyed 426 individuals through Qualtrics

(results available upon request), asking respondents how often

they read news and access social media on desktops or laptops

(PC), as well as smartphones or tablets (mobile). The number

of users who access news on various devices is almost iden-

tical, with 82% of users reading news on a PC at least weekly,

compared to 83% reading news on mobile devices at least

weekly. The difference is only slightly more pronounced for

accessing social media, with 88% doing so on mobile at least

weekly, compared to 73% by PC. Further, we found no evi-

dence of significant substitution of one device type for the

other. It is important to note that this survey was conducted

in February 2020, whereas our data encompasses panelists’

web site visits from 2012 through mid-2016. Mobile device

usage has increased significantly from the years 2012–2016,

while PC usage has largely remained stable, particularly for

news consumption (Walker 2019). In summary, it is reason-

MIS Quarterly Vol. 44 No. 4/December 2020 1629

Kitchens et al./Echo Chambers and Filter Bubbles

Table 3. Independent Variables

Measure Description

Facebook referrals

Number of visits to news sites attributable as referrals from each platform; that is, web pages in

the domains facebook.com, reddit.com, and twitter.com (count)

Reddit referrals

Twitter referrals

Time on Facebook

Duration of all visits to each platform (hours)

Note: In all models, this is Log

transformed such that the coefficient represents the impact on

the odds ratio of a doubling of time spent.

Time on Twitter

Time on Reddit

Direct visits Number of direct visits to news sites (count)

Total time online Total duration of all website visits (hours). Transformed Log

in all models.

Table 4: Outcome Measures for an Individual’s Within-Period News Site Visits

Dimension Construct Measure Description Minimum Maximum

Diversity

Information

source

quantity

Distinct news

sites

Number of unique news

sites visited (count)

No news site visits

Visit all 177 news

sites in the sample

Information

source

dispersion

Slant

dispersion

The time-weighted standard

deviation of political slant of

news sites visited

All visited news sites

have the same

political slant

50% of visits to the

most liberal news

site, 50% to the

most conservative

Information

source parity

Reverse Gini

Index

(1 - Gini Index)*100

calculated based on time

spent on each visited news

site (0 to 100 range)

One news site

predominates with

minimal time on

others

An equal amount

of time spent at

multiple news sites

Information

source variety

Audience

variety

Time-weighted mean variety

of news site visits based on

the frequency of overlapping

site visitors (0 to 100 range)

Visit a small set of

news sites that have

the same readership

Visit multiple sites

with minimal

audience overlap

Slant

Information

source slant

Mean slant

Time-weighted average of

political slant for visits to

any of 177 news sites (-100

to 100 range; lower values

more liberal, higher values

more conservative)

Either all visits to

single most-centrist

site or visits to a

perfectly balanced

set of sites

All visits to either

the most liberal

news site or the

most conservative

news site

Diversity

and slant

Cross-cutting

content

consumption

Cross-cutting

proportion

Time-weighted percentage

of news sites visited with

political slant scores

opposite to a user’s base

ideology (0 to 100 range)

All news sites visits

in the same category

of political slant

A user spends a

significant amount

of time at news

sites opposing

their base ideology

able to conclude that the data represents a meaningful portion

of users’ news browsing and social media activity (possibly

even a significant majority given the period of data collec-

tion).

Measures and Analysis

To test our hypotheses, we estimated fixed effects within-

person models over a panel comprised of 4-week periods. For

each individual in our sample, we calculated measures per

nonoverlapping four-week period of their tenure (see Tables

3 and 4 for a summary of measures). To better understand

information source consumption, we focused on information

diversity and slant of news consumption among active users.

Limiting our analyses to user-periods containing at least one

news site visit resulted in an unbalanced panel of 185,548

individuals with a total of 1,096,480 user-period observations

for an average of 5.9 observations per individual in our

sample. Summary statistics and correlations are reported in

1630 MIS Quarterly Vol. 44 No. 4/December 2020

Kitchens et al./Echo Chambers and Filter Bubbles

the Appendix. In such a within-person model, the results may

be interpreted as estimated differences for an individual,

comparing periods when they use social media platforms more

to those periods in which that same individual uses them less.

This design avoids the potentially problematic comparison of

light and heavy platforms users to each other and helps to

isolate the impacts of changes in an individual’s platform use

from other characteristics of the user.

The independent variables relate to the use of three social

media platforms (Facebook, Twitter, and Reddit) that provide

links to news sites. We measured the number of visits to news

sites occurring via referrals from each platform to assess the

effect of these referrals on diversity and slant. Also, we mea-

sured the number of visits made directly to the home page of

a news site (for instance, by typing the URL of the news site

directly into the browser navigation bar). Because these

direct visits are made independent of exposure to platform-

prioritized information sources, they provide a baseline

against which to assess the impact of diversity and slant of

news site visits referred by social media platforms. Beyond

the primary effect of news site referrals, we also expected that

general usage of these platforms may influence news con-

sumption, and therefore also measured the time individuals

spent on each platform, as well as their total time online.

Finally, as a control for use of other popular information

discovery platforms, we include time and count variables for

individuals’ use of email and search.

Information Diversity

We adopted four distinct but complementary measures that

together provide a holistic view of information source diver-

sity of news site visits, as described in Table 4. First, the

number of distinct sites visited by each user during each a

period measures information source quantity. Next, we

measure slant dispersion as separation on the horizontal

dimension of partisan slant by calculating the time-weighted

standard deviation of the slant of news sites visited by a user

during a period (measurement of slant itself is detailed in the

next section). Intuitively, this represents the spread of a user’s

news consumption along the ideological continuum. Third,

we measure information source parity as the vertical distribu-

tion of news source consumption, calculated as the reverse-

coded Gini index of time spent by a user on each distinct news

site within a period.

If a user visits 10 diverse sites, but

spends 99% of their time on one of these, the level of diver

sity is still effectively low, which will be captured by this

measure.

Information source variety represents diversity across any

number of dimensions or categories. From the perspective of

news, this could be by a topic (politics, sports, celebrity),

presentation mode (short-form prose, long-form prose, info-

graphic, video), exposition (fact reporting, opinion, editorial),

etc. Instead of attempting to identify all possible underlying

factors of variety, we used the site visit history of all 3 million

panel members to estimate these latent factors that drive users

to visit different sites. The intuition behind this measure is

that sets of news sites with higher levels of audience variety—

that is, a lower overlap of actual audiences—are more likely

to provide diverse information. To calculate this measure, we

first defined a co-visitation network of all users to all news

sites. For each pair of news sites p and q, we calculated v

the number of users who visited both sites, and defined an

intermediate measure of the audience overlap of these sites,

/ min(v

, v

); note 0 # v

/ min(v

, v

) # 1.

Time-weighted

audience variety is then calculated across all pairs of news

sites visited by user i in period t:

Partisan Shift

Hypotheses 2a and 2b concern the relationship between social

media use and change in the slant of information sources

consumed. For outcomes related to slant, we based the

measurement on slant scores for 177 commonly visited online

news sites published in Shore et al. (2018), as detailed in the

Appendix. We calculated the time-weighted average slant of

online news sites visited by a user during a given period. Our

scores are scaled from their range of ±3.602 to a range of

±100 for ease of exposition. Because a change in partisan

slant is relative to a referent base ideological position for each

user, we also categorized individuals as being in conservative,

centrist, or liberal terciles of our sample by calculating a time-

weighted average slant score per individual based on all of

their news sites visits while on the panel. We used a categori-

cal, rather than continuous, measure following prior research

that has shown qualitative differences between liberals and

conservatives (e.g., Bakshy et al. 2015; Bail et al. 2018).

The Gini index is a measure of concentration, so it must be reverse coded

to reflect increasing diversity.

The choice of min(v

, v

) prevents relatedness scores from being biased

downward when comparing sites with large discrepancy in visitation rates.

We also used the maximum and average of v

and v

with similar results.

We are fortunate to be able to calculate this measure due to the rich nature

of our data. Harrison and Klein (2007) suggest the Blau index as a more

generic measure of variety, which we also tested with results consistent with

the measures presented.

MIS Quarterly Vol. 44 No. 4/December 2020 1631

Kitchens et al./Echo Chambers and Filter Bubbles

Figure 3. Alternative Measures of Cross-Cutting Content (Illustrated for Liberal Users)

We acknowledge that assigning a single score to describe the

political slant of all content on a given news site is a simplifi-

cation. Any given article on a news site may express a

political ideology that is well-aligned or misaligned with the

perceived or actual ideology typical for that site. However,

we build on prior work showing that slant scores provide

insight into general user behaviors (Gentzkow and Shapiro

2011; Shore et al. 2018). Indeed, a survey conducted by Pew

Research Center of 12,043 adults concludes that “deep

partisan divisions exist in the news sources Americans trust,

distrust and rely on” (Jurkowitz et al. 2020, p. 4). Thus, when

users choose to consume information sources, the dominant

ideological position of the information source (in this case, a

news site) is a salient factor in their decision.

To further confirm the external validity of assigning slant

scores to online news sites, we surveyed 426 individuals

regarding their political affiliation and likelihood to read news

from a set of websites. Analysis of these survey responses

demonstrates a clear relationship between an individual’s

political affiliation and their likelihood to visit various news

sites. The correlation between the slant scores of news sites

and the coefficients of simple linear regressions of political

affiliation on the likelihood to visit them is 0.73. These

results (available upon request) suggest that individuals per-

ceive the site that hosts a news article as a strong signal of

potential alignment with their political affiliation.

Cross-Cutting Content Consumption

Finally, we calculated the percentage of cross-cutting content

consumed by each user. As discussed earlier, changes in this

measure reflect a change in both diversity and slant. Because

there is no generally accepted approach for identifying cross-

cutting content, we calculated the proportion of content in

three distinct ways as illustrated in Figure 3. These alterna-

tive measurements incorporate different thresholds for how

different an information source’s slant needs to be from a

user’s baseline position in order to be considered cross-

cutting. First, for users in conservative or liberal terciles, we

calculated the proportion of content they consumed that is in

the opposite tercile (i). Second, we calculated the proportion

of content users in these terciles consume that was at least

across the median (ii). Third, we calculated the proportion of

content users in these terciles consumed in the opposite or

centrist tercile (iii).

Results

Information Diversity

To test H1, we used fixed effects within-person models to

estimate the association between the use of popular social

media platforms and four measures of information diversity

(results shown in Tables 5 and 6). All models control for

direct news site visits and total time spent online, as well as all

time-invariant user characteristics removed via the fixed effect

within transformation. Because of the relatively large sample

size, as suggested by Lin et al. (2013) we report coefficient

confidence intervals for each of our models across a range of

subsample sizes in Appendix Figure A1. Additionally, we

present an intuitive analysis of effect sizes in the discussion

section.

For the first measure of information diversity, the number of

distinct news sites visited within a period (Table 5, Panel A),

we found that each referral from a social media platform to a

news site was associated with an increase in the number of

distinct news sites visited by an individual within the same 4-

week time period. By comparison, the estimated association

between the number of direct visits and total distinct news site

visits was very low. This result is consistent with the intuitive

understanding that, whereas direct site visits are more likely

to be return visits to a previous site, individuals are exposed

to additional information sources through their use of social

media. Nonetheless, we found significant variation among

platforms in the strength of the association between referrals

1632 MIS Quarterly Vol. 44 No. 4/December 2020

Kitchens et al./Echo Chambers and Filter Bubbles

Table 5: Association of Platform Use and Information Source Diversity: Quantity and Separation

Panel A: DV = Distinct News Sites Panel B: DV = Slant Dispersion

I II III IV I II III IV

Facebook referrals

0.268*** 0.261*** 0.299*** 0.260***

(0.0012) (0.0012) (0.0052) (0.0053)

Reddit referrals

0.884*** 0.767*** 0.354*** 0.196***

(0.0051) (0.0055) (0.0215) (0.0235)

Twitter referrals

0.506*** 0.476*** 0.014 -0.014

(0.0077) (0.0078) (0.0328) (0.0332)

Search referrals

0.418*** 0.407*** 0.377*** 0.343***

(0.0009) (0.0009) (0.0039) (0.0039)

Email referrals

0.086*** 0.085*** 0.039*** 0.035**

(0.0025) (0.0025) (0.0106) (0.0106)

Direct visits

0.018*** 0.012*** 0.018*** 0.012*** 0.001 -0.005*** 0.001 -0.004***

(0.0001) (0.0001) (0.0001) (0.0001) (0.0004) (0.0004) (0.0004) (0.0004)

Time on Facebook

(Log

of hours)

0.197*** 0.062*** 0.562*** 0.432***

(0.0033) (0.0029) (0.0123) (0.0124)

Time on Reddit

(Log

of hours)

1.803*** 0.804*** 1.418*** 1.090***

(0.0164) (0.0156) (0.0611) (0.0665)

Time on Twitter

(Log

of hours)

0.425*** 0.186*** 0.229*** 0.116**

(0.0110) (0.0097) (0.0410) (0.0414)

Time on search

(Log

of hours)

0.572*** 0.240*** 0.988*** 0.707***

(0.0047) (0.0042) (0.0177) (0.0178)

Time on email

(Log

of hours)

-0.135*** -0.062*** -0.209*** -0.143***

(0.0046) (0.0040) (0.0171) (0.0171)

Total time online

(Log

of hours)

0.614*** 0.418*** 0.289*** 0.297*** 1.301*** 1.126*** 0.680*** 0.688***

(0.0024) (0.0021) (0.0032) (0.0028) (0.0089) (0.0089) (0.0119) (0.0119)

Intercept

-0.019 0.318*** 0.373*** 0.459*** 0.325*** 0.620*** 1.062*** 1.130***

(0.0118) (0.0101) (0.0119) (0.0103) (0.0432) (0.0429) (0.0443) (0.0440)

Observations 1,096,480 1,096,480 1,096,480 1,096,480 1,096,480 1,096,480 1,096,480 1,096,480

Individuals 185,548 185,548 185,548 185,548 185,548 185,548 185,548 185,548

R-squared 0.0920 0.3349 0.1273 0.3404 0.0231 0.0384 0.0306 0.0423

Standard errors shown in parenthesis below estimated coefficients.

*p < 0.05, **p < 0.01, ***p < 0.001

and distinct news sites. For example, for every 10 Reddit

referrals there are almost 8 distinct news sites visited during

the period, whereas 10 Facebook referrals relates to less than

3 new distinct news sites visited.

We also found that time spent on each platform was signi-

ficantly and positively associated with the number of distinct

news sites visited, with relative estimated effect sizes among

platforms substantially similar to that of referrals. The sole

exception was time spent on email, which was associated with

a slight reduction in the number of distinct news sites visited.

Because time variables are base-2 logged, their coefficients

represent the effect of a doubling of time spent on each

platform. A user doubling their typical amount of time spent

using Reddit visited 0.8 additional distinct sites in a period,

while a user doubling time on Facebook visited only 0.06

additional distinct sites.

We report the results for slant dispersion in Panel B of Table

5. Among social media platforms, we found that the use of

Reddit was associated with the most substantial magnitude

increases in slant dispersion, particularly for time spent on

each platform. Facebook referrals and time on Facebook

both had moderate positive associations with slant dispersion.

MIS Quarterly Vol. 44 No. 4/December 2020 1633

Kitchens et al./Echo Chambers and Filter Bubbles

Table 6. Association of Platform Use and Information Source Diversity: Variety and Parity

Panel C: DV = Audience Variety Panel D: DV = Reverse Gini Idex

I II III IV I II III IV

Facebook referrals

0.700*** 0.557*** 0.887*** 0.823***

(0.0200) (0.0203) (0.0104) (0.0105)

Reddit referrals

0.699*** 0.098 0.824*** 0.425***

(0.0822) (0.0897) (0.0426) (0.0465)

Twitter referrals

-0.231 -0.368** 0.373*** 0.259***

(0.1251) (0.1266) (0.0648) (0.0656)

Search referrals

1.758*** 1.591*** 1.432*** 1.355***

(0.0148) (0.0150) (0.0077) (0.0078)

Email referrals

0.144*** 0.138*** 0.518*** 0.513***

(0.0405) (0.0405) (0.0209) (0.0210)

Direct visits

0.016*** -0.005** 0.015*** -0.003 0.059*** 0.039*** 0.058*** 0.040***

(0.0017) (0.0017) (0.0017) (0.0017) (0.0009) (0.0009) (0.0009) (0.0009)

Time on Facebook

(Log

of hours)

1.861*** 1.524*** 1.093*** 0.664***

(0.0469) (0.0473) (0.0247) (0.0245)

Time on Reddit

(Log

of hours)

4.572*** 4.108*** 3.587*** 2.730***

(0.2334) (0.2538) (0.1228) (0.1315)

Time on Twitter

(Log

of hours)

1.055*** 0.643*** 1.160*** 0.654***

(0.1566) (0.1579) (0.0824) (0.0818)

Time on search

(Log

of hours)

5.021*** 3.758*** 2.813*** 1.716***

(0.0675) (0.0680) (0.0355) (0.0352)

Time on email

(Log

of hours)

-1.438*** -1.110*** -0.693*** -0.474***

(0.0654) (0.0653) (0.0344) (0.0338)

Total time online

(Log

of hours)

7.761*** 7.072*** 5.152*** 5.161*** 4.781*** 4.145*** 3.233*** 3.256***

(0.0340) (0.0342) (0.0456) (0.0453) (0.0179) (0.0177) (0.0240) (0.0235)

Intercept

5.869*** 6.981*** 8.859*** 9.119*** -3.339*** -2.273*** -1.519*** -1.260***

(0.1650) (0.1639) (0.1691) (0.1680) (0.0869) (0.0848) (0.0890) (0.0870)

Observations 1,096,480 1,096,480 1,096,480 1,096,480 1,096,480 1,096,480 1,096,480 1,096,480

Individuals 185,548 185,548 185,548 185,548 185,548 185,548 185,548 185,548

R-squared 0.0549 0.0718 0.0640 0.0769 0.0797 0.1264 0.0909 0.1304

Standard errors shown in parenthesis below estimated coefficients.

*p < 0.05, **p < 0.01, ***p < 0.001

Finally, Twitter use was associated with minimal change in

slant dispersion. In comparing the analyses for distinct site

visits (Panel A) and slant dispersion (Panel B), the R

for the

latter is lower than the former. Nonetheless, there is nearly a

doubling in R

from Panel B Model I including only direct

visits and total time online to Panel B Model IV also

including social media usage. In all, we conclude that the

relationships between slant dispersion and the use of Reddit

and Facebook are modest but nontrivial.

Panels C and D in Table 6 provide results for the information

source diversity measures of audience variety and reverse

Gini index, respectively. Among the social media platforms,

we found again that the use of Reddit was associated with the

most substantial magnitude increase in both measures, fol-

lowed by Facebook. Similarly, Twitter was associated with

minimal changes. In summary, in considering all four diver-

sity measures, we conclude that for Facebook and Reddit

there is consistent support of hypothesis H1b, predicting a

1634 MIS Quarterly Vol. 44 No. 4/December 2020

Kitchens et al./Echo Chambers and Filter Bubbles

broadening association between platform use and information

source consumption. For these platforms, there is no support

of the alternative H1a. With inconsistent sign and signifi-

cance as well as lower magnitude of observed effects

associated with Twitter usage, we find weak, mixed support

for either H1a (narrowing) and for H1b (broadening).

Partisan Shift

To measure shifts in the slant of news consumed, we divided

users into conservative, centrist, and liberal terciles and used

a fixed effects within-person model to estimate the relative

mean slant of news consumed by these users during periods of

higher or lower usage of each social media platform.

Each

variable is interacted with a tercile indicator variable, with the

main effect of the tercile indicator subsumed by the fixed

effect. This effectively allows separate models for each

tercile to be estimated simultaneously within one model. With

slant scores ranging from negative (indicating liberal) to

positive (indicating conservative), a positive coefficient for

users in the conservative tercile indicates a partisan shift (i.e.,

a shift from conservative to more conservative), whereas a

negative coefficient indicates a moderating shift. The oppo-

site signs apply for the liberal tercile.

As shown in Table 7, an increase in referrals from and time

spent using Facebook were associated with a partisan shift in

the slant of news consumed. In periods when conservative

individuals used Facebook more, their news site visits were

more conservative. The same applies to liberal users, but with

lower magnitude. For Reddit, increased referrals were asso-

ciated with a moderating shift for conservative users, yet there

was no association found for liberal users. Twitter usage had

no significant association with changes in slant. Overall,

these results support H2a for Facebook, where greater use was

associated with a partisan shift in information source con-

sumption. The results support H2b for Reddit, with greater

use associated with a moderating shift, albeit for conservative

users only. However, R

values are very low, meaning that

the vast majority of variance in the mean slant of information

sources consumed by users remains unexplained by their

social media usage.

Cross-Cutting

As a final analysis, we estimated the relationship between

social media use and the percentage of cross-cutting content

consumption using three classifications for cross-cutting

content. The three panels in Table 8 (F, G, and H) correspond

with the classifications of cross-cutting illustrated in Figure 3

as alternatives (i), (ii), and (iii), respectively. The results are

largely consistent for all three measurement approaches.

all three cases, Facebook referrals were associated with a

decrease in the percentage of cross-cutting content consumed

by users. However, as highlighted by a shaded band in

Table 8, for time spent on Facebook, results vary depending

upon what threshold is applied to define cross-cutting.

Results ranged from significantly positive for the most strin-

gent categorization (i.e., cross-cutting news is only news from

the opposite side) to significantly negative for the least strin-

gent categorization (i.e., cross-cutting news also includes a

large segment of centrist content). We find little to no evi-

dence for an association of Reddit or Twitter with cross-

cutting.

It is also interesting to note that, although cross-cutting con-

sumption conflates both diversity and slant, the estimated

effects for cross-cutting were more closely aligned with esti-

mated effects for shifts in slant, particularly for Facebook and

the reference platforms of search and email. The R

values

were also very low, which more closely resembles those for

shifts in slant. This stands in contrast to prior research that

considers cross-cutting consumption as a measure of diversity

(e.g., Bakshy et al. 2015). Our findings support the conclu-

sion that a holistic view encompassing separate dimensions of

information source diversity and slant provides a more com-

plete and nuanced understanding of information consumption

than considering cross-cutting content consumption alone.

Discussion

Summary of Findings

In summary, we found that the use of different social media

platforms had varying associations with changes in infor-

mation consumption. Tables 9 and 10 summarize estimated

referral and time coefficients across platforms for each

dependent variable. These coefficients were drawn from the

fully specified Model IV of each panel and are shaded by

magnitude. For comparison we also included the results for

search and email. Use of Facebook and Reddit were each

associated with increases in information source diversity

across all measures, with the estimated effects of Reddit

significantly larger than those of Facebook. Use of Twitter

We tested a variety of divisions, including median split, quartiles, quintiles,

sextiles, and deciles, with varying sizes for the centrist component, finding

results consistent across a wide range of measurement choices.

We also calculated these measures with alternatives that varied both the size

of the partisan groups as well as the categorization of cross-cutting content.

Those results are consistent with variations reported here.

MIS Quarterly Vol. 44 No. 4/December 2020 1635

Kitchens et al./Echo Chambers and Filter Bubbles

Table 7. Association of Platform Use and Mean Slant

Panel E: DV = Mean Slant

Measure Tercile I II III IV

Facebook referrals

Liberal -0.078*** -0.069***

Centrist -0.053* -0.033

Conservative 0.396*** 0.384***

Reddit referrals

Liberal -0.002 -0.019

Centrist -0.242** -0.130

Conservative -0.784*** -0.599**

Twitter referrals

Liberal -0.055 -0.073

Centrist -0.162 -0.110

Conservative -0.132 -0.116

Search referrals

Liberal 0.195*** 0.184***

Centrist 0.074*** 0.076***

Conservative -0.273*** -0.261***

Email referrals

Liberal -0.066* -0.064*

Centrist -0.032 -0.033

Conservative 0.425*** 0.423***

Direct visits

Liberal -0.002 -0.005 -0.002 -0.005

Centrist 0.007** 0.006** 0.007** 0.006**

Conservative 0.016*** 0.015*** 0.016*** 0.016***

Time on Facebook

(Log

Hours)

Liberal -0.160*** -0.137**

Centrist -0.161*** -0.158***

Conservative 0.360*** 0.207***

Time on Reddit

(Log

Hours)

Liberal 0.114 0.121

Centrist -0.847*** -0.713**

Conservative -1.275*** -0.731

Time on Twitter

(Log

Hours)

Liberal 0.142 0.121

Centrist -0.231* -0.235*

Conservative -0.217 -0.170

Time on search

(Log

Hours)

Liberal 0.388*** 0.269***

Centrist 0.080 0.022

Conservative -0.617*** -0.416***

Time on email

(Log

Hours)

Liberal -0.188** -0.144*

Centrist -0.042 -0.022

Conservative 0.394*** 0.247***

Time online

(Log

Hours)

Liberal 0.131*** 0.094** 0.08 0.074

Centrist 0.010 -0.006 0.065 0.063

Conservative -0.188*** -0.183*** -0.175*** -0.156***

Intercept Intercept -4.315*** -4.289*** -4.325*** -4.325***

Observations 1,096,480 1,096,480 1,096,480 1,096,480

R-squared 0.0004 0.0020 0.0007 0.0021

Number of individuals 185,548 185,548 185,548 185,548

Standard errors omitted for space and can be provided upon request.

*p < 0.05, **p < 0.01, ***p < 0.001

1636 MIS Quarterly Vol. 44 No. 4/December 2020

Kitchens et al./Echo Chambers and Filter Bubbles

Table 8. Association of Platform Use and Consumption of Cross-Cutting News

Panel F: DV = Cross-Cutting (user tercile

to opposite news site tercile)

Panel G: DV = Cross-Cutting (user tercile

to opposite news site half)

Panel H: DV = Cross-Cutting (user tercile

to centrist or opposite news site tercile)

I II III IV I II III IV I II III IV

Facebook

referrals

-0.083*** -0.099*** -0.203*** -0.205*** -0.362*** -0.339***

(0.0121) (0.0123) (0.0161) (0.0163) (0.0170) (0.0173)

referrals

0.036 0.012 0.062 0.047 0.144* 0.092

(0.0497) (0.0544) (0.0659) (0.0722) (0.0699) (0.0765)

Twitter

referrals

-0.049 -0.037 -0.010 -0.001 0.009 -0.008

(0.0756) (0.0767) (0.1003) (0.1018) (0.1064) (0.1080)

referrals

0.104*** 0.096*** 0.177*** 0.167*** 0.251*** 0.240***

(0.0090) (0.0091) (0.0119) (0.0121) (0.0126) (0.0128)

referrals

-0.119*** -0.115*** -0.139*** -0.134*** -0.180*** -0.179***

(0.0245) (0.0246) (0.0324) (0.0326) (0.0344) (0.0345)

Direct visits

-0.012*** -0.013*** -0.012*** -0.013*** -0.015*** -0.016*** -0.015*** -0.016*** -0.017*** -0.017*** -0.017*** -0.017***

(0.0010) (0.0010) (0.0010) (0.0010) (0.0013) (0.0014) (0.0013) (0.0014) (0.0014) (0.0014) (0.0014) (0.0014)

Time on

Facebook

0.166*** 0.199*** -0.064 0.008 -0.453*** -0.331***

(0.0283) (0.0288) (0.0376) (0.0381) (0.0398) (0.0405)

Time on

0.178 0.168 0.141 0.098 0.424* 0.342

(0.1406) (0.1539) (0.1865) (0.2042) (0.1979) (0.2166)

Time on

Twitter

-0.098 -0.109 -0.041 -0.072 0.170 0.129

(0.0944) (0.0959) (0.1252) (0.1272) (0.1329) (0.1349)

Time on

0.253*** 0.184*** 0.390*** 0.272*** 0.531*** 0.364***

(0.0407) (0.0413) (0.0540) (0.0548) (0.0573) (0.0581)

Time on

-0.150*** -0.111** -0.257*** -0.196*** -0.295*** -0.208***

(0.0394) (0.0396) (0.0523) (0.0526) (0.0555) (0.0558)

Time online

0.369*** 0.353*** 0.239*** 0.234*** 0.201*** 0.182*** 0.134*** 0.124*** -0.076** -0.093** -0.050 -0.066

(0.0204) (0.0207) (0.0275) (0.0275) (0.0271) (0.0275) (0.0365) (0.0365) (0.0287) (0.0291) (0.0387) (0.0387)

Intercept

6.873*** 6.886*** 7.010*** 7.014*** 15.750*** 15.758*** 15.787*** 15.794*** 29.713*** 29.704*** 29.623*** 29.631***

(0.0992) (0.0994) (0.1022) (0.1022) (0.1316) (0.1318) (0.1355) (0.1355) (0.1396) (0.1398) (0.1438) (0.1437)

Observations 1,092,010 1,092,010 1,092,010 1,092,010 1,092,010 1,092,010 1,092,010 1,092,010 1,092,010 1,092,010 1,092,010 1,092,010

Number of

Individuals

185,282 185,282 185,282 185,282 185,282 185,282 185,282 185,282 185,282 185,282 185,282 185,282

R-squared 0.0005 0.0007 0.0006 0.0008 0.0002 0.0006 0.0003 0.0006 0.0002 0.0011 0.0004 0.0012

Standard errors shown in parenthesis below estimated coefficients.

*p < 0.05, **p < 0.01, ***p < 0.001

MIS Quarterly Vol. 44 No. 4/December 2020 1637

Kitchens et al./Echo Chambers and Filter Bubbles

Table 9. Summary of Estimated Coefficients for Changes in Information Source Diversity

Distinct News Sites Slant Dispersion Audience Variety Gini Index

Hyp.

SupportReferrals Time Referrals Time Referrals Time Referrals Time

Facebook 0.261 0.062 0.260 0.432 0.557 1.524 0.823 0.664 H1a

Reddit 0.767 0.804 0.196 1.090 4.108 0.425 2.73 H1a

Twitter 0.476 0.186 0.116 -0.368 0.643 0.259 0.654 Mixed/low

Search 0.407 0.240 0.343 0.707 1.591 3.758 1.355 1.716

Email 0.085 -0.062 0.035 -0.143 0.138 -1.110 0.513 -0.474

Darker shading represents increasing absolute magnitude of coefficients within DV (coefficients with p > 0.05 omitted).

Green cells with black text are positive estimated coefficients, and orange cells with white text negative.

Table 10. Summary of Estimated Coefficients for Changes in Information Source Slant

Liberal Tercile Centrist Tercile Conservative Tercile Hyp.

Support

Referrals Time Referrals Time Referrals Time

Facebook -0.069 -0.137 -0.158 0.384 0.207 H2a

Reddit -0.713 -0.599 H2b (cons)

Twitter -0.235 None

Search 0.184 0.269 0.076 -0.261 -0.416

Email -0.064 -0.144 0.423 0.247

Darker shading represents increasing absolute magnitude of coefficients (coefficients with p > 0.05 omitted).

Blue cells with white text are negative estimated coefficients (more liberal), red cells with black text are positive (more conservative).

had mixed associations with diversity with low magnitudes.

Use of Facebook was associated with partisan shifts in slant,

notably larger for conservatives than for liberals. Conversely,

use of Reddit was associated with moderating shifts in slant

for conservatives.

The R

values for partisan shift and cross-cutting models (all

less than 1%) were much lower than those for diversity

(ranging from 4% to 34%). Thus, while social media

platforms are associated with changes in each of these

measures, there is much less unexplained variation in diversity

after accounting for platform use. This further underscores

the need to isolate and separately study the underlying

dimensions of information-limiting environments.

To demonstrate the magnitude of effects, we calculated

example estimated effects by platform for a concurrent

increase in both referrals and time. These measures of use

were jointly estimated, and it is intuitive to imagine that

increased referral activity from a platform corresponded with

increased time on the platform, and vice versa. Figure 4

demonstrates the marginal change in information diversity and

slant associated with a user who clicked on an additional six

referrals to news sites from a platform and had a four-fold

increase in time spent on the platform (e.g., an increase from

1 hour to 4 hours). Our measures are calculated per

individual by month, so this may be interpreted as a com-

parison of a month when a user had significant use of a

platform to a month when that user had limited use of that

platform. In a period when an average user has this type of

increase on Facebook (6 more referrals and 4times more time

spent), their standard deviation of slant would be 37% higher,

moving from 6.5 to 9.0. If conservative, their mean slant

would have increased by 2.7; if liberal, it would have

decreased by 0.7.

To illustrate these effects in combination, we provide an

example in Figure 5 for a representative conservative user

with monthly news site visits that have a mean slant of 10 and

standard deviation of 6.5. For this user, a ±2 standard devia-

tion range of online news consumption (e.g., -3 to 23) encom-

passes both USA Today (-0.4) and the New York Post (21.5).

In a month with 6 additional news referrals from and 4 times

more time spent on Facebook than average, their consumption

would shift to a more conservative mean, but also broaden.

Their estimated range of ±2 standard deviations (-4.6 to 31.2)

would increase by encompassing some additional news

sources to the left, but even more to the right (almost

including Fox News). If not for the rightward, partisan shift

in mean, the increase in diversity alone would have resulted

in a range of -7.9 to 27.9 (thereby including CNN to the left).

A similar example is also shown in Figure 5 for the combina-

1638 MIS Quarterly Vol. 44 No. 4/December 2020

Kitchens et al./Echo Chambers and Filter Bubbles

Figure 4. Social Media Platform Use, Diversification, and Partisan Shift

Figure 5. Effects of Increased Platform Use for a Representative Conservative User

MIS Quarterly Vol. 44 No. 4/December 2020 1639

Kitchens et al./Echo Chambers and Filter Bubbles

tion of broadening diversity with a moderating shift associated

with an increased use of Reddit. These stylized examples

demonstrate that associations with diversity and slant differ in

important ways by platform.

Implications for Conceptualizing Echo

Chambers and Filter Bubbles

We identify five contributions that this study makes to the

conceptualization of echo chambers and filter bubbles. First,

we adopt a generative framing of information-limiting envi-

ronments as reflecting concerns raised in discussion of these

captivating, yet elusive, phenomena. By considering more

nuanced ways that social media usage may shape information

consumption, we highlight both the hopes and fears for online

technology. On one hand, there is a widespread fear that

homophily and algorithmic filtering will amplify personal

preference for congenial, ideologically congruent content and

foster information-limiting environments that exclude

opinion-challenging information (Pariser 2011; Sunstein

2017). On the other, there is the hopeful potential for social

media and other information discovery platforms to expand

individuals’ information consumption by connecting them

with a broad range of relevant information sources. To ap-

propriately address this tension, we move beyond a limiting,

muddled question of whether information-limiting environ-

ments exist to a more generative framing of how platforms

influence information source consumption.

Second, by examining in detail the constituent elements of

diversity and slant of information consumption we provide a

deeper understanding of the phenomenon. There is a long-

standing interest in understanding preferences for opinion-

challenging versus opinion-confirming information (see Hart

et al. 2009). This interest has often informed the under-

standing of filter bubbles and echo chambers as being

manifest in a binary outcome: the presence or absence of

cross-cutting content consumption (Bakshy et al. 2015; Mutz

and Martin 2001). The joint consideration of diversity and

slant provides a richer depiction of the variety of ways that

platforms may help shape exposure, engagement, and con-

sumption of information sources. Indeed, empirically we find

a range of outcomes across both dimensions.

Third, we demonstrate the value of conceptualizing the

relationship between platform use and information source

consumption at an individual-level, using panel data to

identify within-user effects. Prior research has often focused

on ideological segregation at a population-level by comparing

the distribution of information source consumption among

categories of users (Gentzkow and Shapiro 2011). For

example, conservative and liberal platform users prefer dif-

ferent online news sources (Flaxman et al. 2016; Jacobson et

al. 2016). However, a cross-sectional observation of aggre-

gated user preferences is insufficient to fully address how

platform usage shapes consumption. Relationships observed

at a group level do not necessarily hold for individuals within

the same groups (Freedman 1999); thus, individual-level

effects are best understood by analyzing individual-level data.

Our inquiry complements studies that focused on individual-

level outcomes associated with usage of Facebook (Bakshy et

al. 2015) and Twitter (Shore et al. 2018). We extend this

prior work further by concurrently analyzing individual-level

usage across multiple social media platforms. By considering

variation in platform use and information source consumption

for the same individual, we provide a more nuanced expli-

cation of the differential effects of various platforms.

Fourth, incorporating a multifaceted view of diversity

(Harrison and Klein 2007) provides additional insights into

information source consumption. Social media platforms are

important pathways for individuals to discover new sources of

information, and ideological differences are just one form of

differentiation among information sources. In defining infor-

mation source diversity in terms of separation, variety, and

disparity among information sources, we extend prior work

that focused on dominant ideological stance of an information

source as a source of information diversity (Shore et al. 2018).

The consistency of our results suggests there are additional

ways to reliably conceptualize and measure information

source diversity beyond political ideology, which may be

particularly useful for studying contexts where dominant

ideology of an information source is not evident or relevant.

Finally, our conceptual model illustrates the importance of

considering how the underlying characteristics of platforms

may each uniquely shape information consumption. Our

differentiation of usage by platform expands on prior work

that considered multiple platforms together, such as the com-

bination of email plus multiple social media platforms as a

single technology category by Flaxman et al. (2016). We find

wide variation in outcomes among the platforms that Flaxman

et al. aggregated together, thereby demonstrating the value of

considering them individually. By analyzing patterns of

results across different platforms, we can make inferences

about the impacts of individual features and mechanisms.

Next, we present possibilities following this line of thought.

Implications of Mechanisms for

Determining Information Sources

The variation of results by platform provides some suggestive

evidence about the most salient factors shaping information

source consumption. In comparing observed outcomes by

1640 MIS Quarterly Vol. 44 No. 4/December 2020

Kitchens et al./Echo Chambers and Filter Bubbles

Compared to

Direct Visits

Moderating

Shift

Change

Partisan

Shift

Narrowing

Diversity

No Change in

Diversity

Twitter

Broadening

Diversity

Search Reddit

Facebook

Figure 6. Platform Use, Diversification, and Partisan Shift

platform (summarized in Figure 6) with how these platforms

prioritize the content provided to users (identified in Figure 2

and Table 2), we draw four broad inferences on the relative

importance of social network homophily and various

approaches to algorithmic filtering.

First, our findings provide insight about the potential role of

social influence in partisan shifts. On Facebook, as well as

the reference platform of email, the information sources a user

is exposed to are strongly influenced by other individuals with

whom the user has a strong sense of social identification and

affiliation (e.g., friends, family, and coworkers). In contrast,

social affiliation is somewhat less salient on Twitter (where

users form nonreciprocal relationships) and Reddit (where

content is largely organized in topic-based communities).

From this lens, a partisan shift is consistent with concerns that

platforms may reinforce social influence within heterogenous

social groups, and, thereby cause a hardening of partisan

beliefs (Sunstein 2002, 2009). In addition, Facebook and

email also both utilize some form of engagement-based

curation—Facebook’s curated news feed and email’s prior-

itized inbox—that could also contribute to a partisan shift

(Pariser 2011).

Second, for Reddit, as well as the reference platform of

search, algorithmic filtering is more heavily influenced by an

individual’s topics of interest than by their social network or

prior engagement with content. A broadening effect here is

consistent with prior work arguing that, in the course of inter-

actions on other topics, individuals experience an incidental

exposure to additional sources of news and information

(Fletcher and Nielsen 2018a; Weeks et al. 2017).

Third, our results highlight trade-offs inherent in the study of

information source consumption. By isolating within-person

variation over time in both platform use and news site con-

sumption, we are able to quantify variation among multiple

social media platforms and draw inferences about the impact

of platform characteristics in relation to consumption

behaviors. Yet, our ability to draw inferences about indi-

vidual characteristics and attitudes is limited. For example,

we are unable to assess the impact of selective exposure due

to individual preference for opinion-confirming content

(Festinger 1964; Hart et al. 2009). Further, the relatively

small amount of variation in partisan shifts explained by

platform usage in our analyses suggests there are significant

unobserved factors that shape consumption. Because no

single study can capture all of the factors that influence con-

sumption decisions, particularly when considering multiple

platform use by heterogeneous individuals, it is important to

further develop the conceptualization of information-limiting

environments in ways that can drive future empirical research,

as well as help integrate and reconcile findings.

Finally, the variation in our results across platforms implies

that rather than making broad-brush assertions about how they

may be information-limiting or information-expanding, more

focused attention should be paid to the range of individual

factors that determine how and to what extent platforms may

shape users’ choices. Further, our findings represent average

effects for a typical platform user, not a universal outcome for

all users. Platforms themselves are not monolithic; there may

be wide variation in how people use each platform as well as

changes in platform implementations over time. For example,

the impact of using interest-oriented features such as Face-

book pages (Jacobson et al. 2016; Skjerdal and Gebru 2020)

may differ from using only the primary Facebook news feed.

Thus, future research should take into account not only dif-

ferences across platforms but also differences among users in

how they use those platforms.

As noted, Twitter only implemented engagement-based feed curation in

2016, after our data collection.

MIS Quarterly Vol. 44 No. 4/December 2020 1641

Kitchens et al./Echo Chambers and Filter Bubbles

Practical Implications

Our findings have implications for a practical understanding

of information consumption. We find no support for a

prevalence of the popular, intuitive understanding of echo

chambers and filter bubbles. Specifically, we find no evi-

dence that the use of social media is limiting the information

sources that users choose to consume—instead, with respect

to information diversity we find a range of outcomes from

effectively neutral (no change) to positive (broadening

diversity). Nonetheless, we do find important variation

between platforms whose use was associated with moderating

shifts (Reddit) or partisan shifts (Facebook).

Of the platforms we consider, Facebook represents the most

complete embodiment of features that may impact users’ news

consumption choices. It hosts user-generated content, sug-

gests additions to users’ social networks, and provides several

content engagement options that influence the algorithmic

curation of information. Given the high level of personali-

zation that Facebook deploys in determining what content to

expose users to, we reiterate calls for further study of

Facebook’s algorithmic filtering (Sunstein 2017; Vaidhyana-

than 2018). Likewise, more research is needed to understand

the impacts of Twitter’s recent implementation of more

comprehensive algorithmic filtering. Yet, the study of these

features is stymied by a lack of transparency of platform

operations. As platform providers are unlikely to unilaterally

provide outside researchers with sufficient data on the imple-

mentation and effects of algorithmic filtering, there is an

important role for public policy oversight of these influential

platforms.

Limitations and Future Research

We acknowledge that our study has multiple limitations

related to the scope of the conceptual model, the scope of the

empirical test, and the research method. First, our conceptual

model focuses on individual and platform characteristics that

explain variation between platforms. There are additional

characteristics, including level of interest in politics (Dubois

and Blank 2018; Lawrence et al. 2010), reaction to content

topics (Baldassarri and Bearman 2007), and intensity of

platform use (Shore et al. 2018), that may help explain varia-

tion among individual users of the same platform. Also, our

scope is limited with regard to platform characteristics. Tech-

nology providers frequently update algorithms to determine

the most engaging and relevant information sources and

content to present to users; more research is needed to

encompass the full range of characteristics that influence this

presentation (Johnson et al. 2019). Likewise, our hypotheses

are focused on information source consumption. More

research is also needed to understand the antecedents and

consequences of exposure, engagement, and consumption.

For example, there is considerable research on opinion

formation that could be expanded upon to examine how usage

of social media relates to changes in opinions.

Second, the scope of empirical tests of our model is con-

strained by available data. The majority of the pathways in

our conceptual model are only observable at scale by platform

providers. For example, we could not observe the information

sources an individual was exposed to or engaged with on a

platform, nor other characteristics of those information

sources and their presentation. More research is needed to

better understand how exposure and engagement relate to the

information sources an individual chooses to consume.

Likewise, we were unable to observe how individuals varied

in their social networks and topic interests. Although the

within-person research design strengthens the robustness of

our results, these individual characteristics may change over

time and may even be shaped by platform usage. For these

reasons and others, it is impossible to establish causality

within our research design. The associations we identify are

informative of the variation among platforms and illustrative

of the importance of nuance in measurement. But more

research is required to establish causal effects and establish

the mechanisms that influence these outcomes.

Third, there are numerous limitations related to our data

source and measurements. Our sample only included adults

residing in the United States and may not be representative of

other populations. Another data limitation is that we only

observed desktop use of information discovery platforms and

online news sites. As discussed, we believe this still repre-

sents an important portion of user activity, but particularly as

mobile usage continues to replaces desktop use, additional

research is needed regarding how behavior on mobile devices

may differ. Another key operationalization in our study

involved attributing a slant score to all content on a single

news site domain. While we established that the news domain

provides a meaningful signal to the user, this is ultimately a

simplification of the potential information value of that con-

tent. Further, our measures of ideological diversity and news

source slant are most relevant in high media choice environ-

ments with a dominant liberal, centrist, and conservative

ideological configuration of public opinion; they may not

generalize to other settings.

Conclusion

The question of how social media platforms shape the con-

sumption of information and may foster the creation of

1642 MIS Quarterly Vol. 44 No. 4/December 2020

Kitchens et al./Echo Chambers and Filter Bubbles

information-limiting environments is crucial to society. Our

study provides a deeper conceptualization of this phenom-

enon, supported by insights obtained through observation of

user behaviors across various platforms in a natural environ-

ment, interacting with their naturally occurring social network,

without artificial intervention by researchers. As implied by

the complex variety in social media platforms, we found

considerable variation in observed impacts among them,

which helps explain prior conflicting conclusions, but also

leads to new unanswered questions. Based on our concep-

tualization and findings, we see a strong need for future

research to focus on the nuance of how the details of platform

implementations and individual choices combine to influence

a variety of outcomes that underlie the concerns that have

been raised in the discussion of echo chambers and filter

bubbles.

Acknowledgments

The authors would like to thank the McIntire School of Commerce

Foundation and the McIntire School of Commerce’s Center for

Business Analytics for their support as well as participants of the

Virginia Research Seminar Series and the UMD Center for the

Advanced Study of Communities and Information for their valuable

feedback. The authors would also like to thank the senior editor,

associate editor, and reviewers for their extremely helpful comments

and suggestions that greatly improved the quality of this work.

References

Adamic, L. A., and Glance, N. 2005. “The Political Blogosphere

and the 2004 U.S. Election: Divided They Blog,” in Proceedings

of the 3

International Workshop on Link Discovery, New York:

ACM, pp. 36-43.

Allcott, H., and Gentzkow, M. 2017. “Social Media and Fake News

in the 2016 Election,” Journal of Economic Perspectives (31:2),

pp. 211-236.

Bail, C. A., Argyle, L. P., Brown, T. W., Bumpus, J. P., Chen, H.,

Hunzaker, M. F., Lee, J., Mann, M., Merhout, F., and

Volfovsky, A. 2018. “Exposure to Opposing Views on Social

Media Can Increase Political Polarization,” Proceedings of the

National Academy of Sciences (115:37), pp. 9216-9221.

Bakos, J. Y. 1997. “Reducing Buyer Search Costs: Implications

for Electronic Marketplaces,” Management Science (43:12), pp.

1676-1692.

Bakshy, E., Messing, S., and Adamic, L. A. 2015. “Exposure to

Ideologically Diverse News and Opinion on Facebook,” Science

(348:6239), pp. 1130-1132.

Baldassarri, D., and Bearman, P. 2007. “Dynamics of Political

Polarization,” American Sociological Review (72:5), pp.

784-811.

Barberá, P. 2015. “Birds of the Same Feather Tweet Together:

Bayesian Ideal Point Estimation Using Twitter Data,” Political

Analysis (23:1), pp. 76-91.

Boxell, L., Gentzkow, M., and Shapiro, J. M. 2017. “Greater

Internet Use Is Not Associated with Faster Growth in Political

Polarization among US Demographic Groups,” Proceedings of

the National Academy of Sciences (114:40), pp. 10612-10617.

Bozdag, E., and van den Hoven, J. 2015. “Breaking the Filter

Bubble: Democracy and Design,” Ethics and Information

Technology (17:4), pp. 249-265.

Bright, J. 2018. “Explaining the Emergence of Political Frag-

mentation on Social Media: The Role of Ideology and

Extremism,” Journal of Computer-Mediated Communication

(23:1), pp. 17-33.

Burt, R. S. 1992. Structural Holes: The Social Structure of

Competition, Boston: Harvard University Press.

Burton-Jones, A., Recker, J., Indulska, M., Green, P. F., and

Weber, R. 2017. “Assessing Representation Theory with a

Framework for Pursuing Success and Failure,” MIS Quarterly

(41:4), pp. 1307-1333.

Cooper, M. 2003. Media Ownership and Democracy in the Digital

Information Age, Center for Internet & Society, Stanford Law

School.

Dubois, E., and Blank, G. 2018. “The Echo Chamber Is Over-

stated: The Moderating Effect of Political Interest and Diverse

Media,” Information, Communication & Society (21:5), pp.

729-745.

Festinger, L. 1964. Conflict, Decision, and Dissonance (Vol. 3),

Stanford, CA: Stanford University Press.

Flaxman, S., Goel, S., and Rao, J. M. 2016. “Filter Bubbles, Echo

Chambers, and Online News Consumption,” Public Opinion

Quarterly (80:S1), pp. 298-320.

Fletcher, R., and Nielsen, R. K. 2016. “Is Social Media Use

Associated with More or Less Diverse News Use?,” November

25 (https://rasmuskleisnielsen.net/2016/11/25/is-social-media-

use-associated-with-more-or-less-diverse-news-use/; accessed

May 27, 2018).

Fletcher, R., and Nielsen, R. K. 2018a. “Are People Incidentally

Exposed to News on Social Media? A Comparative Analysis,”

New Media & Society (20:7), pp. 2450-2468.

Fletcher, R., and Nielsen, R. K. 2018b. “Automated Serendipity,”

Digital Journalism (6:8), pp. 976-989.

Fortunato, S., Flammini, A., Menczer, F., and Vespignani, A. 2006.

“Topical Interests and the Mitigation of Search Engine Bias,”

Proceedings of the National Academy of Sciences (103:34), pp.

12684-12689.

Freedman, D. A. 1999. “Ecological Inference and the Ecological

Fallacy,” International Encyclopedia of the Social & Behavioral

Sciences (6), pp. 4027-4030.

Frey, D. 1986. “Recent Research on Selective Exposure to Infor-

mation,” in Advances in Experimental Social Psychology (Vol.

19), L. Berkowitz (ed.), Cambridge, MA: Academic Press, pp.

41-80.

Gabszewicz, J. J., Laussel, D., and Sonnac, N. 2002. “Press Adver-

tising and the Political Differentiation of Newspapers,” Journal

of Public Economic Theory (4:3), pp. 317-334.

Garrett, R. K. 2009. “Echo Chambers Online? Politically Moti-

vated Selective Exposure among Internet News Users,” Journal

of Computer-Mediated Communication (14:2), pp. 265-285.

MIS Quarterly Vol. 44 No. 4/December 2020 1643

Kitchens et al./Echo Chambers and Filter Bubbles

Gentzkow, M., and Shapiro, J. M. 2011. “Ideological Segregation

Online and Offline,” The Quarterly Journal of Economics

(126:4), pp. 1799-1839.

Goel, S., Mason, W., and Watts, D. J. 2010. “Real and Perceived

Attitude Agreement in Social Networks,” Journal of Personality

and Social Psychology (99:4), pp. 611-621.

Gray, P. H., and Cooper, W. H. 2010. “Pursuing Failure,” Organi-

zational Research Methods (13:4), pp. 620-643.

Harrison, D. A., and Klein, K. J. 2007. “What’s the Difference?

Diversity Constructs as Separation, Variety, or Disparity in

Organizations,” Academy of Management Review (32:4), pp.

1199-1228.

Hart, W., Albarracín, D., Eagly, A. H., Brechan, I., Lindberg, M. J.,

and Merrill, L. 2009. “Feeling Validated Versus Being Correct:

A Meta-Analysis of Selective Exposure to Information,”

Psychological Bulletin (135:4), pp. 555-588.

Herlocker, J. L., Konstan, J. A., Terveen, L. G., and Riedl, J. T.

2004. “Evaluating Collaborative Filtering Recommender

Systems,” ACM Transactions on Information Systems (22:1), pp.

5-53.

Himelboim, I., McCreery, S., and Smith, M. 2013. “Birds of a

Feather Tweet Together: Integrating Network and Content

Analyses to Examine Cross-Ideology Exposure on Twitter,”

Journal of Computer-Mediated Communication (18:2), pp.

154-174.

Iyengar, S., Lelkes, Y., Levendusky, M., Malhotra, N., and

Westwood, S. J. 2019. “The Origins and Consequences of

Affective Polarization in the United States,” Annual Review of

Political Science (22), pp. 129-146.

Jacobson, S., Myung, E., and Johnson, S. L. 2016. “Open Media or

Echo Chamber: The Use of Links in Audience Discussions on

the Facebook Pages of Partisan News Organizations,” Infor-

mation, Communication & Society (19:7), pp. 875-891.

Janis, I. L. 1982. Groupthink: Psychological Studies of Policy

Decisions and Fiascoes (Vol. 349), Boston: Houghton Mifflin

Company.

Jehn, K. A., Northcraft, G. B., and Neale, M. A. 1999. “Why

Differences Make a Difference: A Field Study of Diversity,

Conflict and Performance in Workgroups,” Administrative

Science Quarterly (44:4), pp. 741-763.

Johnson, S. L., Gray, P., and Sarker, S. 2019. “Revisiting IS

Research Practice in the Era of Big Data,” Information and

Organization (29:1), pp. 41-56.

Jürgens, P., and Stark, B. 2017. “The Power of Default on Reddit:

A General Model to Measure the Influence of Information

Intermediaries: The Influence of Information Intermediaries,”

Policy & Internet (9:4), pp. 395-419.

Jurkowitz, M., Mitchell, A., Shearer, E., and Walker, M. 2020.

“U.S. Media Polarization and the 2020 Election: A Nation

Divided,” , January 24 (https://www.journalism.org/2020/01/24/

u-s-media-polarization-and-the-2020-election-a-nation-divided/).

Kaminskas, M., and Bridge, D. 2016. “Diversity, Serendipity,

Novelty, and Coverage: A Survey and Empirical Analysis of

Beyond-Accuracy Objectives in Recommender Systems,” ACM

Transactions on Interactive Intelligent Systems (7:1), pp. 1-42.

Klapper, J. T. 1960. The Effects of Mass Communication, Glencoe,

IL: Free Press.

Lawrence, E., Sides, J., and Farrell, H. 2010. “Self-Segregation or

Deliberation? Blog Readership, Participation, and Polarization

in American Politics,” Perspectives on Politics (8:1), pp.

141-157.

Lazer, D. 2015. “The Rise of the Social Algorithm,” Science

(348:6239), pp. 1090-1091.

Lazer, D. M. J., Baum, M. A., Benkler, Y., Berinsky, A. J.,

Greenhill, K. M., Menczer, F., Metzger, M. J., Nyhan, B.,

Pennycook, G., Rothschild, D., Schudson, M., Sloman, S. A.,

Sunstein, C. R., Thorson, E. A., Watts, D. J., and Zittrain, J. L.

2018. “The Science of Fake News,” Science (359:6380), pp.

1094-1096.

Li, K. G., Mithas, S., Zhang, Z., and Tam, K. Y. 2019. “How Does

Algorithmic Filtering Influence Attention Inequality on Social

Media?,” in Proceedings of the 40

International Conference on

Information Systems, Munich.

Lin, M., Lucas Jr., H. C., and Shmueli, G. 2013. “Research

Commentary—Too Big to Fail: Large Samples and the p-Value

Problem,” Information Systems Research (24:4), pp. 906-917.

McDonald, D., and Dimmick, J. 2003. “The Conceptualization and

Measurement of Diversity,” Communication Research (30:1), pp.

60-79.

Messing, S., and Westwood, S. J. 2014. “Selective Exposure in the

Age of Social Media: Endorsements Trump Partisan Source

Affiliation When Selecting News Online,” Communication

Research (41:8), pp. 1042-1063.

Mitchell, A., and Jurkowitz, D. 2014. “Social, Search & Direct:

Pathways to Digital News,” Pew Research Center (https://www.

journalism.org/2014/03/13/social-search-direct/).

Mutz, D. C. 2006. “How the Mass Media Divide Us,” in Red and

Blue Nation? Characteristics and Causes of America’s

Polarized Politics (Vol. 1), P. S. Nivola and D. W. Brady (eds.),

Washington, DC: Brookings Institution Press.

Mutz, D. C., and Martin, P. S. 2001. “Facilitating Communication

across Lines of Political Difference: The Role of Mass Media,”

American Political Science Review (95:1), pp. 97-114.

Negroponte, N. 1996. Being Digital, New York: Vintage Books.

Nickerson, R. S. 1998. “Confirmation Bias: A Ubiquitous

Phenomenon in Many Guises,” Review of General Psychology

(2:2), pp. 175-220.

Page, S. E. 2010. Diversity and Complexity (Vol. 2), Princeton, NJ:

Princeton University Press.

Pariser, E. 2011. The Filter Bubble: What the Internet Is Hiding

from You, New York: Penguin Press.

Prawesh, S., and Padmanabhan, B. 2011. “The ‘Top N’ News

Recommender: Count Distortion and Manipulation Resistance,”

in Proceedings of the 5

ACM Conference on Recommender

Systems, New York: ACM, pp. 237-244.

Schkade, D., Sunstein, C. R., and Hastie, R. 2007. “What Hap-

pened on Deliberation Day,” California Law Review (95), pp.

915-940.

Schmidt, A. L., Zollo, F., Vicario, M. D., Bessi, A., Scala, A.,

Caldarelli, G., Stanley, H. E., and Quattrociocchi, W. 2017.

“Anatomy of News Consumption on Facebook,” Proceedings of

the National Academy of Sciences (114:12), pp. 3035-3039.

Shore, J., Baek, J., and Dellarocas, C. 2018. “Network Structure

and Patterns of Information Diversity on Twitter,” MIS Quarterly

(42:3), pp. 849-872.

1644 MIS Quarterly Vol. 44 No. 4/December 2020

Kitchens et al./Echo Chambers and Filter Bubbles

Skjerdal, T., and Gebru, S. 2020. “Not Quite an Echo Chamber:

Ethnic Debate on Ethiopian Facebook Pages during Times of

Unrest,” Media, Culture & Society (42:3), pp. 365-379.

Sohn, M.-W. 2001. “Distance and Cosine Measures of Niche

Overlap,” Social Networks (23:2), pp. 141-165.

Stroud, N. J. 2010. “Polarization and Partisan Selective Exposure,”

Journal of Communication (60:3), pp. 556-576.

Sunstein, C. R. 2001. Republic.com, Princeton, NJ: Princeton

University Press.

Sunstein, C. R. 2002. “The Law of Group Polarization,” Journal

of Political Philosophy (10:2), pp. 175-195.

Sunstein, C. R. 2009. Going to Extremes: How Like Minds Unite

and Divide, New York: Oxford University Press.

Sunstein, C. R. 2017. # Republic: Divided Democracy in the Age

of Social Media, Princeton, NJ: Princeton University Press.

Vaidhyanathan, S. 2018. Antisocial Media: How Facebook Dis-

connects Us and Undermines Democracy, New York: Oxford

University Press.

Van Alstyne, M., and Brynjolfsson, E. 1996. “Could the Internet

Balkanize Science?,” Science (274:5292), pp. 1479-1480.

Van Alstyne, M., and Brynjolfsson, E. 2005. “Global Village or

Cyber-Balkans? Modeling and Measuring the Integration of

Electronic Communities,” Management Science (51:6), pp.

851-868.

Walker, M. 2019. “Americans Favor Mobile Devices over Desk-

tops and Laptops for Getting News,” Pew Research Center,

November 19 (https://www.pewresearch.org/fact-tank/2019/

11/19/americans-favor-mobile-devices-over-desktops-and-

laptops-for-getting-news/).

Weeks, B. E., Lane, D. S., Kim, D. H., Lee, S. S., and Kwak, N.

2017. “Incidental Exposure, Selective Exposure, and Political

Information Sharing: Integrating Online Exposure Patterns and

Expression on Social Media,” Journal of Computer-Mediated

Communication (22:6), pp. 363-379.

Welch, M. J., Cho, J., and Olston, C. 2011. “Search Result Diver-

sity for Informational Queries,” in Proceedings of the 20

International Conference on World Wide Web, pp. 237-246.

Wojcieszak, M. E., and Mutz, D. C. 2009. “Online Groups and

Political Discourse: Do Online Discussion Spaces Facilitate

Exposure to Political Disagreement?,” Journal of Communication

(59:1), pp. 40-56.

About the Authors

Brent Kitchens is an assistant professor at the University of

Virginia McIntire School of Commerce. He received his Ph.D. from

the University of Florida Warrington College of Business

Administration. His research interests include analytics of online

interactions, with a focus on applications for societal good and

public health. His research has been published at a variety of

outlets, including MIS Quarterly, Information Systems Research,

Journal of Management Information Systems, and IEEE Trans-

actions on Knowledge and Data Engineering.

Steven L. Johnson is an associate professor at the McIntire School

of Commerce, University of Virginia. He received his Ph.D. from

the Robert H. Smith School of Business, University of Maryland.

His research interests include online communities and other social

media that support open innovation with an emphasis on network

analysis. His research has appeared in outlets such as MIS

Quarterly, Organization Science, and Information Systems

Research.

Peter Gray is a professor at the University of Virginia’s McIntire

School of Commerce, where he teaches courses that connect ideas

about innovation, information technology, and strategic manage-

ment. His research focuses on the collaborative impacts of social

media, organizational networks, virtual teams, and online com-

munities. His research has been published in a range of leading

journals, including MIS Quarterly, Information Systems Research,

Management Science, Sloan Management Review, Organizational

Research Methods, California Management Review, the Journal of

Management Information Systems, Information & Organization, the

Journal of Strategic Information Systems, Information Technology

& People, and others.

MIS Quarterly Vol. 44 No. 4/December 2020 1645

Kitchens et al./Echo Chambers and Filter Bubbles

Appendix

Shore et al. (2018) identified the slant of 186 news URLs, which we used as the basis for multiple measures in this paper. Unless otherwise

noted, we used the scores they identify as is. However, because their unit of analysis was a URL associated with a Twitter account, there were

certain necessary exceptions or deviations from their list as described in Table A1. First, a small number of the URLs they identified contained

both a domain and a path or page (e.g., buzzfeed.com/news). In order to cleanly identify direct versus other visits, for these sites we either

removed the path and used only the domain, or removed the entry from the list if the domain was already included on the list with a different

slant score. Another small subset of sites either functioned as more than just a news site (aol.com, reddit.com) or were simply a sub-site of a

portal or social media outlet (facebook.com/cnsnewscom, news.yahoo.com). Again, in order to avoid bias and conflation in our measures, these

sites were omitted. All URLs not listed in Table A1 were included as is.

Table A1. Changes in URL Inclusion Compared to Shore et al. (2018)

URL Issue/Resolution

cnn.com/politics

Includes path. This entry omitted. Separate no-path entry retained.

huffingtonpost.com/black-voices/

huffingtonpost.com/politics

rollingstone.com/politics

wsj.com/opinion

buzzfeed.com/news

Includes path. This entry used with path removed.

cbc.ca/news/

nature.com/news

pbs.org/newshour

pewresearch.org/fact-tank

pjmedia.com/instapundit/

membership.politifact.com Includes nonrelevant subdomain. Entry used without subdomain.

aol.com General site containing more than news. Removed.

facebook.com/cnsnewscom Facebook page, not news site. Removed.

finance.yahoo.com

news.msn.com Subdomain of search/portal site. Removed.

news.yahoo.com

reddit.com Social media site. Removed from news site list

1646 MIS Quarterly Vol. 44 No. 4/December 2020

Kitchens et al./Echo Chambers and Filter Bubbles

Table A2. Correlations (1,092,480 User-Periods over 185,548 Users)

Distinct News Sites

Slant Dispersion

Audience Variety

Reverse Gini

Slant

Cross-Cutting

Direct News Visits

Search Referrals

Email Referrals

Facebook Referrals

Reddit Referrals

Twitter Referrals

Time on Search

Time on Email

Time on Facebook

Time on Reddit

Time on Twitter

Slant Dispersion 0.49

Audience Variety 0.50 0.62

Reverse Gini 0.64 0.46 0.78

Slant 0.07 0.29 0.05 0.06

Cross-Cutting 0.04 0.16 0.07 0.01 -0.06

Direct News Visits 0.30 0.07 0.09 0.21 0.09 -0.02

Search Referrals 0.52 0.19 0.24 0.33 0.00 0.01 0.15

Email Referrals 0.17 0.06 0.07 0.14 0.02 0.00 0.11 0.12

Facebook Referrals 0.36 0.18 0.15 0.24 0.04 0.01 0.10 0.15 0.13

Reddit Referrals 0.21 0.05 0.06 0.07 -0.02 -0.01 0.01 0.04 0.00 0.04

Twitter Referrals 0.17 0.04 0.04 0.08 0.00 0.00 0.06 0.09 0.07 0.11 0.02

Time on Search 0.18 0.11 0.14 0.15 0.01 0.02 0.04 0.18 0.10 0.06 0.00 0.02

Time on Email 0.12 0.09 0.11 0.12 0.02 0.02 0.04 0.08 0.17 0.05 0.00 0.01 0.60

Time on Facebook 0.13 0.14 0.13 0.12 0.02 0.05 0.00 0.05 0.01 0.25 0.01 0.02 0.08 0.06

Time on Reddit 0.14 0.05 0.06 0.06 -0.02 -0.01 0.00 0.02 0.00 0.02 0.54 0.01 0.00 -0.01 0.01

Time on Twitter 0.06 0.02 0.04 0.04 -0.01 0.00 0.00 0.04 0.00 0.02 0.01 0.20 0.02 0.00 0.06 0.02

Time Online 0.26 0.17 0.22 0.26 0.01 0.04 0.10 0.16 0.09 0.13 0.04 0.04 0.44 0.39 0.35 0.05 0.09

MIS Quarterly Vol. 44 No. 4/December 2020 1647

Kitchens et al./Echo Chambers and Filter Bubbles

Table A3. Summary Statistics for Users on Panel for at Least 365 Days Across all 4-Week User-Periods

Including > 0 News Visits (1,092,480 User-Periods over 185,548 Users)

All Users

Liberal Tercile

(Users with all-time mean

slant < -10.45)

Centrist Tercile

(Users with all-time mean

slant b/w -10.45 and -3.16)

Conservative Tercile

(Users with all time mean

slant > -3.16)

Median Mean SD Median Mean SD Median Mean SD Median Mean SD

Distinct news sites 2 2.97 3.80 2 2.94 3.81 2 2.90 3.28 2 3.09 4.35

Diversity - Slant dispersion 1.02 6.53 10.46 0.79 5.07 7.93 1.30 5.55 8.75 0.98 9.24 13.66

Diversity - Audience variety 63.09 42.95 40.62 60.84 41.73 40.56 65.10 43.96 39.92 62.64 42.95 41.50

Diversity - Reverse Gini 6.30 19.66 23.35 3.48 19.59 23.65 10.00 19.78 22.56 3.67 19.59 23.98

Slant -7.58 -4.35 20.75 -16.03 -15.36 11.16 -7.58 -7.03 11.70 -0.44 10.12 27.59

Cross-cutting 0.00 8.59 23.78 0.00 3.87 14.32 N/A 0.00 23.97 35.76

Direct news visits 0 3.26 29.28 0 2.26 20.09 0 2.26 17.34 0 5.51 44.68

Search referrals 0 1.08 3.10 0 0.97 2.74 0 1.20 2.99 0 1.03 3.55

Email referrals 0 0.11 1.18 0 0.15 1.40 0 0.08 1.01 0 0.12 1.14

Facebook referrals 0 0.45 2.40 0 0.58 2.67 0 0.31 1.50 0 0.50 2.94

Reddit referrals 0 0.03 0.55 0 0.07 0.84 0 0.03 0.45 0 0.01 0.20

Twitter referrals 0 0.02 0.33 0 0.02 0.38 0 0.01 0.23 0 0.01 0.38

Time on search 1.56 4.19 10.36 1.36 3.61 9.48 1.81 4.59 10.87 1.49 4.29 10.56

Time on email 0.40 2.61 7.89 0.38 2.39 7.31 0.45 2.68 8.09 0.36 2.74 8.19

Time on Facebook 0.57 4.74 10.78 0.66 4.60 10.22 0.52 4.52 10.41 0.54 5.15 11.74

Time on Reddit 0.00 0.09 1.25 0.00 0.19 1.91 0.00 0.07 0.97 0.00 0.02 0.44

Time on Twitter 0.00 0.13 1.39 0.00 0.14 1.22 0.00 0.16 1.53 0.00 0.09 1.38

Time online 28.91 46.44 57.30 27.18 44.18 54.38 30.80 48.15 57.66 28.38 46.62 59.60

Measures per user

Age 33 36.94 15.35 32 35.69 15.08 33 36.16 14.84 37 38.98 15.91

Gender 51.55% male 47.39% male 53.60% male 53.65% male

1648 MIS Quarterly Vol. 44 No. 4/December 2020

Kitchens et al./Echo Chambers and Filter Bubbles

Figure A1. 95% CI Plots for Coefficients Estimated Using Random Subsamples

MIS Quarterly Vol. 44 No. 4/December 2020 1649

1650 MIS Quarterly Vol. 44 No. 4/December 2020