Depolarization Through Social Media Use (A Dynamic Panel Analysis for the
Netherlands)
The rise of online social media has changed how people are exposed to information and
news. Facebook, Instagram, Twitter, YouTube, and others provide users with instant and varied
sources of information, which is also often filtered by algorithmic systems. The selective nature
of information flows on online social media has attracted scholarly attention. On the one hand,
prominent scholars have argued that the use of online social media use would fuel insularity in
political communication. People would end up communicating within “echo chambers” and
“filter bubbles”, in which individuals are selectively exposed to content that reinforces
previously held beliefs (Sunstein, 2018; Pariser, 2011).1 On the other hand, scholars such as
Mutz (2006) and Benkler (2008) have argued that the lower costs of accessing information and
increased choice promoted by online social networks would actually lead to greater exposure
to diverse ideas. With the escalation of the ideological divide worldwide, polarization is often
attributed to the rise of social media use. In recent decades, political polarization has been at
the forefront of popular and scientific debate. While some level of political division is
important to incite public debate (Mouffe, 2011), increasing polarization can impede not only
compromise in the design and implementation of public policies but also the effective
functioning of democracies (Fishkin, 2009; Sunstein, 2001). Thus, there is an increased
concern that social media use would exacerbate political polarization.
Evidence, however, is still mixed. While some studies found that social media use can
reinforce polarization (Levy, 2021; Quattrociocchi et al., 2016), others have found no effect
and even that social media use would actually help to alleviate polarization (Bakshy et al.,
2015; Dubois & Blank, 2018; Barberá, 2014; Beam et al., 2018). As literature reviews on the
topic have highlighted (Tucker et al., 2018; Kubin & von Sikorski, 2021), explanations for such
disparate evidence are related to the fact that the relationship between social media and political
polarization is likely heterogeneous and complex to estimate. Observational studies struggle
with the fact that social media use measures are based on self-reports, leaving open the
possibility of endogeneity and biased estimates. Endogeneity may arise from unobserved
factors that may correlate with political polarization and SM use or reverse causation. One may
assume that polarized individuals are more likely to use SM to seek information, which would
create a causal arrow running from polarization to SM use and would bias estimates (Tucker et
al., 2018). Many have reverted to experiments due to endogeneity issues.
Adding to the complexity is that polarization is seen as both a state and a dynamic
process. Polarization is, as DiMaggio et al. (1996) put it, “both a state and a process”. One of
the difficulties in accounting for such a dynamic process is the lack of longitudinal data on
political attitudes. Although a great deal has been written about the dynamics of political
polarization2 most of the studies exploring the effects of social media use on political
polarization are based on static analyses.3 To our knowledge, the only empirical studies
exploring the causal relationship between political polarization and social media over time are
Lee et al. (2018); Nordbrandt (2021); Barberá (2014). The contribution of this paper lies in
combining the advantages of a rich panel data set with an advanced econometric identification
method to explore the effects of social media use on political polarization in the Netherlands.
We employ a system generalized method of moments (System-GMM) estimator applied to a
dynamic panel data model. The System-GMM is acknowledged as the most efficient method
to estimate dynamic panel models that suffer from endogeneity. The method has been widely
applied in areas that typically suffer from endogenous explanatory variables, for example, in
economic growth (Bayraktar-Sa ˘glam, 2016), finance (Levine et al., 2000), and education
(Castelló-Climent & Mukhopadhyay, 2013). It has also been applied to examine political
polarization and its relation to globalization (Fang et al., 2021), political instability (Alt &
Lassen, 2006) and energy consumption (Apergis & Pinar, 2021).
The system-GMM allows for the dynamic nature of political polarization and rigorously
addresses the likely endogeneity of the relationship between social media and polarization by
employing (internal) instrumental variable (IV) techniques. In addition, it controls for
unobserved, time-invariant, and individual-specific effects. Thus, the panel data analysis we
conduct has numerous benefits over the cross-sectional analyses used on social media and
political polarization (Iyengar et al., 2012). Findings suggest that - contrary to popular
assumptions - social media use attenuates rather than drives political polarization.
Our work relates to two strands of literature. The first is the empirical analysis of “echo
chambers” and “filter bubble” theories. Theories of why SM would fuel polarization are related
to the information flow and network configurations shaped by this new environment. The
second is the body of empirical literature that allows for the dynamic nature of political
polarization.
The rise of SM has led to concerns that this information environment populated by
personalized recommendation features and algorithmic filtering would reinforce political
polarization. Popular scholars have theorized this environment where information would flow
in the so-called “echo chambers” and “filter bubbles” would reinforce the preconceived beliefs
of users (Pariser, 2011; Sunstein, 2018).
The new generation of Internet filters looks at the things you seem to like – the actual
things you’ve done, or the things people like you like – and tries to extrapolate. They are
prediction engines, constantly creating and refining a theory of who you are and what you’ll
do and want next. Together, these engines create a unique universe of information for each of
us – what I’ve come to call a filter bubble – which fundamentally alters the way we encounter
ideas and information. (Pariser, 2011, p.9)
Thus, according to Pariser, while traditional media allows individuals with random
viewpoints, SM’s algorithmic filtering limits exposure to differing perspectives.4 While
offering a narrower view of the political debate, “bubble filters tend to dramatically amplify
confirmation bias - in a sense, they are designed to do just that” (Pariser, 2011, p. 88). From
the author’s perspective, an increase in political polarization would be thus a direct
consequence. Such argumentation is in line with Sustein’s echo chamber theory. According to
Sustein, a central factor behind polarization is the existence of a “limited argument pool” - one
that is skewed in a particular direction.
If your Twitter feed consists of people who think as you do, or if your Facebook friends
share your convictions, the argument pool will be sharply limited. Indeed, shifts should occur
with individuals not engaged in the discussion but instead consulting only ideas — on radio,
television, or the Internet — to which they are predisposed. Such consultations will tend to
entrench and reinforce preexisting positions — often resulting in extremism. (Sustein, 2017,
p.135)
Considering these arguments, an increasing number of studies have tested empirical
support for these theories. To date, evidence using SM data remains mixed. Examinations of
selective exposure have shown that Facebook’s algorithm may indeed limit exposure to
counter-attitudinal news (Levy, 2021). Indeed, Quattrociocchi et al. (2016) shows that the
spreading of information on Facebook tends to be confined to like-minded communities.
However, a number of studies found that SM users are actually frequently exposed to diverse
and cross-cutting viewpoints (Bakshy et al., 2015).
Barberá (2014), in particular, shows that SM users are embedded in diverse ideological
networks and that exposure to political diversity reduces political polarization. Analysis using
surveys is also mixed. While some scholars found that SM use indirectly contributed to
polarization Lee et al. (2018), more recent evidence did not find any evidence that SM use
affects polarization over time (Nordbrandt, 2021). Beam et al. (2018) and Johnson et al. (2017),
in particular, identified negative effects of social media use polarization. A recent review of
literature on the topic suggested that one of the explanations for such mixed results is the
heterogeneity of the countries, platforms, and political issues analyzed in such studies (Kubin
& von Sikorski, 2021). Our work also relates to a body of research that allows for the dynamic
nature of political polarization. Many aspects of political attitudes are theorized as persistent.5
However, not many empirical studies accounted for the impacts of political polarization over
time. Green & Yoon (2002) was the first empirical study to apply a dynamic panel model to
analyze political behavior at an individual level. They were interested in exploring the
persistence of party identification over time. In other words, they focused on testing if the
process of party identification had a memory, i.e., if past levels had reverberating effects on
current party identification at an individual level. This was modeled by including lags of the
dependent variable (party identification) as a regressor in the model.6 They found that the
coefficient of the lagged dependent variable was not statistically different from zero and thus
concluded that party identification was not persistent at the micro-level, in contrast to what was
proposed by previous theories
Green and Yoon (2002) represented a major advance over cross-sectional studies.
However, it is now known that the Anderson–Hsiao first-difference estimator used for the
dynamic panel model is biased and imprecise. Arellano (1989) shows that estimators that use
instruments in levels are preferred to Anderson–Hsiao estimators that use instruments in
differences. Thus, the Anderson–Hsiao estimator is inefficient compared with other more
recently developed estimators that employ a system of equation framework - namely the
System-GMM estimator7. In a replication exercise, Wawro (2002) has drawn attention to the
costs of using less efficient estimators for a dynamic panel data analysis. He found that Green
and Yoon (2002) were possibly wrong in dismissing that party identification had a memory
over time. Some of the model’s specifications of Green and Yoon (2002) were unreliable after
applying the System-GMM estimator with the appropriate specification tests.
It has now been demonstrated that the System-GMM is the most efficient method to
estimate dynamic panel data models. The method was applied recently to analyse political
polarization and its relation to globalization (Fang et al., 2021), political instability
(Elbahnasawy et al., 2016), renewable energy consumption (Apergis & Pinar, 2021) and
transparency (Alt & Lassen, 2006). Our study adds to the literature on the causal effects of SM
use on political polarization by applying a dynamic panel model. Most of the studies exploring
the effects of SM use in political polarization are based on static analyses8. To our knowledge,
the only empirical studies exploring the causal relationship between political polarization and
SM over time are Lee et al. (2018); Nordbrandt (2021); Barberá (2014). Our work pursues a
complementary aspect to these studies by applying a method that allows for the dynamic nature
of political polarization and rigorously addresses the likely endogeneity of the relationship
between SM and polarization by employing instrumental variable (IV) techniques and
individual-specific effects.
The System-GMM estimates of the effects of SM use on our main measure of
polarization - namely the ideological polarization - are reported in Table 4.1. We run separate
models for each measure of SM use. In the first column, we report estimates of the model
measuring the effect of using SM, in the second column, the model measuring the effect of the
intensity of SM use, and in the third column the model measuring the effect of the frequency
of SM use.10 In every regression, we use the two-step system-GMM method with Windmeijer
(2005) robust standard error, and Principal Component Analysis to control for the proliferation
of the instruments (Mehrhoff, 2009; Kapetanios & Marcellino, 2010). To check the validity of
the instruments and the model’s fit we analyze Hansen’s J-statistic specification tests, Arellano-
Bond test for first-order and second-order autocorrelation, and the Kaiser-Lawyer-Olkin
measuring of sample adequacy (KMO). The test of overidentification is based on the Hansen J
statistic. The Hansen test results do not reject the null hypothesis that the instruments are valid
for all the specifications used (Hansenp p > 0.05 for all estimates), indicating that instruments
are valid for all regressions. As for the Arellano-Bond test for autocorrelation, the results reject
the null hypothesis of the absence of first order autocorrelation (AR(2) p < 0.00 for all
estimates) and do not reject the null hypothesis of the absence of second order autocorrelation
(p > 0.05 for all estimates). The insignificance of all the second order autocorrelation test results
implies that there was no second-order serial correlation of the error term for all the regression
models. Finally, the Kaiser-Meyer-Olkin (KMO) measure of sample adequacy for PCA shows
values higher than 0.5. That is, we have confidence that the factor analysis used is adequately
adjusted to the data. In sum, the tests indicate good specification quality
Findings lend support to the argument that using SM attenuates ideological polarization.
In Table 4.1, reading and viewing SM has a significant and negative effect on polarization,
controlling for demographic, individual characteristics and prior polarization (dummy variable
b = - 0.06, p < 0.001). Furthermore, spending more hours per week reading and viewing SM
and using SM with a greater frequency is also associated with lower ideological polarization
(intensity variable b = - 0.04, p < 0.001 and frequency variable b = - 0.05, p < 0.001).
Returning to our research question — the effect of social media use on political
polarization - there are two competing theories. On one hand, various authors have argued that
social media usage would reinforce polarization, because they would facilitate communication
between like-minded individuals, the strengthening of communities with same ideologies and
the development of imbalanced flows of information. On the other, some voices have argued
that social networks could have the effect of exposing individuals to opposing views and to a
more pluralistic set of information, what could lead to the adoption of more centrist positions
and the alleviation of political polarization. This study lends support to the second strand of
research. Our findings suggest that social media use attenuates rather than drives polarization,
a result that holds for different measures of social media usage - dummy (yes vs. no), intensity
(time spent), and frequency. Reading and viewing social media has a significant and negative
effect on polarization. More hours spent reading and viewing social media per week, and
greater frequency of social media use are also associated with lower polarization. Our findings
seem to support Mutz’s argument of cross-cutting social networks (Mutz, 2006, 2002)
according to which social media is associated with weaker opinions and identities because it
would promote a greater understanding of other’s people perspectives. The research developed
in this paper has important limitations and several questions remain open for future research.
First, the research focused on the quantitative aspects of online communication and skipped
any discussion regarding the qualitative characteristics of messages disseminated through
social media. Careful qualitative analysis is important to illuminate the types of content that
circulate in social networks and assess the possibilities these technologies open to the
dissemination of extreme messages, hate speech and misinformation. Second, the study did not
take into account the heterogeneity across different social media platforms. Because each social
network has a specific method of disseminating information and reaches a particular audience,
it is important to investigate if some platforms might promote different effects on political
polarization and, if so, what are the reasons that cause these effects.
