The Impact of Twitter on Lawmakers’ Decisions
Reza Mousavi & Bin Gu
Arizona State University
{mousavi,bin.gu}@asu.edu
Abstract
Organizations have used social media extensively to engage customers, but little is
known if such engagement truly influences organizations’ decisions to make them closer to their
customers.
This paper studies this question in a unique context – the impact of U.S.
Representatives’ Twitter engagement on their voting behavior in The Congress. In particular,
we consider whether the adoption and frequent use of Twitter make Representatives to vote more
in line with the political orientation of their constituents. We constructed a panel data for 445
Members of the 111th U.S. House of Representatives across a period of 24 months. We exploit
the variation in joining Twitter across Representatives to identify the impact of joining Twitter
on voting behavior. Using a difference-in-difference approach, we found that the adoption of
Twitter makes Representatives to vote more in line with their constituents. That is, in districts
where Representatives are more conservative than the constituent, Twitter adoption results in
less conservative votes by Representatives. Similarly, in districts where Representatives are more
liberal than the constituent, Twitter adoption results in less liberal votes by the Representatives.
We also found evidence that Representatives who generate more contents on Twitter platform
tend to vote closer to the political orientation of their constituents.
Keywords: Online Social Networking, Societal Impact of IS, Decision-Making in
Politics, U. S. House of Representatives, Difference-in-Difference Model
1
1. Introduction
Online social networking (OSN) platforms are profoundly changing the way we
communicate, collaborate, and make decisions. The enormous impacts of these platforms on the
societies can be observed in numerous examples: microblogging platforms such as Twitter which
have been widely credited as a key enabler of Arab Spring, Spain and Portugal movements in
2011, and more recently Turkey and Brazil movements in 2013.
Moreover, OSN platforms facilitate the participations of consumers and the public in
business (Edvardsson et al. 2011; Mousavi and Demirkan 2013) and government decision
making processes (Bertot et al. 2010; Linders 2012). However, little is known about the degree
to which such participations truly affect decision outcomes. While organizational decisions are
often not observable to the public and difficult to quantify, the U.S. political system provides a
rare exception where the most important decisions taken by U.S. politicians – voting decisions –
are observable to the public and have been carefully analyzed and quantified by political
scientists. Additionally, the low cost of adoption and the wide reach of OSN platforms have
convinced almost every American politician to be active on these platforms. A 2012 study by
Greenberg revealed that nearly 98% percent of the U.S. Representatives are active users of
online social networks (Greenberg 2012). Moreover, the analysis of the contents of the posts by
Representatives revealed that the majority of them are position taking posts. Furthermore, online
social media not only help lawmakers to communicate their messages to the constituents, but
also provide the constituents with a channel to interact with their representatives in a convenient
and casual way. In addition, the fact that these conversations are publicly available to other
constituents creates a dynamic system that facilitates transparency and accountability. Hence, the
adoption of online social networks by politicians could facilitate the influence of constitutes on
2
the political process. Therefore, in this research, we use the U.S. Representatives’ voting
decisions to assess to what degree the public influences organizational decisions through OSN
platforms. Specifically, we examine how the adoption and the use of OSN platforms influence
politicians’ political orientations. In particular, we examine whether the adoption of online social
networking moves politicians’ voting orientations closer to the views of their constituents.
The organization of this paper is as follows. Section 2 reviews the literature of online
social networking and its impact on decision making. We present our data and variables in
Section 3. The focus is on our empirical model in section 4. In Section 5, we present descriptive
statistics along with the results of our analyses. We discuss our findings and conclude with
limitations and potential extension of this study in the subsequent sections.
2. Literature Review
The interest of IS researchers in the realm of online platforms during last few years
provided the IS literature with rigorous examples of research studies in social media domain. The
objective of these studies, in a broad sense, can be divided to two categories: studies that focus
on consumer activities in social media, and studies that focus on how social media influences
firm and organization decisions.
Yan and Tan (2010) studied the impact of involvement in an online healthcare social
community on patients’ mental health. Their results suggested that patients who are active on the
OSN platform gain benefits from learning from other users. Learning from other patients helps
the users in the self-management processes and therefore is associated with better health
outcomes. Wu (2013) found that information-rich networks can influence the decisions made by
3
users. Burt (1992) theorized that three distinct information benefits drive this effect: access,
timing, and referrals. An information-rich network would allow the users to access more
information in a timely manner. Furthermore, the users can obtain recommendations from trusted
acquaintances. These mechanisms may change the users’ decisions and, therefore, their
performance (Wu 2013). These findings are in line with similar studies in other disciplines.
Marketing literature, for instance, hosts extensive research studies about the influence of wordof-mouth in online social networking platforms on the consumer purchasing decisions. Through
a unique natural experimental setting resulting from information policy shifts at Amazon.com,
Chen et al. (2011) studied the impacts of online reviews on consumer choices and revealed that
word of mouth and observational learning may influence consumer behavior. In another study,
Goh et al. (2013) found that engagement in social media brand communities leads to a positive
increase in purchase expenditures. By analyzing the communications among the users, Goh and
colleagues concluded that social media contents could impact consumer purchase behavior
through embedded information and persuasion. The influence of social media activities on
consumer behavior has also been studied by Rishika et al. (2013). The results of this study
further supported the positive influence of customer participation in a firm's social media efforts
on the frequency of customer visits and, therefore, firm’s profitability. It is worth noting that,
although the majority of the studies in this area provide evidence for the intended outcomes of
online platforms, scholars have also found evidence for some unintended outcomes relevant to
these platforms. For instance, Chan and Ghose (2014) studied whether the entry of Craigslist, a
major online personals ad site designed to facilitate consumer to consumer transactions,
increases the prevalence of HIV in the United States. Their results indicate that the entry of
4
Craigslist is related to a 15.9 percent increase in HIV cases. Intended or unintended, the main
findings of these studies support the impact of online platforms in a variety of contexts.
The second stream of studies emphasizes on the influence of social media on firm and
organization decisions. Wu (2013) revealed that the adoption of a work-related social networking
tool may impact employees’ decisions and performance. By studying the changes in employees’
networks and performance before and after the introduction of a social networking tool, she
revealed that the adoption influenced the quality of the decisions and, therefore, employees’
performance. Luo et al. (2013) studied the relationship between social media and firm equity
value using vector autoregressive models. Their results suggested that social media-based
metrics (web blogs and consumer ratings) are significant leading indicators of firm equity value
and have a faster predictive value than conventional online media. These findings support the
transformative power of social media as noted in (Aral et al. 2013).
Overall, the majority of studies in this area have focused on the network attributes of
social media or the flow of information within the networks. A new user who joins a network has
the opportunity to seek new information within the network. Whether the user is a patient who
seeks relevant information to deal with the illness, or a buyer who seeks information about a
certain type of product, information-rich networks could help her in achieving her goals. For
politicians, an information-rich network is a network that allows them to seek information about
the citizens. After all, politicians are representing the citizens and need their votes to sustain. We
believe that OSNs can be regarded as information-rich networks for politicians since these
networks contain overwhelming information about the citizens, their behaviors, and their
preferences.
5
3. Data
To study the impact of the adoption of OSN on the voting behavior of politicians, we
constructed a panel data for 445 Members of the 111th U.S. House of Representatives across a
period of 24 months using 3 disparate datasets. Summary statistics of the dataset are represented
in Table 1.
Spanning the 111th Congress, we estimate the monthly measure of Representatives’
political orientation based on votes cast by each Representative in a given month. We use
Weighted Nominal Three-Step Estimation (WNOMINATE), a widely used estimation model in
political science, for our estimation (Poole et al. 2011). 1 WNOMINATE is “a scaling procedure
that performs parametric unfolding of binary choice data.” (Poole and Rosenthal 1985) Given a
matrix of binary choices by individuals (for example, Yea or Nay) over a series of Parliamentary
votes, WNOMINATE produces a configuration of legislators and outcome points for the Yea
and Nay alternatives for each roll call using a probabilistic model of choice. WNOMINATE
creates a spectrum of scores ranging from -1 to +1, with -1 representing the most Liberal
Representative and +1 representing the most Conservative Representative (Figure 1). It is worth
mentioning that the WNOMINATE scores have been employed by numerous social scientists to
study the behaviors of the politicians based on their voting records (Aldrich and Battista 2002;
Aldrich et al. 2014; Lupu 2013).
1
Please refer to Appendix A for further information about WNOMINATE and our estimation procedure.
6
To capture the dates Representatives adopted Twitter, we made API calls to Twitter API
and Sunlight Foundation’s Congress API, which helped us to link Representatives’ Twitter
accounts to their legislative data. Out of 445 Representatives, 246 had Twitter accounts by the
time we made the calls. Among the 246 Representatives, 189 joined Twitter during the 111th
Congress (January 2009 – December 2010). With this data, we constructed a binary twitter
adoption indicator for a Representative for a given month. For every month, the value of this
binary variable is 1 if the Representative joined Twitter before or during that month and zero
otherwise.
We adopted Cook’s partisan voting index (PVI) developed and introduced by Charlie
Cook in 1997 as a measure of constituents’ political orientation. Cook’s PVI is a measurement of
how strongly a United States congressional district leans toward the Democratic or Republican
Party, compared to the nation as a whole. Cook employs the last two presidential election results
as a baseline for gauging the political orientation of each congressional district (Cook and David
7
2014; Nunnari 2011). We obtained Cook’s PVI for each congressional district during 111th
Congress. Cook’s PVI for the 111th Congress ranges from 0 to 29 for Republican states and
from 0 to 41 for Democrat states. During the 111th Congress, New York's 15th and 16th districts
with a PVI score of D+41 were the most liberal districts while Alabama's 6th and Texas's 13th
districts with a PVI of R+29 were the most conservative districts. We rescaled Cook’s PVI to
match our WNOMINATE measures. In our dataset Cook’s PVI ranges from -1 to +1 with -1
being the score for the most liberal congressional district and +1 being the score for the most
conservative district.
4. Empirical Methodology
The adoption of Twitter by Representatives over time creates a natural experiment setting
that allows the comparison of difference in voting behavior before and after joining Twitter. We
exploit the variation in joining Twitter across Representatives as the basis for identifying the
impact of adopting Twitter on voting behavior. This strategy has been implemented in numerous
research studies including (Chan and Ghose 2014; Dranove et al. 2003; Jin and Leslie 2003). To
assess the effect of Twitter adoption on Representatives’ voting behavior, we employ the
following fixed effect model:
𝑦𝑦𝑖𝑖𝑖𝑖 = 𝛽𝛽𝑥𝑥𝑖𝑖𝑖𝑖 + 𝛼𝛼𝑖𝑖 + 𝑇𝑇𝑡𝑡 + 𝑢𝑢𝑖𝑖𝑖𝑖
Specification (1)
where i is the index for Representatives and t is the index for time, t = 01-2009, 02-2009, … , 122010; 𝑦𝑦𝑖𝑖𝑖𝑖 is Representative i’s political orientation at time t; 𝑥𝑥𝑖𝑖𝑖𝑖 is the binary variable for
adopting Twitter, meaning that 𝑥𝑥𝑖𝑖𝑖𝑖 = 1 if Representative i has a Twitter account at time t and
zero otherwise; 𝛼𝛼𝑖𝑖 is the fixed effects for Representative i, T is a vector of time fixed effects. In
this model, β is the difference-in-difference estimate of the impact of adopting Twitter on voting
8
behavior of Representatives. If β>0, then joining Twitter has made the Representatives to vote
more conservatively. If β<0, joining Twitter has made the Representatives to vote more liberally.
The Representative fixed effects control for observed and unobserved variations across
Representative such as age, gender, longevity of service, and constituents’ characteristics.
To assess whether the adoption of Twitter makes Representatives to vote more in line
with the political orientation of their constituents, we add a moderating variable (Figure 2) to
Specification (1) to measure the political difference between each Representative and his
constituents:
𝑦𝑦𝑖𝑖𝑖𝑖 = 𝛽𝛽1 𝑥𝑥𝑖𝑖𝑖𝑖 + 𝛽𝛽2 𝑥𝑥𝑖𝑖𝑖𝑖 × 𝑃𝑃𝑖𝑖 + 𝛼𝛼𝑖𝑖 + 𝑇𝑇𝑡𝑡 + 𝑢𝑢𝑖𝑖𝑖𝑖
Specification (2)
where 𝑃𝑃𝑖𝑖 is the political difference between Representative i and her constituent. We obtained 𝑃𝑃𝑖𝑖
by calculating the difference between the mean of WNOMIATE scores of Representative i over
time periods t = 0 , 1, …, 𝑡𝑡 ∗ − 1 and her district’s PVI. Here, 𝑡𝑡 ∗ is the time period during which
Representative i joined Twitter. In other words, 𝑃𝑃𝑖𝑖 is the difference between political orientation
of Representative i and that of her constituent prior to joining Twitter. A positive 𝑃𝑃𝑖𝑖 value
denotes that Representative i is more conservative than her constituent, and a negative 𝑃𝑃𝑖𝑖 denotes
that Representative i is more liberal than her constituent. Therefore if 𝛽𝛽2 is negative and
significant, the Representatives’ political orientation moves closer to that of their constituents
and if 𝛽𝛽2 is significant and positive the Representatives’ political orientation moves further away
from that of their constituents. For instance if Representative i is less conservative than her
constituent before the adoption, the political difference (𝑃𝑃𝑖𝑖 ) would be negative. In this case if
𝛽𝛽2 is negative and significant, Representative i becomes more conservative after the adoption
and therefore moves closer to the constituent.
9
5. Results
Table 1 presents the summary statistics for the dataset. The mean and standard deviation
for Representative’s political orientation is consistent with prior studies (Poole et al. 2011). The
positive mean of Representative’s political orientation denotes that the 111th Congress was
slightly leaned toward conservative side of the political spectrum. Twitter adoption is a binary
variable and equals to zero if the Representative does not have a Twitter account and 1
otherwise.
Table 1. Summary Statistics
Observations
Mean
Std. Dev.
Min
Max
10680
0.011
0.547
-1
1
Twitter adoption
10680
0.404
0.491
0
1
Constituent’s political
orientation (Cook’s PVI)
10680
0.039
0.414
-1
1
Political difference
10680
-0.045
0.266
-0.658
0.507
Representatives’ political
orientation
(WNOMINATE)
10
As explained earlier, we employed Cook’s PVI as a measure of constituent’s political
orientation. Cook’s PVI has been rescaled to match WNOMINATE scores. Similar to
WNOMINATE, Cook’s PVI ranges from -1 to +1 with -1 being the most liberal district and +1
being the most conservative district. The positive mean of constituent’s political orientation
denotes that the constituents were slightly leaned toward the conservative side of the political
spectrum during the 111th Congress.
Political difference is obtained by subtracting constituent’s political orientation from the
temporal mean of Representative’s political orientation before the Twitter adoption. The negative
mean of this variable denotes that the Representatives were slightly less conservative than the
constituents. Political difference ranges from -0.658 to +0.507. As noted earlier, the larger the
political difference value, the more conservative the Representative as compared to the
constituent. The smaller the political difference value, the more liberal the Representative as
compared to the constituent.
According to Table 2, the constituents of Representatives who adopted Twitter during the
111th Congress had a mean political orientation of 0.028. The constituents of Representatives
who did not adopt Twitter at all during the 111th Congress had a slightly higher mean political
orientation (0.053) meaning that the adoption of Twitter by Representatives from less
conservative districts was slightly higher than that of Representatives from more conservative
districts.
11
Table 2. Comparison of means between eventual adopters and non-adopters
Variable
Representatives’ political orientation
(WNOMINATE)
Constituent’s political orientation
(Cook’s PVI)
Difference between means of political
orientation of Representatives and
constituents
Period
Adopter
Before
adoption
-0.229
(0.521)
After
adoption
0.141
(0.545)
Throughout
th
111
Congress
0.028
(0.425)
Before
adoption
-0.257
Nonadopter
-0.029
(0.525)
0.053
(0.399)
-0.082
After
adoption
0.113
Table 2 also reveals interesting information about the Representatives’ political
orientation. Compared to non-adopters, eventual adopters had much lower mean political
orientation before they adopted Twitter. However after the adoption, they had a higher mean
political orientation. That is, adopters became more conservative after joining Twitter.
Comparison of the means between Representatives and constituents in adopters’ districts reveal
that the difference between Representatives and their constituents diminished after the adoption.
According to Table 2, however, the absolute value of the difference is still smaller in the nonadopter districts. Table 3 represents the results of our empirical analysis. Model 1 and Model 2
are based on Specification (1). The difference between Model 1 and Model 2 is the inclusion of
time-fixed effects in Model 2. Twitter adoption is positive and significant in both models. That
is, joining Twitter is associated with the Representatives becoming more conservative.
12
Table 3. Impact of Twitter adoption on Representatives’ left-right political orientation
(dependent variable = Representative’s political orientations)
Variables
Model 1
Model 2
Model 3
Model 4
0.029
0.135***
0.034**
0.112**
Twitter adoption
(0.030)
(0.015)
(0.014)
(0.016)
Twitter adoption × political
-0.204***
-0.194***
difference
(0.048)
(0.039)
Robust
√
√
√
√
Cluster
√
√
√
√
Time-fixed effects
√
√
Adj. R-squared
0.017
0.489
0.024
0.516
N
9724
9724
9724
9724
Specification
FE
FE
FE
FE
* Significant at 0.05, ** Significant at 0.01, *** Significant at 0.001
Model 3 and Model 4 are based on Specification (2) in which we included the interaction
term between Twitter adoption and political difference. We included the time-fixed effects in
Model 4 but not in Model 3. Although Twitter adoption is not significant in Model 4, the
interaction between political difference and Twitter adoption is negative and significant. That is,
political difference moderates the relationship between Twitter adoption and Representatives’
political orientation. Model 4, which includes the interaction term and the time-fixed effects, is
the research model of this study as represented in Figure 2. According to the results of Model 4,
Twitter adoption is not significant anymore. Instead, the interaction between twitter adoption and
political difference is significant and negative. This means, after Twitter adoption, the political
difference between Representatives and the constituent shrinks. Put it differently,
Representatives behave (make decisions) more in line with their constituents after adopting
Twitter.
To test for heteroskedasticity we exploited modified Wald test for group-wise
heteroskedasticity in fixed effects regression model. Since we identified the presence of
13
heteroskedasticity, we employed Eicker-Huber-White robust standard errors in all models. To
decide between random and fixed effect models, we employed Hausman test. The result was in
favor of the fixed effect model (Prob>chi2 = 0.072). We also tested for the time fixed effects and
identified the presence of time fixed effects. Therefore we generated time dummies to obtain the
two level fixed effects estimators (Model 2 and Model 4).
We also tested for cross-sectional dependence using Breusch-Pagan LM test of
independence and Pesaran cross-sectional dependence. Pesaran cross-sectional dependence test
is used to test whether the residuals are correlated across subjects. Cross-sectional dependence
can lead to bias in tests results (also called contemporaneous correlation). The results of our tests
did not signal the presence of cross-sectional dependence. We further employed Wooldridge test
for autocorrelation in panel data; which signaled the presence of serial correlation. Therefore we
used clustered errors in all models. Moreover, instead of the binary Twitter adoption variable, we
employed a variable called Twitter experience. Twitter experience is simply equal to the
common logarithm of the number of days the Representative has been active on Twitter plus 1
(ranges from 0 to 2.844 with a mean of 1.453 and 1.165 standard deviation).
We replicated our analysis by replacing Twitter adoption with Twitter experience in
Specifications 1 and 2. Table 4 represents the results of the analysis. According to Model 8 in
table 4, Twitter experience is not significant in our full model. The interaction between Twitter
experience and political difference is significant and negative. This means that the difference
between political orientation of the Representatives and that of their constituents shrinks as
Representatives gradually gain experience using Twitter.
14
Table 4. Impact of Twitter experience on Representatives’ left-right political orientation
(dependent variable = Representative’s political orientation)
Variables
Model 5
Model 6
Model 7
Model 8
Twitter experience
0.065***
(0.006)
0.033***
(0.009)
Twitter experience × political
difference
0.064***
(0.007)
0.020
(0.010)
-0.110***
(0.019)
-0.101***
(0.017)
Robust
√
√
√
√
Cluster
√
√
√
√
√
Time-fixed effects
√
Adj. R-squared
0.054
0.491
0.063
0.532
N
9724
9724
9724
9724
FE
FE
FE
FE
Specification
* Significant at 0.05, ** Significant at 0.01, *** Significant at 0.001
To further examine the impact of Twitter adoption on Representatives’ political
orientation and the political difference between Representatives and their constituents, we
performed another set of analyses based on Sun and Zhu (2013). We use the following
specification to determine the impact of Representatives’ Twitter adoption on political
orientation:
𝑦𝑦𝑖𝑖𝑖𝑖 = 𝛽𝛽0 + 𝛽𝛽1 𝑄𝑄𝑖𝑖 + 𝛽𝛽2 𝑄𝑄𝑖𝑖 × 𝑥𝑥𝑖𝑖𝑖𝑖 + ∑24
𝑗𝑗=1 𝛾𝛾𝑗𝑗 𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀ℎ𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝐷𝑗𝑗 + 𝜖𝜖𝑖𝑖𝑖𝑖
Specification (3)
where 𝑄𝑄𝑖𝑖 is a dummy that takes the value of 1 if Representative i is an eventual adopter, and 0
otherwise. We call this variable “Adopter”. 𝛽𝛽2 is our difference-in-differences estimator that
captures the adoption’s effect on political orientation of the Representatives. 𝑥𝑥𝑖𝑖𝑖𝑖 is the binary
variable for adopting Twitter, meaning that 𝑥𝑥𝑖𝑖𝑖𝑖 =1 if Representative i has a Twitter account at
15
time t and zero otherwise (This variable is the same as in Specifications 1 & 2, but this time
instead of “Twitter adoption” we call it “Twitter status” to prevent confusion with variable
“Adopter”). We run the model in Specification (3) with two different response variables. For the
first response variable, we use Representatives’ political orientation (the same as Specifications 1
& 2). For the second response variable, we use a variable called political misfit. To construct this
variable, we calculate the Representatives’ mean political orientation before the adoption and
after the adoption. Then we subtract Cook’s PVI from each of these two values. Therefore, we
will have one value for the political difference before the adoption and one value for the political
difference after the adoption. Then, we calculate the absolute values of these values and call this
variable political misfit. Political misfit, simply, measures the absolute difference between
Representatives’ political orientation and their constituents’ political orientation before and after
the adoption. A decrease in political misfit means that Representative is more aligned with his
constituent in terms of political orientation.
We employed Eicker-Huber-White robust standard errors in all models. We also included
dummies for each month from January 2009 to December 2010 to control for changes in all
Representatives’ propensity to shift in the Liberal- Conservative spectrum through changing their
voting behavior. We further employed Wooldridge test for autocorrelation in panel data; which
signaled the presence of serial correlation in models 9 and 10. Therefore we clustered the error
terms at the Representative level in models 9 and 10. We introduced Representative-level fixed
effects to control for time-invariant, unobserved Representative characteristics. Fixed effects
allow us to focus on changes in behavior over time for any given Representative, rather than the
absolute levels. Fixed effects, however, do not control for time-variant unobservables that may
be correlated with the decision to adopt Twitter. These time-variant unobservables could lead,
16
for example, to different trends over time for adopters and non-adopters. Therefore, we construct
a time-variant instrumental variable to account for the effects of time-variant unobservables.
Valid instruments need to correlate with the decision to adopt and affect the dependent variable
only through the adoption decision. We construct the instrumental variable by counting the
number of peers from Representative i’s state who joined Twitter at each time period. The idea is
that the choice to use Twitter among Representatives from the same state may be correlated
(Golbeck et al. 2010; Peterson 2012). We call this variable 𝑁𝑁𝑁𝑁𝑁𝑁𝑁𝑁ℎ𝑏𝑏𝑏𝑏𝑏𝑏_𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝐸𝑖𝑖𝑖𝑖 and incorporate
it in Model 11 and Model 14 by using both fixed effects and 2SLS specifications. It is worth
mentioning that this variable is not correlated with neither political orientation nor political
misfit (p-value>0.05). However, it was correlated with the regressor in both models (pvalue<0.001).
Table 5 reports our regression results. Model 9 reports the results based on ordinary least
squares (OLS) regression. On average, a Representative who adopts Twitter during 111th
Congress is less conservative than a Representative who does not by 18.8 percentage points.
However after the adoption, the Representative’s political orientation shifts toward the
conservative edge of the spectrum by 35 percentage points on average. Model 10 reports the
results with the fixed effects (FE) specification. The results are similar to those in Model 9. The
variable Adopter drops from the regression, because its value does not vary over time. In both
models 9 and 10, the coefficients of the interaction variable, Adopter × Twitter status, reflect the
average effect of the adoption on the treated group. Model 11 reports the results with both fixed
effects and instrumental variables using two-stage least-squares (2SLS).
17
Table 5. Impact of Twitter on Representatives’ political orientation and political misfit
(DV=Representative’s political
orientation)
Model 9
Model 10 Model 11
(DV=political misfit)
Model 12 Model 13 Model 14
Adopter
-0.188***
(0.054)
-0.017***
(0.005)
Adopter × Twitter
status
0.353***
(0.047)
0.045***
(0.009)
0.120*
(0.056)
-0.031**
(0.005)
-0.038***
(0.002)
-0.037**
(0.013)
Robust
√
√
√
√
√
√
Cluster
√
√
Time-fixed effects
√
√
√
√
√
√
Adj. R-squared
0.873
0.489
0.496
0.018
0.019
0.019
N
9320
9320
9320
10537
10537
10537
Specification
OLS
FE
FE/2SLS
OLS
FE
FE/2SLS
Variables
Note 1: We employed Eicker-Huber-White robust standard errors in all models.
Note 2: Within panel R-squared are reported in FE models.
* Significant at 0.05, ** Significant at 0.01, *** Significant at 0.001
The results in Model 11 show that the adoption’s effect becomes stronger after we correct
for endogeneity with the instrumental variable. Similar to Model 9, Model 12 reports the results
based on ordinary least squares (OLS) regression. However, the dependent variable in Model 12
is political misfit. On average, the political misfit for a Representative who adopts Twitter during
111th Congress is around 2 percentage points less than that of a Representative who does not.
This percentage further decreases by an additional 3 percentage points after the Representative
adopts Twitter. Model 13 reports the results with fixed effects (FE), which are similar to those in
Model 12. The variable Adopter drops from the regression, because its value does not vary over
18
time. In both models 12 and 13, the coefficients of the interaction variable, Adopter × Twitter
status, reflect the average effect of the adoption on the treated group. Model 14 reports the results
with both fixed effects and instrumental variables using two-stage least-squares (2SLS). The
results in Model 14 show that the adoption’s effect does not change much after we correct for
endogeneity with the instrumental variable. To further study the impact of Representatives’
Twitter activities on their voting decisions, we replace the variable Twitter status with the log
transformed number of tweets posted by each Representative during each month (ranges from 0
to 5.956 with a mean of 0.751 and 1.295 standard deviation). This variable would inform us in
understanding to what extent the use of Twitter would impact Representatives’ roll call votes.
Table 6 represents the estimation results.
Table 6. Impact of Representatives’ Twitter activity on political orientation and political
misfit
Variables
Adopter
(DV=Representative’s political
orientation)
Model 15
Model 16
-1.117***
(0.069)
(DV=political misfit)
Model 17
Model 18
-0.243***
(0.013)
0.020**
(0.007)
0.020*
(0.009)
-0.022**
(0.007)
-0.022***
(0.005)
Robust
√
√
√
√
Cluster
√
√
Time-fixed effects
√
√
√
√
Adj. R-squared
0.879
0.488
0.920
0.011
N
9320
9320
10537
10537
Specification
OLS
FE
OLS
FE
Adopter × log
number of tweets
Note 1: We employed Eicker-Huber-White robust standard errors in all models.
Note 2: Within panel R-squared are reported in FE models.
Note 3: Standardized variables are employed in all models.
* Significant at 0.05, ** Significant at 0.01, *** Significant at 0.001
19
According to the results in tables 4, 5, and 6, the use of OSN platforms by
Representatives is associated with more conservative votes. More importantly, Representatives
who adopted Twitter during the 111th Congress move closer to the political orientation of their
constituents.
Table 7 provides first-stage estimation results for models 11 and 14 in Table 5 to
illustrate the instrumental variable’s relevance. In Model 19, we did not include the time-fixed
effects. In Model 20, we added the time-fixed effects to the model. We find that the instrument is
highly correlated with becoming a Twitter adopter, and these results are statistically significant at
the 1% level in both models. The overall Wald chi-squared test or F-test for the instrument in
each model is also highly significant.
Table 7. First-stage regressions and instrument relevance (DV = Adopter ×
Twitter status)
Model 19
Model 20
0.033***
0.015***
Neighbor Effect × Twitter status
(0.003)
(0.003)
√
Robust
√
√
Time-fixed effects
Adj. R-squared
N
Specification
0.02
0.05
10680
10680
FE
FE
*** Significant at 0.001
The instrumental variable approach that we have taken to address the selection problem
relies on validity of our instrument. Because we are unable to empirically evaluate the
exogeneity of the instrument, we also undertake a propensity-score matching approach (DiPrete
20
and Gangl 2004; Leuven and Sianesi 2014; Sun and Zhu 2013) to evaluate the impact of
potential selection effects. The instrumental variable approach and the propensity-score matching
approach rely on different sets of assumptions. The instrumental variable approach relies on
exogenous variables to purge the effects of unobservables on the decision to adopt. Propensityscore matching corrects for selection bias by matching adopters with non-adopters based on
observables. Under propensity-score matching scheme, we used Representative’s age, gender,
seniority in Congress, percent of party-favored votes, number of sponsored bills, number of
cosponsored bills, percent of missed votes, and constituent’s mean household income, percent of
high school graduates, percent of white population, and political orientation as attributes to be
matched upon. Samples are matched based on the nearest neighbor algorithm within a caliper
size of 0.05, with replacement. Table 8 provides the descriptions and summary statistics for these
variables.
Table 9 reports the results of our models under the propensity-score matching scheme.
Twitter status estimate in Model 21 remains positive and statistically significant. In particular,
this estimate indicates that a 12.9 percentage point increase in Representatives’ political
orientation follows Twitter adoption. Twitter status estimate in Model 22 remains negative and
statistically significant. In particular, this estimate indicates that a 1.7 percentage point decrease
in political difference follows Twitter adoption by Representatives. It is worth mentioning that in
Model 22, political difference for Representative i at time t is obtained by calculating the
absolute value of the arithmetic difference between Representative i’s political orientation at
time t and his constituent’s political orientation.
21
Table 8. Summary Statistics for Matching Variables
Variable
Description
Age
Representative's age
Gender
Seniority
Party Vote
Sponsorship
Cosponsoriship
Missed Votes
Household
Income
High school
Graduate
White
Political
Orientation
Observations Mean
Representative's gender (1 if
male)
The number of years
Representative has been in the
body (House or Senate) of
which he or she is currently a
member.
The percentage of votes in
which the Representative's
position agreed with the
majority position in his or her
party.
The number of bills sponsored
by the Representative while in
that particular role.
The number of bills
cosponsored by the
Representative while in that
particular role.
The percentage of votes in
which the Representative was
eligible to vote but did not.
Mean logarithm household
income in Representative's
district.
% of high school graduates in
Representative's district
% of white population in
Representative's district
Constituent’s political
orientation
22
Std.
Dev.
Min Max
10680
57.164 10.293
28
86
10680
0.829
0.376
0
1
10320
11.935
9.130
1
56
10320
94.12
4.369
10320
18.997 13.348
0
84
10320
339.691 144.759
0
966
0
93.4
70.83 99.09
10320
4.693
6.990
10680
10.827
0.252 10.084 11.532
10680
85.028
6.761
10680
76.529 17.667 16.04 98.12
10680
0.039
0.414
55.1
-1
96.2
1
Table 9. Regression results under propensity-score matching scheme
Model 21
Model 22
DV: Representative’s
DV: Political
Political Orientation
Difference
0.129***
-0.017*
Twitter status
(0.008)
(0.007)
Robust
√
√
Time-fixed effects
√
√
Representative-fixed
effects
√
√
0.96
0.83
N
10320
10320
Specification
OLS
OLS
Adj. R-squared
*** Significant at 0.001
* Significant at 0.05
Note: Robust standard errors are reported in parentheses below coefficient values.
6. Discussion
Social influence network theory suggests that a social network user’s initial opinion or
behavioral assessment might change due to the information cascaded in the social network
(Friedkin 1998). Furthermore, by communicating and interacting with one another, people create
social influences that affect their opinions, attitudes, and behaviors (Fang et al. 2013; Iyengar et
al. 2011). In this study we find that Representatives’ presence in Twitter platform directs them
toward the conservative side of the political orientation spectrum. According to figure 3,
although the politicians became more conservative due to their presence in Twitter sphere, their
average political orientation has shifted toward the middle of the spectrum, which signals a
higher political balance in the 111th House of Representatives. The comparison between the
mean political orientation of the constituents with that of the adopters shows that Representatives
who adopted Twitter during the 111th Congress moved closer to their constituents in terms of
political orientation.
23
The main finding of this study relates to the interaction between Twitter adoption and
political difference (Model 4). This variable is significant and negative in our model. This means
that in districts where the Representative is more conservative than the constituent Twitter
adoption results in less conservative voting behavior by Representative. That is, the
Representatives adjusts their political orientation according to the constituents’ political
orientation. Similarly, in districts where the Representative is more liberal than the constituent,
Twitter adoption results in less liberal behavior by the Representative. Again, the
Representatives are adjusting their political orientation according to the constituents. As nicely
coined by Bartels (2013), “the public tends to act as a thermostat”, as OSN platforms enable the
constituents to be heard by their representatives in the Congress.
Our findings in Model 4 and Model 8, have been supported by the results of models 9
through 18 and models 21 and 22. According to the results of Model 11 in table 5,
Representatives who eventually adopted Twitter tended to shift toward the conservative
spectrum even when we corrected for the endogeneity of adoption. Furthermore, the results of
24
models 12 and 13 in table 5 reveals that the political misfit between Representatives and their
constituents shrinks as they adopt Twitter platform. The results of Model 14 in table 5 and Model
18 in table 8 support this finding even when we correct for the endogeneity of the adoption. Our
results also suggest that not only the adoption of Twitter by Representatives may impact their
voting decisions, but also the extent to which they generate content in OSN platforms would
influence their voting decisions.
It is worth mentioning that although we do not have an estimate for political orientation
of Twitter users, studies show that Americans deviated from liberalism and became more
conservative during the 111th Congress (Bartels 2013; STIMSON 2013). As figure 4 illustrates,
American’s liberalism policy mood declined from 62.6% from 2008 to 58.2% in 2011.
Meanwhile according to Figure 5, the conservatism policy mood among Americans has
been on the rise during the same period.
25
7. Conclusion
Previous studies suggest that online social networking has caused numerous societal,
economic, and cultural changes. However, the impact of online social networking activities on
politics and policy making has not been adequately tapped. To pursue the goal of studying the
impact of online social networking on the voting behavior of politicians, we constructed a panel
data for 445 Members of the 111th U.S. House of Representatives across a period of 24 months
using three disparate datasets. We collected the voting records of the Representatives, the date
they created their Twitter accounts and their constituent’s political orientation. Using a
difference-in-difference approach, we revealed that the adoption of Twitter by Representatives
directs them toward the conservative side of political spectrum. Furthermore, we found that the
relationship between adoption and orientation shift is moderated by the political difference
between Representatives and their constituents. That is, in case of a Representative whose
political orientation is more conservative than that of her constituent, the adoption directs her
26
toward the liberal side to be more aligned with the constituent. In case of a Representative whose
political orientation is more liberal than that of her constituent, adoption leads her to a more
conservative point in the spectrum so that the Representative and the constituent are further
politically aligned. We also found evidence that Representatives who create more content in
Twitter sphere move closer to their constituents in terms of political orientation. More analysis
supported our finding that Representative’s voting behavior becomes more aligned with the
constituent’s political orientation after the adoption of Twitter by Representative. 2
8. Limitations and Future Research
One of the limitations of this study is related to the sample. U.S. Representatives are elite
politicians whose decision making in politics would differ from regular citizens. Thus,
generalizability of these findings shall be limited to the elite politician population.
To better extend the realm of this research, one may study the dynamic network of
politicians in social media. Such network can be built based on friendships or conversations in
online social networking platforms. A dynamic network analysis may shed more light on the
underlying mechanism that causes the change in political orientation. Moreover, we only
extracted and analyzed data from Twitter platform due to its popularity in political domain. A
good extension of this study would be studying the impact of adoption of other OSN platforms
by politicians on their political orientation. Although Twitter is sometimes perceived as a
broadcasting medium rather than a social network, it shares certain features with other OSN
2
It is worth noting that in a separate study we found that Representatives who adopted Twitter by the end of The
111th Congress had a higher chance for being re-elected in The 112th Congress. We studied the impact of
Representatives’ social media adoption on the results of the next election by collecting data about Twitter,
Facebook, and Youtube adoption by Representatives. After controlling for a number of Representative-specific and
constituent-specific factors, we found that Twitter and Facebook adoption are positively associated with the
Representative being elected for the next term. Youtube adoption however, did not yield any significant results. The
results of this study will be reported and discussed in a separate paper.
27
platforms. For instance, Twitter enables the users to follow and be followed by others. Or, a
Twitter user only receives the tweets from her following list on her home page. While some
users might decide to only broadcast their own ideas, prior studies show that Twitter users read
tweets posted by people they follow. For instance, a study by Liu and colleagues (2014) shows
that during the 111th Congress (from January 2009 to December 2010), more than 30% of the
posts on Twitter were either replies or retweets. Twitter also recommends out of network users
based on the current network of followers/ followings. According to Mousavi and Gu (2014), a
strong relationship exists in the following/ follower network of Representatives in 113th House
of Representatives in Twitter platform. Replicating this study in other OSN platforms with
different set of features may enable the researchers to examine information-richness of the
networks and thus elaborate on the mechanism of influence on greater detail.
Another interesting extension of this study would be the study of the posts by
Representatives. A granular dataset that includes the actual social networking activities (e.g.
tweets, number of friends, number of followers, …) would allow the researchers to elaborate on
the underlying mechanism of social media influence on political orientation and decision
making.
Last but not least, a dynamic monthly measure for constituents’ political orientation
could have helped us to construct a more accurate measure for political difference and political
misfit. However, to our best knowledge, all of the measures developed for constituents’ political
orientation are for four years or more time periods. It is worth mentioning that similar to Cook’s
PVI the majority of these measures are developed, at least partly, based on the presidential
elections data (Kernell 2009).
28
9. References
Aldrich, J. H., and Battista, J. S. C. 2002. “Conditional Party Government in the States,”
American Journal of Political Science, (46:1), pp. 164–172.
Aldrich, J. H., Montgomery, J. M., and Sparks, D. B. 2014. “Polarization and Ideology: Partisan
Sources of Low Dimensionality in Scaled Roll Call Analyses,” Political Analysis
(Forthcomin).
Aral, S., Dellarocas, C., and Godes, D. 2013. “Introduction to the Special Issue —Social Media
and Business Transformation: A Framework for Research,” Information Systems Research
(24:1), pp. 3–13.
Bartels, L. 2013, September 30. “Americans are more conservative than they have been in
decades,” The Washington Post .
Bertot, J. C., Jaeger, P. T., and Grimes, J. M. 2010. “Using ICTs to create a culture of
transparency: E-government and social media as openness and anti-corruption tools for
societies,” Government Information Quarterly (27:3), pp. 264–271.
Burt, R. 1992. Structural Holes: The Social Structure of Competition, Cambridge, MA: Harvard
University Press.
Chan, J., and Ghose, A. 2014. “Internet’s Dirty Secret: Assessing The Impact of Online
Intermediaries on HIV Transmission,” MIS Quarterly (Forthcomin).
CHEN, Y., WANG, Q., and XIE, J. 2011. “Online Social Interactions: A Natural Experiment on
Word of Mouth Versus Observational Learning,” Journal of Marketing Research (XLVIII),
pp. 238–254.
Cook, C., and David, W. 2014. “Recalibrating Ratings for a New Normal,” PS: Political Science
& Politics (47:2), pp. 304–308.
DiPrete, T. A., and Gangl, M. 2004. “Assessing Bias in the Estimation of Causal Effects:
Rosenbaum Bounds on Matching Estimators and Instrumental Variables Estimation with
Imperfect Instruments,” Sociological Methodology (34:1), pp. 271–310.
Dranove, D., Kessler, D., Mcclellan, M., and Satterthwaite, M. 2003. “Is more information
better? The effects of ‘report cards’ on health care providers,” The Journal of Political
Economy (111:3), pp. 555–588.
Edvardsson, B., Tronvoll, B., and Gruber, T. 2011. “Expanding understanding of service
exchange and value co-creation: a social construction approach,” Journal of the Academy of
Marketing Science (39:2), pp. 327–339.
29
Fang, X., Hu, P. J., Li, Z., and Tsai, W. 2013. “(2012). Predicting Adoption Probabilities in
Social Networks. Information Systems Research, 24(1), 2013.,” Information Systems
Research (24:1), pp. 128–145.
Friedkin, N. E. 1998. A Structural Theory of Social Influence, New York: Cambridge University
Press, p. 231 pages.
Goh, K.-Y., Heng, C.-S., and Lin, Z. 2013. “Social Media Brand Community and Consumer
Behavior: Quantifying the Relative Impact of User- and Marketer-Generated Content,”
Information Systems Research (24:1)INFORMS, pp. 88–107.
Golbeck, J., Grimes, J. M., and Rogers, A. 2010. “Twitter use by the U.S. Congress,” Journal of
the American Society for Information Science and Technology , p. n/a–n/a.
Greenberg, S. R. 2012. “Congress + Social Media,” Austin, TX, pp. 1–55.
Iyengar, R., Van den Bult, C., and Valente, T. 2011. “Opinion leadership and social contagion in
new product diffusion,” Marketing Science (30:2), pp. 195–212.
Jin, G. Z., and Leslie, P. 2003. “The Effect of Information on Product Quality: Evidence from
Restaurant Hygiene Grade Cards,” The Quarterly Journal of Economics (118:2)Oxford
University Press, pp. 409–451.
Kernell, G. 2009. “Giving Order to Districts: Estimating Voter Distributions with National
Election Returns,” Political Analysis (17:3), pp. 215–235.
Leuven, E., and Sianesi, B. 2014, February 12. “PSMATCH2: Stata module to perform full
Mahalanobis and propensity score matching, common support graphing, and covariate
imbalance testing,” Statistical Software Components, Boston College Department of
Economics.
Linders, D. 2012. “From e-government to we-government: Defining a typology for citizen
coproduction in the age of social media,” Government Information Quarterly (29:4), pp.
446–454.
Liu, Y., Kliman-Silver, C., and Mislove, A. 2014. “The Tweets They are a-Changin’: Evolution
of Twitter Users and Behavior,” in Proceedings of the Eighth International AAAI
Conference on Weblogs and Social Media, , pp. 305–314.
Luo, X., Zhang, J., and Duan, W. 2013. “Social Media and Firm Equity Value,” Information
Systems Research (24:1), pp. 146–163.
Lupu, Y. 2013. “The Informative Power of Treaty Commitment: Using the Spatial Model to
Address Selection Effects,” American Journal of Political Science (57:4), pp. 912–925.
30
McCarty, N., Poole, K. T., and Rosenthal, H. 2008. Polarized America: The Dance of Ideology
and Unequal Riches, The MIT Press, , p. 256.
Mousavi, S., and Demirkan, H. 2013. “The Key to Social Media Implementation: Bridging
Customer Relationship Management to Social Media,” in 2013 46th Hawaii International
Conference on System Sciences, , January 1, pp. 718–727.
Mousavi, S., and Gu, B. 2014. “The Role of Online Social Networks in Political Polarization,” in
Twentieth Americas Conference on Information Systems, Savannah, GA, pp. 1–11.
Nunnari, S. 2011. “The Political Economy of the U.S. Auto Industry Crisis,” Working Paper, pp.
1–37. http://www.columbia.edu/~sn2562/nunnari_autobailout.pdf
Peterson, R. D. 2012. “To tweet or not to tweet: Exploring the determinants of early adoption of
Twitter by House members in the 111th Congress,” The Social Science Journal (49:4), pp.
430–438.
Poole, K., Lewis, J., Lo, J., and Carroll, R. 2011. “Scaling Roll Call Votes with WNOMINATE
in R,” Journal of Statistical Software (42:14).
Poole, K. T., and Rosenthal, H. 1985. “A Spatial Model For Legislative Roll Call Analysis,”
American Journal of Political Science (29:2), pp. 357–384.
Poole, K. T., and Rosenthal, H. 2007. Ideology & Congress, New Jersey: Transaction Publishers,
p. 361.
Rishika, R., Kumar, A., Janakiraman, R., and Bezawada, R. 2013. “The Effect of Customers’
Social Media Participation on Customer Visit Frequency and Profitability: An Empirical
Investigation,” Information Systems Research (24:1)INFORMS, pp. 108–127.
STIMSON, J. A. 2013. “Policy Mood,” http://www.unc.edu/~jstimson/Welcome.html, .
Sun, M., and Zhu, F. 2013. “Ad Revenue and Content Commercialization: Evidence from
blogs,” Management Science (59:10), pp. 2314–2331.
Wu, L. 2013. “Social Network Effects on Productivity and Job Security: Evidence from the
Adoption of a Social Networking Tool,” Information Systems Research (24:1), pp. 30–51.
Yan, L., and Tan, Y. 2010. “An empirical study of online supports among patients,” Working
Paper, pp. 1–39. http://papers.ssrn.com/sol3/papers.cfm?abstract_id=1697849
31
Appendix A: Estimating WNOMINATE Scores
To estimate WNOMINATE scores for Representatives, we employed a software package
designed to estimate Poole and Rosenthal WNOMINATE scores in R. This package uses a
logistic regression model to analyze a series of Parliamentary votes. According to Poole et al.
(2011), WNOMINATE assumes probabilistic voting based on a spatial utility function, “where
the parameters of the utility function and the spatial coordinates of the legislators and the votes
can all be estimated on the basis of observed voting behavior (p. 1). One of the key inputs of this
program is the roll call matrix for The 111th House of Representatives. The roll call matrix is the
result of two sets of variables: an ideal point for each Representative that stands for their
ideology or vote preference, and separate yea and nay locations for each roll call. It is assumed
that the Representatives have an ideal point of each of these two dimensions. As explained in
Poole and Rosenthal (2007) and widely used in political science literature (Aldrich and Battista
2002; Aldrich et al. 2014; Lupu 2013), the first dimension can be interpreted as the LiberalConservative spectrum. The second dimension picks up social issues such as civil rights for
African- Americans during. According to McCarty et al. (2008), this dimension is no longer
important. Therefore, the estimates on the first dimension were used in our study.
Since we need WNOMINATE scores for each Representative during each month, we created roll
call matrices with Representatives’ votes cast during each month of the study. Legislators who
voted fewer than twenty times during each month were excluded from estimation and were
treated as missing observations. Along with the roll call matrix WNOMINATE program requires
other inputs, most of which are set by default as reported in (Poole et al. 2011). The only thing
that needs to be set is the argument polarity, which is used to direct the results in the desired
direction. The polarity is set by specifying a Representative to be positive in each dimension.
Since in general researchers wish to orient Conservatives on the right and Liberals on the left,
positive in this case refers to Conservative side of the spectrum. Since there are two dimensions,
the polarity needs to be set for each dimension. Therefore, we need one fiscally Conservative
Representative (Republican Representative) to set the polarity on the first dimension and one
socially Conservative Representative on the second dimension. We decided to use
Representative Eric Cantor (R-VA 7th district) for the first dimension and Representative Walter
Jones, Jr. (R-NC 3rd district) for the second dimension. The reason is that Representative Cantor
32
has high score on the first dimension (fiscally Conservative) but low scores on the second
dimension. In contrast, Representative Jones has low score on the first dimension but high scores
on the second dimension (socially Conservative). Using these settings we estimated first
dimension scores for each Representative for each month and used as a measure for ideological
polarization as reported in the manuscript.
33