CSCW 2013 Talk, including supplementary slides
As Twitter becomes a more common means for officials to communicate with their constituents, it becomes more important that we understand how officials use these communication tools. Using data from 380 members of Congress’ Twitter activity during the winter of 2012, we find that officials frequently use Twitter to advertise their political positions and to provide information but rarely to request political action from their constituents or to recognize the good work of others. We highlight a number of differences in communication frequency between men and women, Senators and Representatives, Republicans and Democrats. We provide groundwork for future research examining the behavior of public officials online and testing the predictive power of officials’ social media behavior.
3. Descriptive Statistics
N Median S.D. Range
Overall 35,361 1,090 2,134 18-8,893
By Females 5,535 760 873 60-3,677
By Males 29,826* 1,155 2,262 18-8,893
By Republicans 21,253* 1,228 2,544 51-8,893
By Democrats 13,648 825 799 18-3,005
By Independents 460 3,372 0 3,372
By Representatives 28,834* 1,055 2,266 18-8,893
By Senators 6,527 1,219 1.396 165-5,927
* Marks groups who were significantly more active than
their counterparts.
4. Data
• 380 members of Congress
• 12/20/2011 – 2/29/2012
Coding
• Getting Code Agreement: 6 coders, 791 tweets, excluding
RTs
• Training Data: 526 tweets with binary values for each of 5
codes
• Coding Process: Mallet’s MaxEnt classifier on 30,373
tweets
5. Tweets as Speech-Acts
• “in saying something, we do
something” – Austin, 1962
• Performance
• Goal-orientation
7. Directing to Information – 41%
The Bureau of Labor Statistics
reports today that the US economy
added 200,000 jobs in Dec.
Unemployment falls to 8.5%.
http://t.co/WHZO7RaR
(Rep. Andre Carson, D-IN)
8. Positioning – 22%
President Obama is again bypassing Congress-
this time to give amnesty to an untold number of
illegal immigrants- http://t.co/KhqoQBCQ
(Rep. Walter Jones, R-NC)
House Republicans refused to let me speak on
House floor today. GOP needs to return to work on
#payrolltaxcut. Video: http://t.co/YwZFxwWb
(Rep. Jim Moran, D-VA)
9. Narrating – 7%
I'm talking to CNN's @randikayecnn at
1:15pm ET and MSNBC's
@mitchellreports at 1:45pm ET please
tune in! #nhprimary #FITN
(Rep. Debbie Wasserman Schultz, D-
FL)
10. Thanking – 2%
Thank u Matt Strawn for the
successful leadership u gave to
IaGOP Enjoy a rest. Pls continue to
help us in someway to ur liking
(Sen. Chuck Grassley, R-IA)
11. Requesting Action – 1%
RSVP to my Immigration Forum
with Rep. Luis Gutierrez this
Saturday in Brooklyn
http://t.co/qTcWugs
(Rep. Yvette Clark, D-NY)
26. Moderated effects – Twitter-action and sub-group – upon following
Speech-Act Narrating Positioning Providing Requesting Thanking
info action
Male
GOP - - -
Female
GOP - - -
Male
Dem - -
Female
Dem - -
F-stat 18.47*** 17.72*** 18.10*** 17.88*** 18.52***
R2 0.02 0.02 0.02 0.02 0.02
27. Action tweets’ effects on audience size
80%
60%
40%
20%
thanks
0%
request action
-20% providing info
-40% positioning
narrative
-60%
-80%
-100%
Male Female Male liberals Female
conservatives conservatives liberals
28. Predicting Voting Behavior using Frequency of
Positioning
Dependent variable DW-NOMINATE
Speech-act Positioning (raw)
Male GOP 0.001*
(baseline)
Female GOP -0.101
Male Dem -0.206***
Female Dem -0.160***
F-stat 27.47
R2 0.27
29. Takeaways
• Men, Republicans, Representatives
more active
• Broadcast mechanism
• Implicitly campaigning all the time
• Effects on audience not uniform
• More positioning, more polarized
31. Ongoing Work
• Is Congress polarized like the public?
• Does Twitter provide an alternate path to
influence?
• How do politicians interact with their
constituents?
• How do constituents interact with their
politicians?
• What’s happening in the EU? South Korea?
32. Contact us
• Libby Hemphill (libby.hemphill@iit.edu; @libbyh)
• Jahna Otterbacher (jotterba@iit.edu)
• Matt Shapiro (mshapir2@iit.edu)
Illinois Institute of Technology
@casmlab
http://www.casmlab.org/projects/publicofficials/
https://twitter.com/CaSMLab/lists
34. Why study Congress?
• > 90% adoption rate
• ~650 tweets per day
• Reaching > 35K followers
• Plenty of hype
• No traditional media corporation mediating conversation
between officials and constituents
35. Coding
Golbeck, Grimes, and Rogers Our Study
• Getting Code Agreement: • Getting Code Agreement:
3 coders, 200 tweets 6 coders, 791 tweets,
• Coding Process: 3 coders excluding RTs
each coded 2/3; 4,626 • Training Data: 526 tweets
tweets
with binary values for
• Agreement: Included only
each of 5 codes
tweets with identical
codes from two coders • Coding Process: Mallet’s
• Codes: Tree scheme, MaxEnt classifier on
some branches mutually 35,361 tweets
exclusive
36. Cohen’s
Code Definition kappa N (%)
Narrating Telling a story about their day, 0.83 2,069
describing activities (7%)
Positioning Situating one's self in relation to 0.87 6,728
another politician or political (22%)
issue, may be implied rather than
explicit
Directing to Pointing to a resource URL, telling 0.70 12,468
information you where you can get more info (41%)
Requesting action Explicitly telling followers to go do 0.70 299
something online or in person (not (1%)
just visiting a link but asking them
to do something like sign a
petition, apply, vote) - look for
action verbs
Thanking Says nice things about or thanks 0.90 667
someone else, e.g. (2%)
congratulations, compliments
38. Comparing tweet frequency
Model 1 Model 2 Model 3 Model 4
Male 943.309 597.863 593.203 585.361
Republican 1110.179 1105.994 1201.930
Senate -113.932 -134.147
Days in Office 0.077
Constant 1080.438 704.562 732.038 427.333
r2 0.026 0.087 0.088 0.101
All coefficients significant; p < 0.001
41. Predicting Voting Behavior using Relative Frequency
of Positioning
Dependent variable DW-Nom.
Speech-act Positioning
Male GOP 0.073
(baseline)
Female GOP -0.109*
Male Dem -0.211***
Female Dem -0.156***
F-stat 28.9
R2 0.29
42. Positioning tweets’ effects on
DW-NOMINATE – by subgroup
0.1
0.08
0.06
0.04
0.02
0
-0.02
-0.04
Male Female Male liberals Female liberals
conservatives conservatives
43. Future Work
• Government Responsiveness
• Constituent lobbying efforts
• @ replies from MoCs
• Civic Engagement
• Voting records
• Non-voting political activities
Editor's Notes
CampaigningWebsites – informative, not positioning
12/22/2011 – 3/14/2012Internet blackout – 1/18/2012State of the Union – 1/24/2012Giffords resigned – 1/25/2012Kirk stroke – weekend of 1/20-1/22/2012Super Tuesday – 3/6/2012Solar storm – 3/8/2012FoxNews What’s Happening Now?
Cohen’s kappas of 0.7 or more5 action codes and one “other”MaxEnt – assumes 50/50 tweet is narrative or not, learns constraints to apply380 members were active Twitter users who had ever mentioned someone else in Congress – part of a larger project about how Congress engages each other and their constituentsStopped before Super Tuesday
We see tweets as speech-acts in the sense that Austin described in How to Do Things with Words. By that I mean that we treat each tweet as an utterance with elements of performance and goal-orientation. By tweeting, officials are doing something whether its directing our attention to a certain information source or trying to get us to do some activity. And we used speech-acts to develop those 5 codes I mentioned earlier. Those 5 codes correspond to various goals officials have in posting a tweet
Pointing to a resource URL, telling you where you can get more infoEven when “providing information” officials make choices about where to direct our attention – nearly all information tweets have URLs in them, implying the official wants us to visit that URL. His/her act then is not just to transfer information but to direct our attention and action toward that particular source of information.IN 7th, political family (Grandmother was a Rep), IndianapolisFollowing Yun’s talk – we see lots of informational, media use by Congress. Stay tuned for results about social, relational uses
Situating one's self in relation to another politician or political issue, may be implied rather than explicitJones – NC 3rd, Outer Banks and Atlantic coastMoran – VA 8th, NoVA including Arlington
Telling a story about their day, describing activitiesFL 23rd, Chair of the DNC, Miami area
Says nice things about or thanks someone else, e.g. congratulations, complimentsHouse in the 70’s, Senate since 1981
Explicitly telling followers to go do something online or in person (not just visiting a link but asking them to do something like sign a petition, apply, vote) - look for action verbsNY 9th and 11th since 2007 – BrooklynGutierrez – IL 4th, Cook County, including (west) Chicago
137 voting records + mention data380 mentionersThis is an ugly slide that you’re not supposed to read. The takeaway here is that we did detect differences among groups about how various speech acts affected the size of their audience.
I find these effects of speech acts on audience size, measured in followers, easier to get graphically, and we see they differ between parties and sexes. Republicans with big audiences don’t request much action, but they do thank and congratulate. Democrats, on the other hand, have smaller audiences when they do a lot of thanking and congratulating.
We use just the first dimension of DW-NOMINATE, a scale from -1 to 1 that roughly maps liberal to conservative. Its based on roll call votes and is widely used to talk about polarization in Congress.The effect is stronger for Democrats, and we saw before that positioning actually reduces the size of Female Dems’ audiences.
Franking rules
Congress less polarized than political blogosphereA couple people asked Andrew at our poster about how politicians respond to people who lobby them via Twitter. That’s a great question and one we’re just beginning to answer.
Not to say that there’s nothing mediating the connections between politicians and their constituents, but that the traditional media role is usurped on Twitter.100% of the newly elected MoCs in the 113th Congress have Twitter accounts.
Cohen’s Kappa for human ratersN (%) for machine coder
Mean accuracy using 10-fold cross-validationNaïve Bayes: This classifier infers the label of a tweet based on the conditional probabilities of the words it contains. Suppose the text of the tweet is “Meeting with Senator Smith this morning.” In order to determine whether or the tweet should be labeled as “narrative,” the classifier determines the probability of a tweet in general being labeled “narrative,” and multiplies this by the conditional probabilities of each word in the tweet occurring, given that the tweet is in a fact a narrative tweet. This would then be compared to the conditional probability of the set of words being used in a tweet that is not a narrative tweet. This classifier is “naïve” in the sense that it assumes that words occur independently of one another, which clearly is not the case.Decision tree: A classifier that models the process of asking questions about the input tweet, in order to determine its likely action code. For instance, at each node in the tree, we would determine whether or not a particular word is present in the tweet (e.g., Does the tweet contain a URL? Does the word “thank” occur? If so, does the word “you” also occur?)Maximum entropy: Researchers have applied maximum entropy to text classification problems in an attempt to get around the independence assumptions of naïve Bayes, which were mentioned above (Nigam). Maximum entropy models begin with the assumption that uniform distributions are preferred (i.e., assume a 50/50 chance that a tweet is “narrative” or not). They use training data to learn constraints to be applied to this distribution. Nigam and colleagues report that in many cases, maximum entropy outperforms naïve Bayes, however, it does have a tendency toward overfitting in cases where data is sparse (i.e., there are few positive examples of a tweet of a given class).
137 voting records + mention data380 mentionersThis is an ugly slide that you’re not supposed to read. The takeaway here is that we did detect differences among groups about how various speech acts affected the size of their audience.
We use just the first dimension of DW-NOMINATE, a scale from -1 to 1 that roughly maps liberal to conservative. Its based on roll call votes and is widely used to talk about polarization in Congress.
For the most part, more positioning correlates with being more polarized in one’s voting records. That holds for both genders and in both countries.