Violence, Sectarianism and Patterns of Communication in Yemen
MURI Presentation Christia, Dahleh, Jadbabaei, Leskovec, 1
Slide 2
Question What is the effect of violence on a societys structure
of communication? Examine effects of exogenous violent shocks
indigenous violent events on patterns of communication in Yemen.
Specifically, look at the effects of drone strikes and Arab spring
protests on cell phone communication. For the Arab spring protests
we can also look at the effect of communications on protest
activity Builds on recent work that uses CDRs to examine the change
in social networks of communication after rapidly evolving and
transient events such as emergencies. 2
Slide 3
Patterns of Communication 3 years of call records metadata from
January 2010-January 2013 Anonymized data on: Caller Location Date
and time of call Duration of call Call or sms Increase in cell
phone penetration from roughly 40% to over 70% in that time span.
No 3G in Yemen and only 15% internet penetration. 3
Slide 4
Users and Calls per Day 4
Slide 5
Data on Violent Incidents Exogenous: 73 Drone
StrikesIndigenous: 432 Arab Spring Events 5
Slide 6
6 Drone Strikes Jan 2010-Jan 2013
Slide 7
7 Bulk of Strikes in 2012-Jan 2013
Slide 8
8
Slide 9
9
Slide 10
Analytical Approach Address our question by looking at: Volume
of calls Timing of calls Direction of calls (incoming vs. outgoing)
Duration of calls Mobility effect in calls Individual level network
connections. Also look at how violent events compare to other
important but largely peaceful events Look at whether phone data
reveals the countrys divisions. 10
Slide 11
Day of Week Call Patterns 11 Friday Prayer
Slide 12
First Strike: 24 May 2010 Tower 2: 3.7 miles from strike Tower
1: 0.6 Miles from strike 12 Marib: Approximately 17,000 people,
2,200 unique numbers
Slide 13
Marib Strike May 24 th 2010 Entire CountryTower 0.6 miles from
strike 13 Spike
Slide 14
Direction of Call Traffic May 24 th 2010 14
Slide 15
Duration vs. Time Plotted Average Call Duration vs. Time of day
Slight decrease in average call duration right after strike i.e.
people make shorter phone calls This decrease only happens in
outgoing calls 15 Average duration is a lot longer at night than
during the day (2 mins vs. ~5 mins) After drone strike, average
duration is much shorter than the average at that time of day
Slide 16
Duration After Strike Calls in the 1 hour after the strike
happened Red normal night, long-tailed, Blue May 24 th,
narrow-bands 16
Slide 17
Al-Awlaki Strike September 30 th, 2011 in the middle of nowhere
He had pulled over for breakfast on side of road Drone Strike 37
Miles Away 34 Miles Away Nothing for 30 miles here 17
Slide 18
Al-Awlaki Strike No spike at all at the 2 nearest towers 18 The
day of the strike was quite calm Note the standard pattern on
Friday, even on the day Awlaki was killed
Slide 19
Another Strike - Lawdar Lawdar 1/30/2012, Killed 15 militants
Arab Spring happening in the same time period 19 Just one set of
towers in town, 0.3 miles away
Slide 20
Lawdar 20 The outlier on January 30 th A jump in calls around
11:00PM, consistent with other spikes (incoming/outgoing/within
tower/duration), but not too large comparatively... Large
differences in total volume for Mondays over just a 2 month
interval (not monotonically increasing each week) To zero out
influence of strike, normalized call volume Probably not a good
idea in the long run, but helps here Important next step:
Distinguish drone strikes in war-torn vs. peaceful areas Power
outage
Slide 21
Measurable Strikes 73 are between Jan 2010 and Jan 2013 10 in
Zinjibar, where antennas were shut off as under Al Qaeda control;
for 1 no location data Of the remaining 62: 33 strikes less than
2.5 miles from a tower, 41 less than 5 miles, 51 less than 10 miles
Remaining 11 greater than 10 miles (unlikely to see a spike)
21
Slide 22
Anecdotal (so far) findings Based on the handful of cases we
have looked at: Drone strike effect on communications appears very
localized and quite contained in terms of time. Hits appear to
happen at night which probably helps with keeping the effect more
contained. Strike seems to have shortening effect in terms of
duration of outgoing calls. There also appears to be an increase in
incoming calls. But all this very tentative as based only on a few
cases. 22
Slide 23
Next Steps See if we could predict the time of a strike by a
spike/deviation from the standard level for that day of the week
for the towers near the strike for all 62 strikes in the dataset.
Look for missing strikes. Quantify local spikes for other
activities such as religious holidays, celebrations, other cultural
events, and leverage them in connection to violent events. Look at
whether tower outages are in any consistent way linked to drone
strikes. Look at individual level data (who called whom). Do the
same analysis for Arab Spring events. There, also look at whether
there is any network behavior that seems to affect events. 23