0.4
0.08
Prevalence of violent events
Average no. of violent events
0.07
0.3
0.2
0.06
0.05
0.04
0.1
0.03
2014
2016
2018
2014
Time
2016
2018
Time
Figure 2: Violent events during the sample period. The vertical dashed lines show the
beginning and end of the Facebook campaign. Data source: ACLED
on political violence and protest (defined as having a political purpose or motivation).
Because ACLED uses a different categorisation of events, it allows me to further explore
heterogeneity in conflict types.
The events in ACLED are categorized violent events, demonstrations, and non-violent
actions. My main focus is on violent events, which are further classified as battles, explosions/remote violence, and violence against civilians. Most frequent event type is battles,
and more specifically armed clashes. Most frequent actor types are state forces and political militias. Figure 2 plots the time series of violent events in ACLED data. The left
panel shows the monthly number of violent events in Myanmar, and the right panel shows
the monthly share of townships with conflict events. During the sample period, every
month on average 5% of townships experienced at least one violent event. Unlike GDELT,
ACLED data exhibits a slightly increasing trend in conflict occurrence. Appendix Figure
B.3 shows the frequency of different types of violent events.
Figure 3 shows the geographic distribution of conflict events in the two sources. Both
panels map the population weighted number of conflict events between June, 2016 and
end of August, 2017. The figures show that conflict events are more pronounced in the
peripheral areas, and particularly in Rakhine state (in Western Myanmar) which is home
to majority of the Rohingya, and in the Shan (North-Eastern Myanmar) and Kachin
states (Northern Myanmar). ACLED contains much less conflict events than GDELT,
and the events are more geographically concentrated on the northern and eastern parts
of the country. In a related study, Manacorda and Tesei (2020) compare GDELT with
ACLED and Social Conflict Analysis Database (another manually compiled dataset). The
authors show that, assuming that the probability that an event is correctly reported is
larger than the probability of incorrect reporting, true reporting is more likely in GDELT
data than in the manually compiled datasets.
9