Take-home Exercise 2: Geospatial Analytics for Peace
Where armed conflict clusters in Myanmar, how its geography is changing, and whether it spreads. 330 townships, January 2021 to September 2025.
3. A war that changed its means
Violence climbed through 2021 and peaked with Operation 1027 in November 2023. Explosions and remote violence overtook battles in 2023, driven by the military’s air campaign: 88 air or drone strikes in 2021, 2,977 in 2024. An air strike lands where the military chooses, not where the fighting is.
4. Where violence clusters
All political violence, p < 0.05, 999 permutations.
Global Moran’s I is 0.67. A township’s violence is very well predicted by its neighbours’.
One dominant core: the dry zone of Sagaing, Magway and Mandalay, with a second belt down the Thai border through Kayin, Mon and Tanintharyi.
Cold spots: the Ayeyarwady delta, Yangon and eastern Shan.
No High-Low outliers at all. Violence in Myanmar does not form islands; it comes in neighbourhoods.
82 to 96% of townships keep their class under raw counts or 6-nearest-neighbour weights. About half survive a false discovery rate correction: the cluster cores.
5. Every kind of violence, the same core
Local Moran’s I by type. Battles add the ethnic fronts in Chin, Kachin and Kayah. Violence against civilians is the most concentrated: 88% of its High-High townships are in the dry zone. Fatalities follow battles.
6. Which way the trends run
Mann-Kendall test on each township’s 19-quarter Gi* series, p < 0.05.
Cooling: Yangon. 44 of its 45 townships have a falling Gi*. The urban resistance of 2021 is gone from the map.
Warming: the middle belt. Bago West, Mandalay, Magway, Bago East and Tanintharyi, where the war arrived late as the dry-zone insurgency spread and new fronts opened.
Sagaing barely registers, because it was hot from 2021 and stayed hot. A trend test cannot see a level, which is the gap emerging hot spot analysis fills.
7. Emerging hot spots
Intensifying, the most urgent: Dawei, Yesagyo, Seikphyu, Madaya, Myingyan, Ngazun. Hot in nearly every quarter and still rising, five of six in the dry zone.
Consecutive, a front that moved and held: Kyaukkyi and Shwegyin through Kyaikto and Bilin to Kawkareik and Myawaddy on the Thai border.
Oscillating hot spots fill Sagaing and Magway. Not unstable: they were cold in the coup quarter, before the resistance existed, and have been hot since.
Persistent cold spots are the Shan hills, including Wa and Mongla.
Fine classes are fragile: 62% of townships keep their label on a monthly cube, 78% keep hot/cold/neither. Every label was checked against an independent Gi* series, because sfdep 0.2.5 can attach labels to the wrong townships.
8. Violence spreads to neighbours
Value-added: a spatial Markov chain. Each township-quarter is active (3 or more events) or quiet. Does a quiet township’s chance of becoming active depend on its neighbours?
Yes, by a factor of four: 9% with no active neighbour, 36% with most neighbours active.
It is not just “a violent region”. With fixed effects for state and quarter and the township’s own history, the odds ratio is 4.1 (bootstrap 2.8 to 6.6). Shuffling townships’ histories across the map never reproduces the gap (p < 0.002). It holds at 1, 3 and 6 events.
Out of sample it is a modest signal: AUC 0.73 from own history, 0.75 with neighbours.
9. Fronts, not sparks
Since Operation 1027, isolated outbreaks have almost stopped. A quiet township with no active neighbour ignited 13% of the time before the offensive and 4% after. Onset next to active neighbours did not change (36% and 36%).
New violence now appears at the edge of existing violence: a war of consolidated fronts rather than the scattered uprising of 2021.
The map: townships quiet in 2025 Q3, shaded by the share of neighbours active. The warmest are where the evidence says to look next, most of them in Magway, Sagaing, Kachin and Mandalay. Not a validated forecast.