A week of escalation,
visible in the composition.
When two or more distinct ACLED sub-event types converge in the same district-week, the district sits in a persistently high-conflict cluster, averaging 8.8 times the Sahel-wide clean baseline attack rate, exclusive of triggering types. The signal identifies where. It does not specify when.
Abstract
Early-warning systems for political violence typically score districts on the intensity of a single sub-event type. We propose a complementary signal: the compositional diversity of distinct sub-event types in the same district-week. Across approximately 1.17 million district-weeks in 13 Sahelian and West African countries (1 January 2010 to 26 March 2025), districts where KAFD and IED co-occur in the same week sit in a persistently high-conflict cluster, averaging 8.8 times the Sahel-wide clean baseline attack rate, exclusive of triggering types (N = 292, p < 0.001).
The broader threshold dose-response is monotonic: district-weeks with two or more distinct types carry a 2.3× baseline multiplier at t+1, rising to 7.5× at five or more. The pattern replicates across Central Sahel, Lake Chad Basin, and Coastal West Africa (2.6×, 8.3×, 10.2× respectively) and survives Bonferroni correction at p = 0.0011 across 44 specifications. We treat the composite signal as a district identifier, not a timing forecast. The temporal escalation coefficient (1.2×) is reported as provisional pending a panel regression with district fixed effects. An MIT-licensed Python pipeline accompanies the working paper for independent replication.
Table 1 — Threshold dose-response
| Specification | Multiplier (t+1) | N district-weeks | Bonferroni p | Intensity |
|---|---|---|---|---|
| Composite threshold: 2 or more distinct types | 2.3× | 13,984 | < 0.001 |
|
| Composite threshold: 3 or more distinct types | 4.0× | 4,312 | < 0.001 |
|
| Composite threshold: 4 or more distinct types | 5.7× | 1,477 | < 0.001 |
|
| Composite threshold: 5 or more distinct types | 7.5× | 482 | < 0.001 |
|
| KAFD+IED pair · district identification | 8.8× | 292 | < 0.001 |
|
Monotonic dose-response across composite thresholds, plus the KAFD+IED pair-specific district-identification result. All comparisons survive Bonferroni correction at p = 0.0011. Specific pair-type combinations beyond KAFD+IED are available in the authenticated dashboard; see the methods note for full operational definitions and exclusive-outcome corrections.
Identification robustness — threshold sweep
| N | Flagged | TP | FP | FN | TN | Precision | Recall | F1 | Lift |
|---|
Lift compares flagged districts' non-flagged-week attack rate to the Sahel-wide clean baseline (seed-42 sample of 8,000 district-weeks). Outcome excludes KAFD and IED sub-event types (N4 decontamination). Districts whose every week is a signal week are excluded from the confusion matrix and counted separately. Bonferroni-corrected threshold p < 0.00102 across 49 specifications. Tests robustness of the where claim — not a timing forecast.
Methods
1,510 admin-2 districts across 13 Sahelian and West African countries. Weekly observations, 1 January 2010 to 26 March 2025. Approximately 1.17 million district-weeks and 100,184 ACLED events with 201,641 reported fatalities.
Count of distinct ACLED sub-event types per district-week. Composite signal defined as two or more co-occurring types. Lead-lag mean comparison across t−4 to t+5, two-sided Mann-Whitney U, referenced to a clean baseline of non-composite district-weeks.
Outcome categories: Armed clash, Attack, Suicide bomb, Shelling and artillery. KAFD and IED removed from the outcome set to eliminate predictor-outcome overlap. 44 comparisons across threshold, region, era, and pair dimensions. Bonferroni-corrected at p = 0.0011.
The exclusive metrics (8.8× district identification, 4.0× dose-response at three or more types, 1.2× temporal escalation) remove KAFD and IED from the outcome count. This eliminates mechanical double-counting of triggering events inside the outcome variable and produces the methodologically conservative headline. The inclusive metrics are retained in the manuscript appendix.
Citation & related work
- [1] Raleigh, C., Linke, R., Hegre, H. & Karlsen, J. (2010). Introducing ACLED: An Armed Conflict Location and Event Dataset. Journal of Peace Research, 47(5), 651–660.
- [2] Odjidja, E. N. (2026). Methods Note: Composite co-occurrence detection for conflict escalation in the Sahel. PRAXIS Working Paper 01, v0.1. Read online.
- [3] ACLED (2026). Armed Conflict Location & Event Data Project. Data export used: 2026-03-26. acleddata.com.
- [4] Bonferroni, C. E. (1936). Teoria statistica delle classi e calcolo delle probabilita. Pubblicazioni del R Istituto Superiore di Scienze Economiche e Commerciali di Firenze.
- [5] Mann, H. B. & Whitney, D. R. (1947). On a test of whether one of two random variables is stochastically larger than the other. Annals of Mathematical Statistics, 18(1), 50–60.