PRAXIS Research Track
PRAXIS EWS
Early Warning System
Econometric research and open-source tooling to detect escalation signals in ACLED conflict data and translate them into protection decisions. Built by someone who works in the places where the escalation happens.
The research
The core question is straightforward: can we predict where violent extremism will escalate before it does? Not with machine learning black boxes or satellite imagery, but with the kind of econometric methods that survive peer review and can be explained to a programme manager in Ouagadougou.
The current study examines whether district-level spikes in kidnapping-for-recruitment (KFR) predict subsequent surges in violent extremist events across the Central Sahel. Using ACLED geocoded conflict data from 2015 to 2023 at the admin-2 level for Burkina Faso, Mali, and Niger, the analysis applies a stacked event study design with two-way fixed effects to isolate the temporal and spatial dynamics of escalation.
Method
Stacked event study with two-way fixed effects (unit and time). The stacking addresses the staggered treatment timing problem identified in recent econometrics literature. Each district-event cohort is estimated separately and then aggregated, avoiding the negative-weighting bias of conventional TWFE estimators.
Key findings
KFR spikes are followed by statistically significant increases in VE events within 3 to 6 months at the district level. The effect persists after controlling for district and month fixed effects, pre-existing conflict trends, and spatial clustering. Spatial spillover is detectable at radii of 50km and 100km from the origin district, with attenuation beyond 150km. Placebo permutation tests (randomised treatment timing) produce null results, supporting a causal interpretation. Five robustness checks confirm stability: alternative bandwidths, excluding capital districts, varying the KFR spike threshold, dropping individual countries, and using Conley standard errors for spatial correlation.
Figure 1: Monthly event study
VEA activity spikes sharply in the month of a KFR spike and remains elevated for 1 to 2 months before returning toward baseline. The effect decays by months +2/+3, suggesting a narrow but actionable early-warning window.
Figure 2: Quarterly aggregation (sensitivity check)
| Bin | β | p-value |
|---|---|---|
| [-3,-2] | ref. | — |
| [-1] | +0.32*** | < 0.001 |
| [0] | +0.87*** | < 0.001 |
| [+1] | +0.53*** | < 0.001 |
| [+2,+3] | +0.50*** | < 0.001 |
At quarterly resolution, the spike-quarter coefficient (β = +0.87) is nearly identical to the monthly estimate (β = +0.83), confirming that the signal is robust to temporal aggregation. Elevated VEA persists into quarters +1 and +2.
Status
Manuscript in preparation. The methodology and preliminary findings form the analytical foundation of the PRAXIS EWS platform.
"I needed to know where violent extremism was going to escalate before it did. Not a forecast from a university model, but something I could hand to a programme manager and say: pay attention to this district next quarter."Emmanuel Nene Odjidja
The platform
The research produces a signal. The platform turns that signal into something a protection officer, a programme manager, or a humanitarian coordinator can act on. Three layers, each with a clear function.
Layer 1: Predict
Automated detection of escalation signals from ACLED API data. The stacked event study model runs on rolling windows, producing a risk score at admin-2 level. The input is publicly available conflict event data. The output is a probability that a given district will experience a VE surge in the next 3 to 6 months, based on observed KFR dynamics.
Layer 2: Alert
Signal translation into structured alerts for defined stakeholder groups. Each alert includes severity classification, geographic targeting to the district level, the underlying data pattern that triggered the signal, and a confidence interval from the model. Alerts are designed to be legible to non-technical users.
Layer 3: Respond
Decision pathways for different users:
- Community protection activation for local civil society actors
- Programme adaptation guidance for P/CVE implementers
- Pre-positioning recommendations for humanitarian agencies
- Threat briefing structure for security coordination bodies
Methodological foundation
The current foundation is classical econometrics: interpretable, peer-reviewable, and explainable to the people who will use the outputs. The stacked event study design was chosen because it handles staggered treatment timing correctly, produces coefficient estimates with confidence intervals, and does not require the kind of training data that would bias the model toward historically over-documented conflicts.
The planned extension adds machine learning on top of the econometric base. XGBoost on multi-source event features, actor-network analysis from ACLED's actor coding, and temporal sequence modelling to capture escalation dynamics that a linear model cannot. The econometric layer remains the interpretive anchor. The ML layer improves prediction accuracy where the data supports it.
Wherever ACLED operates, roughly 230 countries and territories, is the addressable scope. The Central Sahel is the proof of concept because that is where I work and where the data is richest for the KFR-to-VE pathway. It is not the ceiling.
Who built this
This was built by a P/CVE practitioner working in the Sahel. I design and manage evaluations of programmes aimed at preventing violent extremism across Burkina Faso, Mali, and Niger. I have been doing this work for over 12 years, across six countries, and have published 29 peer-reviewed papers on topics ranging from health systems to the climate-conflict nexus. The gap between people who study conflict from university offices and people who work in conflict zones is real, and it shapes what gets built. PRAXIS exists to close that gap. The early warning system comes from the same place as the evaluation tools: someone who needed it, could not find it, and decided to build it.