Composite co-occurrence detection for conflict escalation in the Sahel

Section 01Study design

The analysis addresses a single empirical question: do signals observable in open-source conflict-event data predict next-week violence in fragile settings? The unit of analysis is the admin2 district-week: a row of data represents one second-level administrative district (e.g. a province, department, or cercle) observed over one calendar week. The outcome of interest is the count of violent-conflict attacks in that district during the following week (t+1).

We test two classes of predictor. Single-type thresholds: the predictor is a binary indicator that at least one event of a given sub-event type (e.g. KAFD, IED) occurred in the district that week. Composite signals: the predictor is a binary indicator that at least two distinct conflict sub-event types co-occurred in the district that week. The composite signal is the central hypothesis of this study.

The comparison is descriptive: for each predictor, we report the ratio of mean next-week attacks between treatment district-weeks (predictor = 1) and control district-weeks (predictor = 0), with 95% confidence intervals and a non-parametric Mann-Whitney U test statistic. The formal econometric analysis, which applies fixed-effects and additional controls, is reported separately in the forthcoming manuscript. This methods note documents the descriptive evidence.

Section 02Data

Source

All event data is drawn from the Armed Conflict Location & Event Data Project (ACLED). Events are coded by ACLED analysts under a published codebook; we rely on ACLED’s own classification of event_type and sub_event_type without re-coding. Access is under ACLED’s standard research licence; raw event data is not redistributed in the GitHub repository.

Geographic scope

Thirteen countries are included, selected for ACLED coverage density, geographic contiguity, and policy relevance to counter-violent-extremism work:

• Central Sahel: Burkina Faso, Mali, Niger
• Lake Chad Basin: Nigeria, Cameroon, Chad
• Coastal West Africa: Ghana, Côte d'Ivoire, Togo, Benin, Senegal, Guinea
• Peripheral: Mauritania

Admin2 boundaries follow the standard ACLED geocoding. The total number of distinct admin2 districts appearing at least once in the panel is 1,510.

Time window

Events with event_date between 1 January 2010 and 26 March 2025 are included. The window spans just over 15 calendar years. The total number of events retained is 100,184, with a combined 201,641 reported fatalities.

Event types

The composite predictor uses the eight ACLED sub-event types listed in bold below. Remaining sub-event types (protests, strategic developments, etc.) are not used to construct the composite signal but do contribute to outcome counts where classified as violent.

Counts derived directly from the ACLED export dated 2026-03-26 (SHA pending). “Other sub-event types” aggregates 12 remaining categories including arrests, protest-with-intervention, agreements, shelling/artillery, disrupted weapons use, excessive force against protesters, territory transfers other than overtake, change-to-group, non-violent territory transfer, grenade, HQ established, and ACLED “Other”. The composite predictor operates only on the eight bolded categories.

Section 03Variable definitions

Composite signal C_d,t

For district d and week t, the composite signal is defined as:

# Binary composite indicator
C_d,t = 1  if |{e: e in {KAFD, IED, Clash, Attack, Looting,
                  Sexual_violence, Territory, Suicide_bomb}
              and count(e, d, t) >= 1}| >= 2
       else 0

The composite signal equals 1 if two or more distinct sub-event types each had at least one event in the district that week. Repeated events of the same type do not contribute. The threshold of 2 is the primary specification. Robustness to thresholds of 3, 4, and 5 is reported in Section 5.

Next-week attack rate Y_d,t+1

The outcome is the count of violent events (all eight analysis categories, summed) in district d during week t+1, divided by the number of district-weeks at risk in the panel. This yields a per-district-week mean, not a per-capita rate; exposure is held constant by design.

Baseline Y_0

The baseline attack rate is Y_0 = 0.241 attacks per district-week, roughly one attack every four weeks in a typical district. All multipliers reported in the interactive visualisation reference this baseline. The construction (which district-weeks are eligible, and whether recent signal history is excluded) is documented in the GitHub repository.

Exclusive signal

For single-type tests, the exclusive signal restricts treatment district-weeks to those in which only the indexed event type occurred (all other seven types had zero events). This isolates the standalone predictive power of each type. Without this restriction, apparent single-type signals are confounded by concurrent multi-type occurrences.

Section 04Statistical methods

Primary test

For each predictor, we compute the ratio of means R = E[Y_d,t+1 | predictor=1] / E[Y_d,t+1 | predictor=0]. Confidence intervals are reported alongside each multiplier in the interactive visualisation on the research page (exact bootstrap specification to be documented in the GitHub repository). We additionally compute a non-parametric Mann-Whitney U statistic comparing the distribution of Y_d,t+1 across the two groups. The Mann-Whitney test makes no distributional assumptions and is robust to the highly skewed, zero-inflated outcome distribution characteristic of conflict-event data.

Multiple-comparison correction

We apply a Bonferroni correction at family-wise α = 0.05 across the 44 distinct tests in the primary and sensitivity analyses (eight exclusive single-type tests, four threshold tests, three regional tests, three period tests, four pair-type tests, plus related descriptive comparisons). The per-test significance threshold is therefore p = 0.05 / 44 = 0.00114. Results reported as “p < 0.001” survive this correction.

Temporal-ordering test

For the specific claim that kidnapping (KAFD) spikes precede violent-extremism attacks within a calendar month, we implement a within-month Wilcoxon signed-rank test comparing the timing of KAFD events against VE-classified attacks in the same district-month. The test returns p = 0.68, indicating no detectable lead of KAFD over VE attacks at the within-month resolution. This is a null result and is a central motivation for the composite-signal approach.

Section 05Results

Single-type exclusive signals (null)

Table 02 reports the exclusive signal test for each of the eight analysis event types. Each row compares next-week attacks in district-weeks where only the indexed type occurred against the clean Sahel baseline. All exclusive multipliers fall between 0.6x and 1.3x. No single event type predicts next-week escalation when concurrent attacks are excluded.

“Exclusive” restricts to district-weeks with only that event type. “Concurrent” restricts to district-weeks with that type and at least one other type. The gap between columns is the evidence for the composite-signal hypothesis.

Composite threshold dose-response

Table 03 reports the main result. As the number of co-occurring event types rises, the next-week attack multiplier rises monotonically. At 2 types the multiplier is 2.3x baseline. At 5 or more types it is 7.5x. Full lead-lag curves (t-4 through t+5) and confidence intervals for each threshold are plotted in the interactive visualisation on the research page.

Intermediate-threshold multipliers and 95% CIs plotted in full on the research page interactive visualisation; numerical values pending extraction from the GitHub repository. The 2.3x (N=13,984) and 7.5x (N=482) endpoints anchor the dose-response. All thresholds yield p < 0.001 against the clean baseline.

Regional replication

Table 04 reports the composite signal (2+ types) within three regional sub-samples. The signal replicates in every region with statistically significant positive multipliers. The Lake Chad Basin and Coastal West Africa produce larger multipliers than Central Sahel despite smaller sample sizes, suggesting the mechanism is not specific to any one conflict ecology.

Coastal West Africa sample size is small (N=11 composite district-weeks in the primary window) and the estimate should be treated as indicative, not confirmatory. Coastal frontier results are treated separately in Section 6 under the coastal-discovery analysis.

Temporal stability

The sample splits into three time periods (2016 to 2019, 2020 to 2022, 2023 to 2025) in the interactive visualisation. Absolute baseline violence has risen across the window, consistent with documented Sahel escalation. The relative signal strength (the incremental spike at t = 0 above the pre-signal corridor) is stable across periods. Full multipliers and district-week counts by period are available in the interactive chart on the research page under the “Period” sensitivity tab. Numerical values for inclusion in this table are pending extraction from the GitHub repository.

Combination specificity (KAFD+IED)

Table 06 reports the headline combination-specific result. The KAFD+IED pair produces 5.5x next-week attacks and 4.6x next-week fatalities (N=292 district-weeks, p < 0.001). Districts where both KAFD and IED have co-occurred in any week sit persistently at 8.8x the clean Sahel baseline. The elevation holds even when triggering event types are excluded from the outcome count, which means those districts are already elevated on non-triggering violence.

KAFD+IED multipliers confirmed against the interactive visualisation. Other pair-type multipliers pending extraction from the GitHub repository. Sample sizes for non-KAFD+IED pairs are taken from the viz data constants. The viz reports that all pair types produce elevated t+1 rates with consistent within-district incremental spikes at t=0, supporting a convergence-driven (rather than pair-specific) mechanism.

Country-level scope

Table 07 shows the full 13-country panel with total event counts from the ACLED export. Three Coastal West African countries (Ghana, Côte d'Ivoire, Benin, Togo) and the broader Coastal zone show systematically higher composite ratios, consistent with the coastal-discovery finding in Section 6.

Event totals derived directly from the ACLED export (2026-03-26 vintage). KAFD/IED and composite-ratio values reported here come from the viz data constants. Guinea, Mauritania and Senegal were excluded from the decomposition in the visualisation and are therefore missing disaggregated values; updating these is pending. Composite ratio is defined in the GitHub repository (operationally: ratio of composite district-weeks to baseline district-weeks within country).

Section 06Robustness

Coastal frontier discovery

Table 08 reports a discovery finding outside the primary hypothesis space. In four coastal West African countries (Togo, Benin, Ghana, Côte d'Ivoire), IED events during 2023 to 2025 predict escalating violence over the next three weeks. Multipliers rise from 5.2x at t+1 to 7.3x at t+3. The result draws on 28 of 38 district-week cases in the period. Sample size is small, so treat this as a lead for follow-up, not a confirmatory result.

Actor-type decomposition

The signal strength correlates inversely with actor composite-frequency. High-frequency composite-capable actors (communal, Ambazonian) produce strong spatial effects (“district identification”) but weaker temporal escalation; low-frequency composite-capable actors (ISWAP, JNIM, IS Sahel) produce weaker spatial effects but stronger temporal escalation. This inverse pattern is consistent with the convergence-as-mechanism interpretation: when unusual-for-the-actor capabilities converge, the next-week signal is strongest.

Exclusive-outcome correction

The temporal multipliers in the main interactive visualisation use an inclusive outcome (all next-week violence, including the triggering types). This inflates absolute values relative to an exclusive-outcome specification that removes the triggering types from the outcome count. Relative ordering is preserved under both specifications; the dose-response pattern and regional replication hold. A fully exclusive-outcome table is available in the GitHub repository.

Section 07Limitations

• No causal identification. Multipliers are associations, not causal effects. District-weeks are not randomly assigned to composite-signal status. Confounders (ongoing campaigns, election cycles, seasonal factors) may co-determine both the signal and next-week violence. We report ratios, not treatment effects.
• Dependence on ACLED coding. Event classification rests on ACLED’s analyst pipeline. Miscoding at the sub-event-type level would bias composite-signal construction. The effect is likely conservative: misclassified events typically remain within the violence category, so composite signals are under- rather than over-detected.
• Attribution uncertainty in frontier zones. Coastal West Africa and other low-density contexts have fewer cross-sources for event verification. Actor attribution is noisier. Results from these zones should be treated as directional rather than precise.
• Sahel proof-of-concept. The 13-country window is not globally representative. Generalisability of the method to East Africa, Central Asia, or Latin America is an open empirical question. The method is portable wherever ACLED coding is applied.
• Exclusive-signal test removes sample. Restricting to district-weeks with a single event type reduces statistical power relative to unrestricted tests. The null results on exclusive tests are therefore interpreted conservatively.
• No individual-level causation. The unit of analysis is the district-week. No inference is made about individual attackers, victims, or events.

Section 08Replication

Analysis code is available at github.com/emmaodjidja-sys/praxis under the MIT Licence. Raw ACLED event data is not redistributed; users must obtain it under ACLED’s standard licence at acleddata.com.

The source dataset used for the results reported here is the ACLED export dated 2026-03-26, covering 13 Sahel countries from 2010-01-01 to 2025-03-26, 100,184 events, 201,641 fatalities. A specific replication walkthrough (setup, dependencies, expected runtime, hash-pinned intermediate outputs) is being prepared alongside this working paper.

Readers who wish to reproduce specific tables should consult the repository README; where a specific command invocation is not yet documented, please open an issue or contact the corresponding author.

Section 09Ethics & data governance

The analysis is descriptive and observational. No human-subjects intervention is performed. No individual-level identifying data is used: ACLED events are aggregated at the admin2-week resolution before any analysis, and no individual perpetrator, victim, or target is identified in outputs. Ethnic labels are deliberately excluded from all PRAXIS outputs, including this document and the TREMOR early-warning system, despite their presence in some ACLED actor fields. This reflects an operational judgement that ethnic targeting outputs carry disproportionate dual-use risk.

ACLED data is used under the project’s standard research licence. The licence prohibits redistribution of the raw data; the GitHub repository provides scripts that transform a locally-held ACLED export into the analysis panel.

Section 10Suggested citation

@techreport{odjidja2026praxis,
  author      = {Odjidja, Emmanuel Nene},
  title       = {Methods Note: Composite Co-occurrence Detection
                 for Conflict Escalation in the Sahel},
  institution = {PRAXIS},
  type        = {Working Paper},
  number      = {01},
  year        = {2026},
  month       = {4},
  version     = {0.1},
  url         = {https://www.emmanuelneneodjidja.org/praxis/research/methods/}
}

Plain-text: Odjidja, E.N. (2026). Methods Note: Composite co-occurrence detection for conflict escalation in the Sahel. PRAXIS Working Paper 01, v0.1. Available at https://www.emmanuelneneodjidja.org/praxis/research/methods/

Section 11Version history

Subsequent revisions will be tracked here. Readers may subscribe to notifications.

Section 12Author & contact

Emmanuel Nene Odjidja. Principal Investigator, Violence Prevention Econometrics. 12+ years of programme-evaluation experience across Burkina Faso, Mali, Niger, South Sudan, and Burundi. 29 peer-reviewed publications spanning health systems, programme evaluation, and the climate-conflict nexus. Section Editor, Journal of MultiDisciplinary Evaluation.

Correspondence: message the author. Public profile: emmanuelneneodjidja.org. Publications: Google Scholar.