Method & Data
How the observatory collects, reads, checks and traces what is said about the Sahel's conflicts; how reliable each step is; what the studies cannot claim; and the data behind every page.
Jump to: Method · Reliability · What we cannot claim · Data · Figures · Changes log
Method
Each step links to its source document. The documents are the project's internal design notes, with operational sections removed (rendered at build time).
1. Collect
We archive public Telegram channels and a set of regional news websites in French, Arabic, Russian and English. Every collected post keeps its source, the time it was collected and its earlier versions, so a figure can always be traced back to the posts behind it.
Channel discovery. The channel list started from a seed registry built from published attribution research: government and research reports that name Telegram channels and websites active in Mali, Burkina Faso and Niger, sorted into strata (attributed, institutional, alternative-media and mainstream baseline). Further candidate channels were found by following forwards out of the collected channels, then checked by hand for relevance to the Sahel; candidates that matched only on a keyword (for example "Niger" for the Niger Delta) were rejected. The roster grew in reviewed waves to 78 channels in September 2026. The per-channel list is held back until the principal investigator confirms channel naming.
2. Read
Posts are read in their own language. Stories are grouped into clusters, and a machine labeller assigns three labels to each story: the actor alignment, the function of the post and its theme. The labels are tested against human coders who follow a written coding guide; the reliability figures are reported below, each kept separate.
- Coding guide v2
- Reliability report 1: independent round-1 inter-coder reliability
- Reliability report 2: post-adjudication residual agreement
- Reliability report 3: machine-versus-gold accuracy
Cluster stability test. Every month the story clusters are re-fitted on resampled data and compared with the previous fit (adjusted Rand index, gate 0.75 per language). The August 2026 run gave fr 0.47, ar 0.19, ru 0.30, en 0.21 and the October run was lower; the gate fails and the cause (generation drift versus instability) is under study with the principal investigator. No cluster-based result is published while the gate fails.
3. Check
What the channels say is set beside what happened on the ground. Conflict events come from the Uppsala Conflict Data Program (UCDP); GDELT's machine-coded press reports are a second source, used to show how well each area and incident is covered by the press, never merged into the UCDP record. The map shows both, with a coverage class for every incident.
- UCDP conflict events: ingest and alignment
- GDELT as a second conflict-event source
- Conflict incident map: design
4. Trace
Forwards and copied text show how a story moves between channels, languages and strata, and how fast. The incident study follows sixteen rule-chosen incidents through the channels; the news and conflict study compares attention over time with recorded violence.
Reliability
Three different measurements are reported, and they must never be merged into one "reliability" number. Independent round-1 inter-coder reliability says how well two coders agree when they label the same posts on their own. Post-adjudication residual agreement says how much disagreement was left after the disputes were resolved; it is not an inter-coder reliability figure. Machine-versus-gold accuracy is a third, separate measurement again: how often the machine labeller matches the resolved human labels. All three are Krippendorff's alpha.
Loading…
What we cannot claim
Incident event study
- No cause and effect: the study describes how material about an incident travelled, not what it caused or what caused it.
- The material comes from the 78 Telegram channels and the web sites we collect. It is not "Sahel media" or "the Russian information space" as a whole.
- Views are not persuasion, and Telegram view counts can be inflated; reach means "displayed", not "believed".
- Coverage is estimated only against the candidates the two search routes produced; material neither route surfaced is unmeasured.
- Incident types follow UCDP's coding (e.g. Moura is an army–JNIM battle in UCDP), not how the media framed the events.
- Patterns describe these sixteen rule-chosen incidents; they are not estimates for incidents in general.
- Russian involvement before 2024 is identified through UCDP's source headlines, so incidents whose sources did not name Russian forces are missed.
- Speed is measured in days from the first possible day of the incident, not in hours; incidents with a date range carry that uncertainty.
- Relevance was judged by two AI agents (Codex and Gemini) with human adjudication of their disagreements; the search being validated also relies on AI embeddings, so the check is not fully independent. Agreement between the agents is reported.
- Views multiples compare incident posts with a baseline sample from the 90 days before the window; because the baseline is older, channel growth alone can push a multiple above 1, so a multiple above 1 is not evidence of extra attention.
News and conflict descriptive study
- No causal claim in either direction. Correlations and lead-lag peaks describe co-movement, not influence.
- No claim about the population of Sahel media. The corpus is the registry we chose, added in waves, and shares are shares of that corpus.
- No claim about what unlabeled stories say. Composition covers the labeled variants as of the run date, and the labeled share is reported per year, country and actor.
- No claim about regions where news or events were not geolocated. Country-only UCDP events and unlocated docs are counted Sahel-wide or at country level only, never in a region, and their shares are shown. (amended 2026-10-15 at review: c_tensor carries no country for country-only events, so they count Sahel-wide.)
- No claim about precise timing below a month at regional level, or below a week nationally.
- No claim that a doc mentioning a region is about that region. Geolocation is mention-based, so regional attention means "named the region".
- No comparison with ACLED. It is a frozen reference, not part of this series.
- No claim about intent, coordination, or who is behind a channel. Alignment is a label on stories, not an attribution of actors.
Data
Loading…
Per-channel files are held back until the principal investigator confirms channel naming; see the changes log.
Figures
Every chart on the site has a "download SVG" link; the files carry the run id in their names.
Changes log
Every change to the data shown on this site is recorded here, newest first.
Loading…