How the data is built
162,315 validated incidents from 1991 to today — sourced from 3 violence-type categories, 57 motive codes, and 9 cross-cutting themes. This page documents how those numbers are made.
Step 1
The research desk monitors 24 Bangladeshi newspapers, news agencies, and
broadcasters daily. Reported incidents of violence — including those
non-violent acts that meet the BPO threshold of organised conflict —
are entered with the source URL preserved as eventurl.
Step 2
Each incident is tagged against the BPO taxonomy: a violence type (and a secondary type if compound), motive incident codes, the perpetrator and target groups, primary and secondary actors, a cross-cutting theme where applicable, and geocoded to the upazilla level when location data is available.
Step 3
Every coded incident passes through a two-stage editorial check
before it appears in any public lens — only rows with isvalidate=1
and ispublish=1
flow through to the Observatory. Drafts and rejected rows are
never exposed.
Transparency
Computed live across 162,315 validated incidents. ≥75% · 40–75% · <40%
| Field | Populated | Fill rate | % |
|---|---|---|---|
| eventdate · Date of incident | 162,315 / 162,315 | | 100.0% |
| lat / lon · Geocoordinates | 162,304 / 162,315 | | 100.0% |
| division · Division (8) | 162,315 / 162,315 | | 100.0% |
| district · District (64) | 162,315 / 162,315 | | 100.0% |
| upazilla · Upazilla (564) | 162,315 / 162,315 | | 100.0% |
| viotypeone · Primary violence type | 162,315 / 162,315 | | 100.0% |
| viotypetwo · Secondary violence type | 162,315 / 162,315 | | 100.0% |
| perpgroup · Perpetrator group | 162,315 / 162,315 | | 100.0% |
| targgroup · Target group | 162,315 / 162,315 | | 100.0% |
| actorone · Primary actor | 162,315 / 162,315 | | 100.0% |
| actortwo · Secondary actor | 162,315 / 162,315 | | 100.0% |
| mtvincidentone · Primary motive | 162,315 / 162,315 | | 100.0% |
| mtvincidenttwo · Secondary motive | 162,315 / 162,315 | | 100.0% |
| crosscutting · Cross-cutting theme | 162,315 / 162,315 | | 100.0% |
| eventdesc · Event description | 162,310 / 162,315 | | 100.0% |
| eventurl · Source article URL | 161,077 / 162,315 | | 99.2% |
| kiltotal · Killed count | 162,315 / 162,315 | | 100.0% |
| injtotal · Injured count | 162,315 / 162,315 | | 100.0% |
| arresttotal · Arrested count | 162,315 / 162,315 | | 100.0% |
A low fill rate doesn't mean the field is unreliable — it means it isn't recorded for every incident, usually because the source article didn't supply that detail.
Taxonomy
19 types across 3 categories.
Category-1
Category-3
Indiscriminatory use of force by law enforcing agencies. Indiscriminatory
57 codes across 10 categories.
Armed conflict
Crime
Custom-Single
Economic competition and natural resources
Governance
Identity and social tensions
International terrorism
Law & Order
Political
Sexual and gender-based violence
9 themes that overlay the violence-type taxonomy.
What to keep in mind
Roughly 31% of all incidents involve police as a primary actor. This reflects two things: police-involved events are reliably reported in the press (more so than civil disputes), and police are the most common single actor in the country's day-to-day security landscape. Filter on actor type if you want to see the non-police picture.
Coverage before 2012 is sparse — the dataset starts in 1991 but 99% of records sit in 2012 onward, when the research desk was monitoring a stable daily news corpus. Year-over-year comparisons spanning 2011/2012 should be read as a coverage change, not a real surge in violence.
kiltotal,
injtotal,
and arresttotal
are populated only when the source article supplied a numeric
count. Aggregated totals are therefore lower bounds, not exact
ground truth.
Lat/lon coordinates are populated for ~99.9% of incidents, but they snap to the upazilla centroid where the source didn't specify a finer location. Cluster and heat maps reflect upazilla density rather than street-level precision.
