Methodology
How we define, count, and aggregate.
This page exists because we believe you should understand exactly what our numbers mean before you rely on them.
01 What is an "incident"?
A user-submitted report. Unverified at the time of submission. The report may be reviewed by an operator or partner agency after submission. The report is the data unit — not a verified crime or confirmed event. Gorebet data represents what community members chose to report, not an objective count of events that occurred.
02 What is a "neighborhood"?
A community group boundary defined by member choice, not an administrative woreda. The two may not align. Woreda-level aggregates in Gorebet use the official sub-city and woreda hierarchy of Addis Ababa city administration and other Ethiopian cities.
03 What is a "collective efficacy proxy"?
The neighborhood reaction rate (incidents reacted to divided by incidents submitted, in a 30-day window) and co-membership count. This is a behavioral proxy derived from Sampson et al. (1997) — not a validated psychometric instrument equivalent to the Collective Efficacy Scale. We do not claim it measures collective efficacy directly.
04 Aggregation threshold
No neighborhood or woreda statistic is published for groups with fewer than a minimum number of reports in the trailing 90-day window. Below this threshold, data is suppressed or shown as a city-wide aggregate only. This threshold is set to protect individual privacy and prevent over-interpretation of sparse data. The specific threshold is published in our data documentation.
05 Street-level precision suppression
Publicly visible maps show neighborhood-level or woreda-level clusters, not individual report pins at street resolution. Sensitive categories — gender-based violence, missing persons, and child-involved incidents — are additionally coarsened to the woreda level regardless of volume. Individual incident locations are available only to verified partner agencies under governance agreements.
06 Bayesian shrinkage
Woredas with few reports have their safety scores pulled toward the city-wide mean, not treated as zero-incident. This adjustment reflects the statistical reality that silence in a low-adoption area is not evidence of safety. Technically: score estimates use a hierarchical model that treats each woreda as a draw from a city-level prior, with the posterior pulled toward that prior as observation count decreases.
07 Report deserts
Areas with implausibly low reporting rates relative to population are flagged with a "coverage limited" indicator on public-facing materials, not treated as low-incident. We use population-weighted expected reporting rates and flag areas where observed rates fall more than two standard deviations below expectation. These areas receive outreach priority, not silence.
08 Temporal decay
More recent incidents carry greater weight in safety scores. Weight follows an exponential decay function: an incident from yesterday carries full weight; an incident from 90 days ago carries approximately 5% of that weight. This reflects the empirical finding that crime patterns shift and historical events become less predictive over time.
09 What we have NOT validated
We have not run a causal study of Gorebet's impact on crime rates. That requires partner-agency ground-truth data and a properly identified study design (pre-registered, with a control group). We have not validated our collective-efficacy proxy against the Sampson Collective Efficacy Scale survey instrument. Our AI classification accuracy has not yet been independently published or audited.
10 Fairness audit commitment
We commit to an annual fairness audit comparing score distributions by: language group of reporter (Amharic versus English), approximate gender where derivable from user profile data, and woreda socioeconomic proxy. The first audit will be published within six months of the platform reaching sufficient scale for meaningful analysis. Fairness audits use disparate impact ratio, calibration-by-group, and equalized odds metrics at minimum.
Questions about our methodology?
Researchers, NGOs, and partner agencies are welcome to request detailed methodology documentation, aggregated data, or a methodology review meeting.
Request data access →