0–<15 minutes
Police-reported personal-injury road collisions between the archived alert time inclusive and 15 minutes later exclusive. This remains the primary endpoint and is not displaced by a more striking secondary window.
Methodology
Designed to find a null result as readily as a spike.
The analysis is deliberately pre-specified around a narrow primary endpoint and uses matched controls, placebo timestamps and sensitivity tests to reduce retrospective pattern hunting.
0–<5 / 30 / 60
These secondary windows were defined as sensitivity analyses to test whether any apparent effect is extremely short-lived or persists. They are exploratory/sensitivity results, not independent confirmatory endpoints, and are interpreted with explicit awareness of multiple-window testing.
Matched + placebo
Same weekday and clock time around the alert date, known-alert exclusions and random same-weekday pseudo-alert timestamps. Empirical p-values arise from the matched/placebo reference distribution rather than a simple parametric significance test.
Multiplicity and p-value interpretation
Testing several pre-specified secondary windows increases the opportunity for an apparently unusual value to occur somewhere by chance. Secondary-window p-values are therefore not promoted to confirmatory findings and are not interpreted as though each window were the sole test performed. The public output must disclose the reference-set or resampling count used to generate an empirical p-value whenever that count is available in the validated analysis bundle.
Replication discipline
A secondary-window pattern in one alert is not described as replicated unless the same window, outcome definition and analysis logic are applied to another alert. Where historical reporting is coarser or a comparison is not fully like-for-like, the Observatory will say so rather than imply a clean replication.
Interpretation guardrails
Advance publicity, traffic volume, weather, road conditions, geography and non-simultaneous handset receipt can all influence results. Statistical association is therefore not presented as proof that the Emergency Alert caused an incident.
Safeguarding
Domestic-abuse analysis requires suitably anonymised timestamped aggregate data. Month-level street-crime data is not treated as evidence for minute-level event effects.