The Newest Edge of a Crime Map Isn’t Less Crime — It’s Missing Records

A dashboard can be refreshed this morning and still be wrong about last week. “Refreshed” means the data was re-downloaded — not that it describes everything that happened. In this extract, 44.2% of records were published within 24 hours of the recorded offense start. 67.7% made it within 72 hours, 83.6% within a week, and 95.6% within 30 days. The last few percent trickle in over months.

Which means the newest stretch of any occurrence-date chart is structurally incomplete. That reassuring dip at the right edge of the graph? Part of it is records that haven’t landed yet. The recent period isn’t a smaller version of history. It’s a different, unfinished product.

The shape of the delay is a long tail, not a wall. About one in five records arrives within six hours and just under a quarter more within a day — so a little over two in five are in by the next morning. Another 23.5% land in days one to three, 15.9% in days three to seven, and the remaining 16.4% spread across everything from a week to more than a month. There is no single filing deadline; there is a distribution, and its slow end is long.

The typical record isn’t the problem. The median lag is 32.7 hours — annoying, but workable. The problem is the tail. The 90th percentile is 292.5 hours, about twelve days. The 95th is 622.6 hours: nearly four weeks.

The distribution also has an offense shape. Motor-vehicle theft’s median is 16.7 hours, burglary’s is 32.5, and all-other-larceny runs to 41.1, while arson surfaces in about 3. The slow categories are exactly the ones most likely to still be missing from a fresh map, which is why one global buffer under-protects some offense types and over-protects others. The companion survival study takes that split apart offense by offense.

If you suspect the exact-midnight timestamps inflate the tail, they do — a little. Excluding them, the median drops to 28.4 hours, the 90th percentile to 250.2 hours, and the share taking more than a week falls from 16.4% to 14.5%. Better. Still slow.

The chain that ends in publication starts upstream, and the lag splits geographically. Median offense-to-report intervals run from 9.7 hours around Roxhill to 41.9 in Capitol Hill, 78.6 where the block label is unknown, and a 331.4-hour mean in Queen Anne. A single global buffer is therefore too coarse: the same 24-hour cutoff that is generous in one area is lenient in another.

That variation is what recording-practice research predicts. Recording practices vary by place and time within a single agency, and open-data assembly work documents how revisions, re-geocoding, and late entries ripple through counts — the same forces lengthening the tail measured here.

It also points at a larger, invisible delay. A record can only show up late if it was reported at all; the dark figure of unreported crime means a live view cannot see the offense that never entered the system, the one reclassified away, or the one still queued. Publication lag is the delay you can measure; the detection layer upstream is the one you cannot. Label the first and stay modest about the second.