Framing & context

The date the paperwork arrived

Two dates on every record, and the chart is drawn on the one about the form.

a.k.a. report date · received date · event date versus report date · date of record · notification date · entry date · onset versus report · the date the database sorts by · created_at

Anything assembled out of reports carries at least two dates: the day the thing happened and the day somebody wrote it down. Only the first is about the world. The second is about a form being filled in — and it is the one almost every system sorts, indexes and charts by, because it is the one the system can be certain of. The event date arrives from a human memory and is sometimes blank; the received date is stamped by the machine, is never missing, and is therefore the default. So the axis quietly stops being time and becomes administrative time, in which nothing happens until it is reported. Every change in how eagerly people report — a broadcast, a lawsuit, a letter from the manufacturer, a new online form, a staff member who chases things up, a hotline number read out on air — lands on the chart as a change in the phenomenon, drawn in the right unit, at the right height, in the right year. Two neighbours sit close enough to be worth holding apart, and the difference in each case is where the damage lands and whether waiting repairs it. The unfinished period also comes from reporting lag, but it bites only the final bin, because that is the one still filling up; wait a quarter and it heals itself. The immature cohort buckets events by the cohort they belong to rather than by a date at all, so its shortfall reaches back years along the whole right-hand end. This one puts the distortion anywhere in the series, usually somewhere in the middle, and it is permanent: a spike drawn on report dates will still be in exactly that year a decade from now, because the records really were filed then. Nothing is incomplete, nothing is provisional, and nothing is waiting to arrive.

How to spot it

Count the date columns. A register of reports almost always has two or more — occurred and received, onset and notified, incident and filed, transaction and posting — and the chart will have been built on whichever one the query defaulted to.
A spike with near-vertical sides in a series of human events. Things that happen in the world have onsets spread across the population; things that get reported have a news cycle, and a news cycle has an edge.
Go inside the spike and plot it by month. A rise over three or four months with a single peak and a fall back to the old baseline is the shape of attention. A production line, an epidemic or a wear-out failure does not have that shape.
Ask what changed about reporting in that window rather than about the world: a broadcast, a recall notice, a class action, a compensation scheme, a simplified web form, a regulator writing to everybody affected, a campaign telling people how to complain.
Draw both series on one axis. If the two peaks sit in different places, the gap between them is the lag, and the chart was measuring the lag. If the report-date line is spikier than the event-date line, the extra spikiness is entirely reporting behaviour.
Compute report date minus event date and look at the distribution rather than the average. A long right tail means records are still arriving years later, and any bucket built on the report date is a mixture of every year that came before it.
The field names are the confession: created_at, received_date, date_reported, entry_date, logged, filed, notified, posted. None of those words is a synonym for happened.
Watch for the event date being self-reported, because it usually is. That makes it the better axis and not a clean one — it is remembered, rounded to a month, and only exists for the people who eventually filed.

The fix

Plot the date the event happened, and say in the caption which date you used, the way you would print a unit. In most registers this is a one-word change to a query and it is the whole remedy: the field is already on the record, it was simply not the one the tool reached for. Where the event date is missing on some rows, say how many and show them rather than dropping them silently — a count of records set aside, printed beside the chart, is worth more than a tidier picture. Then draw what you know about the looking. Put the attention events on the plot as labelled rules — the broadcast, the filing, the letter, the day the form went online — so a reader can see which part of the series is the phenomenon and which part is the publicity; a spike that lines up with one of them to the month has explained itself, and a spike that lines up with none of them is a finding. Be honest that the better axis is not a clean one, because it rarely is: a series rebuilt on event dates still contains only the events somebody eventually reported, so its own tail is shaped by who kept reporting, and moving the axis exposes the artefact rather than removing it. Saying that out loud is the difference between a corrected chart and a chart that has moved its problem somewhere harder to see. Where decisions depend on the recent end, the lag is a thing you can measure and correct for: fit the distribution of report-minus-event from the years that have settled, inflate the unsettled ones by it, and draw the correction as an interval rather than a line — actuaries have called this loss development for a century and epidemiologists call it nowcasting. And keep the words straight on the page, because it is the cheapest safeguard there is: reported, recorded, received, logged and filed are five different words from happened, and a title that says “complaints received” instead of “failures” makes the whole mistake impossible to make.

In the gallery

Dark navy slide headed “Field quality review · monthly pack · demonstration slide” on the left and “2007 model year · 459 consumer complaints · NHTSA component: structure” on the right. The headline reads “The dashboards started failing in 2014.” A bordered gold chip beneath it reads “Demonstration · drawn by us from the federal complaint database”. The standfirst says every complaint the regulator holds for this model year under this component code is on the slide — 459 of them, none dropped, one make, one model, one model year, one component, no sampling and no smoothing; that the count axis is linear and starts at zero with nothing truncated, broken, clipped or logged; and that eight quiet years were followed by 248 complaints in 2014 alone, 6.9 times the year before and 54% of everything ever filed under this code. The main panel, “Complaints about the 2007 dashboard, by year”, subtitled “Count of complaints on the federal register · component ‘structure’ · linear count axis starting at zero · every complaint held for this model year is in the chart, none dropped, nothing smoothed or rescaled”, is a column chart on a count axis running 0 to 250 with gridlines every 50. Twenty-one columns run from 2006 to 2026 reading 3, 6, 8, 4, 15, 6, 15, 36, 248, 35, 6, 18, 16, 15, 9, 5, 6, 2, 4, 1 and 1. All are blue except the 2014 column, which is gold and towers over everything else, labelled 248, with an annotation beside it on two lines reading “248 complaints in one year” and “against 36 the year before”; the final 2026 column is hatched to mark a part-year. Every other column is a stub along the bottom. Three panels sit below. “Inside the spike”, subtitled “Complaints per month, 2014”, is twelve columns reading 3, 5, 5, 7, 10, 18, 29, 66, 41, 36, 17 and 11, with August picked out in gold at 66. “Share arriving in 2014”, subtitled “Percent of all 459, zero-based axis”, is two horizontal bars — 2014 at 54.0% in gold and all other years at 46.0% in blue — under a line reading “One year holds more than half the register”. “The review, tabulated”, subtitled “Counts read straight off the register”, is a monospace list giving 2012 complaints 15, 2013 complaints 36, 2014 complaints 248, year on year ×6.9, peak month Aug · 66, then a rule and all years all complaints 459, share filed in 2014 54.0%, complaints dropped 0, status ESCALATE, closing with a note that the count rose almost sevenfold in a year and fell back just as fast and is treated as a 2014 build-quality event and escalated to the supplier review. Five cards run down the right. “The finding, in one figure” prints a large gold ×6.9 beside “complaints in 2014 against 2013 — 248 against 36, the sharpest rise anywhere in the series”. “Read-out for the review” says a step change this large in a single year is a manufacturing event rather than drift and recommends pulling 2014 build records for the dashboard line and opening a supplier review. “The set as read” tabulates make and model Toyota Camry, model year 2007, component STRUCTURE, complaints 459, span of register 2006–2026 with an asterisk, mentioning “dash” 377, mentioning “dashboard” 323, sampling none. “Why nothing here is bent” notes one population, one unit, one component, a linear count axis from zero, every year drawn at the same width, every complaint in the chart and the 2026 column marked as a part-year, ending “Whatever is wrong with this slide, it is not the drawing.” “How the years were counted” explains that each complaint is placed in the year its record was received by the regulator — the date-filed field, grouped by calendar year — and that every chart in the pack is built the same way. Footer chips read “Every complaint included”, “One model year, one component”, “Linear axis from zero”, “Nothing smoothed” and “Nothing dropped”, above a footnote naming the source, giving the counts of records containing “dash” and “dashboard”, listing the yearly and monthly counts, noting that the register was pulled on 25 September 2026 so the 2026 column is a hatched part-year, and stating that the review, the status line and the recommendation are invented while every figure behind them is the regulator’s own.

The 248 complaints of 2014 that describe eight years of dashboards

The date the paperwork arrived The self-selected sample The count that measures the looking