Measles and the vaccine

Fifty rows, eighty-one columns, and one vertical discontinuity in 1963. This is the form Tynan DeBold and Dov Friedman published in the Wall Street Journal in 2015, and it is the strongest argument a heat map has ever made: the colour is an incidence rate, and the argument is that it changes all the way down the chart at once.

The chart

Point at a cell and the strip above names the state, the year and the rate. Nothing in the drawing is doing rhetorical work — no annotation arrows, no highlighted states. The four markers are dates, and everything else is 3,500 measured numbers. Three of the dates are interventions; the fourth, 1989, is the one moment the record moves the wrong way.

This is a reconstruction, not a reproduction: the record here runs 1921–2001, which is what Project Tycho covers. Our World in Data's version runs to 2022 by splicing CDC data on top, and starts at 1929.

The data

Real, measured, freely licensed. Case counts are Project Tycho's digitisation of the US Nationally Notifiable Disease Surveillance System — 436,932 weekly rows, CC BY 4.0. Populations are decennial census counts from Wikipedia, CC BY-SA 4.0. Both are pulled and reduced by website/scripts/fixtures/measles.mjs, which runs by hand and commits a 28 KB fixture.

The reduction is where all the judgement is, and every rule changes the picture:

State-level rows only. 238,827 of the 436,932 rows are city rows nested inside their state's rows. Summing naively double-counts every state that had a reporting city.
Non-cumulative rows only. A further 84,414 restate a running total.
An absent week is a zero, not a hole — see below, because this one is genuinely counter-intuitive.
Fifty states. Territories are dropped, which also removes a Virgin Islands 1975 figure — 53,009 cases from 8 reported weeks, against a population near 60,000 — that cannot be right.

Why "how complete is this year?" is the wrong question

The obvious quality check is to count how many of a year's 52 weeks a state reported, and throw away the thin ones. Summing a 37-week year next to a 52-week year makes a thin year look like a quiet year, and Texas 1964 reports 37 weeks and 39,353 cases.

That rule is wrong, and it fails in the most damaging possible way. Only 836 of the 113,691 surviving rows carry an explicit zero — a week with no cases is simply absent from the file. So weeks-reported tracks the disease, not the surveillance:

era	median weeks reported	median cases/year
1928–1939	49	3,726
1950–1962	48	4,239
1971–1980	24.5	160
1991–2001	2	4

A ≥40-week completeness rule reads the disease's own disappearance as missing data and deletes the entire post-1990 record — the half of the chart that carries the finding. Applied here it dropped coverage to 47.6% and left nothing at all after 1995.

What works instead is the reporting span: a state is under surveillance from its first reported year to its last, silent weeks inside that span are zeros, and years outside it are no data. Coverage goes to 87.8%, and the 400 remaining holes are real — which is what puts Alaska's and Hawaii's pre-statehood gaps on the chart without special-casing them.

The cost, stated: in the pre-vaccine era a silent week may be a lost report rather than a genuine zero, so those years can undercount. Texas 1964 is at worst ~30% low — comfortably inside one band of a log ramp.

Why fifty rows, when they all say the same thing

A fair objection: every row does the same thing at the same time, so what is the second dimension buying? A single national line would show the same cliff.

The uniformity is the evidence. A national line is one observation, and one observation invites the obvious alternative explanations — sanitation, nutrition, a change in how cases were reported, the natural multi-year cycling of an epidemic disease. Fifty jurisdictions with different climates, demographics, health systems and economies, all breaking at the same vertical line, is a natural experiment repeated fifty times. Comparing each state's mean incidence before the vaccine with its own mean after:

1955–62 → 1966–73, 48 states with both windows
states whose incidence fell	48 / 48
fell by more than 2×	47 / 48
median drop	9.1×
range	1.7× to 34.9×

The rows are not there to show you difference. They are there to show you there isn't any — which is a claim a line chart cannot make.

And then the second era flips it. 73% of the 1989–91 resurgence sits in five states (California, Texas, Ohio, Pennsylvania, Puerto Rico), and you can only see that because the rows are separated. The same chart is doing two opposite jobs in its two halves: proving an effect is universal, then showing that a later failure was not.

The one place the record goes the wrong way

Three of the four markers are interventions. The fourth, 1989, is the counter-example — and it is the most interesting thing on the chart, because it is the moment the trend reverses.

year	reported cases
1988	2,812
1989	10,062
1990	16,261
1991	5,668
1992	1,776
1993	391

A six-fold jump off the 1988 floor, peaking in 1990, gone by 1993. It shows up on the chart as a scatter of blue-green cells reappearing in the pale field, heaviest in California, Texas, Ohio and Pennsylvania.

Two distinct things were happening at once, and it is worth not conflating them:

A coverage hole. The resurgence was concentrated in preschool-aged children in poor urban areas who had never received a first dose. The 1980 school-entry mandates — the third marker — worked so well for school-age children that they masked the fact that under-fives in inner-city neighbourhoods were not being reached. The mandate bit at kindergarten; the outbreak was in the years before it. CDC reports incidence among children under five running some 15 times the 1981–88 median at the 1990 peak, and several times higher again among American Indian, Hispanic and Black children than among non-Hispanic white children.
A small rate of primary vaccine failure. Separately, outbreaks among already-vaccinated school-age children showed one dose was not quite enough — tolerable at high coverage, not tolerable once susceptibles accumulate. That is what prompted the two-dose MMR schedule, not the preschool outbreak.

The near-flatline from 1993 onward is the response working: two doses, plus a concerted push on first-dose coverage in exactly the populations the 1980 mandates had missed.

References. CDC, History of Measles, which puts the episode at over 55,000 cases and 123 deaths; MMWR, Public-Sector Vaccination Efforts in Response to the Resurgence of Measles Among Preschool-Aged Children — United States, 1989–1991; and Atkinson WL et al., "The resurgence of measles in the United States, 1989–1990", Annual Review of Medicine 1992 (PMID 1580601).

These are provisional counts, and they run low

Worth knowing before reading any absolute number off this chart. CDC puts 1989–91 at over 55,000 cases; the weekly reports underlying this chart sum to 31,991 — about 58%.

That is not a cleaning error. Project Tycho digitises the weekly provisional notifiable-disease reports, and those are systematically lower than the final annual figures published after late reports and reconciliation. Every cell here is a provisional count.

It matters for what you can ask the chart. Comparisons within it — this state against that one, this decade against that one — are what the drawing is for and are unaffected by a roughly consistent shortfall. An absolute rate read off a single cell is a floor, not a measurement.

Build it

Two props carry this chart, and both exist because of it.

<HeatMap
  series={series}
  columns={ROWS} // 50 states, one row each
  colors={RAMP} // YlGnBu reversed, 8 stops
  scale="log"
  noData="hatch"
  domain={[0, 3000]}
/>

`scale="log"`, or the chart says nothing

Incidence runs from about 2,900 per 100,000 before the vaccine to under 1 after it. Linear banding over eight colours makes each band 360 wide — so every value below 360 lands in the same band, and that is the whole record after 1965. The chart would be a dark block, a cliff, and then a uniform pale field with no structure in it.

scale="log" bands by equal ratio instead: eight bands over four orders of magnitude, each a factor of ~3.7. Bands are taken on log1p of the offset from the domain floor rather than on log, so a value at the floor is a real band instead of -Infinity — and zero is not an edge case here, it is most of the 1990s.

`noData="hatch"`, or missing reads as zero

400 state-years have no data. Painting nothing leaves the background showing through, and on a ramp whose floor is pale that reads as the bottom of the scale — so a state that had not begun reporting looks exactly like a state reporting zero cases. In the late record, which is full of genuine zeros, those two must not look alike.

noData="hatch" strokes diagonals in the theme's grid colour. It strokes rather than fills on purpose: any fill colour can be read as a value, and the point is to say this is not a value.

Incidence is computed in the page, not baked into the fixture

The fixture ships raw case counts plus decennial census populations, and the page divides:

const p = populationAt(state, year); // geometric between census anchors
return p === null ? null : (n / p) * 100_000;

Populations are interpolated geometrically — a population multiplies rather than adding a fixed number of people a year, and a straight line visibly undershoots mid-decade in fast-growing states. Decennial anchors are enough precisely because the scale is log: each band is a factor of ~3.7, and a few percent of interpolation error cannot move a cell across one.

Options to try

Option	What it does	Reach for it when
`scale="linear"`	Equal-width bands	Never on this data — it is the failure the log scale exists to prevent, and worth looking at once
`noData="blank"`	Holes show the background	A hole means "outside the record" and cannot be confused with a low value
`domain` unpinned	Ramp spans the visible extent	Never with a colour scale you want stable — it re-means every colour as the window moves, silently
Raw counts instead of incidence	Colour carries cases, not rate	You want to see outbreak size; California will dominate every year
`orientation="horizontal"`	States across x, years down y	Rarely — 50 categories across is unreadable, and this is the orientation the form settled on

The chart​

The data​

Why "how complete is this year?" is the wrong question​

Why fifty rows, when they all say the same thing​

The one place the record goes the wrong way​

These are provisional counts, and they run low​

Build it​

scale="log", or the chart says nothing​

noData="hatch", or missing reads as zero​

Incidence is computed in the page, not baked into the fixture​

Options to try​

See also​