Stations and people: a complete history

How the station layer was made

This is, as near as the records allow, every public passenger station Great Britain has had: 9,156 of them, 2,990 still open. Each one has a place on the map, the year it opened, the year it closed if it did, and the year it came back if it reopened.

Most of the dates come from one book, Michael Quick's chronology of passenger stations, which railway historians treat as the standard reference. In the published dataset each date cites the page it came from. Where Quick is silent, mainly heritage lines and the London Underground, the dates come from Wikipedia and Wikidata, read in more than one way and compared. A few hundred stations have no record of their closing at all. For those we took the year their line closed, or, where even that is unknown, a typical year for a station of their age. These estimates are flagged in the data. Taken together they are about right; for any one station they may be wrong. Six dates in seven are backed by two separate records that agree.

A station is drawn in a year if it had opened by then and had not yet closed. One that closed and reopened goes dark in between. Hover any station for its dates.

How the people layer was made

The blue haze is the census: every one since 1801, placed on the map. A census is taken every ten years, so in between the map shows the nearest one; there was no census in 1941. Getting a census onto a map means two things: having the count for each place, and knowing where that place was. Neither is given for the old ones.

The counts. For the censuses since 1971 the statistics offices publish counts for small areas, ready to use. For 1851 to 1901 the counts come from a Cambridge research dataset of the registration sub-districts the Victorian census was reported by, about two thousand areas, each a handful of parishes. For 1911, 1951 and 1961 they exist only by local government district. For everything else we read them off the printed census volumes: the 1831 volume, which carries the counts of 1801, 1811, 1821 and 1831 parish by parish, the 1841 volume, and for 1921 and 1931 the County Reports, again parish by parish. Scotland is counted by civil parish throughout, from National Records of Scotland's own parish tables; for 1921 to 1971 those were read from scans in the same way, with a parish that failed its check estimated between its 1911 and 1981 figures instead. Those pages were scanned and read by machine, and every row had to prove itself before it was kept: a parish row stands only when its printed males and females add up to its printed total, and a page's rows must add up to the printed page total. The transcriptions are published with the dataset so any figure can be checked against the page it came from. The totals for every year come out within a fraction of a percent of the published national figures.

The places. A modern small area comes with the point its people cluster around, so it goes straight onto the map. The older areas do not. A Victorian registration district, typically a market town and the countryside around it, is too big to mark as one point, so we spread its count over the hamlets, villages and towns inside it, in proportion to the place names the Ordnance Survey printed there on its maps of around 1900: a town's name draws more of the count than a hamlet's. Those maps are the closest thing there is to a complete picture of where people lived in that era. For the parish counts of 1801 to 1841 and 1921 to 1931, each parish is placed by finding its name among those same map names, looking first inside the district an 1852 index of parishes assigns it to, so that one of England's many Suttons lands in the right county. A parish whose name cannot be found is spread over its district. Scotland is placed by its civil parish boundaries throughout.

How wrong it is. The spreading method was tested on the 1971 census, where the true positions are known down to a few streets: the error grows with the size of the areas a year's counts come in, by about 8% of their width, and that is the error band each modelled year carries. Four years, 1851, 1861, 1921 and 1931, can also be checked directly against parishes placed by name, and each comes out well inside its band, so the band is a ceiling. The haze is still not where every household was. It is each count spread over our best reading of its area. The colour scale is logarithmic, so a city centre and a market town can both be seen instead of the cities swamping everything else.

How the distances were worked out

Take everyone in the country in a given month, measure the straight line from where they live to the nearest station open that month, and sort those distances from shortest to longest. The 50th percentile is the person halfway along that line: half the country lived closer to a station than they did, half further. The 90th percentile is the person nine tenths of the way along, and so on. We did this for every month from 1825 to today. The people of each month are a blend of the two nearest censuses, which is why the chart moves smoothly while the map, which shows one census at a time, changes in steps. In the published dataset the census years also carry an error band, which adds up every source of error we could measure or bound. Read it as the size of the known errors, not as a probability.