Boundaries from the wrong decade

New York City had 2,167 census tracts in 2010 and 2,324 in 2020. Drawing the ACS 2019-2023 income estimates, which are keyed to 2020 tracts, on the 2010 lines raises no error and leaves 134 tracts grey for a reason the map cannot show.

Change the layout

Layout

Median household income, ACS 2019-2023, on 2020 and on 2010 tract boundaries

2020 map: 2,324 tracts drawn, 2,192 coloured, 132 with no estimate. 2010 map: 2,167 drawn, 1,956 coloured, 77 with no estimate, 134 with no row at that id.

Nothing on the 2010 map is flagged as wrong. The legend is right, the colours are in range, the city is the right shape, and 1,956 of 2,167 tracts carry a real ACS 2019-2023 estimate. Of the 211 grey ones, 77 have no estimate and are grey on the 2020 map too, and 134 are tracts the Bureau retired after 2010. The map cannot show which is which, so the mistake takes the same colour as the honest gap.

Source: US Census Bureau, American Community Survey 2019-2023 5-year estimates, table B19013, keyed to 2020 tracts. Boundaries: US Census cartographic boundary files for 2020 (cb_2020_36_tract_500k) and 2010 (gz_2010_36_140_00_500k). Counts are computed from the tract ids in the two boundary files and the estimates.

How the join fails

2,033 tract ids appear in both vintages, 134 were retired after 2010, and 291 are new in 2020. Joining 2020-keyed values to 2010 geometry therefore lands for most tracts and misses for the rest, and the miss does not stop anything. It shows up as a scatter of grey polygons.

Grey already means something on this map. 132 tracts have no income estimate on the correct 2020 map: 81 have no residents, 12 have fewer than 50, and 39 have a withheld estimate. On the 2010 map 134 more go grey because their id has no row in the estimates at all. On the finished map the two look identical, and the legend cannot separate them either.

The mismatched map is not obviously worse. It is the right city, with a plausible income gradient and a correct legend. It is wrong in two ways. Some tracts are absent, and the tracts that did survive can have different boundaries under the same id, so a value drawn there describes an area that no longer exists. The Bureau redraws tracts mostly after population change, and the error is assuming that an id means the same place in both vintages.

What the id lists can and cannot tell you

For each outcome of joining ACS 2019-2023 values to the two tract vintages: how many ids, what happens to the join, and what the count does not establish.
Outcome	Tracts	What the join does	What this does not establish
Id in both vintages	2,033	The join lands. An ACS 2019-2023 value drawn on the 2010 polygon of this id gets a colour and raises nothing.	That the polygon is the same polygon. An id surviving a decade says the Bureau reused the number, not that it left the boundary alone, so the value may describe an area with different edges, and nothing on screen can show that.
Id in 2010 only	134	The join misses. The polygon is drawn in the no-data grey, beside tracts with no estimate painted identically.	That the place stopped existing. The tract was retired as a unit, usually split or merged after population change, and its households are on the 2020 map under other ids.
Id in 2020 only	291	Nothing is drawn for it on the 2010 map, so a reader of that map never learns these tracts exist.	That the place is new. A new id is a new reporting unit, not new ground.

Separating an unchanged tract from a redrawn one that kept its id needs the Census Bureau’s tract relationship file, which the data here does not include, so it is not shown. Inferring the split from how the two geometries overlap would be a modelled relationship, and presenting a modelled relationship as a published one is the error this page is about.

Nor is the 2010 panel a change map. It is one period of estimates on two geometries. A real change map needs two comparable estimation periods, boundaries harmonised between them, rules for uncertainty and, for a money measure, an inflation adjustment.