๐Ÿงญ Concept Census & demographics Data cleaning Census & Demographics ยท step 12 of 20 Updated

Boundary Changes Over Time: Why Two Census Years Do Not Line Up

Problem statement

Census geography is redrawn every decade. Between 2010 and 2020 the United States went from 74,134 census tracts to 85,528. The Census Bureau's relationship file records 22,893 tract codes that were new in 2020 and 11,499 codes from 2010 that were retired.

A join on GEOID between a 2010 table and 2020 boundaries matches 62,634 tracts and looks as if it mostly worked. It did not. Measured against the relationship file, 8,138 of those matching codes cover an area that changed by more than 1%. The code survived; the place it describes did not.

Reading the change is harder than it looks. Taken literally, the same relationship file says 43.6% of 2010 tracts overlap more than one 2020 tract. Most of that is digitising noise: a quarter of the file's rows are slivers smaller than 1% of both tracts, with a median area of 6,195 mยฒ. A comparison across census years has to separate real redrawing from those slivers, renamed codes and recoded counties before it can say anything about change.

Quick answer

Load the tract relationship file, ignore overlaps below a threshold, and count how each 2010 tract maps to 2020:

import pandas as pd

rel = pd.read_csv("tab20_tract20_tract10_natl.txt", sep="|", dtype=str,
                  encoding="utf-8-sig")
area = [c for c in rel.columns if c.startswith("AREA")]
rel[area] = rel[area].apply(pd.to_numeric)

part = rel["AREALAND_PART"] + rel["AREAWATER_PART"]
rel["share_of_10"] = part / (rel["AREALAND_TRACT_10"] + rel["AREAWATER_TRACT_10"])
rel["share_of_20"] = part / (rel["AREALAND_TRACT_20"] + rel["AREAWATER_TRACT_20"])

real = rel[(rel["share_of_10"] >= 0.01) | (rel["share_of_20"] >= 0.01)]
targets = real.groupby("GEOID_TRACT_10")["GEOID_TRACT_20"].nunique()
print(f"{(targets > 1).mean():.1%} of 2010 tracts feed more than one 2020 tract")

With a 1% threshold the national figure falls from 43.6% to 21.8%. Classified, the 74,134 tracts of 2010 were 53,073 one-to-one, 7,991 split, 4,904 merged and 8,166 part of a more complex rearrangement.

Table of 2010 tracts classified as one-to-one, split, merged or complex at a 1% threshold, for the United States and Delaware.
Most tracts carry over, but the ones that change are where growth and decline happened.

Step-by-step solution

1. Know which units are meant to be stable

The Census Bureau describes tracts as "small, relatively permanent statistical subdivisions of a county", with a population between 1,200 and 8,000 and an optimum of 4,000. Block groups sit inside tracts with 600 to 3,000 people. Blocks are rebuilt from scratch every census.

Stability is therefore a matter of degree. A block code from 2010 means nothing in 2020. A tract code usually means the same place. The two have to be treated differently.

2. Recognise the four kinds of change

Every 2010 tract falls into one of four patterns:

  • One-to-one โ€” the same area, usually the same code.
  • Split โ€” growth pushed a tract past the size limit and it became several. Tract 402.02 in Kent County, Delaware, became 402.04, 402.05 and 402.06.
  • Merged โ€” decline pushed tracts below the limit and they were combined.
  • Complex โ€” boundaries moved so that pieces of several old tracts form several new ones.

In Delaware, with a 1% threshold, the 218 tracts of 2010 were 128 one-to-one, 29 split, 24 merged and 37 complex. The split, merged and complex tracts held 411,829 of the state's 897,934 people in 2010, which is 46%.

3. Read the relationship file row by row

The tract relationship file has one row for every pair of tracts that share any area. Each row names the 2020 tract and the 2010 tract and gives the land and water area of their intersection:

GEOID_TRACT_20  GEOID_TRACT_10  AREALAND_PART  share_of_10  share_of_20
09001010201     09001010201          11254407     0.999574     0.999943
09001010201     09001010300               650     0.000067     0.000057

The second row is typical. Two tracts in Fairfield County, Connecticut, share 650 mยฒ of land: 0.0067% of the old tract. Nothing moved between them in any meaningful sense.

4. Separate real change from slivers

The threshold you choose changes the answer a great deal:

overlap threshold   2010 tracts feeding >1   2020 tracts drawing on >1
none                        32,288 (43.6%)             28,273 (33.1%)
0.1%                        23,954 (32.3%)             17,455 (20.4%)
1%                          16,157 (21.8%)              7,652  (8.9%)
5%                          11,968 (16.1%)              2,477  (2.9%)

With no threshold, 28,360 tracts are classified as complex. At 1% that falls to 8,166. The slivers come from lines being redrawn, not from tracts being reorganised. When the same road boundary is digitised a little differently in two vintages, the file records a thin shared strip.

5. Judge change by people, not by area

Area is a poor measure of what matters. In Delaware, 141 of the 432 relationship rows (33%) contained nobody at all in 2010, measured by allocating 2010 block populations to each piece.

A rule based on people โ€” a 2010 tract counts as divided only if at least 1% of its residents end up in a second 2020 tract โ€” flagged 45 Delaware tracts. The 1% area rule flagged 66, and the 5% area rule flagged 45. For this state, a 5% area threshold happened to agree with the population rule. Do not assume that holds elsewhere; compute it from blocks where you can.

6. Watch for codes that change without the boundaries

In 2022 Connecticut replaced its eight counties with nine planning regions as county-equivalents, and every tract GEOID in the state changed. Compared by geometry, 875 of the state's 884 tracts in the 2022 TIGER/Line file are within 1% of a 2020 tract with the same tract code. The GEOIDs have no match at all: 0 of 884 are shared.

A join that fails for an entire state is a coding change. A join that succeeds for a tract whose area changed is a boundary change, and that one gives no warning.

7. Choose how to compare the two years

Once you know what changed, there are three defensible choices:

  • Restrict the comparison to one-to-one tracts. This is simple, but in Delaware it drops the 46% of 2010 residents who lived in tracts that changed.
  • Crosswalk one year onto the other's boundaries using blocks. This is the accurate route, covered in the how-to guide on comparing census years.
  • Aggregate both years to units that did not change, such as counties.

Weighting by area alone is not on the list. In Delaware it moved 5% of the population into the wrong 2020 tract, and it reversed the direction of change โ€” growth shown as decline or decline as growth โ€” in 31 of the 262 tracts.

Bar chart of the share of 2010 tracts feeding more than one 2020 tract: 43.6% with no threshold, 32.3% at 0.1%, 21.8% at 1% and 16.1% at 5%.
The threshold is a choice, and it moves the answer by a factor of almost three.

Code examples

Example 1 โ€” loading the relationship file with overlap shares 

import numpy as np
import pandas as pd


def load_tract_relationship(path):
    """The 2020-2010 tract relationship file with overlap shares."""
    rel = pd.read_csv(path, sep="|", dtype=str, encoding="utf-8-sig")
    area = ["AREALAND_TRACT_20", "AREAWATER_TRACT_20", "AREALAND_TRACT_10",
            "AREAWATER_TRACT_10", "AREALAND_PART", "AREAWATER_PART"]
    rel[area] = rel[area].apply(pd.to_numeric)
    part = rel["AREALAND_PART"] + rel["AREAWATER_PART"]
    rel["share_of_10"] = part / (rel["AREALAND_TRACT_10"] + rel["AREAWATER_TRACT_10"])
    rel["share_of_20"] = part / (rel["AREALAND_TRACT_20"] + rel["AREAWATER_TRACT_20"])
    return rel

Reading with dtype=str keeps the GEOIDs intact, and utf-8-sig strips the byte-order mark the Census Bureau puts on the first header. Without that, the first column is named ๏ปฟOID_TRACT_20.

Example 2 โ€” classifying each 2010 tract 

def classify_changes(rel, threshold=0.01):
    """Label each 2010 tract one-to-one, split, merged or complex."""
    real = rel[(rel["share_of_10"] >= threshold) | (rel["share_of_20"] >= threshold)]
    targets = real.groupby("GEOID_TRACT_10")["GEOID_TRACT_20"].nunique()
    sources = real.groupby("GEOID_TRACT_20")["GEOID_TRACT_10"].nunique()
    real = real.assign(n_targets=real["GEOID_TRACT_10"].map(targets),
                       n_sources=real["GEOID_TRACT_20"].map(sources))
    per_tract = real.groupby("GEOID_TRACT_10").agg(n_targets=("n_targets", "first"),
                                                   max_sources=("n_sources", "max"))
    per_tract["change"] = np.select(
        [(per_tract.n_targets == 1) & (per_tract.max_sources == 1),
         (per_tract.n_targets > 1) & (per_tract.max_sources == 1),
         (per_tract.n_targets == 1) & (per_tract.max_sources > 1)],
        ["one-to-one", "split", "merged"], "complex")
    return per_tract

National results at three thresholds:

threshold   one-to-one    split   merged   complex
none            32,177    3,928    9,669    28,360
1%              53,073    7,991    4,904     8,166
5%              60,125    9,496    2,041     2,472

Example 3 โ€” codes that survived while their tract changed 

def same_code_different_shape(rel, tolerance=0.99):
    """GEOIDs used in both years whose area changed by more than 1 - tolerance."""
    same = rel[rel["GEOID_TRACT_10"] == rel["GEOID_TRACT_20"]]
    changed = same[(same["share_of_10"] < tolerance) | (same["share_of_20"] < tolerance)]
    return sorted(changed["GEOID_TRACT_20"])
moved = same_code_different_shape(load_tract_relationship("tab20_tract20_tract10_natl.txt"))
print(len(moved))
8138

These are the tracts a GEOID join matches without complaint. A time series built on them compares two different areas under one label.

Explanation

Why tracts are redrawn every ten years

Tracts are defined by population. The Census Bureau's glossary notes that they "occasionally are split due to population growth or merged as a result of substantial population decline". A tract that has grown well past 8,000 people no longer fits the criteria, so the next census divides it.

The Delaware split above kept its parent's base number: 402.02 became 402.04, 402.05 and 402.06, so the children look related. Nothing in a merged or complex rearrangement's codes tells you which old tracts it drew on; only the relationship file does.

Why the relationship file is full of slivers

The relationship files are computed by intersecting two sets of polygons. Any two lines that represent the same boundary but are digitised slightly differently produce a thin intersection. Across the country that adds up to 32,223 rows, 25.5% of the file, each under 1% of both tracts.

The file reports them faithfully because it does not know the lines were meant to be the same. Filtering them is the user's decision, which is why every count in this guide is quoted at a stated threshold.

Why a surviving code is not a promise

A GEOID identifies a tract within one vintage. Nothing in the numbering guarantees that the area stays fixed across vintages: a boundary can move a street over, or take in part of a neighbour, while the code stays the same.

Measured, 8,138 of the 62,634 codes used in both 2010 and 2020 changed shape by more than 1%. For analyses that need a fixed area, the relationship file is the reliable test, not the code.

Why area weighting is not a comparison method

Area weighting assumes people are spread evenly within a tract, and they rarely are. A tract that loses a large rural area to a neighbour loses little population but a lot of area. Area weighting then moves people out with the land.

In Delaware that put 45,362 of 897,934 people, or 5.05%, into the wrong 2020 tract. The state total was still exact, because area weights always preserve totals. That is why this error never shows up in a total check.

Why Connecticut changed every code at once

County codes are part of every tract GEOID. When planning regions (county codes 110โ€“190) replaced the old counties (001โ€“015) as Connecticut's county-equivalents in 2022 data products, every tract, block group and block GEOID in the state changed. The tract codes themselves survived, and the geometry barely moved.

Recoding the county part of the GEOID restores the join. No areal method is needed.

Two panels: 8,138 tract codes kept across 2010 and 2020 despite changed areas, and Connecticut tracts whose codes all changed in 2022 while 875 of 884 kept their shape.
Test the geometry with the relationship file; never infer it from the codes.

Edge cases or notes

  • Blocks are not comparable across censuses. Delaware had 24,115 blocks in 2010 and 20,198 in 2020, and the numbering was rebuilt.
  • Relationship files are area-based. They contain no population, so decide splits by allocating block counts to the pieces.
  • The threshold is a choice. Quote it with any count of changed tracts; at none, 1% and 5% the national share of dividing tracts was 43.6%, 21.8% and 16.1%.
  • Water counts in the total area. Coastal tracts can look changed because a shoreline was redrawn; test land area separately.
  • Treat a GEOID as unique only within its vintage. 11,499 codes from 2010 were retired, and 22,893 appeared for the first time in 2020.
  • County-equivalents change too. Connecticut in 2022 is the large recent case, and every tract GEOID in the state changed with it.
  • Zero-population tracts exist. Delaware has four in 2020, with codes in the 9800 and 9900 ranges, so growth rates for them are undefined.
  • A one-to-one tract can still gain or lose a block. "One-to-one at 1%" means under 1% of area moved, not zero.

FAQ

How many census tracts changed between 2010 and 2020?

It depends on the threshold. Ignoring overlaps below 1% of either tract, 21.8% of 2010 tracts fed more than one 2020 tract. Counting every sliver, 43.6% did.

Does the same GEOID mean the same area in both years?

Not reliably. Of the 62,634 tract codes used in both 2010 and 2020, 8,138 cover an area that changed by more than 1%.

Where does the Census Bureau publish the changes?

In relationship files under the rel2020 directory of www2.census.gov. There is one for tracts and one for blocks, each listing every pair of old and new units with the area they share.

Why do the relationship files list so many tiny overlaps?

Because they intersect two separately digitised sets of lines. A quarter of the national tract file's rows are under 1% of both tracts, with a median of 6,195 mยฒ.

Can I just use area weighting to compare years?

It preserves totals and misplaces people. In Delaware it moved 5% of the population into the wrong tract and reversed the direction of change for 31 of 262 tracts.

Why did every Connecticut tract GEOID change in 2022?

Connecticut's planning regions replaced its counties as county-equivalents, and the county code is part of every tract GEOID. The tract geometry was almost unchanged.