How to Map CDC PLACES Health Estimates by Census Tract in Python

On this page

Problem statement

CDC PLACES publishes estimates of 40 health measures for every census tract in the United States: diabetes, obesity, smoking, lack of health insurance, depression, and more, each with a 95% confidence interval. It is the most detailed open health data for small areas in the US, served through a Socrata API that answers SQL-like queries without a key. Mapping it takes a query, a join to tract boundaries and a choice of what to show. The estimates come from a statistical model, and their intervals are wide enough to matter when comparing neighbouring tracts.

This page downloads PLACES diabetes estimates for the 1,328 tracts of Cook County, Illinois, joins them to 2020-vintage tract boundaries, maps them, and measures how often two tracts' estimates cannot be told apart.

Quick answer

Query the tract dataset with a $where clause and join on the tract GEOID:

def soql(dataset, where, select="*", group=None, page=5000):
    rows, offset = [], 0
    while True:
        params = {"$select": select, "$where": where, "$limit": page, "$offset": offset}
        params.update({"$group": group} if group else {"$order": ":id"})          # stable paging
        part = requests.get(f"https://data.cdc.gov/resource/{dataset}.json", params=params, timeout=120).json()
        rows += part; offset += len(part)
        if len(part) < page:
            return pd.DataFrame(rows)

dia = soql("cwsq-ngmh", "countyfips = '17031' AND measureid = 'DIABETES'",
           "locationname, data_value, low_confidence_limit, high_confidence_limit, totalpop18plus, year")
dia = dia.astype({"data_value": float, "low_confidence_limit": float, "high_confidence_limit": float, "totalpop18plus": int})
gdf = TRACTS.merge(dia, left_on="GEOID", right_on="locationname", how="left")
print(f"{len(dia)} tract estimates for {dia.year.iloc[0]}; joined {gdf.data_value.notna().sum()} of {len(TRACTS)} tracts")
print(f"diagnosed diabetes among adults: {dia.data_value.min():.1f}% to {dia.data_value.max():.1f}%, median {dia.data_value.median():.1f}%")

1328 tract estimates for 2023; joined 1328 of 1331 tracts
diagnosed diabetes among adults: 2.0% to 32.3%, median 11.4%

Choropleth of diagnosed diabetes among adults for 1,328 census tracts in Cook County, Illinois, in five Fisher-Jenks classes from 2.0 to 32.3 percent: the lowest rates along the lakefront north of the Loop, the highest in the West Side and the South Side of Chicago and the southern suburbs; three tracts, including the airports, have no estimate. — PLACES diabetes estimates for Cook County. The pattern follows the city's well-known divides.

Step-by-step solution

1. Find the measures and their codes

The dataset is long: one row per tract and measure. Ask the API what is there instead of guessing codes:

m = soql("cwsq-ngmh", "countyfips = '17031'", "measureid, measure, count(*) as n", group="measureid, measure")
print(f"{len(m)} measures, for example:")
print(m.sort_values("measureid").head(6)[["measureid", "measure"]].to_string(index=False))

40 measures, for example:
measureid                                                                              measure
  ACCESS2                       Current lack of health insurance among adults aged 18-64 years
ARTHRITIS                                                               Arthritis among adults
    BINGE                                                          Binge drinking among adults
   BPHIGH                                                     High blood pressure among adults
    BPMED Taking medicine to control high blood pressure among adults with high blood pressure
   CANCER                                           Cancer (non-skin) or melanoma among adults

2. Look at the tracts with no estimate

A few tracts in the boundary file have no row in PLACES:

missing = gdf[gdf.data_value.isna()]
print(f"{len(missing)} tracts without an estimate; their land area: {missing.ALAND.sum() / 1e6:.1f} km²; tract IDs: {missing.GEOID.tolist()}")

3 tracts without an estimate; their land area: 23.1 km²; tract IDs: ['17031381700', '17031980100', '17031980000']

Tracts with no or very few residents, such as parks, airports and water, get no estimate. Leave them blank on the map rather than filling them with zero.

3. Read the confidence intervals

Every estimate has a 95% interval. Its width shows how precisely the model pins down each tract, and decides which differences are meaningful:

dia["ci_width"] = dia.high_confidence_limit - dia.low_confidence_limit
print(f"interval width: median {dia.ci_width.median():.1f} points, range {dia.ci_width.min():.1f}–{dia.ci_width.max():.1f}")
pairs = dia.sample(2000, replace=True, random_state=1).reset_index(drop=True).join(dia.sample(2000, replace=True, random_state=2).reset_index(drop=True), rsuffix="_b")
overlap = (pairs.low_confidence_limit <= pairs.high_confidence_limit_b) & (pairs.low_confidence_limit_b <= pairs.high_confidence_limit)
print(f"random pairs of tracts whose intervals overlap: {overlap.mean():.0%}")

interval width: median 2.8 points, range 0.5–6.3
random pairs of tracts whose intervals overlap: 35%

Chart of 1,328 tract estimates of diabetes ranked from lowest to highest, each with its 95 percent interval as a shaded band: the band is about 3 points wide for most tracts, so tracts near each other in the ranking overlap; median 11.4 percent. — Each estimate with its interval. Most of the ranking is within overlapping intervals.

Code examples

The data used on this page

CDC PLACES census tract data (dataset cwsq-ngmh on data.cdc.gov, 2025 release), model-based estimates from the Behavioral Risk Factor Surveillance System (BRFSS) and Census data, and the 2023 Census cartographic boundary file of Illinois tracts (2020 vintage), Cook County only.

from pathlib import Path

import geopandas as gpd
import numpy as np
import pandas as pd
import requests

Path("data").mkdir(exist_ok=True)
F = Path("data/cb_2023_17_tract_500k.zip")
if not F.exists():
    F.write_bytes(requests.get("https://www2.census.gov/geo/tiger/GENZ2023/shp/cb_2023_17_tract_500k.zip", timeout=300).content)
TRACTS = gpd.read_file(F).query("COUNTYFP == '031'").to_crs(3435)                  # Illinois East state plane, feet
print(f"{len(TRACTS)} Cook County tracts (2020 vintage)")

1331 Cook County tracts (2020 vintage)

Example 1: two measures on the same tracts

ob = soql("cwsq-ngmh", "countyfips = '17031' AND measureid = 'OBESITY'", "locationname, data_value").astype({"data_value": float})
both = dia.merge(ob, on="locationname", suffixes=("_diabetes", "_obesity"))
print(f"{len(both)} tracts; correlation of diabetes and obesity estimates: {both.data_value_diabetes.corr(both.data_value_obesity):.2f}")

1328 tracts; correlation of diabetes and obesity estimates: 0.90

Explanation

PLACES does not survey every tract. It fits a multilevel model to the BRFSS telephone survey, relating each health outcome to age, sex, race and ethnicity and to county and state effects, and then applies the model to each tract's Census population. The estimates are therefore predictions: they inherit the tract's demographic make-up and the survey's patterns, and they are smooth where the population is similar. The confidence intervals express the model's uncertainty, not sampling error in the tract itself.

Where a PLACES tract estimate comes from.

That makes PLACES excellent for seeing broad patterns and for planning, and weak for ranking individual tracts or for studying how a health outcome relates to the demographics that went into the model. Two neighbouring tracts a point or two apart are usually not distinguishable. Classes on the map should be wide enough that tracts within a class are comparable, and the interval width belongs in the legend or a companion map.

Edge cases or notes

PLACES traps.

Tract vintage: the 2025 release uses 2020 tracts; older releases used 2010 tracts, which do not join to 2020 boundaries without a crosswalk.
Crude and age-adjusted: the dataset carries crude prevalence at tract level; county-level data also offer age-adjusted values.
Socrata limits: page with $limit and $offset, and order by :id so pages do not overlap.
Numbers as text: the JSON API returns numbers as strings; cast them before arithmetic.
Year: the measure year (here survey year 2023) differs from the release year.

Internal links

Why do my health estimates look like a poverty map? — what model-based estimates contain
Small area rates explained — why small areas need special methods

FAQ

How do I download CDC PLACES data in Python?

Query https://data.cdc.gov/resource/cwsq-ngmh.json with Socrata parameters ($where, $select, $limit, $offset) using requests, and build a DataFrame from the JSON.

Are PLACES tract estimates measured or modelled?

Modelled: they are small area estimates from the BRFSS survey and Census data, not counts from each tract.

Which census tracts does PLACES use?

The 2025 release uses 2020 census tracts, so it joins directly to current Census tract boundaries on the 11-digit GEOID.

How should I show PLACES confidence intervals on a map?

Use classes wider than typical interval widths, show interval width in a companion map, or mark tracts whose intervals overlap the county value.