Data

The dataset contains 27 health and socioeconomic indicators for each of Chicago’s 77 community areas: birth outcomes, mortality rates, infectious disease, childhood lead exposure, and census measures such as poverty, income, and unemployment. Most indicators cover roughly 2005–2011, and income is in 2011 dollars, so these results describe the city around 2010, not today.

Exploring the data

Per capita income

All indicators (standardized)

Per capita income and gonorrhea rates

Sex n Pearson r p-value Spearman ρ
Females 65 -0.442 0.000 -0.475
Males 65 -0.413 0.001 -0.354

Gonorrhea rates are strongly and negatively associated with income. The relationship is not linear: rates are near zero across most higher-income areas and very high in a group of low-income areas. Spearman’s rank correlation is included because it doesn’t assume a linear relationship.

Handling missing values

Indicator Missing areas
childhood_blood_lead_level_screening 1
childhood_lead_poisoning 1
gonorrhea_in_females 12
gonorrhea_in_males 12

14 community areas have at least one missing value, almost all in the gonorrhea columns. Most of them are low-incidence areas on the Northwest and far Southwest Sides, where small counts were likely suppressed. Two approaches are compared:

  • Imputed: missing values are estimated with random-forest imputation (missForest).
  • Zero-filled: missing values are set to 0.

Both versions are standardized (mean 0, SD 1) before PCA and clustering.

Principal component analysis

  • PC1 (51% of variance) is an overall hardship measure. Nearly every health outcome loads on it in the same direction: homicide, firearm deaths, gonorrhea, unemployment, cancer, infant mortality, teen births, low birth weight, and diabetes. Per capita income and early prenatal care load on the opposite side. Community areas that do badly on one of these tend to do badly on all of them.
  • PC2 (17%) is a separate dimension made up of crowded housing, adults without a high school diploma, high birth and fertility rates, lead screening, and tuberculosis. This pattern is typical of lower-income immigrant neighborhoods, and it is largely independent of the mortality and violence measured by PC1.

Choosing the number of clusters

Suggested number of clusters
Method Imputed Zero-filled
Silhouette (max) 2 2
Gap statistic (firstSEmax) 3 6
NbClust (majority vote) 2 2

The silhouette method and the NbClust majority vote both point to two clusters, and the elbow plots show no sharp bend after k = 2 or 3. The gap statistic suggests more, especially for the zero-filled data. One possible reason is that areas with gonorrhea rates set to 0 sit unrealistically far from everyone else and get split off into their own group. Below, both k = 2 and k = 3 are examined; k = 3 turns out to be the more informative split.

k-means clustering

Dataset k Cluster sizes Avg. silhouette Between SS / total SS
Imputed 2 30 / 47 0.36 0.42
Imputed 3 28 / 17 / 32 0.27 0.53
Zero-filled 2 30 / 47 0.36 0.41
Zero-filled 3 28 / 17 / 32 0.27 0.52

The imputed and zero-filled datasets produce identical cluster memberships at both k = 2 and k = 3, so the choice of missing-value method doesn’t affect the conclusions. The plots and profiles below use the imputed data.

Two clusters

Three clusters

Cluster profiles (k = 3)

Indicator High hardship Low income, low mortality Higher income
Community areas 28.0 17.0 32.0
Per capita income ($) 16,338.6 16,394.2 37,407.4
Unemployment (%) 20.6 11.6 7.8
Below poverty level (%) 30.4 19.3 12.0
No high school diploma (%) 22.7 37.6 12.2
Crowded housing (%) 4.7 9.3 2.8
Homicide (per 100k) 37.2 9.5 5.9
Firearm-related deaths (per 100k) 32.8 8.8 6.9
Infant mortality (per 1,000) 13.0 6.1 6.0
Low birth weight (%) 14.3 7.5 7.7
Teen birth rate (per 1,000) 75.0 53.4 26.5
Diabetes-related deaths (per 100k) 92.3 65.1 57.7
Cancer deaths, all sites (per 100k) 242.2 154.8 173.3
Gonorrhea, females (per 100k) 1,766.4 210.7 250.5
General fertility rate 71.5 82.4 58.3
Tuberculosis (per 100k) 7.3 9.2 5.2
Cluster Community areas
High hardship Auburn Gresham, Austin, Avalon Park, Burnside, Calumet Heights, Chatham, Douglas, East Garfield Park, Englewood, Fuller Park, Grand Boulevard, Greater Grand Crossing, Humboldt Park, New City, North Lawndale, Oakland, Pullman, Riverdale, Roseland, South Chicago, South Deering, South Shore, Washington Heights, Washington Park, West Englewood, West Garfield Park, West Pullman, Woodlawn
Low income, low mortality Albany Park, Archer Heights, Armour Square, Avondale, Belmont Cragin, Brighton Park, Chicago Lawn, East Side, Gage Park, Hermosa, Lower West Side, McKinley Park, Montclaire, South Lawndale, West Elsdon, West Lawn, West Ridge
Higher income Ashburn, Beverly, Bridgeport, Clearing, Dunning, Edgewater, Edison Park, Forest Glen, Garfield Ridge, Hegewisch, Hyde Park, Irving Park, Jefferson Park, Kenwood, Lake View, Lincoln Park, Lincoln Square, Logan Square, Loop, Morgan Park, Mount Greenwood, Near North Side, Near South Side, Near West Side, North Center, North Park, Norwood Park, O’Hare, Portage Park, Rogers Park, Uptown, West Town

Interpretation

Two clusters. The k = 2 solution splits the city along PC1 into 30 high-hardship community areas and 47 lower-hardship ones. The high-hardship cluster is almost entirely made up of predominantly Black community areas on the South and West Sides, including Englewood, West Englewood, Austin, North Lawndale, East and West Garfield Park, Washington Park, and Roseland. Between the two clusters, rates of homicide, gonorrhea, and infant mortality differ by roughly 2x to 8x, while educational attainment and crowded housing are similar.

Three clusters. At k = 3 the lower-hardship group splits along PC2, and the high-hardship cluster stays almost the same:

  1. High hardship (28 areas): per capita income of $16,300, unemployment of 20.6%, homicide at 37.2 per 100k, and infant mortality of 13 per 1,000. These areas have the worst value on nearly every mortality and chronic-disease indicator.
  2. Low income, low mortality (17 areas): largely Latino and immigrant neighborhoods such as South Lawndale (Little Village), Lower West Side (Pilsen), Gage Park, Brighton Park, and Belmont Cragin. Their income ($16,400) is the same as the high-hardship cluster, but homicide (9.5 per 100k), infant mortality (6.1), and low birth weight (7.5%) resemble the higher-income cluster. They stand out instead on crowded housing (9.3%), adults without a high school diploma (37.6%), fertility, and tuberculosis. This matches the pattern often described as the “immigrant health paradox.”
  3. Higher income (32 areas): the North Side lakefront and downtown (Lincoln Park, Lake View, the Loop, Near North Side), Northwest Side neighborhoods such as Edison Park and Norwood Park, and Southwest Side areas such as Beverly and Mount Greenwood. Income is $37,400, and these areas have the best values on almost every indicator.

Main takeaway: income alone doesn’t account for health outcomes across Chicago’s community areas. Clusters 1 and 2 have nearly the same per capita income, but homicide is about four times higher and infant mortality about twice as high in the high-hardship cluster.

Caveats.

  • The clusters are only moderately separated (average silhouette of 0.36 for k = 2 and 0.27 for k = 3). They are better read as segments along a continuum than as sharply separate groups. The areas closest to a cluster boundary at k = 3 (lowest silhouette width) are New City, Rogers Park, Logan Square, Calumet Heights, Humboldt Park.
  • The indicators are mostly from 2005–2011 and are aggregated to large community areas, so they can hide variation within each area.
  • k-means gives every indicator equal weight. Several indicators are highly correlated (homicide and firearm deaths, gonorrhea in females and males), so that dimension effectively counts more than once.