Place-based Economic Segregation Metrics

• Place-based measures of economic segregation are analytic frameworks that quantify the spatial concentration of socio-economic groups using traditional indices and advanced methodologies.
• They integrate classical indicators like ICE with information-theoretic metrics and spatial Bayesian models to assess local imbalances and guide policy interventions.
• These methods are applied in both residential and mobility contexts, providing actionable insights to address health and economic disparities across urban areas.

Place-based measures of economic segregation quantify the spatial distribution and concentration of socio-economic groups within clearly defined geographic units—such as tracts, neighborhoods, or counties—using indices and frameworks that explicitly respect spatial structure, demographic extremity, and the multi-scale nature of segregation. These measures have evolved beyond classic indices to incorporate advances in information theory, random walks, Bayesian inference, network science, and urban mobility analytics.

1. Classical Place-based Indices: ICE and Representation/Exposure

Several foundational indices remain widely used for operationalizing place-based economic segregation. The Index of Concentration at the Extremes (ICE) and the representation/exposure indices are particularly notable.

ICE (Index of Concentration at the Extremes): For a given unit $i$, ICE is defined as

$\mathrm{ICE}_i = \frac{A_i - P_i}{T_i}$

where $A_i$ is the number of “privileged” (e.g., non-Hispanic Whites in the top income decile), $P_i$ is the number of “deprived” (e.g., non-Hispanic Blacks in the bottom income quintile), and $T_i$ is the total with known income. $\mathrm{ICE}_i = +1$ indicates total privilege concentration, $-1$ complete deprivation, and $0$ a balance (Xu et al., 2023, Xu et al., 2023).

Representation Index: For group $\alpha$ in areal unit $i$: $\mathrm{ICE}_i = \frac{A_i - P_i}{T_i}$0 This ratio quantifies local over- or under-representation relative to the city's group share. Exposure between groups $\mathrm{ICE}_i = \frac{A_i - P_i}{T_i}$1 and $\mathrm{ICE}_i = \frac{A_i - P_i}{T_i}$2 is then: $\mathrm{ICE}_i = \frac{A_i - P_i}{T_i}$3 with $\mathrm{ICE}_i = \frac{A_i - P_i}{T_i}$4 implying spatial attraction, $\mathrm{ICE}_i = \frac{A_i - P_i}{T_i}$5 repulsion (Louf et al., 2015).

Both indices are straightforward to compute from census tabulations and are foundational for identifying local neighborhoods of over- or under-concentration, as well as for designing clustering algorithms to identify contiguous “class neighborhoods.”

2. Information-Theoretic and Decomposable Indices

Recent research emphasizes information theory for flexible, fully decomposable analysis at varying spatial and economic scales.

Divergence Index (Kullback–Leibler): For discrete economic classes $\mathrm{ICE}_i = \frac{A_i - P_i}{T_i}$6: $\mathrm{ICE}_i = \frac{A_i - P_i}{T_i}$7 where $\mathrm{ICE}_i = \frac{A_i - P_i}{T_i}$8 is the group-$\mathrm{ICE}_i = \frac{A_i - P_i}{T_i}$9 share in unit $A_i$0, and $A_i$1 is the group share region-wide. The regional index is aggregate: $A_i$2 This index is strictly nonnegative, zero when local and regional compositions match, and fully decomposable into between- and within-region terms (Roberto, 2015, Simpson et al., 2018).

Mutual Information Framework: For individual economic label $A_i$3 and spatial unit $A_i$4: $A_i$5 This quantifies, in “nats,” the average information gain about an individual’s economic class from their location (Sahasrabuddhe et al., 28 Nov 2025). This framework generalizes seamlessly to any number of income categories or nested spatial scales, and supports exact decompositions: $A_i$6 Distinguishing between- and within-region segregation is essential for multi-scale policy analysis.

3. Incorporating Spatial Structure and Autocorrelation

Place-based indices are subject to spatial autocorrelation, which classical non-spatial models neglect. To address this, contemporary research employs spatial Bayesian modeling, especially for ICE.

Bayesian Reformulation of ICE: Treat the privileged and deprived counts $A_i$7 as (conditionally independent) Binomial variables given latent probabilities $A_i$8. These probabilities are modeled as: $A_i$9 where $P_i$0 incorporates a spatial random effect (e.g., Besag–York–Mollié CAR prior), enabling the estimation of spatially smoothed ICE values with full posterior uncertainty, effectively borrowing strength across neighboring units and correcting for small-area noise (Xu et al., 2023).

Two-Stage Spatial Bayesian Model: Xu and Tabb extend this to model both the latent health structure (as spatial factors) and the spatially varying association of ICE with premature mortality, yielding posterior maps that highlight geographic patterning and permit direct inference on local links between economic segregation and health outcomes (Xu et al., 2023).

4. Network-Based and Mobility-Driven Indices

Network approaches address economic segregation in relational (rather than only areal) terms, capturing both static and dynamic patterns.

Network Entropy (Commuting Networks): For commuting flows $P_i$1 from tract $P_i$2 to $P_i$3, global and local entropy measures are: $P_i$4

$P_i$5

Decomposition by income (high- and low-income subnetworks) enables examination of experiential segregation, revealing strong correlations between entropy differences and ICE indices in employment and residential contexts (Iyer et al., 2023).

Exposure Segregation (ES, Mobility Data): At the individual level, ES is the Pearson correlation

$P_i$6

where $P_i$7 is an individual's SES, $P_i$8 is the mean SES in their realized encounter set (e.g., mobile phone co-locations). ES thus captures segregation in “realized exposure” space, not just home locations, and can be computed across urban scales (Nilforoshan et al., 2022).

Mixing Matrices from Place Visits: Using digital traces, city-wide mixing matrices $P_i$9 (probability that a visit by a class-$T_i$0 individual is to a class-$T_i$1 place) yield indices of diagonality/assortativity ($T_i$2), upward-bias, and dispersion—operationalizing mobility-driven economic segregation distinct from residential segregation (Hilman et al., 2021).

5. Multi-Scale, Topological, and Random Walk Approaches

Place-based analysis increasingly leverages non-Euclidean spatial frameworks and multi-scale methods.

Generalized Jensen-Shannon Divergence (GJSD): The normalized divergence between local distributions $T_i$3 for spatial units $T_i$4, $T_i$5: $T_i$6 This can be extended to clusters of multiple tracts, mitigating Modifiable Areal Unit Problems and supporting topological interpretations via tract adjacency graphs, with decay patterns parameterized by scaling laws. Clustering algorithms apply modularity optimization to these divergence-weighted networks for unsupervised identification of homogeneous regions (Kirkley, 2020).

Random Walks on Spatial Graphs: Two nonparametric measures, spatial variance and local spatial diversity, are estimated by the dynamics of random walks over adjacency networks of areal units labeled by income class. The key measures are: $T_i$7

$T_i$8

where $T_i$9 is the time-averaged exposure profile, $\mathrm{ICE}_i = +1$0 is the probability a walker remains within class $\mathrm{ICE}_i = +1$1 after $\mathrm{ICE}_i = +1$2 steps. These measures are scale-invariant and robust to city size, facilitating direct comparison across contexts (Sousa et al., 2020).

6. Extensions: Amenity Complexity, Urban Mixing, and Measurement Challenges

Amenity Complexity: The economic-complexity “method of reflection” is adapted to the urban context (amenity–neighborhood matrices), quantifying neighborhood or amenity “complexity” via higher-order cross-specialization. Regression models show that these complexity indices, rather than raw diversity or ubiquity of amenities, are strong predictors of realized socio-economic mixing as traced by GPS mobility (Juhász et al., 2022).

Measurement and Interpretation Issues:

• All place-based measures face the challenge of defining and aggregating spatial units (tracts, blocks, grid cells), with significant implications (MAUP).
• Binning of income and smoothing strategies alter results; sensitivity analyses and decomposition are essential.
• Null-model or randomization baselines are critical for interpretation—especially in entropy or random walk-based approaches.
• Robust estimation in the presence of small-area sampling noise or population heterogeneity often requires Bayesian or empirical-Bayes smoothing.

Comparison of Prominent Indices

Index/Framework	Input Data	Main Feature	Spatial Coupling
ICE	Census/categorical	Extremes, direct interpretation	Optional (Bayesian)
Divergence Index	Categorical/density, census	KL divergence, full decomposability	No, can be spatialized
Representation/Exposure	Census, categorical	Null-model, co-location/attraction	No
Information Theory ($\mathrm{ICE}_i = +1$3)	Any, counts/densities	Mutual info, multi-scale	Implicit in structure
Network Entropy	Commuting, OD flows	Structural diversity, node entropy	Graph topology
ES (Mobility)	GPS traces, home SES	Realized exposure segregation	Implicit mobility
Random Walk	Areal unit graph, labels	Multi-scale, scale-invariant	Full
Amenity Complexity	POI data, mobility	Economic-complexity, mixing	Amenity visitation net

7. Policy and Research Implications

• Place-based measures directly inform urban planning, health equity, and policy interventions by localizing both privilege and deprivation in empirical, spatially explicit terms.
• Spatial smoothing and Bayesian modeling have become standard for uncertainty quantification and for correcting boundary effects or sparse-data instability.
• Network and entropy-based indices provide insight into “experienced” and “dynamic” segregation, not just static residential patterns.
• Multi-scale and decomposable information-theoretic approaches are necessary for capturing the hierarchical and nested structure of urban and regional segregation.

Ongoing work addresses the limits and interpretability of different place-based measures, with a trend toward hybrid models that integrate spatial statistics, network analysis, Bayesian inference, and high-resolution behavioral data to yield robust, actionable segregation profiles suited for diverse policy and research applications (Sahasrabuddhe et al., 28 Nov 2025, Xu et al., 2023, Kirkley, 2020, Iyer et al., 2023, Nilforoshan et al., 2022).