Index of Dissimilarity
- Introduced by
- Otis Dudley Duncan & Beverly Duncan
- Year
- 1955
- Type
- Index of evenness of two groups across units
- Family
- Segregation measurement
What does it do?
The index of dissimilarity, often called the Duncan segregation index, measures how unevenly two groups — such as two racial or occupational groups — are distributed across a set of units like neighborhoods, schools, or occupations.
It ranges from 0, when both groups have identical distributions across units, to 1, when the units are completely segregated, and has the intuitive interpretation of the share of one group that would have to relocate to achieve an even distribution.
The intuition behind the method
Imagine a city's neighborhoods and two groups of residents. If every neighborhood had the same group mix as the city overall, there would be no segregation. The dissimilarity index measures the gap from that ideal by adding up, neighborhood by neighborhood, how over- or under-represented each group is. The clever part is its interpretation: the index equals the fraction of one group's members who would need to move to different neighborhoods to make every neighborhood mirror the citywide composition. A value of 0.6 means 60 percent would have to relocate to erase the unevenness.
History
Otis Dudley Duncan and Beverly Duncan introduced and rigorously justified the index in their 1955 American Sociological Review paper, comparing the available segregation indices and arguing that the dissimilarity index best captured occupational and residential segregation. It became the dominant measure for decades. Douglas Massey and Nancy Denton's 1988 Social Forces paper situated it within a five-dimensional framework of segregation (evenness, exposure, concentration, centralization, clustering), clarifying that dissimilarity measures only evenness and motivating complementary indices.
When to use it
Use the index of dissimilarity to measure the evenness with which two groups are spread across units — residential segregation across neighborhoods, school segregation, occupational sex or race segregation. It is the standard measure of the evenness dimension and is easy to compute and communicate. It is appropriate only for comparing exactly two groups (multigroup settings need entropy-based measures), and it does not capture other dimensions of segregation such as isolation, clustering, or concentration. It is sensitive to the number and size of units (the checkerboard and modifiable-areal-unit problems) and ignores the spatial arrangement of units, so it should be reported alongside complementary measures.
Limitations and considerations
- Defined for only two groups at a time; multigroup segregation requires entropy or other indices.
- Captures only evenness, ignoring isolation, exposure, clustering, and concentration.
- Aspatial: it ignores where units sit relative to one another, so two very different spatial patterns can score identically.
- Sensitive to the number and size of units (modifiable areal unit problem) and biased upward when units are small or counts sparse.
Source conditions do not replace independent scientific review or validation by an analysis engine.
Strengths
- Intuitive interpretation as the share of a group that must relocate to achieve evenness.
- Simple to compute from readily available cross-tabulated counts.
- Bounded on [0,1], enabling clear comparison across cities, periods, or domains.
- The long-standing standard measure of the evenness dimension, ensuring comparability with a vast literature.
Applications
- Measuring residential segregation between racial or ethnic groups across census tracts.
- Quantifying school segregation across schools or districts.
- Assessing occupational sex or race segregation across occupations.
- Tracking change in segregation over decades for the same area.
- Comparing segregation levels across cities or metropolitan areas.
How it works
How it works
For each unit i, compute that unit's share of the entire group A population (a_i divided by total A) and its share of the entire group B population (b_i divided by total B). These are distributional shares, not within-unit proportions, so they each sum to one across all units.
Formula
\frac{a_i}{A} = \text{unit } i\text{'s share of group A},\quad \frac{b_i}{B} = \text{share of group B}The index is half the sum over all units of the absolute difference between the two groups' shares. Taking absolute differences ensures over- and under-representation do not cancel, and the factor of one-half rescales the total to the [0,1] range.
Formula
D = \frac{1}{2}\sum_{i=1}^{n} \left| \frac{a_i}{A} - \frac{b_i}{B} \right|A value of 0 means the two groups are distributed identically across units (no segregation on this measure), while a value of 1 means no unit contains members of both groups (complete segregation). Intermediate values indicate the degree of unevenness.
Formula
D \in [0,1]:\ 0 = \text{even},\ 1 = \text{complete segregation}The index has a direct substantive reading: it equals the proportion of one group that would have to move to other units, with no compensating moves required, for both groups to be evenly distributed. This 'displacement' interpretation makes it the most widely reported measure of the evenness dimension of segregation.
Formula
D = \text{share of one group that must relocate for evenness}Common pitfalls
- Comparing D across cities with very different numbers or sizes of units without recognizing the upward small-unit bias.
- Using D for more than two groups by averaging pairwise values, which has no clean interpretation.
- Treating a low D as 'no segregation' when isolation or spatial clustering may remain high.
- Ignoring the checkerboard problem: D is unchanged if segregated units are rearranged in space, so it misses clustering.
Frequently asked questions
What does a dissimilarity index of, say, 0.5 mean?
It means that 50 percent of one group would have to move to different units for both groups to be evenly distributed — that is, for every unit to have the same group composition as the whole area. Conventionally, values below 0.3 are considered low segregation, 0.3–0.6 moderate, and above 0.6 high, though these thresholds are rules of thumb rather than hard cutoffs.
Why can the index only compare two groups?
The formula contrasts two distributional profiles by summing their absolute differences. With three or more groups there is no single pair of profiles to compare, and averaging pairwise dissimilarities loses a clean interpretation. Multigroup segregation is better measured with entropy-based indices such as Theil's information index, which decompose total segregation across all groups simultaneously.
What is the small-unit (checkerboard) bias?
When units are small and contain few people, random variation alone makes them look more segregated, biasing the index upward. Additionally, because the index ignores spatial arrangement (the checkerboard problem), rearranging segregated units into clustered versus dispersed patterns leaves D unchanged. Both issues mean D should be interpreted cautiously and paired with spatial or clustering measures.
Related methods
Sources
Duncan, O. D., & Duncan, B. (1955). A methodological analysis of segregation indexes. American Sociological Review, 20(2), 210–217.
Open DOI referenceCopied.Massey, D. S., & Denton, N. A. (1988). The dimensions of residential segregation. Social Forces, 67(2), 281–315.
Open DOI referenceCopied.