Ecological Correlation¶

A correlation computed on aggregated group-level units that need not equal — and can reverse the sign of — the individual-level correlation, so it warrants no claim about the individuals inside the groups.

Version: v3 · 2026-09-28 · History
Domain-specific #: 302
Domain group: Formal Sciences
Origin domain: Experimental Design & Statistics
Subdomain: Aggregation and Inference Fallacies → Experimental Design & Statistics

Core Idea¶

An ecological correlation is computed between variables measured at the group level — countries, counties, tracts — rather than on the individuals within them. Robinson showed in 1950 that it need not equal, and can even reverse the sign of, the individual-level correlation: the 1930 census showed +0.53 between state proportions foreign-born and literate, but −0.11 at the individual level. Aggregation is lossy — it discards within-group variance — so when a common third factor drives both between-group variation and group composition, the coefficient reflects that factor, not the individual association.

How would you explain it like I'm…

The Group Pattern Trick

Say the classes that have more tall kids also get better spelling scores. That does not mean tall kids are better spellers. Maybe the tall kids are bad spellers and something else about those classes helps. Ecological correlation is a pattern found between whole groups, and it can tell a different story from the one about each person.

Group Patterns Aren't People Patterns

A correlation is when two things tend to go up or down together. An ecological correlation is one measured using whole groups, like states or schools, instead of individual people. The tricky part is that the group pattern can be very different from the person pattern, and it can even point the opposite way. In 1930 U.S. census data, states with more immigrants had more people who could read. But individual immigrants were actually slightly less likely to read than other people. So a group pattern can't tell you what is true for each person.

Aggregate-Level Correlation

An ecological correlation is computed on aggregated units, such as countries, counties, or schools, rather than on the individuals inside them. W. S. Robinson showed in 1950 that it need not match the individual-level correlation and can even have the opposite sign. In 1930 census data, the cross-state correlation between the share foreign-born and the share literate was +0.53, but the individual-level correlation between being foreign-born and being literate was −0.11. The cause is aggregation: averaging within groups throws away within-group variation. A third factor that shapes which people live in which groups can then drive the group-level pattern. Concluding something about individuals from such a correlation is called the ecological fallacy.

An ecological correlation is a correlation between group-level variables, such as proportions or means over geographic or institutional units. It is not a correlation among the individuals who make up those groups. Robinson (1950) demonstrated formally that the two can diverge in magnitude and sign. His 1930 census example showed a state-level correlation of +0.53 between the proportion foreign-born and the proportion literate, against an individual-level correlation of −0.11. Mechanistically, aggregation is lossy: it discards within-group variance and keeps only between-group variance. If between-group variation in both variables is driven by a common factor that also governs group composition, the ecological correlation tracks that factor rather than the individual association. The size and direction of the distortion depend on how groups are drawn, on the within-group distributions, and on confounding of membership. Crucially, the distortion cannot be detected from the aggregate data alone. The inferential rule follows: an ecological correlation never by itself warrants a claim about individual-level association, and making such a claim is the ecological fallacy.

Scope of Application¶

Because it is a statistic-plus-inferential-rule, not a mechanism, ecological correlation applies wherever a correlation is computed on aggregated units and a claim is risked about the units inside them.

Epidemiology — country-level exposure correlations against disease prevalence, misread onto persons.
Political science and sociology — precinct returns read onto individual voters.
Economics — cross-country GDP-per-capita correlations versus household relationships.
Education research — school-level test-score correlations used as student-level proxies.
Public health and policy — aggregate findings a covariate cannot rescue; only micro-data can.

Clarity¶

Naming it forces apart the unit of analysis at which a number was computed from the unit of claim an interpreter wants to make. Labelling the group-level coefficient ecological, and the unlicensed leap the ecological fallacy, makes the smuggled inference visible. It also supplies a triage rule: any correlation is presumed silent about every level but the one it was measured at, until the within-group structure is modelled.

Manages Complexity¶

Aggregate data are everywhere and any pair can be correlated then misread. The construct compresses an open-ended verification problem to one tracked attribute — the aggregation level — and one default: presumed silence about other levels. A whole class of policy and journalistic claims is pruned by that move before any modelling.

Abstract Reasoning¶

Its moves are mostly prohibitive: a boundary-drawing level rule (group coefficients warrant nothing about individuals), a negative diagnostic about non-recoverability (the individual sign cannot be recovered from the aggregate), an interventionist move (only micro-data fixes it, not a covariate), and a classificatory move placing look-alikes — Simpson's paradox, the modifiable areal unit problem, the ecological inference problem, ordinary confounding.

Knowledge Transfer¶

Ecological correlation is a statistic-plus-inferential-rule, so it transfers literally to any field that correlates aggregated units and risks a unit-level claim — same coefficient, same sign-reversal hazard, same presumed-silence default, same micro-data remedy. The boundary is instrument-reach versus over-reading (the fallacy is the over-read the term forbids). A genuinely cross-domain lesson about scale-dependent relations belongs to the parents Aggregation, Correlation, and Scaling and Scale Dependence, not this named construct.

Relationships to Other Abstractions¶

Current abstraction Ecological Correlation Domain-specific

Parents (2) — outgoing typed relations

Ecological Correlation is a kind of Correlation Prime

Ecological correlation is a correlation specialized to variables measured on partition-aggregated groups rather than individuals.
Ecological Correlation is a kind of Partition Dependence of Aggregates Prime

Ecological correlation is the correlation-coefficient species of partition dependence: grouping fixes the retained between-group variance.

Children (1) — abstractions linked here by typed relations

Fallacy of Division Domain-specific presupposes, conditional Ecological Correlation

In its ecological statistical branch, division presupposes a group-level correlation that the reasoner then reads onto individuals.

Hierarchy paths (2) — routes to 2 parentless roots

Ecological Correlation → is a kind of Correlation

Show alternative path (1)

Neighborhood in Abstraction Space¶

Ecological Correlation sits in a sparse region of the domain-specific corpus (80^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (2769 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-10-25