Ecological Studies

Overview

An ecological study analyses groups, places or time periods rather than individual people as the primary units. Exposure and outcome information may be aggregated for districts, countries, hospitals, schools, calendar years or other collective units. Ecological designs are useful for population-level exposures, policy comparisons, surveillance, hypothesis generation and evaluation of broad contextual patterns.[1–3]

The design cannot automatically answer an individual-level question. A district with greater transport disadvantage and lower postpartum follow-up may not show that the individual women experiencing transport disadvantage were the same women who missed testing. Moving between levels without justification creates the ecological fallacy.

Why this matters

Many public-health determinants operate at group level: legislation, air pollution, health-service distribution, food environments, urban design and economic policy. Individual-level studies may be unable to represent these exposures adequately.

At the same time, aggregate data can hide substantial within-group variation. Ecological associations can arise from:

• genuine contextual effects;
• individual composition;
• group-level confounding;
• measurement and coding differences;
• spatial or temporal autocorrelation;
• migration and boundary changes; or
• mathematical coupling between denominators.

Key concepts

The group is the unit of analysis

In a purely ecological analysis, each row might represent one district. Variables are group summaries such as mean travel time, testing coverage and clinic density. The sample size for the primary model is the number of districts, not the number of women whose data contributed to the district summaries.

Aggregate and contextual variables

An aggregate variable summarises individual characteristics, such as percentage uninsured. A contextual variable describes the group environment, such as number of clinics per 100,000 population or a district policy. Some variables combine both interpretations.

Ecological and atomistic fallacies

The ecological fallacy occurs when a group-level association is assumed to apply to individuals. The atomistic fallacy occurs when an individual-level association is assumed to explain group-level differences. Both levels can contain distinct mechanisms.

When to use this approach

DECISION GUIDE

Questions suited to ecological analysis

• Does the exposure operate at policy, environmental, service-system or population level?
• Are groups or time periods the meaningful intervention or decision units?
• Is the objective to compare population rates, monitor trends or generate hypotheses?
• Are definitions and data quality sufficiently comparable across groups?
• Can major group-level confounders and spatial or temporal structures be considered?
• Will conclusions remain at the group level unless multilevel data support more?

Ecological designs are especially useful for natural experiments or policy changes across jurisdictions, but stronger quasi-experimental structures—such as controlled interrupted time series or difference-in-differences—may be needed for causal evaluation.

Research question and methodological fit

A defensible ecological question is:

Across districts, is average travel time to an authorised testing facility associated with district postpartum glucose-testing coverage?

This differs from:

Are individual women with longer travel times less likely to complete testing?

The second question requires individual exposure and outcome linkage, potentially within a multilevel design.

Design, setting and study population

DESIGN PROFILE

Multiple-group ecological study

Core structure: Group-level exposure and outcome summaries are compared across several populations during a defined period.

Best suited for: Contextual determinants, policies, service distribution, surveillance and population comparisons.

Important limitations: Ecological fallacy, group-level confounding, small number of units, heterogeneity, autocorrelation and data comparability.

Other ecological structures include:

• time-trend studies within one population;
• mixed time-and-place studies;
• spatial studies;
• ecological comparisons before and after policy change; and
• multilevel studies combining individual and group data.

Sampling and data sources

Groups may form a census of jurisdictions rather than a sample. This does not remove uncertainty about measurement or the broader inferential target. If only 12 districts exist, complex models with many predictors are not supported merely because each district contains thousands of residents.

PROTOCOL CHECKLIST

Ecological data-source specification

• Define group boundaries and the study period.
• State whether groups are a census or sample of a broader set.
• Verify comparability of outcome numerators and denominators.
• Describe exposure construction and geographic resolution.
• Address boundary changes, migration and missing jurisdictions.
• Prespecify contextual confounders and temporal or spatial structures.
• Record data provenance, revisions and linkage procedures.

Variables, constructs and measurements

Rates should use compatible populations and time windows. Crude counts can mislead when district population sizes differ. Standardisation may be needed for age or other structural differences.

Ecological dataset structure

illustrative

Table 1. One row per district

district_id	testing_coverage_pct	mean_travel_time_min	clinics_per_100000	urban_pct	median_income
D01	64.2	18	3.8	82	4,900
D02	41.7	47	1.9	35	3,100
D03	55.8	31	2.4	58	3,700

Note: Values are illustrative. The analytical sample contains three districts in this excerpt, not the women contributing to the percentages.

Exposure misclassification can occur when an average masks distribution. Two districts with the same mean travel time may have very different proportions living more than one hour away.

Types of ecological comparison

Different ecological structures support different interpretations:

• Multiple-group comparison: several areas are compared during one period.
• Time-trend study: one population is observed across several time points.
• Mixed study: several populations are observed across time.
• Spatial study: geographic pattern, proximity or spatial dependence is central.
• Ecological intervention or policy study: group outcomes are compared around a population-level change.

A simple before-and-after ecological comparison is vulnerable to secular trends, concurrent policies and changes in measurement. Where the aim is policy evaluation, controlled interrupted time-series, difference-in-differences or synthetic-control methods may offer a more explicit counterfactual structure.

The ecological literature describes legitimate uses including population monitoring, evaluation of group-level exposures and generation of hypotheses.[4] Multilevel analysis is conceptually distinct because it combines variables measured at more than one level and keeps individuals linked to their contexts.[5]

Denominators, standardisation and uncertainty

A district percentage is itself an estimate. Districts with small denominators may have much more sampling or random variation than large districts. Scatterplots should display denominator size or uncertainty where useful, and models should not treat all group estimates as equally precise without justification.

Rates require compatible numerators and denominators. Hospital events assigned by treatment location should not be divided by residential populations without considering cross-boundary care. Exposure and outcome periods should allow a plausible lag, particularly for environmental or policy effects.

Standardisation can reduce confounding by population composition, but an age-standardised district rate remains a group-level summary. If the standard population or age bands change across releases, trend comparisons may be artefactual.

Spatial smoothing can stabilise noisy small-area rates, but smoothed maps are model-based estimates. The method, neighbourhood structure and uncertainty should be reported, and users should not confuse apparent visual precision with directly observed data.

Procedures and data management

Data from administrative systems should be harmonised before analysis. Investigators should compare definitions, collection systems, denominator updates and missingness across jurisdictions. Maps and rates must use the same boundary version.

Where small counts may reveal individuals, suppression or aggregation rules are needed. Repeated release of overlapping small cells can still enable disclosure.

Bias, validity and quality assurance

Other biases include:

• confounding by urbanisation or deprivation;
• different coding practices by district;
• numerator-denominator mismatch;
• migration between exposure and outcome periods;
• spatial spillover between neighbouring districts;
• temporal trends common to all groups; and
• modifiable areal unit problems, where results change with geographic aggregation.

APPRAISAL CHECKLIST

Appraising an ecological study

• Is the question explicitly group-level?
• Are exposure, outcome and covariates measured for compatible groups and periods?
• Is the number of independent groups sufficient for the planned model?
• Were population size, standardisation and uncertainty addressed?
• Were spatial or temporal dependencies considered?
• Do conclusions avoid unsupported individual-level interpretation?

Visualisation and interpretation

Maps are powerful but can imply patterns that depend on class intervals, colour scales and boundary size. Display denominators or uncertainty alongside rates, use accessible palettes and avoid ranking small areas when differences are mostly random. A scatterplot with labelled influential groups may reveal that an apparent association is driven by one capital district or data-quality outlier.

Outlying groups should be investigated rather than automatically removed. They may represent coding differences, unique policy contexts or genuine high-leverage observations. Any exclusion or alternate model should be reported as a sensitivity analysis with substantive justification.

Ethical and governance considerations

Interface with statistical analysis

Descriptive analysis should map or graph group-level exposure and outcome with denominators and uncertainty. Weighted regression may account for differing precision of group estimates, but weighting choices should match the estimand. Spatial or temporal correlation can invalidate ordinary independence assumptions.

The ecological correlation is not an individual correlation. Model coefficients describe changes in group outcomes associated with group exposures. Standardisation and multilevel approaches are separate methodological decisions.

Multilevel studies can separate but not automatically identify levelsMMed extension

Linking individual women to district characteristics permits individual and contextual variables in one model. This can reduce ecological ambiguity, but causal interpretation still depends on correct timing, confounding control, measurement and sufficient variation at each level. A multilevel model is not a universal cure for ecological bias.

Ecological analyses should also distinguish exposure assignment from outcome residence. A policy may be implemented by workplace, school, treatment facility or home district; selecting the wrong geography can misclassify the relevant context. When people routinely cross borders for care or employment, flow data and sensitivity analyses using alternative geographic definitions may be necessary.

Where group definitions are policy-relevant, stakeholder review can identify boundary changes, referral patterns and local data anomalies that are not visible in the analytical file.

Worked example

WORKED EXAMPLE

A bounded district-level analysis

The team defines 25 districts using 2027 administrative boundaries. Testing coverage is calculated from eligible deliveries and documented tests in the same year. Travel time is estimated from populated grid cells to the nearest authorised facility and summarised as both mean time and percentage of residents beyond 45 minutes.

The primary analysis relates district testing coverage to the percentage beyond 45 minutes, adjusting for urbanisation, median income and clinic density. Confidence intervals reflect the uncertainty of district coverage estimates, and residual spatial pattern is assessed. The conclusion states that districts with poorer geographic access had lower coverage, not that individual travel time caused non-completion.

Suggested protocol or reporting wording

Common mistakes

Knowledge check

A district with more clinics has lower mortality. Can we conclude that individuals living nearer a clinic have lower mortality?

Answer and reasoning

No. The observed association is between district summaries. Individual distance and mortality are not paired, and district-level confounding or composition may explain the pattern.

What is the sample size in a regression with one row for each of 18 districts?

Answer and reasoning

For the primary ecological model, the analytical sample size is 18 districts, even if the rates were calculated from many residents. The precision of each district estimate may vary, but residents are not independent rows in that model.

Why can geographic aggregation change an ecological result?

Answer and reasoning

Different boundaries combine populations and exposures differently. A relationship seen across large districts may weaken, strengthen or reverse when smaller areas are used—the modifiable areal unit problem.

What wording preserves the correct level of inference?

Answer and reasoning

State that districts with greater exposure had higher or lower district outcome rates. Do not substitute “people with greater exposure” unless individual-level data and analysis support that conclusion.

Key take-home messages

SUMMARY

Key take-home messages

• Ecological studies analyse groups, places or times as the primary units.
• They are valuable for contextual exposures, policies and population surveillance.
• Group-level associations do not automatically describe individual-level relationships.
• Data comparability, denominators, boundaries and autocorrelation require explicit attention.
• Conclusions must remain at the level supported by the design.

References

1. Morgenstern H. Ecologic studies in epidemiology: concepts, principles, and methods. Annu Rev Public Health. 1995;16:61–81. https://doi.org/10.1146/annurev.pu.16.050195.000425
2. Greenland S, Robins J. Ecologic studies—biases, misconceptions, and counterexamples. Am J Epidemiol. 1994;139(8):747–760. https://doi.org/10.1093/oxfordjournals.aje.a117069
3. Wakefield J. Ecologic studies revisited. Annu Rev Public Health. 2008;29:75–90. https://doi.org/10.1146/annurev.publhealth.29.020907.090821
4. Morgenstern H. Uses of ecologic analysis in epidemiologic research. Am J Public Health. 1982;72(12):1336–1344. https://doi.org/10.2105/AJPH.72.12.1336
5. Diez Roux AV. A glossary for multilevel analysis. J Epidemiol Community Health. 2002;56(8):588–594. https://doi.org/10.1136/jech.56.8.588