The ecological fallacy
Why the sign can flip
Group averages discard the within-group variation, which is where the individual-level relationship lives. Countries with higher average fat intake can have higher average cancer rates while, inside each country, the individuals eating most fat are not the ones getting cancer. Nothing about this requires error or confounding; it follows from aggregation itself.
What a reversed trend looks like
Why adjustment does not rescue it
Adjusting for the group mean of a confounder controls confounding only if every individual-level relationship in the model is linear. Risk models in epidemiology are rarely linear, so the ecological estimate remains biased after adjustment, and adding more group-level covariates does not converge on the individual answer.
When group-level analysis is legitimate
- When the EXPOSURE is genuinely a group property: a law, a tax, a policy, an air quality level that everyone in the area shares.
- When the question is about the group: does this region's programme reduce this region's rate?
- For hypothesis generation, stated as such, where individual data do not exist.
- For monitoring trends over time, where no individual-level claim is made.
- In multilevel designs where individual and group data are analysed together, and the two levels are kept distinct.
The mirror image
The atomistic fallacy runs the other way: inferring a group-level effect from individual data, which fails when an intervention changes the whole environment. A vaccine's individual efficacy does not tell you a population's benefit once herd effects are included.
What to watch out for
The fallacy is not that group-level data are noisy; it is that the group-level association can have the OPPOSITE SIGN to the individual one, and can do so even with no confounding and no measurement error at all. Adding covariates does not fix it: adjusting an ecological analysis for group-level averages of confounders leaves bias behind whenever the individual-level relationship is nonlinear, which it usually is.
Clinical example
Counties with higher average radon levels show lower lung cancer rates. The individual-level relationship is the opposite, and it is the aggregation, together with the nonlinear individual risk model and the distribution of smoking, that produces the reversal.
- Invited commentary: ecologic studies--biases, misconceptions, and counterexamples (opens in a new tab) · Am J Epidemiol · doi 10.1093/oxfordjournals.aje.a117069
- The potential for bias in Cohen's ecological analysis of lung cancer and residential radon (opens in a new tab) · J Radiol Prot · doi 10.1088/0952-4746/22/2/302