Aggregate evidence: do not infer individual behavior
A relationship between group averages does not by itself identify a relationship between individuals within those groups. Before writing that customers, workers or firms behaved in a particular way, check the unit of analysis. The missing individual cross-tabulation cannot be reconstructed from two group percentages without additional assumptions or evidence.
Match the claim to the unit
Suppose a research memo compares regions with high digital-service adoption and regions with strong household spending. That can describe a regional association. It does not establish that the digital-service users are the households increasing their spending. Different people inside each region may account for the two aggregate measurements.
The CDC discussion of ecological studies explains, in its epidemiological setting, why aggregate associations may not apply to individual cases. The same logical issue motivates the hypothetical business example here. CDC is not endorsing an investment method, and this article does not transfer any medical finding to financial behavior.
A hypothetical town with fixed totals
Imagine a town of 100 households. Sixty use a digital service and sixty increased their spending. Let x be the number who did both. The intersection can range from 20 to 60: at least twenty overlap because the two totals sum to 120, but the aggregate totals do not tell you where within that range the true overlap lies.
| Possible allocation | Users who increased spending | Nonusers who increased spending | Association within town |
|---|---|---|---|
| x = 20 | 20 of 60 | 40 of 40 | Users less likely to increase |
| x = 60 | 60 of 60 | 0 of 40 | Users more likely to increase |
Both allocations reproduce sixty users and sixty spending increases. The direction of the individual association reverses while those aggregate totals remain identical. Even knowing the direction of a regional correlation would not supply this missing joint distribution inside the town.
Derive the feasible intersection
For a population of N with A members in one category and B in another, the overlap is bounded by max(0, A + B - N) and min(A, B). These are arithmetic feasibility bounds, not confidence intervals. They show what the margins permit, not which allocation is more probable or what caused any household’s decision.
Additional cross-tabulations can narrow the possibilities, but check that they refer to the same people and period. An adoption survey from one sample and spending records from another cannot automatically be joined person by person. Similar region names or matching population totals do not establish record-level identity.
Avoid the reverse overreaction
Group-level evidence is not worthless. A regional policy, infrastructure constraint or local market condition may be a group-level research target. The problem is changing the unit of the conclusion without supporting data. Specify whether the claim concerns a region, an institution, a company or an individual account, and use evidence suited to that level.
Likewise, individual data alone may not fully represent group mechanisms. Preserve relevant context instead of insisting that every useful question must be reduced to a person-level comparison. The goal is alignment between the estimand and the observation, not a universal preference for the smallest available unit.
Design the next evidence request
Ask what linked information would distinguish the competing allocations: a matched panel, a joint survey question or a carefully defined administrative cross-tabulation, subject to appropriate access and privacy constraints. Do not claim that a model has recovered individual behavior merely because it generates plausible-looking rows from aggregate margins.
Even with individual observations, selection, timing and confounding remain separate concerns. Finding that users increased spending more often would establish an association under that sample’s definitions, not automatically a causal effect of the service. Keep the missing-joint-data problem and the causal-identification problem distinct.
Write a bounded research conclusion
State the observed group relationship, identify the individual inference that remains unresolved and describe which evidence could resolve it. In the example, the two sixty-percent figures support no unique claim about users’ spending. That narrower conclusion is useful because it prevents a regional chart from being mistaken for proof of customer behavior.
FAQ
Does matching two regional percentages identify the same people?
No. Marginal percentages do not supply the overlap. You need linked observations or defensible additional information.
Should aggregate evidence always be rejected?
No. It can directly answer group-level questions. Keep the unit of the conclusion consistent with what was actually measured.