State Lines, Blurred Findings: The Hidden Cost of Geographic Aggregation in US Data Research

There is a particular kind of analytical confidence that emerges when a dataset is clean, the sample sizes are large, and the geographic unit of analysis is the US state. Fifty observations. Tidy boundaries. Familiar labels. It feels rigorous. In practice, it frequently is not.

The state is one of the most overused units of analysis in American research, and its prevalence reflects convenience far more than it reflects statistical appropriateness. When researchers aggregate individual- or county-level records up to the state level, they are not simply summarizing data — they are discarding variation. In many cases, that discarded variation is precisely where the meaningful signal lives.

What Aggregation Actually Does to Your Data

Aggregation is not neutral. When you compute a state-level mean, you are collapsing a distribution into a single number. The spread of that distribution — its internal heterogeneity — vanishes from the record. Two states can share an identical aggregate value while harboring wildly divergent underlying realities.

Consider diabetes prevalence. If Texas and Colorado both report a statewide adult diabetes rate of approximately 11 percent, a state-level analysis treats them as functionally equivalent. But Texas contains both urban counties in the Dallas-Fort Worth metro area with rates below 9 percent and rural counties in the Rio Grande Valley where rates exceed 17 percent. Colorado similarly spans Denver's urban core and its agricultural eastern plains, where health infrastructure and demographic composition differ substantially. The state average conceals a within-state range that is larger, in many cases, than the between-state range the researcher is trying to study.

This is not a corner case. It is the norm across virtually every domain of US social and health research.

The Ecological Fallacy and Its Practical Consequences

The formal name for the error that follows from this kind of aggregation is the ecological fallacy: the mistaken inference that relationships observed at the group level hold at the individual level, or that aggregate patterns accurately represent the sub-aggregate units from which they were constructed.

The concept dates to W.S. Robinson's 1950 work on literacy and nativity, but its practical relevance has only grown as administrative datasets have made state-level summaries easier to produce and disseminate. The fallacy does not require carelessness. It can emerge from entirely reasonable analytical choices that happen to be made at the wrong scale.

In public health research, ecological fallacy has contributed to misleading findings on the relationship between income inequality and health outcomes, on opioid prescribing patterns, and on vaccine uptake. In each case, state-level correlations pointed in one direction while county- or ZIP-code-level analysis revealed a more complicated — and more actionable — picture.

In economic policy research, state-level employment figures can mask metropolitan-rural divergence that is directly relevant to federal program targeting. A state reporting modest unemployment may contain a mid-sized city experiencing a structural collapse in manufacturing employment, surrounded by rural counties with persistently high joblessness. Neither group is visible in the aggregate.

Why Researchers Reach for the State Level Anyway

Understanding the problem requires acknowledging the pressures that produce it. Several forces systematically push researchers toward state-level aggregation regardless of its analytical appropriateness.

First, many federal datasets are publicly released at the state level, with sub-state data restricted or suppressed for privacy reasons. The American Community Survey, for example, provides reliable single-year estimates only for geographies above certain population thresholds, which often means researchers default to states even when county-level data would serve the research question better.

Second, policy variables — minimum wage laws, Medicaid expansion decisions, state income tax structures — are themselves defined at the state level, creating a natural pull toward matching the unit of analysis to the unit of policy variation. This is sometimes appropriate. It becomes problematic when researchers treat the state-level policy variable as a license to aggregate all outcome variables to the same geography.

Third, statistical power calculations are easier to defend with n=50 than with n=3,000 counties or n=74,000 census tracts, even when the latter provide a more honest representation of the data structure.

A Framework for Choosing the Right Unit Before You Commit

The appropriate unit of analysis should be determined by the research question, not by the availability of a convenient summary file. Before aggregating, researchers should work through the following considerations.

Define the causal mechanism at stake. If the question concerns how a state-level policy affects individual behavior, the individual remains the appropriate unit of outcome measurement, with the state-level variable entered as a covariate. If the question is about state fiscal capacity, the state may be the correct unit throughout.

Assess within-unit heterogeneity before collapsing it. Compute the coefficient of variation for key variables at the sub-state level before aggregating. If within-state variance accounts for a large share of total variance — a threshold of 40 percent or more is a reasonable starting point — state-level aggregation will materially distort any correlation or regression estimated on the resulting data.

Match geographic granularity to the intervention or exposure being studied. Environmental exposures, for instance, operate at the neighborhood or tract level. Aggregating pollution data to the state level and then correlating it with health outcomes is almost always methodologically indefensible, regardless of how many states are in the sample.

Consider multilevel modeling as an alternative to aggregation. Hierarchical linear models allow researchers to simultaneously estimate within-state and between-state effects without discarding sub-state variation. When the data structure is genuinely nested — individuals within counties within states — the model should reflect that structure.

Document the aggregation decision explicitly. Any publication or report that uses state-level summaries should include a clear statement of what sub-state variation has been collapsed, and whether the research question was adequately addressed at that scale. Reviewers and readers are entitled to that information.

When State-Level Analysis Is Defensible

This article is not an argument that state-level analysis is always wrong. There are research questions for which the state is the correct and appropriate unit. Studies of interstate regulatory variation, analyses of state government fiscal behavior, and research on the administrative capacity of state agencies all have legitimate reasons to treat the state as the primary unit of observation.

The problem is not the state as a geographic unit. The problem is the reflexive application of that unit to questions that operate at finer scales, driven by convenience, data availability, or an underestimation of how much analytical damage aggregation inflicts.

The Responsibility of the Analyst

Data professionals working with US geographic data carry a specific responsibility: to resist the pull of the convenient summary and to interrogate whether the unit of analysis they have chosen is actually capable of answering the question they have posed. The state boundary is a political artifact. It has no inherent relationship to the distribution of health outcomes, economic activity, educational attainment, or environmental exposure. Treating it as though it does is not a minor methodological footnote — it is a foundational decision that shapes every inference that follows.

The aggregation trap is not sprung by bad intentions. It is sprung by the accumulated weight of small, defensible-seeming decisions that collectively produce an analysis incapable of reflecting the world it purports to describe. Recognizing that trap — and building the habit of checking for it before the analysis is designed, not after — is among the more valuable skills a data professional in the current research environment can develop.