03 · What You Need to Know
The Individual Is Not Always the Correct Scale for Explaining a Collective Outcome
Start with an individual-level relationship
Suppose faculty members are nested within universities.
The study asks:
Are faculty members with greater AI self-efficacy more likely to adopt generative AI?
Both variables belong to individuals.
If the analysis finds a positive relationship, the supported conclusion is:
Within the population and model studied, faculty with greater self-efficacy tend to report greater personal adoption.
That is an individual-level finding.
The atomistic fallacy occurs when that relationship is transferred upward
Suppose researchers then state:
Universities with higher faculty self-efficacy will therefore have higher institutional AI adoption.
The conclusion now concerns universities.
The level has changed without directly estimating the university-level relationship.
Supported individual conclusion
Faculty with greater self-efficacy report greater personal AI adoption.
Atomistic inference error
Therefore, universities with higher faculty self-efficacy necessarily have greater institutional AI adoption.
Why is the upward inference unsafe?
Because group outcomes may depend on properties that have no direct individual counterpart.
University-level adoption can depend on:
- institutional policy;
- procurement systems;
- technology infrastructure;
- leadership priorities;
- governance structures;
- budget allocation;
- collective norms;
- coordination across departments.
These are not simply personal characteristics averaged across faculty.
Groups can have emergent properties
Some characteristics exist only because people interact within a collective.
Examples include:
- team cohesion;
- organizational culture;
- network density;
- institutional governance;
- collective norms;
- organizational centralization.
Knowing every member's individual attitude may not fully determine these collective properties.
This is one reason multilevel theory distinguishes lower-level characteristics from higher-level constructs.
The fallacy is the reverse of ecological fallacy
| Fallacy |
Observed relationship |
Unsupported inference |
|
Ecological fallacy
|
Group level |
Individual level |
|
Atomistic fallacy
|
Individual level |
Group level |
The ecological fallacy moves downward.
The atomistic fallacy moves upward.
Both involve transferring a finding across levels without adequate evidence.
The atomistic fallacy receives less attention, but the problem is equally real
Subramanian and colleagues revisited Robinson's classic ecological-correlation work and argued that overemphasizing ecological fallacy can encourage the opposite mistake: assuming that individual-level analysis is inherently the only meaningful form of analysis. Their multilevel reanalysis highlighted that individual relationships can depend importantly on higher-level context.
Recent methodological discussions likewise describe ecological and atomistic fallacies as opposite forms of invalid cross-level generalization.
An individual-level coefficient answers an individual-level question
Suppose:
Faculty self-efficacy → Faculty adoption
has a positive coefficient.
This tells you something about variation among faculty members under the specified model.
It does not automatically tell you:
University average self-efficacy → University adoption rate.
The second relationship needs to be estimated at the university level or within an appropriate multilevel structure.
The aggregate of an individual characteristic becomes a different variable
Suppose individual self-efficacy is Xᵢⱼ.
The university mean is:
Xᵢⱼ and X̄ⱼ are related but not interchangeable.
One describes an individual.
The other describes a feature of the group's observed composition.
Within-group and between-group relationships can differ
Suppose within every university, more self-efficacious faculty adopt AI more frequently.
At the university level, however, universities with greater average self-efficacy might not have greater adoption.
Why?
Perhaps highly self-efficacious faculty are concentrated in institutions with restrictive governance.
Perhaps institutions with lower average confidence provide mandatory training and powerful infrastructure that increase adoption.
Context can reshape the between-university pattern.
This is why the same relationship can differ across levels of analysis.
The individual relationship could even point in the opposite direction from the group relationship
Imagine that within universities:
Higher workload → Greater individual burnout.
Yet across organizations, those with higher average workloads might have lower average burnout because high-workload organizations also provide better pay, staffing, autonomy, or support.
Both patterns can coexist.
The group relationship cannot be deduced from the individual coefficient.
Groups are not merely arithmetic averages of members
Some group outcomes are constructed directly from individual behavior, such as:
percentage of faculty adopting AI.
Even then, the relationship between group-average X and group-average Y does not necessarily equal the individual relationship between X and Y.
Aggregation changes the source of variation.
Other group outcomes, such as governance quality or organizational culture, are even less reducible to individual averages.
Composition and context jointly shape groups
Suppose universities differ in adoption.
Some of that difference may reflect composition:
Universities contain different types of faculty.
Some may reflect context:
Universities provide different policies, resources, structures, and incentives.
A purely individual-level analysis can overlook the contextual part of the explanation.
Individual characteristics may have different consequences in different contexts
Suppose AI self-efficacy predicts adoption strongly in universities with excellent infrastructure but weakly in universities where approved tools are unavailable.
The individual-level relationship is conditional on institutional context.
A pooled individual analysis may therefore conceal meaningful differences among groups.
This is an example of a cross-level interaction.
The atomistic fallacy can arise from methodological individualism
A study may assume that all collective outcomes can be explained by individual characteristics.
For some questions, individual composition is indeed central.
But automatically reducing organizational, neighborhood, or societal phenomena to individuals can omit contextual mechanisms.
Subramanian and colleagues argued that understanding individual outcomes may itself require attention to the contexts in which individuals are embedded.
A group-level theory requires group-level constructs
Suppose researchers want to explain why universities differ in innovation.
Individual faculty creativity may matter.
But a theory of university innovation might also involve:
- institutional strategy;
- resource allocation;
- collaborative structure;
- leadership;
- incentive systems;
- external partnerships.
An analysis limited to individual creativity may therefore answer only part of the question.
An individual predictor can still contribute to a group explanation
Avoiding the atomistic fallacy does not mean individual characteristics are irrelevant to groups.
Individual behaviors can aggregate or combine into collective outcomes.
For example:
Faculty adoption decisions → Institutional adoption rate
But researchers should specify how that bottom-up composition works.
Some bottom-up relationships are definitional
If institutional adoption rate is defined as the percentage of faculty adopting AI, individual adoption mathematically composes the group outcome.
This relationship is largely definitional.
By contrast:
Individual faculty experimentation → Emergence of an institutional innovation culture
is a substantive theory of emergence.
The second claim requires evidence about interactions, shared norms, institutionalization, and temporal development.
Aggregation alone does not demonstrate emergence
Suppose researchers average personal innovation scores and call the result “organizational innovation culture.”
They have created an aggregate.
They have not necessarily measured a shared culture.
The transition from individual-level data to a group-level construct requires a composition argument.
Shared group constructs require more than individual internal consistency
If organizational climate is measured from employee surveys, researchers may need evidence about:
- the group referent;
- within-group agreement;
- between-group variation;
- reliability of group scores;
- the number of respondents per group.
A scale's reliability among individuals does not by itself establish a valid organizational-level construct.
The atomistic fallacy can appear in intervention recommendations
Suppose individual self-efficacy predicts AI adoption.
Researchers conclude:
Universities should therefore improve institutional adoption by providing self-efficacy training.
That recommendation may be reasonable, but the individual association alone does not establish that changing faculty self-efficacy will change the university-level outcome.
The intervention may encounter:
- policy restrictions;
- technical barriers;
- procurement constraints;
- leadership resistance;
- lack of integration into institutional systems.
A group-level intervention claim requires more than an individual-level association.
Individual causation does not automatically imply an aggregate causal effect of the same magnitude
Even if changing individual X truly affects individual Y, scaling that intervention across an organization can introduce:
- spillovers;
- resource constraints;
- peer effects;
- saturation;
- organizational adaptation;
- feedback processes.
The collective outcome may therefore differ from a simple multiplication of individual effects.
Interaction among individuals can produce non-additive group outcomes
Suppose team performance depends on coordination.
Adding one highly skilled person may improve performance in one team but disrupt another if communication and role structures differ.
Group outcomes can therefore be nonlinear functions of individual characteristics.
Distribution can matter more than the mean
Two organizations can have the same average competence but different distributions.
Organization A:
everyone has moderate competence.
Organization B:
half have very high competence and half have very low competence.
The same average may produce different collective functioning.
If collaboration requires every member to meet a minimum level, the weakest-member distribution may matter more than the mean.
Network position can matter too
Suppose only a few employees have advanced AI expertise.
If those employees occupy central mentoring or leadership positions, they may influence organizational adoption substantially.
If they are isolated, the same number of experts may have little institutional effect.
Individual attributes alone do not capture where those individuals sit in the group structure.
Institutional processes can suppress strong individual tendencies
Faculty may personally want to adopt AI but face institutional prohibition.
Employees may be highly innovative but work under rigid procedures.
Students may be strongly motivated but lack access to resources.
These examples demonstrate why individual tendencies cannot always be projected upward into collective outcomes.
Institutional processes can also amplify weak individual tendencies
Strong support systems may allow people with only moderate initial confidence to adopt new practices.
Thus, a group can achieve high aggregate adoption even when the average individual predisposition is not especially strong.
The group context changes what individual characteristics become behaviorally consequential.
Sampling only one organization cannot establish an organizational relationship
Suppose you survey 5,000 employees from one university.
You can learn a great deal about individual variation within that institution.
You cannot estimate how universities differ from one another because you observed only one university.
Watch Out
Many individuals inside one or a few groups provide rich lower-level information but little or no replication for estimating higher-level relationships.
Even several groups may provide limited higher-level information
Suppose 3,000 faculty members are sampled from eight universities.
Individual coefficients may be estimated with considerable precision.
A university-level relationship still depends on variation across only eight universities.
Higher-level inference therefore requires its own sample-size reasoning.
A group-level hypothesis should be written with group-level nouns
Compare:
Faculty with greater self-efficacy report greater adoption.
with:
Universities with higher average faculty self-efficacy have greater institutional adoption.
The first is individual-level.
The second is group-level.
Writing both explicitly makes it obvious that one does not substitute for the other.
A multilevel design can estimate both relationships
Suppose faculty are nested within universities.
Researchers can include:
Individual self-efficacy relative to the university mean
and:
University-average self-efficacy
as separate predictors.
This allows the data to reveal whether the within-university and between-university relationships differ.
A simple within-between model makes the distinction visible
That pattern directly demonstrates why the atomistic inference would have been unsafe.
Multilevel modeling does not make every group conclusion causal
Estimating a between-university coefficient does not automatically identify a causal institutional effect.
Universities may differ in unmeasured ways, faculty may select into institutions, and institutional variables may themselves be consequences of earlier adoption.
Level alignment solves one problem. Causal identification remains another.
Temporal sequencing is especially important for emergence claims
If researchers argue that individual behavior creates organizational culture, they should ideally observe:
individual behaviors first → collective processes develop → later group-level outcome
A one-time cross-sectional survey provides limited evidence about such bottom-up temporal dynamics.
The atomistic fallacy is not the same as generalizing from a sample to a population
Statistical generalization asks whether findings from sampled individuals apply to a broader population of individuals.
Atomistic fallacy concerns something different:
changing the level of the inference from individuals to groups.
A perfectly representative sample of individuals can still fail to establish a group-level relationship.
It is also different from individual-level confounding
An individual relationship may be confounded.
Even if confounding were perfectly addressed, the resulting individual causal effect would not automatically determine a separate group-level association involving group averages or collective outcomes.
Level-of-analysis reasoning remains necessary.
Policy decisions often need evidence at several levels
If the intervention can operate on individuals and institutions, a multilevel evidence base is more informative.
For example:
Individual evidence: Does training improve faculty capability?
Institutional evidence: Do governance and infrastructure improve university adoption?
Cross-level evidence: Does infrastructure help trained faculty translate capability into actual classroom use?
These questions address different parts of the same implementation problem.
Do not respond to ecological fallacy by making the opposite error
Because ecological fallacy is widely taught, researchers sometimes conclude that only individual-level relationships are trustworthy.
That position is too strong.
Subramanian and colleagues' reanalysis of Robinson's classic data emphasizes the value of multilevel thinking and argues that individualistic inference can also produce a distorted account when contextual differences are ignored.
Individual and contextual explanations are often complementary
Rather than asking:
Is the outcome caused by individuals or organizations?
a better question may be:
How do individual characteristics and organizational contexts jointly contribute to the outcome?
This does not mean every study needs multilevel modeling. It means the level of explanation should match the phenomenon being claimed.
