Canonical definition
A composition model specifies how lower-level properties, interactions, relationships, distributions, or processes combine to support a construct defined at the level of a bounded collective. It establishes the theoretical and analytical relationship between observable evidence and the collective state being inferred.
Claim status: The cross-level composition problem is established in multilevel theory and measurement scholarship. The canonical definition above and its role within CSI are a research-program synthesis informed principally by Chan (1998) and Kozlowski and Klein (2000); the CSI-specific qualification requirements remain original propositions pending empirical validation.
Why composition matters
Collective-State Inference operates across levels. Evidence may originate in individual contributions, dyadic exchanges, network structure, temporal sequences, or contextual conditions, while the target construct exists at the collective level. A composition model explains how those lower-level observations become meaningful evidence of that higher-level condition (Chan, 1998; Kozlowski & Klein, 2000).
The model follows from the construct. CSI does not assume that aggregation is always invalid, nor that relational complexity is always superior. A mean may be theoretically appropriate for one construct and misleading for another. The question is whether the chosen composition rule matches the construct and survives validation against credible alternatives.
Requirements of a defensible composition model
Begin with theory
The composition rule must reflect what the collective state is claimed to mean, not merely what the available data make easy to calculate.
Separate evidence from target
The model must identify the level at which evidence is observed and the level at which the construct is defined.
Permit comparison
The proposed model should be compared with simpler baselines and alternative composition rules.
Preserve boundary conditions
The meaning of the composition rule may depend on task, time, hierarchy, membership, and environment.
Major forms of composition
The additive, consensus, referent-shift, and dispersion distinctions are grounded in established composition-model typologies; relational and process forms extend the logic to configurations and dynamics that may be constitutive of the target construct (Chan, 1998; Morgeson & Hofmann, 1999; Kozlowski et al., 2013).
| Composition form | Core logic | Potential CSI use |
|---|---|---|
| Additive | Member-level values combine to represent the collective level. | Appropriate when the construct reflects the total or average amount of a property and dispersion is not theoretically central. |
| Consensus | A shared collective property requires sufficient within-collective agreement. | Useful for constructs such as shared climate or common interpretation when agreement and referent-shift conditions are met. |
| Dispersion | Variation, disagreement, or polarization is itself meaningful. | Useful when heterogeneity, fragmentation, or fault-line structure defines the collective condition. |
| Relational | The construct depends on who interacts with whom and how relationships are configured. | Useful for cohesion, influence concentration, subgroup structure, reciprocity, brokerage, and coordination patterns. |
| Process or temporal | The collective property develops through sequences, feedback, adaptation, and path dependence. | Useful when escalation, recovery, persistence, volatility, or timing is central to the state. |
| Configurational | Multiple conditions combine in distinct but potentially equivalent patterns. | Useful when no single feature is sufficient and different combinations can support a similar collective state. |
When aggregation may be appropriate
Aggregation can be a legitimate composition rule when theory predicts an additive or consensus-based relationship and the relevant statistical conditions are satisfied. For example, a collective construct may reasonably depend on the average level of member experience when members use a common collective referent and agreement is sufficient (Chan, 1998).
CSI therefore does not reject averages categorically. It rejects unexamined aggregation: calculating a mean and treating it as a validated collective state without establishing why the mean represents the construct, whether disagreement matters, and whether richer alternatives add explanatory value.
When relational or temporal models are necessary
Some collective states cannot be adequately represented by the amount of individual-level evidence alone. Two collectives may have the same average participation yet differ sharply in reciprocity, centralization, subgroup separation, persistence, or recovery. In those cases, the arrangement and sequence of interaction are part of the construct-relevant evidence, consistent with relational accounts of collective structure and dynamic emergence (Moody & White, 2003; Kozlowski et al., 2013).
A relational or temporal model is warranted only when theory predicts that those patterns matter and empirical comparison shows that the added complexity improves construct validity, explanatory power, or generalization.
Illustrative comparison
Suppose two teams report the same average level of confidence in a strategic decision. Team A shows broad agreement, distributed participation, and stable coordination. Team B contains two polarized subgroups, with decisions dominated by a single broker. The shared mean obscures substantively different collective configurations.
A consensus model may characterize Team A as aligned. A dispersion or relational model may characterize Team B as fragmented despite the same average. The correct inference depends on the construct definition and validation evidence, not on the convenience of one summary statistic.
Validation implications
- Compare baselines: test the proposed composition model against means, counts, and context-free alternatives.
- Test agreement where required: consensus constructs need evidence that aggregation is meaningful.
- Test incremental value: relational or temporal complexity should improve prediction, explanation, or construct validity.
- Test boundary conditions: determine when the composition rule changes across tasks, cultures, time windows, or collective structures.
- Preserve uncertainty: competing composition models may remain plausible when evidence is limited.
Common composition errors
- Treating a group average as a collective construct without a referent or agreement test
- Assuming all disagreement is measurement noise rather than a meaningful collective property
- Using a complex network model without demonstrating value beyond simpler summaries
- Mixing individual, dyadic, team, and organizational levels in one unlabeled estimate
- Ignoring changes in membership or observation window that alter the composition process
Relationship to the CSI construct system
The collective establishes the bounded unit. The collective state defines the latent target. Emergence explains how a meaningful higher-level condition can arise. The composition model specifies how lower-level evidence relates to that condition. CSI then estimates the state and evaluates the estimate through collective-level validation.
Selected scholarly foundations
The principal scholarly foundations for this article are Chan (1998) on alternative composition models, Kozlowski and Klein (2000) on multilevel theory, Morgeson and Hofmann (1999) on collective constructs, Kozlowski et al. (2013) on temporal emergence, and Moody and White (2003) on relational cohesion. The CSI-specific requirements and decision rules presented here are original research-program propositions informed by those foundations.
Research status
RA-005 documents the current CSI position on composition models and is aligned with the pre-submission theoretical manuscript. The article remains a foundational draft. Future versions will add construct-specific decision rules, formal tests for alternative composition models, and evidence from the planned operationalization program.
Preferred interim citation
Clark, E. D. (2026). Composition Models. CSI Reference, RA-005, Version 0.2. Collective-State Inference Research Program.
See the Scholarly Sources registry, Editorial and Citation Policy, and Reference version history.