03 · What You Need to Know
Unit and Level Answer Two Related but Different Questions
What is the unit of analysis?
The unit of analysis is the specific type of entity represented by the cases your study analyzes and about which it seeks to make substantive claims.
Common units include:
- individual students;
- teachers;
- employees;
- households;
- classrooms;
- schools;
- organizations;
- countries;
- research articles;
- events;
- dyads or relationships.
If your research question asks whether students with greater self-efficacy report greater engagement, the unit of analysis is the student.
If it asks whether universities with stronger AI governance have greater institutional adoption, the unit is the university.
That is the same underlying question addressed when determining what your study is actually studying.
What is the level of analysis?
The level of analysis refers to the broader scale or hierarchical location at which the phenomenon is examined.
Common broad levels in social research include:
-
micro or individual level;
-
group or team level;
-
organizational or institutional level;
-
community level;
-
societal, national, or macro level.
The labels vary by discipline. Sociology often uses micro, meso, and macro. Organizational research may speak of individual, team, and organization levels. Education may distinguish students, classrooms, schools, and education systems.
The central idea is that level describes where the phenomenon sits within a larger hierarchy or scale.
One level can contain many different units of analysis
This is the easiest way to see why the concepts are not identical.
| Unit of analysis |
Possible level of analysis |
| Student |
Individual |
| Teacher |
Individual |
| Employee |
Individual |
| Patient |
Individual |
| Classroom |
Group |
| Work team |
Group |
| Household |
Group |
| University |
Organizational |
| Hospital |
Organizational |
| Company |
Organizational |
| Country |
Societal or macro |
Student and employee are not the same unit, but both are located at the individual level.
Classroom and work team are not the same unit, but both may occupy a group level.
The distinction is partly disciplinary
Not all research traditions use these terms identically.
Some scholars use unit of analysis and level of analysis almost interchangeably. Others distinguish them more sharply, using unit for the specific entity and level for the scale at which that entity is situated.
Because usage varies, a manuscript should not rely on terminology alone when the distinction affects the design.
Watch Out
There is no benefit in turning terminology into a vocabulary contest. If your field uses “level of analysis” differently, follow disciplinary conventions but state clearly what entities are analyzed and what hierarchical levels your variables and conclusions represent.
A research question can reveal both the unit and level
Consider:
Do faculty members with greater AI self-efficacy adopt generative AI more frequently?
Unit of analysis: faculty member
Level of analysis: individual
Now consider:
Do universities with stronger AI policies have higher institutional adoption rates?
Unit of analysis: university
Level of analysis: organizational
The nouns identify the units. Their place in the broader hierarchy identifies the levels.
The unit of observation is a different concept again
Suppose individual faculty members complete a questionnaire about institutional AI support, and responses are aggregated to characterize each university.
Faculty members are units of observation.
Universities may be the units of analysis.
The analysis occurs at an organizational level.
Thus, three concepts can be distinguished:
| Concept |
Question |
Example |
|
Unit of observation
|
Where does the information come from? |
Faculty member |
|
Unit of analysis
|
What specific entity is being compared? |
University |
|
Level of analysis
|
At what broader hierarchical scale is that entity analyzed? |
Organizational level |
This is why the distinction between unit of analysis and unit of observation should remain separate from the level-of-analysis question.
Levels become especially visible in nested data
Consider:
Students → classrooms → schools
The units are students, classrooms, and schools.
The corresponding levels might be described as:
Level 1: student
Level 2: classroom
Level 3: school
Here, researchers sometimes use the specific unit names themselves as level labels.
This is another reason the terminology can blur in practice.
What matters is that each layer is recognized and that variables are attached to the correct layer.
Individual-level variables describe individuals
Examples include:
- student motivation;
- faculty self-efficacy;
- employee job satisfaction;
- individual age;
- personal technology adoption;
- patient symptoms.
These variables vary among individual people.
If faculty self-efficacy is measured separately for each faculty member, it is an individual-level variable even though faculty members work inside universities.
Group-level variables describe groups
Examples include:
- class size;
- team composition;
- organizational policy;
- school resources;
- university governance structure;
- national research expenditure.
These variables describe properties of higher-level entities rather than personal attributes of each member.
The distinction between individual-level and group-level variables therefore follows naturally from the concept of analytical levels.
A variable's level is part of its theoretical meaning
Consider self-efficacy and collective efficacy.
Individual self-efficacy: a person's belief in their own capability.
Collective efficacy: a group's shared belief in its collective capability.
The variables are related conceptually but are not merely the same score placed at different levels.
Likewise:
Individual perception of support
and:
organizational support climate
represent different constructs unless the theory and measurement explicitly connect them.
Level is therefore not only a statistical issue. It helps define what the construct means.
The same word can hide different levels
“Performance” could mean:
- an individual's job performance;
- a team's collective performance;
- an organization's financial performance;
- a country's economic performance.
“Adoption” could mean:
- whether an individual faculty member uses AI;
- whether a department has adopted an AI platform;
- whether a university has institutionalized AI practices.
Researchers should therefore define the entity and level rather than assuming the variable name makes them obvious.
Level of theory should match level of variables
If a theory proposes that university policy shapes institutional adoption, both constructs are organizational-level.
If it proposes that faculty confidence shapes personal adoption, both constructs are individual-level.
If it proposes that university policy shapes faculty behavior, the theory crosses organizational and individual levels.
Making those levels explicit helps prevent conceptual frameworks from connecting variables whose meanings do not align.
Level of measurement should also match the construct
Suppose researchers want to measure organizational AI-support climate.
They survey individual faculty members.
That is not automatically a problem because individuals can provide information about higher-level constructs.
But the measurement should clearly refer to the organizational phenomenon.
Compare:
“I personally receive adequate AI support.”
with:
“Faculty members in this university receive adequate support for using AI.”
The first item emphasizes an individual experience. The second explicitly uses the university as the referent.
Data source and construct level can therefore differ
A manager can provide information about firm strategy.
A teacher can provide information about school climate.
A student can provide information about classroom practices.
The respondent exists at the individual level, but the construct may exist at a higher level.
The important requirement is conceptual and measurement alignment.
Aggregating lower-level responses changes the analytical level
Suppose students rate classroom climate and responses are averaged within classrooms.
Individual student ratings are lower-level observations.
The classroom mean is a group-level variable if the construct and aggregation are justified.
This is the process addressed when researchers aggregate individual-level data to the group level.
Changing levels changes the source of variation being compared
This is crucial.
At the individual level, researchers compare one person with another.
At the group level, researchers compare one group with another.
Those are different statistical contrasts.
Suppose individual employees with higher autonomy are more satisfied than colleagues with lower autonomy.
That does not automatically imply that organizations with higher average autonomy have higher average satisfaction.
The first result compares people. The second compares organizations.
A relationship can differ across levels
The association between two variables at the individual level may be stronger, weaker, absent, or even reversed at the group level.
This occurs because aggregation changes what variation is represented.
Individual-level variation reflects differences among people within and across groups.
Group-level variation reflects differences among group averages or characteristics.
This is why a relationship can be different at different levels of analysis.
Ecological fallacy is a level-of-analysis error
Suppose countries with higher internet access have higher average educational achievement.
That country-level relationship does not establish that individuals with greater internet access necessarily achieve more within those countries.
Inferring the individual relationship directly from the national relationship risks the ecological fallacy.
The problem is not merely that the variables were aggregated. The conclusion has moved from one level of analysis to another without sufficient evidence.
The atomistic fallacy moves in the opposite direction
Suppose individual faculty members with greater AI self-efficacy report greater AI adoption.
That does not automatically establish that universities with higher average faculty self-efficacy will have greater organizational AI institutionalization.
Moving from an individual-level relationship to a group-level conclusion can produce an atomistic or individualistic fallacy.
Both errors arise from treating evidence from one level as though it automatically applies at another.
A cross-level relationship deliberately connects levels
Consider:
University infrastructure → Faculty AI adoption
The predictor is organizational-level.
The outcome is individual-level.
The theory therefore makes a deliberate cross-level claim.
This is legitimate when the relationship is clearly specified and the analysis respects the hierarchy.
It is the defining issue in a cross-level relationship.
Cross-level moderation involves a level changing a lower-level relationship
Suppose faculty self-efficacy predicts personal AI adoption more strongly in universities with extensive infrastructure.
The focal relationship:
Self-efficacy → Adoption
exists at the individual level.
The moderator:
University infrastructure
exists at the organizational level.
The study therefore asks whether a higher-level characteristic changes a lower-level relationship.
This is more complex than simply including both variables in the same regression equation.
Mixed-level research questions should be written explicitly
Consider:
Does institutional support affect faculty adoption?
The wording is ambiguous.
Does “institutional support” mean an individual's perception of support?
Does it mean a university-level support climate?
Does “faculty adoption” mean average adoption across a university or each faculty member's behavior?
A clearer formulation would state:
Do faculty members working in universities with stronger AI-support climates report greater individual AI adoption?
This makes the organizational predictor, individual outcome, and cross-level structure visible.
Such clarification helps avoid the problems that arise when research questions mix individual- and group-level explanations without specifying their levels.
Multilevel models can preserve variation at several levels
When lower-level units are nested within higher-level units, multilevel models can partition variation across those levels.
For example, faculty adoption may vary because:
- faculty members differ from one another within universities;
- universities differ from one another;
- individual relationships vary across universities.
A multilevel model can represent these sources rather than forcing all variation into one level.
A simple variance decomposition illustrates the idea
This distinction is one reason ignoring hierarchy can misrepresent both coefficients and uncertainty.
The intraclass correlation can help describe clustering
In a simple random-intercept context, researchers may estimate an intraclass correlation coefficient, or ICC, describing the proportion of outcome variation associated with differences between clusters.
An ICC does not determine the theoretical level of a construct by itself, but it can reveal that clustering is empirically consequential.
A small ICC does not automatically mean level is irrelevant
Even modest clustering can matter for standard errors, and a theoretically important group-level predictor can still be relevant when the unconditional between-group variance is not enormous.
Researchers should not decide whether a level exists only from one threshold applied to an ICC.
Theoretical level and empirical variance are related questions, not the same question.
The number of units at each level matters
Suppose a study includes 4,000 students nested within only six schools.
There is substantial individual-level information but very limited information for comparing schools.
A school-level predictor cannot be estimated with the same informational basis as an individual-level predictor merely because thousands of student observations exist.
The effective information for higher-level questions depends strongly on the number and variability of higher-level units.
Level of analysis affects sample-size planning
If your primary hypothesis concerns individuals, you need adequate information at the individual level while accounting for clustering.
If it concerns organizations, you need enough organizations.
If it concerns a cross-level interaction, you need adequate variation at both levels.
There is no single overall N that captures every requirement in a multilevel study.
Level of analysis is not simply determined by statistical software
Running a multilevel model does not automatically make the research multilevel theoretically.
You might use a mixed-effects model solely to account for clustering while still asking an entirely individual-level substantive question.
Conversely, a study may pose a genuinely organizational-level theory but analyze individual data incorrectly with a single-level regression.
Theory, measurement, and analysis should therefore be aligned rather than allowing the selected statistical package to define the research problem retroactively.
The level of the conclusion should match the level of the evidence
If the analysis compares individuals, conclusions should describe individuals unless an additional argument supports a broader claim.
If the analysis compares organizations, conclusions should describe organizations.
If the model estimates a cross-level relationship, the conclusion should preserve both levels.
Individual-level conclusion
Faculty members with greater self-efficacy reported greater personal AI adoption.
Organization-level conclusion
Universities with stronger AI infrastructure had higher average adoption rates.
Cross-level conclusion
Faculty members in universities with stronger infrastructure reported greater individual adoption.
These statements sound similar but make different claims.
A study can have more than one level of analysis
A study need not choose permanently between individual and organizational analysis.
If its research questions deliberately concern both, it can contain multiple analytical levels.
For example:
Level 1 question: Does faculty self-efficacy predict individual adoption?
Level 2 question: Does university infrastructure predict average adoption?
Cross-level question: Does infrastructure alter the individual self-efficacy–adoption relationship?
This is one reason a study can have more than one unit of analysis.
Do not confuse level of analysis with level of measurement
The phrase “level of measurement” can also refer to nominal, ordinal, interval, and ratio measurement scales.
That is a different concept entirely.
In multilevel research, researchers may also use “level of measurement” informally to indicate the level at which a construct is measured, but this should not be confused with nominal-versus-ordinal measurement terminology.
Context matters, and explicit wording is helpful.
Do not confuse analytical level with statistical significance
A variable does not become an organizational-level construct because it is significant in an organization-level model.
Nor does an individual-level construct become group-level merely because its group mean predicts something.
Level comes from the theoretical definition, referent, data structure, and analytical target, not from the p-value.
The best distinction is practical rather than philosophical
If the terminology feels abstract, ask two questions:
1. What exact kind of entity is each case?
That identifies the unit.
2. Where does that entity sit within the hierarchy of the phenomenon?
That identifies the level.
For example:
Entity: faculty member
Level: individual
Entity: department
Level: group or suborganizational
Entity: university
Level: organizational
The distinction becomes useful because the study can now describe exactly where its variables, questions, and conclusions belong.
