03 · What You Need to Know
The Level of a Variable Is Part of What the Variable Means
What is an individual-level variable?
An individual-level variable describes a characteristic, perception, behavior, outcome, or exposure that can differ from one person to another.
Examples include:
- student motivation;
- faculty AI self-efficacy;
- employee burnout;
- individual achievement;
- personal technology adoption;
- age;
- job satisfaction;
- individual income.
If two faculty members in the same university can legitimately have different values of the variable, it is often individual-level.
For example, one faculty member may feel highly confident using generative AI while another does not. Their self-efficacy scores describe the individuals themselves.
What is a group-level variable?
A group-level variable describes a property of a collective entity.
The group might be:
- a classroom;
- a team;
- a department;
- a school;
- a university;
- a company;
- a hospital;
- a neighborhood;
- a country.
Examples include:
- class size;
- team composition;
- school resources;
- formal university AI policy;
- organizational size;
- departmental workload norms;
- national research expenditure.
These variables describe the collective rather than one member's personal state.
Individual-level variable
Varies among people within the same group.
Group-level variable
Describes a property of the group itself or a characteristic assigned to the group as a whole.
The unit of analysis helps identify the variable's level
If the unit of analysis is an individual faculty member, variables such as self-efficacy and personal adoption are naturally individual-level.
If the unit of analysis is the university, variables such as institutional policy and university-wide adoption rate are organizational-level.
This is why understanding the difference between level of analysis and unit of analysis helps clarify variable classification.
Some group-level variables are directly measured
Not every higher-level variable comes from aggregating individual responses.
Some variables exist directly at the group level.
Examples include:
- number of students in a classroom;
- department budget;
- presence of an institutional policy;
- organization size;
- school accreditation status;
- national GDP;
- hospital bed capacity.
These variables can often be observed directly for the higher-level entity.
Other group-level variables are constructed from individual responses
Some constructs such as organizational climate, team cohesion, or classroom climate are often measured through individual reports.
Suppose faculty members rate:
“In this university, faculty receive strong support for experimenting with new digital tools.”
If the construct is theoretically defined as a shared university-level support climate, researchers may aggregate individual responses to create a university-level measure, provided the aggregation is justified.
This is fundamentally different from simply saying that every individual perception is itself an organizational variable.
Individual perception and group climate are not automatically the same construct
Consider:
“I feel supported by my university.”
This measures a person's perceived experience.
Now consider:
“Faculty in this university generally receive strong support for AI adoption.”
This item uses the group as the referent.
The first may remain an individual-level perception even if responses are averaged. The second is more explicitly designed to capture a shared group property.
Watch Out
Averaging individual responses changes the numerical level of the score, but it does not automatically transform an individual-level construct into a valid group-level construct. The theoretical definition and measurement referent still matter.
Group-level constructs need a theoretical rationale
If researchers claim that an organization has a particular climate, culture, collective efficacy, or shared norm, they should explain why the property belongs to the collective.
For example, a support climate might be defined as a shared perception that organizational policies and practices encourage employee innovation.
The group-level theory therefore concerns:
What is shared or characteristic of the organization?
rather than:
How does one employee personally feel?
Aggregation requires more than calculating a mean
Suppose 30 employees within each organization rate team climate.
The group mean is easy to compute:
But a mathematically correct mean is not enough to establish that 4.20 validly represents “organizational climate.”
Researchers may need to examine whether members show meaningful agreement and whether groups actually differ from one another.
This is the issue addressed when deciding when individual responses can be treated as a group-level measure.
Within-group agreement and between-group variation answer different questions
Two ideas are often relevant when individual responses are aggregated.
Within-group agreement asks whether people within the same group give sufficiently similar ratings to justify treating the group as having a shared property.
Between-group variation asks whether different groups actually differ enough for the group-level construct to be analytically meaningful.
A university climate measure is more compelling when faculty within each university show meaningful consensus and universities differ from one another.
These are not identical statistical questions.
ICC statistics can help describe group-related variation
In multilevel research, intraclass correlation coefficients are often used to describe clustering.
Conceptually:
Different ICC formulations serve different purposes, including reliability of group means and assessment of clustering.
Researchers should therefore report and interpret the version appropriate to their measurement and design rather than treating “the ICC” as one universal statistic.
Agreement statistics answer another part of the aggregation problem
Measures such as within-group agreement indices may be used to evaluate whether respondents within the same group give sufficiently similar ratings relative to an expected null distribution.
No single threshold universally proves that aggregation is justified.
Agreement evidence should be interpreted alongside construct definition, item referent, group size, between-group variability, reliability, and prior theory.
The same variable name can refer to different levels
Consider “AI adoption.”
At the individual level:
Faculty AI adoption may mean whether one faculty member uses AI in teaching.
At the department level:
Departmental AI adoption may mean the extent to which a department has institutionalized AI practices.
At the university level:
University AI adoption may refer to formal integration across institutional systems.
These are related concepts, but they are not automatically interchangeable measures of the same thing.
Individual-level and group-level variables represent different sources of variation
Suppose employees' workload is measured individually.
An employee's score can be decomposed conceptually into:
1. whether that individual works more than colleagues in the same organization;
2. whether the organization itself tends to impose greater workloads than other organizations.
These are different comparisons.
The first is within-group variation.
The second is between-group variation.
A single raw employee workload score contains information about both unless the model separates them.
Group-mean centering can help separate within- and between-group relationships
In some multilevel models, researchers subtract the group mean from each individual's score:
The group mean can then be included separately as a higher-level predictor.
This allows the researcher to distinguish:
within-organization effect: are employees with higher workload than their colleagues more burned out?
from:
between-organization effect: are employees in generally higher-workload organizations more burned out?
Those two effects do not have to be equal
This point is crucial.
Suppose within organizations, employees who work more than their colleagues report greater burnout.
At the organization level, however, average workload might have a weak relationship with average burnout because organizations with heavy workloads also provide better compensation or support.
The same variable names can therefore produce different relationships at different levels.
This is why relationships can differ across levels of analysis.
An individual-level predictor can have a group-level consequence
Suppose individual employee turnover intentions are aggregated into an organizational turnover-risk indicator.
Researchers may examine whether organizations with higher average turnover intention subsequently experience higher actual turnover.
The theoretical claim is now group-level even though the original measurements came from individuals.
The transition from individual data to group inference should be explicit.
A group-level predictor can have an individual-level consequence
Suppose university AI policy affects individual faculty adoption.
The predictor exists at the organizational level.
The outcome exists at the individual level.
This is a cross-level relationship.
Such questions are common in education, organizational research, public health, and sociology because people operate within contexts that can influence their behavior.
Group-level moderators can change individual-level relationships
Suppose individual self-efficacy predicts AI adoption more strongly in universities with strong infrastructure.
Self-efficacy and adoption are individual-level variables.
Infrastructure is an organizational-level moderator.
The question is whether a higher-level context changes a lower-level relationship.
This is cross-level moderation.
Individual-level variables can also be used to explain group composition
A group may be characterized partly by the distribution of individual attributes among its members.
For example, researchers might examine:
- average team experience;
- proportion of senior faculty;
- diversity of disciplinary backgrounds;
- variance in member expertise.
These are group-level composition variables derived from lower-level characteristics.
The group mean is only one possible composition summary. Dispersion, minimum, maximum, proportion, or diversity indices may be more theoretically appropriate depending on the question.
A group-level construct is not always the average individual construct
Suppose team diversity is the research variable.
Diversity is inherently a property of the composition of the team. It is not obtained by asking each person “how diverse are you?” and averaging the responses.
Similarly, class size, network density, organizational centralization, and policy presence are collective properties that have no direct individual equivalent.
Group-level research therefore extends beyond aggregation.
Composition and contextual effects should be distinguished
A group can affect an individual because of who its members are or because of a collective context that exists beyond those individual characteristics.
Suppose students in high-achieving schools perform better.
That pattern might arise because the schools simply contain many high-achieving students, because the schools provide stronger instructional environments, or both.
Multilevel analysis can help distinguish individual composition from contextual effects when the design and measurements support that distinction.
Levels should be specified before interpreting coefficients
A coefficient involving “support” is difficult to interpret unless readers know whether support means:
- one person's perception;
- group-average perception;
- a shared support climate;
- an objective institutional policy;
- a cross-level contextual variable.
The variable label alone is insufficient.
Level is part of the variable's substantive definition.
Mixing levels can create misleading research questions
Consider:
Does institutional support influence faculty adoption?
What does “institutional support” mean?
If it is one person's perception of support and adoption is that same person's behavior, the relationship may be entirely individual-level.
If support is an organization-level climate and adoption is individual, the relationship is cross-level.
If both variables are university averages, the relationship is organizational-level.
The wording needs enough specificity to reveal the intended structure.
This is why mixing individual-level and group-level explanations can create conceptual ambiguity.
Ignoring groups can underestimate statistical uncertainty
Suppose 1,000 students are drawn from 20 schools.
Students within the same school may be more similar than students from different schools because they share teachers, policies, resources, or local contexts.
If the analysis treats all 1,000 students as completely independent, standard errors can be too small.
Recognizing the group structure is therefore statistically important even when the focal variables are individual-level.
Aggregating everything to the group level can create the opposite problem
Suppose researchers average all student variables within each school and analyze only 20 school means.
This eliminates individual-level variation.
If the original question concerns why some students perform better than others within schools, aggregation has discarded the information needed to answer it.
Neither ignoring the groups nor collapsing everything into group averages is automatically correct.
Ecological fallacy arises when group-level findings are interpreted as individual-level findings
Suppose universities with higher average faculty AI competence have higher overall AI adoption rates.
That does not prove that, within each university, faculty members with greater competence are more likely to adopt AI.
The group-level association may differ from the individual-level relationship.
Making the individual inference from aggregate evidence risks the ecological fallacy.
Atomistic fallacy runs in the opposite direction
Suppose individual faculty self-efficacy predicts personal AI adoption.
That does not automatically establish that universities with higher average self-efficacy will have higher institutional adoption.
Using individual-level evidence to make group-level claims can produce the atomistic fallacy.
Both errors occur when researchers cross levels without evidence.
Group-level sample size is not the same as individual sample size
Suppose 2,400 faculty members are distributed across 12 universities.
For individual-level questions, the study contains substantial faculty information, though clustering matters.
For university-level questions, however, only 12 universities are available.
Watch Out
Thousands of individual observations do not create thousands of higher-level units. Group-level effects depend on the number and variation of the groups themselves.
Group-level variables need adequate variation across groups
If every university in a sample has essentially the same AI policy, university policy cannot explain much between-university variation even if it is theoretically important.
Likewise, if nearly all classrooms have the same class size, estimating a classroom-level class-size effect will be difficult.
A meaningful higher-level analysis requires both enough groups and enough variation among them.
The distinction matters during sampling
If the primary question is organizational-level, researchers need to sample organizations, not merely many people from a few organizations.
If both individual and organizational questions matter, sampling should provide adequate information at both levels.
This can affect:
- number of groups;
- number of respondents per group;
- how groups are selected;
- balance across groups;
- power for cross-level effects.
It also matters during interpretation
Consider these three conclusions:
Individual level: Faculty members with greater AI self-efficacy report greater adoption.
Organizational level: Universities with higher average AI self-efficacy have higher institutional adoption rates.
Cross level: Faculty members in universities with stronger institutional support report greater adoption.
They refer to related topics but answer different questions.
A good manuscript preserves those differences rather than collapsing everything into “self-efficacy and support influence AI adoption.”
