Multilevel Structural Equation Modeling

Traditional statistical models often force you into a compromise: you can analyze complex relationships between latent variables using structural equation modeling (SEM), or you can account for nested data structures using multilevel modeling (MLM). But what happens when your research question demands both? Multilevel structural equation modeling (MSEM) solves this problem by integrating these two powerful frameworks. It allows you to simultaneously model measurement and structural relationships at multiple levels, providing a nuanced understanding of how individual characteristics and their broader contexts—like students within schools or employees within teams—interact to produce observed outcomes.

The Core Problem: Nested Data and Latent Variables

Data in the social, behavioral, and organizational sciences are rarely independent. Students are nested within classrooms, employees within departments, and patients within clinics. This nesting creates dependence in your data; individuals within the same group are often more similar to each other than to individuals in different groups. Ignoring this violates the independence assumption of standard SEM, leading to underestimated standard errors, inflated Type I error rates, and potentially misleading conclusions.

Concurrently, many constructs of interest—such as intelligence, organizational climate, or depression—are not directly observable. These latent variables must be measured indirectly using multiple observed indicators (e.g., survey items). SEM excels at modeling these latent constructs, separating measurement error from true score variance. MSEM brings these two necessities together. It formally partitions the total variance in your observed variables into within-group (individual-level) and between-group (cluster-level) components, and it can model latent structures at each of these levels.

Fundamental Components and Notation

To grasp MSEM, you must understand how it decomposes your data. For any observed variable $y_{ij}$ for individual $i$ in group $j$, the model specifies: $$y_{ij}=y_{W,ij}+y_{B,j}$$ Here, $y_{W,ij}$ represents the individual's deviation from their group mean (the within-group component), and $y_{B,j}$ represents the group mean itself (the between-group component). This decomposition is applied to all indicators in your model.

This leads to the specification of two parallel models:

1. The Within-Group Model: This describes the relationships among variables at the individual level, controlling for group membership. For example, it models how a student's self-efficacy (a latent variable) relates to their individual achievement score.
2. The Between-Group Model: This describes the relationships among variables at the cluster level. It models how the school-average level of self-efficacy relates to the school-average achievement score, or how a school's resources (a latent variable at the group level) predict that average achievement.

The key advance of MSEM over simpler multilevel approaches is that both the within and between models can include their own latent variables, factor structures, and structural paths. This allows you to test, for instance, whether the factor structure of a questionnaire is equivalent across levels.

Specifying and Testing a Multilevel SEM

Building an MSEM typically follows a logical sequence of steps to ensure your model is correctly identified and your interpretations are sound.

Step 1: Establishing Non-Independence. Before fitting a full MSEM, you must confirm that nesting is an issue. Calculate the intraclass correlation coefficient (ICC) for your key outcome variables. The ICC represents the proportion of total variance that lies between groups. An ICC above 0.05 to 0.10 generally warrants a multilevel approach. For a variable $y$, the ICC is computed as: $$ICC=\frac{{\sigma }_{\text{between}}^2}{{\sigma }_{\text{between}}^2+{\sigma }_{\text{within}}^2}$$ where ${\sigma }^2$ represents variance components from a null (intercept-only) model.

Step 2: The Unconditional Measurement Model. This step asks: "Does my measurement model hold at both levels?" You specify your hypothesized factor structure (e.g., a two-factor model for a survey) and test it as a multilevel confirmatory factor analysis (MCFA). You will assess model fit separately for the within and between structures. Often, you test for measurement invariance across levels—does the same factor model apply to both individual deviations and group means?

Step 3: Adding Structural Paths. Once you are confident in the measurement model, you introduce the structural relationships. You can specify paths between latent variables at the within level, at the between level, or even cross-level paths (though these are specified differently). For example, your final model might show that at the individual level (within), student engagement predicts achievement, while at the school level (between), collective teacher efficacy predicts the school's average achievement.

Step 4: Interpretation. Interpretation is always level-specific. A parameter estimate in the within-group model tells you about individual-level processes, controlling for group membership. An estimate in the between-group model tells you about group-level processes. You might find a relationship is strong at one level but nonexistent at the other, which is a critical substantive finding that simpler models would miss.

Common Pitfalls

Ignoring Between-Group Measurement. A major error is assuming your latent construct operates identically at the group level. Aggregating individual survey items to a group mean creates a "shared" or "global" construct that may differ conceptually from the individual-level construct. For example, "climate" at the group level is an emergent property, not just an average of individual perceptions. Failing to model this can lead to aggregation bias or ecological fallacies. Always test the between-group measurement model separately.

Misinterpreting the ICC. A low ICC does not always mean multilevel modeling is unnecessary. If your primary research question involves a group-level predictor (e.g., school funding), you still need the between-group model to test it correctly. Conversely, a high ICC for your predictors can lead to multicollinearity between the within- and between-group components, making estimates unstable. Centering your predictors (e.g., group-mean centering) is often essential.

Overlooking Model Complexity and Power. MSEMs are complex and require large sample sizes. You need a sufficient number of groups (often 50+ for reliable between-level estimation) and individuals per group. With too few groups, the between-group model is underpowered, and estimates may not converge. Start with simpler models (e.g., a multilevel path model with observed variables) before adding full latent structures at both levels.

Confusing Cross-Level Interaction with Mediation. If you hypothesize that a group-level variable (e.g., school climate) moderates an individual-level relationship (e.g., between stress and performance), you are testing a cross-level interaction. This is distinct from a multilevel mediation model, where a variable may mediate relationships at different levels. Using the wrong specification will misrepresent your theoretical mechanism.

Summary

• MSEM unifies multilevel modeling and SEM to analyze data that are both nested and involve latent variables, partitioning variance and relationships into distinct within-group and between-group components.
• The approach begins by decomposing observed variables into within and between parts, allowing for the specification of separate—and potentially different—measurement and structural models at each level.
• Critical preliminary steps include calculating ICCs to justify the multilevel approach and testing the unconditional multilevel measurement model to ensure your constructs are valid at both the individual and cluster levels.
• Interpretation is strictly level-specific. A path in the within-model describes individual processes, while a path in the between-model describes group-level processes, preventing ecological fallacies.
• Successful application requires attention to power (particularly a sufficient number of groups), careful centering of predictors, and a clear theoretical distinction between individual-level and emergent group-level constructs.