Structural Equation Modeling (SEM)

What Is Structural Equation Modeling?

Structural Equation Modeling (SEM) is a multivariate statistical technique that combines factor analysis and multiple regression to test theoretical models involving observed (measured) and latent (unobserved) variables. It enables researchers to specify, estimate, and evaluate complex relationships among variables, while also accounting for measurement error.

SEM is widely used in psychology, sociology, economics, education, marketing, and many other fields where theories involve multiple constructs that cannot be directly measured.

Core Concepts

Observed vs. Latent Variables

Observed variables are directly measured (e.g., survey items, test scores). Latent variables represent abstract constructs such as intelligence, satisfaction, or socioeconomic status, and are inferred from a set of observed indicators.

Path Diagrams

SEM models are commonly visualized using path diagrams. Rectangles denote observed variables, ovals denote latent variables, arrows indicate causal direction, and double-headed arrows represent covariances or correlations.

Model Specification

A model consists of two parts:

• Measurement model: defines how latent variables are measured by observed indicators (similar to confirmatory factor analysis).
• Structural model: specifies relationships among latent constructs (similar to path analysis).

Steps in Conducting SEM

1. Develop a theoretical model articulate hypotheses about how constructs relate.
2. Translate theory into a path diagram draw the model showing latent and observed variables and the arrows that represent hypothesized relationships.
3. Choose indicators select observed variables that reliably reflect each latent construct.
4. Collect data ensure sample size is adequate (commonly a minimum of 10 observations per estimated parameter, though rules vary).
5. Specify the model in software enter the diagram or define equations in a program such as AMOS, LISREL, Mplus, or the lavaan package in R.
6. Estimate parameters usually by maximum likelihood (ML) or robust alternatives like GLS or Bayesian estimation.
7. Assess model fit examine fit indices (e.g., , CFI, TLI, RMSEA, SRMR) to determine how well the model reproduces the observed covariance matrix.
8. Modify and refine if fit is inadequate, consider theoretically justified modifications (adding paths, correlating errors) and reestimate.
9. Interpret results evaluate estimated coefficients, standard errors, and significance levels to test hypotheses.
10. Report findings include the model diagram, fit statistics, parameter estimates, and discussion of implications.

Key Assumptions

• Linearity: relationships among variables are linear.
• Multivariate normality: observed variables are normally distributed, especially for ML estimation.
• Independence of observations: each case is independent of all others.
• Correct model specification: the hypothesized model must reflect the true underlying structure.
• Measurement reliability: indicators should have acceptable reliability (Cronbachs .70 is a common rule).

Software Options

Several programs implement SEM. Below is a brief comparison:

Illustrative Example

model <- '  # measurement model  JS =~ JS1 + JS2 + JS3  OC =~ OC1 + OC2  Perf =~ Perf1 + Perf2  # structural model  OC ~ a*JS  Perf ~ b*OC + c*JS'fit <- sem(model, data = mydata)summary(fit, fit.measures=TRUE, standardized=TRUE)

Interpreting Results

When reading SEM output, focus on three aspects:

1. Fit indices: A nonsignificant suggests perfect fit, but is sensitive to sample size. Complement it with incremental fit indices (CFI, TLI) where values above .90 or .95 indicate acceptable/good fit, and absolute error indices (RMSEA, SRMR) where RMSEA<.08 and SRMR<.08 are commonly accepted.
2. Parameter estimates: Look at standardized coefficients (). Values of .10, .30, and .50 are often interpreted as small, medium, and large effects, respectively. Check standard errors and pvalues to assess statistical significance.
3. Indirect effects: Multiply the relevant path coefficients (e.g., ab) to obtain the mediated effect. Many programs report indirect effects and associated confidence intervals automatically.

Advantages of SEM

• Simultaneous analysis: Multiple dependent relationships are estimated together.
• Measurement error: Latent variables allow researchers to separate true score variance from error.
• Theory testing: Provides a rigorous framework for falsifying or supporting theoretical models.
• Flexibility: Handles hierarchical, longitudinal, and multigroup data.

Limitations and Common Pitfalls

• SEM does not prove causation; causal claims require experimental or longitudinal designs.
• Model fit can be improved by adhoc modifications, but such changes must be theoretically justified.
• Violation of assumptions (nonnormality, small sample size) can bias estimates, especially with ML estimation.
• Complex models with many parameters can become underidentified, meaning they lack enough information for unique solutions.

Conclusion

Structural Equation Modeling is a powerful tool for researchers who wish to evaluate elaborate theoretical frameworks involving both measured and unmeasured constructs. By integrating measurement and structural components, SEM provides a more realistic representation of the underlying phenomena than traditional regression techniques. Successful application hinges on solid theory, careful model specification, sufficient sample size, and honest reporting of fit and limitations. When used appropriately, SEM can illuminate the pathways through which variables influence each other, guide evidencebased interventions, and advance knowledge across many scientific domains.

For further reading see: Kline, R. B. (2016). Principles and Practice of Structural Equation Modeling. Guilford Press; Byrne, B. M. (2016). Structural Equation Modeling with Mplus. Routledge.