Structural Equation Modelling

Structural equation modelling (SEM) is a family of statistical techniques used to specify and evaluate complex relationships among observed and latent variables. It combines elements of measurement modelling and path analysis, allowing researchers to examine how theoretical constructs are measured and how those constructs are hypothesized to relate to one another within an integrated model. (PubMed Central (PMC))

For example, a researcher may propose that service quality increases customer satisfaction, which subsequently increases customer loyalty. Service quality and satisfaction may themselves be latent constructs measured through multiple questionnaire items. SEM allows the researcher to represent both the measurement of these constructs and the hypothesized relationships among them.

This ability makes SEM particularly useful for theory-driven research involving multiple constructs, mediation and measurement error. However, the sophistication of SEM can create a misleading impression that directional arrows or statistically significant paths establish causality. They do not. Causal interpretation depends on research design, temporal ordering, assumptions, theory and alternative explanations—not simply on estimating a structural equation model. The role of SEM in causal inference has consequently been debated extensively in methodological literature. (PubMed Central (PMC))

On this page:

• Structural equation modelling explained simply
• What SEM is
• Measurement model vs structural model
• Observed and latent variables
• Exogenous and endogenous variables
• CFA vs SEM
• Path analysis vs SEM
• SEM model specification
• Model identification
• Estimation
• Evaluating the measurement model
• Evaluating the structural model
• Direct, indirect and total effects
• Mediation in SEM
• Model fit
• Why significant paths do not establish causality
• Model comparison and respecification
• Dudovskiy SEM Model Development Framework
• Application example
• Advantages and limitations
• Common mistakes
• SEM in business research
• SEM in the age of AI
• When to use SEM
• Dissertation example
• Exam tip

Question	Short answer
What does SEM analyse?	Systems of relationships among observed and/or latent variables
Can SEM include latent constructs?	Yes
What does the measurement model address?	Relationships between constructs and their indicators
What does the structural model address?	Hypothesized relationships among variables or constructs
Is CFA part of SEM?	Yes
Can SEM estimate several equations simultaneously?	Yes
Can SEM estimate indirect effects?	Yes
Does good model fit prove the theory?	No
Does a significant path prove causation?	No
Should structural paths be interpreted before measurement quality is evaluated?	Generally, no

Structural Equation Modelling Explained Simply

Imagine a retailer wants to understand why customers remain loyal. The researchers propose the following theory:

Service Quality → Customer Satisfaction → Customer Loyalty

They also believe Service Quality affects Customer Loyalty partly through Customer Satisfaction.

The difficulty is that concepts such as service quality, satisfaction and loyalty cannot be measured directly. Each is represented using several questionnaire items. SEM allows the researcher to deal with these two levels of the problem together.

First comes the measurement question:

Do the questionnaire items adequately represent Service Quality, Customer Satisfaction and Customer Loyalty?

Then comes the structural question:

Are the hypothesized relationships among Service Quality, Customer Satisfaction and Customer Loyalty supported by the data?

The distinction is fundamental. A sophisticated structural model is difficult to interpret convincingly if the constructs entering that model have not been measured adequately.

Dudovskiy Research Assistant

Not sure if structural equation modelling is the correct methodology for your dissertation?

Enter your topic to receive a free, tailored methodology preview.

Dissertation topic

Generate Free Methodology Preview → Free · No account required

Learn more about Dudovskiy Research Assistant

What Is Structural Equation Modelling?

Structural equation modelling is a broad statistical framework for representing and estimating systems of relationships among variables. SEM can incorporate observed variables, latent variables, measurement error and multiple equations within the same analytical framework. It developed from traditions including factor analysis and path analysis and now encompasses a wide range of specific models. (PubMed Central (PMC))

A major strength of SEM is its ability to distinguish a theoretical construct from the indicators used to measure it. Instead of treating an observed questionnaire score as a perfectly measured representation of customer trust, for example, a latent-variable model can represent Customer Trust through multiple indicators while explicitly modelling measurement error.

SEM can also estimate multiple relationships simultaneously. A variable may operate as an outcome in one equation and as a predictor in another. This makes SEM especially useful for investigating theoretical systems containing mediators, multiple dependent variables and interconnected relationships. (PubMed Central (PMC))

SEM is therefore better understood as a modelling framework than as one statistical test.

Measurement Model vs Structural Model

One of the most important distinctions in SEM is between the measurement model and the structural model.

The measurement model addresses:

How are theoretical constructs represented by observed indicators?

The structural model addresses:

How are the variables or theoretical constructs hypothesized to relate to one another?

Consider three latent constructs:

Perceived Value
→ Value Item 1
→ Value Item 2
→ Value Item 3

Customer Satisfaction
→ Satisfaction Item 1
→ Satisfaction Item 2
→ Satisfaction Item 3

Customer Loyalty
→ Loyalty Item 1
→ Loyalty Item 2
→ Loyalty Item 3

These relationships form the measurement component.

The researcher may then hypothesize:

Perceived Value → Customer Satisfaction → Customer Loyalty

and perhaps:

Perceived Value → Customer Loyalty

These paths form the structural component.

SEM integrates these components, allowing theoretical relationships among constructs to be investigated while accounting for the way those constructs are measured. The combination of latent measurement and structural relationships is a defining feature of much SEM research. (PubMed Central (PMC))

Observed and Latent Variables

An observed variable is directly measured. Questionnaire responses, revenue, age and number of purchases can all be observed variables.

A latent variable represents a construct that is not observed directly but is inferred through measured indicators. Examples include brand trust, organizational commitment, perceived value and employee engagement.

Suppose Brand Trust is measured using four questionnaire statements. The responses to those statements are observed variables, while Brand Trust is the latent construct they are intended to represent.

This distinction allows SEM to address measurement error more explicitly than analyses that simply replace a construct with a single observed score. The ability to incorporate latent variables and measurement error is one of the important features of SEM. (PubMed Central (PMC))

Not every SEM must contain latent variables, however. Path analysis can model structural relationships entirely among observed variables and can be understood as part of the broader SEM family.

Exogenous and Endogenous Variables

SEM also distinguishes between exogenous and endogenous variables.

An exogenous variable is not explained by another variable within the specified model. It functions as a predictor at the point where it enters the model.

An endogenous variable is explained, at least partly, by one or more other variables in the model.

Consider:

Service Quality → Satisfaction → Loyalty

If nothing predicts Service Quality within the specified model, Service Quality is exogenous. Satisfaction is endogenous because Service Quality predicts it. Loyalty is also endogenous because it is predicted by Satisfaction and potentially Service Quality.

These labels refer to the variable’s role within the specified model, not to an inherent property of the variable itself.

Confirmatory Factor Analysis vs Structural Equation Modelling

CFA and SEM are closely related, but they should not be treated as synonyms. CFA focuses primarily on the measurement structure. It evaluates whether observed indicators provide an adequate representation of theoretically specified latent factors. A broader SEM can incorporate that measurement model and add hypothesized structural relationships among the latent constructs.

The progression can therefore be represented as:

CFA:
Indicators → Latent Constructs

SEM:
Indicators → Latent Constructs → Relationships among Constructs

CFA belongs within the wider SEM framework. SEM can subsequently extend measurement modelling into questions concerning structural relationships, mediation and multiple interconnected outcomes. (PubMed Central (PMC))

This explains why measurement evaluation commonly precedes interpretation of structural paths.

Path Analysis vs Structural Equation Modelling

Path analysis and latent-variable SEM also share a common intellectual lineage.

Path analysis examines systems of relationships among observed variables. Researchers can specify multiple directional paths and estimate direct and indirect relationships.

Latent-variable SEM extends this logic by incorporating constructs represented through multiple indicators and explicitly modelling measurement components.

A simplified distinction is:

Path analysis:
Observed variable → Observed variable → Observed variable

Latent-variable SEM:
Indicators → Latent construct → Latent construct → Indicators

The boundary is not that one uses arrows while the other does not. Both can use path diagrams and systems of equations. The crucial difference is whether a measurement model involving latent variables is incorporated.

SEM Model Specification

SEM begins with model specification: translating substantive theory into an explicit statistical model.

Researchers need to determine which indicators measure which constructs, which structural paths are hypothesized, which variables covary, and which parameters are constrained or freely estimated.

These decisions should normally have theoretical justification before the results are examined. Otherwise, SEM can deteriorate into a search for whichever arrangement of arrows happens to fit the current dataset best.

Suppose theory proposes that employee autonomy influences job satisfaction, which subsequently influences turnover intention. The researcher should be able to explain why those relationships are proposed and why the directional ordering makes substantive sense.

The diagram is therefore not merely a convenient picture of the analysis. It is a formal representation of the theoretical claims being evaluated.

Model Identification

Before the parameters of an SEM can be estimated uniquely, the model must be identified.

Identification concerns whether sufficient information and appropriate constraints exist to obtain unique estimates of the unknown model parameters. Problems can arise when the researcher attempts to estimate more relationships than the available information can support or specifies a model whose parameters cannot be uniquely distinguished.

Identification is especially important in complex latent-variable models because a model can appear theoretically sensible while remaining statistically unidentified.

The practical principle is straightforward:

A model must be estimable before its parameter estimates can be meaningfully interpreted.

Software may detect many identification problems, but researchers should understand the underlying model rather than interpreting successful software execution as evidence that the model itself is methodologically sound.

Estimation in SEM

Once the model is specified and identified, its parameters can be estimated. Maximum likelihood estimation is widely used in covariance-based SEM, while robust maximum likelihood, weighted least squares and other estimators may be appropriate under different conditions. SEM software packages provide a range of estimation procedures because characteristics such as indicator type, non-normality and missing data can affect which estimator is appropriate. (PubMed Central (PMC))

The estimator is therefore a methodological choice rather than a purely technical setting.

A dissertation should identify the estimation approach used and explain it where the characteristics of the data make the choice consequential. An estimator suitable for approximately continuous and normally distributed variables should not automatically be assumed suitable for every ordinal questionnaire dataset.

Evaluating the Measurement Model

When an SEM contains latent constructs, researchers should establish that those constructs are represented adequately before interpreting structural relationships among them.

This usually involves examining the measurement model through CFA. Relevant evidence can include factor loadings, factor correlations, residuals, model-fit measures, reliability evidence and evidence concerning construct validity.

This stage addresses a fundamental question:

Are the constructs being measured well enough for their structural relationships to be meaningfully interpreted?

Suppose Customer Trust is poorly represented by its indicators. A statistically significant path from Customer Trust to Customer Loyalty may then be difficult to interpret because uncertainty exists about what the latent Customer Trust variable actually represents.

Measurement quality is therefore not a preliminary nuisance to be cleared before reaching the “interesting” structural model. It is part of the substantive argument.

Good Structural Fit Does Not Rescue a Poor Measurement Model

A structural model can appear theoretically compelling and produce significant paths while its measurement foundation remains weak.

Imagine that Employee Engagement strongly predicts Organizational Commitment in an SEM. If the items intended to represent Employee Engagement actually combine several poorly distinguished dimensions and show problematic measurement behaviour, the meaning of the structural coefficient becomes uncertain.

The problem cannot be repaired simply because the overall SEM displays acceptable global fit.

SEM combines measurement and structural components precisely because errors in measurement affect interpretation of relationships among theoretical constructs. Research describing SEM emphasizes its ability to integrate measurement models with structural relationships rather than treating the two as independent analytical worlds. (PubMed Central (PMC))

A defensible SEM therefore asks first:

What are we measuring?

and only then:

How do those constructs relate?

Evaluating the Structural Model

Once the measurement component is defensible, attention can move to the hypothesized structural relationships.

Researchers examine the estimated path coefficients, their uncertainty and statistical significance, the direction and magnitude of relationships, indirect effects where relevant, and the overall coherence of the structural model.

Suppose the theoretical model proposes:

Service Quality → Satisfaction → Loyalty

The structural analysis may show a strong positive relationship between Service Quality and Satisfaction and another between Satisfaction and Loyalty. It may also estimate a remaining direct relationship between Service Quality and Loyalty.

The interpretation should not stop at whether individual p-values fall below 0.05. Researchers should consider whether the direction and magnitude of the relationships correspond to theoretical expectations, whether estimates are sufficiently precise, whether indirect effects are supported and whether plausible alternative explanations remain.

Direct, Indirect and Total Effects

SEM is particularly useful for decomposing relationships into direct, indirect and total effects.

Consider:

X → M → Y

with an additional direct path:

X → Y

The direct effect represents the relationship between X and Y represented by the direct structural path.

The indirect effect represents the relationship transmitted through the mediator M.

The total effect combines the relevant direct and indirect pathways.

SEM can estimate systems containing multiple such relationships simultaneously, which is one reason it is widely used for mediation models. (PubMed Central (PMC))

This capability becomes particularly valuable when mediators themselves are latent variables or when several mediating mechanisms are investigated within the same theoretical system.

Mediation in SEM

Mediation addresses whether the relationship between one variable and another operates partly through an intermediate variable.

Suppose researchers hypothesize:

Digital Service Quality → Customer Trust → Repurchase Intention

Customer Trust is the mediator. The theory proposes that better digital service quality contributes to trust and that trust subsequently contributes to repurchase intention.

SEM can estimate the direct and indirect pathways simultaneously and can incorporate latent measurement models for the constructs. (PubMed Central (PMC))

However, estimating a statistically significant indirect effect does not by itself demonstrate a causal mechanism. Mediation is fundamentally concerned with process and temporal ordering, and causal interpretation depends on research design and assumptions. Methodological discussions of mediation explicitly caution that mediation analysis does not necessarily establish causality. (PubMed Central (PMC))

This becomes especially important when all variables are measured simultaneously in a cross-sectional questionnaire.

Evaluating SEM Model Fit

SEM researchers commonly evaluate the correspondence between the specified model and observed data using several forms of fit evidence.

Frequently reported measures include:

• chi-square;
• Comparative Fit Index (CFI);
• Tucker-Lewis Index (TLI);
• Root Mean Square Error of Approximation (RMSEA); and
• Standardized Root Mean Square Residual (SRMR).

As discussed in CFA, these statistics should not be transformed into universal pass/fail rules. Model characteristics, sample properties and the nature of misspecification can influence their behaviour.

Global fit should also be considered alongside local evidence. A model can have acceptable overall fit while containing weak loadings, large residuals, implausible coefficients or theoretically questionable paths.

The purpose of model-fit assessment is therefore not to obtain permission to publish the model. It is to investigate how adequately the specified model represents the observed relationships.

Statistical Significance Does Not Establish a Causal Model

SEM diagrams frequently contain arrows, and those arrows can create a powerful visual impression of causality.

Suppose a cross-sectional survey produces a statistically significant path:

Employee Engagement → Job Performance

That coefficient does not establish that increasing engagement would cause an increase in performance. Reverse causation may be possible, omitted variables may influence both constructs, and the simultaneous measurement of engagement and performance may provide limited evidence about temporal ordering.

The methodological literature contains a longstanding debate about causal interpretation of SEM with nonexperimental data. A consistent lesson is that statistical modelling cannot substitute for substantive theory and research design. (PubMed Central (PMC))

SEM evaluates implications of a specified model under assumptions. Drawing a directional arrow does not create the conditions required for causal identification.

A more defensible interpretation of cross-sectional observational SEM is often:

The observed relationships are consistent with the hypothesized structural model.

That is different from claiming:

The SEM proves that X causes Y.

Model Comparison

Researchers sometimes have more than one theoretically plausible explanation for the observed relationships.

For example:

Model A:
Service Quality → Trust → Loyalty

Model B:
Service Quality → Loyalty → Trust

The existence of competing models can be informative because it forces researchers to distinguish what the data support from what theory and research design establish.

Some models may be compared statistically when appropriate, while information criteria and other evidence can also assist model comparison. However, selecting the best-fitting model among several candidates does not prove that the selected model represents the true causal process.

Model comparison is strongest when competing models arise from genuine theoretical alternatives rather than being generated simply to find a better-fitting configuration.

Model Respecification and Overfitting

SEM software can provide modification indices and other diagnostics identifying parameters that might improve model fit if freed.

These tools can be valuable for diagnosing misspecification. They can also encourage researchers to keep altering the model until desirable fit statistics appear.

Suppose a modification index suggests adding a path that was never anticipated theoretically. Adding it may improve fit substantially, but the researcher now needs to explain why that relationship should exist. Statistical improvement alone is not substantive justification.

Repeated modification using the same dataset also increases the risk that the model becomes tailored to sample-specific patterns.

Theory-guided respecification can therefore be legitimate, but extensive post-hoc modification should be reported transparently and, where possible, evaluated using independent data.

Dudovskiy SEM Model Development Framework

The Dudovskiy SEM Model Development Framework synthesizes established SEM principles into a practical sequence connecting theoretical development, measurement and structural analysis. It does not propose a new statistical form of structural equation modelling. Its purpose is to prevent researchers from interpreting structural relationships before establishing what the underlying constructs represent and how adequately they are measured.

Research Theory → Construct Definition → Measurement Model → Measurement Evaluation → Structural Model → Model Estimation → Structural Evaluation → Theory-Guided Diagnosis → Defensible Theoretical Model

The central principle is:

A defensible SEM analysis establishes how constructs are measured before interpreting how those constructs relate to one another.

1. Research Theory

Begin with the substantive theoretical problem. Identify why the proposed relationships should exist and what previous theory or evidence supports the hypothesized model.

SEM should evaluate a theoretical argument rather than manufacture one from patterns discovered after the data are collected.

2. Construct Definition

Define the theoretical constructs precisely. Determine what each construct represents and what it does not represent before selecting indicators.

Ambiguous constructs create ambiguous structural relationships regardless of how sophisticated the subsequent modelling becomes.

3. Measurement Model

Specify how observed indicators represent the latent constructs. Identify which indicators belong to each construct and how the constructs are allowed to relate within the measurement model.

This converts conceptual definitions into an explicit empirical measurement structure.

4. Measurement Evaluation

Evaluate whether the proposed constructs are represented adequately. Factor loadings, model fit, residual evidence, reliability and relevant validity evidence can contribute to this judgment.

If measurement is not defensible, structural interpretation should not simply proceed as though the problem does not exist.

5. Structural Model

Specify the theoretically hypothesized relationships among the constructs. This may include direct paths, mediation, multiple outcomes or other structural relationships appropriate to the research question.

The direction of the arrows should come from substantive reasoning and research design rather than statistical convenience.

6. Model Estimation

Estimate the specified model using an approach appropriate to the data and model.

Estimation is the computational stage, but its results depend on all of the conceptual and methodological decisions made beforehand.

7. Structural Evaluation

Examine the structural coefficients, uncertainty, indirect effects where relevant, model fit and correspondence between results and theoretical expectations.

The question is not merely whether paths are statistically significant. It is whether the overall pattern provides meaningful evidence concerning the theoretical model.

8. Theory-Guided Diagnosis

Investigate unexpected findings, residual problems and potential misspecification. Modifications should have substantive justification rather than being introduced solely to improve fit statistics.

This stage distinguishes methodological diagnosis from statistical optimization.

9. Defensible Theoretical Model

The final model should combine defensible measurement, appropriate statistical analysis and cautious theoretical interpretation.

The researcher should be able to explain how constructs were measured, why structural paths were proposed, what evidence supports the model, what limitations remain and how strongly the results can be interpreted.

The framework therefore changes the objective from:

“Which SEM paths are significant?”

to:

“Is the complete chain from theory to measurement to structural interpretation defensible?”

Application of Structural Equation Modelling: an Example

Consider a researcher investigating why consumers continue using subscription-based digital services. Theory proposes that perceived service quality increases customer trust, customer trust increases satisfaction, and satisfaction subsequently increases continuance intention. The researcher also proposes an indirect pathway from service quality to continuance intention through trust and satisfaction.

Each construct is measured using several questionnaire items adapted from established scales. The researcher therefore begins by specifying and evaluating the measurement model rather than immediately interpreting the structural relationships.

CFA indicates that the four latent constructs are represented adequately by their indicators after theoretically justified measurement decisions. Reliability and relevant validity evidence are also considered. Only after this measurement structure is judged defensible does the researcher specify the structural paths derived from the theoretical model.

The SEM indicates positive associations along the hypothesized sequence, and the indirect relationship between service quality and continuance intention is estimated through the intermediate constructs. The researcher examines coefficient magnitudes and uncertainty rather than treating statistical significance as the sole criterion.

Because the study uses cross-sectional observational survey data, the dissertation does not claim that SEM has demonstrated the proposed causal sequence. Instead, the findings are described as consistent with the hypothesized theoretical relationships, while longitudinal or experimental research is recommended for stronger investigation of temporal and causal processes.

The value of SEM in this example is therefore not simply that several arrows can be estimated simultaneously. It is that measurement and structural theory are represented within one coherent analytical framework.

Advantages and Limitations of Structural Equation Modelling

SEM allows researchers to investigate complex theoretical systems that would be cumbersome to analyse through a series of isolated regressions. Multiple dependent relationships can be estimated simultaneously, latent constructs can be represented using multiple indicators, and direct and indirect effects can be examined within a coherent model. These capabilities make SEM particularly valuable when substantive theory concerns interconnected processes rather than one predictor and one outcome. (PubMed Central (PMC))

Explicit modelling of measurement is another major strength. Business and social research frequently depends on abstract constructs such as trust, satisfaction and commitment. SEM allows researchers to distinguish those constructs from their imperfect observed indicators and to incorporate measurement error within the analytical model.

The same flexibility creates considerable scope for misspecification. Researchers make decisions about constructs, indicators, structural paths, estimators, parameter constraints and model modifications. A technically estimable model can still embody weak theory, inadequate measurement or implausible causal assumptions.

SEM also demands adequate data relative to the complexity and characteristics of the model. Sample-size requirements cannot be reduced reliably to one universal rule because they depend on factors such as model complexity, indicator reliability, parameter sizes, missingness, distributional characteristics and estimation method.

Most importantly, SEM does not compensate for a weak research design. A complex model fitted to cross-sectional observational data remains cross-sectional observational evidence. Statistical sophistication cannot create temporal ordering, eliminate omitted-variable bias or turn association into causation.

Common Mistakes When Using Structural Equation Modelling

Beginning with the structural paths while treating measurement as a technical preliminary reverses the logic of latent-variable SEM. If constructs are represented poorly, relationships among those constructs become difficult to interpret regardless of how impressive the structural coefficients appear.

Another problem arises when researchers equate acceptable global fit with confirmation of the complete theoretical model. Model fit indicates how well aspects of the specified structure correspond to the observed data under the model assumptions; it does not prove that the theoretical explanation is uniquely correct.

Directional arrows can also encourage causal overstatement. A path from X to Y is a model specification, not evidence by itself that X causes Y. The causal interpretation of SEM, particularly with nonexperimental data, has been debated precisely because theory, design and identifying assumptions remain indispensable. (PubMed Central (PMC))

Modification indices create a different temptation. Repeatedly adding paths or correlated residuals until fit becomes acceptable can transform theory testing into data-driven model construction. If modifications are required, their theoretical basis and exploratory character should be reported honestly.

Finally, reporting only whether paths are statistically significant can obscure substantive interpretation. Effect magnitude, uncertainty, indirect relationships, measurement quality, model assumptions and research design all contribute to understanding what an SEM actually demonstrates.

Structural Equation Modelling in Business Research

SEM is particularly well suited to many business-research questions because business theories frequently involve latent constructs connected through multistage relationships.

A marketing researcher might hypothesize:

Service Quality → Perceived Value → Satisfaction → Loyalty

An organizational researcher might investigate:

Leadership Support → Employee Engagement → Organizational Commitment → Turnover Intention

A technology-adoption study might propose:

Perceived Ease of Use → Perceived Usefulness → Adoption Intention

In each case, several constructs may be represented through multiple questionnaire indicators, while the theory concerns relationships among those constructs.

SEM allows the measurement and structural components to be analysed within an integrated framework. It can also accommodate mediation and multiple outcomes, capabilities that have contributed to its widespread use across social and behavioural research. (PubMed Central (PMC))

Its popularity does not mean SEM should automatically be selected whenever a dissertation contains several hypotheses. The method should follow the theoretical and measurement problem rather than serve as a marker of statistical sophistication.

Structural Equation Modelling in the Age of AI and Digital Research

Generative AI has sharply reduced the technical barrier to SEM. Researchers can generate lavaan, Mplus or other syntax, translate path diagrams into code, interpret output, troubleshoot convergence problems and receive explanations of direct and indirect effects within seconds.

This makes SEM more accessible, but it also magnifies an existing methodological risk: the software can estimate a model that the researcher cannot genuinely defend.

AI can suggest adding a covariance, deleting an indicator or introducing a structural path because the modification improves statistical fit. Unless the researcher understands the theoretical consequences, this can rapidly convert a confirmatory model into an optimized representation of one dataset.

AI-generated theoretical models present an even deeper problem. A language model can produce a plausible diagram connecting trust, satisfaction, perceived value and loyalty, complete with academic-sounding hypotheses. Plausibility is not theoretical justification. The relationships still require grounding in relevant literature, conceptual reasoning and an appropriate research design.

The strongest use of AI is therefore as a methodological critic and technical assistant. It can check syntax, explain alternative estimators, identify assumptions, challenge causal interpretations and ask whether a proposed modification has theoretical support.

SEM makes this distinction especially important because computational complexity can disguise conceptual weakness. The researcher should be able to defend the model without relying on the authority of either statistical software or AI.

When to Use Structural Equation Modelling

SEM may be appropriate when:

• the research involves theoretically specified relationships among several variables or constructs;
• important constructs are latent and measured through multiple indicators;
• researchers need to model measurement error explicitly;
• a measurement model and structural model need to be evaluated within an integrated framework;
• several dependent relationships need to be estimated simultaneously;
• the research involves theoretically justified mediation or indirect effects;
• the researcher wants to compare plausible theoretical models;
• the sample, measurements and research design are adequate for the proposed model;
• substantive theory is sufficiently developed to justify the model specification.

SEM is not automatically preferable to regression or simpler statistical methods. If the research question can be answered adequately with a simpler model, additional complexity does not necessarily improve the study.

Dissertation Example

A dissertation titled “The Influence of Digital Service Quality on Customer Loyalty: The Mediating Roles of Trust and Satisfaction” proposes four latent constructs measured through multi-item scales. Based on previous literature, the researcher hypothesizes that digital service quality contributes to trust, trust contributes to satisfaction, and satisfaction contributes to customer loyalty. Direct and indirect relationships are specified before analysis.

The methodology chapter first explains how the latent constructs were operationalized and justifies CFA for evaluating the measurement model. Factor loadings, relevant model-fit evidence, reliability and construct-validity evidence are examined before structural relationships are interpreted. The researcher explicitly distinguishes this measurement stage from evaluation of the structural hypotheses.

After the measurement model is judged adequate, the structural model is estimated. Standardized path coefficients and their uncertainty are reported, while indirect effects are evaluated for the proposed mediating relationships. Model fit is interpreted alongside the theoretical coherence of the estimated relationships rather than being treated as a simple pass/fail test.

Because the dissertation is based on cross-sectional questionnaire data, the researcher avoids claiming that SEM proves the causal sequence implied by the arrows. The findings are described as supporting or being consistent with the hypothesized relationships within the limitations of the design. Longitudinal research is recommended to investigate temporal ordering more convincingly.

The methodology chapter therefore demonstrates the full logic of SEM:

Theory → Measurement → Measurement Evaluation → Structural Model → Structural Evaluation → Cautious Interpretation

Exam Tip

If asked what distinguishes SEM from ordinary regression, do not answer simply that “SEM analyses more variables.”

A stronger answer explains that SEM provides a general framework for modelling systems of relationships and can combine latent-variable measurement models with structural relationships among variables or constructs. It can also estimate multiple equations and direct and indirect effects simultaneously. (PubMed Central (PMC))

Remember the distinction:

Measurement model → How are the constructs measured?

Structural model → How are the constructs hypothesized to relate?

And remember the methodological order:

Establish measurement before interpreting structure.

The final principle is especially important:

SEM can evaluate a theoretically specified pattern of relationships; it cannot turn an inadequate research design into causal evidence.

Build a methodology you can explain and defend

Structural equation modelling involves interconnected decisions about theory, measurement, model specification, estimation and interpretation. Dudovskiy Research Assistant can help you determine whether SEM fits your dissertation and structure the methodological reasoning behind the model you need to defend.

References

Bollen, K.A. (1989). Structural Equations with Latent Variables. New York: Wiley.

Kline, R.B. (2023). Principles and Practice of Structural Equation Modeling. 5th ed. New York: Guilford Press.

Westland, J.C. (2019). Structural Equation Models: From Paths to Networks. Cham: Springer. (PubMed Central (PMC))

Tarka, P. (2018). An overview of structural equation modeling: its beginnings, historical development, usefulness and controversies in the social sciences. Quality & Quantity, 52, 313–354. (PubMed Central (PMC))

Gunzler, D., Chen, T., Wu, P. and Zhang, H. (2013). Introduction to mediation analysis with structural equation modeling. Shanghai Archives of Psychiatry, 25(6), 390–394. (PubMed Central (PMC))