aiwiki.page
English
Mathematics / Probability & Statistics / structural-equation-modeling

Structural Equation Modeling

Structural equation modeling combines measurement models and systems of equations to analyze relationships among observed and latent variables.

41 keywords3 linked from7 not yet writtenWritten by AI
StatisticsEconometricsCovariance matri…Identity MatrixMatrix TransposePrincipal compon…Statistical Iden…Degrees of Freed…Structural…

Structural equation modeling (SEM) is a family of methods in statistics for representing and estimating relationships among observed variables and unobserved, or latent, variables. A model specifies equations for how variables depend on one another, together with assumptions about their variability, measurement error, and unexplained influences. Its defining feature is the integration of a measurement model, which connects latent constructs to observations, with a structural model, which describes relationships among variables. Models containing only observed variables also belong to the family.

SEM is widely used in psychology, sociology, education, management, and related fields where concepts such as ability, socioeconomic position, or organizational commitment cannot be measured directly. It also has important roots in econometrics, where systems of simultaneous equations represent interdependent economic processes. By fitting an entire system rather than isolated regressions, SEM makes assumptions about measurement and dependence explicit and allows many of their observable implications to be tested jointly.

The most familiar form is covariance-based SEM: parameters are estimated so that a model-implied covariance matrix, and often a mean vector, reproduces the corresponding features of the data. Broader formulations accommodate categorical outcomes, nonlinear relationships, multilevel observations, longitudinal processes, and Bayesian estimation. A fitted model is not automatically a causal explanation; interpreting its paths causally requires assumptions and research designs that cannot generally be established from model fit alone.

Historical foundations and development

SEM developed from several traditions rather than from a single invention. One was path analysis, introduced by the geneticist Sewall Wright in work published in 1918 and developed systematically in his 1921 paper “Correlation and Causation.” Wright used diagrams and equations to decompose observed associations into contributions from hypothesized pathways. Under appropriate linearity and disturbance assumptions, a correlation between two variables could be expressed as a sum of products of path coefficients. This gave researchers a way to articulate complex explanations without treating every association as a separate relationship.

A second foundation was factor analysis. Charles Spearman’s 1904 work on general intelligence helped establish the idea that correlations among test scores could arise from shared unobserved influences. Factor models separated common variation from variation specific to individual indicators. Later developments distinguished exploratory approaches, in which much of the loading structure is estimated with limited prior restriction, from confirmatory approaches, in which the researcher specifies which indicators measure which factors. The latter became a central component of modern SEM because it allowed substantive hypotheses about measurement to be evaluated directly.

A third source was the study of simultaneous equations in economics. Economic quantities such as price and quantity may be jointly determined, making an ordinary regression coefficient difficult to interpret as a structural parameter. Work associated with the Cowles Commission, including Trygve Haavelmo’s 1944 “The Probability Approach in Econometrics,” clarified the role of probabilistic assumptions, identification, and estimation in such systems. The distinction between structural equations and reduced-form relationships remains important in SEM: equations intended to describe mechanisms imply observable distributions, but those distributions need not uniquely reveal the mechanisms.

Karl Jöreskog’s 1969 work on maximum-likelihood confirmatory factor analysis and the development of LISREL with Dag Sörbom during the 1970s helped integrate these traditions computationally. Researchers could estimate latent measurement structures and relationships among factors within one framework. Covariance structure analysis became increasingly accessible as computing improved, while later software expanded support for means, categorical responses, missing observations, and more complex sampling structures.

The field subsequently diversified. Peter Bentler’s work contributed to estimation and model evaluation, and software such as EQS, Amos, Mplus, OpenMx, and the R package lavaan made different formulations available to broad research communities. A parallel tradition associated with Herman Wold developed partial least squares path modeling. Although frequently discussed under the SEM label, that tradition generally constructs weighted composites and uses estimation objectives different from those of conventional common-factor, covariance-based SEM.

Mathematical formulation and path diagrams

A general linear SEM can be expressed compactly using two equations. Let yy be a vector of observed variables and η\eta a vector containing the model’s latent variables. A measurement equation is

y=ν+Λη+ϵ,y=\nu+\Lambda\eta+\epsilon,

where ν\nu contains indicator intercepts, Λ\Lambda contains factor loadings, and ϵ\epsilon contains measurement residuals. A structural equation is

η=α+Bη+ζ,\eta=\alpha+B\eta+\zeta,

where α\alpha contains structural intercepts, BB contains directed relationships among latent variables, and ζ\zeta contains disturbances. Variables without modeled predictors can be represented by rows of zeros in BB. An alternative traditional notation separates exogenous latent variables from endogenous ones and gives their coefficients a separate matrix.

If I−BI-B is invertible, the structural equation has the reduced form

η=(I−B)−1(α+ζ).\eta=(I-B)^{-1}(\alpha+\zeta).

Assuming zero-mean disturbances and measurement residuals, and assuming the latter are uncorrelated with ζ\zeta, the model implies

μ(θ)=ν+Λ(I−B)−1α\mu(\theta)=\nu+\Lambda(I-B)^{-1}\alpha

and

Σ(θ)=Λ(I−B)−1Ψ(I−B)−TΛT+Θ,\Sigma(\theta) = \Lambda(I-B)^{-1}\Psi(I-B)^{-\mathsf T}\Lambda^{\mathsf T} +\Theta,

where Ψ=Cov⁡(ζ)\Psi=\operatorname{Cov}(\zeta), Θ=Cov⁡(ϵ)\Theta=\operatorname{Cov}(\epsilon), and θ\theta collects the free parameters. Here II is the identity matrix, and the superscript T\mathsf T denotes a matrix transpose. These expressions show why SEM is often described as covariance structure analysis: a comparatively small parameter set generates a structured pattern of observable means and covariances.

Path diagrams provide a graphical representation of these equations. Rectangles conventionally denote observed variables, circles or ellipses denote latent variables, single-headed arrows denote directed coefficients, and double-headed arrows denote covariances or variances. An omitted arrow usually represents a fixed-zero restriction, not merely a relationship that has been left unreported. A curved arrow between disturbances, for example, allows residual association after the modeled predictors have been taken into account.

The graphical notation is useful but not self-sufficient. A diagram may omit intercepts, distributional assumptions, equality constraints, or details of how exogenous variables are handled. The same-looking arrow can also have different interpretations in different models. In a purely associational model it denotes a regression relationship; in a causal model it is intended to represent an effect of intervention. A complete specification therefore requires equations and assumptions as well as a diagram.

Models without directed cycles are called recursive when the relevant disturbance conditions also hold. Nonrecursive models allow feedback, such as reciprocal relationships between economic variables. Invertibility of I−BI-B permits a reduced-form solution, but it does not by itself establish a dynamically stable process. Stability requires additional assumptions about how adjustment unfolds over time.

Latent variables and the measurement model

A latent variable is not directly observed but is represented through its relationship with observable indicators. In confirmatory factor analysis (CFA), the researcher specifies the indicator–factor pattern in advance. For example, three vocabulary tasks might load on a verbal-ability factor and three spatial tasks on a spatial-ability factor. The two factors may correlate, while cross-loadings are fixed to zero. The model then predicts both the covariances within each group of tasks and those between the groups.

A factor loading expresses the expected change in an indicator for a one-unit change in the latent variable, holding the other modeled influences fixed. Its numerical value depends on the factor’s scale. A common convention fixes one loading per factor to 1, making that indicator a marker variable. Another fixes the factor variance to 1. These conventions change the units of the latent variable, not the observable implications of an otherwise equivalent model. Standardized loadings rescale variables using their estimated variances, but their magnitude depends on the population and on the fitted model.

In a simple one-factor model,

yj=νj+λjη+ϵj,y_j=\nu_j+\lambda_j\eta+\epsilon_j,

the indicator variance is

Var⁡(yj)=λj2Var⁡(η)+Var⁡(ϵj),\operatorname{Var}(y_j) = \lambda_j^2\operatorname{Var}(\eta) + \operatorname{Var}(\epsilon_j),

provided the factor and residual are uncorrelated. The first term is factor-explained variance. The second can include random measurement error, indicator-specific stable influences, and other omitted sources of variation. Consequently, a measurement residual is not necessarily synonymous with purely random error. Correlated residuals may represent shared wording, common methods, or repeated administration, but allowing them changes the measurement theory.

Reflective and formative specifications embody different ideas. In a reflective model, the latent variable explains covariance among its indicators: changes in the construct are expected to influence the measurements. In a formative specification, measured components help define or produce a construct. An index of material resources, for instance, may combine income, housing conditions, and durable-goods ownership without requiring those components to behave as interchangeable manifestations of one common cause. Formative specifications require particular attention to identification and to relationships with other variables.

Latent factors also differ from principal components. A principal component is a weighted combination of observed measurements chosen to account for observed variance; a common factor is part of a model for the sources of covariance. A component therefore does not automatically separate common variation from measurement-specific variation.

Measurement quality is broader than statistical fit. A well-fitting factor model does not alone establish that a scale covers the intended domain, excludes competing interpretations, or behaves appropriately across populations. Conversely, imperfect global fit may arise from a localized measurement feature rather than a failure of every aspect of the construct. SEM makes these measurement assumptions inspectable, but it does not replace substantive evidence about what the indicators measure.

Identification, constraints, and observational equivalence

Identification concerns whether distinct parameter values imply distinct distributions of the observed variables. A model can be mathematically specified yet fail to identify its parameters. For instance, multiplying a factor by a constant and dividing all its loadings by that constant leaves its observable predictions unchanged unless a scale constraint is imposed. Fixing a marker loading or factor variance removes this particular indeterminacy, but other identification problems may remain.

For pp continuous observed variables, the covariance matrix contains p(p+1)/2p(p+1)/2 distinct elements. If the model estimates qq independent covariance parameters, its nominal degrees of freedom are

df=p(p+1)2−q.df=\frac{p(p+1)}{2}-q.

Including an unrestricted observed mean vector adds pp observed moments; estimating corresponding mean parameters uses those moments. Nonnegative degrees of freedom are generally a necessary counting condition for an ordinary covariance model, not a sufficient proof of identification. Two different parameters may influence the observed moments in exactly the same way even when the overall count appears adequate.

A one-factor model with three indicators illustrates this distinction. Fixing the factor variance to 1 leaves three loadings and three residual variances: six parameters for six observed covariance elements. Under suitable nonzero-loading and regularity conditions, the model is locally identified and just identified. It has no covariance degrees of freedom with which to test its overall fit. Boundary cases, including zero loadings, can undermine the generic identification argument.

A two-factor model with three indicators per factor provides an overidentified example. Fixing one loading on each factor leaves four free loadings, two factor variances, one factor covariance, and six residual variances. Its 13 free parameters are fitted to 21 distinct covariance elements, giving 8 degrees of freedom when all cross-loadings and residual covariances are fixed to zero. These restrictions create testable implications, but those implications depend on the precise measurement structure.

Local identification can be studied through the Jacobian matrix of the mapping from parameters to implied moments. A rank deficiency indicates that some parameter changes cannot be distinguished through the observed moments. Global identification is stronger: it excludes distinct parameter solutions anywhere in the relevant parameter space. Sign reversals of latent variables, symmetries, and feedback structures can create complications beyond simple parameter counting.

Even fully identified models can be observationally equivalent to other models with different interpretations. With only two observed variables and unrestricted variances, a regression of one on the other can reproduce the same covariance matrix as the reverse regression. More complex diagrams can also share all observable covariance implications. Identification within a selected model therefore does not establish that the selected model is uniquely supported among all possible explanations. Additional measurements, temporal information, interventions, or theoretically justified restrictions may distinguish models that cross-sectional covariance data cannot separate.

Estimation, distributional assumptions, and missing observations

In covariance-based SEM, estimation searches for parameter values that make the model-implied moments close to the sample moments. For continuous data, maximum likelihood under a multivariate normal distribution is a standard approach. Ignoring the mean structure for simplicity, a common discrepancy function is

FML=log⁡∣Σ(θ)∣+tr⁡ ⁣[SΣ(θ)−1]−log⁡∣S∣−p,F_{\mathrm{ML}} = \log|\Sigma(\theta)| + \operatorname{tr}\!\left[S\Sigma(\theta)^{-1}\right] - \log|S| - p,

where SS is the sample covariance matrix, vertical bars denote determinants, and tr⁡\operatorname{tr} denotes the trace. Numerical optimization minimizes this function subject to the model’s restrictions.

Normal-theory maximum likelihood can provide consistent estimates of correctly specified mean and covariance parameters under broader conditions than exact normality, but its conventional standard errors and test statistics may then be incorrect. Robust procedures adjust estimated parameter uncertainty and model tests for certain forms of nonnormality. “Robust” does not mean assumption-free: these procedures do not automatically correct omitted variables, dependence between observations, severe model misspecification, or inappropriate treatment of measurement scales.

Binary and ordinal indicators require special treatment. A common model assumes an underlying continuous response divided by thresholds into observed categories. Limited-information methods fit thresholds and associations such as polychoric correlations. Diagonally weighted least squares uses a simplified weight matrix during estimation, with appropriate versions supplying robust standard errors and adjusted tests. Full-information categorical models instead construct likelihoods for the observed responses, often requiring numerical integration over latent variables. These approaches differ in computational demands and assumptions.

Missing observations can be handled using full-information maximum likelihood, which evaluates each case’s likelihood using the variables observed for that case. Under an appropriate missing-at-random assumption and a correctly specified model, this avoids discarding every partially observed record. Missing at random means that missingness does not depend on unobserved values after conditioning on the relevant observed information; it is not equivalent to missing completely at random. Multiple imputation is another approach, provided the imputation model is compatible with the analysis and incorporates relevant predictors of missingness. Neither method automatically resolves missingness that depends on unobserved values.

Bayesian inference combines a likelihood with prior distributions to obtain a posterior distribution over model parameters. Computation often uses Markov chain Monte Carlo. Priors can express substantive knowledge or stabilize weakly estimated models, but they also influence results and require examination. Posterior predictive checks assess whether replicated data resemble the observations in relevant respects.

All these estimators can encounter computational problems. Nonconvergence, nearly singular information matrices, negative residual variance estimates, or latent correlations outside admissible bounds may indicate weak identification, sparse information, misspecification, or optimization difficulties. A program’s production of numerical estimates is therefore not sufficient evidence that the fitted solution is valid.

Assessing model fit and uncertainty

Model evaluation concerns both the overall restrictions and the individual estimates. For normal-theory maximum likelihood, a scaled version of the minimized discrepancy function—commonly (N−1)FML(N-1)F_{\mathrm{ML}} under a usual covariance convention—has an asymptotic chi-squared distribution under the correctly specified model and regularity conditions. The associated hypothesis test evaluates exact fit: whether all population restrictions imposed by the model hold simultaneously.

The chi-squared test is sensitive to sample size. With a very large sample, small discrepancies may produce rejection; with a small sample, important discrepancies may escape detection. This is not a defect unique to SEM but a consequence of testing exact restrictions with different amounts of information. Failure to reject does not establish that the model is true, and rejection does not identify which assumption is responsible.

Several descriptive indices supplement the test. The root mean square error of approximation, or RMSEA, estimates a degree-of-freedom-adjusted measure of population discrepancy. A common estimate is

RMSEA^=max⁡(χ2−dfdf(N−1),0).\widehat{\mathrm{RMSEA}} = \sqrt{ \max\left( \frac{\chi^2-df}{df(N-1)},0 \right) }.

Its confidence interval conveys sampling uncertainty, which can be substantial for small samples or low-degree-of-freedom models. The comparative fit index and Tucker–Lewis index compare the fitted model with a baseline model, usually one imposing independence among observed variables. The standardized root mean square residual summarizes standardized discrepancies between observed and implied associations.

Frequently quoted cutoffs—such as a comparative fit index near .95, RMSEA near .06, or standardized residual measure near .08—derive from particular simulation settings and are not universal acceptance rules. Their behavior depends on loading strength, model size, estimator, distribution, and the type of misspecification. A model may satisfy a conventional cutoff while containing an important incorrect path or measurement restriction. Conversely, an index may look unfavorable because of features that are not central to the substantive question.

Local diagnostics are therefore essential. Residual covariance patterns identify portions of the data that the model reproduces poorly. Modification indices estimate the potential improvement from freeing a fixed parameter, usually through a local score-test approximation. Expected parameter changes describe the likely size of that change. These diagnostics can reveal plausible omissions, but repeated data-driven modifications can capitalize on chance and transform a confirmatory analysis into an exploratory one.

Parameter uncertainty is examined through standard errors, confidence intervals, and, where appropriate, bootstrap distributions. Nested models can be compared using likelihood-ratio tests, although robust estimators require appropriate adjusted difference tests rather than naive subtraction of corrected statistics. Information criteria compare relative fit and complexity among models fitted to compatible data and likelihoods. Cross-validation evaluates how well a specification generalizes beyond the observations used to develop it, addressing a question different from in-sample exact fit.

Structural paths, mediation, and causal interpretation

A structural coefficient is first a parameter in a specified system of equations. It becomes a causal effect only when the system can defensibly represent how outcomes would change under intervention. Causal inference therefore requires more than a good covariance fit: it requires assumptions about temporal order, omitted causes, measurement, selection, and the meaning of changing a variable.

A simple mediation model illustrates the distinction. Let an exposure XX predict a mediator MM, which in turn predicts an outcome YY:

M=iM+aX+uM,M=i_M+aX+u_M,
Y=iY+c′X+bM+uY.Y=i_Y+c'X+bM+u_Y.

In a linear additive model without an exposure–mediator interaction, the model-based indirect effect is abab, the direct effect is c′c', and the total effect is c′+abc'+ab, under the relevant structural assumptions. These expressions provide an algebraic decomposition. Their interpretation as causal mediation additionally requires conditions concerning confounding of the exposure–mediator, exposure–outcome, and mediator–outcome relationships.

Randomizing an exposure can address confounding of that exposure, but it does not automatically randomize the mediator. An unmeasured common cause of mediator and outcome can still bias bb and thus the product abab. Some causal mediation definitions also require assumptions excluding exposure-induced confounders of the mediator–outcome relationship. These conditions are not certified by a significant indirect effect or by a well-fitting path diagram.

Indirect effects often have asymmetric sampling distributions because they are products of coefficients. Symmetric normal approximations may therefore be inadequate in some settings. Bootstrap intervals or model-based simulation can better represent this asymmetry, although their validity still depends on the sampling structure and model. A statistically detectable indirect effect does not establish that the proposed mediator is the only mechanism, and a nonsignificant total effect does not logically exclude indirect pathways of opposite signs.

Directed acyclic graphs clarify the distinction between conditioning and intervention. Conditioning on a common effect can induce an association between its causes, while adjusting for a mediator changes the estimand from a total relationship to a more restricted one. Structural equations can express the mechanisms underlying such graphs, but conventional SEM diagrams do not always make all causal assumptions explicit. Correlated disturbances, for example, may represent unmeasured common causes.

Reciprocal systems introduce further difficulties. A pair of equations in which XX predicts YY and YY predicts XX generally needs additional identifying restrictions, such as valid instrumental variables or excluded predictors. A single cross-sectional covariance matrix usually cannot determine reciprocal causal effects without such information. Temporal ordering, experimental design, and sensitivity analyses provide evidence complementary to model fit rather than substitutes for it.

Multiple groups and measurement invariance

Multiple-group SEM estimates a common model framework across populations or conditions while allowing selected parameters to vary. Groups might represent language versions of a questionnaire, treatment conditions, or educational settings. The central statistical device is an equality constraint: a parameter such as a factor loading can be estimated separately in each group or constrained to a common value.

Comparisons involving latent constructs depend on measurement invariance. If an indicator relates differently to a construct across groups, a difference in observed scores may reflect measurement differences rather than a difference in the construct itself. Invariance analysis examines which aspects of the measurement model can reasonably be treated as shared. It does not establish that groups are identical or that all cultural and contextual influences have been removed.

A common sequence begins with configural invariance, which uses the same general pattern of factor–indicator relationships. Metric invariance adds equality constraints on factor loadings. Scalar invariance additionally constrains indicator intercepts in continuous-indicator models. Strict invariance also constrains residual variances. These stages address different questions: comparable loading units are important for comparing structural relationships, while latent mean comparisons ordinarily require suitable intercept restrictions as well as loading restrictions.

The reason for scalar restrictions can be seen from

E(yg)=νg+Λgμg,E(y_g)=\nu_g+\Lambda_g\mu_g,

where gg indexes groups. If intercepts νg\nu_g are allowed to change freely, observed mean differences can be attributed either to intercept differences or to latent mean differences μg\mu_g. Appropriate constraints separate these possibilities. A reference group’s latent mean is commonly fixed to zero, while means in other groups are estimated relative to that origin. Factor scaling must also be coordinated across groups.

Complete invariance is not always supported. Partial invariance models release selected constraints while retaining enough invariant measurement structure to identify the intended comparisons. Their interpretation depends on which parameters are released and which indicators anchor the latent scale. A finding of partial invariance is therefore more informative when reported as a specific pattern than as a simple pass-or-fail label.

For ordinal indicators, thresholds replace the simple role played by continuous-indicator intercepts, and invariance constraints interact with latent-response scaling. The continuous-data testing sequence cannot always be transferred unchanged. Large-scale comparisons may also use approximate invariance or alignment methods, which permit limited departures from exact equality under their own assumptions.

Structural equality tests follow a separate logic. Once the measurement scale is adequately linked, regression coefficients, factor variances, covariances, and means can be compared through explicit constraints. A coefficient being statistically significant in one group and nonsignificant in another does not itself demonstrate a between-group difference; the difference must be estimated or tested directly. Group imbalance, limited precision, and different factor distributions can affect these comparisons even when the measurement specification is otherwise appropriate.

Longitudinal, multilevel, and nonlinear extensions

Longitudinal SEM represents measurements collected from the same units over time. An autoregressive path links a variable to its later value, while a cross-lagged path links one variable to a later measurement of another. These models can describe temporal prediction while accounting for specified prior measurements. Their coefficients depend on measurement intervals and included variables, so a path estimated from annual observations is not automatically comparable to one estimated from daily observations.

The traditional cross-lagged panel model can mix stable between-person differences with within-person fluctuations. For example, people who are generally more confident may also generally perform better, even if temporary increases in confidence do not predict subsequent improvements in performance within the same person. Random-intercept cross-lagged models explicitly separate stable person-specific components from occasion-specific deviations. This changes the question addressed by the lagged coefficients, but it does not eliminate time-varying confounding or guarantee a causal interpretation.

Latent growth models describe individual trajectories through latent intercepts and slopes. A linear example is

yit=η0i+tη1i+ϵit,y_{it}=\eta_{0i}+t\eta_{1i}+\epsilon_{it},

where η0i\eta_{0i} is person ii’s latent level at the chosen time origin and η1i\eta_{1i} is that person’s rate of change. Their means describe the average trajectory, their variances describe differences between individuals, and their covariance describes how initial level relates to change. Quadratic terms, piecewise slopes, or freely estimated time scores allow other trajectory shapes. The meaning of an intercept depends on where time is centered.

Multilevel SEM addresses nested observations, such as students within schools or repeated assessments within individuals. It separates covariance components at different levels and can place different factor or structural models at each level. A predictor’s association across schools need not equal its association among students within a school. Merely adjusting standard errors for clustering does not estimate this separation; it addresses dependence in uncertainty calculations while leaving the modeled association structurally unchanged.

Nonlinear extensions include latent interactions, nonlinear growth, and categorical outcomes. A model in which the effect of one latent variable depends on another requires integrating over the latent predictors or otherwise accounting for uncertainty in their values. Substituting estimated factor scores into an ordinary interaction regression generally does not reproduce a joint latent-variable analysis. With nonlinear links, indirect effects and marginal effects may vary across predictor values, and the simple linear product-of-coefficients interpretation requires qualification.

Other extensions include mixture SEM, dynamic SEM, and continuous-time formulations. Mixture models posit unobserved subpopulations but can sometimes approximate skewness or other distributional features rather than discover substantively distinct classes. Dynamic models connect SEM with time-series analysis and state-space approaches. Continuous-time models describe evolution between observations and can accommodate unequal intervals, provided their assumptions about the underlying process are appropriate.

Research design, computation, and reporting

The feasibility of an SEM depends on the information in the data, not simply on the number of observations divided by the number of parameters. Strong indicators, balanced response categories, sufficient variation, and well-separated constructs can support more precise estimation. Weak loadings, rare categorical responses, high factor correlations, missing observations, or complex clustering can make a nominally large sample uninformative for particular parameters. Consequently, universal rules such as a fixed minimum of 200 cases or ten observations per parameter are not reliable design criteria.

Monte Carlo simulation provides a model-specific way to study power and estimation performance. A proposed population model generates repeated datasets under plausible conditions; each dataset is analyzed using the intended estimator. The resulting convergence rates, parameter bias, interval coverage, and rejection rates reveal how the planned analysis behaves under those assumptions. Simulations can vary uncertain features, including loading strength, missingness, cluster size, or the magnitude of an indirect effect.

Computational checks are part of the analysis rather than a final formality. These include convergence diagnostics, examination of residual variances and factor correlations, verification that the implied covariance matrix is admissible, and inspection of parameter uncertainty. Starting values or alternative optimization settings can help diagnose difficult likelihood surfaces, but a numerically improved solution does not repair an unidentified or substantively incoherent model. Bayesian models require additional checks of simulation mixing and the influence of priors.

Transparent reporting specifies the observed variables, measurement scales, sample structure, missing-data treatment, estimator, parameter constraints, latent scaling, and any robust corrections. It distinguishes unstandardized estimates from standardized ones and reports uncertainty rather than only significance labels. Fit statistics are most interpretable when accompanied by degrees of freedom, the baseline model definition where relevant, residual diagnostics, and explanations of modifications made after examining the data.

Model development and confirmation should also be distinguished. Searching among many residual correlations or paths can improve apparent fit while reducing replicability. Preregistered specifications, independent validation samples, or explicitly separated exploratory and confirmatory stages help clarify the evidential status of a model. Regularization can shrink or select parameters in large models, but tuning and selection introduce additional uncertainty and do not by themselves supply a substantive measurement theory.

Partial least squares SEM requires particularly clear labeling because composite-based and common-factor methods can target different quantities. A weighted score defined by its indicators is not interchangeable with an unobserved common cause of those indicators. Neither a large model nor a predictive objective alone determines which formulation is appropriate. Interpretation ultimately depends on matching the model’s mathematical objects—factors, composites, disturbances, and directed coefficients—to the research question and the process by which the observations were produced.