MerQuR+Statistics Glossary

Statistics Glossary → Structural Equation Modeling (SEM)

Structural Equation Modeling (SEM)analysis

An analysis that defines concepts which cannot be measured directly through several observed variables and tests the hypothesised direct and indirect causal relations among those concepts together, in a single model.

Concept

Many research questions concern relations between concepts that cannot themselves be measured directly: does workload increase burnout, and does burnout in turn feed the intention to quit? Each of these concepts is measured with several questionnaire items, and the items contain measurement error. Structural equation modelling (SEM) estimates at once which items measure which concept (the measurement model) and how the concepts affect one another (the structural model). Relations are thus estimated free of measurement error, and the fragmented picture produced by running several separate regressions disappears.

The model writes the relations between variables as a set of equations. The measurement part resembles confirmatory factor analysis (item = loading · factor + error); the structural part consists of regressions among latent variables (burnout = β · workload + disturbance). The covariance matrix implied by these equations is compared with the observed one, and fit is judged with chi-square, CFI, TLI, RMSEA and SRMR. The effect of one variable on another through a mediator is expressed as an indirect effect, the product of the two paths (a · b).

SEM does not prove causation: the researcher draws the arrows from theory, and the data only say whether this structure contradicts the observed relations. Several models with arrows in different directions may fit the same data equally well. An adequate sample (usually several hundred people for complex models), a measurement model that works well on its own first, and theoretical justification for any changes made to improve fit are essential for a trustworthy result.

When to use it

  • When a theory-derived path model with several latent variables and mediators is to be tested as a whole, separating direct from indirect effects.
  • If only the factor structure of a scale is being tested, confirmatory factor analysis is enough; if every variable is observed with a single measure, path analysis or multiple regression can be used.
  • If the aim is prediction rather than theory testing, the sample is small, or constructs are defined by formative indicators, variance-based PLS-SEM may be preferred.

An everyday example

In a hospital, 400 nurses answer questions on workload, supervisor support, burnout and intention to leave, each concept measured by three or four items. The researcher's model is: workload raises burnout, supervisor support lowers it, and burnout raises the intention to leave. SEM tests whether this network fits the data and might show, for example, that most of workload's effect on the intention to leave runs not directly but through burnout (an indirect effect). That finding suggests the hospital should consider measures to reduce burnout, not just the number of shifts.

Parameters in MerQur

Model syntax

Concept: The model syntax is the theoretical model written as text: which items measure which latent variable, which variable predicts which, and which variables are freely correlated are each stated. Every parameter to be estimated, and whether the model can be tested at all, follows from this specification.

Everyday example: In the nurse study you write lines such as 'burnout is measured by three burnout items', 'burnout is predicted by workload and supervisor support' and 'intention to leave is predicted by burnout'. An arrow left out is a claim that that effect is zero.

What it does: Defines the model with measurement (=~), regression (~), (co)variance (~~), intercept (~ 1) and defined-parameter (:=) lines; the same label gives an equality constraint, a number a fixed value; the first indicator is the marker. The model should be prespecified from theory (Kline, 2016).

Effect on the result: Determines the model itself, the parameter table, the degrees of freedom and fit.

On the validation data: PoliticalDemocracy chi-square(35) = 38.125, p = 0.329. With d1-d3 equality labels chi-square(38) = 40.179. Indirect effect ind := a*b = 1.2418 (delta SE 0.3554), total effect 1.8141 (SE 0.3736).

→ Measurement model and structural model

Estimator (ML, MLR)

Concept: The estimation method determines the principle by which parameters are computed and how their uncertainty is measured. Maximum likelihood (ML) assumes multivariate normality; its robust variant (MLR) gives the same coefficients but corrects the standard errors and fit test for non-normality.

Everyday example: If most nurses ticked the top end of the burnout items, the answers are skewed. MLR estimates the effect of workload the same way but shows more honestly how precise that effect is.

What it does: ML is normal-theory maximum likelihood (expected-information SE); MLR gives the same estimates with Huber-White sandwich SE and the Yuan-Bentler scaled chi-square. Used for non-normal continuous data or together with FIML (Yuan and Bentler, 2000).

Effect on the result: Does not change the estimates; SE, z, p, confidence intervals, the scaled chi-square and the scaled CFI/TLI/RMSEA change.

On the validation data: PoliticalDemocracy: scaled chi-square = 41.401 (scale 0.9209), scaled CFI 0.9901; dem60 ~ ind60 SE 0.3423 instead of 0.3992.

→ Robust ML estimators (MLM, MLR; Satorra-Bentler)

Missing data (Listwise deletion, FIML)

Concept: Listwise deletion removes anyone missing any variable. Full information maximum likelihood (FIML) uses each person's available answers and gives less biased results when missingness can be explained by the observed variables (MAR).

Everyday example: Night-shift nurses may have left the last questions blank more often out of tiredness. Listwise deletion drops them and strips the sample of its most exhausted group, whereas FIML keeps using the questions they did answer.

What it does: FIML lets each case contribute to the likelihood with its observed variables, and a mean structure is added. If an exogenous observed variable has missing values, MerQur also models its distribution (noted in the output). FIML is recommended under MAR (Enders, 2010).

Effect on the result: n, the estimates, SE, chi-square and AIC/BIC (the mean parameters are counted).

On the validation data: bfi (exogenous 'age' complete): listwise n = 2617, FIML n = 2800; ind = 0.00204 instead of 0.00208. With exogenous 'education' missing, MerQur uses n = 2800. With complete data the FIML chi-square is the same as ML and SE come from the observed information (0.3973).

→ Full information maximum likelihood (FIML)

Group variable (multi-group)

Concept: Multi-group analysis estimates the same model separately in several groups. This allows you to test whether the relations differ between groups (e.g. an effect being stronger in one group).

Everyday example: If the model is estimated separately for intensive-care and outpatient nurses, you can see whether workload affects burnout more strongly in intensive care.

What it does: Estimates the model separately in each group (configural); a mean structure is added. Used for between-group comparison and before invariance testing.

Effect on the result: Parameters per group, the total chi-square and degrees of freedom.

On the validation data: HS, school group: chi-square(48) = 115.851, CFI 0.9234, RMSEA 0.0969.

→ Measurement invariance (configural, metric, scalar, strict)

Measurement invariance

Concept: Measurement invariance tests whether the scales measure the same concept in the same way across groups; the steps are the same structure, equal loadings, equal intercepts and equal residual variances. This condition should hold before comparing structural relations or means between groups.

Everyday example: If the item 'I feel burnt out' describes ordinary tiredness for intensive-care nurses but a serious condition for outpatient nurses, the two groups' burnout scores cannot be compared directly. The invariance test reveals this.

What it does: Fits the configural, metric, scalar and strict steps in turn; gives the chi-square difference (Satorra-Bentler 2001 under MLR), CFI and RMSEA differences and score tests. Interpreted together with the Chen (2007) criteria (CFI difference <= -0.01, RMSEA difference >= 0.015).

Effect on the result: Adds an invariance table; without a selected group it has no effect and the warning "no group variable selected; invariance not tested" is given.

On the validation data: Metric chi-square difference(6) = 8.192 (p = 0.224), scalar difference(6) = 40.059, strict chi-square(69) = 181.511; under MLR the metric scaled difference 6.567 (p = 0.363).

→ Measurement invariance (configural, metric, scalar, strict)

Partial invariance (free parameters)

Concept: Partial invariance means exempting the few parameters that behave differently across groups from the equality constraints. As long as the remaining parameters are held equal, the group comparison can stay meaningful.

Everyday example: If only one burnout item works differently in the two units, that item's loading or intercept is freed, and the intensive-care versus outpatient comparison continues with the other items held equal.

What it does: Specifies the parameters exempt from equality constraints ('visual =~ x2; x7 ~ 1').

Effect on the result: In the metric, scalar and strict steps the listed parameters stay free.

On the validation data: scalar chi-square(58) = 148.587 (full scalar 164.103).

→ Measurement invariance (configural, metric, scalar, strict)

Confidence Level (%90, %95, %99, custom)

Concept: Sets how much assurance the intervals for path coefficients and defined parameters (e.g. indirect effects) carry. A higher level gives a wider interval.

Everyday example: If the effect of workload on burnout is 0.45, you can say with 95% confidence that the true effect lies between, say, 0.33 and 0.57; asking for 99% widens the range.

What it does: Sets the level of the Wald confidence intervals of the parameters and := lines and of the bootstrap percentile interval. The RMSEA interval is fixed at 90%.

Effect on the result: Only the interval width.

On the validation data: PoliticalDemocracy dem60 ~ ind60 = 1.483: 95% [0.7007, 2.2653], 90% [0.8265, 2.1395], 99% [0.4549, 2.5111], 85.5% [0.9013, 2.0647].

→ Confidence level and confidence interval

Bootstrap (defined parameters)

Concept: An indirect effect is the product of two path coefficients, and the sampling distribution of this product is usually skewed, so a symmetric formula-based interval can mislead. The bootstrap draws many new samples from the data with replacement, recomputes the effect in each, and builds the interval from the percentiles of that distribution.

Everyday example: If the indirect effect of workload on intention to leave through burnout is 0.18, thousands of resamples are drawn from the 400 nurses' surveys to see the range over which this effect wanders. If the interval excludes zero, the indirect effect is considered reliable.

What it does: Computes a percentile confidence interval for := parameters by case resampling (MerQur seed 42; 0 = delta method only). Recommended because of the skewed distribution of indirect effects; B >= 1000 (Preacher and Hayes, 2008).

Effect on the result: Only adds the bootstrap interval and SE under the := lines; the main table does not change.

On the validation data: PoliticalDemocracy ind: B = 500 [0.631, 1.924], SE 0.335; B = 2000 [0.567, 1.956], SE 0.352. The difference between seeds is within the Monte Carlo error.

→ Bootstrap (resampling)

Extra output: Modification indices

Concept: A modification index shows roughly how much the chi-square would drop if a parameter fixed at zero were freed. It helps locate the source of misfit, but changing the model on the basis of the index requires theoretical justification and confirmation in a new sample.

Everyday example: The index may suggest adding a relation between the errors of two burnout items. If the two items ask almost the same thing, the addition is defensible; otherwise it merely writes a chance feature of this dataset into the model.

What it does: Gives the score (LM) test, EPC and standardized EPC for fixed-zero candidates (listwise).

Effect on the result: Additional table only; when off there is no table, the other results are the same.

On the validation data: For PoliticalDemocracy the MI table of 20 candidate free parameters is given.

→ Modification indices

Example on real data

Data: Holzinger and Swineford 1939, lavaan package

Variables: Model syntax: latent factors visual, textual and speed (x1-x9) with paths textual ~ visual and speed ~ visual + textual

  • Question: Does visual ability predict textual ability, and do visual and textual ability together predict the speed factor (a path model with three latent factors)?
  • Model fit: χ²(24, N = 301) = 85.31, p < .001; CFI = .931, TLI = .896, RMSEA = .092 [.071, .114], SRMR = .065. The fit is weak; since the structural part is saturated, the fit is the same as the CFA.
  • Structural paths: textual ← visual b = 0.504 (SE 0.093, p < .001, β = .46); speed ← visual b = 0.297 (SE 0.078, p < .001, β = .43); speed ← textual b = 0.053 (SE 0.053, p = .318, β = .09) is not significant.
  • Explained variance: R² textual = .21, R² speed = .23; about one fifth of the variability in the speed factor is explained by visual ability.
  • Additional finding: The largest modification index is visual =~ x9 (MI = 36.41); the weak fit comes from the measurement model.
  • APA: χ²(24, N = 301) = 85.31, p < .001, CFI = .931, TLI = .896, RMSEA = .092, 90% CI [.071, .114], SRMR = .065; textual ← visual: β = .46, p < .001; speed ← visual: β = .43, p < .001; speed ← textual: β = .09, p = .318.

Related concepts

Bootstrap (resampling)Measurement invariance (configural, metric, scalar, strict)Measurement model and structural modelRobust ML estimators (MLM, MLR; Satorra-Bentler)Full information maximum likelihood (FIML)Modification indicesIndirect effect (mediation)Score (Lagrange multiplier) test

The numbers in the examples are MerQur's output on independent, publicly available validation data; see the MerQur Technical Note for details.