ACADEMIC WRITING GUIDE

How to Conduct and Report Structural Equation Modeling in a Dissertation (With Worked Examples)

By Acadelyra Editorial TeamPublished

Structural equation modeling (SEM) lets you examine a theory about relationships between constructs while also modeling how those constructs are measured. For a dissertation, that can mean investigating whether perceived academic support is associated with engagement directly and indirectly through study confidence, using several indicators for each construct.

The challenge is connecting the research question, measurement evidence, model assumptions, estimation and interpretation. A diagram with arrows is only the beginning. A defensible report explains why the arrows belong there, whether the model can be estimated, what the results imply, and which explanations the design cannot distinguish.

This guide follows that process through one reproducible latent-variable example. All inputs and results in the worked example are synthetic teaching material. No participants were recruited, and the numbers are not empirical dissertation findings. The example uses continuous indicators and conventional maximum likelihood; it is not a template for every survey or research design.

What is SEM, and when should a dissertation use it?

SEM is a family of models that represents relationships among observed variables and, where specified, latent variables. A latent construct is inferred from indicators rather than directly observed. Full latent-variable SEM combines a measurement model with structural regressions among constructs.

Use it when the research question concerns a theoretically defensible system of relationships and the measurement design supports that system. Examples include multiple related outcomes, indirect pathways, or relationships among constructs measured with error. Simpler regression or path analysis may answer your question adequately when the variables are observed scores and a latent measurement model is unnecessary or unsupported.

SEM does not create a sound theory, repair an unsuitable instrument, or establish causation from observational associations. Specify the proposed relationships using your conceptual framework before inspecting the results. Include plausible competing explanations and justify any omitted paths.

Distinguishing related modeling tasks
ApproachMain questionWhat is explicitly modeled?
Path analysisHow are observed variables related in a specified system?Regressions among measured variables; ordinary score-based models do not separately estimate item measurement error
Confirmatory factor analysisDoes a specified measurement structure describe the indicators?Indicator loadings, factor variances and covariances, and specified residual structure
Full latent-variable SEMHow are constructs related, given the measurement model?Measurement parameters and structural regressions estimated together
Exploratory factor analysisWhat common-factor structure might describe this item set?An exploratory loading structure rather than the prespecified structural paths in this guide

These are related members of a broader modeling family, rather than completely separate statistical worlds. If the indicator structure is still uncertain, start with the reasoning in our EFA guide. If the task is testing a proposed measurement structure, our CFA guide provides the detailed measurement workflow. Here, the emphasis is the structural question that follows.

Separate the measurement and structural models

For a continuous reflective indicator, the measurement model can be written as item = intercept + loading × construct + residual. A reflective specification treats the indicators as manifestations of the construct. It is a substantive assumption about what the items mean, not a choice made solely because the software accepts it.

The structural model relates the constructs themselves. In this guide, X represents perceived academic support, M represents study confidence, and Y represents academic engagement:

M = aX + disturbance_M
Y = cprime X + bM + disturbance_Y
Indirect association through M = a × b
Total association = cprime + a × b

These centered, linear equations omit intercepts for readability. The decomposition assumes this additive model, compatible scales, and the specified disturbance relationships. Interactions, nonlinear links, feedback loops or additional mediators require different interpretation and sometimes a different definition of the effect.

The indicator residuals and construct disturbances are different quantities. An indicator residual represents variation outside its assigned factor. A structural disturbance represents construct variation unexplained by its specified predictors. Neither should automatically be interpreted as pure random error.

Specify and identify the model before estimation

Write down the indicator assignments, structural paths, control variables, residual covariances and fixed restrictions. Explain the timing of measurements and why the proposed mediator belongs between the predictor and outcome. Choose controls from substantive knowledge of confounding and the target relationship; indiscriminately controlling for variables can itself introduce bias.

Give every latent construct a scale. One approach fixes a marker indicator loading to one and estimates the factor variance. Other identification choices exist, but switching conventions changes unstandardized coefficients. Report the convention and retain it across model comparisons.

Identification asks whether the specified parameters can, in principle, be uniquely recovered from the available information. Counting observed moments and free parameters is necessary, but it is not sufficient: redundant paths, inadequate indicators or reciprocal relationships can still create identification problems. A positive degrees-of-freedom count does not guarantee a usable model.

In the worked covariance-only model, there are nine observed indicators. The count is 9 × 10 / 2 = 45 distinct covariance moments. The model has 21 free parameters: six non-marker loadings, nine indicator residual variances, one exogenous factor variance, two structural disturbance variances, and three regression paths. Therefore 45 − 21 = 24 degrees of freedom. Fixing the direct path to zero removes one free parameter and gives 25 degrees of freedom. These counts exclude a mean structure, which this example does not estimate.

Also distinguish identification from convergence. A program can stop successfully near a poorly determined or improper solution. Examine warnings, standard errors, the information matrix, admissibility and stability before interpreting coefficients.

Plan sample size, assumptions and missing data

There is no universal sample size that makes SEM valid. Requirements depend on the measurement quality, model complexity, effect sizes, distributions, missingness and design. Muthén and Muthén's Monte Carlo approach examines parameter recovery, interval coverage and power under an explicitly specified model. For mediation, plan around the indirect effect itself; adequate power for a single path does not guarantee adequate power for its product with another path.

Start with defensible parameter ranges, simulate the intended analysis at candidate sample sizes, and record convergence, improper solutions, bias, interval coverage and power for the main estimand. Include weaker loadings, smaller effects and realistic missingness as sensitivity conditions. Our power-analysis guide explains the broader planning logic; an SEM simulation must additionally reflect the measurement and structural model.

Before fitting, inspect coding, reverse-scored items, distributions, unusual observations, missingness patterns and relationships among predictors. Examine whether the linear functional form is plausible. Near-collinearity between constructs can make structural coefficients unstable even when the overall model appears to fit well.

Account for dependence. Students nested in classes, repeated observations within students, or complex survey sampling require suitable multilevel, longitudinal, survey or cluster-adjusted methods. Conventional single-level standard errors and resampling individual rows are inappropriate when they ignore the sampling structure.

Describe missing data by item and analysis variable. For suitable continuous-data models, full-information maximum likelihood can use available observations under an adequately specified model and a defensible missing-at-random assumption. Multiple imputation is another option when the imputation model is compatible with the analysis. Missing at random is conditional on observed information; it does not mean that missingness has no pattern, and it cannot be established from observed data alone. Consider sensitivity analysis when unobserved causes of missingness are plausible. Our missing-survey-data guide develops these decisions.

Choose the SEM framework and estimator

Covariance-based SEM evaluates the covariance structure implied by a specified model. Partial least squares SEM is commonly used with weighted composites. Common factors and composites represent different construct definitions; they should not be interchanged merely to obtain favorable results.

Henseler and Schuberth explain that the distinction is more nuanced than a strict software divide: covariance-based methods can accommodate appropriately specified composites, and consistent PLS addresses common-factor estimation under its assumptions. State the construct model, estimation method and purpose. Small sample size or nonnormality alone does not establish that PLS is appropriate or solve a weak design. Evaluation criteria and coefficients from the different approaches are not automatically interchangeable.

Estimator decisions for covariance-based SEM
Data or analysis issueDefensible starting pointReporting caution
Complete, approximately normal continuous indicatorsConventional maximum likelihood (ML)Report the likelihood and covariance conventions, standard errors and model test
Continuous indicators with nonnormalityA suitable robust ML method, such as MLRIdentify robust standard errors and scaled tests; robust estimation does not repair misspecification
Binary or ordinal endogenous indicatorsAn appropriate categorical estimator, often WLSMVModel thresholds and identify categorical parameters correctly; do not treat ordered categories as automatically continuous
Missing continuous observationsSuitable full-information ML or compatible multiple imputationExplain assumptions, auxiliary information and sensitivity analysis
Clustered or repeated observationsAn estimator and model accounting for the designState the unit of analysis and dependence correction

The official lavaan estimator and categorical-data documentation supports these software distinctions. In lavaan, declaring endogenous indicators as ordered selects WLSMV; its categorical workflow does not currently provide full-information ML. Check the exact method supported by the version you use. The example below deliberately stays with a complete, continuous, single-level normal-ML covariance analysis.

Worked example: support, confidence and engagement

Imagine a theory proposing that support is associated with confidence and that confidence is associated with engagement, alongside a possible direct support–engagement relationship. The illustrative sample size is N = 400, chosen for teaching, not justified by a completed power study.

Hypothetical constructs and indicator meanings
ConstructIndicator namesIllustrative content
X: Perceived supportx1, x2, x3Availability, approachability and usefulness of academic support
M: Study confidencem1, m2, m3Confidence planning study, understanding tasks and persisting with difficulty
Y: Engagementy1, y2, y3Sustained effort, participation and preparation

These are conceptual labels, not a validated questionnaire. All indicators are treated as continuous. The complete positive-definite covariance input appears in the R code below and uses divisor N, with no mean vector. It was constructed with small discrepancies from the fitted simple structure, so the analysis is not an exact-fit demonstration.

The model fixes the first loading of each construct to one, estimates the remaining loadings and indicator residual variances, and sets cross-loadings and indicator residual covariances to zero. The structural disturbances are uncorrelated with each other and with X. The fit is admissible: the estimated variances are positive and the information matrix is nonsingular. That is a prerequisite, not evidence that the constructs are valid in practice.

Evaluate model fit and local diagnostics

Conventional normal-ML fit for the same synthetic covariance input
ModelChi-square (df)pCFITLIRMSEA (90% CI)SRMR
Full: direct and indirect paths32.122 (24)0.1240.9960.9940.029 (0.000–0.053)0.016
Restricted: direct path fixed to zero42.537 (25)0.0160.9910.9870.042 (0.018–0.063)0.032

These are conventional, unscaled normal-ML results. CFI and TLI compare the fitted model with an independence baseline; RMSEA assesses approximate discrepancy per degree of freedom; SRMR summarizes standardized covariance residuals. The reported RMSEA intervals are 90% intervals obtained by inverting the noncentral chi-square distribution. Here SRMR uses the lower covariance triangle including the diagonal, standardized by observed variances, with no mean residuals.

The independence baseline has chi-square = 1993.348 and df = 36. For the full model, the conventional formula gives RMSEA = sqrt((32.122 − 24) / (400 × 24)) = 0.029; calculations retain full precision before rounding. A lower interval bound of zero is a boundary result, not a guarantee of no misspecification.

Do not decide that a model is correct because a set of indices passes familiar cutoffs. Goretzko, Siemund and Sterner show why fit evaluation depends on the measurement model and data characteristics. Examine the residual pattern, loadings, factor correlations, disturbance variances and substantive plausibility alongside overall fit.

In this example the largest absolute standardized covariance residual is 0.060, between x3 and y2. This is a residual covariance divided by the observed standard deviations, not a residual z statistic. It shows a localized discrepancy that deserves attention even when the global summary looks favorable. In a real study, inspect item wording, method effects and omitted cross-loadings before considering a change. Modification indices and expected parameter changes can help locate restrictions to investigate; they do not supply the theoretical justification for freeing them. Disclose changes made after seeing the data and seek validation in new data.

Interpret paths, indirect effects and uncertainty

Structural paths and effects from the full synthetic model
Path or effectUnstandardized BSEApproximate 95% CI for BFully standardized
Support → confidence (a)0.5130.0460.423–0.6030.621
Confidence → engagement, conditional on support (b)0.6300.0680.496–0.7640.631
Support → engagement, conditional on confidence (cprime)0.1630.0500.066–0.2610.198
Support → confidence → engagement (a × b)0.3230.0430.240–0.4070.392
Support → engagement, total (cprime + a × b)0.4870.0460.396–0.5770.591

B denotes an unstandardized coefficient. SE and the intervals refer to that same unstandardized scale; the last column uses fully standardized latent-variable coefficients. The path intervals use normal-ML expected-information standard errors. The indirect and total intervals use the first-order delta method with the joint parameter covariance matrix. They are approximate 95% Wald intervals, not bootstrap intervals, and the standardized column has no intervals attached.

For a, a one-unit higher X is associated with 0.513 units higher M. For b, a one-unit higher M is associated with 0.630 units higher Y, conditional on X. The direct coefficient is the X–Y association conditional on M. These units depend on the chosen marker indicators; they are not automatically original questionnaire scale points.

Using the unrounded estimates, indirect = 0.512778 × 0.630433 = 0.323 and total = 0.163456 + 0.323272 = 0.487. The decomposition concerns this linear additive model. Do not multiply a standardized path by an unstandardized path, or add coefficients expressed on incompatible scales.

For the completely standardized solution, a one-standard-deviation higher X is associated with 0.621 standard deviations higher M. Standardization uses total factor variances, including explained variation in endogenous constructs. It is useful for describing scale, but it does not remove confounding or make coefficients directly comparable across different populations and models.

The model accounts for 38.6% of the variance in M and 59.4% in Y. These are model-based R-squared values, not the percentages of students helped or of engagement caused by support. With correlated predictors, do not obtain Y's R-squared by simply adding squared standardized paths.

The indirect estimate is positive and its approximate interval excludes zero. That supports a nonzero model-based indirect association under the stated assumptions. A significant total effect is not a prerequisite for investigating an indirect effect, and a nonsignificant direct path would not prove that mediation is complete.

Because products of coefficients can have asymmetric sampling distributions, a suitable bootstrap or a joint-parameter Monte Carlo interval may give better uncertainty estimates than a symmetric delta interval. For raw independent observations, a case-resampling bootstrap refits the entire measurement and structural model for each resample. Report the interval type, number of attempted and usable resamples, random seed and convergence failures. Clustered data require a design-appropriate resampling scheme. A covariance matrix alone cannot support a nonparametric case bootstrap; no such bootstrap was performed here. Parametric simulation would make additional distributional assumptions and must be labeled accordingly.

For the delta calculation, Var(ab) ≈ b² Var(a) + a² Var(b) + 2ab Cov(a,b). Ignoring the covariance between a and b can change the uncertainty estimate. Our independent calculation gives SE(ab) = 0.042609, with the approximate interval 0.240 to 0.407. An interval for the total effect must additionally incorporate covariance with the direct path; adding the endpoints of separate intervals is invalid.

Compare a theory-led alternative

The restricted model fixes cprime to zero while keeping the measurement structure unchanged. Compared with the full model, delta chi-square(1) = 10.415, p = 0.00125. Under conventional ML, that nested comparison rejects the zero-direct-path restriction at the usual significance level. The restricted model nevertheless retains favorable approximate indices. This is why a single fit cutoff cannot decide whether a particular theoretical restriction is defensible.

Do not subtract scaled robust chi-square statistics as though they were ordinary ML tests. Use the appropriate corrected difference test, with compatible estimators, observations, variable definitions and identification conventions. For nonnested alternatives, a nested chi-square difference test is not applicable; consider theory, comparable likelihood-based criteria where appropriate, prediction and the design's ability to discriminate explanations.

An additional limitation is central here: allowing all three structural paths among these three constructs makes the latent covariance structure unrestricted. The full SEM is covariance-equivalent to the corresponding correlated three-factor CFA. Its global degrees of freedom come from the measurement restrictions, not an additional test of the proposed causal ordering. Reversing arrows can produce equivalent covariance explanations. The restricted comparison tests an additional covariance restriction; it still does not establish temporal or causal mediation.

Reproduce the example in R

The code supplies the entire synthetic covariance matrix rather than relying on an unavailable dataset or a random generator. It follows the official lavaan covariance-input, SEM and mediation syntax. R is unavailable in this preparation environment: this code was reviewed but not executed. The displayed numbers were calculated with Python/NumPy and independently reproduced through a separately parameterized CFA fit, numerical derivatives and transformation back to structural coefficients. Minor last-digit differences can occur across software versions and optimizers.

library(lavaan)
# Synthetic teaching input; covariance divisor N, not N - 1.
lower <- '
  1.84
  1.224 1.5404
  1.08 0.918 1.41
  0.72 0.612 0.54 1.45
  0.768 0.5508 0.486 0.9 1.21
  0.576 0.4896 0.432 0.8 0.72 1.19
  0.72 0.612 0.54 0.744 0.6696 0.5952 1.5
  0.792 0.6732 0.484 0.8184 0.73656 0.65472 1.1 1.66
  0.684 0.5814 0.513 0.8018 0.63612 0.56544 0.95 1.045 1.5025
'
S <- lav_getcov(lower, names = c('x1','x2','x3',
                                'm1','m2','m3',
                                'y1','y2','y3'))
N <- 400L
measurement <- '
  X =~ 1*x1 + x2 + x3
  M =~ 1*m1 + m2 + m3
  Y =~ 1*y1 + y2 + y3
  M ~~ 0*Y
'
model <- paste0(measurement, '
  M ~ a*X
  Y ~ b*M + cprime*X
  indirect := a*b
  total := cprime + a*b
')
restricted <- paste0(measurement, '
  M ~ a*X
  Y ~ b*M + 0*X
')
fit <- sem(model, sample.cov = S, sample.nobs = N,
           sample.cov.rescale = FALSE, likelihood = 'normal',
           estimator = 'ML', information = 'expected',
           meanstructure = FALSE, fixed.x = FALSE)
fit0 <- sem(restricted, sample.cov = S, sample.nobs = N,
            sample.cov.rescale = FALSE, likelihood = 'normal',
            estimator = 'ML', information = 'expected',
            meanstructure = FALSE, fixed.x = FALSE)
summary(fit, fit.measures = TRUE, standardized = TRUE, rsquare = TRUE)
parameterEstimates(fit, ci = TRUE, standardized = TRUE)
fitMeasures(fit, c('chisq','df','pvalue','cfi','tli','rmsea',
                  'rmsea.ci.lower','rmsea.ci.upper','srmr'))
lavTestLRT(fit, fit0)
lavInspect(fit, 'converged')
lavInspect(fit, 'cov.lv')
resid(fit, type = 'cor')
modindices(fit, sort. = TRUE)
# Explicit expected-information delta intervals, using 1.96.
p <- coef(fit)
V <- vcov(fit)
g <- setNames(rep(0, length(p)), names(p))
g['a'] <- p['b']; g['b'] <- p['a']
indirect <- unname(p['a'] * p['b'])
se_indirect <- sqrt(drop(t(g) %*% V %*% g))
g['cprime'] <- 1
total <- unname(p['cprime'] + indirect)
se_total <- sqrt(drop(t(g) %*% V %*% g))
rbind(indirect = c(B = indirect, SE = se_indirect,
                  lower = indirect - 1.96*se_indirect,
                  upper = indirect + 1.96*se_indirect),
      total = c(B = total, SE = se_total,
                lower = total - 1.96*se_total,
                upper = total + 1.96*se_total))
packageVersion('lavaan')
sessionInfo()

The covariance rescaling argument is deliberate: the supplied matrix already uses divisor N. Passing an ordinary divisor-(N−1) covariance matrix while suppressing rescaling would change the likelihood convention. For real raw data, supply your dataset and a suitable missing-data and estimator specification instead of copying the synthetic covariance inputs. Do not add se = "bootstrap" to this covariance-only analysis and describe the resulting procedure as case resampling.

Retain the full loading, variance, covariance, residual and convergence output alongside the selected reporting tables. The short table above illustrates the structural question; it does not replace a complete measurement-model report.

Reliability, validity and competing explanations

Review item content and measurement evidence before trusting paths between constructs. Report loadings and uncertainty, explain a suitable model-based reliability coefficient for each intended composite, and examine convergent and discriminant validity alongside substantive evidence. The semTools reliability documentation makes the dependence on factor-model assumptions explicit. High reliability does not establish validity, and a strong association between factors may indicate inadequate distinction between them.

Measurement analysis and structural analysis should tell a coherent story. If you revise items after examining structural results, disclose the revision and distinguish exploration from confirmation. This worked example does not establish content validity, report a completed reliability study, or test alternative indicator assignments.

When comparing structural paths or latent means across groups or occasions, first establish the measurement comparability needed for that comparison. Our measurement-invariance guide addresses the relevant constraints. Different comparisons require different levels of comparability; a shared questionnaire title is insufficient.

Also consider omitted confounding, reverse direction, common measurement methods and selection into the sample. Cain, Zhang and Bergeman examine how timing, variable stability and model specification affect mediation analysis, including the limitations of cross-sectional models for processes unfolding over time. Measuring a proposed mediator and outcome at the same occasion does not establish that one preceded the other. Longitudinal measurement or experimental manipulation can strengthen a design, but neither automatically removes every causal assumption. State the assumptions needed for your target causal interpretation and separate them from what model fit tests.

Dissertation methodology and results paragraphs

The following paragraphs describe this teaching analysis. Replace every study-specific statement with what you actually planned, measured and analyzed. Do not represent these synthetic outputs as your own collected results.

Methodology example

A covariance-based latent-variable SEM was specified to examine the association between perceived support and academic engagement directly and indirectly through study confidence. Each construct had three continuous indicators. The first loading for each construct was fixed to one, and the remaining loadings, indicator residual variances, factor variance, structural disturbance variances and regression paths were estimated. Cross-loadings and residual covariances were fixed to zero. The teaching analysis used a synthetic divisor-N covariance matrix with an illustrative N = 400, conventional normal maximum likelihood, expected-information standard errors and no mean structure. Overall fit and covariance residuals were evaluated, and a prespecified alternative fixed the direct path to zero. Indirect and total associations were defined algebraically and assessed using approximate first-order delta-method intervals. This synthetic example contains no participant findings and includes no case-resampling bootstrap.

In a real dissertation, also document recruitment, inclusion and exclusion rules, sample-size planning, missingness handling, design corrections, the software and version actually used, diagnostic findings and changes from the analysis plan. Never claim that sample-size simulations, distribution checks or sensitivity analyses occurred merely because they would have been desirable.

Results example

The synthetic full model converged to an admissible solution: chi-square(24) = 32.122, p = 0.124, CFI = 0.996, TLI = 0.994, RMSEA = 0.029 (90% CI 0.000–0.053), and covariance SRMR = 0.016. The unstandardized direct association was B = 0.163 (approximate 95% CI 0.066–0.261). The indirect association through confidence was 0.323 (approximate delta-method 95% CI 0.240–0.407), and the total association was 0.487 (approximate delta-method 95% CI 0.396–0.577). Fixing the direct path to zero worsened fit, delta chi-square(1) = 10.415, p = 0.00125. These model-based associations do not establish a causal mediation process. The full structural parameterization is covariance-equivalent to a correlated three-factor measurement model.

In your discussion, explain the magnitude and uncertainty in relation to the research question. Consider whether different plausible models or measurement assumptions would change the interpretation. A statistically detectable path is not automatically educationally meaningful.

Common SEM reporting mistakes

Reporting problems and useful corrections
MistakeWhy it weakens the reportBetter practice
Reporting only fit indices and a diagramHides measurement quality and parameter uncertaintyReport identification, loadings, variances, path estimates, uncertainty and diagnostics
Treating cutoffs as a verdictA favorable summary can hide an important restriction or residualAssess global fit, local discrepancies and specific alternatives together
Calling every product causal mediationA fitted association does not supply temporal ordering or eliminate confoundingState the design, estimand and causal assumptions explicitly
Labeling a delta interval as bootstrapMisrepresents the procedure and its assumptionsName the actual uncertainty method and disclose what was executed
Mixing standardized and unstandardized resultsChanges scales within one calculationDecompose effects on a single compatible scale and label every column
Adding paths until fit improvesConverts confirmatory analysis into an undisclosed specification searchExplain each change and seek independent validation
Declaring nonsignificance proof of no effectIgnores precision and the range of compatible effectsInterpret the estimate and interval, including meaningful effects still compatible with the data

Frequently asked questions

Can I use SEM with a small dissertation sample?

Possibly, but there is no general exemption from identification, precision or power requirements. A simpler well-supported model can be more defensible than an elaborate poorly estimated one. Use model-specific planning and report the limits imposed by the actual sample.

Is CFA the same as full SEM?

CFA is a measurement-model application within the SEM family. Full latent-variable SEM additionally specifies structural regressions among constructs. Some structural parameterizations, including this full example, are covariance-equivalent to a correlated CFA, so the distinction in purpose does not guarantee different global fit.

Must every direct path be significant before testing mediation?

No. Specify the indirect effect of interest and estimate its uncertainty. The direct, indirect and total effects answer different questions. Opposite-signed pathways can also offset each other in a total effect, so a total-effect significance screen can miss a substantively relevant indirect association.

Should I report standardized or unstandardized coefficients?

Usually report unstandardized estimates with their uncertainty and add clearly labeled standardized estimates when they help interpretation. Explain the latent scaling convention and standardization used. Do not attach an unstandardized confidence interval to a standardized estimate.

Can good SEM fit prove my theory?

No. Fit assesses compatibility with restrictions in the specified model under estimation assumptions. Competing models can fit similarly or be covariance-equivalent. Design, measurement quality, external evidence and sensitivity analysis remain necessary.

Can I bootstrap an indirect effect from a covariance matrix?

A covariance matrix does not contain the individual cases needed for a nonparametric case-resampling bootstrap. Parametric simulation or a joint-parameter Monte Carlo method answers a related uncertainty question under additional assumptions. Identify the method accurately and do not present simulated resamples as observed participants.

Sources and further reading

Support for your own analysis

If you need help making the analysis plan, output and dissertation explanation consistent, explore Acadelyra's academic support services. Our role is to help you understand methodological choices, review reporting and improve clarity while preserving your authorship and responsibility. Your dissertation must describe analyses you can explain, inputs you are entitled to use, and work permitted by your institution.