A Transparent and Verifiable Workflow for Reporting PLS-SEM

1. Six evidence layers in an assessable report

An assessable PLS-SEM report explains decisions, their justification and how results change under different analytical conditions. A long list of indices cannot compensate for ambiguous specification or undisclosed settings.

Hair et al. (2019) recommend systematic reporting from method justification through measurement and structural assessment. Benitez et al. (2020) emphasize construct type, specification and confirmatory or explanatory objectives. Current practice also requires out-of-sample assessment with PLSpredict and CVPAT (Shmueli et al., 2019; Liengaard et al., 2021).

These six layers organize reporting; they are not six independent acceptance criteria. Favorable prediction does not fix invalid measurement, and acceptable measurement does not establish structural causality.

2. Document methods before inspecting result tables

Write methods before selecting favorable results. Preserve a record of data, model, software and settings versions to prevent measurement, bootstrap and prediction tables from describing different models.

Justify PLS-SEM through objectives and model characteristics. Small or skewed samples do not automatically require it. Justify sample size with power or an appropriate inverse-square-root approach rather than the 10-times rule alone (Kock & Hadaya, 2018).

Causal language must match the design. Cross-sectional convenience samples generally support conditional-association interpretations. Endogeneity checks identify certain biases but do not turn observation into experimentation (Hult et al., 2018).

• Prespecify explanatory, predictive or combined objectives.
• Justify construct types and measurement directions theoretically.
• Justify sample size using power or an appropriate method.
• Log every data and model change.
• Preserve software versions and settings in text or screenshots.

3. Report measurement according to construct type

For reflective models, report outer loadings and CIs, internal consistency reliability, AVE and discriminant validity. HTMT detects a lack of discriminant validity better than Fornell-Larcker under many simulated conditions (Henseler et al., 2015). Cutoffs are reference points; retention decisions also require construct content and uncertainty.

For composites or formative models, prioritize content coverage, indicator collinearity, outer weights and CIs. Loadings can inform absolute contributions when weights are nonsignificant. Do not remove an indicator solely for a small weight if this narrows the conceptual domain (Benitez et al., 2020).

Draw a separate conclusion for each construct. A blanket statement that measurement is acceptable conceals problematic constructs and the evidence used.

4. Read structural results in a controlled sequence

Inspect inner VIF before interpreting paths. Then report β, bootstrap CIs, p-values if included in the inferential plan, f², and R² or adjusted R² for every endogenous variable. Coefficients describe conditional direction and magnitude; f² describes explanatory changes when a predictor is removed; R² describes in-sample explanation.

Mediation, moderation, multigroup and higher-order analyses need their own procedures. Report indirect effects and CIs rather than inferring mediation from two component p-values. For moderation, explain interaction construction, any standardization, and simple slopes or predictions at meaningful levels.

Mark changes made after inspecting results as post hoc. Explain added paths, removed indicators or altered controls and assess sensitivity against the original specification.

5. Make bootstrapping reproducible

Bootstrap results depend on settings and random seed. Stating that bootstrapping was performed is insufficient to reproduce CIs. Report subsamples, seed, sign-change settings, interval type and test direction.

SmartPLS defaults to percentile CIs and offers bias-corrected and accelerated CIs. Keep the choice consistent with the analysis plan rather than changing it after seeing results (SmartPLS, n.d.-a).

A one-sided test requires a theoretically justified directional hypothesis stated beforehand. Two-sided testing is more transparent for nondirectional hypotheses or effects that may go either way.

6. Assess prediction and robustness

PLSpredict uses k-fold cross-validation and compares PLS-SEM errors against benchmarks. SmartPLS defaults to 10 folds, but each fold must contain enough observations for estimation. Report folds, repetitions, seed, error measure and benchmark (SmartPLS, n.d.-b; Shmueli et al., 2019).

CVPAT tests cross-validated loss differences between PLS-SEM and a benchmark or two models. Inspect the difference’s sign and magnitude before its p-value. A significant result favoring higher loss does not support the model (Liengaard et al., 2021).

Robustness checks should address plausible threats: alternative specifications, controls, endogeneity, heterogeneity, missing-data treatment or error metrics. State what changed and whether the main conclusion changed (Hult et al., 2018).

7. Package tables, figures and files for verification

Give each table a purpose. Do not mix reflective measurement results with formative weights. Structural tables should align β, CIs, f² and hypothesis decisions in a consistent order. Prediction tables must identify the indicator, loss, benchmark and cross-validation settings.

A minimal reproducibility package includes analysis data or substitute metadata, a codebook, model files, settings records, original output and a decision log. If confidentiality restricts data, preserve a synthetic-data generation procedure or enough descriptive material to inspect the analytical logic.

Use versioned filenames and do not overwrite originals. Include a table in a manuscript only when it can be traced to the correct data, model and run.

• Use consistent rounding in tables and text.
• Keep construct and indicator names consistent across diagrams, output and manuscript.
• Give figures captions, alt text and software versions where relevant.
• Preserve original data and version the analysis files.
• Include a codebook and cleaning log.
• Include model files and algorithm-setting screenshots.
• Preserve original Algorithm, bootstrap and PLSpredict/CVPAT output.
• Clearly identify post hoc changes.

8. APA 7 reporting template and submission checklist

Use this as a fill-in framework. Retain only procedures actually performed and replace every bracket with study information.

• Do not justify the method solely by small samples or skewed data.
• Do not justify sample size solely by the 10-times rule.
• Describe sampling, missingness, outliers and all exclusions.
• Distinguish reflective, formative and higher-order constructs.
• Report the SmartPLS version, Algorithm settings and bootstrap settings.
• Report β, CIs, f², R² and adjusted R² for each endogenous variable.
• Assess out-of-sample prediction using PLSpredict and CVPAT when appropriate.
• Identify benchmarks, folds, repetitions, seeds and loss functions.
• Disclose post hoc analyses and unfavorable findings.
• Keep causal language within the design’s limits.
• Match in-text citations and references in both directions.
• Make manuscript tables traceable to original runs.