Citation, validation, methods & ethics

APA (7th ed.)

BibTeX

RIS (EndNote / Mendeley / Zotero)

DOI (v1.0): 10.5281/zenodo.20640225 · Concept DOI (always latest version): 10.5281/zenodo.20640224 · archived on Zenodo (author ORCID 0000-0001-7682-2244). A peer-reviewed software paper is in preparation; once published it will become the preferred citation. Machine-readable metadata: CITATION.cff.

Every statistical routine in SmartSEM is validated against established open-source reference implementations before release. The full harness (reference generators in Python, comparison tests in Node.js, and the validation dataset) is downloadable below so any researcher can reproduce the audit.

Total: 251 / 251 checks passing on the v1.0 release build. The validation dataset is a simulated HRM survey (N = 350) with a known population model (TL → JS → OC; POS moderating TL → JS; true moderated-mediation index > 0), so recovery of the designed effects is itself part of the audit.

⬇ Download validation package (dataset + scripts + reference values)

• CB-SEM: maximum-likelihood estimation of the RAM-specified covariance structure (McArdle & McDonald, 1984), F_ML discrepancy (Jöreskog, 1969; Bollen, 1989) minimised by BFGS with analytic gradients; standard errors from the inverse observed information; fit indices χ², CFI, TLI (Bentler, 1990), RMSEA with 90% CI via the noncentral χ² distribution (Browne & Cudeck, 1992; Steiger, 1990), SRMR; thresholds per Hu & Bentler (1999).
• PLS-SEM: Lohmöller's (1989) algorithm — Mode A outer estimation, path inner weighting (Wold, 1982); reliability ρ_A (Dijkstra & Henseler, 2015), composite reliability ρ_c, AVE; discriminant validity via Fornell & Larcker (1981), HTMT (Henseler, Ringle & Sarstedt, 2015) and cross-loadings; SRMR for composite models; assessment guidelines per Hair, Hult, Ringle & Sarstedt (2022).
• Path analysis: OLS on standardized mean composites with PROCESS-equivalent equations (Hayes, 2018).
• Mediation: bootstrapped specific indirect effects (Preacher & Hayes, 2008), serial chains (Hayes model 6), pairwise contrasts, Zhao, Lynch & Chen (2010) typology.
• Moderation: two-stage standardized interactions; simple slopes (Aiken & West, 1991); Johnson–Neyman significance regions (Johnson & Neyman, 1936; Bauer & Curran, 2005).
• Conditional (moderated) mediation: Hayes models 7 and 14; index of moderated mediation (Hayes, 2015) with bootstrap CIs.
• Bootstrapping: case resampling, percentile and bias-corrected CIs (Efron, 1987), fixed seeds for reproducibility.
• Endogeneity: two-stage least squares; Durbin–Wu–Hausman augmented-regression test; weak-instrument F (Stock & Yogo, 2005); Sargan (1958) over-identification.
• Common method bias: Harman single-factor; unmeasured latent method construct (Podsakoff, MacKenzie, Lee & Podsakoff, 2003) with equality-constrained loadings and multi-start; full-collinearity VIF (Kock, 2015).
• EFA: principal-axis factoring with SMC starts, varimax and promax rotation, Kaiser (1974) KMO, Bartlett's sphericity, Horn's (1965) parallel analysis.

• Every report is stamped with the SmartSEM version and analysis date; cite the version you used.
• Bootstrap and permutation routines use fixed seeds — re-running the same data, model, and settings reproduces identical confidence intervals.
• The complete computation runs client-side in standard JavaScript shipped to your browser — the implementation is inspectable by anyone (view source: engine.js).
• The validation package above lets third parties re-derive our reference values from public Python libraries.

• Your data never leaves your device. All statistics are computed in your browser; nothing is uploaded or stored. This supports GDPR/UAE PDPL compliance and most institutional ethics (IRB) requirements for secondary analysis. Researchers remain responsible for the consent and ethics approvals governing their own data collection.
• AI features are optional and transparent. Hypothesis parsing, interpretations, theory matching, and the Results-section draft call a server-side AI (Groq, Llama 3.3). These requests contain only your hypothesis text and aggregate statistics — never raw, row-level data.
• AI-assistance disclosure. AI-generated text in SmartSEM is drafted from your computed results, but large language models can err. Verify every number against the tables, edit the prose, and disclose AI assistance where your target journal's policy requires it (e.g., COPE guidance and major publishers' AI policies). Scientific responsibility for the manuscript rests with the authors — SmartSEM cannot be a co-author.
• Statistical responsibility. SmartSEM automates computation, not judgment. Model specification, assumption checking, and interpretation remain the researcher's responsibility; for high-stakes results we encourage cross-checking in a second package (the validation page documents the agreement you should expect).
• No tracking of analysis content. Site analytics record page usage only (e.g., that an analysis was run), never variables, models, or results.

• CB-SEM estimator is normal-theory ML; robust (MLR) and categorical (WLSMV) estimators are on the roadmap — under non-normality use the bootstrap CIs (default) rather than normal-theory p-values.
• Missing data are handled by listwise deletion per analysis (FIML/EM imputation planned).
• Moderation and conditional mediation use the two-stage score approach (factor scores / LV scores / composites) rather than latent product-indicator estimation.
• PLS measurement is reflective (Mode A); formative (Mode B) assessment is planned.
• Multigroup comparison uses pairwise z-tests; permutation MGA and MICOM invariance testing are planned.
• No modification indices; respecification should be theory-led.
• Very large datasets (>~50,000 rows × many items) may be limited by browser memory.

v1.0 — 10 June 2026. First public beta. Three estimation methods (CB-SEM ML, PLS-SEM, PROCESS-style path analysis); hypothesis-driven model building with AI parsing; data screening; EFA; reliability & validity suite; bootstrap (percentile/BC); simple, serial, parallel and moderated mediation; moderation with simple slopes, Johnson–Neyman and 3D surfaces; 2SLS endogeneity checks; common-method-bias battery; multigroup; total effects; SmartPLS-style diagram with HD export; TheoryLens integration (best-fit theories overall and per relationship); APA tables and .doc reporting; 251-check validation suite.