A practical guide for journal articles and theses: run partial least squares structural equation modeling on StatFas, check each quality criterion in order, and write the results in APA 7 style.
Open the PLS-SEM toolThe example below uses the built-in data on the tool (constructs SUP, EFF and WEL, N = 300). All numbers in the write-up templates are illustrative. Replace them with your own output.
Use a CSV, TXT or TSV file with a header row. Name each questionnaire item as a construct prefix plus a running number (sup1, sup2, sup3). Use numeric codes only. Rows with a missing value in any model indicator are dropped, so check how many cases that costs you. Click Load example data to practise first.
The tool detects constructs from the column names and writes the measurement lines for you:
SUP =~ SUP1 + SUP2 + SUP3 + SUP4 EFF =~ EFF1 + EFF2 + EFF3 + EFF4 WEL =~ WEL1 + WEL2 + WEL3 + WEL4
=~ is reflective (Mode A). Change it to <~ for a formative construct (Mode B). Decide this from theory: indicators that are interchangeable manifestations of the construct are reflective, and indicators that jointly define it are formative.
Write one line per outcome, as outcome ~ predictors, from your hypotheses:
EFF ~ SUP WEL ~ SUP + EFF
This model has a direct path from SUP to WEL and an indirect path through EFF.
Keep path weighting. Tick PLSc when your constructs are reflective. Set Bootstrap samples to 5000 for the results you report (1000 is fine for exploring). Leave Multimethod comparison and CVPAT ticked.
Read the output from top to bottom. The sections below follow the same order, because measurement quality must be established before you interpret the structural paths.
Reflective constructs are judged on four criteria. Formative constructs are judged on weights, collinearity and significance instead.
| Table in the tool | Criterion | Common guideline |
|---|---|---|
| Outer loadings | Indicator reliability | ≥ .708; consider deleting items below .40, and those between .40 and .70 only if composite reliability or AVE improves |
| Reliability and convergent validity | Composite reliability, ρA | .70 to .95 (above .95 suggests redundant items) |
| Reliability and convergent validity | AVE | ≥ .50 |
| HTMT, Fornell–Larcker, cross-loadings | Discriminant validity | HTMT below .85 (conservative) or .90, ideally with a confidence interval that excludes 1 |
| Weights and VIF (formative) | Collinearity, relevance | VIF below 3 (5 at most); keep nonsignificant indicators if their loading is high |
A well-fitting model does not necessarily predict well, so assess both. PLSpredict (10-fold): Q²predict above 0 means the model beats a naive mean prediction, and indicator errors lower than the linear-model benchmark show good predictive power. CVPAT tests whether the model predicts better than an indicator-average benchmark. A significantly negative loss difference favours your model. The path-level table shows which relationships contribute to prediction.
Also read the Multimethod comparison. If PLS, PLSc, equal weights and sum scores agree on a path, it is unlikely to be an artifact of one estimator. If they diverge, check measurement quality and specification, and report the divergence openly.
APA 7 rules that apply here: italicize statistical symbols (N, p, R², f²); no leading zero for values that cannot exceed 1 (p = .021, R² = .31, β = .42); report exact p to three decimals, and p < .001 below that; two decimals for coefficients is standard, three is also common in PLS-SEM; table number in bold on one line, title in italic title case on the next; a "Note." line below each table.
The hypothesized model was tested with partial least squares structural equation modeling (PLS-SEM) using StatFas (2026). We used the path weighting scheme with consistent PLS (PLSc) for the reflective constructs. Significance and confidence intervals were obtained with 5,000 bootstrap resamples (percentile method). The model was estimated on N = [300] complete cases. We chose PLS-SEM because the study aims to [explain and predict key target constructs / includes formative constructs / tests a complex model] (Hair et al., 2011).
All reflective indicator loadings exceeded .70 (range = .74–.89; see Table 1). Composite reliability (.86–.91) and ρA (.85–.92) were within the recommended .70–.95 range, and average variance extracted exceeded .50 for every construct (.61–.72), supporting reliability and convergent validity. All HTMT values were below .85 (maximum = .62), indicating discriminant validity (Henseler et al., 2015).
Table 1
Reliability and Convergent Validity of the Measurement Model
| Construct | Items | Loadings | CR | ρA | AVE |
|---|---|---|---|---|---|
| SUP | 4 | .74–.86 | .88 | .87 | .65 |
| EFF | 4 | .76–.89 | .91 | .92 | .72 |
| WEL | 4 | .75–.83 | .86 | .85 | .61 |
Note. CR = composite reliability; AVE = average variance extracted. Values are illustrative.
Collinearity was not a concern (all inner VIF < 1.5). Support was positively related to efficacy (β = .52, 95% CI [.42, .61], p < .001, f² = .37) and to well-being (β = .21, 95% CI [.08, .34], p = .002, f² = .04). Efficacy was also related to well-being (β = .44, 95% CI [.31, .56], p < .001, f² = .16). The model explained 27% of the variance in efficacy and 35% in well-being, both weak to moderate. The indirect effect of support on well-being through efficacy was significant (β = .23, 95% CI [.16, .31], p < .001), and because the direct effect remained significant, efficacy partially mediated this relationship (SRMR = .05).
PLSpredict showed positive Q²predict values for all indicators of well-being (.08–.19), and the PLS-SEM prediction errors were lower than the linear-model benchmark for all indicators, indicating high predictive power. CVPAT indicated that the model predicted significantly better than the indicator-average benchmark (average loss difference = −0.21, p < .001). Path estimates were consistent across PLS, PLSc, equal weights and sum scores (Table 2).
Download the path diagram (SVG or PNG) from the tool and cite it as Figure 1, with a title in italic title case (for example, Structural Model With Standardized Estimates) and a note explaining the symbols and significance stars. Mention in your limitations that cross-sectional data limit causal claims and that results should be replicated.
The Download APA 7 report (Word) button writes the narrative, tables and figure for you. Edit the method statement so it fits your study.
Use 5,000 for the results you publish. Fewer is fine while you explore the model.
Tick it when the constructs are reflective and you want estimates that mimic a common-factor model. Standard PLS treats constructs as composites and tends to overestimate loadings.
Not automatically. Delete it only if removal raises composite reliability or AVE above the threshold, and keep content validity in mind.
Report SRMR, but justify the model mainly by explanation (R², f²) and prediction (PLSpredict, CVPAT). For theory confirmation and global fit, CB-SEM is usually preferred.
Hair, J. F., Ringle, C. M., & Sarstedt, M. (2011). PLS-SEM: Indeed a silver bullet. Journal of Marketing Theory and Practice, 19(2), 139–152.
Henseler, J., Ringle, C. M., & Sarstedt, M. (2015). A new criterion for assessing discriminant validity in variance-based structural equation modeling. Journal of the Academy of Marketing Science, 43(1), 115–135.
Jöreskog, K. G., & Wold, H. (1982). The ML and PLS techniques for modeling with latent variables. In K. G. Jöreskog & H. Wold (Eds.), Systems under indirect observation (Part 1, pp. 263–270). North-Holland.
Sharma, P. N., Liengaard, B. D., Hair, J. F., Sarstedt, M., & Ringle, C. M. (2023). Predictive model assessment and selection in composite-based modeling using PLS-SEM. European Journal of Marketing, 57(6), 1662–1677.
PLS-SEM tool · SEM tutorial · CFA tutorial · Mediation tutorial · Sample size and power