Free PLS-SEM online: measurement model, structural model, bootstrapping and HTMT

How to use the tool

1. Prepare the file. One row per respondent, one column per indicator, with column names made of a construct code plus a number (PU1, PU2, SAT1…). Declare reverse-worded items in their own box so the tool reverse-scores them.
2. Measurement model. The tool groups indicators by column name; delete any items you have already dropped. Every construct is reflective (Mode A) by default. Only list a construct as formative if the original scale was designed that way.
3. Structural model. One path per line, X -> Y, in the order of your hypotheses H1, H2… If you drew the model in the Research Model Drawer, click “Import from Model Drawer” to bring the arrows across (construct codes in the diagram must match the data columns).
4. Bootstrapping. Keep 5,000 subsamples and note the seed: the same data and the same seed always give the same results. Bootstrapping runs in the background, so a model with 300 responses and 18 indicators takes a few seconds on an ordinary laptop.
5. Export. Download a Word file (tables, path diagram, results text and references), the tables as CSV, or the construct scores for further analysis elsewhere.

What the tool computes

• The PLS algorithm (Lohmöller, 1989) with SmartPLS defaults: path weighting scheme, standardised data, initial weights of 1, stop when the total change in weights is below 10⁻⁷ or after 300 iterations. Missing values are replaced by the indicator mean (the SmartPLS default) or handled by listwise deletion.
• Measurement model: outer loadings and weights, indicator reliability (λ²), Cronbach's alpha, rho_A (Dijkstra & Henseler, 2015), composite reliability rho_c, AVE, the Fornell–Larcker criterion, HTMT, cross-loadings and outer VIF; for formative constructs, weights, their significance and VIF.
• Structural model: path coefficients, R², adjusted R², f², inner VIF, specific indirect effects for each chain, total indirect and total effects, and Q² by blindfolding (omission distance D = 7).
• PLSpredict (Shmueli et al., 2016, 2019): 10-fold cross-validation repeated 10 times, Q²predict, and RMSE and MAE of PLS against a linear regression benchmark (LM) for every endogenous indicator.
• SRMR for the saturated and estimated models, threshold < .08 (Henseler et al., 2014; Hu & Bentler, 1999), and an optional HTMT2 (Roemer et al., 2021) in place of HTMT.
• Bootstrapping (5,000 subsamples by default, fixed seed): standard errors, t, two-tailed p and percentile or BCa confidence intervals for paths, loadings, weights, R², indirect and total effects and HTMT.

Worked example: the ECSI model on the mobi data

The mobi data set (250 mobile phone customers, 1–10 scale; Tenenhaus et al., 2005) has six constructs (Image, Expectation, Perceived Value, Satisfaction, Complaints with one item, and Loyalty) and nine paths. With 5,000 bootstrap subsamples, and point estimates identical to seminr:

• The measurement model has problems in places: CUSL2 has a loading of .202; Expectation has α = .452 and AVE = .481; the HTMT between Satisfaction and Loyalty is .957.
• Structural model: Satisfaction → Loyalty β = 0.488 (95% CI [0.316, 0.639]); Complaints → Loyalty β = 0.066 (95% CI [−0.054, 0.181]), not significant. R² is .590 for Satisfaction and .458 for Loyalty.

The example shows why the measurement model comes first: in a real thesis you would deal with CUSL2 and the high HTMT before interpreting any path coefficient. The full data and the number-by-number comparison with seminr are on the verification page.

How to read the output

1. Outer loadings of .708 or above. Indicators between .40 and .708 are removed only if removal brings composite reliability or AVE above the threshold; below .40, remove them. Re-run the model after any removal.
2. Reliability (alpha, rho_A, rho_c) of .70 or above but not above .95; AVE of .50 or above.
3. Discriminant validity: HTMT below .85 (.90 for conceptually similar constructs) is the main criterion (Henseler et al., 2015); the Fornell–Larcker criterion and cross-loadings are supplementary. The tool also prints the upper bound of the bootstrap confidence interval for each HTMT value.
4. Structural model: inner VIF below 3, then path coefficients with p values (or confidence intervals that exclude zero), R², f² and Q². Indirect effects are judged by their bootstrap confidence intervals and classified following Zhao et al. (2010).
5. PLSpredict: every Q²predict must exceed zero; then count the indicators for which PLS has a lower RMSE than LM (use MAE if the prediction errors are highly skewed): all means high, most means medium, a minority means low and none means no predictive power. SRMR below .08 is reported only as a supplement.

How to report PLS-SEM results in APA 7

The tool writes the results section in the usual order: loadings, reliability and convergent validity, HTMT, path coefficients, R², f², Q². A typical sentence reads: “Satisfaction had a significant positive effect on Loyalty (β = 0.488, 95% CI [0.316, 0.639]).” Before pasting it into your thesis, check three things: your method chapter states the number of bootstrap subsamples, one- or two-tailed testing and the handling of missing values, matching this run; any threshold that is not met is reported and discussed, not removed from the table; and non-significant hypotheses are described as “not supported”. Cite the tool as: Research Corner. (2026). Online PLS-SEM tool [Computer software]. https://gocnghiencuu.com/en/tools/pls-sem/

How it compares with SmartPLS

Point estimates (loadings, weights, paths, R², alpha, rho_A, rho_c, AVE, HTMT, VIF) follow the standard PLS algorithm, so they can differ from SmartPLS only by rounding. Bootstrap results differ slightly between all programs, and even between two SmartPLS runs with different seeds, because resamples are drawn at random; standard errors usually differ only in the third decimal. PLSpredict also splits folds at random, so RMSE, MAE and Q²predict differ from SmartPLS by a few thousandths. The tool does not yet offer PLSc, higher-order constructs, interaction-term moderation, multi-group analysis (MGA) or MICOM. If your department requires SmartPLS output, use this free SmartPLS alternative to check your model first, then run SmartPLS with the same specification.

Frequently asked questions

Do I need to install SmartPLS?

Not to estimate the model: the algorithm runs in your browser with no licence, and your data never leaves your computer. If your study needs PLSc, higher-order constructs, MGA or MICOM, you will still need SmartPLS or the R packages seminr or cSEM.

What sample size does PLS-SEM need?

The “10 times the number of arrows” rule has been criticised; use the inverse square root method (Kock & Hadaya, 2018) or G*Power in the sample size calculator. The tool flags samples below 155 responses, the inverse square root minimum for a smallest expected path coefficient of .20.

Is this an HTMT calculator as well?

Yes. HTMT (or HTMT2) is computed for every pair of reflective constructs with at least two indicators, with bootstrap upper confidence bounds, and judged against the .85 and .90 cut-offs of Henseler et al. (2015).

How is PLSpredict different from blindfolding Q²?

Blindfolding re-predicts omitted data points but still estimates the model on almost the whole sample, so it is not true out-of-sample prediction. PLSpredict estimates the model on 9 of 10 folds, predicts the remaining fold, repeats this 10 times and compares the errors with a simple linear regression. Hair et al. (2019) recommend reporting PLSpredict when the study has a predictive aim.

Why do my bootstrap results differ from a classmate’s?

Because of a different seed or number of subsamples. With the same data, model, number of subsamples and seed, the tool gives identical results to the last digit, so report the seed in your method chapter.