Run the PLS algorithm and bootstrapping

With the model specified, the next step is estimation. PLS-SEM does not need distributional assumptions to estimate the model, but it also has no formula for standard errors. That is why every PLS-SEM analysis has two parts: the PLS algorithm, which estimates the coefficients, and bootstrapping, which tests whether they are significant.

How the PLS algorithm works

The algorithm alternates between the measurement and the structural model until the estimates stop changing:

1. Start with equal weights and compute each construct score as a weighted sum of its standardised items.
2. Inner approximation: compute a proxy for each construct from the neighbouring constructs. With the path weighting scheme (the default and recommended choice), the proxies use regression coefficients for predecessors and correlations for successors.
3. Outer approximation: update each indicator’s weight. For reflective (mode A) constructs, the weight is the correlation between the indicator and the inner proxy.
4. Repeat until the weights change by less than the stop criterion (10⁻⁷), then estimate loadings and path coefficients with the final construct scores.

The sample model converges after 5 iterations. If a model needs close to the maximum of 300 iterations, something is wrong, such as an item with almost no variance or a construct with indicators that do not belong together.

Bootstrapping

Bootstrapping draws many random samples of 300 respondents with replacement from the original 300, re-estimates the whole model in each subsample, and uses the spread of the estimates as the standard error. From it come t-values, p-values and confidence intervals. Recommendations:

• 5,000 subsamples for final results (500 for a quick look while you build the model).
• Percentile confidence intervals, the current default recommendation (Aguirre-Urreta and Rönkkö, 2018). Bias-corrected intervals are an option when the bootstrap distribution is skewed.
• Two-tailed tests unless you have directional hypotheses and your field accepts one-tailed tests; state your choice.
• A fixed random seed makes the bootstrap results reproducible. With a different seed, the second or third decimal of the standard errors may change, but not the conclusions.

The first results

The run takes about 15 seconds and reports R² = 0.413 for behavioural intention, R² = 0.241 for use behaviour and an SRMR of 0.046. The four drivers explain 41% of the variance in intention, and intention plus effort expectancy explain 24% of the variance in use. These are typical values for technology acceptance research, and you will interpret them properly in Lesson 6.

Weighting schemes and other settings

The Advanced settings let you choose the factorial or centroid weighting scheme. Results rarely differ much, but the path weighting scheme gives the highest R² and is the standard; change it only with a reason. The prediction settings control blindfolding (omission distance 7, meaning every seventh data point is omitted in turn; choose a distance that does not divide the sample size) and PLSpredict (10 folds and 10 repetitions are typical).

PLS-SEM or PLSc?

Standard PLS-SEM estimates constructs as composites. When your reflective constructs are meant as common factors and you want estimates comparable with CB-SEM, consistent PLS (PLSc) corrects the loadings and paths for measurement error. For most applied studies with composite-based reasoning, standard PLS-SEM is appropriate.

What to report about estimation

Readers need enough detail to reproduce your analysis. The Export tab writes a methods paragraph for you, which you adapt:

The research model was estimated with PLS-SEM in PLS-SEM Studio (Xavier, 2026), using the path weighting scheme (maximum 300 iterations, stop criterion 10⁻⁷). Significance was tested with bootstrapping using 5,000 subsamples (two-tailed, α = .05) and 95% percentile confidence intervals.