Assess the structural model and test hypotheses

With a sound measurement model, you can turn to the hypotheses. The structural model assessment has a fixed sequence (Hair et al., 2022): collinearity, significance and relevance of the paths, explanatory power, predictive power, and, if needed, model comparison.

Step 1: collinearity

Path coefficients are estimated with regressions, so strongly correlated predictor constructs can distort them. Inner VIF values should be below 3 (5 at most). In the sample study the highest is 1.167, so collinearity is not an issue.

Step 2: path coefficients and hypotheses

Path coefficients are standardised, between −1 and +1. A path is significant when its bootstrap confidence interval excludes zero (equivalently, p < .05 and t > 1.96 for a two-tailed test). All six hypotheses are supported. Performance expectancy is the strongest driver of intention (β = 0.351), and intention strongly drives use (β = 0.440).

Significance is not relevance

H6 is significant, but look closer: β = 0.107, the CI almost touches zero (0.006), and f² = 0.013, below the 0.02 threshold for a small effect. Effort expectancy has a statistically detectable but practically marginal direct effect on use. Report this nuance; reviewers appreciate it.

The f² effect size measures how much R² drops when a predictor is removed: 0.02 small, 0.15 medium, 0.35 large. PE (0.182) and BI (0.219) have medium effects; EE, SI and TR on BI have small effects.

Step 3: explanatory power (R²)

R² = 0.413 for BI and 0.241 for UB. The often-quoted guidelines (0.75 substantial, 0.50 moderate, 0.25 weak) come from marketing; in consumer behaviour research, R² around 0.20 is already considered high, and in technology acceptance, intention models typically explain 40 to 60%. Interpret R² against comparable studies in your field, and report adjusted R² (0.403 and 0.236), which corrects for the number of predictors.

Step 4: predictive power

R² is in-sample explanatory power. To show that a model predicts, use out-of-sample methods:

• Blindfolding Q²: values above zero indicate predictive relevance; 0.25 and 0.50 mark medium and large. Here Q² = 0.332 for BI and 0.179 for UB.
• PLSpredict (Shmueli et al., 2019), the current standard: the model is estimated on part of the data and predicts the held-out indicators. Compare its prediction errors (RMSE or MAE) with a naive linear regression benchmark (LM). If PLS has lower errors for all or most indicators, predictive power is high or medium. The Studio runs PLSpredict on the Prediction tab, but not for models with interaction terms; to run it here, remove the PE × TR moderation and run the main-effects model.

Step 5: model fit

The SRMR (standardised root mean square residual) is 0.046 for the saturated model and 0.047 for the estimated model, below the 0.08 guideline. Be careful: model fit in PLS-SEM is debated, because PLS does not minimise a fit function. Report SRMR if your journal expects it, but do not base your evaluation on it.

Reporting

Inner VIF values were below 1.17, so collinearity was not a concern. Bootstrapping with 5,000 subsamples supported all six hypotheses (Table 4). Performance expectancy (β = 0.351, p < .001, f² = 0.182), effort expectancy (β = 0.224, p < .001), social influence (β = 0.185, p < .001) and trust (β = 0.181, p < .001) positively affected behavioural intention, which in turn affected use behaviour (β = 0.440, p < .001, f² = 0.219). The direct effect of effort expectancy on use was significant but small (β = 0.107, p = .043, f² = 0.013). The model explained 41.3% of the variance in intention and 24.1% in use; Q² values of 0.332 and 0.179 indicated predictive relevance.