What is PLS-SEM, and when should I use it?
PLS-SEM estimates relationships between constructs that are each measured by several indicators, such as questionnaire items. It suits prediction-oriented research, complex models with many constructs and paths, and models with formative constructs. Covariance-based SEM, run in software such as AMOS, is better suited to confirming an established theory using model fit. Choose the method that matches your research aim.
How do I assess the measurement model?
For reflective constructs, assess indicator reliability, internal consistency, convergent validity and discriminant validity. These are common guidelines; interpret them in light of the construct, research design and discipline.
- Indicator loadings should be 0.708 or above.
- Internal consistency reliability, measured by Cronbach's alpha and composite reliability, should be 0.70 or above.
- Convergent validity, measured by average variance extracted (AVE), should be 0.50 or above.
- Discriminant validity, measured by the HTMT ratio, should be below 0.85, or below 0.90 for conceptually similar constructs (Henseler, Ringle and Sarstedt, 2015).
How do I assess the structural model?
First check collinearity: variance inflation factors (VIF) of 3 or below are ideal, and above 5 signal a problem. Then test each path's significance with bootstrapping. Report R² for the variance each outcome construct explains, and f² for each predictor's effect size. Finally, assess out-of-sample predictive power with PLSpredict (Shmueli et al., 2019), where a Q²predict above zero indicates predictive relevance.
What if an indicator's loading is below 0.708?
Don't delete it automatically. Consider removing an indicator with a loading between 0.40 and 0.708 only if doing so raises the construct's composite reliability or AVE above its threshold, and its content isn't essential to the construct. Indicators below 0.40 are usually removed. Report every deletion and your reason for it.
Can I test mediation and moderation in SmartPLS?
Yes. For mediation, test the specific indirect effect with bootstrapped confidence intervals: if the interval excludes zero, the indirect effect is significant. For moderation, add an interaction term to the model, then interpret it with a simple slope plot. Explain both in terms of your hypotheses, not just the output.
How large a sample do I need for PLS-SEM?
There is no single sample-size number that fits every PLS-SEM model. The inverse square root method (Kock and Hadaya, 2018) uses the smallest path coefficient you expect to detect; for a 5% significance level and 80% power, its estimate is the next whole number above the square of 2.486 divided by that coefficient's absolute value. Use a conservative, theory- or pilot-based estimate for the smallest meaningful path, and justify the assumptions. Treat this as a planning estimate and consider a power analysis or simulation for complex models. The 10-times rule alone can give inaccurate estimates.
N = ceil((2.486 / |βmin|)²)
For example, if the smallest path you need to detect is 0.20, the estimate is 155 participants. This is a heuristic planning method, not a guarantee of adequate power for every model.