When to use PLS-SEM
PLS-SEM (partial least squares structural equation modelling) tests relationships between latent constructs measured by multiple indicators. It is often chosen when the aim is prediction or theory development, when the model is complex, when the sample is relatively small or when the data is not normally distributed.
Covariance-based SEM (for example in AMOS or lavaan) is more common for confirming established theories and testing overall model fit. Choose based on your objective and justify it with method sources, and check your supervisor's preference.
Preparing data
Clean your data first: check missing values, outliers, straight-lining responses and coding errors, then save the file as CSV for import into SmartPLS. Check that your sample size is adequate; a common rule of thumb is the "10-times rule", but power analysis is more reliable.
Make sure indicator names are clear and consistent, and that reverse-coded items are recoded. Draw the model in SmartPLS, linking each construct to its indicators and each hypothesis to a path between constructs.
Measurement model: reliability and validity
Assess the measurement model before testing hypotheses. Common checks, following guidance such as Hair et al.:
- Indicator reliability: outer loadings of about 0.708 or higher.
- Internal consistency: composite reliability (and Cronbach's alpha), usually between 0.70 and 0.95.
- Convergent validity: average variance extracted (AVE) of at least 0.50.
- Discriminant validity: HTMT below about 0.85 to 0.90, or the Fornell-Larcker criterion.
Remove or revise weak indicators carefully, and explain any changes. Do not delete items just to improve numbers.
Structural model and bootstrapping
Once the measurement model is acceptable, check for collinearity (VIF values below about 3 are preferable, and below 5 is often accepted). Then run bootstrapping, commonly with 5,000 subsamples, to test whether path coefficients are statistically significant.
Report path coefficients, t-values and p-values or confidence intervals, R-squared for explanatory power, effect sizes (f-squared) and predictive relevance (Q-squared or PLSpredict). Interpret each hypothesis in terms of direction, size and significance.
Reporting results
Describe your sample, then report the measurement model results in a table (loadings, alpha, composite reliability, AVE) and discriminant validity results. Then report the structural model with a path diagram and a table of hypotheses with the decisions: supported or not supported.
Explain the findings in relation to theory and prior studies, and note limitations such as sample size, sampling method and cross-sectional design. Cite the sources of the thresholds you use.
Common mistakes
Mistakes to avoid:
- Using PLS-SEM without justifying why.
- Skipping the measurement model checks.
- Reporting only significant results.
- Deleting indicators without explanation.
- Ignoring common method bias and collinearity.
- Over-interpreting small effects.
- Mixing up reflective and formative constructs.
If you are unsure, ask your supervisor and check method sources before interpreting results.