PLS-SEM has no distributional assumptions to lean on, so SmartPLS tests significance by resampling: bootstrapping draws thousands of samples (with replacement) from your data, re-estimates the model each time, and uses the spread of those estimates to judge which paths are real. The output tables confuse many scholars — here is how to read them, path by path.
Run it right before you read it
- Use 5,000 subsamples (the reporting standard; 10,000 for final runs costs little).
- Choose two-tailed testing at 0.05 unless you have directional hypotheses you committed to in advance.
- Select percentile bootstrap confidence intervals (current methodological guidance prefers them over the older bias-corrected option in most situations).
- Complete the measurement model first — bootstrapping structural paths is meaningless while validity is unsettled.
The path coefficients table: your hypotheses, one row each
Each structural arrow gets a row with five numbers. Original sample (O) is the path coefficient from your actual data — its sign and size are what you interpret substantively. Sample mean (M) is the average across bootstrap runs (should sit close to O; a large gap hints at instability). Standard deviation (STDEV) is the bootstrap standard error. t-statistic = O/STDEV. p-value is the significance verdict.
What passes, what fails
- t ≥ 1.96 (two-tailed) ⇔ p < 0.05 — the path is significant at the conventional level; t ≥ 2.58 ⇔ p < 0.01; t ≥ 3.29 ⇔ p < 0.001.
- Confidence interval excluding zero — the modern, preferred criterion: if the 2.5%–97.5% interval doesn't straddle zero, the path is supported. It always agrees with the two-tailed p at matching levels, and reviewers increasingly ask for the CI columns explicitly.
- A significant path with a tiny coefficient (say β = 0.06) is statistically real but practically marginal — interpret size, not just stars. Path coefficients around 0.1 are weak, 0.2–0.3 moderate, 0.5+ strong in most management models.
Beyond the paths: what else to bootstrap-read
- Indirect effects (specific, not total) — for mediation hypotheses: a significant specific indirect effect whose CI excludes zero is the mediation evidence; see mediation and moderation explained.
- Outer weights/loadings — for formative constructs, bootstrapping tells you which indicators contribute significantly.
- R² and f² come from the standard estimation, not bootstrapping — report them alongside: R² for explained variance of each endogenous construct, f² (0.02/0.15/0.35 = small/medium/large) for each predictor's contribution.
Bootstrapping is random resampling, so t-values shift slightly between runs — that's normal, not an error. Results stabilise as subsamples increase. Fix the random seed (or report 'results based on 5,000 subsamples') and never re-run repeatedly hunting for a p that dips under 0.05 — that is p-hacking.
The reporting format reviewers expect
One table, one row per hypothesis: path, β (original sample), t, p, 95% CI, and the supported/not-supported verdict. Example: H1: Perceived usefulness → Adoption intention, β = 0.42, t = 6.31, p < 0.001, CI [0.29, 0.55] — supported. Round to two decimals, state the subsample count and CI method in the table note, and interpret the size of each supported effect in the text, not just its stars.
If your model's results look unstable — signs flipping, wild t-values, mediation that appears and disappears — the problem usually sits upstream in the measurement model or sample size, not in bootstrapping itself. That diagnosis is exactly what SmartPLS mentoring is for; the Statistical Test Selector and sample size calculator cover the planning end.
Frequently asked
How many bootstrap subsamples should I use in SmartPLS?+
5,000 is the accepted reporting standard; use 10,000 for your final run if time permits. Fewer (500) is fine only for quick exploratory checks — never for reported results.
What t-value is significant in SmartPLS?+
For two-tailed tests: 1.96 at the 0.05 level, 2.58 at 0.01, and 3.29 at 0.001. One-tailed thresholds are lower (1.645 at 0.05) but require a directional hypothesis committed before analysis.
Should I report p-values or confidence intervals?+
Both. The percentile bootstrap CI is the methodologically preferred evidence (a path is supported when its CI excludes zero), while t and p remain the familiar convention. Reporting β, t, p and the 95% CI per path satisfies every reviewer.
phdguide's mentors are senior academics, former supervisors, statisticians and publication specialists with 25+ years' average experience guiding MBA, MPhil and PhD scholars from topic to viva.
Ethical, compliant guidance: We provide academic support, mentoring, analysis, editing and structuring — not authorship. Your work stays compliant with university policies.