PLS-SEM · Step 2

Assessing a structural model in SmartPLS: read β, R², f², Q², T and P to conclude your hypotheses

Bootstrapping finished and every table is full of numbers. Which hypotheses are supported, how do you write the conclusion, and what happens if the P value is above 0.05?

⏱ ~6 min readSource: Hair et al. (2022)Updated:

Your scales pass, you run bootstrapping, and out come table after table: Path Coefficients, T Statistics, P Values, R Square, f Square… Which hypotheses are supported? How do you write the conclusion? And what if your favorite hypothesis comes back with a P value of 0.21? This guide covers assessing a structural model in SmartPLS per Hair et al. (2022): reading the 9 indicators correctly, building your hypothesis table and handling unexpected results honestly.

Structural model: functional and social value significantly affect perceived value, time risk does not, R² equals 0.684
H1 and H2 are supported; H3 has P = 0.389, so it is not.
🧭 The order of assessment

① Check collinearity (VIF) → ② explanatory and predictive power (R², f², Q²) → ③ test the hypotheses (β, T and P from bootstrapping).

1VIF: checking collinearity in the structural model

Collinearity arises when predictors of the same dependent variable correlate too strongly with each other, which distorts the path coefficients. SmartPLS calculates VIF for each set of predictors pointing at one dependent variable.

Example
Predictor → Purchase IntentionVIF
Perceived value (PV)1.84
Trust (TR)2.12
Service quality (SQ)2.45

All are below 3, so there is no collinearity problem.

  • VIF < 3: Ideal
  • 3 ≤ VIF < 5: A possible problem; keep an eye on it
  • VIF ≥ 5: Critical collinearity

2R²: what is a good value?

R² is the share of variance in the dependent variable explained by its predictors.

Example

Functional, Social and Environmental Value explain Perceived Value with R² = 0.684: 68.4% of the variance is explained and 31.6% comes from factors outside the model.

  • 0.75: Substantial
  • 0.50: Moderate
  • 0.25: Weak
✦ Do not panic over a low R²

These levels are only a guide. Hair et al. (2022) note that in consumer behavior research, an R² of around 0.20 can already be considered high. Compare with studies in your own field.

3Adjusted R²

A version of R² that accounts for the number of predictors, since adding any variable raises R² even if it means nothing.

Example

R² = 0.705 and adjusted R² = 0.697 (a gap of 0.008): the predictors all contribute. With R² = 0.72 but adjusted R² = 0.54, the model may carry surplus variables.

4f²: which predictor has the strongest effect?

From Cohen (1988), f² measures each predictor’s own contribution to the dependent variable’s R², by comparing R² with and without it.

Example
Pathf²
PV → PI0.42
Social Value → PV0.18
Physical Risk → PV0.04

PV → PI is a large effect, Social Value medium and Physical Risk small.

  • 0.02: Small
  • 0.15: Medium
  • 0.35: Large

5Q²: the model’s predictive power

R² tells you how well the model explains the data you have; Q² tells you how well it predicts. In SmartPLS 3, the Stone–Geisser Q² comes from blindfolding; SmartPLS 4 assesses predictive power with Q²predict through the PLSpredict procedure.

Example
Endogenous constructQ²
Perceived Value0.412
Purchase Intention0.327

Both are above 0, so the model has predictive relevance.

  • Q² > 0: Predictive relevance (0.25 and 0.50 mark medium and large)
  • Q² ≤ 0: No predictive relevance

6Paths: your hypotheses

Each arrow between two constructs is one hypothesis to test.

Example
  • H1: Functional Value → Perceived Value
  • H2: Social Value → Perceived Value
  • H3: Perceived Value → Purchase Intention

7Path coefficients (β): direction and strength

Standardized regression coefficients from −1 to +1: the sign gives the direction and the absolute value the strength.

Example
Pathβ
Functional Value → PV0.521
Social Value → PV0.247
Physical Risk → PV-0.184

A strong positive effect, a weaker positive effect and a negative effect.

8T statistics, from bootstrapping

T statistics come from bootstrapping and test whether β differs significantly from 0. Note that the PLS algorithm gives no T or P values, so bootstrapping is essential (Hair et al., 2022 recommend 10,000 samples for final results).

  • T > 1.645: Significant at the 10% level
  • T > 1.96: Significant at the 5% level
  • T > 2.576: Significant at the 1% level

9P values: is the hypothesis supported?

The P value is the probability of seeing a result at least this extreme if the null hypothesis (H₀: β = 0) were true. The smaller it is, the stronger the evidence for your hypothesis.

  • P < 0.05: Significant; the hypothesis is supported
  • P < 0.01 / < 0.001: Highly / very highly significant
  • P ≥ 0.05: Not statistically significant
Decision chart: T above 1.96 and P below 0.05 means supported, otherwise review and discuss
How to conclude a hypothesis from T and P.

10Building the hypothesis table

HypothesisβTPConclusion
H1: FV → PV0.52112.846<0.001Supported
H2: SV → PV0.2474.963<0.001Supported
H3: TR → PV-0.0520.8610.389Not supported
✍️ How to word it

“Functional value has a positive effect on perceived value (β = 0.521; T = 12.846; p < .001), so hypothesis H1 is supported.” Write “p < .001” rather than “p = .000”.

11P value above 0.05: what now?

An unsupported hypothesis is not an error. It is a finding. Before you conclude, check:

  1. The measurement model: does it pass (AVE, HTMT)? Weak scales weaken coefficients.
  2. Collinearity: a high VIF can “share out” an effect among predictors.
  3. Sample size: do you have enough power for a small effect?
  4. Indirect effects: the variable may work through a mediator rather than directly.
  5. Context: sample characteristics, culture and timing may differ from the original study.
⚠️ Report it honestly

Do not add, remove or edit responses to make a P value look better. A committee will value a convincing discussion of why a hypothesis was not supported over a table of nothing but “supported”.

12Testing mediation in SmartPLS

After bootstrapping, look at Specific Indirect Effects. An indirect effect is significant when P < 0.05 (the 95% confidence interval excludes 0). Then compare it with the direct effect:

  • Indirect significant, direct not significant → full mediation.
  • Both significant → partial mediation.
  • Indirect not significant → no mediation.

Plan your results chapter before you have data

On The Light Data’s model canvas, draw your model with positive and negative hypotheses and moderators, pick SmartPLS and read the path coefficients, bootstrap and R² tables on simulated data. It helps you picture how to present your results. Simulated data is for learning only and never replaces real data.

🎯 Reading the results in 30 seconds
  • VIF < 3 · compare R² and Q² with your field · f² shows which predictor is strongest.
  • β gives direction and strength · T > 1.96 and P < 0.05 means the hypothesis is supported.
  • T and P come from bootstrapping; P > 0.05 is a finding to discuss, not an error.
①
Back to step 1
Assessing the measurement model
The 6 criteria to check before the structural model.
→

Frequently asked questions

Is a P value above 0.05 a wrong result?

No. It is a result: the data does not support that hypothesis. Review the data, the measurement model and the sample size, then report it honestly and discuss why in the context of your study.

What is a good R² value?

0.75, 0.50 and 0.25 are often described as substantial, moderate and weak (Hair et al., 2022), but compare with studies in your field; in consumer behavior research, an R² of around 0.20 can already be considered high.

How many bootstrap samples should I use?

Hair et al. (2022) recommend 10,000 bootstrap samples for final results; a smaller number is fine for quick checks.

Where do I find Q² in SmartPLS 4?

SmartPLS 4 assesses predictive power with Q²predict through the PLSpredict procedure; the Stone–Geisser Q² from blindfolding was the familiar route in SmartPLS 3. Values above 0 indicate predictive relevance.

How do I test mediation in SmartPLS?

After bootstrapping, check Specific Indirect Effects. If the 95% confidence interval excludes 0 (or P < 0.05), the indirect effect is significant; compare it with the direct effect to decide between full and partial mediation.