Viva verification sheet

Working notes behind the ratings. Each item below is arithmetic done on the candidate’s own reported numbers, so it can be checked in front of the candidate.

1. R² is not reconcilable with the path coefficients

In a standardised model, R² = Σ (βᵢ × rᵢ), where rᵢ is the correlation between predictor i and the outcome.

Using Table 4.16 betas and Table 4.13 correlations with AI Adoption:

Reported R² = 0.864. An R² above 1.0 is impossible. The same check on the AI Readiness block gives an implied R² of 0.452 against a reported 0.864.

Ask: which SmartPLS run produced Table 4.16, and which produced Table 4.15?

2. p-values do not match the t-statistics

The three non-significant paths are correctly reported. Only the significant ones are wrong — consistent with manual entry rather than a software fault. Every “p = 0.001” result is in fact significant only at 0.05.

3. Two indirect effects ≠ a × b

H12 and H13 are also declared supported although their a-paths are non-significant (p = 0.287 and p = 0.434).

4. HTMT cannot be smaller than the construct correlation

HTMT is a disattenuated correlation and is therefore always ≥ r. These two tables cannot come from the same model.

5. Fornell–Larcker diagonals ≠ √AVE

Three values inflated by exactly 0.100; three correct.

6. Spearman correlations vs PLS latent correlations — same data, opposite signs

Chapter 4 offers a theoretical explanation for the negative TR–AIA relationship (§4.7) and a theoretical explanation for the positive one (§4.10), and retains both.

7. The EFA is a Harman single-factor test, read backwards

Bartlett’s df = 15. Since df = p(p−1)/2, p = 6 — the analysis was run on the six construct composites, not on the 30 items. Eigenvalues sum to 6.000, confirming this.

One component extracting 77.4% of variance across six constructs is the classic signature of common method variance (threshold: 50%). The thesis reports it as evidence of “satisfactory construct dimensionality.”

There is no other CMB test in the thesis.

8. Descriptive means are below the scale midpoint

Seven-point scale → neutral = 4.0.

All six are below neutral, i.e. mild disagreement. Chapter 4 reads these as respondents having “moderately agreed” and “generally tended to agree.”

Note also the tension with Table 4.1, where 100% of firms report AI implemented (33.6% “fully integrated”) while AIA itself averages 3.014.

9. Sample vs stated eligibility criteria

Section 3.3.3 excludes: firms <10 employees, firms >250 employees, and firms that have fully implemented AI.

Appendix A offers “Less than 10” and “250+” as response options, so no screening filter was applied. The questionnaire also has no non-adopter option, and reports categories (“Other” ownership, n = 65; “Education/Training”, n = 21) that do not appear in the instrument, while omitting one that does (Manufacturing).

Questions to put to the candidate

1. Please open the SmartPLS project and show the structural model report — do the R², betas, t-statistics and p-values in Table 4.16 match the software output?
2. How were Tables 4.12 and 4.13 generated, given that they are mutually inconsistent?
3. Table 4.10 shows TR negatively correlated with AI Adoption and Table 4.16 shows a strong positive path. Which is correct?
4. AIA4, AIR2 and AIR4 were dropped. On what criterion, and why is this not reported?
5. How were the 92 out-of-scope firms handled, and why were full AI adopters retained when the criteria exclude them?
6. Why does the questionnaire have no “not adopted” option in a study whose problem statement is that adoption is low?
7. Chapter 5 states TR significantly and positively influenced AI Readiness. Chapter 4 reports β = −0.442, p = 0.287. How did this arise?
8. Several scale sections refer to “the reliability value of turnover intention.” Where did that text come from?

Bottom line

Current ratings total 75/120 = 62.5% → Pass with Major Revisions (40–69%).

The conceptual framing, literature work and written expression are at minor-revision standard. The mark is held down almost entirely by Chapter 4 and its knock-on effects in Chapter 5 (items 11a, 11b, 13a — 6 marks across three items).

If the candidate produces clean SmartPLS output at the viva showing these are transcription errors in the tables rather than defects in the analysis, items 11a, 11b and 13a could reasonably be re-rated to 3–4, taking the total to roughly 84/120 = 70% and into the minor-revision band. If the output does not reconcile, the model must be re-estimated and the major-revision route is the correct one.