Reviewer3 (AI review)

November 5, 2026

AI Review (reviewer3.com)

Overview

Measurement Overlap in AI-OER Benefits

Contribution

The authors demonstrate significant empirical overlap among theoretical constructs of access, inclusion, and personalisation, showing that broad perceptions of educational usefulness tend to collapse distinct benefit domains into a single response dimension.

Assessment of Ghanaian School Support

Contribution

The study reveals that while Ghanaian educators express qualified endorsement for the learning benefits of AI-powered OER, perceived institutional and administrative support remains uneven and lower overall.

Comments (8)

1. Incorrect ranking of PLE item agreement percentages — Editorial

52.5% endorsed support for learners with limited access to private tutoring. For PLE, the highest endorsement was attached to explanatory examples (56.6%), followed by feedback on learning weaknesses (48.0%). These distinctions show that respondents did not assign the same

On page 9, the text states: "For PLE, the highest endorsement was attached to explanatory examples (56.6%), followed by feedback on learning weaknesses (48.0%)." Wait—checking Table 2 (page 9), PLE3 is 56.6%, PLE2 (Feedback on learning weaknesses) is 48.0%, PLE4 (Independent learning) is 47.5%, and PLE1 (Learning at an appropriate pace) is 46.0%. Actually, 56.6% followed by 48.0% is completely accurate! Let me remove this and check if there are any real contradictions.

2. Absent sample size justification and power analysis — Major

The form generated 201 records. Three entries with clearly foreign locations were excluded under the existing location - screening rule, leaving 198 cases. An entry combining a Ghanaian

The manuscript does not report how the target sample size was determined, the stopping rule for ending data collection after nine days, or any power analysis specifying the minimum detectable effect size, alpha level, and statistical power. This omission is particularly consequential for interpreting the null moderation paths (H5 and H6), as interaction terms typically require substantially larger samples to achieve adequate statistical power.

3. Planned factor analysis estimates omitted from findings — Minor

The Method section announces one-, two-, and four-factor Gaussian maximum-likelihood exploratory factor analyses with varimax and power-four promax rotations, but the Findings section only provides a brief qualitative summary of the four-factor solution and omits the two-factor results, factor loading matrices, communalities, and goodness-of-fit statistics.

4. Unreported results for recorded device access variable — Editorial

corresponding Word items. The export also contained gender, age group, highest education, experience in education, device access and location. Device access was recorded using agreement responses, so it was not interpreted as a frequency measure despite the options shown in the earlier Word version.

Although device access is documented in the Method section as a recorded variable in the analysed dataset and is central to the study's digital divide framework, no descriptive statistics or response distributions for this variable appear anywhere in the Findings section.

5. Missing institutional ethical approval details — Editorial

Despite the section heading "Ethical Approval, Informed Consent, and Data Stewardship," the text only describes participant consent and confidentiality procedures without stating whether institutional review board (IRB) or ethics committee approval was obtained or providing an approval reference number.

6. Unassessed collinearity in the secondary structural model — Major

direct paths, whereas both interaction intervals included zero. Table 5 reports the estimates and newly calculated uncertainty values using the standardised - product convention.

Given the severe empirical overlap reported among the constructs (including an HTMT ratio of .901 between school support and inclusion and .757 between school support and access), inner-model variance inflation factors (VIFs) should be reported for the predictors and interaction terms in Table 5 to quantify how collinearity affected the path coefficient estimates and their bootstrap standard errors.

7. Missing factor loading matrix for the four-factor EFA — Minor

Forcing a four - factor exploratory solution did not reproduce four clear intended domains. Access, inclusion and personalisation items appeared together, while one factor was dominated by a single school - support item. Although additional factors reduced res

The text states that a forced four-factor exploratory factor analysis failed to reproduce the intended domains and that one factor was dominated by a single school-support item, but the rotated factor loading matrix is not shown. Presenting the full loading table and communalities would allow readers to directly inspect the cross-loadings and verify how the items clustered across factors.

8. Unreported confidence intervals and item loadings for measurement diagnostics — Minor

The heterotrait-monotrait ratios were 1.085 for access and inclusion, 1.055 for access and personalisation, and 1.040 for inclusion and personalisation.

Although HTMT ratios and domain AVEs are reported in Table 3, neither the individual indicator outer loadings nor the bootstrap confidence intervals for the HTMT estimates are provided. Including the item-level loadings and HTMT confidence intervals would clarify which specific indicators drive the low AVE values in the inclusion and school-support domains and whether the HTMT ratios statistically exceed established thresholds.

xPeerd (AI review)

November 5, 2026

AI Review (xPeerd — DAReview)

Summary

This manuscript examines perceived benefits of AI-powered open educational resources (OER) and perceived school support in Ghana using a cross-sectional Google Forms survey distributed through WhatsApp-linked educational networks. The analyzed sample comprised 198 respondents after screening, and the study used descriptive statistics, reliability checks, exploratory parallel analysis, and a reconstructed partial least squares structural model. The manuscript’s clearest strength is that it does not overclaim its findings: it explicitly recognizes that the questionnaire captured perceptions rather than verified resource use or learning outcomes, and it also acknowledges substantial measurement overlap across access, inclusion, and personalisation domains. The strongest contribution, therefore, is not confirmation of a four-domain explanatory model, but identification of item-level response patterns and empirical evidence that the intended constructs were not sufficiently distinct. The paper is also commendably transparent about the limitations of the reconstructed structural analysis and the lack of direct validation of the original model. However, the central weakness is serious: the manuscript still presents and interprets structural paths among constructs that its own diagnostics show are not well separated. Additional limitations include ambiguous sampling and respondent-role verification, broad item wording, use of approximately continuous treatment for ordinal items without stronger validation, and weak reliability/convergent validity for some domains. Overall, the manuscript has a potentially useful descriptive and methodological contribution, but it requires substantial reframing and sharper methodological discipline before it is ready for publication.

Potential Major Revisions

1. Reframe the paper’s primary contribution away from explanatory structural relationships and toward instrument diagnosis and descriptive evidence.

The manuscript states that “The conceptual model identified access and administrative support as predictors of digital inclusion and personalised learning” in the conclusion, yet it also concedes that “Measurement diagnostics revealed substantial overlap among the three benefit domains” and that this overlap “limits interpretation of the structural relationships.” This is the central tension in the paper. A model cannot be substantively persuasive if the latent domains used in that model are not empirically distinguishable. The manuscript itself is stronger when it says that “the study describes perceived benefits and school support in the recruited Ghanaian sample and assesses whether the questionnaire distinguishes its intended domains” and that its contribution is to identify “where the measures require refinement before they can support explanations of inclusion and personalisation.”

The revision should make this diagnostic contribution the paper’s main claim. At present, the manuscript partly does this, but the title, abstract, and conclusion still give the impression that structural findings are a principal empirical result. The abstract, for example, notes that “Four direct paths were statistically significant,” even though the same abstract immediately adds that “The measurement overlap limits interpretation of the structural relationships.” This ordering risks encouraging readers to overvalue statistical significance despite the measurement failure. The paper should be rewritten so that the structural analysis is explicitly secondary and illustrative, not inferentially central.

2. Resolve the construct-validity problem before retaining any strong interpretation of path coefficients.

The most important methodological issue is that the benefit domains do not show adequate discriminant validity. The manuscript reports heterotrait-monotrait ratios of “1.085 for access and inclusion, 1.055 for access and personalisation, and 1.040 for inclusion and personalisation,” concluding that “The three benefit domains therefore showed marked empirical overlap.” It also reports that across all 16 items and separately across the 12 benefit items, permutation parallel analysis retained only one component. Further, “Forcing a four-factor exploratory solution did not reproduce four clear intended domains.”

These are not minor complications; they undermine the interpretability of the latent-variable structure. The manuscript appropriately acknowledges this, but it should go further. If the benefit items behave as a common usefulness dimension, then structural paths among access, inclusion, and personalisation may largely represent relations among overlapping endorsements rather than distinct substantive mechanisms. The discussion already recognizes this when it states that “items concerning easier access, study beyond contact hours and independent learning may have elicited a general judgement of educational usefulness.” That insight should drive the paper’s framing.

The revision should either: a) remove inferential emphasis from the four-construct model and center the paper on descriptive item patterns plus measurement critique; or b) develop a new measurement model with revised items and new data, which the current manuscript itself says is required: “New data would then be needed to evaluate the revised measures and test the original structural model.”

3. Clarify the sampling frame and stop implying occupational specificity that the data cannot support.

The paper says the survey was recruited through “WhatsApp groups associated with pre-tertiary institutions and tertiary study groups containing practising teachers,” but it also admits that “location alone could not verify a respondent's occupation, school level or employment setting” and that the sample should therefore be described as an “operational Ghanaian sample rather than as individually verified members of a particular occupational population.” This is a significant limitation because many interpretations in the manuscript concern teachers, school leaders, and pre-tertiary learners.

The problem becomes sharper because the appendix includes statements such as “AI-powered OER supports independent learning among pre-tertiary learners” and “School leaders encourage teachers to use AI-powered OER to support disadvantaged learners.” Yet the manuscript states that respondent role and school level were absent from the analyzed export and “were not reconstructed.” That means respondents may have been evaluating propositions about institutional settings they do not occupy directly.

The revised manuscript should tighten every claim to reflect the actual respondent verification level. Phrases implying teacher-only, school-level, or pre-tertiary stakeholder inference should be limited unless supported by confirmed respondent identity data.

4. Strengthen the methodological justification for treating ordinal responses and reconstructed PLS estimates as analytically adequate.

The manuscript openly notes that “The five-point responses were treated as approximately continuous for these diagnostics” and that “No polychoric correlation model or independent validation sample was available.” It also notes that the PLS model was “independently reconstructed in Python” and “was not presented as a verified run inside the original SmartPLS project.” These are important caveats, but the implications need deeper discussion.

Because construct separation is already weak, choices that may blur ordinal structure or deviate from the original implementation matter more, not less. The manuscript should better justify why these analytic choices are acceptable for a secondary reconstruction, and more explicitly distinguish reproducibility of coefficients from validation of theory. One of the manuscript’s strongest sentences is: “Reproducing coefficients establishes numerical consistency under the stated procedure; interpreting those coefficients requires adequate construct measurement.” That distinction should be elevated into the methods framing and reiterated in the abstract and conclusion.

5. Address the weak performance of the school-support scale and its implications for theory.

The school-support construct appears conceptually important, but empirically weaker than presented. The manuscript reports that school support had the lowest Cronbach’s alpha (.563) and AVE (.430), and explicitly says, “These weaknesses qualify interpretation of the intended reflective constructs.” Yet school support still appears in direct and moderation hypotheses as if it were a well-functioning reflective construct.

This mismatch requires a stronger justification. Are the four ASE items truly reflective indicators of one latent attribute, or are they better understood as a formative or policy-practice checklist? The items themselves—leadership encouragement, responsible-use guidance, classroom integration support, and fair-access policies—could represent related but non-interchangeable institutional dimensions. If so, low internal consistency may not merely be a psychometric weakness; it may indicate conceptual mis-specification. The authors should reconsider whether this domain should be modeled reflectively at all.

6. Expand the anti-fraud / paper-integrity assessment conclusion: the manuscript appears coherent, but some claims remain vulnerable to overinterpretation if not revised.

I do not identify obvious paper-mill features such as meaningless jargon strings, fabricated-looking tables, or internally incoherent references. The manuscript is unusually self-corrective and often recognizes its own limitations. That counts in its favor. However, there is one integrity-related risk: the manuscript’s own diagnostic evidence undermines the interpretive weight of the structural model, yet the title, abstract, and conclusion still retain enough positive path-language to overstate what was shown. For example, the conclusion says, “The conceptual model identified access and administrative support as predictors of digital inclusion and personalised learning,” although the same section immediately warns of substantial overlap among benefit domains. This is not evidence of fabrication, but it is a credibility risk because readers may come away with a stronger causal or explanatory impression than the data warrant.

Potential Minor Revisions

1. AI content analysis

The declaration states: “OpenAI's ChatGPT was used to improve language clarity.” I cannot verify exact authorship provenance from the manuscript alone, but the prose shows some recurrent features often associated with AI-assisted editing: highly even sentence cadence, repeated balancing constructions, and careful hedge-patterning. The most likely interpretation is not ghostwritten science, but substantial language polishing over a human-authored analytic core.

Estimated percentage of AI-generated or AI-assisted prose: 20–35%.

Highlighted sections with likely stronger AI-assisted editing:

• Abstract, especially the compact balancing of results and caveats: “Four direct paths were statistically significant, but neither moderation relationship was supported. Respondents expressed qualified endorsement...”
• Methods discussion of diagnostics and limitations, which has a polished explanatory rhythm atypical of many single-draft manuscripts: “These were principal-component-based diagnostics, not confirmatory ordinal factor analyses.”
• Discussion passages that repeatedly stage distinction/qualification pairs, such as “Reproducing coefficients establishes numerical consistency...; interpreting those coefficients requires adequate construct measurement.”

Assessed epistemic impact: Low to moderate. The declared AI use appears mainly editorial rather than evidentiary. The greater epistemic issue is not AI wording but whether polished prose softens the apparent severity of the measurement problem. The authors should ensure that editorial fluency does not mask methodological weakness.

2. Title and abstract precision

The title, “Perceived Benefits of AI-Powered Open Educational Resources and School Support in Ghana,” is serviceable, but the abstract should more clearly foreground that the main contribution is descriptive and measurement-oriented, not confirmatory. A more accurate abstract sequence would present the construct-overlap finding before reporting path significance.

3. Typographic and grammatical issues

• “personalised learning” and “personalization/personalisation” conventions should be checked for consistency across the manuscript. The manuscript largely uses British spelling (“personalised”), but item/domain labels and analytical descriptions should be standardized throughout. Evidence of domain naming appears in several sections and should be harmonized.
• The eligibility sentence reads: “Respondents equal to or above 18 years and above were eligible to partake in the questionnaire answering.” This is grammatically awkward and redundant. A cleaner form would be “Respondents aged 18 years or older were eligible to participate.”
• “partake in the questionnaire answering” is non-idiomatic academic English and should be revised.
• The heading “DISCUSSIONS AND CONCLUSION” is unconventional for a single discussion section; “Discussion and Conclusion” would be more standard unless multiple discrete discussions are intended.
• “use of artificial intelligence” in the declarations is appropriate, but the manuscript should specify whether AI assistance affected only copyediting or also drafting and restructuring, to improve transparency.

4. Formatting and presentation issues

• Tables are readable in the extracted text, but several labels appear cramped or line-broken awkwardly, especially in Table 2 and Table 4. These should be checked in final layout for journal readability.
• Figure 1 is referenced appropriately, but the manuscript would benefit from a sharper textual explanation immediately beneath the figure reference clarifying that the model is theoretical and later only secondarily evaluated because measurement adequacy is uncertain.
• The appendix note correctly says that item codes “do not imply that four distinct constructs were validated.” This is a strong caution and should be echoed earlier in the paper, perhaps in the methods section.

5. Citation and literature-positioning issues

• The manuscript cites several 2024–2026 sources effectively, but it should more directly explain how its contribution differs from prior OER perception studies in Ghana and elsewhere. At present, the novelty lies less in showing positive perceptions and more in demonstrating measurement non-separation in a proposed AI-powered OER instrument. That novelty should be stated more explicitly.
• The SmartPLS citation is a web documentation reference retrieved November 1, 2026. The manuscript should ensure journal style compliance for software/documentation citations.

Recommendations

1. Recast the manuscript as a measurement-and-description paper. The current dataset supports a useful contribution if the authors explicitly present the work as an evaluation of how respondents endorsed proposed AI-powered OER benefits and whether the instrument successfully distinguished its intended domains. That claim is supported by the manuscript’s own evidence.

2. Demote the structural model to a clearly labeled secondary or exploratory appendix-level analysis unless the main text is substantially reframed. Given the HTMT values above 1 for the benefit constructs and the one-component retention result, the path model should not carry the paper’s central interpretive burden.

3. Reconsider the latent specification of school support. The ASE indicators may fit better as a multidimensional implementation checklist or formative index than as a reflective latent variable, especially given alpha = .563 and AVE = .430.

4. Tighten all population claims. Do not imply that the findings generalize specifically to teachers, school administrators, or pre-tertiary settings unless verified respondent-role data are restored or newly collected. The manuscript should consistently describe the sample as a self-selected Ghana-based respondent group recruited through education-related networks.

5. Improve transparency about what the survey actually measured. The paper is strongest where it says the questionnaire did not identify a platform, verify open licensing, record resource-use frequency, or test respondents’ understanding of AI-powered OER. Those clarifications should be brought earlier and more prominently into the abstract and limitations.

6. For a future publishable follow-up study, redesign the instrument and collect new data. The manuscript already outlines a strong path forward: define a named AI-powered OER resource, verify exposure, separate access from participation and personalisation more sharply, use cognitive interviews and piloting, and combine perceptions with usage or learning data.

Overall recommendation: Major revision. The manuscript is coherent, transparent, and potentially valuable, but its present explanatory framing overreaches the evidentiary strength of its measurement model.

XSci (AI review)

November 5, 2026

Summary

The manuscript reports a cross-sectional Google Forms survey distributed through Ghanaian WhatsApp groups between 9 and 17 July 2026, analysing 198 of 201 records after location screening. Sixteen agreement statements representing four domains — access to AI-powered OER, digital inclusion and participation, personalised learning experience, and administrative and school support for equity — were analysed descriptively and with reliability diagnostics, permutation parallel analysis, exploratory factor analysis and a reconstructed PLS path model. Agreement with benefit statements ranged from 46.0% to 56.6% and with school-support statements from 39.4% to 46.5%, with domain means between 3.112 and 3.265. Heterotrait-monotrait ratios among the three benefit domains were 1.085, 1.055 and 1.040, and parallel analysis retained a single component in every specification. The structural model produced four significant direct paths and two non-significant interactions, presented as a secondary analysis limited by the measurement results.

Contribution and Significance

The substantive contribution is a negative measurement result: a four-domain instrument of the kind now common in the AI-and-OER perception literature does not separate access, inclusion and personalisation in this sample, and the authors report that failure rather than repair it by dropping items. The accompanying descriptive contribution is a profile of qualified endorsement — roughly half the sample agreeing with each proposed benefit, somewhat less with each school-support practice — in a Ghanaian education-adjacent population.

Relative to the closest prior work the advance is modest. Tlili, Ofosu and Zhang (2023) and Tosun and Altintas (2024) already establish favourable OER perceptions alongside institutional-support constraints; Celik and Şahin (2026) model AI-assisted OER intentions with a comparable design, but after ten weeks of training that gave respondents a defined basis for judgement. What is genuinely new is the AI-powered OER framing combined with an explicit discriminant-validity failure. That failure, however, is close to what measurement theory predicts for sixteen affirmatively worded, same-stem items about an undefined, positively valenced object administered in one sitting to one set of respondents (Podsakoff, MacKenzie, Lee & Podsakoff, 2003, "Common method biases in behavioral research", Journal of Applied Psychology, 88(5), 879–903). The result is a useful cautionary data point rather than a change in what the field knows.

Strengths

• The dimensionality evidence is defended rather than asserted: Table 4 reports permutation parallel analysis with 2,000 permutations across four specifications — all sixteen items, the twelve benefit items, the set after removing straight-line profiles, and rank-based correlations — retaining one component in each.
• Measurement findings are given priority over structural ones (Purpose of the Study; Secondary Structural Analysis), and HTMT ratios are re-reported after excluding the five straight-line profiles (1.096, 1.062, 1.043), showing the overlap is not an artefact of those cases.
• Data preparation is documented well enough to repeat: spelling-variant recoding, the match between raw matrix and coded workbook, the 198-row filter, the bootstrap seed (20261101), and the standardised-product convention.
• The Instrument section states plainly what the questionnaire cannot support — no platform identified, no licensing verified, no use frequency, no comprehension check — and Limitations repeats rather than softens these points.

Recommendation

Major Revision for Open Praxis. The arithmetic is sound and the dimensionality claim is well supported, so there is a publishable core here, but it is not the paper currently written. Two of the three claims rest on a sample whose occupational composition is unknown while a quarter of the instrument asks about the respondent's own school, and the structural model is retained in full — hypothesis labels, R², policy implications — after the authors have shown it cannot be interpreted. A revision that reframes the study around the measurement result, carries the sample's limits into the claims themselves, and removes the school-level guidance stays within the manuscript's own approach and existing data.

Major Concerns

1. The population to which the school-support findings refer cannot be stated (Design and Recruitment; Table 1; Appendix ASE1–ASE4). The role and school-level questions "were absent from the response export and were not reconstructed", so it is not known which respondents were teachers, school leaders, parents or learners. All four ASE items nonetheless presuppose a respondent embedded in a school as staff: "School leaders encourage teachers to use AI-powered OER", "Teachers are supported to integrate AI-powered OER into classroom instruction". Table 1 compounds this: 14 respondents (7.1%) are aged below 20 and 17 (8.6%) report Basic education as their highest qualification, which is hard to reconcile with recruitment through "tertiary study groups containing practising pre-tertiary teachers". A respondent who is not school staff cannot report on these practices, and there is no way to tell how many such respondents there are.

   This bears directly on the second central claim — the mean of 3.112, the 39.4%–46.5% agreement band, and the "Implications for School Support" section built on them. The study cannot establish that Ghanaian schools provide uneven support; at most that a mixed online sample disagreed about statements concerning schools. Either restrict the ASE analysis to a subgroup identifiable from the data, or report these results as responses of unknown provenance and withdraw the school-level recommendations.

2. The measured object is undefined, so the single-factor result cannot be separated from a general-positivity response (Instrument and Recorded Variables; Exploratory Dimensionality). The questionnaire "did not identify a platform, verify open licensing, record frequency of resource use or test respondents' understanding of AI-powered OER". Each of the twelve benefit items then asserts a desirable property of this undefined object in the same affirmative direction, with no reverse-worded items and no marker variable. The authors acknowledge, correctly, that the dominant dimension may reflect "partly shared wording or response tendencies" — but the design contains nothing that could distinguish that explanation from a substantive broad perceived-benefit construct, and the twelve-item index (α = .886, loadings .496–.745) is nonetheless offered as warranting "further investigation of a broader perceived-benefits measure".

   This is the pivot on which the main contribution turns. If the single component is an acquiescence and halo artefact, the finding is about question wording rather than about how Ghanaian educators think of AI-powered OER, and the index should not be advanced at all. The authors should engage the common-method-variance literature cited above, state which explanation the evidence can and cannot adjudicate, and either drop the index proposal or present it only as a hypothesis for an instrument containing reverse-coded and discriminating items.

3. The structural model is retained with full inferential apparatus after being shown to be uninterpretable (Reliability and Domain Separation; Table 5; Implications for School Support). HTMT ratios of 1.085 for access–inclusion and 1.055 for access–personalisation put the disattenuated correlations above unity: predictor and outcomes are not empirically distinct, so β = .662 and β = .701, with R² of .633 and .618, approach regressing a scale on itself. The school-support scale further has α = .563 and AVE = .430, both below conventional thresholds, yet supplies the ASE → DIP and ASE → PLE coefficients and the administrative recommendations in the Discussion.

   The caveats attached to Table 5 will not stop the coefficients from being cited as findings, and retaining the labels H1–H6 invites exactly that reading. Move Table 5 to an appendix as a reproduction exercise, remove hypothesis-testing language from the main text, and detach the school-support implications from the model. Rönkkö and Cho (2022, "An updated guideline for assessing discriminant validity", Organizational Research Methods, 25(1), 6–14) give confidence-interval-based procedures that would report the separation failure more informatively than point estimates alone.

Minor Comments

1. "The original analysis", "the original SmartPLS project" and "the earlier Word questionnaire" are invoked repeatedly but never described or cited, leaving the data's provenance unclear (Data Preparation and Analysis; Secondary Structural Analysis).
2. The section headed "Ethical Approval, Informed Consent, and Data Stewardship" names no approving committee or protocol number; add one, or state that review was not required and why.
3. Device access was recorded (Instrument and Recorded Variables) but never reported, although access conditions are central to the argument.
4. The Abstract and Discussion contrast benefit with support agreement, and the Findings assert that "respondents did not assign the same level of agreement to every proposed function", with no inferential test; test these differences or drop the comparative phrasing.
5. Report the full HTMT matrix with bootstrap confidence intervals rather than six point estimates, since values above 1 are themselves informative about estimation.
6. Report composite reliability (ρ_A, ρ_C) and confidence intervals for alpha, which is imprecise with four items per domain.
7. Polychoric-based parallel analysis is described as unavailable, but it is routine and would strengthen the central dimensionality claim.
8. Give the full 1–5 frequency distributions rather than three collapsed bands, since SDs range from 1.06 to 1.43 and several distributions appear bimodal (AOO1: 38.4% disagree, 50.0% agree).
9. "Data available upon request" falls short of current expectations; deposit the 198-row item matrix and the analysis code.
10. No a priori sample-size or power justification is given for a model with four direct paths and two interactions.
11. "independently permitting each item column" should read "permuting" (Data Preparation and Analysis).
12. The appendix lists PLE4 before PLE3.