A UTAUT2-based study of technology acceptance in GameFi platforms — N = 516
MSc dissertation research · University of Western Macedonia · School of Economic Sciences · Department of Management Science and Technology · MSc "Electronic Business and Digital Marketing" · Kozani, 2025
Author: Georgios Dikos · Supervisor: Prof. Ioannis Antoniadis
The complete quantitative pipeline behind the dissertation: the survey dataset, every Python script used to analyse it, and the raw console output of each run. Nothing is summarised away — if a number appears in the thesis, the script that produced it and the output it printed are both here.
The study extends UTAUT2 (Venkatesh, Thong & Xu, 2012) with four GameFi-specific constructs — Economic Motivation, Risk Perception, Trust in Technology, and Regulatory & Compliance Risks — and tests the resulting model on 516 GameFi users, predominantly from Southeast Asia and Latin America.
The hypothesised research model (Σχήμα 4 in the dissertation, p. 47). Blue = core UTAUT2 constructs, green = GameFi-specific extensions. Solid arrows are hypothesised direct effects; red arrows are hypothesised negative effects; dashed arrows are moderations. This is the model as proposed, not the model as confirmed. Two of the hypothesised negative paths were not supported — Trust → Risk Perception came out at β = +0.129 (p < 0.001) and Regulatory Compliance → Risk Perception at β = +0.394 (p < 0.001), both significantly positive rather than negative. Of the three moderators shown, only Experience was tested; Age and Gender were excluded because the sample offers too little variance to moderate on (91.7% are aged 18–34 and 75.2% are male). See Key findings below for what the data actually supported.
A note on the folder layout. Everything lives under
Users/Velze/MSC UTAUT2/. That path is a leftover from the original local project directory. It is kept deliberately: the dissertation text contains permanent hyperlinks into these exact paths, and renaming them would break every citation in the submitted document. Use the navigation table below rather than the folder tree.
| Section | What's in it |
|---|---|
| Questionnaire (PDF) | The full survey instrument as administered |
| Reliability — Cronbach's Alpha | α for all 13 scales, item-total correlations, α-if-deleted, per-construct data files |
| Factor Analysis | KMO, Bartlett's test, EFA, scree plot, loadings heatmap |
| PLS-SEM Analysis | Data cleaning, measurement model, structural models v1/v2a/v2b, moderation analyses |
| Cluster Analysis | k-means segmentation, elbow and silhouette diagnostics, cluster profiles |
| Cluster Analysis — Thesis version | The k-means run reported in the dissertation (Ch. 4.3). See Two clustering solutions below |
| Demographics | Age, gender, education, income, experience, region — charts and scripts |
| Constructs 1–13 | Folders 1-Performance-Expectancy-(PE) through 13.-Use-Behavior-(UB) — per-question distributions, charts and item wording |
The cleaned dataset is at
PLS-SEM Analysis test 2/utaut2_cleaned_data.xlsx
(516 × 50). The raw export is dataset-utaut2.xlsx in the same folder.
| Sample | 516 valid responses (534-row export; the final 18 rows are frequency tallies, not respondents) |
| Model | 12 latent constructs · 42 Likert indicators · 17 structural paths tested |
| R² Behavioural Intention | 0.712 (adj. 0.706) |
| R² Use Behaviour | 0.037 |
| Strongest predictor of BI | Habit (β = +0.355, f² = 0.216) |
| Then | Price Value (β = +0.249) · Social Influence (β = +0.150) |
| Measurement quality | α ≥ 0.80 on 12 of 13 scales · KMO = 0.949 · Bartlett's p < 0.001 |
1. Habit beats economics. Economic Motivation records the highest mean of any construct (M = 4.13) yet contributes essentially nothing once the other predictors are in the model (β = −0.006, ns). Users say earning is what matters; their intention is actually driven by routine and perceived value for money. Wanting to earn is near-universal in this sample, so it does not discriminate between who intends to keep playing and who does not.
A note on the dissertation abstract. The abstract describes Economic Motivation as the strongest predictor of intention and use. That is a description of the descriptive result: EM is the highest-rated construct in the sample (M = 4.13, above all eleven others). It is not a statement about predictive weight. In the structural model reported in Chapter 4.4 — and in
path_coefficients.xlsxin this repository — EM → BI is β = −0.006 (ns), and the strongest predictor of intention is Habit at β = +0.355. EM is not the strongest predictor of use either (r = +0.15 with usage frequency, against +0.24 for Behavioural Intention and +0.23 for Habit). Where the abstract and the structural model differ, Chapter 4.4 and the outputs in this repository are the authoritative account.
2. Trust runs the wrong way — and that is the interesting part. Trust in Technology has a negative direct effect on intention (β = −0.115, p < 0.01) and Risk Perception a positive one (β = +0.098). Both contradict the standard hypotheses. The moderation analysis explains why:
Trust × (Risk Perception → Behavioural Intention): ΔR² = 0.042, t(512) = −6.14, p < 0.001
Low-trust users: risk → intention slope = +0.659 High-trust users: risk → intention slope = +0.280
Risk raises intention, and trust dampens that effect. This fits a speculative-participation reading rather than a safety-seeking one: in this population, perceived risk appears to function partly as perceived upside.
3. Experience flattens the learning curve, and little else. Of five moderations tested by GameFi experience, only one holds: Experience × (Effort Expectancy → BI), β = −0.325, ΔR² = 0.034. Ease of use matters to newcomers and stops mattering once people are regular users. Social influence, facilitating conditions and habit effects do not vary by experience level.
Fully anonymous. No names, emails, IP addresses, wallet addresses, timestamps or free-text responses are included — those fields were stripped at export. Collected by online questionnaire under GDPR-compliant consent.
42 Likert items (1 = Strongly Disagree … 5 = Strongly Agree) across 12 constructs:
| Construct | Code | Items | Cronbach's α | Mean (SD) |
|---|---|---|---|---|
| Performance Expectancy | PE | 5 | 0.930 | 3.92 (0.83) |
| Effort Expectancy | EE | 4 | 0.886 | 3.38 (1.07) |
| Social Influence | SI | 3 | 0.804 | 3.69 (0.91) |
| Facilitating Conditions | FC | 4 | 0.836 | 4.03 (0.85) |
| Hedonic Motivation | HM | 4 | 0.890 | 3.86 (0.87) |
| Price Value | PV | 3 | 0.887 | 4.02 (0.87) |
| Habit | HB | 4 | 0.924 | 3.92 (0.95) |
| Behavioural Intention | BI | 3 | 0.912 | 4.09 (0.95) |
| Economic Motivation † | EM | 3 | 0.882 | 4.13 (0.94) |
| Risk Perception † | RP | 4 | 0.907 | 4.02 (0.91) |
| Trust in Technology † | TT | 3 | 0.809 | 3.43 (1.19) |
| Regulatory & Compliance Risks † | RC | 2 | 0.895 | 3.31 (1.25) |
† GameFi-specific extension to standard UTAUT2.
Plus Use Behaviour (UB — usage frequency and weekly hours, 2 items, α = 0.666) and six demographic variables.
Sample composition. 89.9% are current GameFi users, 75.2% of them weekly or more. 91.7% are aged 18–34 and 76.6% earn under $1,000/month. 44.2% are from Southeast Asia and 19.2% from Latin America and the Caribbean. 75.2% male.
This is a self-selected sample of active users in emerging markets, not a general population. Every finding above should be read inside that frame.
git clone https://github.com/Jojohorororo/MSC-UTAUT2-Research-GameFi.git
cd "MSC-UTAUT2-Research-GameFi/Users/Velze/MSC UTAUT2"
pip install pandas numpy openpyxl scipy statsmodels scikit-learn factor-analyzer matplotlib seabornEach analysis folder is self-contained — it carries its own copy of the data file its scripts
read, so cd into a folder and run. Suggested order:
cd "PLS-SEM Analysis test 2" && python test.py # cleaning: 534 → 516 rows
python part2.py # measurement + structural model v1
python v2a_analysis.py # v2a: adds RC → BI
python v2b_analysis.py # v2b: GameFi inter-construct paths
python final.py # experience moderations
python tt_rp_moderation.py
cd "../Factor Analysis" && python test1.py
cd "../Cluster Analysis" && python "cluster analysis.py"
cd "../Cronbach's Alpha" && python "1.Cronbach's Alpha Performance-expectancy-(PE).py"Every script writes its results next to itself. The .txt files beside each script are the
console output from the original runs, kept for comparison.
Tested on Python 3.10 and 3.11. The headline figures above were independently re-derived from
utaut2_cleaned_data.xlsx, and the reliability and cluster scripts re-run from a clean clone
reproduce their committed outputs exactly.
One compatibility note: the 13 scripts in Cronbach's Alpha contain a backslash inside an
f-string expression, which is a syntax error before Python 3.12. On older versions, either run
them on 3.12+ or change the line
print(f"\n{'Item':<6} {'Corrected Item-Total':<20} {'Cronbach\'s α if':<15}")to
print("\n{:<6} {:<20} {:<15}".format('Item', 'Corrected Item-Total', "Cronbach's α if"))Everything else runs on 3.10+ unchanged.
Stated plainly, because they affect how the results should be read:
- PLS-SEM is implemented manually — standardised OLS regression per structural equation, not SmartPLS or a dedicated PLS library. Significance is assessed analytically rather than by bootstrapping. Path coefficients are reliable; confidence intervals for indirect effects would need a bootstrap of 5,000+ resamples.
- AVE falls below the 0.50 threshold for SI (0.447) and FC (0.467). Convergent validity for these two constructs is marginal and their coefficients deserve more caution than the rest.
- Use Behaviour is weakly measured. Two self-reported items, α = 0.666, R² = 0.037. The behavioural half of the model is the weakest part of the study: intention is well explained, actual use is not.
- The cluster solution is exploratory. k = 4 was chosen for interpretability as much as fit — the silhouette optimum is k = 2, and at k = 4 the score is 0.271. All 12 constructs differ across the four groups at p < 0.001, but they are soft regions rather than sharply separated segments. See Two clustering solutions below.
- Cross-sectional and self-selected. No causal claims are made or supported.
This repository contains two k-means runs on the same data, in two folders. This is deliberate, and worth understanding before citing either.
Thesis version/ |
Cluster Analysis/ |
|
|---|---|---|
| Preprocessing | raw 1–5 construct means | z-scored constructs |
| Risk-Aware Skeptics | n = 67 (13.0%) | n = 243 (47.1%) |
| Disengaged Users | n = 81 (15.7%) | n = 22 (4.3%) |
| Pragmatic Adopters | n = 192 (37.2%) | n = 171 (33.1%) |
| Confident Enthusiasts | n = 176 (34.1%) | n = 80 (15.5%) |
| Silhouette (k = 4) | 0.271 | 0.228 |
| Reported in the dissertation | yes (Ch. 4.3) | no |
The Thesis version/ run is the one reported in the dissertation. The two scripts are
identical apart from a single step: whether the twelve construct scores are z-scored before
k-means.
The constructs do not have equal variance on the 1–5 scale. Regulatory & Compliance Risks (SD = 1.25) and Trust in Technology (SD = 1.19) spread respondents out considerably more than Performance Expectancy (SD = 0.83) or Facilitating Conditions (SD = 0.85) do. Because k-means minimises squared Euclidean distance, it implicitly weights each dimension by its variance — so in the raw space, trust and regulatory concern drive the partition.
Z-scoring removes that weighting and forces every construct to count equally. The effect is measurable in the results: Trust in Technology goes from being the strongest separator between clusters (a 3.06-point spread from 1.40 to 4.47 on the 1–5 scale) to the weakest (2.13). Once trust stops differentiating, the solution collapses into a general high / medium / low agreement gradient, and the four segments are distinguished mainly by how positively respondents answered overall.
Both are legitimate preprocessing choices, and neither is a mistake. Which one is appropriate depends on the question. This study asks whether GameFi users divide into groups with qualitatively different adoption logics — so the informative solution is the one that lets naturally high-variance constructs separate people, rather than the one that flattens them into a single satisfaction dimension.
Its distinguishing segment is Risk-Aware Skeptics (n = 67): Trust in Technology = 1.40, Risk Perception = 4.73, Behavioural Intention = 4.83. These are people with almost no trust in the technology, high awareness of its risks, and the highest intention to keep using it in the entire sample.
That group is the empirical anchor for the study's central structural finding — the negative TT → BI path and the significant Trust × Risk interaction reported in Chapter 4.4. In the z-scored solution this segment does not appear at all; its trust signal is averaged away, and no cluster has a trust score below 1.52 or above 3.64. The raw-space solution also separates better on silhouette (0.271 vs 0.228).
cd "Users/Velze/MSC UTAUT2/Thesis version"
python "cluster analysis (thesis version).py" # the dissertation's solution
cd "../Cluster Analysis"
python "cluster analysis.py" # the z-scored variant- Cluster sizes drift slightly. The dissertation reports 67 / 81 / 182 / 186; re-running the script today gives 67 / 81 / 192 / 176. The two distinctive segments reproduce exactly, in size and in all twelve construct means to two decimal places. The two larger, more similar groups exchange about six cases — expected behaviour for k-means on weakly separated clusters, and sensitive to the scikit-learn version.
- The F-statistics in the dissertation's Σχήμα 18 do not reproduce. The table reports F values
between 38.67 and 72.45; this script produces 92–447. Every construct is significant at
p < 0.001 in both, so the substantive conclusion — that the four groups differ on all twelve
constructs — holds either way, but the exact F values in that table should not be relied on.
Use
Thesis version/cluster_analysis_results.xlsx(sheetANOVA_Results) for the figures this code actually produces.
You are welcome to use this research. Reuse the dataset, the scripts, the figures or the findings for your own work — academic, commercial or otherwise. No permission needed. Attribution is appreciated:
Dikos, G. (2025). GameFi: Opportunities, Challenges and Prospects. MSc dissertation, University of Western Macedonia, Department of Management Science and Technology.
The full dissertation (125 pages, Greek with English abstract) is in this repository: GameFi - Opportunities, Challenges and Prospects.pdf
If you have questions about the methodology, or would like to discuss the findings, reach out on LinkedIn:
→ linkedin.com/in/george-dikos-6b371a287
I would be glad to hear from researchers building on this, and from anyone who finds something in the data I missed.
Released under Creative Commons Attribution 4.0 International (CC BY 4.0) - free to share and adapt, including commercially, with attribution.