What Shapes FinTech Impact in the Digital Banking Era? Evidence from Kosovo, Albania, and North Macedonia

This section reviews the theoretical and empirical literature related to FinTech adoption and PFI in banking. Broader sustainability themes are discussed as contextual literature where relevant, but they are not treated as an empirically measured construct in this study. The literature review follows a theory-driven structure, moving from the broader role of FinTech in emerging banking systems to the specific mechanisms examined in this study: DM, BTF, FL, and interaction effects. For clarity, the abbreviations used throughout the study are listed in Appendix A.

2.1 Theoretical foundations of FinTech adoption and impact

The theoretical foundation of this study is mainly based on the TAM and the Unified Theory of Acceptance and Use of Technology (UTAUT). TAM explains technology adoption through two key factors: perceived usefulness and perceived ease of use, suggesting that users are more likely to accept a technology when they believe it improves their performance and is easy to use. Similarly, UTAUT explains technology acceptance and usage through factors such as performance expectancy, effort expectancy, social influence, and facilitating conditions [6]. This logic is relevant to FinTech because digital financial services require users to evaluate both the practical benefits and the ease of using digital platforms.

UTAUT extends this perspective by emphasizing performance expectancy, effort expectancy, social influence, and facilitating conditions as key determinants of technology use [7]. Later extensions of TAM also show that adoption decisions may be shaped by social influence, cognitive evaluations, and perceptions of system quality or usefulness [8]. These perspectives justify the inclusion of DM, BTF, and FL in this study.

In the context of FinTech, TAM needs to be extended with behavioral and knowledge-based factors. Users’ adoption decisions are rarely based only on technical usability; instead, they are also influenced by trust, perceived security, perceived risk, digital confidence, social influence, and FL. These factors are particularly important in digital finance because users interact with financial services through digital platforms and often share sensitive personal and financial information [9-12].

2.2 FinTech adoption and Perceived FinTech Impact

FinTech adoption has become an important dimension of banking transformation, especially in emerging and transition economies. FinTech can reduce transaction costs, expand access to financial services, increase payment efficiency, and support financial inclusion. Research on digital finance shows that FinTech changes the organization of financial services by combining new business models, digital platforms, and technological infrastructures [1, 2].

FinTech can support financial inclusion and sustainable development when it is supported by adequate infrastructure, regulation, and institutional safeguards [4, 5]. However, the impact of FinTech should not be treated as automatically positive, because digital financial innovation may also generate risks related to financial stability, consumer protection, cybersecurity, and unequal access to digital capabilities [4].

This study therefore analyzes FinTech beyond simple adoption. Adoption indicates whether users start using digital financial services, while impact reflects whether these services generate meaningful practical and financial benefits. PFI is treated as a broader dependent construct that includes time saving, transaction efficiency, increased use of digital services, better monitoring of monthly expenses, improved personal financial management, and willingness to recommend FinTech services [1, 2].

2.3 Digital Marketing and Perceived FinTech Impact

DM plays an important role in shaping users’ perceptions of FinTech services. In digital finance, marketing does not function only as a promotional tool, but also as an informational mechanism that reduces uncertainty, increases awareness, explains service benefits, and builds familiarity with digital platforms. Social media advertisements, credible online information, application design, online promotions, and continuous digital communication may influence how users perceive the usefulness, ease of use, and reliability of FinTech services. This argument is consistent with technology acceptance theory, which suggests that perceived usefulness and ease of use are central to technology adoption [6, 7].

DM can also strengthen users’ awareness and engagement by communicating the practical benefits of digital financial services, such as time saving, transaction efficiency, easier access to financial tools, and improved financial management. Prior studies on digital finance show that new financial technologies reshape service delivery and customer interaction, making digital communication and platform-based engagement increasingly important for financial service providers [1, 2].

In the context of FinTech, DM may increase perceived value by making users more familiar with service features, promotional benefits, security messages, and platform usability. However, marketing alone may not be sufficient if users do not trust digital providers or lack the FL required to interpret financial information. Therefore, DM is hypothesized to have a positive association with PFI, while the magnitude of this association may vary with users’ behavioral readiness and FL [10-12].

2.4 Behavioral and Trust Factors

The behavioral dimension is central to FinTech adoption because digital financial services involve payments, personal data, savings, banking information, and financial decision-making. Users’ trust, perceived risk, security perceptions, attitudes toward innovation, and social influence play a key role in shaping adoption and impact. Mobile payment research shows that perceived security, performance expectations, innovativeness, and social influence are important determinants of adoption and recommendation intentions [11].

Trust and perceived security are particularly important in digital financial contexts because users must believe that platforms can protect their personal and financial data. Zhao and Bacao [12] showed that perceived security, trust, perceived benefits, and social influence jointly shape mobile payment adoption intentions. Similarly, Appiah and Agblewornu [13] showed that perceived benefits, perceived risks, and trust are important mechanisms explaining FinTech adoption.

In this study, BTF capture perceived security and trust in FinTech platforms, attitudes toward FinTech innovation, preference for digital financial services, and the influence of family and friends. Stronger values on these dimensions are expected to be positively associated with PFI [11-13].

2.5 Financial Literacy and Perceived FinTech Impact

FL is an important cognitive capability that enables users to understand and benefit from FinTech services. Financially literate users are more capable of evaluating risks, comparing financial alternatives, understanding costs and benefits, and making informed decisions. FL has been conceptualized as a form of human capital that supports better financial decision-making and welfare outcomes [9].

In modern financial markets, FL has become increasingly important because individuals face more complex financial choices, many of which are mediated by digital platforms and financial technologies [10]. In the FinTech context, this means that users need not only access to digital tools but also the knowledge and confidence to use them responsibly.

Recent empirical evidence also supports the role of literacy-related capabilities in FinTech adoption. Islam and Khan [14] showed that FL, digital literacy, and financial self-efficacy positively influence FinTech adoption. Sumartini et al. [15] further connected digital FL and self-efficacy with intention to use digital banking through a TAM-based approach. Therefore, FL is hypothesized to have a positive association with PFI because financially literate users may be better positioned to interpret and use digital financial services.

2.6 Interaction effects between Digital Marketing and Financial Literacy

The interaction between DM and FL is important because DM may expose users to FinTech services, but users need FL to interpret and evaluate the information they receive. Without sufficient FL, online promotions and digital communication may increase awareness but may not necessarily lead to meaningful financial outcomes. Financially literate users are more capable of assessing the credibility of digital financial information, comparing alternatives, understanding potential costs and risks, and deciding whether a FinTech service fits their financial needs [9, 10].

This moderating logic is supported by recent studies showing that FL, digital literacy, and financial self-efficacy have positive effects on FinTech adoption and digital banking use [14, 15]. These findings suggest that FL can strengthen the relationship between DM and PFI because users with higher FL are better able to transform digital information into informed financial behavior.

The study also examines a three-way interaction among BTF, FL, and DM. This interaction is based on the argument that PFI is not shaped by isolated factors only, but by the combined influence of digital information, behavioral readiness, and financial capability. DM creates awareness and communicates value; BTF determine whether users feel secure and willing to adopt FinTech services; and FL enables users to evaluate and use these services responsibly [11-13].

The three-way interaction is expected to capture the conditional and potentially non-linear nature of PFI. For example, marketing may be more effective when users trust digital financial platforms and possess sufficient FL, while high FL may also make some users more selective and less influenced by promotional communication. Therefore, the interaction model provides a more nuanced explanation of PFI than a direct-effects model alone [6, 7, 10].

2.7 Research gap and observed-composite analytical framework

Although the literature on FinTech adoption has expanded rapidly, several gaps remain. First, many studies focus on larger or more digitally advanced economies, while evidence from smaller emerging European economies remains limited. Second, studies on the Western Balkans frequently analyze FinTech through macro-level indicators such as digital payment penetration, mobile banking usage, internet access, regulatory quality, FL, and trust in digital finance. While such studies are useful, they do not fully explain how individual users perceive, evaluate, and experience FinTech services [3].

Third, existing research often examines adoption determinants separately, focusing on either technology acceptance, trust, perceived risk, or FL. However, FinTech adoption and impact may reflect the combined roles of digital communication, behavioral readiness, and financial capability. Fourth, limited attention has been given to PFI as a broader outcome that goes beyond intention or usage [1, 2, 4].

To address these gaps, this study specifies an observed-composite analytical framework for Kosovo, Albania, and North Macedonia. DM, BTF, and FL are represented by arithmetic composite scores and are examined in relation to the PFI composite through adjusted main terms and precomputed interaction terms. Country, age, and income are included as observed control variables. The framework is conceptual only in the sense of organizing the regression variables and hypotheses; it does not represent a latent-variable measurement or structural model [6, 7, 10].

image003.png

Figure 1 summarizes the study variables as observed arithmetic composites rather than latent constructs. DM is represented by DM_R (Q1-Q5), and H1 concerns its adjusted association with PFI (PFI_R). The graphical arrow is therefore a schematic representation of the regression association tested in OLS, not a latent structural path and not a causal effect. The theoretical rationale remains grounded in technology-acceptance and digital-marketing literature [1, 2].

BTF form the second explanatory dimension of the model. These factors include risk perception, trust in digital financial platforms, perceived security, attitudes toward FinTech, and social influence. Their inclusion is justified by behavioral finance and trust theory, which suggest that financial decisions are shaped by risk perceptions, uncertainty, psychological factors, and trust in digital environments [3, 4]. Accordingly, users who perceive FinTech platforms as secure, reliable, and socially accepted are expected to experience stronger PFI, representing H2. FL represents the third independent variable. It reflects users’ FL, ability to evaluate risk, and capacity to make informed financial decisions. Since FinTech services require users not only to access digital tools but also to understand financial information and assess potential risks, FL is expected to support more effective and responsible FinTech use [5, 6]. This relationship represents H3.

At the center of the framework is PFI, the dependent observed composite of the study. It is calculated directly from Q15-Q20 and captures use-related benefits, perceived financial-management outcomes, and recommendation/advocacy. The six items are not interpreted as two validated latent dimensions of Usage and Financial Performance. In particular, Q20 (willingness to recommend FinTech) is treated as an advocacy/recommendation indicator rather than as Financial Performance. This observed-composite interpretation is consistent with the focused factor assessment reported in the Results section and avoids imposing a higher-order latent structure that was not estimated [7, 9]. FinTech may support sustainable financial inclusion when regulation, institutional readiness, and digital infrastructure are adequate [5], while FL remains important because knowledge gaps and overconfidence may weaken financial decision-making outcomes [16]. The framework also includes interaction effects. H4 proposes that FL strengthens the relationship between DM and PFI because financially literate users may be better able to interpret and evaluate digital financial information. H5 evaluates the three-way interaction among BTF, FL, and DM. Pearson correlations among the observed composites are reported separately as preliminary bivariate associations and are not treated as an additional formal hypothesis.

Finally, country, age, and income are included as observed control variables because FinTech adoption and impact may vary across socio-demographic and national contexts. Kosovo, Albania, and North Macedonia share a Western Balkan setting but differ in banking digitalization, income levels, digital readiness, FL, and institutional trust. The three-country design broadens regional coverage beyond a single-country study; however, the country-specific regressions are treated as supplementary sensitivity analyses rather than as evidence of statistically established cross-country coefficient differences.

2.8 Theoretical integration and empirical justification of the research model

The synthesis of the reviewed studies provides a concise basis for linking the literature review with the development of the research model. The studies summarized in Table 1 indicate that the literature supporting this research model can be grouped into three main analytical streams. The first stream focuses on technology acceptance and digital finance, showing that FinTech adoption is influenced by perceived usefulness, ease of use, performance expectations, digital innovation, and the transformation of financial services. The second stream emphasizes behavioral and trust-related factors, particularly perceived risk, perceived benefit, security, and user confidence, which explain why individuals may accept or resist digital financial services. The third stream highlights the role of FL and institutional readiness, suggesting that users’ FL, digital capability, and the broader regulatory environment are essential for achieving meaningful FinTech outcomes.

Table 1. Summary of the literature review

A particularly important contribution comes from the technology acceptance literature, which explains FinTech adoption through perceived usefulness, ease of use, performance expectations, and facilitating conditions [6-8], while the FL literature shows that users’ knowledge and capability influence the quality of digital financial decision-making [9, 10].

The literature portrays PFI as a multidimensional phenomenon in which technological innovation interacts with behavioral readiness, financial capability, and institutional context. Foundational studies show that FinTech and digital finance alter financial-service delivery, business models, and market structures [1, 2]. Yet the technology-acceptance literature indicates that innovation alone is insufficient: perceived usefulness and ease of use [6], together with performance expectancy, effort expectancy, social influence, and facilitating conditions [7], shape users’ willingness to engage with new technologies. TAM3 further demonstrates that these perceptions are themselves influenced by broader cognitive and contextual determinants [8]. This theoretical progression supports treating users’ perceptions of PFI as the outcome of both technological exposure and individual evaluation rather than as a purely technical consequence.

Empirical evidence reinforces this interpretation. In mobile-payment settings, usefulness, trust, security, and compatibility are closely linked to adoption [11], while the pandemic context increased the salience of convenience and safety perceptions [12]. More recent FinTech evidence similarly identifies perceived benefits, perceived risks, and trust as central behavioral mechanisms [13]. These findings provide a strong rationale for incorporating BTF alongside DM: digital exposure may increase awareness and engagement, but its relationship with PFI is likely to depend on whether users consider digital financial services credible, beneficial, and sufficiently safe.

Financial capability adds a further explanatory layer. Lusardi and Mitchell [9, 10] established FL as a key resource for informed financial decision-making, particularly as financial markets become more complex and digitalized. FinTech-specific studies extend this logic by showing that FL, digital literacy, and financial self-efficacy are associated with stronger adoption [14], while digital FL and self-efficacy reinforce intentions to use digital banking [15]. Conversely, knowledge gaps and overconfidence may weaken decision quality [16]. FL can therefore be viewed not only as an independent correlate of PFI, but also as a plausible moderator that conditions how effectively individuals interpret and respond to digital financial information.

Finally, the institutional environment places these individual-level mechanisms within a broader system. Evidence from the Balkans emphasizes the importance of payments infrastructure, regulation, literacy, and trust [3], while research on financial inclusion and sustainability shows that FinTech benefits depend on appropriate safeguards, regulatory quality, and technological readiness [4, 5]. Taken together, the literature supports an integrated empirical model in which DM, BTF, and FL are examined jointly, with interaction terms capturing whether their associations with PFI are conditional rather than merely additive. This framework is theoretically coherent because it links technology acceptance, behavioral trust, financial capability, and institutional context within a single explanatory structure while avoiding unsupported causal claims.

This section presents the empirical results. The analysis was conducted in SPSS 27 and begins with descriptive statistics and sample characteristics. Internal consistency was assessed with Cronbach's alpha, and exploratory item structure was examined using PAF with Promax rotation. Formal hypotheses were evaluated with the four-block hierarchical OLS model, followed by interaction probing, diagnostics, user-status sensitivity analysis, and supplementary country-specific models.

4.1 Descriptive statistics

This section presents the empirical results. The analysis was conducted in SPSS 27 and begins with descriptive statistics and sample characteristics. Internal consistency was assessed with Cronbach's alpha, and exploratory item structure was examined using PAF with Promax rotation. Formal hypotheses were evaluated with the four-block hierarchical OLS model, followed by interaction probing, diagnostics, user-status sensitivity analysis, and supplementary countryspecific models. The descriptive statistics of the study variables are presented in Table 4 and Appendix C.

Table 4. Descriptive statistics of study variables

The dependent variable, PFI, is an observed composite based on Q15-Q20. Its mean is M = 3.85 (SD = 0.87). A focused factor assessment of the six items is reported below; the measure is not treated as a CFA-validated latent or higher-order construct.

The DM observed composite (Q1-Q5) reported M = 3.50 (SD = 0.89). The BTF (Q6 and Q8-Q10) reported M = 3.64 (SD = 0.92). Q7R is retained separately as a perceived-risk indicator because its inclusion weakened internal consistency and it showed low communality in EFA.

The FL composite (Q11-Q13) reported M = 3.81 (SD = 0.93). Q14 is retained separately as training exposure because it did not cluster with Q11-Q13 and its exclusion improved reliability. These descriptive results correspond to the observed-composite specification used in the regression analyses.

4.2 Demographic analysis

The demographic analysis describes the realized sample according to country, age, gender, education, and income. Because the sample was not selected to represent national populations, these characteristics are descriptive of the respondents included in the study.

image029.png

Figure 3. Sample distribution across Kosovo, Albania, and North Macedonia

The final analytical sample consisted of 262 respondents: Kosovo n = 163 (62.2%), Albania n = 55 (21.0%), and North Macedonia n = 44 (16.8%). The distribution is substantially unbalanced, with Kosovo accounting for nearly two-thirds of the sample and North Macedonia representing the smallest subgroup (Figure 3). These proportions describe the realized analytical sample only; they are not population weights. The country subsamples were not designed to provide nationally representative estimates, so country-level descriptive and regression results are interpreted as evidence from the surveyed respondents rather than as estimates for the national populations.

Age was available for 260 respondents. The largest group was 21-30 years (n = 95, 36.5%), followed by 31-40 years (n = 55, 21.2%), 41-50 years (n = 43, 16.5%), over 50 years (n = 36, 13.8%), and under 20 years (n = 31, 11.9%); two cases had missing age data (Figure 4).

image030.png

Figure 4. Age distribution

Female respondents represented 62.2% of the analytical sample (n = 163), while male respondents represented 37.8% (n = 99) (Figure 5). Bachelor's and Master's degrees were the predominant education categories in the questionnaire responses; education is reported descriptively and was not included as a control in the final hierarchical model.

image031.png

Figure 5. Gender distribution

Monthly income was available for 257 respondents. The largest group reported €300-€700 per month (n = 104, 40.5%), followed by €700-€1,000 (n = 67, 26.1%), above €1,000 (n = 54, 21.0%), and below €300 (n = 32, 12.5%); five cases had missing income data (Figure 6).

image032.png

Figure 6. Monthly income distribution

Among 231 respondents with valid app-specific selections, Paysera was selected by 62 respondents (26.8%), PayPal by 60 (26.0%), and Wise (TransferWise) by 51 (22.1%). Forty-eight respondents (20.8%) selected 'Bank-based FinTech only.' This category indicates use of bank-provided digital financial applications and is therefore not treated as FinTech non-use (Figure 7). Among the same 231 respondents with valid app-specific selections, Raiffeisen ON was selected by 92 respondents (39.8%), ProCredit Mobile by 65 (28.1%), TEB Mobile by 62 (26.8%), NLB Klik by 53 (22.9%), and BKT Mobile by 50 (21.6%). Five respondents (2.2%) selected 'Independent FinTech only,' indicating independent FinTech use without a bank-based application selection.

image033.png

Figure 7. Independent FinTech application selections

These are multiple-response application categories and are not used alone to define overall FinTech use status (Figure 8).

image034.png

Figure 8. Bank-based FinTech application selections

4.3 Measurement and data-quality checks

This section reports internal-consistency reliability, the FinTech use-status audit, exploratory factor evidence, the focused assessment of PFI, and the common-method diagnostic used before formal hypothesis interpretation. The measurement analyses are exploratory and are not presented as CFA-level construct validation.

4.3.1 Reliability analysis of study constructs

Internal consistency of the observed measures was assessed using Cronbach's alpha; results are presented in Table 5.

DM (Q1-Q5) showed good internal consistency (Cronbach's α = 0.829). The original five-item BTF specification including Q7R produced α = 0.759; excluding Q7R increased reliability to α = 0.796 for Q6 and Q8-Q10. The original Q11-Q14 FL specification produced α = 0.780; excluding Q14 increased reliability to α = 0.817 for Q11-Q13. The six-item PFI block (Q15-Q20) showed high internal consistency (α = 0.890). The exploratory three-item blocks were also internally consistent (Q15-Q17: α = 0.837; Q18-Q20: α = 0.836), but reliability alone does not establish that the two blocks are discriminable constructs.

Table 5. Reliability assessment of observed measures

The reliability results support use of the main composite scores as internally consistent observed measures, while Q7R and Q14 remain separate indicators because they weakened or did not align with their original composite specifications.

Cronbach's alpha is interpreted here as evidence of internal consistency only. It is not treated as proof of convergent, discriminant, or latent-construct validity.

4.3.2 FinTech use-status audit

Because Q15-Q20 concern use-related benefits, financial-management outcomes, and recommendation/advocacy, FinTech use status was audited across both independent and bank-based application responses before sensitivity analysis.

The audit identified 247 respondents (94.3%) with recorded FinTech use and 15 respondents (5.7%) with no recorded FinTech use (Table 6). The 15 cases are not assumed to provide experience-based evaluations of Q15-Q20; their responses may instead be hypothetical or anticipated.

Table 6. FinTech use-status audit in the analytical sample

To assess whether these cases affected inference, the hierarchical model was estimated again after excluding respondents without recorded FinTech use. The user-status sensitivity analysis is reported in Section 4.4.3 and preserves the substantive hypothesis pattern.

4.3.3 Exploratory factor analysis and measurement assessment

EFA was conducted on the 20 questionnaire items using PAF and Promax oblique rotation. The analysis evaluated sampling adequacy, factor retention, communalities, pattern coefficients, cross-loadings, and factor correlations. It is used as exploratory measurement evidence rather than as confirmatory latent-construct validation.

Table 7. Kaiser-Meyer-Olkin (KMO) and Bartlett’s Test

Table 8. Initial eigenvalues and Principal Axis Factoring (PAF) extraction variance

The KMO value was 0.892 (Table 7), indicating very good sampling adequacy. Bartlett's Test of Sphericity was statistically significant, χ²(190) = 2603.962, p < 0.001 (Table 7), confirming that the item correlation matrix was suitable for factor analysis. These results support proceeding with EFA but do not, by themselves, establish construct validity.

Four factors had initial eigenvalues greater than 1 (7.582, 2.265, 1.460, and 1.167). Under PAF, the four retained factors accounted for 53.319% of the extracted variance (Table 8). The scree plot also showed a clear leveling after the fourth factor. Because the factors were correlated, Promax rotation was used rather than an orthogonal Varimax rotation.

As shown in Table 9, the Pattern Matrix shows that the first factor is dominated by Q15-Q20, with coefficients from 0.605 to 0.911. The second factor contains Q1-Q6, with coefficients from 0.420 to 0.866; Q4 (0.529 on Factor 2; 0.328 on Factor 4) and Q5 (0.420 on Factor 2; 0.380 on Factor 4) show cross-loadings. The third factor is defined clearly by Q11-Q13 (0.656-0.853). The fourth factor contains Q7R, Q8-Q10, and Q14 (0.335-0.668). This pattern demonstrates that the empirical item structure does not reproduce the original five theory-imposed dimensions cleanly.

The factor correlations ranged from 0.460 to 0.623, supporting an oblique rotation. Q6 (perceived security/trust) clustered with DM items, while Q14 (training exposure) did not cluster with the core FL items; Q7R also showed weak communality (0.256). These results support a conservative observed-composite specification and caution against treating the theory-imposed dimensions as EFA-validated latent constructs.

Table 9. Pattern matrix from Principal Axis Factoring (PAF) with Promax rotation

Table 10. Summary of exploratory measurement decision criteria

A forced five-factor sensitivity analysis was also estimated to reconcile the empirical solution with the original theoretical model. The fifth initial eigenvalue was 0.881, below 1, and the additional factor was defined mainly by Q4 and Q5 rather than by a theoretically distinct Usage or Performance dimension. The forced solution therefore did not provide convincing support for the original five-factor measurement structure.

Overall, the measurement evidence is interpreted conservatively (Table 10). The four-factor EFA describes how the items cluster in this sample, while the theory-guided composites used in OLS are treated as observed arithmetic scores. EFA is not treated as confirmatory construct validation, and CFA-level convergent or discriminant validity is not claimed.

The scree plot (Figure 9) and the eigenvalue pattern therefore support a four-factor exploratory solution for the present sample. This solution is not presented as a confirmatory latent measurement model; rather, it identifies the principal areas of agreement and mismatch between the empirical item structure and the original theoretical grouping.

image035.png

Because no CFA/SEM measurement model is estimated, construct reliability, AVE, and discriminant-validity statistics are not presented as evidence of latent validity. Such confirmatory indices require a prespecified latent measurement model and are appropriate for future validation on an independent or enlarged sample. The present study limits its measurement claims to internal consistency and exploratory factor evidence for observed composites.

4.3.4 Focused dimensional assessment of Perceived FinTech Impact

To assess whether Q15-Q20 should be represented as separate FinTech Usage and Financial Performance dimensions, the six items were examined using focused PAF. Sampling adequacy was good (KMO = 0.862), and Bartlett's Test of Sphericity was significant, $\chi^2(15)=854.047$, p < 0.001. The first initial eigenvalue was 3.884 and the second was 0.716, indicating one dominant factor.

A one-factor PAF solution accounted for 57.847% of the variance, with loadings from 0.675 to 0.814, and the six-item scale showed high internal consistency (Cronbach's α = 0.890). Although Q15-Q17 and Q18-Q20 were separately reliable (α = 0.837 and 0.836), a forced two-factor extraction did not converge within 25 iterations, the rotated factors correlated strongly (r = 0.727), and Q17 cross-loaded (0.487 and 0.334). The evidence therefore does not support interpreting Usage and Financial Performance as two distinct first-order constructs. Q15-Q20 are retained as one PFI observed composite, and Q20 is classified as recommendation/advocacy rather than Financial Performance.

4.3.5 Common-method variance diagnostic

All 20 questionnaire items were entered into an unrotated one-factor PAF analysis. The first factor had an initial eigenvalue of 7.582 and represented 37.909% of the total initial variance. When a single factor was forced, the extraction sum of squared loadings was 6.952, accounting for 34.759% of the extracted variance; the one-factor loadings ranged from 0.303 to 0.712.

No single general factor therefore accounted for the majority of item variance. This result provides no evidence of a dominant single common-method factor, but it does not rule out common-method variance. Accordingly, the OLS coefficients and interaction estimates are interpreted as contemporaneous associations rather than causal effects.

4.4 Hypothesis testing using four-block hierarchical Ordinary Least Squares regression

The pooled hypotheses were tested using one four-block hierarchical OLS specification. PFI (PFI_R) was the dependent variable. Block 1 entered demographic controls; Block 2 added standardized DM (ZDM_R), BTF (ZBTF_R), and FL (ZFL_R); Block 3 added the three two-way interactions (DM × BTF, DM × FL, and BTF × FL); and Block 4 added the DM × BTF × FL three-way interaction. The regression used N = 255 complete cases (Table 11).

Table 11. Four-block hierarchical Ordinary Least Squares (OLS) model summary

The demographic controls were dummy-coded rather than entered as linear 1-2-3 category scores. Kosovo served as the reference country, while the first age and income categories served as their reference groups. This specification makes the control block explicit and aligns the reported model sequence with the methodology.

Block 1, containing only demographic controls, explained 1.2% of the variance in PFI (R² = 0.012; adjusted R² = -0.024) and was not statistically significant, F-change(9, 245) = 0.339, p = 0.961. Adding the three main predictors in Block 2 produced a substantial improvement: R² increased to 0.311 (adjusted R² = 0.277), with ΔR² = 0.299, F-change(3, 242) = 35.034, p < 0.001.

Block 3 added the three lower-order two-way interactions and increased R² from 0.311 to 0.377 (adjusted R² = 0.338). The additional 6.5% explained variance was statistically significant, ΔR² = 0.065, F-change(3, 239) = 8.348, p < 0.001. Block 4 then added the three-way interaction; the final R² was 0.385 (adjusted R² = 0.343), but the additional 0.8% was not statistically significant, ΔR² = 0.008, F-change(1, 238) = 3.033, p = 0.083. The complete final model was significant, F(16, 238) = 9.294, p < 0.001, with Durbin-Watson = 2.075.

Table 12 reports the coefficients of the final four-block model, including the demographic controls and all focal terms.

Table 12. Final four-block hierarchical regression coefficients

Final hierarchical regression equation:

$\begin{gathered}P F I_R=3.957+\text {controls}+0.109 Z D M_R+0.371 Z B T F_R+0.243 Z F L_R-0.011(D M \times B T F) \\ +0.168(D M \times F L)-0.052(B T F \times F L)-0.048(D M \times B T F \times F L)+\varepsilon\end{gathered}$   (19)

In the final model, BTF were the strongest positive predictor (B = 0.371, β = 0.438, p < 0.001), followed by FL (B = 0.243, β = 0.287, p < 0.001). DM remained positive but was not statistically significant at the 0.05 level (B = 0.109, β = 0.127, p = 0.076). None of the demographic controls reached statistical significance in the final model.

Interaction results:

Among the two-way interactions, only DM × FL was statistically significant (B = 0.168, β = 0.266, p = 0.004). DM × BTF was not significant (β = -0.017, p = 0.833), and BTF × FL was also not significant (β = -0.083, p = 0.421).

The three-way interaction was negative but did not reach the 0.05 significance threshold (B = -0.048, β = -0.176, p = 0.083; 95% CI for B = [-0.102, 0.006]). Its addition in Block 4 likewise did not significantly improve model fit (ΔR² = 0.008, p = 0.083). Accordingly, the three-way term is not treated as evidence of a statistically established combined moderation effect.

The block sequence shows that demographic controls contribute little explanatory power, while the three main observed composites account for the largest incremental increase in R². The two-way interaction block provides a further statistically significant increment, driven principally by the positive DM × FL term.

The hierarchical sequence is therefore interpreted in the following order: demographic controls, main observed-composite associations, all two-way interactions, and the three-way interaction.

The final model therefore explains 38.5% of the variance in PFI, but the evidence does not support attributing additional explanatory power to the three-way interaction after the lower-order terms have been entered.

Interpretation is based on the fully specified hierarchical model, in which controls, main effects, lower-order interactions, and the three-way interaction are evaluated within one coherent specification.

For regression-based hypothesis testing (Table 13), H2 and H3 are supported because BTF and FL are positive and statistically significant in the final model. H4 is supported because DM × FL is positive and significant. H1 is not supported at the conventional 0.05 level in the final model because the direct DM coefficient is positive but non-significant (p = 0.076). H5 is not supported because the three-way interaction is non-significant (p = 0.083).

Table 13. Hypothesis decisions from the four-block hierarchical Ordinary Least Squares (OLS) model

The final hypothesis decisions therefore distinguish significant direct and lower-order conditional associations from the unsupported three-way hypothesis. BTF and FL retain significant positive associations with PFI, and the significant DM × FL term supports the proposed moderating role of FL.

DM does not show an independently significant direct association after adjustment for controls, the other main effects, and interaction terms. The three-way term is also not statistically significant; H1 and H5 are therefore not supported in the pooled hierarchical model.

Substantively, the strongest evidence in the pooled model concerns BTF, FL, and the DM × FL interaction. These results are interpreted as associations within a cross-sectional observed-composite OLS model.

Because the three-way interaction is not statistically significant, no substantive mechanism such as information overload, selectivity, saturation, or diminishing returns is inferred from its negative sign.

Overall, the hierarchical model provides a coherent pooled specification. The final model explains 38.5% of the variance in PFI (adjusted R² = 0.343), with the largest explanatory gain occurring when the three main observed composites are added and a smaller but significant gain when the lower-order two-way interactions are added.

Figure 10 presents standardized coefficients for the focal terms in the hierarchical model. BTF shows the largest positive standardized coefficient (β = 0.438), followed by FL (β = 0.287) and the DM × FL interaction (β = 0.266). DM is positive but non-significant (β = 0.127, p = 0.076), while the three-way interaction is negative but non-significant (β = -0.176, p = 0.083).

image040.png

Figure 10. Standardized coefficients from the four-block hierarchical Ordinary Least Squares (OLS) model

4.4.1 Conditional simple-slope analysis of the interaction between Digital Marketing and Financial Literacy

The analysis first showed that the three-way interaction was not statistically significant (B = -0.048, SE = 0.028, β = -0.176, t = -1.742, p = 0.083, 95% CI [-0.102, 0.006]). In simple terms, this means that the combined effect of the three variables was not strong enough to be considered statistically reliable. The p-value of 0.083 is above the commonly used 0.05 threshold, while the CI includes zero. For this reason, the analysis then focused on the interaction between DM and FL, which was statistically significant (B = 0.168, SE = 0.058, β = 0.266, t = 2.899, p = 0.004, 95% CI [0.054, 0.282]).

This result shows that DM does not affect all FinTech users in the same way. Its effect changes depending on the level of FL. The positive coefficient B = 0.168 indicates that the higher the level of FL, the more positive the effect of DM on PFI becomes. The p-value of 0.004, which is much lower than 0.05, shows that this interaction is statistically reliable.

The simple-slope analysis makes this relationship even clearer (Figure 11). Among users with low FL (-1 SD), the effect of DM was slightly negative but not statistically significant (B = -0.059, SE = 0.079, t = -0.746, p = 0.456, 95% CI [-0.215, 0.097]). In practical terms, this means that among users with lower FL, greater DM was not clearly associated with either a more positive or a more negative PFI. The p-value of 0.456 and the CI including zero show that this effect is highly uncertain.

image041.png

Among users with an average level of FL, the relationship became positive (B = 0.109, SE = 0.061, t = 1.781, p = 0.076, 95% CI [-0.012, 0.229]). This suggests that DM begins to be associated with a more positive PFI when users have an average level of FL. However, p = 0.076 is still above 0.05, and the CI continues to include zero; therefore, this effect is still not considered statistically significant.

The clearest result appears among users with high FL (+1 SD). At this level, DM had a positive and statistically significant effect on PFI (B = 0.277, SE = 0.089, t = 3.113, p = 0.002, 95% CI [0.102, 0.451]). This means that among users with higher FL, an increase in DM is associated with a more positive PFI. The p-value of 0.002 is much lower than 0.05, and the CI is entirely positive, making this result statistically reliable.

Overall, the pattern is clear: when FL is low, the effect of DM is B = -0.059 and is not significant; at the average level, the effect becomes positive at B = 0.109 but is still not statistically significant; whereas at the high level, the effect increases to B = 0.277 and becomes statistically significant. Therefore, the results show that FL strengthens the positive relationship between DM and PFI. In practical terms, DM appears to work better among FinTech users who have stronger FL.

The Johnson-Neyman analysis provided additional information regarding the range of FL values over which the conditional DM effect was statistically significant. The estimated region-of-significance boundaries were approximately -2.28 SD and +0.07 SD relative to the sample mean. Between these two boundaries, the conditional DM slope was not statistically different from zero at the 0.05 level. Above approximately +0.07 SD, the conditional effect became positive and statistically significant, whereas below approximately -2.28 SD, the effect became negative and statistically significant. Because the lower Johnson-Neyman boundary lies within an extreme low-literacy region, the substantive interpretation emphasizes the transition occurring near +0.07 SD.

•Normality of Residuals (Histogram)

The histogram of standardized residuals is approximately centered around zero and provides a visual assessment of the residual distribution (Figure 12). Because heteroskedasticity was detected, residual diagnostics and robust covariance estimates are emphasized rather than relying on normality alone.

image042.png

Figure 12. Histogram of standardized residuals for the Perceived FinTech Impact (PFI) model

4.4.2 Final hierarchical regression model for hypothesis testing

Before interpreting the final hypothesis-testing results, it is necessary to assess whether the hierarchical regression model satisfies the main diagnostic conditions for reliable estimation.

The diagnostic values for the final N = 255 hierarchical model (Table 14) show that the maximum VIF was 4.113 and the maximum Condition Index was 9.143. Standardized residuals ranged from -3.402 to 2.698 and studentized residuals from -3.512 to 2.828. Cook's Distance reached a maximum of 0.168, below the conventional 1.0 screening rule, whereas centered leverage reached 0.443 and Mahalanobis Distance reached 112.599. These values identify several observations that warrant case-level influence review, but no observation was deleted automatically on the basis of a single diagnostic.

Table 14. Regression diagnostic tests for model stability and influence

The standardized residual-versus-standardized-predicted plot (Figure 13) shows residuals distributed around zero without a pronounced curved pattern, supporting approximate linearity. The spread is not perfectly uniform, which is consistent with the formal heteroskedasticity result and motivates robust covariance inference.

image043.png

Figure 13. Standardized residuals versus standardized predicted values

Because heteroskedasticity was detected, the same 16-predictor linear model was re-estimated in GENLIN using a normal identity link and robust sandwich covariance. In simple terms, this additional analysis was carried out to make sure that the statistical conclusions remained reliable even though the variability of the errors was not constant across observations. Importantly, the coefficient estimates did not change, which means that the estimated direction and size of the relationships remained the same. What changed was the way the uncertainty around those estimates was calculated, making the statistical inference more robust.

For DM, the coefficient was positive (B = 0.109), with a robust standard error of 0.0569, a 95% CI from -0.003 to 0.220, and p = 0.056. This suggests that higher levels of DM were associated with a higher PFI. However, the p-value of 0.056 is slightly above the conventional 0.05 threshold, and the CI includes zero. Therefore, although the relationship is positive and close to statistical significance, the evidence is not strong enough to conclude with conventional statistical confidence that DM has a direct independent effect.

A much clearer result was observed for BTF. The coefficient was B = 0.371, with a robust standard error of 0.0709, a 95% CI from 0.233 to 0.510, and p < 0.001. This means that higher levels of behavioral and trust-related factors were associated with higher PFI. The CI is entirely positive, and the very small p-value indicates strong statistical evidence for this relationship. For the audience, the practical message is straightforward: users who show stronger trust and more favorable behavioral attitudes toward FinTech are also more likely to perceive FinTech as having a positive impact.

FL also showed a positive and statistically significant relationship, with B = 0.243, robust SE = 0.0621, 95% CI [0.121, 0.365], and p < 0.001. This indicates that users with higher FL tend to report a more positive PFI. The CI remains fully above zero and the p-value is well below 0.05, providing strong evidence that FL is an important factor in the model. In practical terms, people who better understand financial concepts may be more capable of recognizing and evaluating the benefits offered by FinTech services.

The interaction between DM and BTF (DM × BTF) was very small and not statistically significant (B = -0.011, robust SE = 0.0484, p = 0.819). This means there is no reliable evidence that the effect of DM changes depending on the level of BTF. In other words, trust-related factors do not appear to strengthen or weaken the relationship between DM and PFI in a statistically meaningful way.

In contrast, the interaction between DM × FL was positive and statistically significant (B = 0.168, robust SE = 0.0550, 95% CI [0.060, 0.276], p = 0.002). This is one of the most important findings in the model. The positive coefficient means that the relationship between DM and PFI becomes stronger as FL increases. The p-value of 0.002 is clearly below 0.05, while the CI [0.060, 0.276] is entirely positive. In practical terms, DM appears to be more effective among FinTech users who have greater FL. Users who understand financial concepts more clearly may be better able to interpret DM messages, evaluate the information presented to them, and recognize the potential value of FinTech products and services.

The interaction between BTF × FL was negative but not statistically significant (B = -0.052, robust SE = 0.0684, p = 0.450). Therefore, there is no reliable evidence that FL meaningfully changes the relationship between BTF and PFI. Although the estimated coefficient is slightly negative, the relatively high p-value shows that this pattern could easily be due to random variation in the data.

Finally, the three-way interaction between DM × BTF × FL was also not statistically significant (B = -0.048, robust SE = 0.0313, 95% CI [-0.109, 0.013], p = 0.125). The negative coefficient suggests a small negative estimated three-way effect, but the p-value is above 0.05 and the CI includes zero. This means that there is insufficient evidence to conclude that the combined effect of DM and FL changes further depending on the level of BTF.

Overall, the robust analysis presents a clear pattern. BTF (B = 0.371, robust SE = 0.0709, 95% CI [0.233, 0.510], p < 0.001) and FL (B = 0.243, robust SE = 0.0621, 95% CI [0.121, 0.365], p < 0.001) show strong positive direct relationships with PFI. DM also shows a positive coefficient (B = 0.109, robust SE = 0.0569, 95% CI [-0.003, 0.220], p = 0.056), although its direct effect falls just short of the conventional 0.05 significance level. Most importantly, the significant DM × FL interaction (B = 0.168, robust SE = 0.0550, 95% CI [0.060, 0.276], p = 0.002) shows that the influence of DM becomes more positive when FL is higher. For a non-technical audience, the central message is that FinTech users appear to respond more positively when they both understand financial matters and are exposed to effective DM, while trust and Behavioral Factors independently remain strong contributors to how positively FinTech is perceived.

All demographic controls remained non-significant under robust covariance estimation (smallest p = 0.086). The hypothesis pattern is unchanged: H1 and H5 are unsupported, whereas H2, H3, and H4 are supported.

4.4.3 User-status sensitivity analysis

FinTech use status combines independent and bank-based application responses. Respondents who reported bank-based FinTech use were classified as FinTech users even when they reported no independent FinTech application; 15 respondents (5.7%) had no recorded use in either category.

A case-level audit of the frozen N = 262 analytical sample identified 15 respondents (5.7%) who explicitly reported non-use and had no independent or bank-based FinTech application recorded. Because Q15-Q20 ask about experienced use-related benefits and outcomes, responses from these 15 cases may reflect hypothetical or anticipated evaluations. They were therefore excluded in a sensitivity analysis rather than being assumed to be equivalent to respondents with recorded FinTech use. The remaining user-status sample contained 247 respondents; after applying the same age- and income-complete-case requirements as in the primary model, the sensitivity regression used N = 242.

The identical four-block hierarchical specification was estimated for N = 242 complete user cases using the same pooled-sample standardized predictors, demographic controls, two-way interactions, and three-way interaction. The model remained statistically significant, R² = 0.355, adjusted R² = 0.309, F(16, 225) = 7.741, p < 0.001. The main-effect block remained significant (ΔR² = 0.276, p < 0.001), as did the two-way interaction block (ΔR² = 0.064, p < 0.001), whereas the three-way increment remained non-significant (ΔR² = 0.006, p = 0.146).

The inferential pattern was unchanged. BTF remained positive and significant (B = 0.335, β = 0.404, p < 0.001), FL remained positive and significant (B = 0.242, β = 0.296, p < 0.001), and the DM × FL interaction remained positive and significant (B = 0.163, β = 0.270, p = 0.006). The direct DM coefficient remained positive but non-significant (B = 0.109, β = 0.133, p = 0.078), and the three-way interaction remained non-significant (B = -0.041, β = -0.158, p = 0.146).

Thus, the primary conclusions are not driven by the small number of respondents without recorded FinTech use. Accordingly, the sensitivity analysis supports the stability of the main inferential pattern among respondents with recorded FinTech use, while the 15 non-user responses are not interpreted as validated experience-based assessments of Q15-Q20.

4.5 Preliminary Pearson correlation analysis

Pearson correlation analysis was conducted as a preliminary bivariate assessment of the direction and strength of associations among four observed composites: DM (DM_R), BTF (BTF_R), FL (FL_R), and PFI (PFI_R). The analysis used the full sample of N = 262. These zero-order correlations are descriptive preliminary evidence and are not used to determine support for the formal hypotheses.

Table 15 shows that PFI is positively correlated with DM (r = 0.402, p < 0.001), BTF (r = 0.533, p < 0.001), and FL (r = 0.418, p < 0.001). These coefficients indicate positive bivariate associations but do not establish the adjusted regression effects specified in H1-H3. The explanatory composites are also positively intercorrelated: DM with BTF (r = 0.649, p < 0.001), DM with FL (r = 0.439, p < 0.001), and BTF with FL (r = 0.519, p < 0.001). These correlations provide descriptive context for the multivariable model and do not by themselves test moderation or higher-order interaction hypotheses.

Table 15. Preliminary Pearson correlations among observed composites

Pearson correlations are reported only as preliminary descriptive evidence and are not treated as a separate formal hypothesis.

Formal hypothesis decisions are based exclusively on the correctly specified four-block hierarchical OLS model reported in Section 4.4 and Table 13, where each coefficient is evaluated while accounting for the other terms in the model.

Accordingly, H1 is not supported at the 0.05 level because the adjusted DM coefficient is positive but non-significant (β = 0.127, p = 0.076). H2 and H3 are supported because BTF (β = 0.438, p < 0.001) and FL (β = 0.287, p < 0.001) are positive and significant. H4 is supported by the significant DM × FL interaction (β = 0.266, p = 0.004), whereas H5 is not supported because the three-way interaction is non-significant (β = -0.176, p = 0.083).

This separation prevents bivariate correlations from being interpreted as evidence for adjusted direct associations or interaction hypotheses and aligns the hypothesis decisions with the regression model used for formal inference.

Figure 13 is the standard residual-versus-fitted diagnostic for the final pooled model: standardized residuals are plotted against standardized predicted values. Residuals are generally centered around zero and the plot does not show a pronounced systematic curve, which supports approximate linearity. The vertical spread is not fully constant and several tail observations are visible, so the graph is not used as evidence of model fit; instead, it motivates the formal heteroskedasticity and influence checks reported above.

Since the study includes respondents from three countries, country was included in the pooled regression model as a demographic control variable. Separate country-specific regressions are presented only as supplementary within-country sensitivity analyses and are not used to infer statistically established cross-country differences in regression coefficients or associations.

4.6 Country-specific regression comparison

Separate two-block OLS regressions were estimated for Kosovo, Albania, and North Macedonia using the observed composite measures. In each country, PFI_R was the dependent variable. Block 1 contained age-group dummies (Age2_D-Age5_D) and income-category dummies (Inc2_D-Inc4_D), with the first age and income categories serving as reference groups. Block 2 added DM_R, BTF_R, and FL_R. Country dummies were not included because country is constant within each national subsample. Complete-case analysis yielded N = 156 for Kosovo, N = 55 for Albania, and N = 44 for North Macedonia.

The country-specific models use the same two-block specification and are reported as supplementary within-country sensitivity analyses rather than as formal cross-country comparisons.

The country-specific model-fit results are reported in Table 16.

Table 16. Country-specific regression model fit and incremental explanatory power

The country-specific regression analyses revealed notable differences in model performance across Kosovo, Albania, and North Macedonia. For Kosovo (N = 156), the final regression model was statistically significant and explained a substantial proportion of the variance in the outcome, R² = 0.433, adjusted R² = 0.394, F(10, 145) = 11.082, p < 0.001. After the demographic control variables were entered, the addition of DM (DM_R), BTF (BTF_R), and FL (FL_R) produced a considerable increase in explained variance, ΔR² = 0.406, F-change(3, 145) = 34.613, p < 0.001. The Durbin-Watson statistic of 1.875 indicated no substantial concern regarding residual autocorrelation.

A similar, although weaker, pattern emerged for Albania (N = 55). The final model was statistically significant, R² = 0.379, adjusted R² = 0.237, F(10, 44) = 2.680, p = 0.012, indicating that the predictors collectively accounted for a meaningful proportion of the variation in the dependent variable. The introduction of the focal predictors after the demographic controls increased the explained variance by ΔR² = 0.293, with this change being statistically significant, F-change(3, 44) = 6.912, p = 0.001. The Durbin-Watson value of 1.993 was close to the conventional reference value of 2, suggesting no important residual autocorrelation.

In contrast, the results for North Macedonia (N = 44) were substantially weaker. The final model was not statistically significant, R² = 0.210, adjusted R² = -0.030, F(10, 33) = 0.876, p = 0.564. Moreover, adding DM_R, BTF_R, and FL_R after the demographic controls produced only a limited increase in explained variance, ΔR² = 0.079, and this increment was not statistically significant, F-change(3, 33) = 1.107, p = 0.360. The Durbin-Watson statistic was 2.207, again indicating no major concern regarding residual autocorrelation. Overall, the country-specific findings suggest that the focal predictors demonstrated the strongest explanatory contribution in Kosovo, a moderate but statistically significant contribution in Albania, and no statistically significant contribution in North Macedonia.

The country-specific regression results in Table 17 reveal distinct patterns in the effects of DM, BTF, and FL across Kosovo, Albania, and North Macedonia.

Table 17. Country-specific coefficients for the focal predictors

•Analysis for Kosovo

In Kosovo, DM (Table 17) was not statistically significant (B = 0.004, SE = 0.085, β = 0.004, t = 0.046, p = 0.963, 95% CI [-0.163, 0.171], VIF = 1.823). In contrast, BTF showed a positive and statistically significant association with the outcome (B = 0.489, SE = 0.081, β = 0.523, t = 6.019, p < 0.001, 95% CI [0.328, 0.649], VIF = 1.932). FL was also positively and significantly associated with the outcome (B = 0.179, SE = 0.071, β = 0.192, t = 2.539, p = 0.012, 95% CI [0.040, 0.319], VIF = 1.466). Thus, within the Kosovo subsample, BTF and FL emerged as significant focal predictors, whereas DM did not.

image044.png

Figure 14. Country-specific coefficient profile of the focal predictors in Kosovo

Figure 14 presents the three focal predictors. BTF has the strongest positive association with PFI, with an unstandardized coefficient of B = 0.489 and a 95% CI of [0.328, 0.649]; the effect is statistically significant (p < 0.001). FL also shows a positive and statistically significant association, although smaller in magnitude (B = 0.179, 95% CI [0.040, 0.319], p = 0.012). By contrast, DM is positioned almost exactly at zero (B = 0.004, 95% CI [-0.163, 0.171], p = 0.963), and its CI crosses zero, indicating no statistically significant independent association.

The substantive content of the indices helps explain this visual pattern. BTF combine transaction security and trust (Q6), attitudes toward innovation (Q8), trust relative to traditional banks (Q9), and social influence (Q10), whereas FL comprises basic FL (Q11), digital-investment risk assessment (Q12), and the role of financial education in supporting decisions (Q13). DM includes social-media advertising (Q1), information clarity and credibility (Q2), UI/UX satisfaction (Q3), online promotions (Q4), and online communication (Q5). Overall, the graph and the regression estimates converge on the same conclusion: within the Kosovo subsample, trust-related factors and FL are more strongly associated with PFI than the combined DM index.

•Analysis for Albania

A similar but not identical pattern was observed in Albania (Table 17). DM was not statistically significant (B = -0.062, SE = 0.164, β = -0.060, t = -0.381, p = 0.705, 95% CI [-0.393, 0.268], VIF = 1.763). BTF, however, were positive and statistically significant (B = 0.557, SE = 0.176, β = 0.508, t = 3.159, p = 0.003, 95% CI [0.202, 0.912], VIF = 1.832). FL also showed a positive coefficient, but its effect did not reach statistical significance (B = 0.189, SE = 0.128, β = 0.222, t = 1.476, p = 0.147, 95% CI [-0.069, 0.446], VIF = 1.601). Accordingly, BTF were the only focal predictor that reached the conventional 0.05 significance level in the Albanian subsample.

The coefficient profile (Figure 15) shows that BTF have the strongest positive association with PFI (B = 0.557, 95% CI [0.202, 0.912], p = 0.003). The graph confirms this pattern because the CI remains entirely above zero. This index reflects transaction security and trust (Q6), attitudes toward innovation (Q8), trust relative to traditional banks (Q9), and social influence (Q10), suggesting that these trust-related dimensions are particularly important in the Albanian subsample.

image045.png

Figure 15. Country-specific coefficient profile of the focal predictors in Albania

FL also has a positive coefficient (B = 0.189), but its 95% CI [-0.069, 0.446] crosses zero and the effect is not statistically significant (p = 0.147). This index includes basic FL (Q11), digital-investment risk assessment (Q12), and financial education for informed decisions (Q13).

DM shows a small negative and non-significant coefficient (B = -0.062, 95% CI [-0.393, 0.268], p = 0.705). The index covers social-media advertising, information clarity and credibility, UI/UX satisfaction, online promotions, and online communication (Q1–Q5). Overall, the graph indicates that trust-related factors are the clearest correlate of PFI in Albania, while FL is positive but uncertain and DM remains statistically neutral.

•Analysis for North Macedonia

The results (Table 17) for North Macedonia were weaker, as none of the three focal predictors reached the 0.05 significance threshold. DM displayed a positive but non-significant association (B = 0.360, SE = 0.199, β = 0.437, t = 1.812, p = 0.079, 95% CI [-0.044, 0.764], VIF = 2.426). BTF were also non-significant and carried a negative coefficient (B = -0.202, SE = 0.192, β = -0.257, t = -1.053, p = 0.300, 95% CI [-0.593, 0.189], VIF = 2.493). FL was essentially null in this subsample (B = -0.009, SE = 0.192, β = -0.010, t = -0.048, p = 0.962, 95% CI [-0.399, 0.380], VIF = 1.911). These findings indicate that none of the three focal predictors provided statistically significant explanatory power at the 0.05 level within the North Macedonian subsample.

•Analysis for North Macedonia

The coefficient profile (Figure 16) indicates that none of the three focal predictors reaches statistical significance at the 0.05 level. DM shows the largest positive coefficient (B = 0.360, 95% CI [-0.044, 0.764], p = 0.079), but its CI crosses zero, so the observed positive pattern remains statistically uncertain. This index includes social-media advertising, information clarity and credibility, UI/UX satisfaction, online promotions, and online communication (Q1–Q5). BTF show a negative but non-significant association (B = -0.202, 95% CI [-0.593, 0.189], p = 0.300). This composite reflects transaction security and trust (Q6), attitudes toward innovation (Q8), trust relative to traditional banks (Q9), and social influence (Q10). FL is essentially neutral (B = -0.009, 95% CI [-0.399, 0.380], p = 0.962) and comprises basic FL, digital-investment risk assessment, and financial education supporting informed decisions (Q11–Q13).

Overall, the graph shows wide CI and no statistically reliable focal predictor in North Macedonia, indicating greater uncertainty in this smaller subsample.

image046.png

Figure 16. Country-specific coefficient profile of the focal predictors in North Macedonia

Taken together, Kosovo, Albania and North Macedonia, the country-specific pattern shows that BTF are statistically significant in both Kosovo and Albania, whereas FL is statistically significant only in Kosovo. In contrast, the North Macedonia model does not yield any statistically significant focal predictor at the 0.05 level. These country-specific results should therefore be interpreted as descriptive within-country sensitivity evidence rather than as formal tests of whether regression coefficients differ statistically across countries.

4.6.1 Scope and limits of country-specific estimates

The country-specific analyses are supplementary sensitivity analyses rather than evidence of stable cross-country structural differences. The full sample is substantially unbalanced: Kosovo n = 163 (62.2%), Albania n = 55 (21.0%), and North Macedonia n = 44 (16.8%); the corresponding complete-case regression samples are N = 156, N = 55, and N = 44. The small North Macedonia subsample limits precision and makes strong multi-group conclusions inappropriate. None of the country-specific coefficients, hypothesis decisions, or model-fit statistics is presented as nationally representative evidence. In the pooled regression, Kosovo is the reference category for the Albania and North Macedonia country dummies; this is a coding convention for estimating adjusted contrasts and does not make Kosovo a population benchmark.

Measurement equivalence across countries was not established. The exploratory measurement assessment was conducted on the pooled sample using PAF and internal-consistency reliability; no multi-group CFA was estimated, so configural, metric, and scalar invariance are not claimed. Differences in coefficient magnitude or statistical significance across country regressions therefore cannot be interpreted as statistically demonstrated differences in the underlying relationships. The focal predictors show acceptable VIF values in all three models (1.466-2.493), whereas some demographic dummy controls have elevated VIFs in the smaller Albania and North Macedonia subsamples, further limiting subgroup precision.

Accordingly, the country-specific regressions provide reproducible within-country estimates under one common specification but do not constitute formal tests of equality of regression coefficients across countries. The pooled hierarchical OLS model remains the primary basis for inference, and subgroup estimates are used only to assess descriptive heterogeneity and sensitivity of the pooled pattern.

The supplementary country-specific regressions provide descriptive within-country evidence under a common specification. Kosovo shows significant positive BTF and FL associations; Albania shows a significant positive BTF association; and North Macedonia has no statistically significant focal predictor in the fully adjusted country model. Differences in coefficient magnitude or statistical significance are not interpreted as formal evidence that the underlying relationships differ across countries.

4.7 Summary of hypothesis testing and country-specific evidence

This section summarizes the overall hypothesis-testing results and presents the country-specific evidence to highlight similarities and differences in the observed relationships across Kosovo, Albania, and North Macedonia.

Table 18 consolidates the formal hypothesis decisions and shows that conclusions are derived from the adjusted hierarchical regression rather than from zero-order correlations. H1 is not supported because the DM coefficient, although positive, does not reach the 0.05 threshold in the fully specified model (β = 0.127, p = 0.076). H2 is supported: BTF show the strongest positive adjusted association with PFI (β = 0.438, p < 0.001). H3 is also supported, with FL retaining a positive and statistically significant association after adjustment for controls and the other explanatory variables (β = 0.287, p < 0.001). H4 is supported because the DM × FL interaction is positive and significant (β = 0.266, p = 0.004), indicating that the DM association becomes stronger at higher levels of FL. H5 is not supported because the three-way interaction is not statistically significant (β = -0.176, p = 0.083). Robust covariance estimation leaves all five support decisions unchanged.

Table 18. Summary of hypothesis testing results

Note: Decisions are based on the fully specified four-block hierarchical OLS model. H4 is evaluated through the DM × FL interaction and H5 through the three-way interaction. Heteroskedasticity-robust covariance estimation produced the same decisions (robust p-values: H1 = 0.056; H2 < 0.001; H3 < 0.001; H4 = 0.002; H5 = 0.125).

Table 19 shows that the country-specific evidence is heterogeneous and should be interpreted as descriptive within-country sensitivity evidence rather than as formal cross-country hypothesis testing. In Kosovo, H1 is not supported because DM is essentially null (β = 0.004, p = 0.963), whereas H2 and H3 are supported: BTF (β = 0.523, p < 0.001) and FL (β = 0.192, p = 0.012) are both positively associated with PFI. In Albania, H1 is also not supported (β = -0.060, p = 0.705), while H2 is supported through a strong positive BTF association (β = 0.508, p = 0.003). H3 is not supported in Albania because the FL coefficient is positive but non-significant (β = 0.222, p = 0.147). In North Macedonia, none of H1-H3 is supported at the 0.05 level. DM is positive but non-significant (β = 0.437, p = 0.079), while BTF and FL are also non-significant. This pattern indicates that the pooled findings are not reproduced uniformly in every national subsample. However, the smaller subgroup sizes, particularly for North Macedonia, and the absence of established measurement invariance mean that differences in significance cannot be interpreted as proof that the underlying associations differ statistically across countries.

Table 19. Country-specific evidence for H1–H3

Note: These regressions are supplementary within-country sensitivity analyses. “Supported” denotes a positive coefficient significant at p < 0.05 within that country model. H4 and H5 were not estimated separately by country because the country-specific specifications did not include interaction terms. The table is not a formal test of cross-country coefficient differences.

The three countries are also examined through separate regressions, in which each national subsample is estimated independently and no country serves as a reference category. By contrast, in the pooled hierarchical model, Kosovo is the reference category for the country control because Albania_D and NMacedonia_D are entered as dummy variables while Kosovo is omitted. Consequently, the two country-dummy coefficients compare Albania and North Macedonia with Kosovo on the conditional outcome level, holding the remaining model terms constant. This coding does not test whether the slopes of DM, BTF, or FL differ between countries; such comparisons would require explicit country-by-predictor interaction terms.

Taken together, the two summaries provide a transparent bridge from formal hypothesis testing to the subsequent discussion. The pooled hierarchical model remains the primary basis for inference because it uses the largest common analytical sample and evaluates the focal predictors, controls, and interactions within one coherent specification. The country-specific models add useful sensitivity information by showing where similar or different within-country patterns appear, but they are not treated as evidence of statistically established national differences. This distinction preserves the strength of the supported findings while avoiding overinterpretation of subgroup results. The discussion that follows therefore emphasizes the pooled hypothesis decisions and uses country-level estimates only as supplementary contextual evidence.

Within the cross-sectional sample from Kosovo, Albania, and North Macedonia, PFI is associated with respondents' BTF, FL, and the conditional relationship between DM and FL. The dependent measure is an observed composite of use-related benefits, perceived financial-management outcomes, and recommendation/advocacy. These findings describe contemporaneous perceived associations at one point in time; they do not establish causal effects, persistence of use, retention, or long-term/sustainable adoption.

The final hierarchical model indicates that BTF have the strongest positive standardized association with PFI (β = 0.438, p < 0.001), followed by FL (β = 0.287, p < 0.001). DM is positive but not independently significant at 0.05 (β = 0.127, p = 0.076). The significant DM × FL interaction (β = 0.266, p = 0.004) shows that the association involving DM varies with FL.

Because the residual-variance diagnostic indicated heteroskedasticity, inference was also checked with robust sandwich standard errors. This sensitivity analysis did not change the hypothesis conclusions: DM remained non-significant (p = 0.056), BTF and FL remained significant (both p < 0.001), DM × FL remained significant (p = 0.002), and the three-way term remained non-significant (p = 0.125).

A key conditional result is the positive DM × FL interaction. The three-way interaction is not statistically significant and is not used to support claims about overload, saturation, or diminishing returns. Simple-slope analysis shows that the DM association is non-significant at low FL (B = -0.059, p = 0.456) and at the mean (B = 0.109, p = 0.076), but positive and significant at high FL (B = 0.277, p = 0.002). Johnson-Neyman analysis further indicates that the positive conditional DM association with FL is approximately 0.07 SD above its sample mean, with BTF held at its standardized mean.

The country-specific regressions provide supplementary regional context only. BTF are significant in the Kosovo and Albania subsamples, FL is significant only in Kosovo, and the North Macedonia model is not statistically significant overall. These patterns are not interpreted as formal between-country differences because the national subsamples are unbalanced and measurement invariance was not established. The pooled hierarchical model remains the primary basis for inference, and country patterns describe within-sample associations rather than national population estimates.

Overall, the study extends TAM and UTAUT by integrating marketing, behavioral, and FL dimensions into a broader model of PFI. Future research should use longitudinal or panel designs to evaluate whether FinTech use and perceived benefits persist over time and should directly measure continuance intention, repeated-use frequency, retention, and responsible-use behavior. Such designs could distinguish short-term perceptions from sustained behavioral patterns and provide a stronger basis for evaluating persistence. Future studies should also expand the model by including institutional trust, regulatory awareness, and digital infrastructure. Survey designs should separate actual users from potential users at the screening stage and apply skip logic so that experience-dependent outcome items are administered only to respondents with relevant FinTech use. Longitudinal, temporally separated, or multi-source measurement would additionally reduce common-method concerns and strengthen causal and temporal inference. These additional constructs were not measured in the present model and are proposed strictly as future research directions rather than as evidence-based policy recommendations.

6.1 Recommendations

To distinguish evidence-based implications from broader policy proposals, the recommendations in this section are limited to constructs represented in the empirical model: DM, BTF, FL, and PFI. Based on the observed associations, practical implications for FinTech in Kosovo, Albania, and North Macedonia are limited to the measured user-level dimensions. Higher PFI was observed alongside stronger BTF and FL, while the DM association varied with FL. These findings provide the empirical basis for the recommendations directed at FinTech providers, banks, public institutions, educational organizations, and users.

For banks and FinTech providers, the main implication is to combine DM with financial education and trust-building communication. Promotional campaigns should not only advertise services but also communicate potential user benefits such as time savings, expense monitoring, financial-management support, and reduced uncertainty. Mobile banking and FinTech applications can provide clear educational messages, transparent security information, and user-friendly design that supports informed decision-making.

For public institutions, regulators, and educational organizations, the empirical evidence most directly supports financial-literacy initiatives and clear communication that helps users understand digital financial services, evaluate perceived security, and make informed financial decisions.

For users, FinTech should be viewed not only as a convenient digital tool, but as part of responsible financial management. Users should understand how digital services work, compare information carefully, and use FinTech platforms for budgeting, expense monitoring, and financial planning.

Building directly on these evidence-based implications, several practical actions can be recommended. Banks and FinTech providers can integrate short financial-literacy explanations directly into mobile applications, online banking platforms, and digital campaigns, particularly when introducing new services or features. Promotional messages can be accompanied by concise explanations of service benefits, transaction procedures, perceived-security features, fees where relevant, and the ways in which digital tools can support budgeting and financial management. Providers can also offer short in-app tutorials, frequently asked questions, practical examples, and step-by-step guidance to help users understand digital financial products before making decisions.

Trust-related communication can be strengthened through transparent explanations of security procedures, authentication processes, transaction confirmation, and available customer-support channels. In addition, financial-literacy initiatives can include practical guidance on comparing digital financial products, monitoring expenses, evaluating financial information, and incorporating FinTech tools into regular budgeting and financial planning. Banks and FinTech providers can periodically assess users' perceptions of trust, financial understanding, digital communication, and perceived FinTech benefits in order to identify areas where user education and communication require improvement. These actions remain focused on the constructs measured in the present study and should not be interpreted as evidence of effects arising from unmeasured institutional or regulatory conditions.

Country-specific recommendations are not derived from subgroup coefficient differences. The subgroup regressions are exploratory and do not establish that different strategies are required in Kosovo, Albania, and North Macedonia. Evidence-based implications are restricted to the measured dimensions: FL, trust-related and perceived-security factors, DM/communication, and PFI. Institutional trust, regulatory quality, digital infrastructure, consumer-protection systems, and other macro-level conditions are reserved for future research rather than presented as findings of the current model. Future country-comparative research should use larger and more balanced national samples and formally assess measurement invariance before testing differences in regression coefficients.

Overall, the findings support a user-centered FinTech environment in which digital financial services are understandable, trustworthy, and practically useful for users' financial well-being. Because sustainability and temporal persistence were not directly operationalized, these implications concern current PFI and should not be interpreted as evidence of sustainable, persistent, or long-term adoption. Such claims require direct longitudinal measures of continuance, repeated use, retention, and responsible FinTech behavior.