Prior entrepreneurship research has devoted considerable attention to identifying the antecedents of entrepreneurial intention through frameworks such as the Theory of Planned Behavior (TPB); however, relatively few studies have empirically examined how intention is subsequently converted into actual entrepreneurial behavior within a unified analytical framework (Roos & Botha, 2022). The mechanisms governing the intention - action transition, particularly the role of resource endowments and ecosystem conditions as moderating or enabling factors, remain insufficiently theorised and rarely tested in sequential empirical designs.

Existing entrepreneurship research is disproportionately concentrated in metropolitan or economically advanced contexts. In emerging economies, the intention - action gap is particularly pronounced, reflecting structural misalignments among individual resources, institutional capacity, and entrepreneurial ecosystems (Cao & Shi, 2021; Lone et al., 2026; Roundy, 2017). Vietnam’s entrepreneurial ecosystem has achieved notable progress, ranking 56th globally and 5th in Southeast Asia in 2024, with more than 4,000 innovative start-ups and 208 investment funds (Ministry of Science and Technology of Vietnam, 2024), yet this development remains heavily concentrated in Hanoi and Ho Chi Minh City. Mountainous border regions characterised by geographical isolation, ethnic heterogeneity, and resource constraints constitute a significant research gap, particularly with respect to rigorous quantitative studies (Lone et al., 2026).

Most prior studies rely on single-method designs that either model intention as the terminal outcome or examine behavioral action without capturing its latitudinal determinants as a latent construct. This methodological fragmentation impedes a comprehensive understanding of the full entrepreneurial process. Lao Cai a high-altitude border province in northern Vietnam with 182 kilometres of frontier adjoining China, 27 co-resident ethnic groups, and over 76% of the population residing in rural areas (Lao Cai Statistical Office, 2025) exemplifies this underexplored context. During 2023–2026, approximately 30% of start-up projects supported by the Lao Cai Youth Union failed due to insufficient capital or managerial experience (Lao Cai Youth Union, 2026), signalling that intention alone is insufficient to predict entrepreneurial success.

Accordingly, this study addresses two research questions: (1) Which individual and contextual factors shape entrepreneurial intention among young people in a mountainous emerging economy? (2) To what extent does entrepreneurial intention translate into actual entrepreneurial action? To answer these questions, the study develops a sequential analytical framework in which entrepreneurial intention is first estimated using PLS-SEM and subsequently linked to entrepreneurial decision through Binary Logistic Regression.

This study makes four principal contributions. First, it explicitly models the entrepreneurial process as a sequential transition from intention - action, extending TPB with resource-based and entrepreneurial ecosystem variables while maintaining a unified empirical framework. Second, it introduces a two-stage design combining PLS-SEM (to estimate intention as a latent construct) and Binary Logistic Regression (to predict the probability of actual entrepreneurial decision-making), enabling more comprehensive examination of the intention - action gap than conventional single-method approaches. Third, it provides empirical evidence from a mountainous, multi-ethnic border province in an emerging economy a context that remains substantially underrepresented in the literature. Fourth, the findings offer evidence-based implications for the design of youth entrepreneurship support policies in disadvantaged mountainous regions.

Entrepreneurship is understood as the process by which an individual identifies, evaluates, and exploits business opportunities through the creation of a new economic organisation (Hitt et al., 2011; Sheeran, 2002). Among young people, entrepreneurial activity carries a dual significance: economically, it generates employment and income; socially, it fosters innovation and reduces youth unemployment (Ogamba, 2019). However, a substantial body of entrepreneurship research has long documented an ‘intention - action gap’: many individuals harbour clear entrepreneurial intentions yet fail to convert them into actual behaviour due to insufficient resources, capability deficits, or inadequate support environments (Kautonen et al., 2015; Kuo & Young, 2008). Crucially, this gap is not uniform across contexts, it tends to be more pronounced in resource-constrained and institutionally underdeveloped settings, where structural barriers mediate the translation of intention into action (Cao & Shi, 2021). Clarifying the conversion mechanism, particularly in mountainous regions of emerging economies, therefore carries considerable theoretical and practical importance.

The Theory of Planned Behavior - TPB (Ajzen, 1991) constitutes the most widely applied theoretical framework in entrepreneurial intention research. TPB posits that behavioral intention is determined by three antecedents, attitude toward the behavior (ATT), subjective norms (SN), and perceived behavioral control (PBC) and that behavioural intention is the most proximal and strongest predictor of actual behavior. While TPB has accumulated extensive empirical support across diverse cultural and economic contexts, its explanatory power varies considerably depending on the institutional environment and population under study (Krueger Jr et al., 2000; Liñán & Chen, 2009), underscoring the necessity of context-specific empirical testing.

Attitude toward entrepreneurship (ATT) captures the extent to which an individual evaluates the prospect of becoming an entrepreneur positively or negatively. A preponderance of evidence from developed economies confirms that attitude exerts a robust and stable positive influence on entrepreneurial intention (Dissanayake, 2014; Liñán & Chen, 2009). However, findings in peripheral and emerging economy contexts are less consistent: in settings where entrepreneurship is perceived as high-risk or socially undesirable, attitudinal effects are attenuated (Baluku et al., 2021), and the relationship may be contingent on the availability of supportive institutions. Notwithstanding this variability, the balance of evidence suggests that when individuals regard entrepreneurship as a personally valuable and attractive career pathway, stronger entrepreneurial intentions are likely to follow.

Subjective norms (SN) reflect an individual’s perception of the expectations held by significant others, including family members, peers, and broader society regarding entrepreneurial behavior. (Ajzen, 1991) argues that individuals tend to conform to referent group expectations when perceived social pressure is salient. While the effect of subjective norms on entrepreneurial intention is theoretically well-grounded, it is empirically the most contested of the three TPB components: several meta-analyses report weak or non-significant direct effects in Western contexts (Autio et al., 2001; Schlaegel & Koenig, 2014), whereas studies in collectivist and family-oriented societies, including Vietnam consistently document stronger and more significant subjective norm effects, as family and social endorsement constitute primary sources of motivation and legitimacy for entrepreneurial pursuit (Doanh, 2021; Hoang et al., 2020). Given the collectivist cultural context of Lao Cai, subjective norms are expected to be a meaningful predictor of entrepreneurial intention.

Perceived behavioral control (PBC) denotes an individual’s belief in their capacity to master and regulate the intended behavior, and is consistently recognised as among the strongest predictors of behavioural intention across both developed and developing country settings particularly when the behavior is subject to factors outside the actor’s full control (Ajzen, 1991; Liñán & Chen, 2009). In the entrepreneurship context, PBC is operationalised through self-confidence regarding personal competence, resource availability, and risk-bearing capacity. Critically, while PBC effects are broadly replicated in the literature, their magnitude is moderated by the objective opportunity structure of the environment: in regions with limited access to capital and business support, perceived control may be systematically suppressed regardless of individual disposition (Cao & Shi, 2021; Krueger Jr et al., 2000). Despite this caveat, PBC remains a theoretically and empirically indispensable predictor of entrepreneurial intention.

Beyond the three core TPB components, the Resource-Based View (RBV) emphasises that the formation and enactment of entrepreneurial intention depends on the resources an individual can mobilise, encompassing human capital resources (education, experience) and relational resources (social capital). While TPB captures motivational and attitudinal antecedents, RBV complements this by accounting for the enabling conditions that transform intention into feasible action (Davidsson & Honig, 2003).

Entrepreneurship education (EDU) enables individuals to accumulate knowledge of business processes, planning, financial management, legal requirements, and risk management. Evidence from developed economies consistently demonstrates that entrepreneurship training programs exert positive effects on entrepreneurial cognition and motivation, particularly when theory is integrated with practice (Nabi et al., 2017). However, findings in developing country contexts are more mixed: several studies report that the impact of formal entrepreneurship education is contingent on curriculum quality, instructor competence, and alignment with local market realities, factors that vary considerably across institutional settings (Bae et al., 2014). In the Vietnamese context, where entrepreneurship education has expanded rapidly but unevenly across urban and rural areas, education has been identified as a significant contributor to the formation of young people’s entrepreneurial intentions (Doanh, 2021; Hoang et al., 2020), suggesting that its positive effect is likely to persist, albeit potentially at reduced magnitude in mountainous peripheral settings.

Social capital (SC) refers to the network of social relationships through which an individual can access information, resources, and collaborative opportunities. The broadly positive effect of social capital on entrepreneurial intention is well-established in the literature (Gloor et al., 2009); social networks reduce information asymmetries, lower perceived barriers to entry, and expand access to mentoring and co-entrepreneurship opportunities. Nevertheless, the composition and quality of social networks matter: in geographically isolated communities, social ties are frequently bonding rather than bridging in nature, which may constrain rather than expand entrepreneurial opportunities by limiting exposure to diverse knowledge and external markets (Putnam, 2000). In the Vietnamese context, individuals with higher social capital tend to form entrepreneurial intentions and convert them into entrepreneurial behavior more readily (Bui et al., 2018), though the specific role of network type in mountainous multi-ethnic settings warrants empirical investigation.

Experience and entrepreneurial readiness (EXP) captures the degree of practical preparation an individual has achieved, encompassing prior business experience, knowledge of business formation procedures, and the prior development of concrete business plans or ideas. Studies in developed economies demonstrate that managerial experience and prior venture creation experience substantially enhance entrepreneurial intention by reducing perceived uncertainty and increasing entrepreneurial self-efficacy (Dissanayake, 2014; Stuart & Abetti, 1990). However, in resource-constrained environments, the relationship between experience and intention may be complicated by the prevalence of necessity-driven rather than opportunity-driven entrepreneurship, wherein prior negative business experiences may dampen rather than reinforce intention (Global Entrepreneurship Monitor, 2025). Vietnamese studies nonetheless find that individuals with prior small-scale business experience hold higher entrepreneurial intentions than those without (Hoang et al., 2020), and this pattern is expected to hold within the Lao Cai context.

The entrepreneurial ecosystem is understood as the constellation of actors, institutions, and interactions, including supportive policies, financing, incubators, and advisory networks that collectively create enabling conditions for entrepreneurial activity (Spigel, 2017). Ecosystem support (SUP) encompasses an individual’s access to start-up financing, incubation services, business advisory, and government policy information. Evidence from developed economies consistently demonstrates that institutional and policy support contributes substantially to the realisation of entrepreneurial ideas (Davidsson & Honig, 2003). Yet the effectiveness of ecosystem support is highly context-dependent: in peripheral regions of emerging economies, support mechanisms frequently suffer from limited outreach, low awareness among potential beneficiaries, and misalignment between available instruments and the actual needs of young entrepreneurs (Xu & Dobson, 2019). In Vietnam, the entrepreneurial ecosystem is developing rapidly but continues to require improvements in effectiveness and accessibility, particularly for young people located outside the two major urban centres (Bui et al., 2018; Hoang et al., 2020). This suggests that while ecosystem support is theoretically expected to enhance entrepreneurial intention, the magnitude of this effect may be attenuated in mountainous and underserved regions such as Lao Cai.

Within the TPB framework, behavioral intention is posited as the direct and most immediate precursor of actual behavior (Ajzen, 1991). In the entrepreneurship domain, multiple studies have demonstrated that individuals with clearly formed intentions are more likely to undertake concrete steps, such as drafting a business plan, completing legal registration, or commencing business operations (Kolvereid & Isaksen, 2006; Krueger Jr et al., 2000). Meta-analytic evidence further confirms that the intention behavior link is among the most robust in the behavioral sciences (Sheeran, 2002). However, this relationship is neither automatic nor universal: the degree to which intention converts into action varies substantially across institutional and resource contexts, as structural barriers, including capital constraints, regulatory complexity, and weak support networks may interrupt the intention - action pathway even among highly motivated individuals (Kautonen et al., 2015; Kuo & Young, 2008; Roos & Botha, 2022). In mountainous emerging economies such as Lao Cai, where approximately 30% of supported start-up projects have failed due to capital and managerial deficiencies (Lao Cai Youth Union, 2026), this attenuation risk is particularly salient. Empirical testing of the intention–action relationship within this specific context is therefore both theoretically warranted and practically necessary.

Note: The model integrates three theoretical frameworks: the Theory of Planned Behavior (TPB), comprising attitude toward entrepreneurship (ATT), subjective norms (SN), and perceived behavioral control (PBC); the Resource-Based View (RBV), comprising entrepreneurship education (EDU), social capital (SC), and experience and readiness (EXP); and Entrepreneurial Ecosystem Theory, operationalised as ecosystem support (SUP). All seven constructs (H1–H7) ( Figure 1) are hypothesised to predict entrepreneurial intention (EI), which is estimated via Partial Least Squares Structural Equation Modeling (Stage 1). Entrepreneurial intention is subsequently entered as the primary predictor of entrepreneurial decision (ED) in a Binary Logistic Regression model (Stage 2; H8).

The study was conducted in Lao Cai province a high-altitude border province in northern Vietnam with a natural area of 13,256.92 km², 182 kilometres of land border with China, and a total population of 1,778,785 as of 2025, of whom ethnic minorities constitute more than 61%. In 2025, the province’s Gross Regional Domestic Product (GRDP) growth rate reached 8.5%, exceeding the national average (8.02%). The economic structure continues to shift positively: agriculture, forestry, and aquaculture contribute 16.0%; industry and construction 37.31%; and services 39.16%. The total youth population of the province in 2025 was 319,172, accounting for approximately 17.9% of the total provincial population (Lao Cai Youth Union, 2026). The dispersed, multi-ethnic, and predominantly rural demographic profile of Lao Cai constitutes a theoretically appropriate empirical setting for testing a model of youth entrepreneurship determinants in a mountainous emerging economy.

Sampling strategy and justification: Data were collected through direct surveys of young people across Lao Cai province, employing a non-probability sampling strategy that combined convenience sampling with snowball sampling. This approach was adopted in response to three structural constraints that render probability sampling infeasible in this context. First, the target population, young people with entrepreneurial intentions or experience in Lao Cai constitutes a difficult-to-identify subgroup for which no complete or reliable sampling frame exists. Second, the mountainous and geographically fragmented terrain of the province, with settlements dispersed across 177 commune-level administrative units, renders systematic area-probability sampling logistically prohibitive. Third, the multi-ethnic composition of the population, encompassing 27 co-resident ethnic groups, several of which are concentrated in remote high-altitude communities, further limits the feasibility of random enumeration approaches. Under these conditions, convenience sampling combined with snowball referral from initial respondents to socially proximate peers constitutes the methodological approach most widely recommended for reaching dispersed, difficult-to-access populations in behavioral research (Ting et al., 2025). To mitigate the risk of selection bias inherent in non-probability designs, data collection was conducted across all nine districts of the province and stratified by residential location (urban/rural), age group, and educational attainment to ensure demographic representativeness consistent with the provincial youth profile.

Informed consent and ethical considerations: Before participating in the survey, all respondents were informed about the purpose of the research, the voluntary nature of participation, their right to refuse or withdraw at any stage without penalty, and the confidentiality of the information provided. Verbal informed consent was obtained from all participants prior to questionnaire administration. Verbal consent was considered more appropriate than written consent because data collection was conducted through anonymous face-to-face surveys across multiple rural and mountainous communities in Lao Cai Province. Requiring written consent could have reduced participation and weakened respondent anonymity, while verbal consent adequately ensured that participation was fully informed and voluntary. No personally identifiable information was collected, and all data were anonymized and used solely for scientific research.

Sample size and power analysis: Following data screening and cleaning, a total of 307 valid questionnaire responses were retained for analysis, comprising 196 respondents who had already initiated entrepreneurial activity (63.8%) and 111 respondents who had not yet acted on their entrepreneurial intentions (36.2%). The inclusion of both groups was necessary to measure the latent construct of entrepreneurial intention and to analyse the gap between intention and actual behavior, consistent with the TPB theoretical framework.

Sample adequacy was established through two complementary procedures. First, in accordance with (Hair & Alamer, 2022), minimum sample size requirements for PLS-SEM range from five to ten times the number of observed indicators; the present model comprises 31 observed indicators, yielding a recommended range of 155–310, which the obtained sample of 307 satisfies. Second, and more rigorously, an a priori power analysis was conducted using GPower 3.1 ( Faul et al., 2009). For the PLS-SEM structural model, assuming a medium effect size (f² = 0.15), a significance level of α = 0.05, statistical power of 1 − β = 0.80, and a maximum of seven predictors entering the most complex path, GPower returned a minimum required sample of 103 observations. For the Binary Logistic Regression estimated in Stage 2, assuming an odds ratio of 2.0, α = 0.05, power = 0.80, and a two-tailed test with one predictor, the minimum required sample was 138 observations. The obtained sample of 307 observations therefore exceeds the requirements established by both criteria, providing sufficient statistical power to detect hypothesised effects of medium magnitude and ensuring the reliability, parameter stability, and convergence properties of both estimation procedures.

Regarding sample composition, 90.6% of respondents held secondary school qualifications, 9.1% middle school qualifications, and 0.3% primary school qualifications; in terms of vocational attainment, 41.7% had completed vocational training, 39.1% held university degrees, 16.0% held associate or college-level qualifications, and 3.3% held postgraduate qualifications. Geographically, 76.2% of respondents resided in rural areas and 23.8% in urban areas, closely reflecting the provincial demographic distribution.

All latent constructs were operationalised using reflective multi-item scales adapted from established instruments validated in prior entrepreneurship research. Items were measured on five-point Likert scales ranging from 1 (Strongly disagree) to 5 (Strongly agree). To ensure semantic equivalence between the original English-language instruments and the Vietnamese field context, the questionnaire was subjected to a forward–backward translation procedure (Brislin, 1970): the instrument was first translated into Vietnamese by a bilingual researcher specialising in social science, and subsequently back-translated into English by an independent bilingual researcher without access to the original; discrepancies were resolved iteratively until full semantic equivalence was confirmed. The translated instrument was then reviewed by a panel of five domain experts comprising two entrepreneurship academics, one Lao Cai Youth Union officer with direct programme experience, and two active local entrepreneurs, who evaluated each item for content relevance, phrasing clarity, and cultural appropriateness for the multi-ethnic mountain context. Items identified as ambiguous or contextually incongruent were revised accordingly. A pilot study with 35 respondents drawn from the target population but excluded from the final sample confirmed satisfactory preliminary reliability across all constructs (Cronbach’s α ≥ 0.72) and identified no substantive comprehension difficulties. A comprehensive summary of all constructs, measurement items, and their respective sources is presented in Table 1.

Attitude toward entrepreneurship (ATT) was measured using four items adapted from (Ajzen, 1991; Liñán & Chen, 2009), capturing respondents’ evaluative orientation toward entrepreneurship as a career pathway, including perceived attractiveness, alignment with personal aspirations, and admiration for entrepreneurial role models. Subjective norms (SN) were assessed with three items from (Ajzen, 1991; Liñán & Chen, 2009), reflecting perceived social endorsement of entrepreneurial activity from three referent sources: family, peer networks, and broader society. Perceived behavioural control (PBC) was operationalised through four items adapted from (Ajzen, 1991; Dissanayake, 2014), assessing self-efficacy beliefs regarding knowledge sufficiency, capacity to manage entrepreneurial adversity, process controllability, and immediate start-up readiness. Entrepreneurship education (EDU) was measured using four items from (Dissanayake, 2014) and (Hoang et al., 2020), evaluating programme participation, perceived curriculum utility, knowledge applicability, and institutional facilitation of project-based entrepreneurship experience. Entrepreneurial ecosystem support (SUP) was assessed with four items adapted from (Bui et al., 2018) and (Xu & Dobson, 2019), covering access to start-up financing, organisational support, government policy information, and perceived local ecosystem effectiveness. Social capital (SC) was captured through three items from (Gloor et al., 2009), measuring access to successful entrepreneurial contacts, availability of experiential mentorship, and capacity to mobilise social network resources for entrepreneurial activity. Experience and entrepreneurial readiness (EXP) was operationalised using three items from (Dissanayake, 2014) and (Bui et al., 2018), assessing whether respondents had trialled a small-scale business, understood venture formation procedures, and had developed a concrete business plan or idea. Entrepreneurial intention (EI) was measured with four items adapted from (Ajzen, 1991; Liñán & Chen, 2009), capturing forward-looking intention to start or continue a venture, goal commitment, concrete planning, and a five-year entrepreneurial aspiration. Item wording was adapted to remain applicable to both respondents who had not yet initiated a venture (future-oriented phrasing) and those who had already done so (continuation-oriented phrasing), ensuring conceptual consistency across both subgroups. The binary dependent variable for Stage 2, entrepreneurial decision (ED), was derived directly from the survey screening item: ED = 1 for respondents who had already initiated a business venture, and ED = 0 for those who had not yet done so, consistent with the two-stage intention-to-action design employed in prior entrepreneurship research (Ajzen, 1991; Kautonen et al., 2015; Kolvereid & Isaksen, 2006).

Given that the entrepreneurial decision variable (ED) is binary, the study adopts a two-stage analytical strategy. In Stage 1, PLS-SEM (implemented in SmartPLS) is employed to assess the measurement model, evaluating internal consistency, convergent validity, and discriminant validity and the structural model, reporting path coefficients, R², effect sizes f², and 5,000-resample bootstrapping, with respect to hypotheses H1- H7. In Stage 2, latent variable scores for EI extracted from the PLS-SEM estimation are entered as the sole predictor in a Binary Logistic Regression model (implemented in IBM SPSS) to test hypothesis H8. The logistic regression model takes the form:

Ln[P(ED=1)/(1–P(ED=1))]=β₀+β₁EI+ε

Where P (ED = 1) denotes the probability that a respondent has initiated entrepreneurial activity; β₀ is the intercept; β₁ is the slope coefficient for entrepreneurial intention; and ε is the stochastic error term. Following (Hair & Alamer, 2022), the measurement model is considered adequate when outer loadings ≥0.70, Cronbach’s Alpha and composite reliability (CR) ≥ 0.70, average variance extracted (AVE) ≥ 0.50, and HTMT <0.90; the structural model is considered adequate when VIF < 5, t-statistics >1.96, and p < 0.05. Levels of explanatory power are interpreted as weak (R² ≈ 0.25), moderate (R² ≈ 0.50), and substantial (R² ≈ 0.75), while effect size benchmarks are small (f² ≈ 0.02), medium (f² ≈ 0.15), and large (f² ≥ 0.35; Cohen, 2013).

Table 2 presents the results of the reliability and convergent validity assessment for all reflective constructs in the model. All constructs satisfied the recommended internal consistency thresholds, with Cronbach’s Alpha and composite reliability (ρc) values exceeding the minimum acceptable level of 0.70, indicating strong reliability (Hair & Alamer, 2022). Outer loadings for all observed indicators ranged from 0.728 (EI2) to 0.944 (SC3), all surpassing the 0.708 threshold and confirming sufficient indicator reliability. Moreover, AVE values for all constructs exceeded 0.50, establishing convergent validity and indicating that each construct accounts for more than half of the variance in its respective indicators (Fornell & Larcker, 1981). Notably, SC and SN achieved the highest reliability indices (SC: CA = 0.912, ρc = 0.944, AVE = 0.850; SN: CA = 0.904, ρc = 0.939, AVE = 0.837), while PBC also showed strong reliability (CA = 0.918, ρc = 0.942, AVE = 0.803), reflecting the conceptual precision of these constructs’ operationalisation. VIF values ranged from 1.402 to 3.746, all falling below the conservative threshold of 5.0, confirming that collinearity does not threaten the stability of structural path estimates (Hair & Alamer, 2022). Collectively, these results provide robust empirical support for the reliability and convergent validity of the measurement model.

Discriminant validity was assessed using two complementary criteria: the Fornell-Larcker criterion and the Heterotrait-Monotrait Ratio (HTMT). Fornell-Larcker criterion: as shown in Table 3, the square root of each construct’s AVE (bold diagonal) exceeds its correlations with all other constructs, satisfying the Fornell-Larcker criterion for discriminant validity (Fornell & Larcker, 1981). HTMT criterion: HTMT ratios were computed to corroborate these findings ( Table 4). Unlike the Fornell-Larcker criterion, HTMT is estimated directly from the indicator correlation matrix and is not deflated by measurement error; consequently, HTMT values are systematically higher than inter-construct correlations and provide a more conservative test of discriminant validity (Henseler et al., 2015). Discriminant validity is considered adequate when HTMT remains below 0.90 for conceptually distinct constructs (Henseler et al., 2015), or below 0.85 when constructs are a priori expected to be highly similar (Cheah et al., 2018). All HTMT values in the present study fall below the 0.90 threshold, ranging from 0.060 (SN-SUP) to 0.872 (EDU-EI). The elevated HTMT for the EDU-EI pair (0.872) reflects the theoretically expected conceptual proximity between entrepreneurship education and entrepreneurial intention formation, constructs that are closely linked by definition within the integrated TPB-RBV framework, yet it remains within acceptable bounds. Taken together, results from both criteria provide adequate evidence of discriminant validity for all latent constructs in the model (Hair & Alamer, 2022).

Model fit was assessed using the global fit indices generated by SmartPLS ( Table 5). Both the saturated and estimated models yielded an SRMR value of 0.060, well below the recommended threshold of 0.08, indicating a good level of approximate model fit (Hair & Alamer, 2022). The discrepancy measures (d_ULS = 1.551 and d_G = 0.885) and the model chi-square (1607.844) are reported as supplementary fit information; given the large sample-size sensitivity of the chi-square statistic in PLS-SEM, these indices are interpreted cautiously and in conjunction with SRMR rather than in isolation (Hair & Alamer, 2022; Henseler et al., 2015). The Normed Fit Index (NFI) was 0.746, below the conventional benchmark of 0.90; however, in variance-based structural equation modelling, SRMR is generally considered a more informative indicator of global model fit than NFI, and NFI is known to be conservative and rarely reaches 0.90 in PLS-SEM applications with a comparatively large number of indicators (Hair & Alamer, 2022; Henseler et al., 2015). Collectively, these results indicate that the proposed integrated TPB - resource - ecosystem model provides an acceptable representation of the observed data and is appropriate for subsequent structural model evaluation.

Explanatory Power: The structural model’s explanatory power was evaluated using the coefficient of determination (R²) and effect sizes (f²) in accordance with (Hair & Alamer, 2022). The R² for entrepreneurial intention (EI) was 0.748, with an adjusted R² of 0.742, indicating that the seven predictors (ATT, SN, PBC, EDU, SUP, SC, and EXP) jointly explain 74.8% of the variance in EI, a level classified as substantial according to the benchmarks set out in Section 3.4 (R² ≈ 0.75; (Hair & Alamer, 2022).

Effect Sizes: Effect sizes (f²) were computed to assess the incremental contribution of each predictor to model explanatory power. Following (Cohen, 2013), thresholds of 0.02, 0.15, and 0.35 denote small, medium, and large effects, respectively. EDU recorded by far the largest effect size (f² = 0.281), approaching the large-effect threshold, followed by EXP (f² = 0.137) and PBC (f² = 0.124), both bordering the medium-effect threshold, and SUP (f² = 0.096) and ATT (f² = 0.047), which exhibited small effects. SN (f² = 0.010) and SC (f² = 0.001) contributed negligibly, falling below Cohen’s small-effect threshold ( Table 6). These findings suggest that entrepreneurial intention in this mountainous emerging economy context is principally driven by formal entrepreneurship education and practical readiness, rather than normative social pressure or social-network capital.

Predictive Relevance: PLSpredict: Out-of-sample predictive power was evaluated using PLSpredict (Shmueli et al., 2019), employing a 10-fold cross-validation procedure. PLSpredict compares the prediction errors of the PLS-SEM model against a naïve linear model (LM) benchmark and an indicator-average (IA) benchmark: if the PLS model’s root mean squared error (RMSE) is lower than that of the LM and IA benchmarks across indicators, the model demonstrates genuine predictive power beyond simple benchmarks (Shmueli et al., 2019). As reported in Table 7, the four EI indicators returned Q²predict values between 0.311 (EI1) and 0.529 (EI4), all well above zero and confirming out-of-sample predictive relevance for every indicator. PLS-SEM generated lower RMSE than the LM benchmark for three of the four indicators (EI1, EI2, EI3), and a comparable RMSE for EI4, while consistently outperforming the cruder IA benchmark across all four indicators. At the construct (latent-variable score) level, EI achieved a Q²predict of 0.703 with RMSE = 0.585 and MAE = 0.315, indicating strong aggregate predictive relevance. These results collectively confirm that the structural model possesses both in-sample explanatory power and out-of-sample predictive relevance (Hair & Alamer, 2022; Shmueli et al., 2019).

Hypothesis testing

The structural model was tested using bootstrapping with resampling (no sign change), as implemented in SmartPLS. Path coefficients (β), sample means (M), standard deviations (STDEV), t-statistics, and p-values are reported in Table 6 and discussed sequentially below. Approximate 95% confidence intervals are computed as β ± 1.96 × STDEV (normal approximation); a hypothesis is considered supported when this interval excludes zero and p < 0.05, consistent with current best-practice recommendations for PLS-SEM inference (Hair & Alamer, 2022; Henseler et al., 2015). The structural model was evaluated using SmartPLS bootstrapping with 5,000 resamples. Path coefficients, bootstrap t-statistics, and confidence intervals are presented in Table 8.

H1, which proposed that attitude toward entrepreneurship (ATT) would positively predict EI, was supported (β = 0.130, t = 2.344, p = 0.019, approx. 95% CI [0.022, 0.238]). The interval excludes zero, corroborating the statistical significance of this path, albeit with the smallest magnitude among the five significant predictors. This finding is consistent with the core proposition of TPB (Ajzen, 1991) and replicates prior evidence from Vietnamese and South Asian contexts (Dissanayake, 2014; Liñán & Chen, 2009), confirming that individuals who perceive entrepreneurship as an attractive and personally valuable career option form stronger entrepreneurial intentions, even though in this sample its relative contribution is modest compared with education and experience.

H2, which proposed a positive effect of subjective norms (SN) on EI, was not supported (β = 0.054, t = 0.920, p = 0.357, approx. 95% CI [−0.062, 0.170]). The confidence interval straddles zero, confirming the absence of a statistically significant effect, suggesting that perceived social pressure from family, peers, and community does not exert a direct influence on entrepreneurial intention in this sample. This finding contrasts with some evidence from collectivist societies (Hoang et al., 2020) but is consistent with emerging evidence indicating that younger Vietnamese generations increasingly prioritise personal agency and information-based judgment over social conformity in career decisions. The multi-ethnic composition of Lao Cai’s population may also contribute to this result: normative pressures vary considerably across communities and may be absorbed by more proximal predictors such as education and experience.

H3, which examined perceived behavioural control (PBC), was supported (β = 0.210, t = 4.173, p < 0.001, approx. 95% CI [0.112, 0.308]). The narrow, entirely positive confidence interval underscores the robustness of this effect, which is the third strongest in the model. The high level of statistical significance confirms that perceived self-efficacy and capacity to manage entrepreneurial challenges strongly reinforce intention, in accordance with TPB predictions (Ajzen, 1991; Liñán & Chen, 2009). This finding highlights the practical importance of capability-building interventions that strengthen young people’s confidence in undertaking entrepreneurial activity within resource-constrained environments.

H4, which proposed a positive effect of entrepreneurship education (EDU) on EI, was strongly supported and produced the largest path coefficient in the entire structural model (β = 0.348, t = 4.661, p < 0.001, approx. 95% CI [0.201, 0.495]). The confidence interval is narrow and entirely positive, indicating a stable, highly significant, and substantively large effect (f² = 0.281, approaching Cohen’s large-effect threshold). Access to structured entrepreneurship training, business planning curricula, and vocationally relevant knowledge measurably increases entrepreneurial motivation, consistent with the broader developing-country evidence base (Hoang et al., 2020; Nabi et al., 2017). In Lao Cai, where 41.7% of respondents have completed vocational training, education-based pathways represent the single most influential and scalable mechanism for intention development identified in this study, particularly when programmes integrate contextually relevant and market-oriented content.

H5, which examined the role of ecosystem support (SUP) on EI, was supported (β = 0.195, t = 3.712, p < 0.001, approx. 95% CI [0.093, 0.297]). The confidence interval is narrow and entirely positive, reflecting high bootstrap stability for this path. This finding confirms the enabling function of institutional support factors, encompassing government policy, seed financing access, incubation infrastructure, and business advisory networks, in activating and sustaining entrepreneurial motivation (Bui et al., 2018; Davidsson & Honig, 2003; Xu & Dobson, 2019). The robust predictive role of SUP in this provincial mountain context suggests that young people in peripheral economies are sensitive to institutional commitment signals, and that targeted policy interventions can meaningfully attenuate the intention–action gap.

H6, which proposed that social capital (SC) would positively influence EI, was not supported (β = 0.019, t = 0.362, p = 0.717, approx. 95% CI [−0.081, 0.119]). The confidence interval is wide and centred near zero, and the effect size is negligible (f² = 0.001), indicating that social capital, as operationalised in this study, exerts no discernible direct influence on entrepreneurial intention once education, experience, PBC, attitude, and ecosystem support are accounted for. This finding may reflect the dispersed and informal character of social networks in Lao Cai’s multi-ethnic rural communities, where social ties exist but may not yet function as effective channels for entrepreneurial resource mobilisation, or it may indicate that the influence of social capital operates indirectly, through experience or ecosystem support, rather than directly on intention.

H7, which examined experience and entrepreneurial readiness (EXP), was supported and generated the second largest path coefficient in the structural model (β = 0.242, t = 3.777, p < 0.001, approx. 95% CI [0.117, 0.367]). The confidence interval is narrow and entirely positive, confirming the robustness of this effect. This result identifies EXP as the second most influential driver of entrepreneurial intention in the study, corroborating the argument that prior exposure to business activity, concrete planning, and knowledge of start-up processes substantially reinforces motivational readiness to pursue entrepreneurship (Dissanayake, 2014; Hoang et al., 2020), operating alongside, rather than in place of, formal entrepreneurship education. The finding is particularly significant in the Lao Cai context, where approximately 30% of start-up projects fail due to insufficient experience and managerial capability (Lao Cai Youth Union, 2026), underscoring the urgency of embedding experiential learning within youth entrepreneurship support frameworks in mountainous regions.

To evaluate the conversion of entrepreneurial intention (EI) into actual entrepreneurial decision (ED), Binary Logistic Regression was employed, with ED coded as a binary outcome: ‘1’ for respondents who have initiated a business venture (Entrepreneur) and ‘0’ for those who have not yet done so (Non-entrepreneur) (Hosmer Jr et al., 2013). Latent variable scores for EI derived from the PLS-SEM estimation were used as the sole predictor in this second-stage analysis, consistent with the two-stage analytical approach described in Section 3.4. Model evaluation followed a comprehensive framework encompassing overall significance, pseudo-R² indices, classification performance disaggregated by class, and receiver operating characteristic (ROC) analysis, as recommended for logistic regression reporting in behavioural science (Hosmer Jr et al., 2013).

Overall model significance

The Omnibus Tests of Model Coefficients ( Table 9) confirm that the logistic regression model achieves overall statistical significance (χ² = 38.345, df = 1, p < 0.001), demonstrating that EI makes a meaningful and non-trivial contribution to predicting whether a young person in Lao Cai converts intention into actual entrepreneurial action. The block-0 (constant-only) model already classified the majority class correctly (Wald = 22.910, p < 0.001), while the score test for EI prior to entry (score = 17.434, df = 1, p < 0.001) confirmed that EI should be entered into the model. This result validates the core proposition of TPB that behavioural intention is the most proximal and powerful predictor of overt behaviour (Ajzen, 1991).

Conformity assessment

Pseudo-R² Indices

The model’s explanatory capacity is summarised through pseudo-R² indices in Table 10. SPSS output reports Cox & Snell R² = 0.117 and Nagelkerke R² = 0.161. McFadden’s R², not directly reported by SPSS, can be derived from the reported −2 log-likelihood of the fitted model (363.403) and the null-model deviance implied by the observed class proportions (196/307 = 63.8% “Entrepreneur”), yielding McFadden R² ≈ 0.095. Collectively, these indices indicate that entrepreneurial intention accounts for approximately 9.5–16.1% of the variance in the binary entrepreneurial decision outcome, depending on the index employed. Pseudo-R² values in logistic regression are inherently lower than their OLS counterparts and should not be interpreted against the same benchmarks; values of 0.09–0.16 are generally regarded as indicating acceptable explanatory power for a single-predictor model of a complex behavioural outcome (Hosmer Jr et al., 2013), with the remaining unexplained variance reflecting contextual factors, including access to financing, family obligations, and local market conditions, that are particularly salient in the mountainous frontier economy of Lao Cai.

Classification performance

The Classification Table ( Table 11) reports disaggregated prediction accuracy at the default cut value of 0.500. To provide a complete and transparent assessment, as required when class imbalance is present, four complementary metrics are reported alongside overall accuracy: sensitivity (true positive rate), specificity (true negative rate), positive predictive value (PPV), negative predictive value (NPV), and balanced accuracy.

The overall accuracy of 69.4% should be interpreted with caution given the class imbalance in the sample (63.8% entrepreneur), as this figure is partly inflated by the model’s strong sensitivity to the majority class. The sensitivity of 98.5% confirms that the model correctly identifies virtually all individuals who have made an actual entrepreneurial decision. However, the specificity of 18.0% reveals that the model correctly classifies only a small proportion of non-entrepreneurs, a limitation that is structurally expected in single-predictor intention-to-behaviour models under class imbalance (King & Zeng, 2001). The balanced accuracy of 58.2%, which treats both classes equally regardless of prevalence, provides a more conservative and honest summary of discriminative performance, only marginally above the chance level of 50%. The PPV of 68.0% indicates that approximately two-thirds of those predicted to be entrepreneurs have indeed made that decision, while the NPV of 87.0% confirms higher accuracy in identifying confirmed non-entrepreneurs among the (much smaller) set of cases predicted as such.

ROC analysis and AUC

To evaluate discriminative performance independently of the classification threshold, Receiver Operating Characteristic (ROC) analysis was conducted (Hanley & McNeil, 1982). The Area Under the Curve (AUC) was 0.604 (SE = 0.034, asymptotic p = 0.002, 95% CI [0.537, 0.671]) ( Table 12). Following the conventional benchmarks established by Hosmer and Lemeshow (2000), AUC 0.50–0.70 = poor/no better than chance-adjacent discrimination, 0.70–0.80 = acceptable, 0.80–0.90 = excellent, the present AUC of 0.604 indicates poor-to-modest discrimination between entrepreneurs and non-entrepreneurs across classification thresholds. The 95% confidence interval excludes 0.50, confirming that EI’s discriminative ability is statistically distinguishable from pure chance, but the magnitude of discrimination is weak. This pattern, a highly significant and substantively large odds ratio (Exp(B) = 8.472) combined with a comparatively low AUC, is not contradictory: the Wald test and odds ratio assess the strength of the linear association between EI and the log-odds of ED, whereas AUC assesses the model’s ability to rank-order and separate individual cases, which is additionally constrained here by the limited variance and ceiling effects in the EI latent score among the large majority of already-active entrepreneurs (63.8% of the sample). The result indicates that, although entrepreneurial intention is a statistically robust and theoretically central gateway predictor, it is, on its own, an incomplete classifier of who ultimately starts a business; additional contextual predictors are required to materially improve case-level discrimination.

Regression coefficients and H8 test

Regression coefficient estimates are presented in Table 13. EI exerts a positive and highly significant effect on entrepreneurial decision (B = 2.137, S.E. = 0.567, Wald = 14.211, df = 1, p < 0.001), thereby supporting H8. The odds ratio Exp(B) = 8.472 indicates that for each one-unit increase in entrepreneurial intention score, the odds of having made an actual entrepreneurial decision increase by approximately 8.47 times, ceteris paribus. This constitutes a substantively large effect on the odds scale, establishing EI as a critical gateway mechanism through which cognitive readiness is converted into economic action, even though, as the ROC analysis above demonstrates, this strong association coexists with comparatively modest case-level discriminative power. The finding replicates and extends evidence from (Krueger Jr et al., 2000) and (Kolvereid & Isaksen, 2006), who similarly identified intention as the most direct behavioural antecedent, and is consistent with (Ajzen, 1991) foundational claim that the stronger the intention to engage in a behaviour, the more likely its performance.

The primacy of formal entrepreneurship education in a peripheral entrepreneurship ecosystem. The finding that entrepreneurship education (EDU) emerged as the single strongest predictor of entrepreneurial intention (β = 0.348, f² = 0.281), ahead of experience and readiness (EXP, β = 0.242) and perceived behavioural control (PBC, β = 0.210), invites a theoretical reappraisal of how human capital operates in spatially and institutionally peripheral contexts. In well-resourced urban economies, formal education and informal experience are often found to contribute comparably to entrepreneurial cognition (Van Praag & Versloot, 2007). In Lao Cai, where formal credit markets are shallow, business registration processes remain opaque for ethnic minority youth, and locally accessible role models are scarce, structured entrepreneurship training appears to function as the primary channel through which abstract entrepreneurial knowledge is converted into actionable cognitive readiness, with prior hands-on experience reinforcing, rather than substituting for, this educational channel. The theoretical implication is that human capital theory (Wuttaphan, 2017) requires contextual calibration: in mountainous, institutionally thin settings, the productive returns to formal education on intention formation may exceed those typically reported in metropolitan TPB studies, where education effects are sometimes attenuated by saturation.

Why subjective norms and social capital both failed: a generational and structural argument. The non-significance of both subjective norms (SN: β = 0.054, p = 0.357) and social capital (SC: β = 0.019, p = 0.717) is the finding most in tension with the dominant TPB and resource-based evidence base in East and Southeast Asian collectivist societies (Gloor et al., 2009; Hoang et al., 2020; Liñán & Chen, 2009), and it warrants substantive theoretical explanation rather than dismissal. Several non-exclusive mechanisms are proposed. First, Lao Cai’s population comprises 27 co-resident ethnic groups with fundamentally different normative systems regarding economic risk, family obligation, and occupational prestige; aggregate SN measurement likely obscures contradictory pressures that cancel out at the sample level. Second, the respondents in this study are predominantly young adults who came of age during the post−2015 expansion of smartphone connectivity and social media in mountainous Vietnam, a cohort for whom information-based peer comparison and formal knowledge increasingly substitute for elder-authority deference and informal network ties in career decisions, a pattern of individualisation documented in other rapidly urbanising emerging economies (de Koning et al., 2016). Third, the very low effect size for social capital (f² = 0.001) suggests that, in this context, social ties may operate as a precondition for entering the sample (most respondents likely already have some entrepreneurial contacts) rather than as a differentiator of intention strength once education, experience, attitude, and ecosystem support are accounted for; alternatively, its influence may be substantially mediated through experience and ecosystem support rather than acting directly on intention. Future research using multi-group SEM across ethnic subgroups in Lao Cai and formal mediation testing of SC’s indirect paths would provide a direct test of these hypotheses.

Institutional support and experience as joint amplifiers beyond core TPB. The significant and sizeable contributions of ecosystem support (SUP: β = 0.195, p < 0.001) and experience and readiness (EXP: β = 0.242, p < 0.001) cannot be accommodated within standard TPB, which treats perceived behavioural control as the sole structural antecedent capturing environmental enablers. This finding provides empirical grounding for extending TPB with resource-based and institutional-embeddedness arguments drawn from the social-cognitive entrepreneurship framework (Davidsson & Honig, 2003): young people’s entrepreneurial intentions in Lao Cai are jointly calibrated against what they have been formally taught (EDU), what they have directly experienced (EXP), what they believe they can control (PBC), and the institutional commitment signals they perceive (SUP), with the family/peer/community normative channel (SN) and informal social-network channel (SC) playing a comparatively marginal direct role. The implication for theory is that, in peripheral and transitional economies, education and ecosystem support should be modelled as co-equal structural antecedents of intention formation, alongside experience, rather than as background or post-intention factors.

The intention-to-action gap: a robust gateway predictor with limited classifier power. The logistic-stage results present a theoretically important duality. On one hand, entrepreneurial intention exerts a highly significant and substantively large effect on the odds of entrepreneurial decision (Exp(B) = 8.472, p < 0.001), an odds ratio approximately twice the meta-analytic benchmark of 2–4 typically reported for intention-behaviour conversion in entrepreneurship (Kautonen et al., 2015; Krueger Jr et al., 2000). On the other hand, the model’s overall discriminative ability, as measured by AUC (0.604, 95% CI [0.537, 0.671]), is only modestly above chance and falls in the “poor” range by conventional benchmarks (Hosmer Jr et al., 2013). These two results are statistically compatible rather than contradictory: a large, significant odds ratio establishes that EI is associated with substantially higher odds of having become an entrepreneur, while a low AUC indicates that EI alone cannot reliably rank-order or separate individual cases, in part because the great majority of the sample (63.8%) already comprises active entrepreneurs whose EI scores are concentrated at the upper end of the distribution, compressing the discriminative range available to a single continuous predictor. This pattern is consistent with the Rubicon model of action phases (Gollwitzer, 1990), which distinguishes motivational from volitional processes, and suggests that in peripheral economies the motivational-to-volitional transition is influenced by a substantial set of additional, unmeasured contextual factors (access to capital, family obligations, local market opportunity) that intention alone does not capture at the level of individual case prediction, even though it remains a powerful average-effect predictor.

Theoretical contribution: a contextually calibrated extension of TPB. Taken together, the findings, the dominance of formal entrepreneurship education over experience and core TPB components, the joint collapse of subjective norms and social capital, the amplifying but non-substitutable role of institutional support, and the coexistence of a strong odds ratio with weak case-level discrimination, converge on a coherent theoretical argument: TPB, as specified for mainstream urban samples, systematically misestimates the relative weights of its components, and the relative power of resource-based variables, when applied to peripheral, multi-ethnic, institutionally thin contexts. The present study’s extended model, by integrating formal human capital (EDU), experiential human capital (EXP), and institutional embeddedness (SUP) alongside TPB core components, advances a more ecologically valid specification for mountainous emerging economy settings, and contributes a replicable two-stage analytical template combining PLS-SEM and Binary Logistic Regression, including transparent reporting of both odds-ratio strength and classifier discrimination, for contexts where the ultimate behavioural outcome is a discrete event.

Prioritising formal entrepreneurship education while sustaining experiential channels. The dominance of EDU (β = 0.348) over EXP (β = 0.242) implies that, contrary to a “learning-by-doing-only” policy logic, structured curricula, business-planning training, and institution-delivered entrepreneurship courses remain the single most leveraged policy instrument for building entrepreneurial intention in Lao Cai. This does not diminish the continued importance of experience (the second-strongest predictor): the two should be designed as complementary tracks, embedding paid apprenticeship and project-based learning opportunities directly within formal entrepreneurship education programmes across highland agriculture, medicinal herb processing, eco-tourism, and community handicraft sectors, industries where Lao Cai has genuine competitive assets and where curriculum content and direct practice can reinforce one another.

Why ecosystem support requires provincial-level commitment, not project-level allocation. The significant, structural contribution of ecosystem support (SUP) to intention formation implies that episodic, project-funded support schemes are insufficient. The signal value of institutional commitment depends on its perceived permanence and scope: a three-year pilot incubation centre communicates far less legitimacy than a legally established, budget-sustained provincial youth entrepreneurship infrastructure. Lao Cai provincial authorities should therefore consider embedding entrepreneurship support commitments into provincial economic development law and multi-year budget frameworks, rather than treating them as discretionary spending contingent on central government project cycles.

Social capital and subjective norms as medium-term, indirect investments rather than short-term levers. The non-significance of both social capital (SC) and subjective norms (SN) should not be read as evidence that social networks and family/community endorsement are irrelevant, but rather that, in their current configuration in Lao Cai’s multi-ethnic rural communities, they do not directly drive additional intention once education, experience, control, and ecosystem support are accounted for. The appropriate policy response is not to abandon social capital and normative-engagement initiatives but to redesign them as enablers that operate through education, experience, and ecosystem-support channels, for example, structured inter-ethnic youth entrepreneurship networks that create weak ties (Granovetter, 1983) feeding directly into apprenticeship placement and incubator referral, rather than expecting awareness or endorsement campaigns alone to move intention.

Translating a strong odds ratio with weak classification into a two-track policy response. The combination of a large odds ratio (8.47×) and modest AUC (0.604) implies a two-track policy design: first, broad-based intention-building investment (education, experience, ecosystem support) that reliably raises the average propensity to act, consistent with the strong odds-ratio finding; and second, recognition that intention alone cannot identify, at the individual level, which specific youth are most likely to convert, so policy should pair intention-building programmes with complementary screening criteria, such as access to seed capital, family financial buffer, and local market opportunity, to better target intensive support resources toward those most likely to successfully cross the decision threshold.

Causal identification. The cross-sectional design means that the path coefficients represent predictive associations rather than experimentally identified causal effects. The extended TPB framework tested here is theoretically motivated, but the data cannot rule out reverse causation, for instance, that individuals who have already decided to pursue entrepreneurship retrospectively report higher perceived ecosystem support or education relevance, inflating the SUP and EDU coefficients. Longitudinal panel designs tracking the same individuals from pre-intention measurement through the decision point, or quasi-experimental evaluations exploiting policy variation across districts or time periods, are needed to establish causal ordering with greater confidence.

The generalisability boundary. Lao Cai is a theoretically productive research site precisely because it combines characteristics, ethnic diversity, peripheral geography, thin institutional infrastructure, post-subsistence economic transition, found across mountainous provinces in Vietnam and comparable economies in Laos, Myanmar, Yunnan (China), and the upland regions of Cambodia and Thailand. However, the quantitative parameter estimates, particularly the dominance of EDU and the joint collapse of SN and SC, should not be generalised without replication in other sites. Multi-province comparative designs, ideally using the same instrument, would allow decomposition of within-province versus cross-province variation in the determinant structure of entrepreneurial intention.

Classifier and measurement limitations. The modest AUC obtained for the single-predictor logistic model (0.604) indicates that EI alone has limited individual-level classification value; future studies should incorporate additional predictors (access to financing, risk perception, family support, market opportunity) into the second-stage model to materially improve discrimination, and should report the Hosmer–Lemeshow goodness-of-fit statistic, which was not available in the present SPSS output and should be obtained and reported in any future or revised analysis. The entrepreneurial decision variable was further operationalised using self-reported binary data, which introduces recall bias in retrospective attribution of the decision point and social desirability inflation of “entrepreneur” self-identification in a survey administered under provincial institutional auspices. Future research should cross-validate against administrative business registration records held by Lao Cai’s Department of Planning and Investment, which would also enable survival analysis of the time elapsed between intention formation and formal registration.

Unmodelled heterogeneity. The model does not distinguish among venture types, agricultural, tourism, e-commerce, traditional craft, cross-border trade, that are likely to have distinct determinant profiles in Lao Cai’s diversified highland economy. Risk perception, digital access, and ethnic cultural factors were also excluded. Future research employing latent class analysis or multi-group SEM across ethnic subgroups and venture types would substantially advance contextual specificity, and may reveal that the joint non-significance of SN and SC is not uniform but is driven by specific subgroups for whom normative or network decoupling has already occurred.

Comparison with GEM benchmarks. The present study’s sample is not directly comparable to GEM’s national probability samples, which limits cross-study benchmarking. Future research should align sampling and measurement protocols more closely with GEM’s Total early-stage Entrepreneurial Activity (TEA) and Entrepreneurial Employee Activity (EEA) indices, enabling Lao Cai’s findings to be situated within the global comparative evidence base and tracked over time as institutional conditions evolve.