Internet Trust: Longitudinal Evidence on Socio-Economic and Digital Adoption Behavior

Page 1 of 1 Internet Trust: Longitudinal Evidence on Socio-Economic and Digital Adoption Behavior Angel Valarezo-Unda✉ Javier Capilla‡ Teodosio Pérez-Amaral¶ Alfredo Garcia-Hiernaux§ Rafael López† Abstract ✉DEAEH and ICAE, UCM. Corresponding author. Email: [email protected] ‡Banco de Santander and UCM. ¶DANAE and ICAE, UCM. §DANAE and ICAE, UCM. †DANAE and ICAE, UCM. This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5528118 Preprint not peer reviewed Page 1 of 26 Internet Trust: Longitudinal Evidence on Socio-Economic and Digital Adoption Behavior Abstract Using the 2014 - 2021 waves of Spain’s ICT-H household panel, we track Internet trust for 59,648 internet users across 130,013 person-year observations. A correlated random-effects ordered-probit model shows that improvements in digital skills and first-hand use of transactional services are the strongest predictors of higher trust, while traditional socio-economic markers play a secondary role. Once skill levels are controlled, age differences largely vanish, but women still report lower trust, and the pandemic years register a notable dip, pointing to attitudinal and systemic factors that skill policies alone cannot solve. The findings highlight the need for advanced skill training, guided initial transactions, and robust consumer safeguards, particularly for women and low-income users, if Spain is to close its remaining trust gap and achieve inclusive digitalization. JEL classification: C33, D83, L86, O33, O35. Keywords: Internet trust; digital inclusion; socio-economic determinants; panel data; digital skills; Eco-RETINA; Spain; ICT-H survey The authors acknowledge the support of the HORIZON Research and Innovation Program of the European Union, under grant agreement No 101120657, project ENFIELD (European Lighthouse to Manifest Trustworthy and Green AI). ✉DEAEH and ICAE, UCM. Corresponding author. Email: [email protected] ‡Banco de Santander and UCM. ¶DANAE and ICAE, UCM. §DANAE and ICAE, UCM. †DANAE and ICAE, UCM. This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5528118 Preprint not peer reviewed Page 2 of 26 1. Introduction As digital technologies become increasingly embedded in daily life, trust in the internet (measured as described in Section 3) has emerged as a critical factor influencing individuals’ willingness to engage with digital platforms, share personal data, and adopt online services. The concept of trust in this context refers to users’ perceptions of safety, reliability, and confidence in online interactions and infrastructures (Gefen, 2000; McKnight et al., 2002). Without trust, users may abstain from online transactions, limit their participation in digital services, or disengage from technological innovation altogether. Although internet access and connectivity have steadily improved in most countries, recent data point to a troubling trend: trust in the internet is eroding. As shown in Figure 1, the share of internet users reporting high levels of trust in the internet in Spain has declined consistently from 12% in 2014 to just 4% in 2021. In contrast, the proportion of users expressing low trust increased from 33% to 41% over the same period. Medium trust remains the most common response but has not absorbed the decline in high trust. This pattern reveals a growing polarization, suggesting that for many users, familiarity with digital services has not translated into increased confidence. Figure 1. Trust in the Internet. Percentage of the internet users Source: INE, 2025 From an economic perspective, trust plays a foundational role in environments characterized by information asymmetry and limited verifiability. Akerlof (1970) classic model of the market for "lemons" demonstrates how markets can fail when consumers cannot assess product quality, a dynamic that is particularly relevant in online environments. Varian (2010) elaborates this idea within the context 33 32 34 35 35 35 44 41 55 60 58 58 58 59 52 55 12 8 8 8 7 7 4 4 0 10 20 30 40 50 60 70 2014 2015 2016 2017 2018 2019 2020 2021 Low Medium High This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5528118 Preprint not peer reviewed Page 3 of 26 of computer-mediated transactions, in which trust must be mediated not by personal relationships or inperson inspection, but by technical safeguards, reputation systems, and institutional signals of reliability. A growing body of empirical research has examined the determinants of internet trust, highlighting the role of website quality, security assurances, institutional reputation, and individual-level variables such as age, education, and digital literacy (Corritore et al., 2003; Kim et al., 2008). However, most of these studies rely on cross-sectional data and therefore offer a limited view of how trust evolves over time. This temporal dynamic is critical, particularly in the context of Spain, where user distrust appears to be intensifying despite increasing digital integration. This paper aims to address these gaps using a rich longitudinal dataset from the Spanish National Statistical Institute’s ICT-H household survey. Covering over 262,000 observations from more than 115,000 individuals, the data allow us to explore the socio-economic, demographic, and behavioral determinants of trust in the internet and its evolution over time. Leveraging eight years of longitudinal microdata and a correlated random-effects ordered-probit framework, this study provides a nuanced picture of the factors that shape Internet trust in Spain. The results clarify how digital skills, transactional experience, and persistent socio-economic gaps interact, yielding policy-relevant guidance for narrowing digital disparities and strengthening user confidence in online environments. The remainder of the paper is structured as follows. Section 2 reviews the existing literature on internet trust and digital inclusion. Section 3 presents the data and main variables. Section 4 describes the methodology. Section 5 presents the main empirical results. Section 6 discusses the results, policy implications, and future research avenues. Section 7 concludes. 2. Literature Review Existing work on Internet trust spans three streams. Information-systems research identifies psychological disposition, institutional safeguards, and transactional familiarity as its core ingredients (Gefen, 2000; McKnight et al., 2002). Economic theory views trust as the mechanism that offsets online information asymmetries, substituting physical inspection with encryption, certification, and reputation systems (Akerlof, 1970; Varian, 2010). Empirical studies, both international and Spanish, show higher trust among users facing better platform quality, security cues, higher education, income, and digital skills (Garín-Muñoz et al., 2019; Kim et al., 2008; Lera-López et al., 2011; Valarezo et al., 2018), yet they mostly rely on cross-sectional data and cannot track how trust evolves. By exploiting Spain’s multiwave ICT-H panel, the present study provides the first longitudinal evidence on the dynamics and determinants of Internet trust. This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5528118 Preprint not peer reviewed Page 4 of 26 2.1 Theoretical Foundations of Internet Trust Internet trust is a multifaceted concept encompassing users’ beliefs about the reliability, integrity, and competence of digital services and platforms. McKnight et al. (2002) developed a comprehensive typology of trust in e-commerce, identifying three primary dimensions: disposition to trust, institutional trust, and interpersonal trust. These components influence users' decisions to engage in online interactions, particularly in contexts where uncertainty is high and direct verification is limited. Gefen, (2000) further emphasized the role of familiarity and perceived ease of use as precursors to trust in e-commerce. Users are more likely to trust platforms that offer intuitive interfaces, transparent procedures, and consistent experiences. Trust formation, therefore, is not only psychological but also experiential and transactional, building over time as users interact with digital services. 2.2 Economic Perspectives: Asymmetric Information and Trust Mechanisms In economics, trust has long been associated with the resolution of information asymmetries. Akerlof, (1970) theory on the market for "lemons" illustrates how the absence of reliable information leads to market inefficiencies. His work implies that in markets where users cannot distinguish between highand low-quality services, such as digital platforms, trust becomes essential to facilitate exchange. Varian, (2010) extended this logic to the realm of computer-mediated transactions. In online markets, traditional trust-building mechanisms like personal inspection or repeated interactions are replaced by algorithmic processes. Trust is increasingly constructed through technological and institutional safeguards: encryption protocols, third-party certifications, user reviews, and platform guarantees. These digital trust mechanisms serve as substitutes for interpersonal trust, and their design has direct implications for market participation and digital inclusion. 2.3 Empirical Determinants of Internet Trust Empirical studies have identified a range of individual and contextual factors that influence levels of trust in the internet. Kim et al. (2008) illustrate that trust in online transactions is shaped by perceived website quality, the presence of security assurances, and institutional reputation. Similarly, Corritore et al. (2003) developed a model in which trust arises from a combination of user interface design, user experience, and social context. Socio-demographic characteristics also play a central role. Research shows that younger individuals, those with higher levels of education, and those more frequently engaged with digital technologies tend to report higher levels of internet trust (Gefen, 2000; Kim et al., 2008). Conversely, users in rural areas or with limited digital literacy are more likely to express distrust, highlighting the intersection between trust and digital divides. This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5528118 Preprint not peer reviewed Page 5 of 26 Despite the contributions of these studies, most empirical analyses rely on cross-sectional data. As a result, they offer limited insight into how trust evolves over time or how behavioral factors such as frequency of internet use or exposure to online risks may influence trust trajectories. There is a clear gap in the literature regarding longitudinal approaches that can capture the dynamic nature of trust and its interaction with digital behavior. 2.4 Empirical Evidence from Spain Using ICT-H Survey Data Several studies have utilized the Spanish ICT-H survey to explore internet usage and trust. Lera-López et al. (2011) analyzed the impact of socio-economic, demographic, and regional factors on internet use and frequency, employing binomial and ordered probit models with a Heckman two-stage estimation procedure. Their findings indicate that internet use is mainly associated with education, age, occupation, employment in the service sector, nationality, urban areas, and regional GDP per capita. Frequency of internet usage is positively related to broadband connection, education, internet skills, the methods through which internet skills are acquired, gender, and population size. Pérez-Hernández & Sánchez-Mangas (2011) studied online shopping in conjunction with having internet at home, using the ICT-H survey for the period 2004–2009. They employed pooled individual data and found that higher education and income levels are significant predictors of online shopping behavior. Garín-Muñoz & Pérez-Amaral, (2011) focused on factors affecting e-commerce use in Spain, utilizing cross-sectional data from the ICT-H survey. They identified that trust in the internet, along with sociodemographic variables such as age and education, significantly influences the likelihood of engaging in e-commerce activities. Valarezo et al. (2018) examined the drivers and barriers of Spanish individual consumer adoption of cross-border e-commerce for private use. Their study highlighted the importance of trust in the product and supplier, as well as trust in the channel, as significant factors influencing online purchasing decisions. Pérez-Amaral et al. (2021) analyzed digital divides across consumers of internet services in Spain using panel data from 2007 to 2019. They found that while some digital gaps are narrowing, disparities persist in areas such as age, education, and digital skills, affecting the equitable adoption of internet services. Fernández-Bonilla et al. (2022) identified and estimated the determinants for participating in ecommerce and developing e-trust, as well as the importance of e-trust for e-commerce in Spain. Utilizing a national survey from 2014 to 2019 and implementing a logit model, they concluded that eThis preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5528118 Preprint not peer reviewed Page 6 of 26 trust is a determining factor in e-commerce, and the improvement of equality education helps the growth of online commerce and e-trust. In turn, e-trust encourages the use of all digital resources. These studies underscore the importance of trust and socio-demographic factors in shaping internet usage behaviors in Spain. However, they primarily rely on cross-sectional or pooled data, limiting the ability to observe changes over time. 2.5 Contribution of This Study This study focuses on analysing trust and exploits the eight-wave ICT-H household panel (2014-2021) of the Spanish National Statistical Institute to provide the first longitudinal assessment of Internet trust in Spain. Using a correlated random-effects ordered-probit model, it distinguishes between permanent socio-economic differences and within-person changes, showing, for example, how trust increases when individuals improve their digital skills or adopt transactional online services. Digital-skill quartiles are constructed from twenty-two observed ICT activities that follow the EU competence framework, and usage variables are divided into transactional and informational categories to test alternative trust-formation mechanisms. By quantifying the relative influence of skills, income, and experiential learning, the paper offers evidence that can guide policies aimed at narrowing the trust gap that continues to hinder digital inclusion. 3. Data and Descriptive Statistics This study draws upon annual panel data obtained from the Encuesta sobre Equipamiento y Uso de Tecnologías de Información y Comunicación en los Hogares (Households Survey on Equipment and Use of ICT), conducted by the Spanish National Statistical Institute (INE, 2024), following Eurostat guidelines. The dataset spans the period from 2014 to 2021 and is representative at both national and regional levels, with a built-in elevation factor to ensure statistical validity. The survey’s main objective is to monitor the adoption and usage of ICT services and technologies by individuals and households. Coordinated by Eurostat, the survey is harmonized across EU member states and is made publicly available on an annual basis. The original sample design comprises a rotating panel of 18,000 to 21,000 dwellings per year. Each selected household may be interviewed for up to four consecutive years, with an annual refreshment rate of approximately 30%. Data collection is conducted through a mixed-mode approach, consisting of roughly 60% telephone interviews and 40% face-to-face interactions. While the survey unit is the dwelling, our study aims to track individuals across survey years. Therefore, special care was taken to identify repeated individuals over time, accounting for the fact that the same household might not report the same respondents every year. This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5528118 Preprint not peer reviewed Page 7 of 26 To create a coherent panel structure, a meticulous harmonization procedure was applied to the raw annual microdata. This process ensured consistency in variable definitions across 8 survey waves (eight survey years, but rotation limits each individual to four interviews) and enabled the identification of unique individuals across years. The result is a consolidated longitudinal dataset that allows for robust panel-data econometric analysis. Detailed documentation of this data cleaning and panel construction process is provided in Pérez-Amaral et al. (2021). The raw panel contains 130,013 person-year observations for 59,648 individuals, with each respondent observed in up to four years; 28,724 individuals appear once, 6,827 appear twice, 5,575 appear three times, and 6,938 appear four times. Table 1 summarizes the key explanatory variables used in the analysis, grouped into three main domains: sociodemographic, individual, and economic. Sociodemographic variables include gender and age (categorized into six cohorts). Individual-level attributes include educational attainment, digital skills, level of trust in the internet (the dependent variable), and employment status. Economic standing is proxied by household net monthly income, grouped into four income brackets. Three variables merit further elaboration due to their methodological construction. First, Internet trust (INT_TRUST). The ICT-H questionnaire measures general confidence in the Internet with the item “En general, por favor, indique su grado de confianza en Internet” (“In general, please indicate your level of confidence in the Internet”). Respondents choose one of three ordered responses: “Poco o nada” (little or none), “Bastante” (enough), “Mucho” (high), which we code as 1, 2 and 3. This operationalization follows the notion of “confident expectation of reliability and benevolence in an electronic medium” proposed by McKnight et al. (2002). Second, household income (INCOME) is entered as four ordered brackets (Low, Medium, High and Very-high), defined each survey year from the original bands (below €900, €900–1600, €1600–3000, above €3000). Using ordinal dummies keeps the variable comparable over time because it captures a respondent’s relative position in the annual income distribution. Also, aggregate price movements are absorbed by the year-specific cut-offs and by the set of year dummies included later in the regressions. This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5528118 Preprint not peer reviewed Page 8 of 26 Table 1. Variable dimensions and categories Variable Categories/coding Gender 2 groups: Male, Female Age 6 ordered groups: 16–24, 25–34, 35–44, 45–54, 55–64, 65+ (base) Dimension Sociodemographic Habitat (population size) 3 groups: Densely populated, Intermediate, thinly populated Education 4 ordered groups: None/Primary (base), Secondary, Bachelor’s, Master/PhD Digital Skills 4 ordered groups: Low (base), Medium, High, Very High Internet Trust (outcome) 3 ordered levels: Low (1), Medium (2), High (3) Individual Employment Situation 6 groups: Employed (base), Unemployed, Retired, Student, Housekeeper, Other Economic Income (household net) 4 groups: Low (base), Medium, High, Very High E‑mail use Binary: 1 if e‑mail in last 3 m; 0 otherwise Social‑network use Binary: 1 if participated in social networks; 0 otherwise E‑commerce Binary: 1 if bought online in the last 12 m; 0 otherwise E‑banking Binary: 1 if used Internet banking; 0 otherwise E‑government Binary: 1 if interacted with public services online; 0 otherwise E‑health Binary: 1 if sought health info online; 0 otherwise E‑tourism Binary: 1 if booked travel services online; 0 otherwise Behavioral E‑learning Binary: 1 if used online learning resources; 0 otherwise Cloud storage Binary: 1 if saved files on Internet storage; 0 otherwise Temporal Year Eight dummy variables: 2014 (base) to 2021 Third, the Digital Skills (DIGITAL_SKILLS) variable. The survey does not ask respondents to self‑rate their digital proficiency. Instead, it records whether they performed a list of twenty‑two ICT‑related activities in the previous twelve months. These items correspond to the European Commission’s digital competence framework (European Commission 2019) and are available for every wave from 2014 to 2021. We regroup them into four functional dimensions. Table 2. Digital Skills Dimensions Dimension Illustrative activities (items) Information Copy/move files; cloud storage; public‑service search; product comparison; health search (5) Communication E‑mail; social networking; VoIP/video calls; upload self‑created content (4) Problem‑solving Transfer files between devices; install apps; change security settings; buy/sell on‑line; internet banking; e‑learning (6) Software for content Word processing; spreadsheets; photo/audio editing; multimedia presentations; advanced spreadsheet functions; coding (7) For each respondent‑year observation, we compute the share of activities undertaken, SkillShare = (number of activities performed)/22. The distribution is then divided into annual quartiles, yielding a This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5528118 Preprint not peer reviewed Page 15 of 26 mechanism posited by McKnight et al. (2002) and echoes cross-sectional results in Kim et al. (2008). The sizeable trust premium earned by transactional service use (e-commerce, e-banking) suggests that positive online experiences reinforce confidence, whereas informational services (e-government, elearning) exert weaker effects, an asymmetry also observed in the OECD’s Digital Economy Outlook (2024). The gender gap persists even after controlling for skills and behavior, indicating that attitudinal or riskperception differences (Gefen, 2000) remain to be addressed through targeted digital safety and awareness campaigns. Conversely, the raw age penalty disappears once heterogeneity in skills and adoption is absorbed, implying that chronological age per se is not a barrier; what matters is the skill set and usage repertoire typically clustered by cohort. Income retains an independent effect, consistent with theories linking economic resources to greater tolerance for uncertainty (Akerlof, 1970). Improving digital trust is therefore not solely a question of up-skilling: programmes that lower the perceived stakes of occasional on-line loss (for example, through robust consumer redress mechanisms) would particularly benefit low-income households. Finally, the sharp downturn in 2020–2021 aligns with global surveys reporting heightened privacy concerns during the pandemic (OECD, 2024). Mitigating these macro-level shocks requires systemlevel safeguards (clear privacy policies, visible certification, and responsive enforcement) beyond individual-level interventions. 7. Conclusion We examined over 130,000 dwellings along eight survey waves (2014-2021) using a random-effects ordered-probit approach that accounts for correlated individual heterogeneity. The rotating-panel design, combined with Gauss-Hermite quadrature and clustered standard errors, allowed us to exploit both within and between-person variation while accommodating the dataset’s imbalance. Results converge on three core findings. First, digital capability is the strongest predictor of Internet trust: moving from low to very-high skills raises the latent index by roughly half a unit, about twice the size of any socio-economic effect. Second, economic resources still matter even after skills and behaviour are controlled, very-high-income respondents enjoy an additional +0.30 units of trust. Third, a persistent gender gap of about −0.10 units indicates that women perceive greater online risk. Other patterns are secondary yet revealing. Age penalties evident in a naïve model vanish once skills and usage are added, implying that cohort differences largely reflect capability gaps. Transactional engagement (ecommerce, e-banking, social media) delivers sizable “trust premia,” whereas informational services such as e-government do not. Finally, calendar-year dummies indicate a steady decline in trust from This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5528118 Preprint not peer reviewed Page 16 of 26 2015 and a pandemic-era plunge in 2020, suggesting that macro-level shocks can erode confidence even among skilled users. These findings suggest a three-pronged policy response. First, priority should go to scaling advanced digital-skill programmes, especially for low-income and female users, because upgrading competence yields the largest marginal gains in trust. Second, initial on-line experiences must feel safe and rewarding. Subsidised buyer-protection schemes, low-value voucher campaigns, and “sandbox” environments can turn first-time transactions into confidence-building landmarks. Third, systemic safeguards (visible privacy seals, rapid breach notifications, and vigorous enforcement of dataprotection rules) are needed to buffer the population against trust-eroding shocks that no amount of individual skill can offset. The study is not without limitations. Because the rotating panel provides, on average, only two observations per person, we cannot map complete life-cycle trajectories of trust. Causal claims rely on conditional exogeneity, so future work could exploit natural experiments (for example, phased fibre roll-outs or sudden privacy-regulation changes) to sharpen identification. Longer administrative panels would also permit richer dynamics, while research on spill-overs, such as whether gains in e-commerce trust translate into greater e-government use, could inform integrated digital-inclusion strategies. Overall, our evidence suggests that Internet trust is less reliant on immutable demographics and more on capability, experience, and system integrity; combining these factors is crucial for unlocking the full social and economic benefits of digital participation. Policy implications Skill development first. Investments in advanced digital-skill training can yield the most significant trust gains. Experiential onboarding. Encouraging first-hand, value-adding use cases (e.g., small, protected ecommerce transactions) may create a virtuous circle of use and trust. Targeted support. Tailored programmes for women and low-income users may be useful to close persistent attitudinal gaps. Structural assurance. Transparent and easily verifiable security and redress mechanisms can mitigate the negative impact of macroeconomic shocks. Limitations and future work The rotating-panel design limits the average observation window to two waves, restricting our ability to model long-run dynamics. Future studies could leverage administrative records or device-based This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5528118 Preprint not peer reviewed Page 17 of 26 panels with longer follow-up. Moreover, causal identification would benefit from exogenous shocks to broadband quality or privacy regulation rollout, enabling quasi-experimental estimation. A complementary avenue is to combine the theory-driven core with predictive analytics that can screen a larger pool of candidate inputs. The “Eco-RETINA” regression framework developed by Capilla et al. (2025) achieves high-quality out-of-sample forecasts while keeping computing costs and its environmental footprint low. Applying that tool as a preliminary filter or an external validation step could sharpen variable prioritization and provide policy-oriented prediction benchmarks without sacrificing the explanatory clarity of the econometric models. Funding: This research is funded by the Ministry of Science and Innovation - Spain. Project: Tecnología, Capital Humano, Innovación, Comercio y Recursos Naturales (PID2022-138706NB-I00 THUITER, 2023 – 2026), and the European Commission (ENFIELD Project. Grant agreement ID: 101120657). This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5528118 Preprint not peer reviewed Page 18 of 26 APPENDIX A1. Descriptive Statistics and Panel Decomposition Table 5. Panel Descriptive Statistics Variable Component Mean Std. dev. Min Max Obs. INT_TRUST overall 1.71 .59 1 3 N = 92,964 between .52 1 3 n = 45,077 within .34 0.21 3.21 T = 2.06 AGE overall 4.26 1.58 1 6 N = 130,013 between 1.61 1 6 n = 59,648 within .21 .51 7.26 T = 2.18 GENDER overall 1.45 .50 1 2 N = 130,013 between .49 1 2 n = 59,648 within .08 .70 2.20 T = 2.18 EDUCATION overall 2.14 .96 1 4 N = 129,746 between .92 1 4 n = 59,598 within .27 -0.11 4.39 T = 2.18 INCOME overall 2.21 .94 1 4 N = 105,036 between .89 1 4 n = 52,973 within .34 -0.04 4.46 T = 1.98 EMPLOYMENT status overall 2.47 1.57 1 6 N = 130,013 between 1.47 1 6 n = 59,648 within .65 -1.28 6.22 T = 2.18 Digital skills overall 2.00 1.07 1 4 N = 129,397 between 1.03 1 4 n = 59,532 within .35 -0.25 4.25 T = 2.17 E‑mail overall .79 .41 0 1 N = 92,237 between .39 0 1 n = 44,823 within .19 .04 1.54 T = 2.06 E‑health overall .66 .47 0 1 N = 82,822 between .42 0 1 n = 40,888 within .27 -0.09 1.41 T = 2.03 Social network overall .61 .49 0 1 N = 92,237 between .45 0 1 n = 44,823 within .23 -0.14 1.36 T = 2.06 E‑commerce overall .33 .47 0 1 N = 129,397 between .43 0 1 n = 59,532 within .21 -0.42 1.08 T = 2.17 E‑banking overall .57 .49 0 1 N = 92,237 between .46 0 1 n = 44,823 within .23 -0.18 1.32 T = 2.06 E‑government overall .63 .48 0 1 N = 93,140 between .44 0 1 n = 45,158 within .26 -0.12 1.38 T = 2.06 E‑tourism overall .64 .48 0 1 N = 37,519 between .45 0 1 n = 21,352 within .24 -0.11 1.39 T = 1.76 E‑learning overall .18 .38 0 1 N = 70,087 between .34 0 1 n = 39,144 within .20 -0.49 .85 T = 1.79 Cloud overall .36 .48 0 1 N = 79,958 between .43 0 1 n = 39,968 within .26 -0.39 1.11 T = 2.00 Note: Overall statistics refer to the pooled sample; ‘Between’ and ‘Within’ follow the standard xtsum decomposition, so σ² = σ_b² + σ_w². N = total person-year observations; n = distinct individuals; T = average waves per individual. This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5528118 Preprint not peer reviewed Page 19 of 26 A2. Polychoric Correlation Matrix Table 6. Polychoric Correlation Matrix INT_TRUST AGE GENDER EDUCATION INCOME EMPLOYMENT_SIT DIGITAL_SKILLS EMAIL EHEALTH SOCIAL_NETWORK ECOMMERCE EBANKING EGOVERNMENT ETOURISM ELEARNING CLOUD INT_TRUST 1.00 AGE -0.04 1.00 GENDER 0.06 0.07 1.00 EDUCATION 0.11 0.08 -0.10 1.00 INCOME 0.11 0.15 0.09 0.36 1.00 EMPLOYMENT_SIT -0.04 0.03 -0.09 -0.28 -0.23 1.00 DIGITAL_SKILLS 0.21 -0.26 0.10 0.34 0.20 -0.13 1.00 EMAIL 0.14 -0.13 0.04 0.39 0.25 -0.12 0.71 1.00 EHEALTH 0.05 0.02 -0.21 0.13 0.05 -0.02 0.38 0.27 1.00 SOCIAL_NETWORK 0.13 -0.39 -0.11 -0.07 -0.13 0.03 0.46 0.27 0.14 1.00 ECOMMERCE 0.15 -0.10 0.03 0.15 0.16 -0.10 0.31 0.26 0.12 0.12 1.00 EBANKING 0.16 0.10 0.07 0.32 0.23 -0.28 0.52 0.47 0.20 0.11 0.28 1.00 EGOVERNMENT 0.10 0.06 0.02 0.32 0.18 -0.14 0.56 0.45 0.27 0.07 0.17 0.42 1.00 ETOURISM 0.13 0.05 0.00 0.33 0.30 -0.21 0.37 0.36 0.13 0.04 0.21 0.36 0.32 1.00 ELEARNING 0.09 -0.12 -0.05 0.29 0.14 -0.20 0.56 0.37 0.17 0.12 0.15 0.23 0.32 0.24 1.00 CLOUD 0.16 -0.17 0.07 0.19 0.12 -0.03 0.66 0.44 0.17 0.31 0.22 0.29 0.30 0.26 0.28 1.00 Note: Entries are polychoric correlations, i.e. latent-variable correlations for ordinal and binary items. All absolute values are below 0.71. This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5528118 Preprint not peer reviewed Page 20 of 26 A3. Internet trust distribution by digital services adoption Figure 3. Internet trust by digital-service adoption Note: For each service dummy, 0 = non-user and 1 = user; bar segments show the percentage of that group reporting Low, Medium, or High trust. Within each category the 2014 bar (left) and the 2021 bar (right) stack to 100 %: light = Low, mid = Medium, dark = High Internet trust. 0 10 20 30 40 50 % of category 0 1 2014 2021 2014 2021 INT_TRUST by EMAIL : 2014 vs 2021 Low Medium High 0 10 20 30 40 % of category 0 1 2015 2021 2015 2021 INT_TRUST by EHEALTH : 2015 vs 2021 Low Medium High 0 10 20 30 40 % of category 0 1 2014 2021 2014 2021 INT_TRUST by SOCIAL_NETWORK : 2014 vs 2021 Low Medium High 0 10 20 30 % of category 0 1 2014 2021 2014 2021 INT_TRUST by ECOMMERCE : 2014 vs 2021 Low Medium High 0 10 20 30 40 % of category 0 1 2014 2021 2014 2021 INT_TRUST by EBANKING : 2014 vs 2021 Low Medium High 0 10 20 30 40 % of category 0 1 2014 2021 2014 2021 INT_TRUST by EGOVERNMENT : 2014 vs 2021 Low Medium High 0 10 20 30 40 % of category 0 1 2014 2019 2014 2019 INT_TRUST by ETOURISM : 2014 vs 2019 Low Medium High 0 10 20 30 40 % of category 0 1 2015 2021 2015 2021 INT_TRUST by ELEARNING : 2015 vs 2021 Low Medium High 0 10 20 30 40 % of category 0 1 2014 2020 2014 2020 INT_TRUST by CLOUD : 2014 vs 2020 Low Medium High This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5528118 Preprint not peer reviewed Page 21 of 26 A4. Estimates from four random-effects ordered-probit specifications Table 7 reports estimates from four random-effects ordered-probit specifications. Model 1 contains only socio-demographic variables; Model 2 adds digital-skill tiers and basic adoption dummies; Model 3 augments the equation with the full portfolio of e-services (smaller sub-sample); Model 4 retains the broader sample of Model 2 but introduces calendar-year dummies to capture macro shocks. Table 7. Random‑effects ordered‑probit estimates of Internet trust Model 1 Model 2 Model 3 Model 4 Variable Coef. p-value Coef. p-value Coef. p-value Coef. p-value Age 16–24 0.598*** 0.000 0.230*** 0.000 0.074 0.428 0.167*** 0.000 Age 25–34 0.513*** 0.000 0.208*** 0.000 0.071 0.382 0.062* 0.082 Age 35–44 0.399*** 0.000 0.170*** 0.000 0.037 0.635 0.031 0.343 Age 45–54 0.242*** 0.000 0.100*** 0.003 0.027 0.724 -0.009 0.786 Age 55–64 0.116*** 0.000 0.052* 0.084 0.013 0.854 -0.012 0.667 Female -0.113*** 0.000 -0.126*** 0.000 -0.111*** 0.000 -0.098*** 0.000 Secondary education 0.250*** 0.000 0.100*** 0.000 0.030 0.658 0.066*** 0.002 Bachelor 0.439*** 0.000 0.172*** 0.000 0.089 0.207 0.108*** 0.000 Master/PhD 0.558*** 0.000 0.218*** 0.000 0.130* 0.075 0.147*** 0.000 Income medium 0.121*** 0.000 0.073*** 0.000 0.046 0.241 0.064*** 0.000 Income high 0.309*** 0.000 0.209*** 0.000 0.144*** 0.000 0.171*** 0.000 Income very.high 0.475*** 0.000 0.341*** 0.000 0.239*** 0.000 0.295*** 0.000 Unemployed 0.006 0.744 0.021 0.292 0.042 0.267 0.038** 0.040 Retired -0.063** 0.031 0.005 0.866 0.039 0.566 0.005 0.870 Student 0.139*** 0.000 0.081** 0.039 -0.004 0.949 0.117*** 0.001 Housekeeper -0.127*** 0.000 -0.012 0.716 0.078 0.299 -0.004 0.906 Other employ. -0.098*** 0.001 -0.083*** 0.006 0.015 0.832 0.005 0.873 Digital Skills medium … … 0.342*** 0.000 0.093 0.123 0.246*** 0.000 Digital Skills high … … 0.542*** 0.000 0.244*** 0.000 0.366*** 0.000 Digital Skills very high … … 0.702*** 0.000 0.460*** 0.000 0.520*** 0.000 Email use … … 0.108*** 0.000 -0.021 0.683 0.084*** 0.000 E‑health … … 0.046*** 0.001 -0.026 0.334 … … Social networks … … 0.163*** 0.000 0.164*** 0.000 0.180*** 0.000 E‑commerce … … … … 0.170*** 0.000 0.234*** 0.000 E‑banking … … … … 0.132*** 0.000 0.142*** 0.000 E‑governmet … … … … -0.013 0.671 0.070*** 0.000 E‑tourism … … … … 0.090*** 0.000 … … E‑learning … … … … -0.009 0.745 … … Cloud … … … … 0.088*** 0.000 … … Year 2015 … … … … … … -0.064*** 0.004 Year 2016 … … … … … … -0.116*** 0.000 Year 2017 … … … … … … -0.162*** 0.000 Year 2018 … … … … … … -0.223*** 0.000 Year 2019 … … … … … … -0.188*** 0.000 Year 2020 … … … … … … -0.577*** 0.000 Year 2021 … … … … … … -0.492*** 0.000 Observations 75,348 67,244 20,721 74,763 Individuals 39,651 35,824 14,877 39,414 Log likelihood -61665.7 -53,284.2 -16,374.3 -59,495.3 Wald χ2 2,966.8 4,285.5 657.0 5,746.7 Prob > χ2 0.000 0.000 0.000 0.000 Note: Robust p‑values clustered by individual reported alongside coefficients. Significance: * p<0.10, ** p<0.05, *** p<0.01. Base categories: Age 65+, Male, Education None/Primary, Income Low, Employment Employed, Digital skills Low, Service‑use dummies 0, Year 2014. Model fit. The Wald statistic rises from 2,967 (Model 1) to 4,285 (Model 2) and 5,747 (Model 4), while its p-value remains < 0.001, confirming that each successive block of variables contributes with additional explanatory power. Digital-skill tiers alone improve pseudo-fit by roughly 13% (loglikelihood: –61,666 → –53,284). This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5528118 Preprint not peer reviewed Page 22 of 26 Socio-demographics. In the baseline specification, age correlates negatively with trust once the reference group (65+) is omitted: respondents aged 16-24 are 0.60 latent-index units above the elderly, with the gradient narrowing but still significant through Model 2. After behavioural controls enter (Models 3–4), the age profile disappears; only the youngest cohort remains weakly positive in Model 4, indicating that much of the raw age gap is mediated by skills and usage. A persistent gender gap of about –0.10 units (women less trusting) survives all specifications. Income shows a steep, monotonic gradient: moving from the low to the very-high quartile adds 0.30–0.48 units even in the richest model. Education and labour status. Secondary schooling is associated with a modest 0.07-unit gain in Model 4, rising to 0.15 for postgraduate qualifications. Employment status matters little once skills and adoption are controlled; the unemployed register a small positive coefficient (0.04), while all other nonemployed categories are insignificant. Digital capability. Digital skills are the single most powerful block. Relative to the low-skill baseline, a very high-skill respondent enjoys a 0.70 (Model 2) to 0.52 (Model 4) rise in the latent index, roughly half a standard deviation. Even the medium tier lifts trust by 0.25 on the full model. Service adoption. Frequent users of transactional services report higher trust: e-commerce (+0.23), ebanking (+0.14), and cloud storage (+0.09) remain significant in Model 3, with similar magnitudes in Model 4 (cloud unobserved). Social-network use is consistently positive (+0.16 to +0.18). Egovernment is insignificant in Model 3 but turns positive once year fixed effects are introduced, suggesting its impact was masked by time shocks. Time effects. Model 4 reveals a persistent decline in trust since 2015, intensifying in 2020 (–0.58) and 2021 (–0.49). These shifts underscore the importance of controlling for macro events (most notably the pandemic) when studying attitudinal outcomes. This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5528118 Preprint not peer reviewed Page 23 of 26 A5. Evidence based implications matrix Table 8. Evidence–implication matrix Finding Evidence Implication Digital skills dominate Low→Medium +0.25***, High +0.37***, Very‑high +0.52*** (robust in all models). Skill training sharply raises odds of high trust. Income gradient High +0.17***, Very‑high +0.30***; medium smaller +0.06***. Economic security matters beyond capability. Gender gap Female ≈ −0.10*** in every model. Persistent risk‑perception difference; policies must be gender‑sensitive. Age effect explained by skills Negative age dummies (45+) in Model 1 vanish after controls; only 16–24 remains +0.17***. Age penalty is mediated by skills and use, not birth cohort. Transactional experience boosts trust E‑commerce +0.23***, E‑banking +0.14***, Social networks +0.18***; e‑gov +0.07***. Hands‑on value‑adding use fosters confidence. Macro time trend Year dummies fall from 2015; 2020 −0.58***, 2021 −0.49***. Pandemic‑era shocks markedly erode trust, calling for system‑level safeguards. Note: Coefficients refer to Model 4 (fully controlled ordered‑probit); “***” indicates p<0.01. Where variables are absent in Model 4, the nearest estimate from Model 3 is reported. During the preparation of this work the authors used GenAI (ChatGPT) in the literature review and select sections to enhance readability and language clarity. After using this tool/service, the author(s) reviewed and edited the content as needed and take full responsibility for the content of the published article. This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5528118 Preprint not peer reviewed Page 24 of 26 References Agresti, A. (2018). An Introduction to Categorical Data Analysis, 3rd Edition | Wiley. John Wiley & Sons. https://www.wiley.com/enus/An+Introduction+to+Categorical+Data+Analysis%2C+3rd+Edition-p-9781119405283 Akerlof, G. A. (1970). The Market for “Lemons”: Quality Uncertainty and the Market Mechanism*. The Quarterly Journal of Economics, 84(3), 488–500. https://doi.org/10.2307/1879431 Capilla, J., Alcaraz, A., Valarezo Unda, Á. E., García Hiernaux, A. A., & Pérez Amaral, T. (2025). Eco-RETINA: A green flexible algorithm for model building. https://hdl.handle.net/20.500.14352/117836 Corritore, C. L., Kracher, B., & Wiedenbeck, S. (2003). On-line trust: Concepts, evolving themes, a model. International Journal of Human-Computer Studies, 58(6), 737–758. https://doi.org/10.1016/S1071-5819(03)00041-7 European Data Protection Supervisor. (2018, May 25). The History of the General Data Protection Regulation | European Data Protection Supervisor. https://www.edps.europa.eu/dataprotection/data-protection/legislation/history-general-data-protection-regulation_en Fernández-Bonilla, F., Gijón, C., & De la Vega, B. (2022). E-commerce in Spain: Determining factors and the importance of the e-trust. Telecommunications Policy, 46(1), 102280. https://doi.org/10.1016/j.telpol.2021.102280 Garín-Muñoz, T., López, R., Pérez-Amaral, T., Herguera, I., & Valarezo, A. (2019). Models for individual adoption of eCommerce, eBanking and eGovernment in Spain. Telecommunications Policy, 43(1), 100–111. https://doi.org/10.1016/j.telpol.2018.01.002 Garín-Muñoz, T., & Pérez-Amaral, T. (2011). Factores determinantes del comercio electrónico en España. Boletín Económico de ICE, Información Comercial Española, 3016, 51–65. Gefen, D. (2000). E-commerce: The role of familiarity and trust. Omega, 28(6), 725–737. https://doi.org/10.1016/S0305-0483(00)00021-9 Greene, W. H. (2018). Econometric Analysis Global Edition (8th ed.). Pearson Education Limited. This preprint research paper has not been peer reviewed. Electronic copy available at: https://ssrn.com/abstract=5528118 Preprint not peer reviewed