Robust Predictors of Mobile Phone Reliance for Information Seeking: A Multi-Stage Empirical Analysis and Validation

Abstract

This study examines factors driving reliance on mobile phones as a primary information source. Using Information-Seeking Complementarity Theory (ISCT), which posits that frequent use of diverse media channels builds digital habits that reinforce mobile reliance, we analyze World Values Survey (WVS) Time Series 1981–2022 (v4.0), validated with WVS v5.0 and Integrated Values Survey (IVS). A multi-stage pipeline integrates AdaBoost (R 4.3.1), LASSO/BMA (Stata v17), Histogram Gradient Boosting (Python 3.12.7), and mixed-effects logistic regression. Missing data (DK/NA) were excluded or median-imputed. The final model (AUC-ROC > 0.85) identifies five robust predictors: age (negative), and positive associations with digital mail, online social networks, peer interaction, and radio listening—all stable across methods, datasets, and reverse causality checks. Subgroup analysis reveals stronger effects among males, unmarried individuals, urban residents, and higher education/employment groups. Nomograms enable probabilistic forecasting and policy simulation. By identifying technology-agnostic behavioral drivers validated across three decades of global survey data (1981–2022), with mobile reliance measured from 2010 onward, this work provides a transparent, replicable predictive framework with implications for emerging AI and wearable contexts.

1. Introduction

In recent years, mobile phones have become ubiquitous devices that have transformed how we access information and engage in e-commerce [1]. With the advent of smartphones and the increasing availability of internet connectivity, individuals now have instant access to a wealth of information and online transactions [2]. The development of mobile technology has provided new ways for information seeking and online shopping, revolutionizing our interaction with the digital world [3,4].

1.1. Research Context and Rationale

The use of mobile phones as a source of information significantly impacted how individuals seek and consume information. Research by [5] demonstrated that mobile phones have become primary devices for accessing information, surpassing traditional sources such as desktop computers and laptops [6]. The convenience, portability, and proliferation of mobile applications tailored for specific information needs further enhanced this role [7,8,9].

In addition to information seeking, mobile phones have emerged as powerful tools for e-commerce. The growth of mobile e-commerce (m-commerce) also relies on the increasing number of smartphone users, their affordability, and the availability of mobile-optimized websites and apps [10,11,12]. Studies have shown that mobile phones have become an integral part of the online shopping experience [13]. According to research by [14], mobile e-commerce has experienced substantial growth, with consumers increasingly relying on their mobile phones for product searches, price comparisons, and online purchases.

The rise of m-commerce is due to several factors [15]. First, mobile phones contribute to a seamless and personalized shopping experience, empowering users to inspect products, conduct price comparisons, read reviews, and make straightforward purchases with just a few taps [16,17]. Secondly, integrating mobile payment systems has simplified the payment process, enhancing the convenience and security of mobile transactions [18]. Thirdly, location-based services and personalized recommendations on mobile platforms improved the overall shopping experience by tailoring offers and promotions to individual preferences [19,20]. Fourthly, a diverse range of businesses have emerged and become accessible through digital social networks (SNs) for e-commerce purposes [21].

When first exploring the comprehensive dataset of the WVS (version 4.0), a specific variable emerged. It captured the frequency of using a mobile phone as an information source [22]. This variable exhibited strong correlations with five other variables. Using various data mining tools and techniques, along with additional datasets from the same provider, the current analysis expands the understanding of the above topic [23,24,25].

1.2. Novelty and Contributions

This study addresses a critical gap by providing a robust, data-driven analysis of the key variables that influence a person’s use of a mobile phone as a source of information. Our research focuses on identifying how these core variables (age, frequency of using digital mail, online social networks, peer communication, and radio) are associated with mobile information acquisition behavior. By focusing on these specific predictors, we aim to provide a nuanced understanding of user patterns. The necessity of this research lies in its practical application: the findings are essential for designing user-centric mobile applications and informing policy decisions aimed at fostering digital inclusion. Our comprehensive methodology, which employs multiple analytical techniques to confirm these variables, demonstrates the feasibility and rigor of our approach in creating a reliable predictive model.

Furthermore, this study offers valuable insights for designing user interfaces (UI) for mobile operating systems and dedicated applications, as UI design is the initial point of interaction between users and applications and is crucial for success regardless of complexity, functionality, or performance [26,27,28].

While we employed known software and algorithms, the novelty and significance of this work do not stem from the creation of new computational tools but rather from their unique application and systematic integration to address a critical gap in the literature. Therefore, we can mention a threefold contribution:

(a)

Methodological Fusion: We proposed a new, multi-stage analytical framework centered on feature selection through rigorous triangulation. This approach systematically combines multiple Data Mining (DM) techniques (including HGB, AdaBoost, and Lasso methods) with traditional statistical models (e.g., Logit). Crucially, we empirically demonstrate the superiority of this triangulated approach as a robustness filter over single-method optimization (e.g., HGB). Specifically, while a single HGB selection yielded a higher raw predictive score based on AUC-ROC, this set proved structurally unstable in the Logit framework, experiencing a severe drop in performance and a loss of statistical significance for key variables. Our methodology, by contrast, identifies a minimal set of core predictors that maintain superior performance (e.g., AUC-ROC) and robustness across all tested frameworks, providing an unprecedented level of confidence in the generalizability of the findings.

(b)

Empirical Contribution: To our knowledge, this is one of the first studies to apply such a comprehensive framework to the specific, cross-national WVS/IVS dataset, providing a new empirical understanding of a rapidly evolving social phenomenon. The latter includes uncovering novel predictors like radio usage, reflecting traditional media integration in digital behaviors.

(c)

Practical Tool: One of the final products of this research, the nomogram, is a significant contribution (in itself). It transforms complex research outputs into an intuitive tool for UI designers and policymakers, enabling simulations for optimizing application features (e.g., low-literacy interfaces).

While basic descriptive analyses (e.g., time-series or group-based tabulations on WVS/IVS data) highlight trends in mobile phone reliance across time, regions, and age groups, they lack the capacity to identify strong predictors, address biases such as multicollinearity, or support forecasting. The multi-stage framework presented here fills these gaps by integrating statistical methods to uncover stable associations, achieve high predictive accuracy (AUC-ROC > 0.85), and generate practical tools like nomograms for simulation—offering deeper explanations and actionable insights.

Unlike standard time-series or trend regressions that capture linear shifts over time or across groups, this framework reveals nonlinear effects (e.g., through HGB) and subgroup-specific patterns often missed in aggregate models. It also enables precise probabilistic forecasting and provides visual policy instruments for individual-level simulation—capabilities that extend beyond conventional longitudinal approaches.

Moreover, although the pipeline integrates multiple ML and statistical layers for robustness, the final sparse LOGIT model and nomogram ensure high interpretability and direct policy applicability—transforming complex outputs into intuitive, simulation-ready tools that simpler models with comparable accuracy cannot match in transparency or actionability.

In terms of comparison with a longitudinal panel model, while the latter might capture broad trends in age or digital activity, the repeated cross-sectional structure of WVS/IVS, combined with high noise and collinearity, demands ensemble triangulation to identify stable, non-spurious predictors. Boosting and regularization (Stages 1–4) are conceptually essential to filter robust signals—yielding a sparse, policy-ready model that simpler longitudinal regression cannot guarantee.

2. Short Literature Review

The use of mobile phones as a primary information source has augmented rapidly and transformed individual access and engagement with data in daily life. One can trace this trend back to technological innovations, beginning with the release of the first Personal Digital Assistant (PDA) in 1984, which introduced handheld computing and laid the foundation for mobile information access [29]. Over the decades, significant milestones, such as the launch of smartphones in the early 2000s and the proliferation of mobile internet, have made mobile phones indispensable tools for information-seeking behaviors [30].

2.1. Overview and Main Research Questions

Mobile technology has significantly transformed e-commerce engagement. Research underscores that smartphones are now the primary devices for accessing information, surpassing traditional sources [5,6]. The convenience and portability of mobile phones enable users to access information at any moment and from any location. These cause them to stay connected with the digital world and remain informed about the latest news, updates, and developments [7].

Additionally, mobile applications tailored to specific information needs have enhanced the role of mobile phones as critical information sources [8,9]. In the realm of e-commerce, the proliferation of mobile-optimized websites and apps, together with the increasing number of smartphone users, has fueled the growth of mobile commerce [10,11,12]. Studies highlight that mobile phones have become integral to online shopping, facilitating product searches, price comparisons, and purchases with ease [14].

Furthermore, the broad applicability of mobile phones is evident in various contexts. For instance, research explores mobile phone usage and associations with cultural values, location-induced specificity, and public support for specific communities [31,32,33]. Mobile phones are also a topic of discussion concerning lower trust levels, moral progress, intimate partner violence against women, female egalitarian values in protests, and financial emancipation [23,34,35,36,37].

Given that the use of SN via mobile technologies has generalized exponentially in the last ten years [4,38], the question arises whether access to information with the mobile phone is partially associated with the use of SN, communication and information sharing with friends and people with common interests (Hypothesis No.1 or H1).

In terms of access to relevant information and trust in information sources, a large number of applications available today on mobile devices are designed to verify email conversations and perform sentiment analysis starting from emails. These applications rely on various classification and feature selection algorithms, reflecting the growing academic and practical interest in this attractive field [39,40]. The latter category should also include the use of ensembles of classifiers (e.g., those based on the Boosted Bayesian approach) for similar or related purposes such as email spam classification [41,42,43,44]. Checking e-mail via mobile technologies has become a widespread practice, reflecting the evolution of traditional mail communication into the digital format. Based on the above, the question that arises is whether access to information using a mobile phone is partially associated with access to email (Hypothesis No.2 or H2).

Regarding the use of mobile devices as an information source in higher education, there are opinions suggesting that technology entangles with everything people do. This entanglement with everyday activities means that mobile technology can no longer be separated from pedagogy. Mobile devices repurposed for educational needs have proven to be a promising and successful approach. They are becoming ubiquitous in higher education. However, disparities in technical infrastructure and Internet access within institutions still pose significant accessibility issues [45,46,47]. Given that the use of educational tools via mobile technologies has experienced an unprecedented growth in recent years, another research question arises. Whether access to information via mobile phones is partly associated with the use of e-learning tools and communication with teachers and colleagues (Hypothesis No.3 or H3).

In addition, these days we must consider the use of mobile devices as an information source together with other well-established channels, such as radio, TV, blogs, and YouTube channels [48,49,50,51,52,53]. The same applies in the context of exploring the evolution and impact of virtualization technology on mobile devices [54]. It could increase accessibility by using VMs (with a broader range of traditional applications and related functionalities) on mobile devices [55]. Given that the use of alternative channels such as radio, TV, blogs, and YouTube channels, together with or through mobile technologies, has grown exponentially in recent years, an additional question arises. To what extent is access to information via mobile phones associated with the use of alternative channels (Hypothesis No. 4 or H4)?

This diverse array of studies emphasizes the multifaceted impact of mobile phones beyond merely information access and e-commerce.

2.2. Common Predictors and Contrasting Findings

From a global perspective, the usage of mobile phones as an information source reflects diverse socio-economic, cultural, and technological contexts.

In terms of differences between developed and developing countries regarding mobile phone use, it is worth mentioning some facts. In developing nations, mobile devices often serve as primary internet access points. Consequently, they bypass traditional infrastructures. This phenomenon, known as leapfrogging [56], enables these countries to adopt advanced technologies with minimal reliance on legacy systems. However, challenges persist. The digital divide is influenced by factors such as income, education, and geographic location, affecting the individual ability to access and effectively use information and communication technologies [57,58]. In developed regions, mobile phones are predominantly used for instant access to information, driven by high smartphone penetration and robust internet infrastructure [6]. Conversely, in developing regions, mobile phones often serve as the primary or sole means of accessing information. They help bridge the digital divide where traditional internet access might be limited [32]. This disparity suggests the critical role of mobile technology in enhancing information accessibility across varying socio-economic landscapes.

Moreover, recent developments, such as the implementation of laws restricting smartphone use in schools, reflect growing concerns about the impact of technology on youth [59].

Furthermore, cultural factors influence mobile phone usage patterns. For instance, research indicates significant regional variations, with mobile phones supporting local cultural practices and information dissemination methods [31].

In regions with strong SN engagement, mobile phones facilitate information access. It is also about community building plus social movements, as seen in studies examining public support for LGBTQ (Lesbian, Gay, Bisexual, Transgender, and Queer or Questioning) rights and female egalitarian values in protests [33,36]. Consequently, collectivist societies (e.g., China, Japan, and many countries in Latin America) may emphasize group communication via messaging apps and see greater applicative utility. Individualistic cultures (e.g., U.S., Australia, and the UK) may prioritize personal application design and content [60].

Even more, the latter authors put that in a larger set, namely those five cultural values (individualism/collectivism, masculinity/femininity, power distance, uncertainty avoidance, and long-term orientation) of Hofstede. These five act as mediators in the context of analyzing the continued intention to use mobile social media applications [61].

Thus, masculinity versus femininity might influence user engagement based on cultural preferences for assertiveness vs. collaboration. Masculine cultures (e.g., Japan, Germany, and the U.S.), which value competitiveness and achievement, might use such technologies more for self-promotion and status display. Feminine cultures (e.g., Sweden, Norway, and the Netherlands), which prioritize cooperation and quality of life, may use them more for building relationships and shared experiences [62].

In high power-distance cultures (e.g., Russia, Mexico, and India), there is a greater acceptance of unequal power distribution and authority. The use of such technologies in these contexts might emphasize hierarchical structures and respect for authority figures.

In low power-distance cultures (e.g., Denmark, Sweden, and the Netherlands), people tend to favor equality and open communication. Such use often fosters informal interactions and encourages a more egalitarian approach, where individuals feel free to express opinions and engage directly with others, regardless of status [63].

In high uncertainty avoidance cultures (e.g., Greece, Japan, and Argentina), people prefer structure, rules, and predictability. Such use may reflect a preference for clear guidelines and less risk-taking in online interactions.

In low uncertainty avoidance cultures, such as the UK, Sweden, and the U.S., there is (more) tolerance for ambiguity, flexibility, and experimentation. Such usage behaviors in these cultures might include more innovation, spontaneity, and openness to new trends or ideas [64].

In long-term orientation cultures, such as China, Japan, and South Korea, there is a strong orientation to future rewards, persistence, and planning. The use of such technologies in these cultures may reflect a preference for maintaining long-term relationships and thoughtful, strategic engagement.

In short-term orientation cultures (e.g., the United States, Canada, and many Middle Eastern countries), the emphasis is on immediate results and the present. Such usage behaviors might focus on instant gratification, trends, and short-term interactions [65].

Moreover, usability is a key predictor of the continued intention to use mobile applications at the individual level. The former includes visual support of task goals, cognitive interaction, efficient interaction, functional support of user needs, and ergonomic support. This finding is confirmed by other contributions [60,66,67]. This predictor (under the form of the perceived ease of use) seems to have a greater impact on outcomes related to the adoption of new technologies. It particularly affects individuals with masculine values, measuring a specific degree of preference for achievement, assertiveness, and material success [68,69]. These global variations underscore the need for context-specific strategies when implementing mobile technology. The latter also applies when it comes to optimizing information access. They also help leverage the full potential of mobile phones in diverse cultural and socio-economic settings [70,71].

The literature presents a nuanced understanding of the impact of mobile technology on information-seeking and e-commerce. While some researchers [5,6] emphasize the dominance of mobile phones in information access, studies like those by [10,11] focus on the seamless shopping experiences facilitated by mobile phones. Research by [16,17] underscores the personalized shopping experiences enabled by mobile phones, highlighting the convenience and security offered by mobile payment systems [18].

While various studies explore the convenience of mobile e-commerce, few address the inclusive design principles necessary for elderly users, as underlined by [72,73,74,75,76]. The latter emphasized that mobile application technologies designed to assist seniors and people with disabilities should incorporate accessibility features. These features include automatic reading and adjustment of font and image dimensions, text enlargement options, customized color schemes, audio descriptions and voice commands, accessibility maps, and location awareness.

The analyzed authors [77] mentioned the Older Adult-Friendly Mobile Apps Features Checklist of [78]. The latter contains six considerations and corresponding features for each:

-

Background (avoid abrupt brightness changes, use colors sparingly, and maintain high contrast between foreground and background elements);

-

Browsing (avoid scroll bars and overlapping pop-ups);

-

Cognition (provide sufficient time to read content and simplify decision-making by offering fewer choices, reducing cognitive load);

-

Content (eliminate irrelevant information, highlight key points, use simple and unambiguous language, and ensure consistent, user-friendly layout and navigation);

-

Graphics (avoid animations, ensure all images have alt tags, and use simple, meaningful icons to improve accessibility and clarity);

-

Target (avoid requiring double clicks and use larger, more visible targets to improve usability and accessibility). Judging by this list of additional requirements, access to information via the mobile phone is to the advantage of younger users (Hypothesis No. 5 or H5) [79].

Other studies have examined mobile phone usage in relation to cultural values, preferences for television versus the internet as sources of information, and public support for specific community-related issues [31,33,80]. They highlight the need for further integrated research aimed at developing a holistic understanding of mobile phone usage patterns.

Regulations and privacy policies significantly shape the extent to which individuals use mobile phones as a source of information. Governments worldwide implement varying degrees of regulation concerning data privacy, internet access, and content moderation, which directly influence trust and engagement with mobile platforms [81] for most users. Strict data protection laws, such as the European Union General Data Protection Regulation (GDPR) [82], aim to safeguard user privacy by imposing stringent requirements on data collection and processing. While these measures enhance security, they can also create barriers for information access by limiting personalized content recommendations and cross-platform data sharing. Conversely, in regions with less stringent regulations, users may have greater access to diverse sources of information but at the cost of potential data exploitation, misinformation, and surveillance [83]. Additionally, government-imposed internet restrictions—such as website censorship, social media bans, or internet shutdowns—can severely limit the use of mobile phones for obtaining information, particularly in authoritarian regimes [84]. The balance between privacy protection and information accessibility remains a crucial factor in understanding mobile phone usage (as an information tool), as regulatory frameworks continue to evolve globally.

Moreover, the modulation of mobile phone use by legal frameworks extends beyond broad privacy regulations (such as the GDPR) to include more granular policies that affect content access and application functionality. For example, laws governing data localization, which require data to persist on servers within certain country borders [85,86], can influence the types of cloud-based services and applications [87,88] available to users, potentially limiting access to international information sources. Similarly, intellectual property laws and digital rights management (DRM) can restrict how users consume and share digital content, affecting the utility of mobile devices for accessing media and academic resources [89,90].

Although mobile phones offer significant advantages as sources of information, increasing dependence presents several risks, including misinformation, digital addiction, and reduced critical thinking skills. All have been widely documented [91]. The rapid dissemination of unverified information through mobile-based SN platforms can contribute to misinformation spread, affecting public opinion and decision-making [92]. Additionally, excessive reliance on mobile devices for information may reduce engagement with diverse sources, potentially reinforcing confirmation biases and limiting exposure to alternative perspectives [93]. Therefore, evolving anti-disinformation and hate speech laws in some jurisdictions may lead platforms to implement stricter content moderation, which, while curbing harmful content, can also be perceived as censorship, thereby impacting user trust in the mobile information ecosystem. These legal layers create a complex landscape where the free flow of information continuously changes, demonstrating how legal and technological infrastructures intertwine in shaping mobile information-seeking behaviors [94,95].

In terms of the techniques used, most of the research approaches indicate complementary methods. These methods include logistic regression, decision tree, K-nearest neighbors (KNN), Gaussian Naive Bayes, ADA BOOST, Structural Equation Modeling (SEM), and even sophisticated machine learning (ML) and deep learning (DL)—based methods [43,44,60,96].

2.3. Existing Gaps

The discussion reveals a comprehensive yet fragmented understanding of the impact of mobile technology on information-seeking and e-commerce. Mobile phones are celebrated for their convenience and portability, leading to enhanced information access and e-commerce engagement. Still, there are critical gaps in understanding the predictors of mobile phone usage as an information source.

The gaps in existing studies (

Table A2. Short summarization of representative prior work.

Study (Author, Year)	Focus/Main Findings	Methods	Limitations/Gaps
Levine et al. (2012) [5]	Mobile phones as primary info devices, surpassing desktops; impacts on multitasking/distractibility.	Survey-based analysis of mobile media use.	Limited to U.S. context; descriptive, no robust predictors or global validation; ignores cultural variations.
Sher et al. (2022) [6]	Mobile dominance in info access; time-of-day/weekend patterns in LMS usage.	Analysis of LMS logs from blended learning.	Focus on educational settings; no cross-national data; lacks predictive modeling for broader behaviors.
Yu et al. (2022) [7]	Convenience/portability enhances info access; changes in electronic news habits.	Comparative survey of university students.	Region-specific (likely Asia); no integration of alternative channels (e.g., radio); limited to young demographics.
Neggaz et al. (2023) [8]	Mobile apps for specific info needs; boosted optimization for facial analysis.	ML-based boosting algorithm testing.	Technical focus on algorithms; no behavioral predictors; lacks empirical validation on survey data like WVS.
Alqahtani & Goodwin (2012) [10]	Growth of m-commerce via smartphones; affordability drives adoption.	Literature review and conceptual model.	Descriptive; no quantitative predictors; ignores age/digital divide effects.
Barry & Jan (2018) [11]	Seamless shopping via mobile-optimized sites/apps.	Case study/review of m-commerce trends.	Limited to e-commerce; no global/cross-cultural analysis; lacks robust statistical validation.
Salehan et al. (2018) [31]	Mobile usage linked to cultural values (e.g., individualism); global trends.	Cross-country trend analysis using secondary data.	Aggregate-level; no individual predictors (e.g., age, SN use); limited to associations, not causation.
Kushlev & Proulx (2016) [23]	Mobile info consumption linked to lower trust; social costs.	Empirical study on mobile use and trust.	Focus on negative outcomes; small sample; no predictive tools or multi-method triangulation.
Hoehle et al. (2015) [60]	Cultural values mediate mobile app use intention; usability as key predictor.	SEM on four-country survey data.	Limited countries; no integration with traditional media; ignores elderly/inclusive design gaps.
Pattison & Stedmon (2006) [72]	Inclusive design for older users in mobile phones.	Human factors/ergonomics review.	Conceptual; no empirical global data; highlights but does not address digital divide quantitatively.
Iancu & Iancu (2020) [73]	Mobile tech design for elderly; theoretical overview.	Literature synthesis on accessibility.	Theoretical; no predictive models; limited to design principles, not behavioral predictors.
Kasemsarn et al. (2024) [76]	Museums/apps for older/mobility-impaired; inclusive principles/digital storytelling.	Literature review on heritage accessibility.	Domain-specific (museums); no broad survey validation; lacks tools for policy simulation (e.g., nomograms).

, Appendix A), such as their focus on specific time periods, regions, or demographic groups, highlight the need for broader, more inclusive research. Moreover, the challenges associated with the digital divide and the necessity for inclusive interface design, particularly for elderly users, point to areas requiring further investigation. By integrating findings from large and diverse datasets with demographic diversity and employing robust validation techniques, research could reveal more, in terms of highly generalizable insights and practical and more intuitive tools for policymakers, researchers, and designers. The goal is to support them as effective contributors to the long-term improvement of mobile applications with reliable informational support.

The additional use of nomograms as intuitive and versatile visual tools can support analysts and managers. They can easily identify and eliminate model redundancies during intermediate exploratory stages by performing comparisons in terms of predictor magnitude. Further, nomograms facilitate intuitive calculus and interpretation of the models, calculate the probabilities, and make predictions based on summed predictor scores transformed into probabilities [97,98,99]. When large datasets allow additional continent-based filters facilitate the analysis of how key predictors in a general or global model behave within regional (continental) contexts. When coupled with these filters, the use of nomograms enhanced with annotations, in the form of scores corresponding to the magnitude of predictors for resulting specific models, makes it possible to identify potential deviations from the overall model [100,101].

3. Materials and Methods

3.1. Key Data Selection Steps

This study utilized the comprehensive WVS dataset, specifically version 4.0 (WVS_TimeSeries_4_0.dta for Stata), which consists of 1045 variables and 450,869 raw observations. This dataset served as the primary source for all selection stages. A part of them relied on Stata 17.0 (multi-processing/parallel edition). Stata and all other tools ran on a Windows VM with four virtual CPUs and 32 GB RAM.

To ensure robustness and validity of the results, two additional datasets supported explorations alongside the primary WVS one. The first is IVS, available at https://www.worldvaluessurvey.org/WVSEVStrend.jsp (accessed on 1 March 2025). This independent dataset resulted after merging the EVS Trend File (1981–2017) with the WVS Trend dataset (1981–2022) as detailed in Appendix D (

Table A15. Short description of the IVS dataset.

	IVS	EVS Trend File	WVS Trend File
Survey period	1981–2022	1981–2017	1981–2022
Number of waves	7	5	7
Number of cases	663.965	224.434	442.473
Number of variables	838	635	732
Countries/territories	120	49	108
Number of surveys	464	160	306

Source: World Values Survey, https://www.worldvaluessurvey.org/WVSEVStrend.jsp.

). The second was the most recent version of the WVS dataset v5.0 (WVS_Time_Series_1981–2022_stata_v5_0.dta for Stata), which includes 1046 variables and 443,488 raw observations. These last two datasets (IVS and WVS v5.0) served additional validation and robustness tests. These datasets are similar to the original WVS v4.0. By comparing the results across all three sources, the study validates the generalizability of the conclusions, ensuring their consistency and reliability. This multi-dataset approach helps mitigate potential biases or limitations inherent in the WVS dataset, strengthening the overall credibility of the findings across different (yet similar) data sources. We considered all three datasets for their global coverage, methodological rigor, and established use in cross-national research. Data harmonization across waves and countries followed the official WVS/EVS protocols (Common EVS/WVS Dictionary, 2022) and the merge syntax for Stata fully available and described at https://www.worldvaluessurvey.org/WVSEVStrend.jsp. The official merge script explicitly uses this dictionary for variable replication criteria and loads pre-harmonized EVS (ZA7503_v3-0-0) and WVS (Trends_VS_1981_2022_v4_0) trend files from GESIS/JD Systems. More, it appends datasets, recodes missing data, standardizes labels, and orders variables, ensuring post hoc consistency across waves and countries. Wave/country identifiers and errata corrections confirm the rigorous harmonization protocols outlined, with no discrepancies. Moreover, we validated via mixed-effects models with fixed effects (Stage 8) and tabulations (

Table A7. The support for those five most robust influences considering the chronology of the WVS.

S002VS (Chronology of Waves)	E260B	E260BBin	E253B	E254B	E259B	E261B	E262B	X002
1981–1984	0	0	0	0	0	0	0	11,324
1989–1993	0	0	0	0	0	0	0	27,331
1994–1998	0	0	0	0	0	0	0	76,448
1999–2004	0	0	0	0	0	0	0	59,979
2005–2009	0	0	0	0	0	0	0	84,887
2010–2014	85,013	85,013	0	85,242	85,262	84,826	84,974	87,315
2017–2022	93,326	93,326	90,726	93,253	93,326	92,460	92,740	92,694
Total	178,339	178,339	90,726	178,495	178,588	177,286	177,714	439,978

Source: Own calculations using the following command in Stata 17: tabstat E260B E260BBin E253B E254B E259B E261B E262B X002, by(S002VS) statistics (count).

and

Table A8. The support for those five most robust influences considering the country codes as defined in the WVS.

S003 (Numeric Country Code)	E260B	E260BBin	E253B	E254B	E259B	E261B	E262B	X002
Albania	0	0	0	0	0	0	0	1994
Algeria	1179	1179	0	1178	1184	1161	1165	2482
Andorra	997	997	1001	999	1001	999	999	2007
Azerbaijan	1002	1002	0	1002	1002	1002	1002	3003
Argentina	1995	1995	759	2007	2005	1966	1984	7401
Australia	3206	3206	0	3240	3228	3209	3219	7949
Bangladesh	1183	1183	1119	1172	1177	1124	1128	4220
Armenia	2311	2311	1218	2310	2307	2307	2309	4323
Bolivia	2048	2048	2025	2059	2064	2021	2030	2067
Bosnia Herzegovina	0	0	0	0	0	0	0	1200
Brazil	3208	3208	1729	3219	3216	3193	3212	5889
Bulgaria	0	0	0	0	0	0	0	2073
Myanmar	1200	1200	1200	1200	1200	1200	1200	1200
Belarus	1515	1515	0	1517	1523	1503	1517	4642
Canada	4018	4018	4018	4018	4018	4018	4018	11,076
Chile	1978	1978	990	1984	1990	1983	1986	6700
China	5124	5124	3023	5125	5119	5106	5120	10,827
Taiwan	2448	2448	1223	2440	2451	2452	2450	4459
Colombia	3028	3028	1520	3031	3031	3026	3026	12,069
Croatia	0	0	0	0	0	0	0	1191
Cyprus	1986	1986	988	1982	1975	1967	1971	3048
Czechia	1193	1193	1195	1195	1194	1192	1194	3271
Dominican Rep.	0	0	0	0	0	0	0	414
Ecuador	2397	2397	1184	2396	2398	2391	2390	2402
El Salvador	0	0	0	0	0	0	0	1254
Ethiopia	1206	1206	1099	1189	1209	1086	1095	2730
Estonia	1497	1497	0	1516	1523	1511	1518	2554
Finland	0	0	0	0	0	0	0	2000
France	0	0	0	0	0	0	0	1001
Georgia	1200	1200	0	1201	1201	1200	1200	4710
Palestine	995	995	0	993	998	993	995	1000
Germany	3563	3563	1527	3571	3574	3566	3568	7657
Ghana	1552	1552	0	1552	1552	1552	1552	3086
Greece	1197	1197	1119	1199	1196	1113	1116	1200
Guatemala	1189	1189	1198	1193	1185	1184	1186	2229
Haiti	1937	1937	0	1937	1937	1938	1938	0
Hong Kong	2075	2075	2075	2075	2075	2073	2073	3237
Hungary	0	0	0	0	0	0	0	1656
India	4078	4078	0	4078	4078	4078	4078	12,525
Indonesia	3164	3164	2980	3198	3182	2959	2990	6214
Iran	1495	1495	1496	1498	1497	1496	1496	6681
Iraq	2378	2378	1188	2387	2369	2356	2366	7426
Israel	0	0	0	0	0	0	0	1191
Italy	0	0	0	0	0	0	0	1012
Japan	3734	3734	1330	3755	3737	3730	3734	9523
Kazakhstan	2731	2731	1222	2713	2724	2716	2723	2776
Jordan	2396	2396	1198	2395	2398	2394	2397	4826
Kenya	1254	1254	1245	1245	1253	1224	1240	1259
South Korea	2421	2421	1245	2427	2415	2421	2424	7343
Kuwait	1225	1225	0	1242	1219	1219	1232	1245
Kyrgyzstan	2694	2694	1195	2691	2696	2683	2691	3743
Lebanon	2377	2377	1200	2379	2388	2368	2379	2400
Latvia	0	0	0	0	0	0	0	1200
Libya	3267	3267	1184	3278	3262	3229	3250	3316
Lithuania	0	0	0	0	0	0	0	1009
Macau	1019	1019	1018	1022	1019	1016	1018	813
Malaysia	2613	2613	1313	2613	2613	2613	2613	3813
Maldives	1038	1038	1032	1031	1033	1028	1029	1036
Mali	0	0	0	0	0	0	0	1428
Mexico	3738	3738	1739	3739	3741	3738	3738	11,707
Mongolia	1638	1638	1638	1638	1638	1638	1638	1638
Moldova	0	0	0	0	0	0	0	3037
Montenegro	0	0	0	0	0	0	0	1298
Morocco	1200	1200	1200	1200	1200	1200	1200	4848
Netherlands	3895	3895	1999	3852	3888	3886	3895	5097
New Zealand	1748	1748	980	1793	1796	1744	1777	3963
Nicaragua	1200	1200	1200	1200	1200	1200	1200	1200
Nigeria	2989	2989	1217	2990	2994	2981	2978	7988
Norway	0	0	0	0	0	0	0	2152
Pakistan	3117	3117	1908	3114	3105	3067	3049	5919
Peru	2577	2577	1368	2571	2587	2557	2562	6822
Philippines	2395	2395	1200	2397	2396	2398	2398	4800
Poland	958	958	0	963	962	958	960	4052
Puerto Rico	1099	1099	1102	1106	1089	1082	1105	2991
Qatar	1060	1060	0	1060	1060	1060	1060	1053
Romania	2723	2723	1218	2716	2748	2685	2702	5755
Russia	4177	4177	1774	4245	4219	4167	4213	10,342
Rwanda	1527	1527	0	1527	1527	1527	1527	3034
Saudi Arabia	0	0	0	0	0	0	0	1502
Serbia	1029	1029	1027	1030	1031	1029	1033	4739
Singapore	3975	3975	2006	3976	3976	3975	3976	5463
Slovakia	1197	1197	1196	1191	1195	1195	1195	2761
Vietnam	1200	1200	1200	1200	1200	1200	1200	3695
Slovenia	1061	1061	0	1060	1065	1060	1062	3100
South Africa	3531	3531	0	3531	3531	3531	3531	16,694
Zimbabwe	2710	2710	1179	2710	2713	2676	2679	3713
Spain	0	0	0	0	0	0	0	6319
Sweden	1194	1194	0	1197	1199	1197	1198	7354
Switzerland	0	0	0	0	0	0	0	3845
Tajikistan	1200	1200	1200	1200	1200	1200	1200	1200
Thailand	2669	2669	1489	2658	2662	2663	2672	4216
Trinidad and Tob	992	992	0	989	993	985	987	2000
Tunisia	2376	2376	1191	2385	2403	2370	2376	2410
Turkey	3997	3997	2405	3996	3989	3980	3995	11,686
Uganda	0	0	0	0	0	0	0	1002
Ukraine	2745	2745	1249	2666	2705	2728	2741	6587
North Macedonia	0	0	0	0	0	0	0	2050
Egypt	2717	2717	1194	2718	2722	2715	2717	8770
United Kingdom	2596	2596	2598	2597	2599	2594	2599	4586
Tanzania	0	0	0	0	0	0	0	1146
United States	4729	4729	2562	4735	4727	4737	4720	12,962
Burkina Faso	0	0	0	0	0	0	0	1479
Uruguay	1978	1978	996	1977	1987	1977	1985	3999
Uzbekistan	1500	1500	0	1500	1500	1500	1500	1500
Venezuela	1190	1190	1190	1190	1190	1190	1190	3590
Yemen	955	955	0	981	989	885	890	1000
Zambia	0	0	0	0	0	0	0	1500
Northern Ireland	446	446	445	445	446	445	445	414
Total	178,339	178,339	90,726	178,495	178,588	177,286	177,714	439,978

Source: Own computation using the tabstat command in Stata 17 (tabstat E260B E260BBin E253B E254B E259B E261B E262B X002, by(S003) statistics (count)).

), confirming consistent coding and support.

The methodological fusion followed a hierarchical and iterative process rather than a simple performance-based comparison. Each stage built systematically on the previous one (as shown in Figure 1), starting with broad feature selection (STAGE 1) and progressively narrowing down variables through regularization (RLASSO/CVLASSO), Bayesian averaging, multicollinearity and reverse causality diagnostics, and other validation tests and checks (mixed-effects, neural networks, nomograms) confirming the robustness and generalizability of findings. In this way, the framework operationalized fusion through multi-stage triangulation, integrating data mining, statistical, and machine learning approaches in a sequential–confirmatory design that balanced accuracy with interpretability.

While the intermediate stages of the methodology used (Figure 1 and Figure A1, Appendix A) enhance robustness, the pipeline converges on an interpretable LOGIT model and nomogram (Stage 9), ensuring findings remain transparent and policy-relevant.

Moreover, a table synthesizing and contrasting the conceptual limitations of longitudinal models with the necessity of ensemble methods (Figure 1 and Figure A1, Appendix A) for robust, policy-relevant insights in repeated cross-sectional survey data is available in Appendix A (

Table A1. Longitudinal vs. Multi-Stage Ensemble Approach.

Criterion	Straightforward Longitudinal Model	Multi-Stage Ensemble (This Study)
Data compatibility	Requires balanced panel	Handles repeated cross-sections
Predictor robustness	Prone to collinearity/spuriousness	Triangulated stability
Nonlinear/heterogeneous effects	Assumed linear/homogeneous	Captured (HGB, subgroup analysis)
Causal inference support	Limited	Reverse causality checks (Stage 7)
Policy interpretability	Coefficients only	Nomogram-based simulation

).

The WVS dataset v 4.0 was prepared for the initial selection stage (Stage 1) using the ADA BOOST technique implemented in the Rattle (v 5.5.1) library of R (v 4.1.3, x64) following approaches outlined in previous studies [102,103,104]. We developed, tested, and executed a customized script sequence for this purpose. This sequence, summarized in Figure 1 and well documented in Listing A1 (Appendix A) and implemented using REMDKNA v 1.1, a custom Stata command (.ado file) [105] included several steps, such as removal of the DK/NA values coded as negative values in the WVS dataset by assimilation to NULL (see

Table A9. Frequency of values for the target variable (E260B) before (top) and after (bottom) removing the artificial increase in scales due to the original encoding of DK/NA values as negative numbers by WVS (1remove_DKNA.do or the remdkna command in Stata).

E260B—Information Source: Mobile Phone (B)	Freq. (Before REMDKNA)	Percent (Before REMDKNA)	Cum. (Before REMDKNA)
Missing; Unknown	347	0.08	0.08
Not asked	270,415	59.98	60.05
No answer	1008	0.22	60.28
Don’t know	760	0.17	60.45
Daily	84,763	18.8	79.25
Weekly	18,483	4.1	83.34
Monthly	8301	1.84	85.19
Less than monthly	11,212	2.49	87.67
Never	55,580	12.33	100
Total	450,869	100	-
E260B—Information Source: Mobile Phone (B)	Freq. (After REMDKNA)	Percent (After REMDKNA)	Cum. (After REMDKNA)
Daily	84,763	47.53	47.53
Weekly	18,483	10.36	57.89
Monthly	8301	4.65	62.55
Less than monthly	11,212	6.29	68.83
Never	55,580	31.17	100
Total	178,339	100	-

Source: Own processing using the script sequence in Listing A1.

, Appendix B), adjusting the scales, and creating a binary-derived target variable (E260BBin detailed in Listing A2, Appendix A) based on other research [106]. We have used both forms of the targeted variable throughout the analysis to support selection, validation, and robustness testing. The latter ensured no informational loss or bias (during the preprocessing stage).

The main reason for removing DK/NA values was to avoid artificially increasing the scales, augmenting collinearity, and distorting accuracy measurements, together with other performance metrics for models, and disastrous effects on data mining and model exploration tasks, as other researchers reported [105,107]. Using similar data, the study referred to by the latter reference shows such examples of undesirable effects. These include false indications of redundancy due to artificially inflated variable scales and resulting collinearity, misleading rankings based on exaggerated predictor effect magnitudes, and unnecessary complexity in the variable selection process during the exploratory approach. The underlying rationale of REMDKNA focuses on recognizing key limitations of data imputation practices. These approaches replace missing values with the mean, median, or otherwise generated values, either across the entire dataset or within specific clusters (such as those based on socio-economic criteria). It also depends on the specifics of the data imputation approach applied [108,109]. Still, there is no guarantee for the respondents corresponding to the missing responses (in terms of proper approximation using one or more data imputation approaches). The respondent might not have responded in the way suggested by such approaches. Therefore, the DK/NA values were first assimilated to NULL (by REMDKNA). No bias was allowed by further avoiding data imputation (for ADA BOOST and all other conservative approaches in Stata). While reducing the sample size, this approach was a deliberate and necessary trade-off to ensure the authenticity and reliability of the model inputs, ultimately leading to a more robust and trustworthy final model. Still, when using HGB in Python (Stage 4, Appendix A, Figure A1), a more flexible approach was considered, namely the median simple imputation strategy. The latter replaces each missing value in a dataset with the median value of the variable’s available, non-missing data points in a straightforward and computationally inexpensive way. Unlike the mean, the median as a measure of central tendency is not affected by extreme outlier values.

The output obtained in the initial selection (Stage 1), which leveraged ADA BOOST, served as input for the following stage. Thus, in stage 2, the PCDM tool (version 1.1, 12 July 2022) was deployed using Stata [110]. The methodological foundation for this stage relied on pairwise correlations (Listing A3, Appendix A), a technique corroborated by recent works [42,111,112]. A key aspect considered in this stage was the inclusion of both forms of the target variable.

The results from the second stage subsequently served the third (Stage 3), which utilized two commands from the LASSO pack in Stata, namely RLASSO (version 1.0.10 13 January 2019 in Stata) which served as the principal method for overfitting removal used in this research, and CVLASSO (version 1.0.09 28 June 2019 in Stata) utilized as the base form of cross-validation used in this study [113,114,115]. RLASSO is known for its efficiency in overfitting removal, while CVLASSO performs random cross-validation (Listing A4, Appendix A).

The analytical process continued with BMA (Stage 4), which further refined model robustness based on the PIP (PIP—Posterior Inclusion Probability representing the probability of a predictor being a part of the valid model) value of possible predictors. The multi-collinearity analysis (Stage 5) was the next one. Then the model evaluation based on different types of regressions (Stage 6) and reverse causality checks (Stage 7) using OLogit and OProbit.

Beyond random criteria (CVLASSO), this study also employed well-established mixed-effects models [116] to perform non-random cross-validations (Stage 8). For the binary outcome, this involved using MELOGIT or Multilevel Mixed-Effects LOGIT in Stata (version 1.1.10, 5 May 2021) and MEPROBIT or Multilevel Mixed-Effects PROBIT in Stata (version 1.1.11, 5 May 2021) regressions for the binary form of the outcome and MEOLOGIT or Multilevel Mixed-Effects Ordinal LOGIT in Stata (version 1.3.3, 5 May 2021) and MEOPROBIT or Multilevel Mixed-Effects Ordinal PROBIT in Stata (version 1.3.3, 5 May 2021) for the original format of the target. The variables used for such non-random cross-validations are widely accepted. They are: gender (X001), age (X003), marital status (X007), number of people in household (X013), education level (X025R), employment status (X028), social class (X045), settlement size (X049), country code (S003), and year of survey (S020). We chose these variables above due to their relevance and high proportion of valid observations.

For Stage 5, we examined the remaining influences obtained from the previous stages for collinearity using a matrix with correlation coefficients [117,118]. The correlation coefficients mean an assessment of both the intensity and orientation of the linear relation between pairs of variables. Identifying redundancies or high correlations between the remaining influences relies on analyzing these coefficients. Redundancies or high correlations between predictors can indicate collinearity. The latter can affect the stability and interpretability of the regression models. A matrix with correlation coefficients allowed a comprehensive assessment of the relationships between the remaining influences. By examining the magnitude of the correlation coefficients, it was possible to identify highly correlated and potentially redundant variables. In such cases, we should remove one of the highly correlated variables to mitigate the issue of collinearity and improve model stability and interpretability. The final set of non-redundant influences resulted after considering the correlation coefficients and addressing all collinearity issues. That formed the basis for the subsequent analysis and interpretation of the regression models.

In addition, the collinearity assessments (Stage 5) stood on considering dynamic maximum acceptable limits for VIF (VIF—Variance Inflation Factor measuring collinearity in regression models) in two forms: the first one is focused on identifying pairs of collinear influences [119,120,121,122], while the second had as the primary goal to identify overall multi-collinearity issues considering all remaining influences at once [123,124].

In Stage 6, we employed various regression models for both outcome forms, which served to test the remaining influences. These models included OLS (Ordinary Least Squares), LOGIT, SCOBIT as Skewed LOGIT, and PROBIT for the binary format of the target variable, as well as OLOGIT or Ordered LOGIT, and OPROBIT/Ordinal PROBIT for the original form considering a scale [125,126,127,128]. These different regression models were employed to thoroughly examine the impact of the remaining influences on the target variable and assess their significance.

To conclude, the regression models used address the specific characteristics of the data and the objectives of the analysis. Therefore, mixed-effects models (MELOGIT, MEPROBIT, MEOLOGIT, and MEOPROBIT) considered the target variables in both binary and ordinal form, served for non-random cross-validations (with precisely defined criteria—Stage 8). OLS regressions handled collinearity, while the LOGIT ones served to measure precision, but also to generate prediction nomograms (target in binary format). SCOBIT addressed non-normal distribution disorders, and PROBIT models were used alternatively for binary data. Finally, we used OLOGIT/OPROBIT for the variable to be analyzed in ordinal format, ensuring that each model choice is optimal from this point (the original scale format of the target) as well.

Reporting model evaluation metrics is a mighty aspect of analyzing regression models (Stage 6). In addition to commonly used evaluation metrics such as R-squared, AIC (Akaike Information Criterion), BIC (Bayesian Information Criterion), and p-values (significance), there are additional tools and commands in Stata that can assist in reporting default and custom model evaluation metrics [129]. One such tool is ESTOUT in Stata (version 3.17, 2 June 2014), developed by [130]. ESTOUT provides a comprehensive summary of regression models, including model fit statistics, coefficient estimates, standard errors, confidence intervals, and other model diagnostics. It is a streamlined instrument that helped us generate tables with model results and evaluation metrics for multiple models. Another tool is MEM (Model Evaluation Metrics, version 1.1, 24 August 2022, in Stata), developed by [124]. MEM is a flexible and customizable command for model evaluation in Stata. It helps users define and compute various model evaluation metrics starting from their specific requirements. It supports the generation of tables with coefficients, standard errors, p-values, and other model evaluation metrics for different regression models. ESTOUT and MEM provide Stata users with additional options and flexibility in reporting model evaluation metrics beyond the default options available in Stata. These tools are effective for generating comprehensive and customized tables that summarize the results of regression models. Thus, they make it easier to interpret and perform comparative evaluations in terms of the performance (of different models).

The NOMOLOG (Nomogram generator for LOGIT in Stata) command was also used (version 4.76.0f, 26 August 2014) [131]. The latter is a Stata package that allows for the construction of nomograms [132,133,134,135,136], which are graphical representations of logistic regression models. To install the NOMOLOG package, we can use the following command in Stata: net install st0391_1, from (http://www.stata-journal.com/software/sj15-3). The NOMOLOG command provides a visual image (nomogram) of the predicted probabilities and the effects of different predictor variables on the outcome for a LOGIT model. This nomogram brings an intuitive way to compare collinear variables during some selection steps (Stage 5) by analyzing their relative magnitudes, as reflected in the lengths of the segmented lines. In interpreting the final model (Stage 9), a nomogram allows users to make predictions by aligning predictor variable values with corresponding scores, which then sum up to calculate a total score and probability. Using a nomogram involves drawing imaginary perpendicular lines on the X-axis to locate corresponding scores, enabling a straightforward calculation of probabilities with minimal effort. Supplemental guidance on how to interpret nomograms and turn them into insights or even competitive diagnostic and predictive tools (augmented nomograms) is also available in this paper. The latter relies on using at least two similar examples [137,138].

To ensure robustness in the models and account for regional peculiarities, we performed tabulations by count using tabulate in tabstat in Stata (version 2.9.5, 24 May 2021). The tabulations examined the support (waves and countries) in the WVS dataset [139], ensuring comprehensive representation. They also provided information on the distribution and frequency of observations across different waves and countries. Another scope of these tabulations is to gain insights into the representation and variability of the data, allowing for a comprehensive understanding of the data characteristics.

All Stata scripts (for processing, selection, and analysis) utilized in this study are available online in two persistent Google Drive containers shared and accessible through the following links: https://tinyurl.com/ms3n3s9v (for WVS, version 4.0 and IVS data) and https://tinyurl.com/2utdjpmr (only for WVS data, version 5.0). To ensure seamless downloading of most scripts, the corresponding URL (Uniform Resource Locator) benefited from an update in the specific syntax (&export=download at the end). This modification eliminates the need for additional confirmation steps, allowing for a straightforward one-step download process.

Furthermore, the proposed multi-stage framework was designed for both scalability and reproducibility. Each analytical stage relies on widely available, cross-platform software (R v 4.1.3, Stata v17, and Python 3.12.7) and open-source libraries (e.g., Rattle, scikit-learn). This structure enables other researchers to replicate the entire workflow or extend it to different datasets with minimal adaptation. The full scripts and parameter settings publicly available via the provided persistent links ensure that every step, from data preprocessing to validation, can be transparently re-executed and independently verified.

The complexity of the proposed framework (Figure 1 and Figure A1, Appendix A) is also due to the principal motivation related to robust research and identifying robust predictors (as suggested in the title). Specifically, this is an accepted practice in data science, according to the scientific principles of triangulation, cross-validation, and replicability. Deep learning techniques were not included because the study’s primary goal was to identify stable, interpretable predictors across statistical and machine-learning paradigms. Deep learning models typically function as low-interpretability black boxes and require substantially larger datasets and GPU-based hardware to yield meaningful gains for tabular data. Accordingly, we selected methods better suited to explanatory modeling and cross-method triangulation.

The triangulation framework described in this paper is conceptually distinct from collaborative or hairy modeling approaches. Unlike collaborative modeling, which relies on stakeholder co-creation of a single shared model through workshops or elicitation, and hairy modeling, which integrates heterogeneous real-time data into complex, multi-scale simulations, our method applies multiple independent statistical algorithms (e.g., logistic regression, HGB, AdaBoost) to pre-existing, harmonized survey data (EVS/WVS) for convergent validation of a fixed predictor set. No stakeholder input is used in model construction, and no dynamic system simulation is performed. This methodological triangulation [140] aims solely to reduce method-specific bias and confirm robustness across analytical lenses—not to build consensus or model emergent complexity.

Still, this triangulation framework is dataset-agnostic and applies to any large-scale behavioral survey. While Stata commands such as REMDKNA, PCDM, and VCPR leverage the standardized missing-value system of WVS/EVS for efficient preprocessing, the methodological core (convergent validation via multiple algorithms) is software-independent and can be replicated in R and Python with other data.

We acknowledge that the methodological framework employs a wide range of statistical and DM techniques. Statistical methods (e.g., LOGIT, OLOGIT, OPROBIT, MELOGIT, MEOLOGIT) provide rigorous hypothesis testing and model validation, grounding the analysis in established econometric practices. DM techniques (e.g., REMDKNA, PCDM, NOMOLOG) typically serve data preprocessing, pattern discovery, and feature selection. ML techniques (e.g., Ada Boost, RLASSO, CVLASSO, BMA, HGB, and Brain NN) focus on predictive modeling, classification, or regression. However, many methods (e.g., AdaBoost, HGB) straddle DM categories due to their versatility. This categorization stands on the primary application of these techniques in this study, with justifications grounded in the manuscript (Appendix A) and the literature [141,142,143]. We intentionally chose this approach. It does not add unnecessary complexity and ensures the robustness and validity of our findings from multiple perspectives. The initial use of techniques such as AdaBoost, RLASSO, and CVLASSO (Stages 1, 3) served as a crucial first step for objective feature selection and regularization, helping to identify the most significant predictors from our extensive dataset. Subsequently, we employed diverse models in a multi-stage process (Stages 4, 6, 8, 10), including BMA, HGB, mixed-effects models, and a neural network (Brain NN) to demonstrate that our key findings were not an artifact of a single modeling approach and were consistent across different algorithmic assumptions. This multi-model validation significantly strengthens the confidence in our results. Finally, the development of the nomogram (Stage 9) also aims to bridge the gap between sophisticated data science and practical application. It consolidates the insights from our best predictive model into a simple, visual tool that offers actionable, real-world utility for practitioners and policymakers who may not have a background in data science. In essence, each method serves a distinct purpose, from feature selection and model validation to creating a practitioner-friendly output. Moreover, the robustness of this multi-stage framework stems from methodological triangulation [144], integrating various algorithms (AdaBoost, RLASSO, CVLASSO, BMA, HGB, Brain NN), DM (PCDM, REMDKNA, NOMOLOG), and statistical methods (LOGIT, OLOGIT, OPROBIT, MELOGIT, MEOLOGIT, MEPROBIT) to mitigate single-method limitations like overfitting or bias [145]. This combination cross-verifies predictors. Moreover, AdaBoost ranks features. RLASSO/CVLASSO ensures sparsity. BMA addresses uncertainty, and HGB captures nonlinearities. The advantage of this combination is the fact that it is empirically validated by consistent identification of core predictors across three diverse datasets—World Values Survey (WVS v4.0, v5.0) and Integrated Values Survey (IVS)—with AUC-ROC greater than 0.85 (Stages 6–8), reflecting real-world generalizability [146].

3.2. Additional Tests and Validations

Additionally (Stage 10), the Brain NN pack (version 10.0) in Stata (https://github.com/ThorstenDoherr/brain, accessed on 1 March 2025) is responsible for implementing backpropagation artificial neural networks. It was used in several configurations (varying the number of hidden layers, the training factor eta, and the number of backpropagation iterations, with the batch size set to 100). The goal of the lightweight feed-forward neural networks implemented using Stata’s Brain NN pack was to assess the stability of the selected predictors under a nonlinear model. This neural architecture is computationally efficient for tabular data and complements the triangulated design by providing a non-parametric benchmark without the opacity and data requirements of deep learning models. In addition, Brain NN comes with the benefit of the parallelization of training.

Moreover, additional tests focused on two possible predictors not selected using this approach (ADA BOOST, CVLASSO, and RLASSO). These predictors were X025R—Highest educational level attained: Lower, Middle, and Upper, and X001-Sex: Male and Female. The tests additionally considered BMA (version 1.0, 6 May 2011 in Stata) [147] that, given a few candidate parametric families, generates the Posterior Inclusion Probabilities (PIP) of possible predictors (as a form of quantifying the input uncertainty). More than this, a Histogram-based Gradient Boosting (HGB) algorithm [148] with 10-fold cross-validation, shuffling, parallel execution, and accuracy and importance calculations (top 25 features after three repeats) was tested (stage 4) in Python 3.12.7 via a Spyder IDE (for WVS data in both versions, namely 4.0 and 5.0 (Each .py file (https://tinyurl.com/5bwwa82a and https://tinyurl.com/mpa2hb6f) in both online supporting folders (https://tinyurl.com/ms3n3s9v and https://tinyurl.com/2utdjpmr))). Following the same logic behind CVLASSO in Stata (by default, 10-fold), the primary purpose of using the 10-fold stratified cross-validation approach for HGB [149] is to ensure the robustness of our models. The 10-fold stratified cross-validation process involves partitioning the dataset into 10 equally sized subsets, or folds. During each of the 10 iterations, we reserved a single fold (10% of the data) for testing. The remaining nine folds (90% of the data) are for training the model. The latter repeats 10 times, until every fold has served as the test set exactly once. This method ensures that the class distribution of the target variable is maintained consistently across both training and testing sets, preventing potential bias in the evaluation. We assessed the model performance by calculating the accuracy score for each test fold [150]. Starting from HGB, a partial dependence plot (PDP) [151] drawing the marginal effect of each predictor on the predicted probability and two SHAP (Shapley Additive Explanations) [152] representations were obtained (a global importance chart showing the strongest predictors, and a summary Bee Swarm plot illustrating the distribution and direction for each individual feature—Appendix D, Figure A6a,b and Figure A7). Moreover, we additionally tested different regression types (SCOBIT in Stata—version 1.12.1, 19 December 2018; LOGIT in Stata—version 11.3.1, 25 June 2020; PROBIT in Stata—version 11.3.1, 25 June 2020; OLOGIT in Stata—version 11.4.2, 25 June, 2020; ordered PROBIT or OPROBIT in Stata—version 11.4.2, 25 June 2020; and OLS named regress in Stata—version 1.3.3, 9 December 2020) in models including these two potential candidates. Some other tests (similar ones) relied on the second dataset (IVS), which supports the validations.

Regarding the normality of distributions for the final models, the Shapiro–Wilk test for normality of data in Stata served. However, it was complemented by generating histograms designed for displaying and inspecting the distribution of variables [153,154], ensuring a comprehensive evaluation. The rejection of normality for certain variables justified the use of alternative modeling approaches, such as the SCOBIT regression, which accommodates deviations from normal or logistic distributions. This combined approach strengthens the robustness of the findings by addressing potential distributional concerns.

Regarding heteroskedasticity, the vce(robust) option in Stata supported the generation of robust standard errors [155,156]. The latter is responsible for correcting the error bias that may arise from the non-constant variance in residuals, improving the reliability of coefficient estimates. This approach commonly supports heteroskedasticity-robust inference. However, given the widespread application of robust standard errors in the regression models, the impact of heteroskedasticity on inference is expected to be minimized.

3.3. Demographic Analysis

Demographic analysis is also available to offer a comprehensive understanding of information retrieval behavior through mobile devices. Such an analysis usually includes dedicated stages. Identifying relevant demographic variables is the first step. The latter is followed by reporting the descriptive statistics, tabulations, cross-tabulations, and data visualization for the input variables and each of the demographic categories. The last ones also include statistical tests, regression analyses, and interpretations.

In terms of testing the linearity, scatter plot graphs (between the dependent variable and each from the list of independent ones) have been generated [157,158].

4. Results

4.1. Outcomes of the Main Selection Steps

The following section includes detailed results. They emerged after each selection stage and additional tests (as described in Section 3).

After conducting the initial selection (Stage 1) using ADA BOOST within the Rattle pack of R (accessible at https://rattle.togaware.com), on the exported .csv file of the cleaned WVS dataset, a set of 23 variables emerged.

These 23 variables were identified as influential during the selection process (Stage 1), as depicted in Figure 2.

The following step involved determining the number of valid observations for the target variables at the beginning of the second selection stage (Stage 2). The custom command PCDM (pcdm E260B E260BBin) was used for this purpose, resulting in 178,339 non-NULL records. This information supported the choice of the minimum support parameter for the subsequent launch of PCDM. The minimum support considered was one plus a half of the total valid observations (as a number) for the target variable. That meant approximately 89,170. Additionally, the maximum accepted p-value (significance) is 0.001 (meaning at most one error in a thousand cases). Moreover, 0.1 is the minimum acceptable absolute value of the correlation coefficient (non-negligible correlations).

In the second launch of PCDM, only 22 out of the 23 variables from the first stage were considered, excluding the variable COUNTRY_ALPHA (a string variable). The PCDM command was executed with the following variables: E260B, A104, E116, E250B, E253B, E254B, E255, E258B, E259B, E261B, E262B, H002_03, H006_02, mode, S006, S007, S012, S020, S022, S023, X002, X003, and X003R. The same command also ran with the variable E260BBin.

The intersection of the results from this stage (Stage 2) yielded only 14 variables: E253B, E254B, E259B, E261B, E262B, S007, S012, S020, S022, S023, X002, X003, and X003R.

These variables served as input for the subsequent selections (Stage 3) based on RLASSO and CVLASSO.

In the first run of RLASSO for both forms of the outcome (RLASSO E260B A104 E253B E254B E259B E261B E262B S007 S012 S020 S022 S023 X002 X003 X003R and RLASSO E260BBin A104 E253B E254B E259B E261B E262B S007 S012 S020 S022 S023 X002 X003 X003R), only seven variables emerged (A104, E253B, E254B, E259B, E261B, E262B, and X002). They were the intersection subset for both forms of the target.

In the first run of CVLASSO, starting from the remaining variables, only A104 was dropped. In the second run of RLASSO and CVLASSO, considering both outcome forms and the remaining six variables, no loss emerged.

For a detailed description of these variables, see

Table A3. The names, descriptions, and coding of the most relevant WVS items (version 4.0 of this dataset) used in this study.

Variable	Short Description	Coding Details
E260B	Information source: Mobile phone (target variable—scale form)	1-Daily, 2-Weekly, 3-Monthly, 4-Less than Monthly, 5-Never
E260BBin	Information source: Mobile phone (target variable—binary format)	1-for E260B >=1 and <=4; 0-for E260B = 5
E253B	Information source: Social media (Facebook, Twitter, etc.)	1-Daily, 2-Weekly, 3-Monthly, 4-Less than Monthly, 5-Never
E254B	Information source: Talk with friends or colleagues	1-Daily, 2-Weekly, 3-Monthly, 4-Less than Monthly, 5-Never
E259B	Information source: Radio news	1-Daily, 2-Weekly, 3-Monthly, 4-Less than Monthly, 5-Never
E261B	Information source: E-mail	1-Daily, 2-Weekly, 3-Monthly, 4-Less than Monthly, 5-Never
E262B	Information source: Internet	1-Daily, 2-Weekly, 3-Monthly, 4-Less than Monthly, 5-Never
X002	Year of birth	values between 1886 and 2004

Source: Own findings using seven times the LABEL LIST command in Stata 17.

. For descriptive statistics, refer to

Table A4. Descriptive statistics for the key WVS items (version 4.0 of this dataset) used in this study after removing their DK/NA values and executing all selection steps except collinearity removal but not specifying the non-NULL condition.

Variable	N (Obs.)	Mean	St.Dev.	Min	Median	Max
E260B	178,339	2.63	1.78	1	2	5
E260BBin	178,339	0.69	0.46	0	1	1
E253B	90,726	2.69	1.8	1	2	5
E254B	178,495	2.18	1.42	1	2	5
E259B	178,588	2.81	1.69	1	2	5
E261B	177,286	3.57	1.67	1	5	5
E262B	177,714	2.9	1.8	1	2	5
X002	439,978	1964.43	18.35	1886	1966	2004

Source: Own calculations using the UNIVAR command (version 1.1.2, 1 November 1997) in Stata 17 followed only by the entire list of the variables above (univar E260B E260BBin E253B E254B E259B E261B E262B X002).

, Appendix B.

In the next stage (Stage 4), BMA and HGB preserved all six remaining variables (no loss). In fact, they have also confirmed the intersection between CVLASSO, RLASSO, and HGB when using the 4.0 version of WVS data (top 25 features—https://tinyurl.com/bdcpf3ez, accessed on 1 March 2025—and accuracy scores on each test fold of almost 0.83 or even more (The first .csv file (https://tinyurl.com/yc3vbkry) in the first online supporting folder (https://tinyurl.com/ms3n3s9v))).

During the verification of collinearity (Stage 5), a correlation coefficient matrix was examined (Figure 2—pwcorr E253B E254B E259B E261B E262B X002 using pwcorr—version 3.1.1 1 May 2018 in Stata), specifically exploring the correlations between these six remaining variables E253B, E254B, E259B, E261B, E262B, and X002. The analysis revealed two instances of collinear combinations: E253B-E262B and E261B-E262B. These combinations exhibited moderate to strong correlation coefficients, indicating a high correlation between the predictors [118]. Consequently, in this case, the variable E262B (Information source: Internet) was dropped from the analysis (refer to

Table A5. Descriptive statistics for the key WVS items (version 4.0 of this dataset) used in this study after removing their DK/NA values and executing all selection steps, including collinearity removal and the non-NULL condition specified for all variables.

Variable	N (Obs.)	Mean	St.Dev.	Min	Median	Max
E260B	87,741	2.44	1.73	1	1	5
E260BBin	87,741	0.73	0.44	0	1	1
E253B	87,741	2.7	1.8	1	2	5
E254B	87,741	2.35	1.45	1	2	5
E259B	87,741	3.1	1.69	1	3	5
E261B	87,741	3.55	1.66	1	4	5
X002	87,741	1975.64	16.43	1916	1978	2004

Source: Own calculations using the UNIVAR command (version 1.1.2, 1 November 1997) in Stata 17 followed by the entire list of the variables above and a non-NULL condition for all (univar E260B E260BBin E253B E254B E259B E261B X002 if E260B!=. & E260BBin!=. & E253B!=. & E254B!=. & E259B!=. & E261B!=. & X002!=.).

, Appendix B). This decision resulted from the fact that E262B appears in two collinearity issues: with E253B (Information source: social media) and E261B (Information source: E-mail). Moreover, E261B exhibited a greater magnitude (larger segment in Figure A2, Appendix C).

An additional test (available online at https://tinyurl.com/yc5e86pv) using the VIF (assessed against a maximum limit depending on the model R-squared) supported the confirmation of existing collinearity [119]. It discovered again those two pairs, namely E253B-E262B and E261B-E262B. Moreover, a prediction nomogram with six influences (Figure A2, Appendix C) supported the confirmation of removal for E262B. The latter applies when considering its lower overall graphical effect (when measured against that of E261B) and its equality in such terms with E253B.

The removal of E262B (Information source: Internet) means an expected tiny decrease in the classification accuracy (from 0.8772 to 0.8714). However, it also led to a consistent reduction in redundancy in the form of the maximum value of the correlation coefficient between predictors (from ~0.71—between E262B and E253B to ~0.49—between E261B and E253B), as seen in Figure 3 and in the performance results captured and available at https://tinyurl.com/yc5e86pv.

This removal (E262B) was a necessary step to mitigate multi-collinearity, which can inflate the variance and standard error of coefficient estimates, potentially distorting the reliability of the prediction model. A minor trade-off in classification accuracy consequently resulted from the prioritizing of model interpretability and stability. The latter ensures that the remaining predictors provide unbiased and meaningful contributions.

In addition, a new stage of collinearity checks focused on the remaining five influences. This time, two types of VIF-based collinearity checks have the focus (the first one is focused on identifying pairs of collinear influences (VCPR-version 1.3 21 November 2023 in Stata [122]), while the second had as a goal to identify and report overall multi-collinearity issues considering all five influences at once (MEM-version 1.1 24 August 2022 in Stata) and a computed maximum VIF value measured against the maximum tolerance which is equal with 1/(1−model R–squared) according with [119,120,121,122]).

No further evidence of collinear pairs and multi-collinearity emerged (do script and console.pdf capture with commands, their results, and comments available at the end of the online resources at https://tinyurl.com/2avjc4n9 and https://tinyurl.com/3b9y2d2t). The latter also stands on the tests performed on IVS and WVS, version 5.0 (validation datasets) available at the end of the online resources at https://tinyurl.com/hru48zzs and https://tinyurl.com/2nck8zdr (IVS) and https://tinyurl.com/mrxz3d59 and https://tinyurl.com/27vzw73s, respectively (WVS, version 5.0).

Following the step above, several regression models emerged (Stage 6). These models consistently confirmed the presence of the five variables that remained influential throughout the analysis. The details of these variables, along with their corresponding regression results, are depicted in

Table A6. Different regression models as built using the five most relevant remaining influences identified in the study to determine the relationship between these influences and the target variable based on WVS data (version 4.0).

Model No.	(1)	(2)	(3)	(4)	(5)	(6)	(7)
Target Variable	E260BBin (Information source: Mobile phone—binary form)	E260BBin	E260BBin	E260BBin	E260B (Information source: Mobile phone—scale form)	E260B	E260B
Regression Type	OLS	LOGIT	SCOBIT	PROBIT	OLS	OLOGIT	OPROBIT
E253B (Information source: Social media)	−0.0740 ***	−0.4186 ***	−0.9218 ***	−0.2417 ***	0.3305 ***	0.4410 ***	0.2557 ***
	(0.0010)	(0.0063)	(0.0548)	(0.0035)	(0.0038)	(0.0053)	(0.0030)
E254B (Information source: Talk with friends or colleagues)	−0.0443 ***	−0.2651 ***	−0.4785 ***	−0.1557 ***	0.2092 ***	0.3117 ***	0.1810 ***
	(0.0011)	(0.0064)	(0.0220)	(0.0038)	(0.0040)	(0.0057)	(0.0033)
E259B (Information source: Radio news)	−0.0211 ***	−0.1498 ***	−0.2538 ***	−0.0893 ***	0.0679 ***	0.0957 ***	0.0588 ***
	(0.0008)	(0.0059)	(0.0137)	(0.0034)	(0.0029)	(0.0047)	(0.0027)
E261B (Information source: E-mail)	−0.0533 ***	−0.6341 ***	−1.7006 ***	−0.3199 ***	0.2254 ***	0.3603 ***	0.2183 ***
	(0.0008)	(0.0116)	(0.1161)	(0.0054)	(0.0033)	(0.0054)	(0.0032)
X002 (Year of birth)	0.0028 ***	0.0203 ***	0.0341 ***	0.0121 ***	−0.0093 ***	−0.0134 ***	−0.0082 ***
	(0.0001)	(0.0006)	(0.0015)	(0.0004)	(0.0003)	(0.0005)	(0.0003)
_cons	−4.2237 ***	−33.7885 ***	−50.5590 ***	−20.5299 ***	18.4061 ***
	(0.1693)	(1.2081)	(2.4583)	(0.7010)	(0.6186)
lnalpha			−1.1583 ***
			(0.0650)
cut1						−22.8815 ***	−14.0814 ***
						(0.9543)	(0.5605)
cut2						−22.1901 ***	−13.6788 ***
						(0.9544)	(0.5605)
cut3						−21.8614 ***	−13.4884 ***
						(0.9544)	(0.5606)
cut4						−21.4249 ***	−13.2367 ***
						(0.9543)	(0.5605)
N	87,741	87,741	87,741	87,741	87,741	87,741	87,741
chi-squared		17,218.9490		19,397.2846		30,872.9575	34,542.1763
p	0.0000	0.0000		0.0000	0.0000	0.0000	0.0000
R-squared	0.3287	0.3320		0.3285	0.3983	0.1890	0.1891
AIC	70,518.3523	67,801.3760	67,457.5315	68,153.9530	300,462.5556	177,927.0057	177,884.4839
BIC	70,574.6452	67,857.6689	67,523.2065	68,210.2458	300,518.8485	178,011.4450	177,968.9232
RMSE	0.3616				1.3408
AUC-ROC		0.8714		0.8708

Source: Own calculations in Stata 17 (the source script for generating this table is available at: https://drive.google.com/u/0/uc?id=1m7uq0pMp_4TCnj8uZ7ZmmK72_9rLtI6K&export=download). Notes: Robust standard errors are between round parentheses. The raw coefficients emphasized using *** are significant at 1‰.

, Appendix B. These five variables are: E253B, E254B, E259B, E261B, and X002.

In addition, reverse causality checks (Stage 7 online available at http://tinyurl.com/42v7r8je) were performed using OLOGIT and Ordinal PROBIT (OPROBIT) regressions and the original form of the target variable (E260B). In each regression that considered only one of those five variables above, the latter served as input and target, interchanging these roles with E260B (regression pairs). The R-squared should be larger (a higher amount of variability explained by the model) and (or) AIC and BIC lower (better fit or more information value) in the models in which they served as input versus those in which they act as outcome. The latter would lead to the conclusion that each of those five behaves more as a determinant (cause) of E260B, rather than an effect of it.

All five variables above (all used in the second nomogram—Figure 4) passed these reverse causality checks when considered as input in five models with just one predictor. In contrast, the other five models, where they served as outcome, and the original target (E260B) as input, were tested with OLOGIT (results in

Table A10. Reverse causality checks using OLOGIT regressions.

Model	Model1	Model2	Model3	Model4	Model5	Model6	Model7	Model8	Model9	Model10
Input/Target Var.	E260B	E253B	E260B	E254B	E260B	E259B	E260B	E261B	E260B	X002
E253B (Information source: Social media)	0.6843 ***
	(−0.0045)
E254B (Information source: Talk with friends or colleagues)			0.5078 ***
			(−0.0035)
E259B (Information source: Radio news)					0.1898 ***
					(−0.0028)
E261B (Information source: E-mail)							0.6046 ***
							(−0.0032)
X002 (Year of birth)									−0.0292 ***
									(−0.0003)
E260B (Information source: Mobile phone—scale form)		0.7194 ***		0.3813 ***		0.1727 ***		0.5932 ***		−0.2491 ***
		(−0.0048)		(−0.0027)		(−0.0026)		(−0.003)		(−0.0024)
N	90,405	90,405	177,518	177,518	177,617	177,617	176,719	176,719	174,628	174,628
Chi-squared	23,128.5659	22,836.1391	20,665.4526	20,105.3049	4489.8956	4286.6316	35,243.6861	37,867.409	11,222.997	10,679.8848
P	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000	0.0000
R-squared	0.1375	0.1384	0.053	0.0459	0.0111	0.0094	0.0925	0.0943	0.0262	0.0076
AIC	195,027.077	197,208.765	426,536.286	467,382.566	445,602.806	511,724.316	406,784.477	421,084.352	430,886.692	1,445,048.39
BIC	195,074.137	197,255.825	426,586.72	467,433	445,653.243	511,774.753	406,834.889	421,134.764	430,937.044	1,445,964.8

Source: Own computation in Stata 17 (online script at https://drive.google.com/u/0/uc?id=1LHyJ_x95-ZW49X7bnRDe7DiIt_p0JBBR&export=download). Notes: Robust standard errors are between round parentheses. The raw coefficients emphasized using *** are significant at 1‰. Green font vs. Red font suggest better (smaller AIC and BIC, larger R-squared) vs. worse (vice versa) model for each pair (vertical borders).

, Appendix C) and OPROBIT (results in

Table A11. Reverse causality checks using OPROBIT regressions.

Model	Model1	Model2	Model3	Model4	Model5	Model6	Model7	Model8	Model9	Model10
Input/Target Var.	E260B	E253B	E260B	E254B	E260B	E259B	E260B	E261B	E260B	X002
E253B (Information source: Social media)	0.4029 ***
	(−0.0025)
E254B (Information source: Talk with friends or colleagues)			0.3053 ***
			(−0.0021)
E259B (Information source: Radio news)					0.1157 ***
					(−0.0017)
E261B (Information source: E-mail)							0.3662 ***
							(−0.0019)
X002 (Year of birth)									−0.0181 ***
									(−0.0002)
E260B (Information source: Mobile phone—scale form)		0.4224 ***		0.2291 ***		0.1041 ***		0.3545 ***		−0.1457 ***
		(−0.0026)		(−0.0016)		(−0.0016)		(−0.0018)		(−0.0014)
N	90,405	90,405	177,518	177,518	177,617	177,617	176,719	176,719	174,628	174,628
Chi-squared	25,945.9827	25,431.203	21,247.3917	21,521.2876	4594.0689	4412.5131	37,695.6852	38,542.2283	11,556.7561	11,130.298
p	0	0	0	0	0	0	0	0	0	0
R-squared	0.1355	0.1347	0.0516	0.047	0.0111	0.0094	0.0932	0.0941	0.0264	0.0079
AIC	195,461.138	198,055.951	427,136.336	466,861.672	445,629.349	511,758.999	406,463.445	421,156.312	430,780.016	144,4671.88
BIC	195,508.199	198,103.011	427,186.77	466,912.106	445,679.786	511,809.436	406,513.857	421,206.723	430,830.368	1,445,588.29

Source: Own computation in Stata 17 (online script at https://drive.google.com/u/0/uc?id=1j_zvk3CHmPgq6SvudNhK5dqWfnv0jT_0&export=download). Notes: Robust standard errors are between round parentheses. The raw coefficients emphasized using *** are significant at 1‰. Green font vs. Red font suggest better (smaller AIC and BIC, larger R-squared) vs. worse (vice versa) model for each pair (vertical borders).

, Appendix C). For relatively close values of R-squared, the lower ones of AIC and BIC confirmed these validations.

When performing non-random cross-validations (Stage 8—

Table A12. Non-random cross-validations using MELOGIT and MEOLOGIT regressions.

MELOGIT/MEOLOGIT	Model1	Model2	Model3	Model4	Model5	Model6	Model7	Model8	Model9	Model10	Model11	Model12	Model13	Model14	Model15	Model16	Model17	Model18	Model19	Model20
Outcome	E260BBin	E260BBin	E260BBin	E260BBin	E260BBin	E260BBin	E260BBin	E260BBin	E260BBin	E260BBin	E260B	E260B	E260B	E260B	E260B	E260B	E260B	E260B	E260B	E260B
E253B (Information source: Social media)	−0.4186 ***	−0.4179 ***	−0.4213 ***	−0.4169 ***	−0.4175 ***	−0.4166 ***	−0.4138 ***	−0.4152 ***	−0.4384 ***	−0.4162 ***	0.4408 ***	0.4382 ***	0.4422 ***	0.4385 ***	0.4403 ***	0.4401 ***	0.4407 ***	0.4390 ***	0.4589 ***	0.4371 ***
	(−0.026)	(−0.0066)	(−0.0108)	(−0.0077)	(−0.0521)	(−0.0149)	(−0.0137)	(−0.0197)	(−0.0298)	(−0.023)	(−0.0194)	(−0.0074)	(−0.0158)	(−0.0132)	(−0.0271)	(−0.009)	(−0.0119)	(−0.0141)	(−0.0282)	(−0.0169)
E254B (Information source: Talk with friends or colleagues)	−0.2650 ***	−0.2628 ***	−0.2640 ***	−0.2663 ***	−0.2656 ***	−0.2636 ***	−0.2644 ***	−0.2672 ***	−0.2579 ***	−0.2697 ***	0.3122 ***	0.3101 ***	0.3101 ***	0.3127 ***	0.3120 ***	0.3121 ***	0.3127 ***	0.3134 ***	0.2974 ***	0.3150 ***
	(−0.0286)	(−0.0081)	(−0.0046)	(−0.0088)	(−0.0232)	(−0.0196)	(−0.0064)	(−0.0145)	(−0.0163)	(−0.0165)	(−0.0181)	(−0.0067)	(−0.0051)	(−0.006)	(−0.0145)	(−0.0116)	(−0.0057)	(−0.0137)	(−0.0159)	(−0.0181)
E259B (Information source: Radio news)	−0.1498 ***	−0.1524 ***	−0.1510 ***	−0.1522 ***	−0.1491 ***	−0.1493 ***	−0.1459 ***	−0.1494 ***	−0.1688 ***	−0.1577 ***	0.0963 ***	0.0963 ***	0.0961 ***	0.0973 ***	0.0951 ***	0.0958 ***	0.0887 ***	0.0944 ***	0.1086 ***	0.0977 ***
	(−0.0115)	(−0.008)	(−0.012)	(−0.0089)	(−0.0049)	(−0.0115)	(−0.0075)	(−0.0119)	(−0.0129)	(−0.0047)	(−0.0107)	(−0.0056)	(−0.0085)	(−0.0077)	(−0.0053)	(−0.01)	(−0.0075)	(−0.0133)	(−0.0122)	(−0.0108)
E261B (Information source: E-mail)	−0.6340 ***	−0.6399 ***	−0.6370 ***	−0.6380 ***	−0.6328 ***	−0.6339 ***	−0.6243 ***	−0.6369 ***	−0.6739 ***	−0.6390 ***	0.3608 ***	0.3629 ***	0.3623 ***	0.3623 ***	0.3602 ***	0.3608 ***	0.3516 ***	0.3596 ***	0.4001 ***	0.3619 ***
	(−0.0183)	(−0.0192)	(−0.0078)	(−0.0581)	(−0.0431)	(−0.064)	(−0.0369)	(−0.0209)	(−0.0614)	(−0.0946)	(−0.0014)	(−0.0061)	(−0.0092)	(−0.0093)	(−0.0214)	(−0.0134)	(−0.0105)	(−0.0135)	(−0.0238)	(−0.042)
X002 (Year of birth)	0.0202 ***	0.0262 ***	0.0211 ***	0.0186 ***	0.0203 ***	0.0174 ***	0.0196 ***	0.0202 ***	0.0231 ***	0.0201 ***	−0.0134 ***	−0.0242 ***	−0.0142 ***	−0.0125 ***	−0.0134 ***	−0.0115 ***	−0.0130 ***	−0.0133 ***	−0.0150 ***	−0.0134 ***
	(−0.0012)	(−0.0013)	(−0.0009)	(−0.0028)	(−0.0038)	(−0.0023)	(−0.0013)	(−0.0013)	(−0.0036)	(−0.0049)	(−0.0003)	(−0.0022)	(−0.0009)	(−0.0024)	(−0.0016)	(−0.002)	(−0.0004)	(−0.0018)	(−0.0026)	(−0.0037)
_cons	−33.7671 ***	−45.4276 ***	−35.5040 ***	−30.4154 ***	−33.9865 ***	−28.2662 ***	−32.5541 ***	−33.7607 ***	−39.0467 ***	−33.4129 ***
	(−2.4971)	(−2.6269)	(−1.8546)	(−5.7798)	(−7.852)	(−4.8731)	(−2.6713)	(−2.5541)	(−7.361)	(−10.0081)
var(_cons [X001]) (Gender)	0.0000										0.0004 ***
	(0.0000)										(−0.0001)
var(_cons [X003]) (Age)		0.0360 ***										0.0455 **
		(−0.0073)										(−0.014)
var(_cons [X007]) (Marital status)			0.0069 *										0.0064 *
			(−0.0029)										(−0.0027)
var(_cons [X013]) (Number of people in household)				0.0137										0.0299
				(−0.011)										(−0.0333)
var(_cons [X025R]) (Education level)					0.0000										0.0000
					(0.0000)										(0.0000)
var(_cons [X028]) (Employment status)						0.0108 **										0.0056 *
						(−0.0034)										(−0.0023)
var(_cons [X045]) (Social class)							0.0078 ***										0.0017 **
							(−0.0021)										(−0.0006)
var(_cons [X049]) (Settlement size)								0.0071 *										0.0074 **
								(−0.0031)										(−0.0028)
var(_cons [S003]) (Country code)									0.5129 ***										0.4610 ***
									(−0.1021)										(−0.0887)
var(_cons [S020]) (Year of survey)										0.0747										0.0702
										(−0.0467)										(−0.0416)
N	87,704	87,734	87,379	87,165	87,071	86,856	82,972	86,686	87,741	87,741	87,704	87,734	87,379	87,165	87,071	86,856	82,972	86,686	87,741	87,741
AIC	67,770.6289	67,637.9946	67,480.3968	67,360.5005	67,364.793	67,060.8194	64,400.5809	67,076.2378	62,934.8751	67,257.6101	177,832.158	177,699.841	177,048.931	176,674.575	176,551.535	176,140.688	169,260.844	175,691.124	170,411.14	177,113.989
BIC	67,789.3924	67,703.669	67,536.6649	67,426.1294	67,383.542	67,126.4235	64,447.2122	67,141.8282	63,000.5501	67,304.5208	177,841.54	177,793.662	177,114.577	176,768.331	176,579.658	176,206.292	169,298.149	175,775.455	170,504.96	177,179.664

Source: Own computation in Stata 17 (online script at https://drive.google.com/u/0/uc?id=1_646hyruIgh88WBHTl6tG-c18tF8wd6F&export=download). Notes: Robust standard errors are between round parentheses. The raw coefficients emphasized using *, **, and *** are significant at 5%, 1% and 1‰. Green font suggests models not losing significance depending on the cross-validation criteria specified on the left (inside var(_cons[…])).

and

Table A13. Non-random cross-validations using MEPROBIT and MEOPROBIT regressions.

MEPROBIT/MEOPROBIT	Model1	Model2	Model3	Model4	Model5	Model6	Model7	Model8	Model9	Model10	Model11	Model12	Model13	Model14	Model15	Model16	Model17	Model18	Model19	Model20
Outcome	E260BBin	E260BBin	E260BBin	E260BBin	E260BBin	E260BBin	E260BBin	E260BBin	E260BBin	E260BBin	E260B	E260B	E260B	E260B	E260B	E260B	E260B	E260B	E260B	E260B
E253B (Information source: Social media)	−0.2417 ***	−0.2410 ***	−0.2430 ***	−0.2406 ***	−0.2411 ***	−0.2404 ***	−0.2395 ***	−0.2399 ***	−0.2512 ***	−0.2403 ***	0.2556 ***	0.2539 ***	0.2564 ***	0.2543 ***	0.2553 ***	0.2551 ***	0.2552 ***	0.2546 ***	0.2663 ***	0.2535 ***
	(−0.0167)	(−0.0038)	(−0.0058)	(−0.0057)	(−0.0319)	(−0.0095)	(−0.0091)	(−0.0107)	(−0.0166)	(−0.0132)	(−0.0118)	(−0.0039)	(−0.0077)	(−0.0072)	(−0.0179)	(−0.0054)	(−0.0075)	(−0.0081)	(−0.0161)	(−0.0096)
E254B (Information source: Talk with friends or colleagues)	−0.1556 ***	−0.1543 ***	−0.1550 ***	−0.1564 ***	−0.1560 ***	−0.1546 ***	−0.1553 ***	−0.1568 ***	−0.1487 ***	−0.1582 ***	0.1813 ***	0.1800 ***	0.1802 ***	0.1816 ***	0.1812 ***	0.1811 ***	0.1816 ***	0.1820 ***	0.1710 ***	0.1827 ***
	(−0.0172)	(−0.0047)	(−0.004)	(−0.0057)	(−0.0149)	(−0.0123)	(−0.0045)	(−0.009)	(−0.0096)	(−0.0093)	(−0.0112)	(−0.0039)	(−0.003)	(−0.0038)	(−0.0097)	(−0.007)	(−0.0032)	(−0.0083)	(−0.0093)	(−0.011)
E259B (Information source: Radio news)	−0.0893 ***	−0.0903 ***	−0.0896 ***	−0.0906 ***	−0.0889 ***	−0.0887 ***	−0.0871 ***	−0.0892 ***	−0.0980 ***	−0.0934 ***	0.0592 ***	0.0591 ***	0.0589 ***	0.0597 ***	0.0585 ***	0.0587 ***	0.0549 ***	0.0580 ***	0.0661 ***	0.0605 ***
	(−0.0068)	(−0.0046)	(−0.0067)	(−0.0053)	(−0.0015)	(−0.0067)	(−0.004)	(−0.007)	(−0.0074)	(−0.0039)	(−0.0063)	(−0.0034)	(−0.0051)	(−0.0043)	(−0.0025)	(−0.0056)	(−0.0042)	(−0.0076)	(−0.0071)	(−0.0057)
E261B (Information source: E-mail)	−0.3198 ***	−0.3229 ***	−0.3215 ***	−0.3224 ***	−0.3194 ***	−0.3199 ***	−0.3134 ***	−0.3206 ***	−0.3417 ***	−0.3221 ***	0.2186 ***	0.2201 ***	0.2196 ***	0.2195 ***	0.2182 ***	0.2187 ***	0.2137 ***	0.2179 ***	0.2352 ***	0.2189 ***
	(−0.0048)	(−0.007)	(−0.0027)	(−0.0192)	(−0.0157)	(−0.0195)	(−0.0161)	(−0.0095)	(−0.0239)	(−0.0404)	(−0.0002)	(−0.0034)	(−0.0047)	(−0.0046)	(−0.0116)	(−0.0083)	(−0.0063)	(−0.0081)	(−0.0133)	(−0.024)
X002 (Year of birth)	0.0121 ***	0.0156 ***	0.0126 ***	0.0111 ***	0.0122 ***	0.0105 ***	0.0117 ***	0.0121 ***	0.0135 ***	0.0120 ***	−0.0082 ***	−0.0143 ***	−0.0088 ***	−0.0076 ***	−0.0082 ***	−0.0070 ***	−0.0080 ***	−0.0081 ***	−0.0091 ***	−0.0082 ***
	(−0.0007)	(−0.0008)	(−0.0005)	(−0.0016)	(−0.0022)	(−0.0014)	(−0.0007)	(−0.0007)	(−0.0021)	(−0.0029)	(−0.0002)	(−0.0011)	(−0.0005)	(−0.0015)	(−0.0011)	(−0.0012)	(−0.0003)	(−0.001)	(−0.0016)	(−0.0023)
_cons	−20.5187 ***	−27.3435 ***	−21.6102 ***	−18.5761 ***	−20.6474 ***	−17.3468 ***	−19.8562 ***	−20.4965 ***	−23.1771 ***	−20.4118 ***
	(−1.4951)	(−1.5225)	(−0.9942)	(−3.3779)	(−4.5297)	(−2.8223)	(−1.4287)	(−1.3829)	(−4.1373)	(−5.8688)
var(_cons [X001]) (Gender)	0.0000										0.0002 ***
	(0.0000)										(0.0000)
var(_cons[X003]) (Age)		0.0120 ***										0.0153 ***
		(−0.0024)										(−0.0041)
var(_cons[X007]) (Marital status)			0.0024 *										0.0021 *
			(−0.001)										(−0.0009)
var(_cons[X013]) (Number of people in household)				0.0054										0.0118
				(−0.0052)										(−0.0124)
var(_cons[X025R]) (Education level)					0.0000										0.0000
					(0.0000)										(0.0000)
var(_cons[X028]) (Employment status)						0.0037 **										0.0021 *
						(−0.0012)										(−0.0008)
var(_cons[X045]) (Social class)							0.0029 **										0.0007 **
							(−0.0009)										(−0.0002)
var(_cons[X049]) (Settlement size)								0.0024 *										0.0024 **
								(−0.0011)										(−0.0009)
var(_cons[S003]) (Country code)									0.1660 ***										0.1538 ***
									(−0.0333)										(−0.0285)
var(_cons[S020]) (Year of survey)										0.0256										0.0251
										(−0.0152)										(−0.0144)
N	87,704	87,734	87,379	87,165	87,071	86,856	82,972	86,686	87,741	87,741	87,704	87,734	87,379	87,165	87,071	86,856	82,972	86,686	87,741	87,741
AIC	68,121.2058	67,982.195	67,824.8319	67,704.6795	67,710.0795	67,402.0599	64,729.7165	67,427.8605	63,195.3261	67,594.3494	177,791.077	177,645.439	177,000.599	176,624.35	176,505.075	176,086.464	169,233.635	175,649.294	170,377.29	177,054.715
BIC	68,130.5875	68,047.8695	67,881.1	67,770.3085	67,728.8285	67,467.664	647,76.3478	67,493.4509	63,261.0011	67,641.2602	177,809.84	177,739.26	177,056.867	176,718.105	176,523.824	176,152.068	169,270.94	175,724.254	170,471.11	177,120.39

Source: Own computation in Stata 17 (online script at https://drive.google.com/u/0/uc?id=1ZA4a1Tl14NedNrdxCgkCpeYlcixQbfyg&export=download). Notes: Robust standard errors are between round parentheses. The raw coefficients emphasized using *, * *, and *** are significant at 5%, 1% and 1‰. Green font suggests models not losing significance depending on the cross-validation criteria specified on the left (inside var(_cons[…])).

, Appendix C) using MELOGIT, MEPROBIT (the binary form of the outcome), MEOLOGIT, and MEOPROBIT (the original format of the target as a scale), those five variables (Figure 4) confirmed with no loss of significance. All of them had *** meaning significant at 1‰ on each subset corresponding to each value of those ten cross-validation criteria (variables) that were considered (X001-Sex, X003-Age, X007-Marital Status, X013-Number of people in household, X025R-Education level, X028-Employment status, X045-Social class, X049-Settlement size, S003-Numeric country code, S020-Survey Year). The last one (S020) failed to be considered in the core list of predictors (Figure 4) due to not being present at the intersection of features selected by all approaches used up to stage 4 (including it) in both versions of the WVS dataset considered in this paper (4.0 and 5.0).

The nomogram in Figure 4 (Stage 9) was constructed based on the binary logistic model presented in Table A6, model 2 (Appendix B). This nomogram represents five resilient influences. An example of an assessment leading to a probability greater than 0.8 (on the Total score X-axis) resides in Figure 4. The model stands on a substantial number of intersecting observations and corresponding support. It is about 87,741 cases, accounting for 49.2% of the total valid records for the target variable. The model exhibits a substantial R² value of 0.332, indicating a considerable volume of variance explained, and its classification accuracy is good, as reflected by the AUC-ROC (Area under the Receiver Operating Characteristic Curve as a metric for evaluating the performance of classification models) value of 0.8714.

Beyond these statistical validations, the predictive model illustrated in the nomogram (Figure 4) offers valuable real-world applications for policymakers, researchers, and businesses. It enables stakeholders to assess and enhance mobile phone usage as a source of information across diverse demographic groups. The nomogram clearly depicts the five key influences—birth year, frequency of electronic mail use, digital social network engagement, interpersonal conversations, and radio news consumption. It also includes a built-in simulation scenario based on a middle/average value for each variable (the text box in the center of Figure 4), showing how every single variable contributes when visually and intuitively computing a total score and a corresponding probability value.

To make this utility more concrete, we consider a practical application for UI designers. Using the nomogram, they can simulate how changes in user behavior or demographics directly influence the likelihood of a person using a mobile phone for information. For example, by following the scales in Figure 4, a designer can observe that an increase in the frequency of electronic mail use (a shift from value 3- monthly to 1- daily), translates into a significant increase in the final probability (0.6 instead of 0.35, when dragging the perpendicular on X for E261B, which contributes to a total score of 10.75 instead of 10.5 (Figure 4) and translates into a corresponding probability now greater than 0.9 instead of 0.8 on the Prob axis or ΔP = 10%). Similarly, a comparable increase in engagement with social media or radio news (same shift from 3 to 1) still has positive effects, but with different magnitudes. For E259B as Information from radio news, ΔP = 5% (e.g., a total score of 10.6 instead of 10.5—Figure 4, and a slight increase from 0.8 to 0.85 on the Prob axis for a lower code value, such as 1—Daily). Conversely, a less frequent engagement (a higher code value, such as 5—Never) with these sources decreases the probability. In the case of birth year, the nomogram shows that a shift from an older generation (e.g., 1960) to a younger one (e.g., 2004) directly leads to an increase in probability. By contrast, a decrease in age (e.g., targeting people born between 1945 and 1950) would generate a slight decrease in probability (from 0.8 to 0.75 for a corresponding total score of 10.4 instead of 10.5—Figure 4) or ΔP = −5%. The latter applies to the overall model based on all valid data without considering a specific subset corresponding to a particular location or respondent group (based on precise filters).

To transform this step-by-step simulation above with a consistent increase in overall probability (∑ΔP = 10% for a total score of 10.75) into an actionable policy recommendation plan, we have designed an explicit operationalization (Appendix D,

Table A19. Getting from a step-by-step nomogram-based simulation to policy simulations and action.

Intervention	ΔP (Delta Probability—Example)	Design/Policy Implication
Promote daily email use (E261B)	0.9 − 0.8 = 0.1	Integrate email digests to reduce mobile dependency
Increase radio news access (E259B)	0.85 − 0.8 = 0.05	Cache audio for offline/low-data users
Target older users born between 1945 and 1950 (X002)	0.75 − 0.8 = −0.05	Prioritize desktop/web interfaces
∑ΔP (Sum of all increases/decreases in ΔP)	0.1 or 10%

). When starting from the resulting peculiar models, we expect UI designers to simulate more specific user profiles and justify, for instance, the simplification of navigation and the minimization of data usage, knowing that these features would be most beneficial for users with lower-end devices or limited data plans. Moreover, by understanding that access to electricity and education (for the users) significantly influences their mobile information use, the designer could prioritize features that function effectively with intermittent power or a low-literacy interface.

Thus, the nomograms enable users to evaluate how variations in these factors contribute to the overall probability of utilizing mobile devices for information.

The aligned scales provide a straightforward method for quantifying scores (Figure 4), enabling UI designers, researchers, policymakers, and businesses to identify specific user behaviors and preferences.

Moreover, this simulation-based approach enables designers to move from abstract statistical findings to actionable decisions that directly enhance user experience and increase the potential for widespread adoption of an application.

This information can guide targeted interventions (such as improving digital literacy programs for older adults or optimizing user interface designs to accommodate different demographic needs). Ultimately, this approach fosters more inclusive access to information through mobile technology.

4.2. Results of Additional Tests and Validations

The raw results of using the Brain NN pack in Stata in different setups (Stage 10) are also available (online at http://tinyurl.com/5fpsrv75 and synthesized in

Table A14. Synthetic results of using the Brain NN pack in Stata on the WVS data (version 4.0).

Number of Hidden Layers	Number of Iterations	Training Factor—Eta	R–square	Accuracy (%)
10	1000	0.05	0.7296	83.3578
10	1000	0.15	0.7284	83.3350
10	1000	0.25	0.7170	82.3617
10	2000	0.05	0.728	83.073
10	2000	0.15	0.7244	82.9658
10	2000	0.25	0.7143	82.5441
10	3000	0.05	0.7306	83.4114
10	3000	0.15	0.7273	83.3681
10	3000	0.25	0.7214	83.0957
15	1000	0.05	0.7190	82.8085
15	1000	0.15	0.7213	82.9749
15	1000	0.25	0.7219	83.2222
15	2000	0.05	0.7283	83.2507
15	2000	0.15	0.73	83.3852
15	2000	0.25	0.7285	83.3841
15	3000	0.05	0.7305	83.4285
15	3000	0.15	0.7292	83.3738
15	3000	0.25	0.72	83.0832
20	1000	0.05	0.7292	83.3761
20	1000	0.15	0.7213	83.0957
20	1000	0.25	0.7207	83.155
20	2000	0.05	0.7271	83.1379
20	2000	0.15	0.7234	83.2416
20	2000	0.25	0.7163	82.902
20	3000	0.05	0.7235	83.2302
20	3000	0.15	0.7229	83.2131
20	3000	0.25	0.7232	83.3373

, Appendix C). The same five determinants appear in all NN models, confirming good classification accuracy (The accuracy was computed as (TP + TN)/(TP + TN + FN + FP), where TP stands for True Positives, TN for True Negatives, FP for False Positives, and FN for False Negatives, according to the details provided by the help brain command.) (values greater than 80%, as seen in the right column of Table A14, Appendix C) in all tested configurations. Regardless of these setups, depending on the number of hidden layers, iterations, and training factor, all models based on backpropagation artificial neural networks using the Brain NN pack in Stata brought clear confirmations. These confirmations were related to the classification accuracy. Although less interpretable in all configurations, they indicated (like the LOGIT-based ones—https://tinyurl.com/yc5e86pv) good accuracy (values of more than 80). As seen in Table A14, Appendix C, they range approximately between 82 and 84. Similar results (https://tinyurl.com/2fekv9k3) confirming the same accuracy performance emerged based on tests performed with the Brain NN pack on the latest version (5.0) of the WVS data.

When positioning the five previously confirmed determinants (Figure 4) in the list of results provided in the initial selection stage with Ada Boosting, we will notice that they are at the top. The latter is where the frequencies of use (based on the number of splits when constructing classification trees) are higher (Figure 2). It also suggests that we are dealing with a classification technique that often leads to robust results. Moreover, the SHAP (SHAP values available at https://tinyurl.com/murjffzf) and PDP representations for the HGB model (Appendix D, Figure A6a,b and Figure A7) confirm the dominance of email (E261B) and social media use (E253B), aligning with the results from logistic regressions (Stage 6), and enhancing interpretability beyond nomograms (Stage 9 and Figure 4).

When testing the potential inclusion of X025R (Highest educational level attained: Lower, Middle, and Upper) and X001 (Gender: Male and Female) suggested by previous research, the results indicated a failure. And that in terms of passing the selection tests based on ADA BOOST, CVLASSO, RLASSO, and, additionally, BMA for both forms of the outcome (binary and original scale). Moreover, they lost their significance and added errors in the models, regardless of the type of regression. Two links for the script dedicated to testing these two variables (version 4.0 of WVS data) and captures of the console output as .pdf (indicating commands, their results, and comments) are also included (https://tinyurl.com/2avjc4n9 and https://tinyurl.com/3b9y2d2t—the first part of these online resources). In addition, these two (X025R and X001) also failed the tests performed on the second dataset (IVS—Stage 11, used for validating the findings obtained using WVS data, version 4.0). More details are available as an additional testing script and a corresponding console capture (the first part of the online resources at https://tinyurl.com/hru48zzs and https://tinyurl.com/2nck8zdr).

We agree that a purely algorithmic explanation can be a drawback in a study on social issues. The exclusion of certain variables, such as gender and educational level, which are often considered significant in social research, deserves a more nuanced discussion. We believe that the reason these variables are not selected is not necessarily due to a lack of significance, but rather because their influence on mobile information acquisition is already available in the five robust predictors in our model. For instance, variables related to digital social network engagement and e-mail use may act as proxies (for the educational level and social and economic status). Individuals with higher levels of education or well-defined professional roles are more likely to use these platforms, thereby making these predictors more direct and specific indicators of information-seeking behavior than a broad category like education. Similarly, the birth year variable may serve as a proxy for generational differences in digital literacy and technology adoption, which in turn implicitly account for differences (otherwise easy to associate with gender). By using the most direct and specific behavioral proxies, our model avoids a black box. It also brings a more precise and behavioral explanation for mobile phone use as an information source.

The results of the Shapiro–Wilk test (more details at https://tinyurl.com/556dk485) and those from the histograms (https://tinyurl.com/mr3tryyj) indicate the distribution of variables leading to the rejection of the hypothesis. The hypothesis was that the variables corresponding to the five most robust determinants follow a normal distribution.

Still, the results of other modeling approaches, such as SCOBIT [159] regressions (an alternative arrangement for the disturbances from normality/logistic distribution), indicated robust predictors (Model 3, Table A6, Appendix B and Model 3,

Table A17. Different regression models considering the four most relevant remaining influences common to WVS and IVS data.

Model No.	(1)	(2)	(3)	(4)	(5)	(6)	(7)
Target Variable	E260BBin	E260BBin	E260BBin	E260BBin	E260B	E260B	E260B
Regression Type	OLS	LOGIT	SCOBIT	PROBIT	OLS	OLOGIT	OPROBIT
E254B (Information source: Talk with friends or colleagues)	−0.0561 ***	−0.2827 ***	−0.5311 ***	−0.1692 ***	0.2671 ***	0.3530 ***	0.2078 ***
	(0.0008)	(0.0041)	(0.0163)	(0.0025)	(0.0029)	(0.0039)	(0.0023)
E259B (Information source: Radio news)	−0.0221 ***	−0.1246 ***	−0.2335 ***	−0.0749 ***	0.0710 ***	0.0932 ***	0.0567 ***
	(0.0006)	(0.0037)	(0.0095)	(0.0022)	(0.0023)	(0.0032)	(0.0019)
E261B (Information source: E-mail)	−0.0884 ***	−0.6411 ***	−1.7086 ***	−0.3490 ***	0.3662 ***	0.4825 ***	0.2911 ***
	(0.0006)	(0.0058)	(0.0719)	(0.0029)	(0.0023)	(0.0034)	(0.0020)
X002 (Year of birth)	0.0053 ***	0.0301 ***	0.0544 ***	0.0182 ***	−0.0202 ***	−0.0260 ***	−0.0158 ***
	(0.0001)	(0.0004)	(0.0015)	(0.0002)	(0.0002)	(0.0003)	(0.0002)
_cons	−9.3355 ***	−54.9435 ***	−94.2903 ***	−33.3224 ***	40.4795 ***
	(0.1208)	(0.7245)	(2.4669)	(0.4211)	(0.4431)
lnalpha			−1.1989 ***
			(0.0407)
cut1						−48.7565 ***	−29.5915 ***
						(0.5999)	(0.3570)
cut2						−48.2092 ***	−29.2648 ***
						(0.5998)	(0.3569)
cut3						−47.9520 ***	−29.1118 ***
						(0.5997)	(0.3569)
cut4						−47.5784 ***	−28.8906 ***
						(0.5995)	(0.3568)
N	171,901	171,901	171,901	171,901	171,901	171,901	171,901
Chi-squared		29,375.6194		34,000.0736		46,704.8809	50,964.0623
p	0.0000	0.0000		0.0000	0.0000	0.0000	0.0000
R-squared	0.2490	0.2319		0.2298	0.2919	0.1335	0.1339
AIC	175,061.6152	164,571.0977	163,971.2185	165,014.1210	628,085.6018	377,315.8048	377,144.2028
BIC	175,111.8885	164,621.3710	164,031.5465	165,064.3944	628,135.8752	377,396.2421	377,224.6402
RMSE	0.4026				1.5037
AUC-ROC		0.8160		0.8151

Source: Own calculations in Stata 17 (the source script for generating this table is available online at: https://drive.google.com/u/0/uc?id=1kd8Tns_SUdZ-oRwZkt1nWGtg7_ZAFzuw&export=download and https://drive.google.com/u/0/uc?id=18hco57Q9225fTon0oextCFy6jaacYVlM&export=download). Notes: The raw coefficients emphasized using *** is significant at 1‰.

, Appendix D).

Regarding heteroskedasticity, the results of most regressions performed in this study include robust standard errors (e.g., Table A6, Appendix B, together with most tables in Appendix C). Therefore, we expect the impact of heteroskedasticity on inference to be minimal by correcting the error bias that may arise from the non-constant variance in residuals.

The results of testing the linearity using scatter plot graphs between the dependent variable and each independent variable (test script at https://tinyurl.com/3hrphw7j and console capture at https://tinyurl.com/msspssk5) confirmed the corresponding hypothesis. It involved an additional conversion script (https://tinyurl.com/48frmven) for reversing the scales and setting them to start from 0 while replacing text labels with numbers to avoid errors (Never-0 up to Daily-4).

We decide not to include interaction terms between independent variables for a simple reason: the substantial complexity and volume of interactions (for reporting). Specifically, four categorical predictors (E253B, E254B, E259B, and E261B) each have five distinct values, leading to a combinatorial explosion when considering interactions. Moreover, the year of birth variable (X002) introduces an additional 85 distinct values (As a result of using the tabulate command in Stata, with the specification of the Non NULL condition for both the target variable and those five most robust predictors, namely tabulate X002 if E260BBin!=. & E253B!=. & E254B!=. & E259B!=. & E261B!=. & X002!=.), further complicating the analysis. Given these constraints, the number of interaction terms would be impractically large, making the results difficult to interpret and communicate effectively. Additionally, ensuring statistical robustness across such a vast number of interaction terms would require significantly larger sample sizes, which may not be feasible. More, in terms of statistical redundancy and instability, LASSO (Stage 3) and BMA (Stage 4) systematically penalize redundant terms (any interaction would fail selection due to high collinearity). Even more, including interactions would generate a policy utility compromise by visually overcomplicating the nomogram (our primary policy simulation tool) and requiring multidimensional grids or conditional scoring, reducing accessibility for non-technical audiences, and violating degree-of-freedom constraints in subgroup estimation. Thus, to maintain clarity and focus on the most influential predictors, interaction effects are not present in the reported analysis.

A sensitivity analysis comparing standard omission (N = 87,741) and global median imputation (N = 439,978) was also included (https://tinyurl.com/49a8et8e). All five predictors were retained. While directions remain consistent, the median imputation significantly alters effect sizes and reduces model fit and accuracy (Pseudo R²: from 0.3320 to 0.2101; AUC-ROC: from 0.8714 to 0.7592). This is expected because imputing DK/NA cases with modal responses (median = 2 or Weekly) introduces noise and compresses variance. Therefore, standard omission is preferred for coefficient accuracy, while median imputation increases sample size but underestimates and/or overestimates associations. Consequently, the results are robust in predictor selection, while sensitive in magnitude.

The superiority of this combined, triangulated feature selection framework (Figure 1 and Figure A1, Appendix A) over a single-method approach such as HGB was quantitatively demonstrated (https://tinyurl.com/55s9268j). Although the HGB model using its top 25 features reached a higher predictive score (AUC-ROC: 0.8987) than the five-feature configuration (AUC-ROC: 0.8490) derived from the fusion approach (Figure 4), a critical robustness check showed that this apparent gain was algorithm-dependent. When the same 25 features (18 after removing missing or collinear variables) were applied in a conventional LOGIT model, performance dropped sharply (AUC-ROC = 0.8094). In contrast, the five predictors retained through the triangulated framework achieved a superior and stable AUC-ROC of 0.8714 (Model 2, Table A6, Appendix B; Figure 4). Moreover, two of the 18 predictors (S022 and S016) completely lost statistical significance in the LOGIT model, confirming instability under single-method selection. These results clearly quantify the performance and stability gap in favor of the proposed multi-stage framework, which operates as a robustness filter ensuring that the final predictors remain generalizable and theoretically consistent across both data mining and econometric paradigms.

4.3. Demographic Analysis Outcomes

The results of the additional demographic analysis are available online in the same main container (https://tinyurl.com/ms3n3s9v), more precisely in the “demographic analysis” subfolder.

The visualizations (in a subfolder named “4visualize” and Figure A3, Appendix C) intuitively indicate how the target variable behaves (at the mean). These visualizations consider all seven variables considered in this demographic analysis. The latter means higher values for male-X001 and for not-married respondents-X007. The maximum occurred for four people in a household (X013). There is a positive linear relationship between the target variable and the education level (X025R, measured as Lower, Middle and Upper). The same type of relation as the previous one resulted when considering employment status (X028), re-coded with larger values for full-time jobs. The same applied to the social class re-coded with larger values for the upper class. A similar dependence occurred on settlement size when considering the highest scale value (communities with more than 500.000 people). The age variable (X003) is not present in this demographic analysis. The latter occurs because it derives from the year of birth (X002), which already exists in the core model.

The identified predictors significantly enhance our understanding of mobile phone usage as a source of information. They illustrate how various demographic and behavioral factors converge to influence information-seeking behavior.

Age emerges as a critical determinant. Younger individuals have a higher engagement with mobile technology due to greater digital literacy. Older respondents face challenges linked to less exposure and adaptability to new tools.

Furthermore, gender disparities reveal interesting trends, as male respondents tend to exhibit higher levels of mobile phone use for information access. The latter is due to occupational differences, technology-related confidence, or cultural norms influencing digital adoption. Similarly, individuals with higher levels of education are more likely to rely on mobile phones as a source of information, potentially due to better digital skills and greater exposure to online resources.

Additionally, socioeconomic status plays a crucial role. Those from higher social classes and individuals employed in full-time jobs demonstrate a stronger inclination towards mobile-based information-seeking. The latter can relate to both affordability and the necessity of real-time information access for professional purposes. Similarly, urbanization level is a key factor, as individuals residing in larger settlements, particularly those exceeding 500,000 inhabitants, exhibit a higher reliance on mobile phones for obtaining information. This trend may stem from improved internet infrastructure, increased digital exposure, and a higher concentration of information-based occupations in urban settings.

The frequent use of electronic mail, reliance on social media, interpersonal discussions, and consumption of traditional radio news further underscore the diverse pathways through which people access information. These various channels highlight the different ways people use mobile devices to stay informed. Notably, the interplay between these predictors highlights the need for tailored strategies in interface design. These strategies are crucial for bridging the digital divide, especially for older adults.

5. Discussion

5.1. Main Findings

In reviewing the primary predictors influencing the use of mobile phones as a predominant information source, it becomes evident that age (validation of H5), use of electronic mail for information retrieval (validation of H2), reliance on social media platforms (validation of both H1 and H3), interpersonal discussions with friends or colleagues (validation of both H1 and H3), and traditional radio news consumption (partial validation of H4) play pivotal roles. The latter stands on multiple studies, including [4,38,39,40,45,46,47,48,50,72,79]. Most references to them in the literature correspond to influences or correlations with accessing information via mobile devices, while causal aspects are usually not present. Bounded by their impact (predictors or determinants) and prevalence (magnitude of effects), they can nevertheless highlight the evolving landscape of information consumption through mobile technology.

The highest potential probability for the optimal combination of variable values, represented on the extreme right side of the graph (Figure 4), is remarkably high. It exceeds 0.99 (99%) for a total score close to 11.75. This exceptional scenario corresponds to respondents born in 2004 (X002 = 2004) with an individual nomogram score approaching 10. Additionally, it involves daily use of the electronic mail service (E261B = 1), for a specific score of approximately 0.75, daily reception of radio news (E259B = 1) with a corresponding score close to 0.2, daily conversations with friends or colleagues to stay updated (E254B = 1) yielding a score above 0.3, and daily utilization of digital SN as an information source (E253B = 1) resulting in a score of 0.5. The graphical representation in Figure 4 effectively illustrates the magnitude of marginal effects for the respective variables, offering improved comparability compared to raw coefficients in Model 2 (Table A6, Appendix B).

Furthermore, this nomogram supports the comprehension of the cumulative effect size. The latter applies by visualizing the scaled amplitude. Notably, when considering the descending order of importance, E261B (use of the electronic mail) and E253B (social media—Facebook, Twitter, etc.) emerge as the top two influential variables, followed by X002 (birth year), E254B (Talk with friends or colleagues), and E259B (Radio News).

This order aligns with the findings from the ADA BOOST technique employed in Rattle (see Figure 2), which ranks the variables based on the number of splits when constructing classification trees. Moreover, these results are consistent with existing literature in the field of e-commerce [160,161,162,163,164,165,166].

Regarding the negative link between the target variable and the age (or the positive one between the target and the birth year as proven for both datasets, WVS 4.0—Figure 4 for Model 2 in Table A6, Appendix D, and IVS—Figure A5b for Model 2 in Table A17, Appendix D), this is a common finding that raises a lot of questions about the digital divide and possible strategies for inclusive interface design.

In terms of the digital divide, the afferent type of literacy tends to decrease with age. The latter is due to factors like less exposure to technology, lower adaptability to new tools, and fewer opportunities for training [167,168,169,170,171,172]. Therefore, the fact that this finding (associated with the age of the respondent) has the highest magnitude, as shown in this research and in the nomograms presented, is not accidental. In terms of strategies for inclusive interface design, some authors [73] emphasized that mobile technology design principles and solutions should fit the specific physical and cognitive profile (that of the elderly users). This adoption should consider existing sensor-related, motor, and cognition impairments. They mention user-centered design and also the presence of large screens and buttons (instead of touch screens), significant space between buttons, big fonts, considerable color contrast, speech recognition, voice commands, and voice response. Additionally, they emphasize minimum display items, a simple and intuitive navigation process, navigation assistance, a permanent indication of the current position within the information space, and tasks together with their status while showing the next one [173,174,175,176].

5.2. Support and Additional Checks and Validations

Upon examining the chronological coverage of the five variables and their respective influences in the WVS (Table A7, Appendix B), it becomes apparent that only the last two waves include the relevant questions and corresponding answers. These waves are Wave 6 (2010–2014) and Wave 7 (2017–2022). This outcome is intuitive, considering the significant advancements in smartphone and digital social network applications after 2010 [177,178]. Nevertheless, it is worth mentioning that one exception exists: the variable E253B (Information source: social media—Facebook, Twitter, etc.)—exclusively included in the last wave (Wave 7).

Conducting a similar assessment using country codes (tabstat E260B E260BBin E253B E254B E259B E261B E262B X002, by (S003) statistics (count)—Table A8, Appendix B), it resulted that out of 108 countries, only 83 countries supported most of the variables. The same exception occurred for E253B. The remaining 65 out of those 83 countries included the variable E253B. When using the second dataset (IVS) for validating those five predictors corresponding to the main findings, E253B (Information source—social media) was the only one that did not exist here (

Table A16. Descriptive statistics for the target variable and the four most relevant remaining influences common to WVS and IVS data used in this study after removing their DK/NA values.

Variable	N (Obs.)	Mean	St. Dev.	Min	Median	Max
E260B	171,901	2.64	1.79	1	2	5
E260BBin	171,901	0.68	0.46	0	1	1
E254B	171,901	2.18	1.42	1	2	5
E259B	171,901	2.83	1.7	1	2	5
E261B	171,901	3.58	1.67	1	5	5
X002	171,901	1972.73	16.81	1913	1975	2004

Source: Own calculations using the UNIVAR command (version 1.1.2, 1 November 1997) in Stata 17 followed by the entire list of the variables above and a non-NULL condition for all (univar E260B E260BBin E254B E259B E261B X002 if E260B!=. & E260BBin!=. & E254B!=. & E259B!=. & E261B!=. & X002!=. Note: E253B (Information source: Social media) is not existing at the intersection of WVS and IVS datasets.

, Appendix D vs. Table A5, Appendix B and Table A17, Appendix D vs. Table A6, Appendix B). Therefore, it was simply excluded from validation using IVS. The other four were fully confirmed. We included additional testing and exploration script sequences and some console captures (all available online at https://tinyurl.com/yck9h3yr and https://tinyurl.com/54yj5bne for Stata and https://tinyurl.com/rnkwev5z and https://tinyurl.com/twdpu9fp for both Rattle—Figure A4, Appendix D, and Stata—Figure A5a,b, Appendix D).

Using this second dataset means much more in terms of generalizability. The latter also better supports the findings that hold up across different datasets or populations, and the fact that they are not specific to the first dataset used (WVS).

Although present in the dataset and initially identified through a variable search using the keyword Internet, four other related influences did not exhibit sufficient significance and robustness. Therefore, these influences are not available in the final models depicted in Figure 4. These influences include E282—Political Actions using the Internet: Exploring information related to politics, E283—Political Actions using the Internet: Signing an electronic petition, E284—Political Actions using the Internet: Motivating others to engage in political activities (of any kind). They also include E285—Political Actions using the Internet: Coordinating political events, activities, protests, and similar initiatives. Similarly, the same applies to other variables manually searched using the keyword Information such as E248B—Information source: Daily newspaper, E249—Information source: News broadcasts on radio or TV, E250—Information source: Printed magazines, E251—Information source: In-depth reports on radio or TV, E252—Information source: Books, E258B—Information source: TV news, H010—Government has the right: Monitor all electronic mails and any other information exchange, and H011—Government has the right: Collect information about anyone living in the country. Additionally, variables identified using keywords like Government has the right, the right, or commerce, such as H006_06—Worries: Government wire-tapping or reading my mail or electronic one, H009—Government has the right: Conduct video surveillance of individuals in public spaces, E069_04—Confidence: The Press, and E069_30—Confidence: The Free Commerce Treaty, did not meet the criteria for inclusion in the final models [179,180,181,182,183].

Moreover, only a sublist (of the aforementioned list of variables previously checked) is available in the dataset used for validations and robustness checks (IVS). This subset consisted of E248B, E250, E258B, H010, H011, H009, E069_04, and E069_30. However, E250 and E069_30 did not meet the support criteria for further tests, as they had less than half of the total number of observations available for the target variable. The rest of them, except for E248B and E258B, were not selected when using BMA, CVLASSO, and RLASSO. When included in regressions together with the most robust predictors already identified in WVS 4.0 and available also here (IVS), namely four (E254B, E259B, E261B, and X002) of those five above (beginning of discussion), E248B proved to be collinear with E259B and with E258B. In addition, its lower magnitude (when compared to that of E259B) as seen in another prediction nomogram for selection purposes (additional testing script and console capture available online at https://tinyurl.com/bdh5f5tr and https://tinyurl.com/4h88u7kj), made it unsuitable for inclusion in the final model. An argument for removing E258B is that, in two combinations (only with E259B and E261B, as already confirmed), it loses significance in some regressions. An additional argument for dropping both E248B and E258B is that they did not result in the first place (selections based on the WVS data).

The five selected influences (as illustrated in Figure 4 and Table A6, Appendix B) offer insights into the simplicity requirements for mobile applications. The latter applies when such applications serve as a source of information. These influences, among numerous others present in the extensive data collection, also provide significant ideas about the desired simplicity of the user interface. In essence, this implies that the interface of any mobile application catering to e-commerce should be as straightforward as checking an e-mail inbox or liking a Facebook page or post (the second and third highest impact of e-mail and social media as a source of information, as seen in the nomogram in Figure 4-Model 2 in Table A6, Appendix B). Engaging in direct conversations or video conferencing with loved ones should also be similarly simple, further emphasizing the need for intuitive and easy-to-use interfaces in mobile applications. It should also allow users to effortlessly switch between radio buttons to change preset frequencies or present a user interface designed like a game tailored for younger individuals or children [184,185,186,187,188,189,190]. These examples show the utility and success of employing simplified information, such as user-generated ratings on product quality for new e-commerce transactions [191,192,193].

5.3. Delimitation from Previous Findings

Of course, there are also other studies in which using mobile phones appears to be a topic of interest, but none of these confirm these five predictors (at once) when considering mobile phones as an information source. For instance, when using mobile phones, but in association with cultural values, lower trust, and television versus the internet for information seeking. Additionally, disparities between regions, public support for LGBTQ, moral progress, intimate partner violence against women, female egalitarian values in their role in protests, or financial emancipation are also relevant factors. These topics appear in various studies [23,31,32,33,34,35,36,37,80].

To delineate these insights from the existing research, it is worth mentioning that most of the existing quantitative studies use data limited in terms of time (a questionnaire at a precise moment, a survey applied within a given period/wave). These studies are also limited in space, covering only one region, country, or continent, or at most comparisons between a limited number of such locations.

In addition, most studies do not state the purpose of identifying the core intersecting predictors of using mobile phones as a source of information. The latter is especially true when starting from a dataset so varied in time, space, and socio-demographic characteristics (of respondents). To such arguments, it added that most studies do not simultaneously use cross-validation according to random (10-fold) and non-random criteria. These include socio-economic features, different versions of the same dataset, or even different datasets with similar variables, etc.

Moreover, by offering NOMOGRAMS for probability prediction, impact analysis, and visual interpretation, this study provides practical tools for decision-makers, researchers, and designers. Such tools help professionals to understand and improve mobile applications.

5.4. Potential Implications

This study offers a range of insights with significant implications for various domains, particularly as mobile phones continue to shape information-seeking behaviors worldwide. The validation of predictive models using advanced techniques (such as ADA BOOST, LASSO, BMA, and neural networks) underscores the reliability of these findings. For example, public health organizations might use these methods to identify how different demographic groups access critical information via mobile platforms, ensuring their messages reach the right audiences effectively. As an additional case, a health agency could rely on such models to determine whether SMS alerts or social media campaigns are more effective for different age groups when disseminating emergency health updates.

The longitudinal nature of the data, spanning decades of the WVS, provides a rare opportunity. It allows us to observe the evolving role of mobile phones in society. These insights reveal generational shifts in how technology is used (such as younger users increasingly relying on mobile devices for education and news consumption). Such trends highlight the need for educational tools that cater to the preferences and habits of this demographic.

The cross-cultural scope of the study further uncovers unique regional and cultural drivers behind mobile usage. For instance, in collectivist societies, mobile applications that prioritize group interactions—such as family chat features or community-focused content—may see higher adoption rates.

These findings are equally valuable for businesses seeking to optimize mobile platforms. By integrating features such as personalized recommendations and seamless navigation, companies can enhance user engagement and drive growth, particularly in e-commerce and social media sectors.

Application developers get in touch with a roadmap for user-centric design. The study emphasizes intuitive interfaces and effective and efficient communication tools as crucial elements. For older users, this might mean designing apps with larger text, voice command capabilities, or simplified navigation, ensuring accessibility across diverse user groups. Similarly, marketers can leverage these insights to create campaigns that resonate with their audiences. Highlighting the speed, accessibility, and interactivity of mobile platforms, for example, can effectively target tech-savvy younger consumers through interactive content or content presented in the form of a game (gamified).

The findings also underline the urgent need for digital literacy initiatives, particularly for older adults (the heaviest impact of year of birth and, consequently, age, as observed in the nomograms in Figure 4-Model 2 in Table A6, Appendix B, and Figure A5b-Model 2 in Table A17, Appendix D) and underserved communities. Empowering individuals through community workshops on mobile banking or health apps can bridge the digital divide and foster greater inclusivity. Local governments or non-profit organizations could organize hands-on workshops where seniors practice using mobile banking apps or online health services with guided support. For policymakers, the results suggest actionable strategies to tackle barriers to mobile phone access, such as improving digital infrastructure in rural areas or subsidizing connectivity for low-income families. These measures can ensure equitable access to information and create more informed and connected societies.

The role of mobile phones in encouraging communication is especially critical for remote or isolated communities. Digital connectivity serves as a lifeline in these areas. For example, localized messaging apps can strengthen social ties and support during emergencies. Similarly, the integration of mobile technology into education offers exciting possibilities. When using mobile tools to enhance teaching methods, educators can create dynamic and interactive learning environments that prepare students for a future more based on technology. For example, schools could integrate mobile-friendly learning apps that offer quizzes or interactive lessons, helping students engage with course materials beyond traditional classroom settings.

The demographic patterns observed in the study highlight areas for targeted interventions. Efforts to improve education and employment opportunities, for instance, can amplify the benefits of mobile information-seeking behavior among less-advantaged groups. Urban planners might also focus on tailoring digital infrastructure investments for smaller communities, reducing disparities between urban and rural areas. Furthermore, the data inspires future research into why certain demographic groups, such as unmarried individuals or those in large households, exhibit higher mobile usage. Longitudinal and cross-cultural studies could shed light on how these trends vary or evolve in time.

Moreover, the study mentions the ethical implications of increasing reliance on mobile phones for information. Concerns around data privacy, technological exclusion, and digital inequities must be a topic of interest to ensure fair access and user protection. Transparent data policies, inclusive application designs, and affordable mobile technologies are just a few ways developers and policymakers can navigate these challenges. These actions help foster trust and equity in an increasingly digital world.

In terms of implementing recommendations in resource-limited settings, there are some other relevant emerging ideas. For instance, while this study highlights the transformative potential of mobile phones in information-seeking behavior (confirmed by the Logistic regression-based risk prediction nomogram in Figure 4 and the regression models in Table A6, Appendix B), implementing these insights in resource-limited settings presents distinct challenges. Limited digital infrastructure, affordability barriers, and low digital literacy often hinder widespread adoption of mobile devices. To overcome these obstacles, governments and organizations can prioritize low-cost mobile solutions, such as subsidized internet access, partnerships with telecom providers for reduced data plans, and the promotion of feature phones with essential internet capabilities.

Additionally, leveraging offline-accessible mobile applications—such as pre-downloaded educational content or SMS-based health advisories—can ensure critical information reaches users without requiring constant internet access.

Moreover, digital literacy programs must tailor local constraints by focusing on community-driven initiatives. In areas where formal education is limited, peer-led training workshops or collaborations with local leaders can enhance mobile proficiency. For example, integrating mobile literacy components into existing community gatherings or health campaigns can maximize outreach. Similarly, simplified user interfaces with voice command options and multilingual support can enhance accessibility for users with limited literacy. By adopting context-sensitive strategies, the benefits of mobile technology can extend to even the most underserved populations, fostering greater inclusivity in digital information access.

We also provided a cost–benefit analysis for implementing the recommendations in resource-limited settings (

Table A18. Potential trade-offs of implementing recommendations in resource-limited settings as a short cost–benefit analysis.

Recommendation	Potential Costs	Expected Benefits
Subsidized internet and data plans	Requires government funding or telecom partnerships; potential revenue loss for providers	Increases digital inclusion, enabling access to vital services like education and healthcare
Offline-accessible mobile apps	Development and maintenance costs; limited content updates	Ensures access to critical information without constant internet connectivity
Peer-led digital literacy programs	Training costs; requires community engagement efforts	Empowers users with mobile skills, increasing adoption and effective usage
Simplified user interfaces	Development costs for app redesign; potential limitations in functionality	Enhances accessibility for users with low literacy, improving engagement
Infrastructure investment in rural areas	High initial costs for network expansion and maintenance	Reduces urban-rural digital divide, enabling equitable access to mobile technology
Public–private partnerships	Requires negotiation and long-term collaboration	Shares costs between stakeholders, ensuring sustainable solutions for mobile accessibility

, Appendix D section).

If accounting for implications in developing countries, the discussion partially overlaps with the one focused on resource-limited settings. Both contexts share specific challenges (such as limited infrastructure, low digital literacy, and affordability barriers). In terms of differences, while resource-limited settings may focus more on specific local challenges, the discussion on developing countries often relies on national strategies and a broader understanding (of the long-term economic, political, and cultural dynamics affecting access to technology). In essence, the general recommendations for resource-limited settings apply within developing countries. Still, resource-limited settings may also include specific localized interventions within wealthier nations.

As mobile technology continues to evolve, one potential long-term implication of these findings is the gradual transition from traditional mobile devices to emerging wearable technologies, such as smart glasses, or their simultaneous or interdependent use [194]. The growing reliance on mobile phones for information-seeking suggests a demand for even more seamless, hands-free, and immersive digital experiences. Smart glasses, equipped with Augmented Reality (AR) capabilities and AI-driven assistance, have the potential to revolutionize information access by integrating real-time data overlays into the personal field of vision [195]. The latter could enhance efficiency in various domains (from education and healthcare to navigation and social interactions). Unlike smartphones, which require manual operation and screen interaction, smart glasses offer a more natural way to consume and interact with digital content. For example, instead of unlocking a phone and searching for information, users could receive instant visual updates, voice-activated responses, or even AI-generated summaries displayed in their line of sight. Moreover, the shift towards smart glasses could bridge digital literacy gaps by providing intuitive, gesture-based interfaces and personalized learning experiences. However, challenges such as miniaturization (including battery life, processing power, display technology, connectivity and sensors, and user comfort and design), privacy concerns, affordability, and social acceptance must be of concern before widespread adoption can occur. If overcoming these barriers, smart glasses could eventually replace mobile phones as the dominant tool for information-seeking, signaling a transformative shift in human-technology interaction.

Nevertheless, the core predictors identified in this study are not tethered to any specific device but represent enduring behavioral functions that are likely to persist—or adapt—across technological platforms. Specifically, radio serves as a proxy for passive, ambient exposure to public narratives, a role that may evolve into AI-curated audio streams, podcasts, or AR soundscapes (with strongest effects observed among older adults and rural populations). Email and social network use reflect digital habit formation and routine engagement, functions expected to transition into AI-mediated equivalents such as proactive notifications or personalized generative feeds. Age captures generational imprinting, which will continue to shift as younger cohorts—raised with AR, voice, and wearable interfaces—enter the information ecosystem.

The central finding (that cross-channel media routines and ambient exposure drive reliance on mobile or future wearable information sources) is technology-agnostic and empirically validated across last two waves (internet and smartphone era vs. post-smartphone dominance). Thus, while smart glasses or AI agents may become the dominant interface, the underlying behavioral mechanisms are expected to remain operative. Future iterations of global surveys (WVS, IVS, or successors) should incorporate items on AI assistant dependency, wearable information access, and AR integration to assess whether these new modalities reinforce, reshape, or supplant the current predictor structure.

5.5. Limitations

This study has inherent data-related limitations due to its reliance on self-reported data from the WVS and IVS datasets. The use of this type of data may introduce several biases that warrant critical discussion. Recall bias may occur as participants report on past behaviors, and their memory may not be entirely accurate. Additionally, desirability bias can influence responses, where individuals may consciously or unconsciously provide answers they believe are socially acceptable rather than a true reflection of their actual behaviors. Furthermore, the use of a single survey instrument across diverse cultures can lead to inconsistent definitions or interpretations of survey items. The latter can potentially affect the comparability of the data. The focus on mobile phone usage frequency also simplifies user engagement while failing to differentiate between light and heavy users, which is another limitation. However, no globally comparable objective behavioral data exist for the period 1981–2022. In addition, this multi-stage robustness filter (many independent statistical/ML techniques and three datasets) retains only predictors that replicate across methods, waves, and continents greatly reducing the chance that findings are artifacts of reporting bias. Moreover, reverse-causality checks and non-random cross-validations by country/year further confirm that the identified associations are structurally stable, not transient desirability effects.

We also acknowledge that the datasets stand on secondary data, which is a known limitation. However, our use of the World Values Survey and Integrated Values Survey was a deliberate choice to ensure a broad, cross-national scope, a substantial sample size and global comparability. And that would not be feasible with primary data collection. The design of our rigorous multi-stage methodology allows the mitigation of the inherent drawbacks of self-reported, secondary data, including various biases and inconsistencies across cultures, thereby ensuring the robustness of our findings. Moreover, we additionally mitigate this by: leveraging multi-wave pooling (1981–2022) with wave-fixed effects to approximate longitudinal trends (1); applying subgroup-stratified models to reveal differential predictor strength across demographics (2); replicating core findings in the independent IVS dataset (3). Future primary longitudinal studies in single countries could complement these insights, though at the cost of cross-national comparability.

Although Stage 7 rigorously tests and rules out reverse causality (Table A10 and Table A11), the study remains correlational. We cannot definitively establish why the five predictors influence mobile information reliance. Plausible mechanisms include: Cognitive habit formation (frequent email/social network use builds digital routines that spill over to mobile news-seeking); Social learning (Peer interaction and radio exposure normalize mobile platforms via observational learning); Generational imprinting (Younger birth cohorts adopt mobile-first behaviors due to early-life technology exposure). These mechanisms are inferred from behavioral theory (e.g., Technology Acceptance Model, Diffusion of Innovations) but not empirically isolated, while our contribution is to identify a minimal, globally robust predictor set, a necessary foundation for such targeted research.

Additionally, while the datasets provide a broad global perspective, they may not fully reflect diverse usage patterns or account for regional nuances.

Contextual and cultural factors also present challenges. The findings are robust for the surveyed regions. They may not generalize about other populations or contexts. Urban-rural dynamics, economic disparities, and cultural influences on technology adoption remain underexplored. While regional and cultural variations are significant, the goal was to isolate predictors that are consistent and resilient across these variables. This approach contributes to a deeper understanding of mobile phone usage on a global scale. The country code, although numeric, does not represent an intensity value and lacks a specific scale that can serve in regression models. Including it in this form could lead to misleading conclusions. Therefore, it contributed only to cross-validation. The aim was generalizability (the variable with code S003, namely Numeric country code, on the lower left of Table A12 and Table A13, Appendix C).

Regarding generalizability, the reliance on limited normality assumptions and the exclusive application of SCOBIT regressions as a compensatory approach could potentially constrain the broader relevance of the outcomes.

Moreover, rapid technological evolution, including advancements such as AI and IoT, may shift the relevance of identified predictors over time, limiting the longevity of these insights.

Finally, gaps remain in understanding the deeper mechanisms of mobile phone use. Predictor interactions, such as the combined effects of income, education, and digital literacy, were not entirely examined. Behavioral drivers of information-seeking via mobile phones also require further investigation. The role of emerging technologies (such as their integration with AI and IoT) is another area that needs attention. It will likely transform how users interact with mobile devices for information.

5.6. Addressing the Limitations in Further Directions

Recognizing the limitations of this study opens exciting opportunities for future research. These directions not only address existing gaps but also pave the way for more comprehensive and nuanced insights into mobile phone usage as a source of information.

In terms of data limitation, to overcome the challenges of self-reported data, future studies can integrate objective usage metrics. These could include application activity logs or data about the screen time. This approach would provide a richer, more accurate picture of mobile phone behavior. Additionally, expanding the scope to include multiple datasets beyond the WVS and IVS could capture a broader spectrum of global usage patterns. The latter would ensure findings that reflect the diversity of user experiences.

To expand generalizability and enhance the applicability of findings, future research could target more diverse samples, extending beyond the WVS and IVS. Including populations from underrepresented regions or those with varying social and economic conditions could ensure the conclusions resonate across a broader range of contexts.

In terms of tracking technological evolution, given the rapid pace of technological advancements, longitudinal studies are crucial. By monitoring mobile phone usage over time, researchers can determine whether the predictors identified in this study remain relevant or evolve together with changing technologies and societal norms.

When it comes to exploring cultural nuances, delving deeper into the cultural dimensions of mobile phone usage could uncover fascinating insights. Future research may find how cultural values and practices shape technology adoption and usage patterns. Comparative studies across regions or cultures could also shed light on unique behaviors and preferences, enriching the global understanding of mobile phone use. Moreover, country-dependent derived numerical indicators (such as GDP or the stock market-to-GDP ratio, as well as governance indicators) may contribute to more specialized studies. Such localized metrics will enable more nuanced analysis of how regional factors influence mobile phone usage, with a stronger focus on local peculiarities.

When accounting for context, future studies could examine how factors such as urban versus rural settings or varying economic conditions interact with identified predictors. Unpacking these contextual influences would provide a more layered understanding of mobile phone use.

In terms of bridging access and literacy gaps, access to technology and digital literacy remain critical barriers for many populations. Future research could investigate how these factors affect mobile phone use. The latter is particularly relevant for underserved groups. Identifying these disparities could guide interventions aimed at promoting equitable access to digital resources.

When it comes to understanding predictor interactions, while this study focuses on individual predictors, future research could explore how these factors interact. Using advanced statistical techniques to analyze interaction effects might reveal new insights into how combined influences drive mobile phone usage patterns.

When aiming to unpack behavioral motivations and insights (or better capture why) behind mobile phone usage, qualitative methods such as interviews or focus groups could be employed. These approaches would offer richer behavioral insights, shedding light on the motivations, attitudes, and habits that underpin mobile phones use for information-seeking.

Finally, in terms of integration of emerging technologies, as new technologies like artificial intelligence and the Internet of Things become increasingly integrated into mobile ecosystems, future research could explore their impact on mobile phone usage. Examining how these innovations shape information-seeking behaviors would ensure findings remain relevant in a rapidly evolving technological landscape.

6. Conclusions

This study significantly enhances our understanding of mobile phone usage patterns as a primary source of information. It utilizes a recent and comprehensive dataset from the WVS (Time Series 1981–2022, v4.0, alongside the IVS and WVS—version 5.0, the last two for validation purposes). By employing a diverse range of advanced analytical techniques, this research identifies five robust predictors that are crucial to mobile phone information use. These techniques include the treatment of DK/NA values (REMDKNA) in Stata, ADA BOOST via Rattle in R, PCDM, RLASSO, CVLASSO, and BMA in Stata, HGB in Python, and mixed effects models for non-random cross-validations also in Stata.

The findings underscore that age is a significant determinant, revealing a negative association whereby younger individuals exhibit a greater propensity to use mobile devices for information retrieval. Additionally, the frequency of using alternative information channels positively correlates with mobile phone usage. These channels include digital mail, online SN, communication with friends or colleagues, and radio broadcasts.

The confirmation of these predictors through rigorous analysis of the WVS dataset (version 4.0) adds to the robustness of the findings. Their validation via the IVS and WVS (version 5.0), except for social media, which is absent in the IVS, further strengthens the results.

The classification model developed in this study demonstrates good (near excellent) classification accuracy. It is also free of redundancy and contains core components passing reverse causality checks, thereby ensuring its reliability. Such results, coupled with the robust predictors identified, confirm the core conclusions of our research.

Importantly, this research provides practical tools in the form of nomograms for probability prediction, impact analysis, and visual computation and interpretation. These target policymakers, researchers, and application designers alike. These intuitive visual tools can serve dual purposes. They help analysts identify and eliminate redundancies during exploratory stages and enable managers to interpret patterns, compute probabilities, and make informed predictions based on variable values and their corresponding summed scores. The insights derived from the nomograms provide concrete guidance for developing targeted interventions and more inclusive digital products.

The originality and significance of this study lie in its empirical identification of key factors influencing mobile phone usage for information. This contribution enhances the broader literature on mobile communication and social media.

Notably, the demographic analysis reveals that higher values for the target variable are associated with males, unmarried individuals, and households of four people. Furthermore, the research shows that usage increases with higher education levels, full-time employment, social class, and within communities of over 500,000 people.

These insights are essential for designing effective user interfaces for mobile applications. They provide a foundation for enhancing usability and accessibility tailored to the identified predictors.

Ultimately, this study sheds light on the dynamics of mobile information consumption. Moreover, it also offers practical implications for improving mobile applications, ensuring they are user-friendly and meet the diverse needs of varying demographic groups. Even more, policymakers should monitor whether AI agents and wearables shift reliance patterns, though ambient media (e.g., radio, podcasts) and digital habit formation are likely to remain central. Therefore, the novelty lies in the methodological rigor and the practical, tangible output that extends the usability of our findings beyond the academic community.