Title Page Doctoral Thesis Cognitive Traits and Entrepreneurial Pursuits in the Digital Age: A Multi-Layered Perspective In partial fulfillment of the requirements for the degree of Doctor of Economics and Business Studies (Doctor rerum politicarum, Dr. rer. pol.) Author Philipp Schade Submitted to Justus Liebig University Giessen Department of Business Administration and Economics Research Network Digitalization Giessen, August 10th, 2023 Supervisors Prof. Dr. Monika C. Schuhmacher Chair of the Department for Technology, Innovation, and Start-up Management Prof. Dr. Irene Bertschek Chair of the Department for Economics of Digitalisation II List of Contents List of Contents Acknowledgments .................................................................................................................... V List of Figures ....................................................................................................................... VII List of Tables ....................................................................................................................... VIII List of Abbreviations .............................................................................................................. IX List of Appendices .................................................................................................................. XI 1. General Introduction ........................................................................................................... 1 2. Study 1 – Predicting Entrepreneurial Activity Using Machine Learning ...................... 5 2.1 Introduction .................................................................................................................... 7 2.2 Data and methodology .................................................................................................... 9 2.2.1 Data, feature selection, and data pre-processing .................................................. 9 2.2.2 Methodology ...................................................................................................... 10 2.2.3 Classifier performance evaluation ...................................................................... 12 2.3 Results .......................................................................................................................... 13 2.3.1 Comparison of classifier performance ............................................................... 13 2.3.2 Global feature importance .................................................................................. 15 2.3.3 Additional analysis and robustness check .......................................................... 17 2.4 Discussion and conclusion ........................................................................................... 18 3. Study 2 – Digital Infrastructure and Entrepreneurial Action-Formation: A Multilevel Study .................................................................................................................................... 23 3.1 Introduction .................................................................................................................. 25 3.2 Theoretical framework ................................................................................................. 28 3.2.1 Social cognitive theory ....................................................................................... 28 3.2.2 External enabler framework and venture creation ............................................. 29 3.3 Mechanism-based theorizing: Hypotheses development ............................................. 29 III 3.3.1 Action-formation mechanisms at the individual level and baseline hypotheses .......................................................................................................... 31 3.3.2 EE mechanisms of digital infrastructure and moderation hypotheses ............... 33 3.3.2.1 Moderation of the self-efficacy action-formation mechanism .............. 34 3.3.2.2 Moderation of the fear of failure as an action-formation mechanism ... 36 3.3.2.3 Moderation of the opportunity recognition action-formation mechanism ............................................................................................ 37 3.4 Research methodology ................................................................................................. 39 3.4.1 Data…… ............................................................................................................ 39 3.4.2 Variables and measures ...................................................................................... 40 3.4.2.1 Dependent variable ............................................................................... 40 3.4.2.2 Key explanatory variables ..................................................................... 40 3.4.2.3 Control variables ................................................................................... 41 3.4.2.4 Empirical multilevel model ................................................................... 43 3.5 Results .......................................................................................................................... 43 3.5.1 Descriptive statistics ........................................................................................... 43 3.5.2 Multilevel logistic regression results .................................................................. 43 3.5.3 Additional analyses and robustness checks ........................................................ 50 3.6 Discussion .................................................................................................................... 53 3.6.1 Contribution to the contextual entrepreneurship literature ................................ 54 3.6.2 Contribution to the external enabler framework ................................................ 55 3.6.3 Policy implications ............................................................................................. 56 3.6.4 Limitations and directions for future research ................................................... 58 4. Study 3 – Responding to the Situational Urgency of Digital Transformation: A Multilevel Analysis of CEO Humility and Corporate Venture Capital ........................ 61 4.1 Introduction .................................................................................................................. 63 4.2 Theoretical background ................................................................................................ 66 4.2.1 The attention-based view from the CEO perspective ......................................... 66 IV 4.2.2 CEO humility and CVC investments as a CE action ......................................... 67 4.2.3 Digital transformation as a context of situational urgency ................................ 68 4.3 Hypotheses development .............................................................................................. 70 4.3.1 CEO humility and CVC investments ................................................................. 70 4.3.2 The influence of external situational urgency for digital transformation ........... 71 4.3.3 The influence of internal situational urgency for digital transformation ........... 73 4.4 Method .......................................................................................................................... 74 4.4.1 Data and measures .............................................................................................. 74 4.4.2 Dependent variable ............................................................................................. 75 4.4.3 Independent variable .......................................................................................... 75 4.4.4 Moderator variables ............................................................................................ 76 4.4.5 Control variables ................................................................................................ 77 4.5 Analysis and results ...................................................................................................... 79 4.5.1 Analysis .............................................................................................................. 79 4.5.2 Results…. ........................................................................................................... 80 4.5.3 Additional analyses and robustness checks ........................................................ 83 4.6 Discussion .................................................................................................................... 86 4.6.1 Theoretical contributions .................................................................................... 87 4.6.2 Practical implications ......................................................................................... 88 4.6.3 Limitations and future research .......................................................................... 89 4.7 Conclusion .................................................................................................................... 90 5. Concluding Remarks .......................................................................................................... 92 References ............................................................................................................................... 95 Appendices ............................................................................................................................ 117 Affidavit ................................................................................................................................ XII V Acknowledgments In the following, I would like to take the opportunity to thank all the people who made this doctoral thesis possible. First, I would like to thank my supervisor Prof. Dr. Monika Schuhmacher. I am deeply grateful for the trust she has placed in me and my research at all times. Through her active integration of the Research Network Digitalization into the Chair of Technology, Innovation, and Start-up Management, she has provided a fruitful research environment. I am very grateful that her door was always open and that she brought her perspective, sense of structure, rigor, and questions to the discussions on each of the studies. Her constructive feedback had a tremendously positive impact on each of the three papers. I am also very appreciative to Prof. Dr. Irene Bertschek for immediately agreeing to assist my dissertation as a second supervisor. Beyond that, and besides Prof. Dr. Monika Schuhmacher, I would also like to thank all other members of the Research Network Digitalization. Namely, Prof. Dr. Andreas Bausch, Prof. Dr. Christian Gissel, Prof. Dr. Georg Götz, Prof. Dr. Alexander Haas, and Prof. Dr. Frank Walter. Without the existence of the Research Network Digitalization, this work would not have been possible. For various reasons, I am very fortunate and deeply grateful for my colleague, co-author, and friend Dr. Petrit Ademi. I could not have imagined a person with whom I would have preferred to share an office over the past years. Our conversations have greatly enriched my work. Among many others, I will always remember our intensive discussions and a subsequent conciliatory attitude toward the notion of affordances and external enabler mechanisms. Great appreciation is also due to all my other companions and colleagues. Specifically, Yannick Amend, Dr. Sadrac Cenophat, Dr. Anna-Lena Hanker, Björn Hofmann, Junior Prof. Dr. Tobias Krämer, Victoria Kuharev, Julian Nickel, Dr. Stephan Philippi, Hieu Thieu, Alexandra von Preuschen, Denis Weinecker, and Ferogh Schaich-Zaman who have contributed to my research through various colloquia, doctoral seminars, and brown-bag conversations. I also VI acknowledge Carmen Wagner for taking care of countless administrative affairs, which resulted in more space on my desk for research. I would also like to address further words of thanks to the student assistants Christian Nicolay and Alicia Leona Schwalbach. Their support in teaching and research was a great help. Furthermore, I am appreciative to Dr. Yannik Bofinger, Benjamin Fiorelli, Florian Gärtner, and Dr. Darwin Semmler from the Research Network Behavioral and Social Finance & Accounting. The dialogues with you on various topics related to research and teaching were always enriching. I would like to dedicate another acknowledgment to a person who may not be aware of his contribution to this doctoral thesis. This person is Prof. Dr. Henrik Egbert, for whom I had the opportunity to work as a tutor for microeconomics and economic policy during my bachelor studies and who had an important and lasting influence on my later academic pursuits. Finally, and most importantly, I would like to thank my entire beloved family for their patience, encouragement, and advice. I especially thank my mother Annette and my twin brother Mark. You have been the greatest of all supports in the most diverse areas throughout my entire educational path. Another person I would like to thank explicitly is Lara. Without you, this doctoral thesis and the journal publications would not have been possible. You have been a bulwark of support during my doctoral years and have kept me going in extremely tense times, even if it meant personal privations for you. This is anything but a matter of course. I owe her a great debt of gratitude—Thank you! VII List of Figures Figure 1: Structure of the doctoral thesis ................................................................................... 3 Figure 2: ROC curve comparison for OME ............................................................................. 15 Figure 3: Feature importance rank for predicting OME .......................................................... 16 Figure 4: Theoretical model ..................................................................................................... 31 Figure 5: Interaction plots – Entrepreneurial self-efficacy and digital infrastructure .............. 49 Figure 6: Interaction plots – Fear of failure and digital infrastructure ..................................... 49 Figure 7: Interaction plots – Opportunity recognition and digital infrastructure ..................... 49 Figure 8: Conceptual framework ............................................................................................. 69 Figure 9: Number of CVC investments as predicted by CEO humility and emerging digital competition – 3D surface plot ................................................................................. 83 VIII List of Tables Table 1: ML model performance on the OME “hold-out”- dataset ......................................... 14 Table 2: Correlation matrix for country and individual level variables ................................... 45 Table 3: Mixed-effects multilevel logistic regression model ................................................... 46 Table 4: Descriptive statistics and correlation matrix for CEO-, firm- and industry-level variables ..................................................................................................................... 81 Table 5: Multilevel negative binomial regression .................................................................... 82 Table 6: Related vs. unrelated CVC investments – Multilevel analysis .................................. 85 Table 7: Firm value effect of CVC investments – Lagged fixed-effects panel analysis .......... 86 IX List of Abbreviations AI Artificial intelligence AIC Akaike's information criterion ALE Accumulated local effects ANN Artificial neural network APS Adult population survey AUC Area under the receiver operating characteristic curve AutoML Automated machine learning BoW Bag-of-words CATA Computer-aided text analysis CE Corporate entrepreneurship CEO Chief executive officer CI Confidence intervals CRE Correlated random effects CTA Company primary technology application CVC Corporate venture capital DT Decision tree EE External enabler FE Fixed effects FN False negative FP False positive GEM Global entrepreneurship monitor GCV Global corporate venturing ICC Intra-class correlation coefficients ITU International telecommunication union IML Interpretable machine learning kNN k-nearest neighbor LIME Local interpretable model-agnostic explanation LIWC Linguistic inquiry word count LPM Linear probability model LR Logistic regression LTS Letter to shareholders MCC Mathews correlation coefficient ML Machine learning NB Naïve bayes X NME Necessity-motivated entrepreneurial activity OME Opportunity-motivated entrepreneurial activity RE Random effects RF Random forest PE Private equity R&D Research and development ROC Receiver operating characteristic curve SCT Social cognitive theory SIC Standard industry classification SMOTE Synthetic minority over-sampling technique TN True negative TP True positive VC Venture capital VIF Variance inflation factors VPC Variance partitioning coefficients WVS World values survey XAI Explainable artificial intelligence XGBoost Extreme gradient boosting tree ensemble XI List of Appendices Appendix A: Feature definitions ............................................................................................ 117 Appendix B: Incidence of missing values .............................................................................. 118 Appendix C: Patterns of missing values ................................................................................ 118 Appendix D: Pearson cross-correlation heatmap ................................................................... 119 Appendix E: Confusion matrices ........................................................................................... 119 Appendix F: Error rates of the k-fold cross-validation .......................................................... 120 Appendix G: Default hyperparameter settings ....................................................................... 120 Appendix H: ML model performance on the NME “hold-out”- dataset ............................... 120 Appendix I: ROC curve comparison for NME ...................................................................... 121 Appendix J: Feature importance rank for predicting NME ................................................... 121 Appendix K: Data description and sources ............................................................................ 122 Appendix L: Summary statistics ............................................................................................ 123 Appendix M: Observations and digital infrastructure per country ........................................ 124 Appendix N: Additional analyses and robustness checks ...................................................... 125 Appendix O: Sensitivity test .................................................................................................. 126 Appendix P: CRE and LPM model for country-fixed effects ................................................ 127 Appendix Q: Data description and source ............................................................................. 128 Appendix R: Fixed-effects negative binomial panel regression ............................................ 129 Appendix S: Random-effects negative binomial panel regression ........................................ 130 1 Chapter 1 1. General Introduction The overarching aim of this cumulative dissertation is to shed light and advance our understanding of the nexus between cognitive traits and entrepreneurial pursuits in the digital age. To achieve this goal, we take a multi-layered perspective throughout the course of three separate studies. To this end, we take into account the individual, firm, industry and country levels as units of analysis across the studies conducted. A multi-layered perspective—i.e., looking at the different hierarchical levels—is of particular importance because it enables the elucidation of mechanisms that explain why effects or relationships within and between the levels occur. Moreover, since the phenomenon of “digitalization” can be characterized by various aspects, this doctoral thesis considers “digital” in different ways throughout the studies, viz. digital technology, digital infrastructure, and digital transformation. Therewith, this dissertation does not only provide a multi-layered perspective in terms of the hierarchical levels considered—i.e., depth—but also regarding the breadth of domains that digitalization comprises. The structure of the dissertation is presented in Figure 1. In Study 1 (see Chapter 2), we use digital technology, i.e., machine learning (ML), a subfield of artificial intelligence (AI), that marks the very center of the so-called digital era. Specifically, we apply ML algorithms and pursue a data-driven approach to unravel important agent-centric features for entrepreneurial pursuits. This way, ML enables us to investigate to what extent entrepreneurial activity is predictable, and more importantly, which features best explain the prediction. To this purpose, we apply various supervised machine-learning techniques—decision tree, random forest, extreme gradient boosting tree ensemble, k-nearest neighbor, artificial neural network, and naïve Bayes—as well as perform classical multiple logistic regression to the most comprehensive existing data set in the field of entrepreneurship. This approach enables us to engage in abductive reasoning and estimate the relative performance of the respective ML algorithm in predicting entrepreneurial activity. The benchmarking of the ML techniques reveals 2 that the extreme gradient boosting tree ensemble is the best-performing ML technique in predicting both necessity and opportunity-motivated entrepreneurial activity with the highest overall accuracy and area under the receiver operating characteristic curve. The feature importance rankings suggest that despite psychological self-regulation mechanisms such as cognitions and personal traits (i.e., socio-cognitive traits), macro country-level factors external to the respective individual at the micro level may also play a pivotal role in the engagement in entrepreneurial pursuits. In Study 2 (see Chapter 3), we engage in a hypothetico-deductive reasoning approach. Specifically, this study examines how the level of digital infrastructure of a country shapes the relationships between socio-cognitive traits and entrepreneurial action—so-called action- formation mechanisms. For our hypothetico-deductive approach, we combine the agent-centric social cognitive theory (SCT) (Bandura, 1986; Sherman et al., 2015; Wood & Bandura, 1989) with the external enabler (EE) framework (Davidsson et al., 2020). Given that SCT is rather coarse-grained and lacks a theoretical explanation of how contextual factors shape entrepreneurial action-formation at the micro-level, we augment SCT with the EE framework and engage in EE mechanism-based theorizing, which allows to reason on how a specific contextual factor, in terms of EE, develops through specific (situational) mechanisms. To investigate the triadic reciprocal relationship system between action-formation at the micro-level and digital infrastructure at the macro-level, we apply multilevel modeling. In line with the SCT, our analysis shows that the socio-cognitive traits of entrepreneurial self-efficacy and opportunity recognition increase entrepreneurial action, while fear of failure reduces it. The results further indicate that a country’s level of digital infrastructure is an EE that takes a shaping role in the relationships between socio- cognitive traits and entrepreneurial action. Consistent with our theorizing derived from the EE framework, the findings suggest that, in particular, the resource access and market access mechanisms of digital infrastructure explain the moderating effects. 3 Figure 1: Structure of the doctoral thesis Chapter 1: General Introduction Structure of Doctoral Thesis Chapter 2: Study 1 Chapter 3: Study 2 Chapter 4: Study 3 T it le a n d co -a u th o rs Predicting Entrepreneurial Activity Using Machine Learning with Monika C. Schuhmacher Digital Infrastructure and Entrepreneurial Action-Formation: A Multilevel Study with Monika C. Schuhmacher Responding to the Situational Urgency of Digi- tal Transformation: A Multilevel Analysis of CEO Humility and Corporate Venture Capital with Petrit Ademi, & Monika C. Schuhmacher R es ea rc h q u es ti o n T h eo ry M et h o d Chapter 5: Concluding Remarks To what extent is entrepreneurial activity predictable and what are the most important features? How does a country’s digital infrastructure shape the entrepreneurial action-formation of individuals? How do humble CEOs influence CVC investment activity in the context of urgency for digital transformation? Atheoretical Social cognitive theory (Wood & Bandura, 1989) and external enabler framework (Davidsson et al., 2020) Attention-based view (Ocasio, 1997) Abductive analysis of various machine learning algorithms based on 1,192,818 observations from 99 countries Logistic multilevel modeling based on 344,265 individual-level observations from 46 countries Longitudinal study based of 373 CEOs from 198 firms and 35 industries between 2010 and 2019 (6,908 CVC investments over 1,597 firm-years) S ta tu s Published in Journal of Business Venturing Insights (ABDC: A) Published in Journal of Business Venturing (FT50, ABS: 4, ABDC: A*, VHB-Jourqual: A) Revise & Resubmit in Journal of Management Studies (FT50, ABS: 4, ABDC: A*, VHB-Jourqual: A) L ev el Individual (micro) and country (macro) Individual (micro), firm (meso), and industry (macro) Individual (micro) 4 In Study 3 (see Chapter 4), we examine whether the findings from Study 1 and 2 are transferable to the corporate context. The underlying assumption is that incumbent firms are increasingly fostering corporate entrepreneurial (CE) actions as a means of addressing challenges arising from the ongoing digital transformation of society and business. Existing research in the management and entrepreneurship literature has considerably advanced our knowledge of firm and industry-specific drivers of corporate venture capital (CVC) investment activity. However, the influencing role of CEOs is largely unknown. In digital times, the purposeful instigation of CE actions requires the CEO—as the top decision-maker and head of the organization—to not only uphold the company’s existing strengths but also to identify weaknesses, obtain accurate self-knowledge and promote self-improvement through continual learning (Ou et al., 2018). An auspicious basis for such contemporary executive leadership that gained momentum in the literature is a person-centered cognitive characteristic known as “humility”. In the study, we distinguish between two forms of urgency for digital transformation: (i) emerging digital competition at the external industry-level, and (ii) business model dependence on information and knowledge at the internal firm-level. By reasoning upon the situational mechanism emanated by the internal and external urgency for digital transformation, we aim to provide an understanding of CVC investments as a CE action that humble CEOs at the individual-level foster in the digital era. Through the application of a bag-of-words (BoW) approach for text analysis and multilevel modeling, we provide evidence for CEO humility as an important, yet overlooked, action- formation mechanism for CVC investment activity. While the study finds support for the moderating role of emerging digital competition (i.e., external urgency), the findings suggest that internal urgency for digital transformation originating from the firm’s business model dependence on information and knowledge positively moderates the action-formation mechanism of CEO humility primarily for CVC investment activity in related ventures. 5 Chapter 2 2. Study 1 – Predicting Entrepreneurial Activity Using Machine Learning Coauthors: Monika C. Schuhmacher Relative share: 90% Status: Published in Journal of Business Venturing Insights (ABDC: A) This chapter is available under: Schade, P. & Schuhmacher, M.C. (2023). Predicting entrepreneurial activity using machine learning. Journal of Business Venturing Insights, 19, e00357, https://doi.org/10.1016/j.jbvi.2022.e00357 A previous version of this chapter has been presented at: • Australian Center for Entrepreneurship Research Exchange (ACERE) Conference 2023, Brisbane, Australia • Doctoral Consortium of the Australian Center for Entrepreneurship Research Exchange (ACERE) Conference 2022, Melbourne/Virtual Edition, Australia https://doi.org/10.1016/j.jbvi.2022.e00357 6 Predicting Entrepreneurial Activity Using Machine Learning Abstract This study evaluates the predictability of entrepreneurial activity using machine learning. We compare different supervised machine learning techniques: decision tree, random forest, artificial neural network, k-nearest neighbor, extreme gradient boosting tree ensemble, and naïve Bayes, as well as run the traditional multiple logistic regression for obtaining a baseline and estimating their relative model prediction performance on a Global Entrepreneurship Monitor dataset of 1,192,818 individuals from 99 countries. By comparing different machine learning techniques, we predict out-of-sample opportunity-motivated entrepreneurial activity with an overall accuracy ranging from 70.1% to 91.2%. The results demonstrate that the extreme gradient boosting tree ensemble is superior in predicting opportunity-motivated entrepreneurial activity. Finally, a global surrogate model reveals that knowing an entrepreneur, entrepreneurial self-efficacy, and opportunity recognition are the three most important features for predicting opportunity-motivated entrepreneurial activity. For comparison purposes, we perform the same analyses for necessity- motivated entrepreneurial activity. The results reveal that the extreme gradient boosting tree ensemble is also the best-performing technique in predicting this form of entrepreneurial activity with a 96.5% accuracy. JEL classification: C45, C53, C55, D91, L26 Keywords: Supervised machine learning, classification, prediction, entrepreneurial activity 7 2.1 Introduction Over the last few decades, scholars have attempted to unravel the focal phenomenon of entrepreneurial activity from different perspectives and found that disentangling the entrepreneurial event is extremely complex. Against this background, entrepreneurship scholars have proposed the ignoramus et ignorabimus-like thesis in leading journals that it is highly improbable and doubtful whether research will ever be able to construct a mathematical model that can be used to predict the occurrence of the entrepreneurial event (e.g., Bruyat & Julien, 2001; Churchill & Bygrave, 1990). These scholars argue that if “we want to understand entrepreneurship, our research methodology must be able to handle nonlinear, unstable discontinuities” (Churchill & Bygrave, 1990, p. 28). However, the advancements in the area of artificial intelligence (AI) and machine learning (ML) in recent years have provided researchers with new methodological potentials for constructing models for predicting various human behaviors and offering fine-grained insights into the actual predictability of entrepreneurial events. As such, ML enjoys the greatest popularity in the business world and is used for performing the most complex prediction tasks, especially supervised machine learning techniques, which were designed for this purpose (Obschonka & Audretsch, 2020). The main benefit of ML is its high predictive accuracy, a property that is crucial in many areas of business (van Witteloostuijn & Kolkman, 2019). However, although the disruptive potentials of AI and ML in analyzing (big) data with a large number of observations or high dimensionality have received increasing attention in a variety of research and application fields, they have not undergone much scrutiny in contemporary entrepreneurship research yet (Hastie et al., 2009; Obschonka & Audretsch, 2020; Schwab & Zhang, 2019; Shepherd & Majchrzak, 2022; van Witteloostuijn & Kolkman, 2019). This fact is surprising because AI-based ML provides mathematical approaches that can be applied to mine and analyze the most comprehensive datasets such as the Global Entrepreneurship Monitor (GEM) (Gerasimovic & Bugaric, 2018; Lévesque et al., 2022) and, consequently, 8 challenge long-held assumptions about the predictability of entrepreneurial activity. In particular, ML offers algorithmic approaches that “learn” incrementally from the inferred data to make predictions by accommodating complex and high-order interactions. This way, ML techniques can select the functional form that best predicts the target outcome, whereas in classic statistical methods the functional form must be specified a priori (Arin et al., 2022). Moreover, ML techniques allow for the investigation of the “nuts and bolts, cogs and wheels” (Elster, 1989, p. 3), i.e., mechanisms, that lead to entrepreneurial activity. Understanding these mechanisms resulting from entrepreneurial activity-related features is critical (Cowen et al., 2022; Hedström & Swedberg, 1998), as entrepreneurial activity plays an important role in the economic growth and prosperity of a nation (e.g., Schumpeter, 1934; Wennekers & Thurik, 1999). Therefore, the better scientists and policymakers understand the relevant features and underlying mechanisms of entrepreneurial activity, the better they can leverage these aspects through specific actions. However, this understanding calls for research into effectively predicting entrepreneurial activity. To answer the long-standing research question of whether and to what extent entrepreneurial activity can be predicted, we apply and compare multiple state-of-the-art supervised machines and deep learning1 techniques to a GEM dataset of 1,192,818 individuals from 99 countries. As no large-scale investigation using ML has been conducted to date, the aim of this research is primarily to evaluate whether ML improves the accuracy of entrepreneurial activity prediction. Since causal inferences are not possible due to the cross-sectional nature of the individual-level GEM data, we predominantly seek to determine which supervised ML technique is superior in terms of predictive accuracy and identify which features are the most relevant in predicting entrepreneurial activity. Typically, entrepreneurship research distinguishes entrepreneurship activity either as a necessity-motivated entrepreneurial activity (NME) or an opportunity- motivated entrepreneurial activity (OME) (Amorós et al., 2019). According to this push/pull 1 Since deep learning techniques, such as artificial neural networks, are a special case of ML, we use the term “machine learning” throughout the paper for greater clarity. 9 framework (Storey, 2016), NME is linked to unemployment and economic recession (e.g., Amorós et al., 2019; Shane, 2009). Thus, entrepreneurship research is mostly interested in OME, wherein individuals start a new business venture in pursuit of profit, innovation, and growth (see e.g., Reynolds et al., 2005; Stenholm et al., 2013). Hence, we apply different ML techniques to OME. 2.2 Data and methodology 2.2.1 Data, feature selection, and data pre-processing This study uses the data of six years from the Adult Population Survey by the GEM initiative. We built a comprehensive cross-sectional sample of 1,192,818 individuals from 99 countries by pooling the individual-level GEM data from 2012 to 2017, with different individuals being observed in each year (Verbeek, 2008). As the GEM data provides labeled observations, the dataset is suitable for conducting supervised analyses. The GEM initiative estimates the prevalence rate of entrepreneurial activity across the participating countries (Reynolds et al., 2005). The GEM data have been used in various previous studies examining entrepreneurial activity (e.g., Aidis et al., 2008; Fredström et al., 2020). As we aim to predict OME, we relied on the individuals from the GEM project who provided their assessment of whether they engage in entrepreneurial activity to take advantage of a business opportunity (TEAyyOPP). The GEM specifies OME as a binary feature (1 = yes, 0 = no). To ensure that OME is not simply predicted by the inherently non-orthogonal manifestations of this feature (e.g., total early-stage entrepreneurial activity, total early-stage entrepreneurial activity based on new technology, self-employment, etc.), we have removed these manifestations from the GEM datasets.2 2 For example, if an individual engages in OME (TEAyyOPP), the same person also tends to pursue total early-stage entrepreneurial activity (TEAyy). The ML techniques would, therefore, use TEAyy as the most important feature in predicting TEAyyOPP. 10 We selected and included all agent-centric features from the unaggregated Adult Population Survey that were reported by the GEM. Specifically, we included several human capital endowment features, such as educational and occupational attainments (see e.g., Davidsson & Honig, 2003); socio-demographic characteristics, such as gender, age, household size (hhsize); and self-regulatory mechanisms such as cognitions and personal traits—entrepreneurial self- efficacy (suskill), fear of failure (fearfail), and opportunity recognition (opport) (e.g., Baron, 2004; Mitchell et al., 2002; Shaver & Scott, 1992). Overall, we included a total of 21 explanatory features. These features are queried in the GEM datasets, as entrepreneurship research has shown that these features are significantly associated with entrepreneurial activity3. Feature definitions are presented in Appendix A. From the initial dataset, observations with missing values and those which are string data (i.e., open-ended survey questions) were excluded from the analyses. Additional information on the incidence and patterns of missing data are provided in Appendix B and C. To account for potential survey effects and unobserved temporal heterogeneity, we also included respondent identifiers (setid), country identifiers (country), year of survey (yrsurv), and the developmental stage of countries (CAT_GCR1) in the analyses. 2.2.2 Methodology To predict OME, we utilized various supervised ML techniques. Since the target class, OME, is binary (i.e., unordered discrete response), we applied the most commonly used ML techniques suitable for solving classification problems. Specifically, we applied the following ML techniques: decision tree (DT), random forest (RF), deep artificial neural network (ANN) in the form of a feedforward multilayer perceptron, k-nearest neighbor (kNN), extreme gradient boosting tree ensemble (XGBoost), and naïve Bayes (NB).4 For comparison, we also ran the traditional multiple logistic regression (LR) as a baseline-benchmark model. All ML techniques 3 Features queried as special topics in individual GEM rounds are not taken into consideration. 4 For the sake of brevity, we do not describe the ML methods used in greater detail. For a more comprehensive overview into the individual methods of statistical learning, the reader can refer to Hastie et al. (2009). 11 were used with their default hyperparameter settings.5 The use of default hyperparameter settings ensures c.p. direct comparability of different ML techniques without additional human intervention. Hyperparameter settings for the applied ML techniques are reported in Appendix G. To compare these ML techniques, we split the GEM dataset into a “training & validation” sample and an unseen “hold-out” test sample (Mullainathan & Spiess, 2017). For training and validating the ML models, we used 70% of the sample. For evaluating the final predictive out-of-sample performance of the ML techniques, we used the remaining 30% (see Choudhury et al., 2021). In a balanced dataset, the probabilities of engaging and not engaging in OME are equal. However, since entrepreneurial activity is a rather rare event, the input data is unbalanced. In other words, there are fewer observations for engaging in OME (i.e., 1 = yes) than for not engaging in OME (i.e., 0 = no). This unbalanced data lead to the issue of ML techniques being dominated by the majority class. To address this issue, we resampled the dataset. Specifically, we performed the synthetic minority over-sampling technique (SMOTE) suggested by Chawla et al. (2002), where data in the minority class is generated through over-sampling. This minority over-sampling was achieved by creating synthetic rows in the dataset by extrapolating between a real object of a given class and one of its nearest neighbors. Thus, the SMOTE increased the number of minority class observations, thereby improving the generalizability of the ML techniques (Chawla et al., 2002; Fernandez et al., 2018). Further, to ensure the reasonable predictive performance of different models in out-of-sample prediction, we employed the k-fold cross-validation technique (Geisser, 1975; Stone, 1974). Therewith, the training data were split randomly into k approximately equal-sized subsets of data. These k subsets were used separately as validation data, i.e., a pseudo-hold-out sample, for assessing the predictive ability of the ML model, whereas the other 𝑘 − 1 subsets were used to train the ML model. Moreover, k-fold cross-validation provides a means to the (over)fitting 5 We adjusted the default setting of parameters in case the default values prevented the ML technique from technical functioning (e.g., the DT with the minimum number records per node and the number of threads). 12 conundrum, as cross-validation makes model prediction performance less sensitive to idiosyncrasies in any of the k subsets (Choudhury et al., 2021; Shao, 1993). For our calculation, we used 10-fold cross-validation (k = 10), which is a common choice for k (Choudhury et al., 2021; Kohavi, 1995). Besides cross-validation, we followed the recommendation of Choudhury et al. (2021) and normalized the scale of features to obtain a unit variance of each feature after splitting the data into the “training & validation” sample and “hold-out” partitions by building z- scores.6 With such feature scaling, we ensure that features with greater magnitude do not outweigh features with smaller magnitudes when they are weighted by an ML technique. This can be especially crucial when ML techniques backpropagate information to update weights, such as in ANNs. 2.2.3 Classifier performance evaluation To evaluate the performance of the ML techniques, i.e., classifiers, we rely on different prominent classification performance scores that are calculated based on the confusion matrix. In Appendix D, we provide a Pearson cross-correlation heatmap for all features used in the main analyses. Specifically, we capture true positive (TP), false positive (FP), true negative (TN), and false negative (FN) scores. In this notation, TPs are positive instances correctly predicted by an ML technique as positive; FP is the number of negative cases that an ML technique predicted as positive; TNs are negative instances in which the classifier correctly predicted them to be negative; FNs are the number of positive cases that the classifier incorrectly predicted as negative. Therewith, this 2 × 2 confusion matrix is useful in understanding the balance between FNs and FPs predicted by a specific classifier (Choudhury et al., 2021). Based on these metrics, we rely on the most widely used classifier performance indices— precision, recall/sensitivity, F1-score, and accuracy—to evaluate the out-of-sample performance and compare the overall model prediction performance (Bergstra & Bengio, 2012; Choudhury et 6 To prevent information leakage, z-score normalization and SMOTE were performed after splitting the sample. 13 al., 2021; Mullainathan & Spiess, 2017). Moreover, we use the receiver operating characteristic (ROC) and investigate the area under the curve (AUC) estimates in order to evaluate and compare the overall performance of the group of classifiers. The ROC plots the TP rate of the confusion matrix against the FP rate. The AUC metric reflects the predictability of an ML technique and can be used to compare the superiority of a model. In addition, based on recent publications, we report the Mathews correlation coefficient (MCC) as an additional informative and reliable statistical score for evaluating binary classification tasks (Boughorbel et al., 2017; Chicco & Jurman, 2020). 2.3 Results 2.3.1 Comparison of classifier performance Appendix E (Panel A) represents the 2 × 2 confusion matrices for all fitted ML models. Based on the confusion matrices, the overall classification performance of the respective supervised ML techniques in predicting OME is depicted in Table 1. Table 1 summarizes the out-of-sample performance results achieved by each ML technique with regard to the different performance scores described in section 2.2.3. The results in this table reveal that the RF model obtained a maximum overall accuracy of 91.2% in predicting OME, which is followed closely by the XGBoost model with an accuracy of 91.1%. In comparison, the kNN and the NB classifiers underperformed with a total accuracy of 80.0% and 70.1%, respectively. If we compare the different ML techniques with respect to the AUC, a marginally different picture emerges—the XGBoost model with an AUC value of 0.850 is superior to the other classifiers. A comparison of the ROC curves across all ML techniques with the corresponding AUCs is shown in Figure 2 (AUC estimates are in parenthesis). When considering the MCC score, the kNN model performs the best in predicting OME with a value of 0.289. A comparative look at the different performance scores reveals that employing ML techniques for predicting OME based on the comprehensive GEM data outperforms the baseline, multiple LR model, which shows the second lowest prediction accuracy (72.7%) and AUC (0.799) 14 estimates. The NB technique performs the worst, with a prediction accuracy of 70.1% and an AUC estimate of 0.782. Furthermore, if we explicitly compare the difference between the best-performing ML model (i.e., XGBoost with an AUC of 0.850, accuracy = 0.911, and precision = 0.506) and the baseline LR (AUC = 0.799, accuracy = 0.727, and precision = 0.207), we can state that LR performs 18.4 percentage points (0.911− 0.727 = 0.184 × 100) worse at accurately predicting OME. Looking at the precision score as an estimate for how reliable a model can predict TP instances (i.e., class = 1) of OME—the proportion of relevant instances belonging to the positive class—we also see that LR correctly predicts an actual opportunity-motivated entrepreneur in only 20.7% of all cases. In comparison, the XGBoost model is 29.9 percentage points (0.506 − 0.207 = 0.299 × 100) better than LR in predicting OME; thus, it correctly predicts almost half of the actual OMEs. Given the fact, that LR primarily predicts non-OMEs precisely (class = 0), but fails in predicting positive data instances, the MCC for LR produces an overly optimistic, inflated score of 0.280. Table 1: ML model performance on the OME “hold-out”- dataset ML technique Class Recall Precision F1-score Accuracy AUC MCC DT 0 0.969 0.922 0.945 - - - 1 0.164 0.343 0.222 - - - Total - - - 0.898 0.814 0.188 RF 0 0.993 0.917 0.954 - - - 1 0.083 0.530 0.144 - - - Total - - - 0.912 0.826 0.184 ANN 0 0.920 0.939 0.930 - - - 1 0.388 0.322 0.352 - - - Total - - - 0.873 0.816 0.284 kNN 0 0.819 0.955 0.882 - - - 1 0.601 0.245 0.348 - - - Total - - - 0.800 0.803 0.289 XGBoost 0 0.991 0.918 0.953 - - - 1 0.095 0.506 0.160 - - - Total - - - 0.911 0.850 0.191 NB 0 0.698 0.964 0.810 - - - 1 0.734 0.191 0.304 - - - Total - - - 0.701 0.782 0.259 LR - baseline 0 0.727 0.965 0.829 - - - 1 0.729 0.207 0.322 - - - Total - - - 0.727 0.799 0.280 However, if we consider the most precise ML models (i.e., XGBoost and RF), it must be noted that even these ML techniques are slightly better than a random draw at correctly predicting TPs 15 of OME. The estimates of prediction error rates for each ML technique in the respective fold of the cross-validation can be inferred from Appendix F. Figure 2: ROC curve comparison for OME 2.3.2 Global feature importance To obtain a more detailed understanding of individual features in predicting entrepreneurial activity and enhance the interpretability of the outputs, we report on the global model-agnostic feature importance (Molnar et al., 2020). For this, we rank the importance of explanatory input features of the best-performing supervised ML technique in order to show how important each feature is on average in predicting OME. To estimate feature importance, we employ a global surrogate model method. A surrogate model creates a model that is trained to mimic the behavior of the original model by finding an approximation function (Crombecq et al., 2011). In other words, the surrogate model can make the same predictions as the original model and, thus, can be used to understand how the different input features are related to the final prediction (Crombecq et al., 2011; Gorissen et al., 16 2009). Specifically, as the RF model performs best in true positively predicting OME (see precision score), we perform a surrogate RF model. Global feature importance is determined by counting how often a specific feature was selected for a split in the DTs and identifying the rank of a feature among all other available explanatory input features in the RF model trees. Figure 3 reports the global feature importance rank for predicting OME.7 Figure 3: Feature importance rank for predicting OME This figure suggests that knowing an entrepreneur (knowent), entrepreneurial self-efficacy (suskill), and opportunity recognition (opport) are on average the three most important agent- centric features for predicting entrepreneurial activity. Furthermore, the rank of a country’s developmental stage (CAT_GCR1) also indicates that a nation’s business environment in which an individual is located plays a pivotal role in entrepreneurial activity prediction. On the other end of the continuum, we see that features such as gender and public media coverage of successful entrepreneurs (nbmedia) both play only a minor role in OME prediction. 7 Features with higher values are more important for predicting the target feature (i.e., they have a larger effect on the model). The feature importance scale is relative. 17 2.3.3 Additional analysis and robustness check To validate the predictive performance of the applied supervised ML techniques in predicting OME, we rerun all of the ML models using an automated ML technique (AutoML). Although we previously use the default parameter settings to better compare the used ML techniques, we now allow for hyperparameter optimization within the AutoML. Hyperparameters are a set of adjustable parameters that are unique to each ML technique. These parameters are assigned and tuned manually by the researcher to prevent overfitting or underfitting (i.e., stopping rules for rule-based DTs or the choice of a regularization term added to the loss function to penalize growing model complexity) and to further improve the out-of-sample predictive performance of the ML techniques (Choudhury et al., 2021; Mullainathan & Spiess, 2017). In the AutoML, the hyperparameters are automatically selected and tuned, which reduces the necessity for human interventions and, thus, increases overall comparability (Prüfer & Prüfer, 2020). The AutoML uses a Python-based library to find optimal values for regularization and other hyperparameters. As in the previous ML models, in the AutoML, we also implement 10-fold cross-validation. According to overall prediction accuracy and compared with the AUC values depicted in Figure 2, the results of the AutoML technique are identical to our main findings that XGBoost (AUC = 0.949) and DT (AUC = 0.936) are the best-performing ML techniques, followed by RF (AUC = 0.911). Since we allow for optimal parameterization in the AutoML technique, the OME prediction performance of all the applied ML techniques is higher compared to their performance according to the results in our main analysis. Until now, we have used ML techniques to predict OME. However, according to the push/pull framework (Storey, 2016), entrepreneurial activity can also occur in the form of NME (TEAyyNEC). To account for this distinction and to compare OME vis-à-vis NME, we rerun our entire analyses for NME.8 Appendix E (Panel B) presents the confusion matrix for the fitted ML models on NME. In Appendix H, we list the detailed results of the ML model performance 8 For the prediction of NME, the same default parameter settings were used as for the prediction of OME (see Appendix G). 18 measures for predicting NME. Appendix I depicts the ROC curve comparison for NME. The results unveil that the XGBoost model again obtained a maximum overall accuracy of 96.5% in predicting NME, as well as the highest precision score (0.491) and AUC (0.824). The feature importance rank in Appendix J illustrates that the developmental stage of a country (CAT_GCR1), the country identifier (country), and household size (hhsize) are the most important features in predicting NME. 2.4 Discussion and conclusion In this atheoretical study, we attempt to investigate the comparative performance of multiple supervised ML techniques in predicting OME (and NME). The findings of our utilized ML techniques reveal that entrepreneurial activity can be predicted with a maximum out-of-bag overall accuracy of 91.2% for OME (and 96.5% for NME), without hyperparameter optimization. Therewith, ML techniques outperform the traditional multiple LR. When we strive for optimal hyperparameter values, the predictive accuracy of entrepreneurial activity can be increased. However, concerning the precision, our results also provide suggestive evidence that even the best-performing ML techniques—XGBoost and RF—are still modest at correctly predicting entrepreneurship activity for the hold-out sample. One possible reason for this is that most of the individual-level GEM features are dichotomous. Nevertheless, these findings still provide clear evidence that, contrary to earlier assumptions of scholars, it is possible to construct mathematical models related to entrepreneurship that can correctly distinguish between entrepreneurs and non- entrepreneurs, even with limited information. The surrogate model in this study suggests that knowing an entrepreneur is the most important feature in predicting the occurrence of OME, followed by different self-regulation mechanisms such as cognitions and personal traits. These findings are in line with existing literature, highlighting the paramount importance of social capital and psychological characteristics of individuals as antecedents of entrepreneurial pursuits (e.g., Chen et al., 1998; Liñán & Santos, 2007). 19 The results of the ML techniques reveal that the XGBoost estimator is superior in predicting both OME and NME. Intuitively, this could be because ML techniques such as XGBoost and RF are ensemble methods that are developed primarily to perform two-class prediction tasks (i.e., coded 0 and 1) with structured data, such as the GEM. Since both techniques are tree-based, they could also benefit from the many binary features in the data, when splitting into branches. In addition, XGBoost uses a method known as “boosting“ in which predictors are trained sequentially so that each model in the ensemble strives to minimize the errors of its predecessor. However, if we look at the precision estimates for NME prediction and compare them with the precision values for predicting OME, we can conclude that all ML techniques perform rather poorly at reliably predicting true positives for NME. These results provide suggestive evidence, that the features queried in the GEM are in favor of understanding OME instead of NME, and that crucial concepts predicting NME are presumably missing in the GEM data. Moreover, while we see that agent-centric features are of importance in predicting OME, geospatial factors and national entrepreneurial ecosystems in which individuals are embedded play a superordinate role in predicting NME. Our study mainly contributes to the embryonic yet burgeoning body of ML-based entrepreneurship literature (e.g., Antretter et al., 2019; Prüfer & Prüfer, 2020; Tan & Koh, 1996) that is focused on analyzing and deciphering the complex phenomenon of entrepreneurial activity using ML methods that can effectively model highly non-linear processes. Due to the comparative nature of our study and by highlighting the predictive performance superiority of XGBoost, we provide an initial benchmark for future ML-based studies that seek to further improve the reliable predictability of entrepreneurial activity. As ML has only recently emerged as a tool for econometricians and entrepreneurship researchers, we consider this study as the first of a series of further studies exploring the predictability of different types of entrepreneurship activities, including digital entrepreneurship, female entrepreneurship, social entrepreneurship, corporate entrepreneurial activity, etc., to address research questions such as the following: Which ML 20 technique is superior in predicting other forms of entrepreneurial activity? Which are the most important features in specific forms of entrepreneurial activity? However, research questions are not limited to positive outcomes. Of particular interest might also be investigations into the predictability of different kinds of business failures. Nonetheless, our study has several limitations. First, it is tempting to draw causal conclusions from the findings. However, with utmost clarity, we emphasize that due to the cross-sectional nature of the individual-level GEM data, direct causal inferences and conclusions require extremely careful scrutiny. Despite that, even correlations and non-linear associations can reveal useful underlying structures and mechanisms in the data. Second, although the GEM is currently one of the most comprehensive datasets on individual entrepreneurial activity, a substantial part of the features is binary, which reduces the possibility of performing more in-depth analyses. To draw even more detailed conclusions about the hidden patterns, most frequent interactions, and underlying relationships between the target and a feature, the use of input features at a higher scale level (i.e., continuous or Likert-based scale) would be beneficial. A higher data quality would also allow for providing various partial dependence plots (PDPs), disaggregated individual conditional expectation curves (ICEs), or Friedman’s H-statistic (Friedman & Popescu, 2008) on selected feature pairs to illustrate specific feature effects (Goldstein et al., 2015). These plots help to effectively visualize how a change in a single explanatory input feature changes the outcome prediction, i.e., (conditional) marginal feature effects and higher-order interactions (Friedman, 2001; Zhao & Hastie, 2021). Third, since we use a complete-case analysis approach, the prediction performance indices may be biased toward either over- or underestimation. However, this is only the case if the missing values are not missing completely at random (MCAR). Lastly, since our analyses only use the GEM data, our analyses may suffer from an omitted variable bias. Our findings and limitations also provide directions for future research. First, since we find that the developmental stage of a country (CAT_GCR1) is the 8th most important feature for predicting OME and the most important for NME prediction, country-specific factors seem to 21 play a significant role in entrepreneurial activity. To this end, future research can draw on different theories that take into account country-level factors that were unraveled through a contextual view of entrepreneurship (Welter, 2011), the institutional theory (North, 1990; Williamson, 2000), or the external enabler framework (Davidsson et al., 2020). These contextual factors are proven to be relevant contingencies, as they influence, for instance, human capital or cognitions and their effects on entrepreneurial pursuits (e.g., Autio & Acs, 2010; Boudreaux et al., 2019). These factors provide specific (situational) mechanisms that shape the entrepreneurial action-formation of individuals (Schade & Schuhmacher, 2022). In doing so, even high-dimensional non-linear relationships between contextual factors could be identified. The first vivid example of this fruitful path is provided by Jabeur et al. (2022), who forecasted and examined macro-level determinants of entrepreneurial opportunities. Second, while our study focuses on the comparison of an array of implemented ML techniques, we aim to inspire future research to dive deeper into and engage in the so-called “interpretable machine learning” (IML) or “explainable AI” (XAI) by using various model-agnostic explanation methods such as PDPs/ICEs, accumulated local effects (ALE) plots, local interpretable model-agnostic explanation (LIME) models or SHAPley values to unravel how models arrived at specific decisions and explain hidden, robust, and even anomalous patterns in the data (Choudhury et al., 2021; Molnar et al., 2020; Shepherd & Majchrzak, 2022). With these post hoc methods and higher data quality, future research can effectively infer and understand relationships between different features and their interactions on different target outcomes. This also allows future research to conduct meta-analytic reviews on independent ML studies that focused on similar prediction tasks with different data. Treating a validated ML prediction model as a stylized fact also opens avenues for theory development. Specifically, to explain and theoretically explicate patterns detected in the data, scholars can either use appropriate existing theories or engage in algorithm-supported induction for building entirely new theoretical approaches (Shrestha et al., 2021). Theoretical approaches that attempt to account for the uncovered patterns in the data can be used for classical hypothetico-deductive theory testing. 22 Lastly, since entrepreneurship research is constantly in flux, bringing to light new insights into the entrepreneurial phenomenon, we encourage future scholars to replicate our analysis for even more fine-grained data using sophisticated AI techniques. Declaration of competing interests The authors declare no conflicts of interest with respect to research, authorship, and publication of this article. Funding This research did not receive a grant from any funding agency in the public, commercial, or not-for-profit sectors. Acknowledgments We thank associate editor Andreas Kuckertz and the anonymous reviewer for their helpful comments and suggestions. Moreover, we appreciate the contribution from Marilyn A. Uy during the ACERE22 DC, as well as the comments from Per Davidsson and Dean A. Shephered on an earlier version and idea of this research project. 23 Chapter 3 3. Study 2 – Digital Infrastructure and Entrepreneurial Action-Formation: A Multilevel Study Coauthors: Monika C. Schuhmacher Relative share: 90% Status: Published in Journal of Business Venturing (FT50, ABS: 4, ABDC: A*, VHB-Jourqual: A) This chapter is available under: Schade, P. & Schuhmacher, M.C. (2022). Digital Infrastructure and Entrepreneurial Action- Formation: A Multilevel Study. Journal of Business Venturing, 37(5), 106232, https://doi.org/10.1016/j.jbusvent.2022.106232 A previous version of this chapter has been presented at: • 42nd Babson College Entrepreneurship Research Conference 2022 (BCERC), Waco/Texas, USA • Australian Center for Entrepreneurship Research Exchange (ACERE) Conference 2022, Melbourne/Virtual Edition, Australia • 24th G-Forum (2020) Interdisciplinary Conference on Entrepreneurship, Innovation and SMEs, Karlsruhe/Virtual Edition, Germany https://www.sciencedirect.com/science/article/abs/pii/S0883902622000441 24 Digital Infrastructure and Entrepreneurial Action-Formation: A Multilevel Study Abstract This study investigates how country-level digital infrastructure shapes the relationships between the action-formation mechanisms of socio-cognitive traits, i.e., entrepreneurial self-efficacy, fear of failure, and opportunity recognition, and entrepreneurial action. We amalgamate the agent- centric social cognitive theory with the external enabler framework and apply mechanism-based theorizing to explain how access-related mechanisms provided by digital infrastructure influence entrepreneurial action-formation. Based on a multilevel analysis of 344,265 individual-level observations from 46 countries and an additional robustness analysis of 391,119 individuals from 53 countries, we find that an individual’s proclivity to starting a new venture is contingent upon the level of the digital infrastructure of a country. The empirical results show that a country’s digital infrastructure is an external enabler that moderates the relationship between socio- cognitive traits and entrepreneurial action. JEL classification: L26, D91 Keywords: Entrepreneurial action, digital infrastructure, social cognitive theory, external enabler framework, mechanism-based theorizing, multilevel analysis 25 3.1 Introduction The World Economic Forum (2014) considers digital infrastructure as the backbone of digitalization and a prerequisite for venture creation and economic growth. Digital infrastructure refers to an unbounded, open, and evolving socio-technical system that includes technological and human components, networks, and systems (Hanseth & Lyytinen, 2010; Tilson et al., 2010). Hence, digital infrastructure refers to both applications of information and communication technologies and the associated infrastructure (Autio et al., 2018). Therefore, digital infrastructure is not limited to a distinct set of specific functions or restricted by strictly defined boundaries; rather, it is relational in nature (Tilson et al., 2010). In view of this relational property of digital infrastructure, we argue that digital infrastructure functions as a macro-level external enabler (EE) that provides specific mechanisms that shape individuals’ entrepreneurial action-formation (Davidsson et al., 2020; von Briel, Davidsson, & Recker, 2018). However, although public opinion and policies exalt digital infrastructure as a panacea for paving the way toward a digitally transformed economy and scholars stress its overriding relevance, there is a dearth of empirical research on whether and how a country’s digital infrastructure shapes individuals’ entrepreneurial action-formation as an EE. To shed light on entrepreneurship in the digital age, scholars adorn potentially effective theoretical lenses, such as the notion of digital affordances (Autio et al., 2018; Nambisan, 2017), the EE framework (Davidsson et al., 2020; von Briel, Davidsson, & Recker, 2018), or the digital entrepreneurial ecosystem perspective (Sussan & Acs, 2017). Although some of the concepts are qua design predominantly venture-level theoretical frameworks, they also allow for theorizing the enabling role of digital infrastructure in individual entrepreneurial action-formation. Extant research on individual entrepreneurship provides ample empirical evidence that socio- cognitive traits, namely, entrepreneurial self-efficacy, fear of failure, and opportunity recognition, are essential mechanisms of entrepreneurial action (e.g., Busenitz & Barney, 1997; Shane & Venkataraman, 2000; Zhao et al., 2005). However, several researchers demonstrate that 26 entrepreneurial action-formation also depends on the proximate and distal macro contexts in which individuals at the micro-level operate. These contexts are external to the respective focal phenomenon (i.e., entrepreneurial action) and enable or hinder entrepreneurship (Jack & Anderson, 2002; Spigel, 2017; Welter, 2011). Country-level contexts that have so far been investigated as moderators for the relationship between individual-level socio-cognitive traits and various entrepreneurial pursuits are social (Schmutzler et al., 2018), institutional, economic, and political (e.g., Autio & Acs, 2010; Boudreaux et al., 2019), or cultural (Stephan & Pathak, 2016; Wennberg et al., 2013) conditions, situations, circumstances, and environments. Although such research provides considerable information about traditional macro-level factors, limited effort has been made in theorizing the role of contemporary factors, such as digital technologies and infrastructure, in shaping entrepreneurial action (Nambisan, 2017). This prevalent line of contextual thinking is consistent with the EE framework proposed by Davidsson (2015) and refined by Davidsson et al. (2020). In this framework, an EE refers to the aggregate macro-level circumstance that can shape individual entrepreneurial action-formation and plays a significant role in eliciting or enabling entrepreneurial endeavors. Thus, the framework also considers the crucial role of the individual (Kimjeon & Davidsson, 2021). Furthermore, the EE framework provides mechanisms that specify the benefits derived from external contexts, which can be strategically used for personal purposes regarding entrepreneurship (Davidsson et al., 2020; Kimjeon & Davidsson, 2021). Thus far, the EE framework has primarily been applied to different factors outside the scope of action by individuals, for example, the high-speed railway expansion in China (Chen et al., 2020), investments in physical infrastructure in the United States (Bennett, 2019), blockchain in the global music industry (Chalmers et al., 2021), and digital technologies in the IT sector (von Briel, Davidsson, & Recker, 2018), among others (e.g., Browder et al., 2019; Davidsson et al., 2021; Frederiks et al., 2019). We propose that digital infrastructure serves as an EE to support entrepreneurial action-formation (Davidsson, 2015; Nambisan, 2017). Hence, we theorize that 27 digital infrastructure provides mechanisms that shape the impact of individual socio-cognitive traits on entrepreneurial action. On this basis, the overarching research question is as follows: How does a country’s digital infrastructure shape the entrepreneurial action-formation of individuals? To answer this question, we used a large-scale, cross-sectional dataset comprising 344,265 individual-level observations from 46 countries—and for robustness analysis, a dataset of 391,119 individuals from 53 countries—and employed logistic multilevel modeling. Overall, our study contributes to knowledge accumulation in the following ways. First, our study represents the first empirical investigation on whether and how a country’s digital infrastructure fosters individual entrepreneurial action-formation, thereby responding to numerous calls in the literature to examine entrepreneurship as a multilevel phenomenon (e.g., Busenitz et al., 2003; Terjesen et al., 2016). Second, our study contributes to the contextual entrepreneurship literature. Specifically, as we consider how digital infrastructure—a hitherto unconsidered technological contextual factor— shapes entrepreneurial action-formation, we add a novel technological dimension to the existing classification of contexts proposed by (Welter, 2011). Third, we combine the agent-centric social cognitive theory (SCT) (Bandura, 1986; Sherman et al., 2015; Wood & Bandura, 1989) with the EE framework (Davidsson et al., 2020). Because SCT is rather unrefined and lacks a theoretical explanation of how contextual factors shape entrepreneurial action-formation at the micro-level, we augment SCT with the EE framework, which allows us to reason on how a specific contextual factor, in terms of EE, develops through specific mechanisms. With this theoretical approach, which allows for mechanism-based theorizing, we demonstrate the theoretical usefulness of the EE framework in elaborating on an existing agent-centric theory, thereby responding to calls from entrepreneurship scholars to merge the EE framework with agent-based theories (Davidsson et al., 2020; Kimjeon & Davidsson, 2021). Our results show that variations in country-level digital infrastructure play a significant role in the relationship between agential cognitions and entrepreneurial action at the individual 28 level and, thus, provide an empirically substantiated theoretical elaboration of the emergent EE theory (Fisher & Aguinis, 2017). Specifically, we are the first to show that the EE framework is not only of paramount importance for theorizing at the venture level (Davidsson et al., 2020; von Briel, Davidsson, & Recker, 2018) but also an appropriate theoretical basis for explaining the role of EEs in the individual-level entrepreneurial phenomenon. 3.2 Theoretical framework 3.2.1 Social cognitive theory According to SCT, the effects of an individual’s predispositions, such as cognitions, emotions, etc., are determined and shaped by both individual-level characteristics and the external environment in which individuals find themselves (Bandura, 2015; Mischel, 2004; Wood & Bandura, 1989). Grounded in an agentic perspective, SCT states that environmental and socio- structural factors operate through psychological or self-regulatory mechanisms of the self-system to produce behavioral outcomes (Bandura, 2001). Moreover, according to Lent et al. (1994), external contexts exert either direct or indirect influence by moderating the individuals’ socio- cognitive traits–behavior relationship. Thus, SCT clarifies the basic mechanisms of agency that govern individuals’ behavior and provides a useful framework to understand underlying mechanisms through which individual dispositions lead to behavioral activity (Hmieleski & Baron, 2009; Ng & Lucianetti, 2016). As SCT is centered on the agentic individual’s embeddedness in their environment, the theory also considers the multilevel perspective necessary to understand complex behavioral processes (Hitt et al., 2007). In fact, Jack and Anderson (2002) and (Welter, 2011) demand that research takes into consideration the individual’s external context at a higher macro-level to improve the understanding of the entrepreneurship phenomenon at a lower level of analysis. Therefore, we focus on how the external environment at the macro-level—digital infrastructure, in this case— indirectly influences the socio-cognitive traits–entrepreneurial activity relationship at the micro-level. 29 3.2.2 External enabler framework and venture creation According to the EE framework, an EE is an aggregate macro-level circumstance that can shape the venture creation process and play a significant role in enabling entrepreneurial endeavors. The EE framework classifies different types of EEs that are external to the focal entrepreneurial phenomenon, based on their origins, such as technological, socio-cultural, macroeconomic, regulatory, demographic, and political factors, and the natural environment (Davidsson, 2015; Davidsson et al., 2020; Mair & Marti, 2009). In understanding the EE framework, a mechanism is a relational construct that establishes a link between the external environment and the entrepreneurial agent (Davidsson et al., 2020). Hence, EE mechanisms provide a theoretical means of explaining in detail how external macro- level circumstances and the micro-level entrepreneurial phenomenon are connected. However, whether an EE provides particular mechanisms depends on both the properties of the enabler and the entrepreneurial agent (Davidsson et al., 2020). As SCT lacks specific theoretical explanations on how external factors shape entrepreneurial action-formation, we combine SCT with the EE framework and reason about the underlying mechanisms provided by the EE—here, digital infrastructure—to deduce our hypotheses and further motivate our multilevel approach. 3.3 Mechanism-based theorizing: Hypotheses development Based on the perspective of critical realism, mechanisms are causal structures that generate observable effects and events (Bhaskar, 1997; Henfridsson & Bygstad, 2013; Merton, 1968), by shaping relationships in a given set of elements (Tilly, 2001). Researchers categorize mechanisms as “situational, action-formation, or transformational mechanisms” depending on the level at which the mechanisms function (Hedström & Swedberg, 1998, pp. 22–23; Kim et al., 2016). First, action-formation mechanisms exclusively operate at the micro-to-micro-level (1-1). At this solely individual-centric micro-level, a plurality of psychological and socio-psychological mechanisms operate and provide a rationale for how agential cognitions, personal traits, beliefs, and motivations generate actions (Hedström & Swedberg, 1998). The action-formation mechanism 30 perspective follows the psychological notion of human agency in SCT; it suggests that human action-formation arises from psychological or self-regulatory mechanisms. These self-regulatory mechanisms, in turn, emerge from lower-order cognitive mechanisms, specifically socio- cognitive traits (Bandura, 1989). Second, transformational mechanisms influence or generate macro-level outcomes bottom-up from the micro-level to the macro-level (1-2) (Hedström & Swedberg, 1998; Kim et al., 2016). Third, situational mechanisms cover contextual factors such as technological, political, socio- cultural, and environmental factors that occur top-down and rest at the interacting imbrication between the macro- and micro-levels (2-1). These situational mechanisms determine and shape existing socio-cognitive traits of individual actors at the micro-level (Coleman, 1986, 1994; Hedström & Ylikoski, 2010; Kim et al., 2016; Sarason et al., 2006). Hence, situational mechanisms are a set of plausible hypotheses that can be explanations of some phenomena, whereby the explanation takes the form of interactions between individuals at the micro-level and a circumstance at the macro-level (Coleman, 1994; Schelling, 1998). We assert that the impact of digital infrastructure on entrepreneurial action-formation refers to such situational mechanisms. The understanding of situational mechanisms fully aligns with the notion of the underlying mechanisms of the EE framework, describing “the higher-level relationship between the emergence of new digital technologies as external enablers (i.e. cause) and venture creation activity in a sector (i.e. the effect)” (von Briel, Davidsson, & Recker, 2018, p. 51). Furthermore, this view of situational mechanisms is particularly emphasized in innovation research (Hedström & Wennberg, 2017) and entrepreneurship research concerning entrepreneurial action-formation (Johnson & Schaltegger, 2020; Kim et al., 2016; von Briel, Davidsson, & Recker, 2018). Since the understanding of situational and EE mechanisms are identical, we will refer to both in our theorizing as EE mechanisms. Figure 4 summarizes the theoretical model that we will hypothesize in the following subsections. 31 Figure 4: Theoretical model 3.3.1 Action-formation mechanisms at the individual level and baseline hypotheses According to SCT (e.g., Bandura, 1986; Wood & Bandura, 1989), self-regulation explains how humans feel, think, and behave. Since both entrepreneurial activity and individuals’ socio- cognitive traits are complex phenomena, we adopt the “cognitive approach to entrepreneurship” (Baron, 2004; Mitchell et al., 2002; Shaver & Scott, 1992). Specifically, we focus on the most prevalent goal-directed entrepreneurial socio-cognitive traits at the individual micro-to-micro- level (1-1) that are representations of the external environment captured through individuals’ mental processes (Krueger, 2003) and are generative mechanisms for entrepreneurial action- formation (see e.g., Baron, 1998), namely, entrepreneurial self-efficacy (e.g., Bandura, 1982; McGee et al., 2009), fear of failure (e.g., Caliendo et al., 2009; Langowitz & Minniti, 2007), and opportunity recognition (e.g., Ardichvili et al., 2003; Baron, 2006). An individual’s perceived self-efficacy refers to the self-assessment of one’s capabilities and skills (reflecting the innermost thoughts) to create and run a new business venture (McGee et al., 2009; Zhao et al., 2005). Previous studies provide empirical evidence that (entrepreneurial) self- efficacy, as developed by Bandura (1982), is a precursor of intentions (e.g., Chen et al., 1998; Macro-level 2 (Country) Micro-level 1 (Individual) External enabler Digital infrastructure Action Entrepreneurial activity Socio-cognitive traits - Entrepreneurial self- efficacy - Fear of failure - Opportunity recognition Action-formation mechanisms (1-1) (H1a, H1b, H1c) EE mechanisms (2-1) (H2, H3, H4) 32 Zhao et al., 2005) and the best ex-ante predictor of behavior (Armitage & Conner, 2001; Bagozzi et al., 1989). Aside from favorable traits, there are also socio-cognitive traits that can stifle entrepreneurial action-formation. Under SCT, this is the case when individuals surmise that an intended goal is difficult to achieve and the likelihood of failure is omnipresent (Wood & Bandura, 1989). Venture creation epitomizes a perilous endeavor associated with high uncertainty and risk-taking; as such, fear of failure plays a pivotal role in entrepreneurial action-formation (Caliendo et al., 2009). An amplification of fear of failure negatively affects the probability of individual entrepreneurial action (Arenius & Minniti, 2005; Langowitz & Minniti, 2007; Wagner, 2007). SCT also highlights the human capacity for self-motivation and self-direction by creating goals that serve as motivators and guides for action (Bandura, 1988). In this regard, entrepreneurial scholars have identified opportunity recognition as one of the most fundamental and distinctive personal traits of entrepreneurial action-formation (Kirzner, 1979; Venkataraman, 1997). Individuals who possess traits of opportunity recognition can translate symbolic conceptions into an appropriate course of action (Wood & Bandura, 1989) and may make individuals better at overcoming the inherent opacity of EE mechanisms and foreseeing their benefits (Davidsson et al., 2020; Grégoire & Shepherd, 2012). This idiosyncratic personal trait is described as an entrepreneur’s evolving vision that is malleable and becomes more concrete with the progress of entrepreneurial action-formation (Berglund et al., 2020; Davidsson, 2021). Therewith, entrepreneurial action originates from the subjective perception that the introduction of a new product or service is feasible and worth pursuing (Ardichvili et al., 2003; McMullen & Shepherd, 2006). In summation, on an individual micro-to-micro-level (1-1), SCT suggests that the individual socio-cognitive traits, i.e., entrepreneurial self-efficacy, fear of failure, and opportunity recognition, are essential action-formation mechanisms. Hence, we propose the following baseline hypotheses: 33 Hypothesis 1a (H1a): An individual’s perception of entrepreneurial self-efficacy is positively associated with entrepreneurial activity. Hypothesis 1b (H1b): An individual’s fear of failure is negatively associated with entrepreneurial activity. Hypothesis 1c (H1c): An individual’s opportunity recognition is positively associated with entrepreneurial activity. 3.3.2 EE mechanisms of digital infrastructure and moderation hypotheses To understand the role of digital infrastructure in entrepreneurial action-formation, we identify specific EE mechanisms by drawing upon the ontological properties of digital technologies (von Briel, Davidsson, & Recker, 2018). According to von Briel, Davidsson, and Recker (2018), digital technologies can be divided into two ambivalent properties—specificity and relationality. Specificity refers to the set of possible actions and interactions that can be performed with a technology (DeSanctis & Poole, 1994; von Briel, Davidsson, & Recker, 2018). In contrast, relationality refers to the set of relationships with other actors or end-users that have access to the same technology (Kallinikos et al., 2013; von Briel, Davidsson, & Recker, 2018). In light of this ontology, digital infrastructure is not characterized by specificity but is inherently relational and provides enhanced accessibility to different location-independent resources and markets through direct interactions with geographically dispersed audiences and end-users (Autio et al., 2018; Nambisan, 2017; Tilson et al., 2010; Wasko, 2005). Thus, we argue that digital infrastructure mainly facilitates access-related EE mechanisms (Bruton et al., 2015; Tilson et al., 2010). Drawing on the EE framework, we propose that digital infrastructure particularly provides resource access and market access mechanisms (Kimjeon & Davidsson, 2021; Majchrzak et al., 2013; Podolny, 2001; von Briel, Davidsson, & Recker, 2018). While resource access mechanisms reflect “improved access for the [individual] to a previously existing type of resource,” market access mechanisms are defined as “improved access for the focal [individual] to a previously existing market” (Kimjeon & Davidsson, 2021, p. 4). These access- related EE mechanisms are especially beneficial to those entities that pursue a specific goal 34 (Davidsson et al., 2020). In other words, digital infrastructure enables access to pre-existing resources and markets, thereby facilitating entrepreneurial action-formation of individuals with similar socio-cognitive traits. 3.3.2.1 Moderation of the self-efficacy action-formation mechanism SCT suggests that the extent to which entrepreneurial self-efficacy fosters entrepreneurial action-formation depends on external contexts (Bandura, 1986; Wood & Bandura, 1989). Based on SCT, we theorize that digital infrastructure, with the accompanying resource access and market access mechanisms, is an EE that moderates the effect of entrepreneurial self-efficacy on entrepreneurial action. Specifically, we argue that this positive relationship is stronger when countries have a higher level of digital infrastructure for several reasons. First, countries with a high level of digital infrastructure provide individuals with a market access mechanism, which refers to improved access to existing markets (e.g., capital, product, labor, credit, customer, etc.) (von Briel, Davidsson, & Recker, 2018). Such market access enables individuals to exchange, buy, or sell services and goods with various customers and vendors (Shelton & Minniti, 2018). Therewith, a market access mechanism enables access to global markets and offers the potential to sell to lucrative international customers via easily accessible online marketplaces and platforms. Simultaneously, market access mechanisms reduce distance- related issues, allow leveraging economies of scale, and lower transaction costs. These benefits enhance the probability that personal effort will lead to successful entrepreneurial performance (Chen et al., 1998). As digital infrastructure, and the associated market access mechanism, is considered an institutional arrangement (Leendertse et al., 2021), it also determines the relative rewards and expected returns from engaging in entrepreneurial activity (Baumol, 1996). Therefore, individuals having similar entrepreneurial self-efficacy beliefs but lacking market access have fewer options to offer products or services in the market, thereby lowering expected returns and enhancing incentives to pursue alternative career options, rather than entrepreneurship (Baumol, 1996). In contrast, in countries that provide market access, i.e., where digital 35 infrastructure is high, existing markets are easily accessible and individuals with similar entrepreneurial self-efficacy are likely to not discount the marginal value of future profits. This, in turn, enhances the impact of individuals prone to mobilizing motivation, abilities, and skills in their effort toward entrepreneurial action (Chen et al., 1998; Wood et al., 2016). Second, entrepreneurial action as an act of innovation describes an entrepreneur as an individual that utilizes existing resources necessary to produce and offer a product or service that fills a market gap (Drucker, 1985; Leibenstein, 1968). Thus, resources refer to access to, the possession of, and usage of specific human, financial, and other resources and assets necessary for entrepreneurial action-formation (Bull & Willard, 1993; Mitchell et al., 2000). Thus, entrepreneurial action-formation is the result of self-assessment in which individuals evaluate the perceived availability of resources and the constraints to task performance. The consideration of perceived personal resources, such as capabilities and skills (Gist & Mitchell, 1992), the type and amount of specific resources (i.e., tangible and intangible resources), as well as perceptions of external resource availability required to complete different tasks has been shown to determine and shape entrepreneurial action-formation (Bandura, 1988; Gist & Mitchell, 1992; Krueger, 1993). In countries with a low level of digital infrastructure resources relevant to the entrepreneurial process are hardly accessible. This, in turn, reduces the effect of individuals mobilizing personal resources on entrepreneurial action, thereby lowering the likelihood of becoming entrepreneurially active. Thus, individuals with identical entrepreneurial self-efficacy beliefs will ultimately be less likely to engage in entrepreneurial action in countries without resource access mechanisms in place, i.e., where digital infrastructure is low. In contrast, when resource access is given, i.e., digital infrastructure is high, individuals’ self-efficacy allows them to virtually access more resources to address deficiencies that prevent them from starting a new business venture. In summation, we argue that a high (low) level of digital infrastructure, i.e., countries that (do not) provide resource access and market access mechanisms, reinforces (attenuates) the impact of 36 entrepreneurial self-efficacy on entrepreneurial action. Thus, we propose the following hypothesis: Hypothesis 2 (H2): The positive relationship between entrepreneurial self-efficacy and entrepreneurial activity is moderated by digital infrastructure, such that this relationship will be stronger when digital infrastructure is high than when it is low. 3.3.2.2 Moderation of the fear of failure as an action-formation mechanism Within SCT, the fear of failure negatively affects the probability of individual entrepreneurial activity (Arenius & Minniti, 2005; Langowitz & Minniti, 2007; Wagner, 2007). We theorize that digital infrastructure weakens the negative effect of the fear of failure on entrepreneurial action for some reasons. First, different obstacles pose serious threats to individuals in the entrepreneurial action-formation process (Cacciotti et al., 2016). These obstacles amplify individuals’ fear of failure, for instance, the perception of tangible and intangible resource availability (e.g., Krueger, 2000). The resource access mechanism provided by digital infrastructure offers improved access to existing resources such as human and social capital, which are important conduits of information and resources for resolving uncertainties (Birley, 1985; Engel et al., 2017). For example, individuals can engage in socially persuasive communication and draw on rich informational expertise from various actors and entrepreneurs that are not in close proximity (Kuhn & Galloway, 2015; Nambisan, 2017; Nambisan & Baron, 2007). Hence, we expect individuals with a similar fear of failure, which dissuades them from becoming entrepreneurially active, to virtually receive assistance on a global scale when digital infrastructure is high, thereby weakening the relationship between fear of failure and entrepreneurial activity. Second, founders generally finance their initial entrepreneurial activities with their own money and “love money” from family and friends (e.g., Bygrave et al., 2003). People showing fear of failure will very likely turn away from making use of such financial resources, leading to fewer entrepreneurial actions. Market access allows for more external funding sources. The market access mechanism through digital infrastructure permits mobilizing external funding directly from 37 demand-side backers through social online networks and crowd-funding platforms, without any distance-related friction (Bruton et al., 2015; Eiteneyer et al., 2019). Hence, the market access mechanism permits individuals with similar failure concerns to mitigate the influence of fears stemming from financial security or the lack of ability to finance an entrepreneurial venture (Cacciotti et al., 2016). Third, entrepreneurs typically start with a niche strategy; competing at the fringe of the market always carries the risk that founders fear—that the target market is too small for entrepreneurial survival and success (Cacciotti et al., 2016). Digital infrastructure provides a market access mechanism and, therewith, direct access to different existing (over) regional, independent markets, and customers that go far beyond those facilitated by physical infrastructure (see e.g., Chen et al., 2020). Thus, in countries that provide a market access mechanism through a high- level digital infrastructure, the relation between fear of failure and entrepreneurial activity will be weaker compared to countries without such a market access mechanism (i.e., low-level digital infrastructure). In summation, countries that provide both resource access and market access mechanisms (i.e. high in digital infrastructure) bequeath individuals a similar fear of failure such that the latter can benefit from the provided mechanisms and the negative effect of fear of failure, which prevents individuals from becoming entrepreneurially active, is weakened. Thus, we formulate the following hypothesis: Hypothesis 3 (H3): The negative relationship between fear of failure and entrepreneurial activity is moderated by digital infrastructure, such that this relationship will be weaker when digital infrastructure is high than when it is low. 3.3.2.3 Moderation of the opportunity recognition action-formation mechanism Possessing the socio-cognitive trait of opportunity recognition is an important antecedent that positively increases the likelihood of entrepreneurial action (e.g., Kirzner, 1979; Wood & Bandura, 1989). We theorize that digital infrastructure strengthens the positive effect of opportunity recognition on entrepreneurial action for the following reasons. First, in countries 38 with a high level of digital infrastructure and, thus, market access mechanisms, individuals who possess the socio-cognitive trait of opportunity recognition are able to evaluate new products, services, and/or business model ideas and meet market needs, thereby enhancing the impact of opportunity recognition on entrepreneurial activity. For instance, individuals can access already existing markets, customers, and relevant competitors through digital forums, discussion boards, and other platforms, or (social) networks. With market access, individuals can contact and converse with customers—i.e., demand-side narratives (Nambisan and Zahra, 2016)—that serve as relevant sources in driving the process of evaluating and modifying potential manifestations of envisioned products and services, as well as assessing their demand and use (Davidsson, 2021). Thus, individuals who affirmatively indicate that they possess the socio-cognitive trait of opportunity recognition benefit from the market access mechanisms provided by digital infrastructure through the enhanced ability to determine the viability of a new venture and the value of entrepreneurial action, thereby increasing the likelihood of entrepreneurial activity. Second, in transiting from the subjective perception of an idea to a viable, operational business, individuals who possess the socio-cognitive trait of opportunity recognition benefit from the resource access mechanism provided by digital infrastructure. In countries with a high level of digital infrastructure, the resource access mechanism enables individuals to identify and acquire relevant but missing tangible and intangible resources (Bhagavatula et al., 2010; Haynie et al., 2009; Nambisan & Zahra, 2016), which condition the feasibility and viability of the entrepreneurial endeavor.