Big Data in Leadership Studies: Automated Machine Learning Model to Predict Preferred Leader Behavior Across Cultures

Abstract

With global leadership as the new norm, discussion about followers’ preferred leader behaviors across cultures is growing in significance. This study proposes a comprehensive predictive model to explore significant preferred leadership factors, drawn from the Leader Behavior Description Questionnaire (LBDQXII), across cultures using automated machine learning (AML). We offer a robust empirical measurement of culturally contingent leader behavior and entrepreneurship behaviors and provide a tool for assessing the cultural predictors of preferred leader behavior to minimize predictive errors, explore patterns in the data and make predictions in an empirically robust way. Hence, our approach fills a gap in the literature relating to applications of AML in leadership studies and contributes a novel empirical method to better predict leadership preferences. Cultural indicators from Global Leadership and Organizational Behavior (GLOBE) predict the likelihood of the preferred leader behaviors of “Role Assumption”, “Production Emphasis” and “Initiation of Structure”. Hofstede’s Long-Term/Short-Term Orientation is the most critical predictor of preferences for “Tolerance of Uncertainty” and “Initiation of Structure”, whereas the value of restraint impacts the likelihood of preferring leaders with skills in “Integration” and “Consideration”. Significant entrepreneurial values indicators have a significant impact on preferences for leaders focused on “Initiation of Structure”, “Production Emphasis” and “Predictive Accuracy”. Findings also support earlier studies that reveal age and gender significantly impact our preferences for specific leader behaviors. We discuss and offer conclusions to support our findings that foster development of global business managers and practitioners.

1. Introduction and Literature Review

Society and companies alike are equally dependent on great leadership. Previous research and common sensemaking have long understood that there seem to be strong connections between leadership and societal norms and values [1]. Additionally, scholars do support and envision the idea that the social world is constrained by the normative and factual order: social situations and practices are embedded within cultural practices and values, and thus factual practices (culture as it is) are perceived based on normative values (culture as it should be) [2]. Consequentially, societal members develop and foster cognitive templates for preferred leader behaviors and traits that may reward favorable outcomes [3]. Thus, scholars, practitioners and society would benefit from understanding what drives and leads to favor for certain leadership behaviors that are preferred in society based on cultural practices and cultural values. Additionally, there is limited understanding of the potential of big data analyses and causal and predictive modelling of the existing leadership concepts.

Leadership studies have commonly approached research questions from a leader-centric view through implicit characteristics. To further understand what drives and leads to favor for certain leadership behaviors in society, the Leader Behavior Description Questionnaire XII [4] was created; the LBDQXII employs a follower-centric orientation, still underrepresented in leadership research [5]. A follower-centric view focuses on a follower’s beliefs regarding descriptions of ideal leader behavior. Hofstede’s and GLOBE project’s conceptualization of culture and leadership across cultures [6,7] are also consistently reflected in the LBDQXII.

Northouse (2013) argues that the LBDQXII questionnaire is the most used instrument in leadership research [8]. This resulted in decades of use, and a considerable amount of evidence regarding LBDQXII´s test-retest reliability, construct validity and use in cross-cultural settings [9,10,11,12].

This study’s objective is to create comprehensive prediction models to explore what drives and leads to favor for certain leadership behaviors across cultures: By creating an extensive dataset with indexes drawn from major established cultural indicators together with an entrepreneurship index using automated machine learning (AML), we offer a robust method to minimize predictive errors, explore patterns in the large amount of data and make predictions in an empirically robust way. To our knowledge, this is the first study to offer contributions towards research and practice by its robust empirical measurement of culturally contingent leader- and entrepreneurship behavior through a novel method. It is also the first attempted study to provide a tool for assessing the cultural predictors of preferred leader behavior by calibrated indexes. Compared to traditional linear and additive models, AML offers an advantage by simultaneous use of vast data characteristics and predictive capacity to evaluate multiple models [13]. AML is useful for assessing the cultural and entrepreneurial predictors of preferred leader behavior to minimize predictive errors, explore patterns in the data and make predictions in an empirically robust way [14]. Additionally, AML allows researchers and practitioners to explore patterns and theory testing using big data, benefiting certain fields and topics that have yet to establish full understanding of complex and exact relationships between predictors and the outcome target variables [13,15].

The paper is organized as follows. The following section offers a review of the literature on cultural and leadership constructs in this study as well as of the Global Entrepreneurship Monitor (GEM) [16] literature we apply. Building on this literature, we develop three main hypotheses. We then present methodological aspects of the sampling procedure and an explanation of the research design and methodological process. The subsequent section reports the main results of the study and discusses these results along with the methodological and management implications. The last section notes the limitations of the study and offers suggestions for future research.

1.1. LBDQXII: Leader Behavior Description Questionnaire

Preferred leader behaviors were operationalized using the Leader Behavior Description Questionnaire XII [4]. Based on the work of Hemphill and Coons [17], Stogdill [4] developed an assessment of 12 leader behavior dimensions with the LBDQXII, consisting of 100 items with Likert-type response categories to measure these leadership constructs: When estimating a follower’s preferred leadership style, LBDQXII dimensions of leadership are measured and compared across cultures (see

Table 1. Overview of LBDQXII dimensions.

LBDQXII Dimension	Definition
Representation	measures a follower’s preference for a manager who speaks clearly as the representative of the group.
Demand reconciliation	reflects a follower’s preference for a manager who explicitly reconciles conflicting demands and reduces disorder to the system.
Tolerance of uncertainty	measures a follower’s preference for a manager who can tolerate uncertainty and postponement without anxiety or getting upset.
Persuasiveness	measures to what extent a follower prefers a manager who uses persuasion and argument effectively and exhibits strong convictions.
Initiation of structure	measures to what degree a follower prefers a manager who clearly defines his or her own role and lets followers know what is expected.
Tolerance of freedom	reflects to what extent a follower prefers a manager who allows followers a scope for initiative, decision and action.
Role assumption	measures to what degree a follower prefers a manager who actively exercises the leadership role rather than surrendering leadership to others.
Consideration	depicts to what extent a follower prefers a manager who regards the comfort, well-being, status and contributions of followers.
Production emphasis	measures to what degree a follower prefers a manager who applies pressure for productive output.
Predictive accuracy	measures to what extent a follower prefers a manager who exhibits foresight and ability to predict outcomes accurately.
Integration	measures to what degree a follower prefers a manager who maintains a closely-knit organization and resolves inter-member conflicts.
Superior Orientation	measures to what extent a follower prefers a manager who maintains cordial relations with superiors, has influence with them and is striving for higher status.

Source: Stogdill, Goode and Day [4].

for overview of definitions).

Most approaches to the study of leadership are leader-centric and define implicit characteristics. The LBDQXII questionnaire has more than 50 years of continuous, extensive use [18], and a considerable amount of research supports its test-retest reliability, construct validity and use in cross-cultural settings with accepted reliability and validity [4,9,11,12,19]. We use the LBDQXII dataset to investigate which cultural factors may predict preferred leadership across cultures (see Ljubica, Littrell, Warner-Søderholm and Minelgaite [11] for more details).

1.2. Measuring Culture with Hofstede Indexes

Geert Hofstede’s research has profoundly shaped the understanding of cultural differences in the context of global business and management [6]. His cultural dimensions provide a valuable framework for analyzing how national culture impacts workplace values and behaviors. Despite criticisms, such as for its overgeneralization, Hofstede’s model remains a cornerstone in the field of cross-cultural studies, offering essential insights for organizations operating in an increasingly globalized world. The Six Cultural Dimensions from Hofstede applied in this study are: Power Distance Index (PDI), or the extent to which less powerful members of organizations and institutions accept and expect that power is distributed unequally; Individualism vs. Collectivism (IDV), or the degree to which people prefer being left alone to look after themselves or want to remain closely integrated into groups, usually around the family; Masculinity vs. Femininity (MAS), or the distribution of emotional roles between the genders and the value placed on competitiveness versus caring; Uncertainty Avoidance Index (UAI); or the extent to which members of a culture feel threatened by ambiguous or unknown situations; Long-Term vs. Short-Term Orientation (LTO), or the degree to which a culture emphasizes long-term planning, perseverance and thrift versus short-term gratification and respect for tradition; and Indulgence vs. Restraint (IVR), or the degree to which cultures encourage free expression and gratification of human desires or suppress gratification of needs and regulate it by means of strict social norms.

Past studies have explored whether Hofstede’s model can impact leadership behaviors. However, there are no studies regarding whether these cultural dimensions predict the likelihood of preferred leadership behaviors. Hence, in line with findings from research on the cultural dimensions of workplace values and behaviors, we hypothesize that:

Hypothesis 1 (H1).

Cultural values of power distance, risk avoidance, masculinity, individualism, time and restraint significantly impact a follower’s preferred leadership behavior.

1.3. Project GLOBE Measurement Items

The Global Leadership and Organizational Behavior Effectiveness (GLOBE) project provides a methodology to quantitatively measure leadership. More specifically, this on-going research project is a multi-phase, multi-method program exploring the interrelationships between leadership, organizational culture and societal culture [20]. The theoretical lens of the GLOBE research project is based on synthesis of implicit leadership and motivation theory [21,22], cultural value belief theory [6,23] and structural contingency theory pertaining to organizational form and effectiveness [24,25].

More than 200 researchers from 62 countries representing all major regions of the world collected data and engaged in analyses of the responses of more than 17,000 mid-level managers [26].

In the GLOBE project, culture is defined as shared beliefs, values, interpretations and meanings of significant events, stemming from common experiences among members of the society and as passed on for generations [27]. Evaluation of the GLOBE project’s research in relation to Hofstede’s seminal work leads to the conclusion that, despite the use of different terms to identify cultural dimensions, the cultural dimensions identified in GLOBE research are related conceptually and correlate empirically to Hofstede’s formulated dimensions [28]. One main difference, however, lies in the calibration of GLOBE research, as it offers nine dimensions, leading to a more comprehensive view of societal culture than Hofstede’s model.

Project GLOBE measures each of the 9 variables as both practices and values using “as is” and “should be” scores (see

Table 2. Overview of GLOBE dimensions.

Dimension	Definition
Performance Orientation	The degree to which a collective encourages and rewards group members for performance improvement and excellence.
Future Orientation	The extent to which individuals engage in future-oriented behaviors such as delaying gratification and planning.
Gender Egalitarianism	The degree to which a collective minimizes gender inequality.
Assertiveness	The degree to which individuals are assertive, confrontational and aggressive in their relationships with others.
Collectivism	(Institutional) The degree to which institutional practices encourage and reward collective distribution of resources and collective action. (Group) The degree to which individuals express pride and cohesiveness in their organizations / families.
Power Distance	The degree to which members of a collective expect power to be distributed equally.
Humane Orientation	The degree to which a collective encourages and rewards individuals for being fair, altruistic, generous, caring and kind to others.
Uncertainty Avoidance	The extent to which a collective relies on social norms, rules and procedures to alleviate unpredictability of future events.

Source: House [1].

below). The “as is” scores represent the practices that are observed and reported within a society—how things currently are. On the other hand, the “should be” scores reflect the values of a society—how people believe things ought to be.

It could be expected that followers’ preferred leader behaviors can be impacted by the GLOBE definitions of culture. Specifically, the differences in cultural practices versus cultural values. The predictive nature of preferred leader behavior still merits exploration, even though some research finds some relationship between the GLOBE dimensions and cultural outcomes. Therefore, we predict that:

Hypothesis 2 (H2).

Global leadership and organizational effectiveness values across cultures significantly impact followers’ preferred leader behaviors.

1.4. The Global Entrepreneurship Monitor

Entrepreneurship is clearly seen as an issue directly connected with leadership behaviors [29]; hence, we apply the Global Entrepreneurship Monitor (GEM) [16] data. GEM is an annual research initiative that assesses entrepreneurial activity, attitudes and aspirations across various economies of working age adults [16]. Since its earliest initiative in 1999, several countries have participated in the world’s most authoritative comparative study of entrepreneurial activity in the general adult population [30]. GEM provides insights into entrepreneurship’s significance and collects data on entrepreneurial behavior measuring how attitudes influence entrepreneurial activity. GEM tracks and compares “characteristics, motivations and ambitions of entrepreneurs, and the attitudes societies have toward this activity” [31]. The GEM database (as well as GLOBE) are useful and robust in the formation of empirical connections between factual and normative culture and entrepreneurship at the country level [2].

We use key GEM indicators from its database (for an overview of indicators, see Kelley, Singer and Herrington [31]) to supplement the indicators from Hofstede and GLOBE, complementing past studies of culture and entrepreneurship [2]. In addition to understanding what increases the likelihood of followers’ preferred leader behaviors with cultural dimensions, we also include data on entrepreneurial behavior and engagement attitudes: followers who have beliefs and attitudes that lean towards certain entrepreneurial behaviors and engagement attitudes will have a higher likelihood of favoring these in preferred leader behaviors. Hence, we can expect that the GEM indicators can impact the followers preferred leader behaviors:

Hypothesis 3 (H3).

Strong entrepreneurship values impact followers’ preferred leader behaviors.

2. Methods

2.1. Datasets

This study uses data sampled for the LBDQXII questionnaire [11], complemented with several leading databases on leadership and entrepreneurship such as Hofstede indexes, project GLOBE and GEM. Raw data applied from LBDQXII was gathered from 15 countries [9]. Data from GLOBE, Hofstede and GEM were collected from relevant online research databases, and then matched and indexed to the LBDQXII data gathered for each respective country (for full details, please contact the corresponding author). There are 9005 datapoints (rows, respondents) in total, with 52 variables extracted and matched with the main LBDQ database for the automated machine learning models to test the main hypotheses. The following hypotheses, presented earlier in this article and developed from the extant literature, were applied in the conceptual model below (Figure 1) and tested in the analyses.

2.2. Automated Machine Learning

To explore the data structure and its predictors of the LBDQ dataset, we use automated machine learning (AML). AML helps to minimize predictive errors by exploring patterns in the data and, through these patterns, making predictions. AML aims to find the most optimal solutions between sets of (independent) variables, or ‘features’, towards the (dependent variable) outcome, or ‘target’ [13]. To achieve this, we use DataRobot (version 10.1.x), one of the many leading software developers that allows for processing of numerical, categorical and time-series data [15,32]. We follow the guidelines presented by Larsen and Becker [33] and use DataRobot to select and deploy the most suitable models that accurately predict the target variable through machine learning.

The AML prediction process commonly has three major steps: data pre-processing (or feature engineering), model selection and ‘hyperparameter optimizations’ and, finally, model interpretation and prediction scorings [32].

In step one, DataRobot begins to pre-process the data through automated feature engineering after a target variable has been selected for prediction to best prepare the dataset for machine learning [34]. This step may require additional optimization via the researcher to ‘feature engineer’ the processed data when deciding and choosing the most relevant predictors for the target variable—which we have explained in our theory and literature review section. We use these informative features (independent variables) compiled from the LDBQ, cultural and entrepreneurial databases to predict the target variables.

The second step, model selection, is where DataRobot attempts to find the best model in its inventory with the dataset. It is also the most crucial and innovative step of DataRobot AML for processing the data into partitions: the data is split into segments, called the ‘training data’, and then used to test and validate multiple models on the other segments through learning algorithms [13]. By default, DataRobot employs a 5-fold cross-validation partitioning method with a 20% holdout sample; in practice, DataRobot begins by selecting 16% of the data to find which models in its inventory it should run with 32% of the data. This process repeats with every model in its inventory on its own or combined together (called “blender” models) to obtain predictive accuracy [33]. Therefore, the best models are those that have the minimal amount of prediction errors across the different partitioned data (validation and cross-validation sample), as well as with the 20% holdout sample.

The final third step is model interpretation and prediction scoring: DataRobot presents the outputs through a detailed ‘model leaderboard’ that ranks and recommends the most accurate models and identifies the most important features. DataRobot can also recommend the optimal metric to score the models in the leaderboard. In our results, DataRobot recommended use of the root mean squared error (RMSE) metric, but we also use and include the R squared (R²) metric. RMSE measures the inaccuracy of predicted mean values assuming the target (dependent variable) is normally distributed; when interpreting the metric, a lower RMSE is, in general, better [35]. R² measures how much the model can explain the proportion of total variation of outcomes, as well as a measure of goodness-of-fit for the fitness of data to a regression line. A R² metric of 0% indicates a model that explains no variability in the data, while 100% indicates a model that explains all variability in the data [35]. The metric scores for the validation, cross-validation and holdout samples are recorded and summarized together with the results of the feature impacts. These outputs are then visualized and presented in the results section.

3. Results

3.1. Automated Machine Learning Results

The results from the AML show that cultural indicators from GLOBE predict the likelihood of the preferred leader behaviors of “Role Assumption”, “Production Emphasis” and “Initiation of Structure”. Hofstede’s Long-Term/Short-Term Orientation is the most critical predictor of preferences for “Tolerance of Uncertainty” and “Initiation of Structure”, whereas the value of restraint impacts the likelihood of preferring leaders with skills in “Integration” and “Consideration”. Significant entrepreneurial values indicators have a significant impact on preferences for leaders focused on “Initiation of Structure”, “Production Emphasis” and “Predictive Accuracy”. We also find that age and gender significantly impact our preferences for leader specific behaviors, supporting earlier studies.

3.2. Model Accuracy and Performance

Using the same set of cultural and entrepreneurial indicators, we created multiple AML models to test which of the informative features were most important for the likelihood of each preferred leader behavior; with DataRobot, we created twelve separate AML prediction procedures for each preferred leader behavior, designating each leader behavior as the ‘target feature’. DataRobot recommended the best modelling algorithms from its inventory for deployment regarding each preferred leader behavior [36]. Six different models were selected for best fit for data processing and feature engineering, and each of the blueprints of the modelling algorithms are displayed in Figure 2: the ‘gradient boosted trees regressor’, ‘extreme gradient boosted trees regressor’, ‘keras slim neural network regressor’, ‘generalized additive 2 model’, ‘Nystroen kernel regressor’ and ‘LightGBM random forest regressor’ (see Appendix A Figure A1 for blueprints of each AML procedure). The results from each of the AML prediction procedures and the metric scores are then extracted and recorded using the DataRobot model leaderboard feature.

Table 3. DataRobot feature impact.

	Target Variable
Feature impact (0–100, percent)	REP *	DRE *	ToU *	PER *	IoS *	ToF *	ROA *	CON *	PEM *	PAC *	INT *	SOP *
GLOBE_Uncertainty Avoidance (As Is)	48.23%	2.65%	−1.10%	0.39%	52.03%	7.97%	14.26%	−4.95%	47.59%	−1.02%	5.12%	0.41%
GLOBE_Future Orientation (As Is)	5.15%	4.79%	0.81%	0.83%	33.54%	9.57%	20.61%	11.80%	12.20%	12.82%	2.53%	0.00%
GLOBE_Power Distance (As Is)	5.65%	3.65%	−0.19%	0.15%	16.38%	2.14%	5.51%	16.22%	15.09%	22.78%	7.29%	0.00%
GLOBE_Collectivism Institutional (As Is)	12.48%	2.22%	−0.04%	0.46%	51.33%	5.13%	23.90%	11.69%	31.97%	5.03%	11.46%	24.75%
GLOBE_Human Orientation (As Is)	4.59%	0.15%	−0.87%	0.50%	22.11%	2.89%	51.81%	−1.88%	18.29%	16.65%	12.09%	0.41%
GLOBE_Performance Orientation (As Is)	13.33%	1.38%	−0.01%	0.51%	48.64%	2.94%	10.71%	11.29%	41.39%	2.92%	10.04%	0.00%
GLOBE_Collectivism Group (As Is)	11.26%	0.28%	1.09%	8.41%	41.50%	−0.22%	15.91%	−0.26%	27.88%	57.50%	0.08%	0.00%
GLOBE_Gender Egalitarianism (As Is)	6.33%	0.11%	0.14%	−0.02%	19.47%	7.02%	21.30%	86.23%	14.70%	9.93%	13.56%	1.03%
GLOBE_Assertiveness (As Is)	6.76%	0.19%	−0.19%	0.66%	17.59%	0.01%	14.16%	8.55%	19.82%	29.51%	9.67%	−0.57%
GLOBE_Power Distance (Should Be)	9.10%	0.32%	0.05%	0.00%	20.75%	6.56%	5.76%	−6.55%	22.35%	12.87%	−0.40%	0.47%
GLOBE_Collectivism Institutional (Should Be)	15.04%	0.71%	0.91%	100%	55.50%	11.23%	17.61%	2.11%	28.26%	13.30%	14.08%	0.00%
GLOBE_Humane Orientation (Should Be)	7.28%	0.36%	−0.16%	2.46%	37.20%	2.83%	27.72%	9.40%	54.34%	18.89%	9.42%	0.00%
GLOBE_Power Orientation (Should Be)	3.63%	0.14%	−0.22%	0.08%	43.25%	−0.23%	11.23%	1.70%	31.84%	−0.27%	2.16%	0.00%
GLOBE_Collectivism Group (Should Be)	3.87%	0.06%	−0.03%	0.16%	38.20%	1.42%	9.65%	42.70%	28.71%	0.81%	7.62%	−0.03%
GLOBE_Gender Egalitarianism (Should Be)	10.03%	0.05%	0.64%	3.73%	50.13%	5.53%	12.41%	2.48%	43.14%	18.66%	19.88%	1.63%
GLOBE_Assertiveness (Should Be)	10.76%	5.53%	−0.13%	0.32%	30.97%	5.57%	10.40%	33.99%	35.62%	100%	9.94%	−0.55%
HOFSTEDE_Power Distance Index	4.05%	17.19%	−0.43%	0.04%	19.82%	−0.09%	17.78%	6.36%	19.46%	21.19%	7.42%	0.01%
HOFSTEDE_Individualism vs. Collectivism	2.20%	13.56%	0.81%	0.12%	21.70%	0.28%	23.04%	7.24%	19.00%	14.55%	5.69%	−0.24%
HOFSTEDE_Masculinity vs. Femininity	4.37%	1.10%	2.08%	0.88%	21.18%	0.39%	11.95%	4.93%	11.20%	5.55%	0.11%	1.61%
HOFSTEDE_Uncertainty Avoidance Index	4.45%	1.82%	0.67%	1.62%	18.78%	21.34%	4.16%	9.21%	11.50%	−0.90%	11.69%	0.16%
HOFSTEDE_Long-Term vs. Short-Term Orientation	8.75%	4.60%	63.79%	0.23%	52.93%	8.74%	17.60%	0.06%	43.59%	34.35%	6.51%	22.96%
HOFSTEDE_Indulgence vs. Restraint	6.39%	1.95%	−1.19%	1.20%	25.57%	2.37%	19.50%	36.24%	25.98%	−0.38%	85.68%	7.11%
GEM_Perceived Opportunities	6.80%	100%	0.28%	0.16%	55.59%	1.01%	16.26%	2.36%	36.16%	42.39%	1.14%	0.00%
GEM_Perceived Capabilities	2.97%	1.27%	0.47%	4.21%	30.85%	4.71%	24.40%	6.34%	16.34%	−0.40%	7.43%	0.00%
GEM_Fear of Failure Rate	4.16%	3.88%	−0.49%	0.75%	24.43%	0.36%	19.28%	64.90%	10.32%	10.96%	23.35%	18.16%
GEM_Entrepreneurial Intentions	8.38%	2.11%	−0.06%	1.56%	41.14%	−0.12%	19.45%	6.02%	23.24%	10.28%	0.25%	1.38%
GEM_Total Early_Stage Entrepreneurial Activity (TEA)	5.69%	1.81%	−0.17%	0.17%	41.80%	14.28%	40.38%	25.75%	23.90%	5.27%	32.46%	1.97%
GEM_Established Business Ownership	16.59%	3.78%	0.32%	−0.08%	50.87%	29.56%	46.66%	3.85%	51.09%	13.37%	10.44%	2.25%
GEM_Entrepreneurial Employee Activity	27.26%	0.13%	0.64%	−0.10%	79.14%	28.40%	38.01%	34.39%	67.63%	54.45%	69.08%	−1.26%
GEM_Motivational Index	7.21%	7.73%	−2.43%	0.32%	53.95%	10.43%	48.45%	11.62%	65.87%	12.50%	2.83%	−0.10%
GEM_Female/Male TEA	12.69%	2.83%	−0.80%	0.02%	42.23%	0.30%	23.82%	−5.78%	37.24%	16.56%	0.38%	0.64%
GEM_Female/Male Opportunity_Driven TEA	27.14%	0.12%	1.27%	1.46%	91.18%	100%	51.86%	51.56%	68.13%	90.02%	17.88%	0.00%
GEM_High Job Creation Expectation	13.26%	0.39%	−0.35%	0.08%	27.35%	18.60%	14.47%	15.40%	20.46%	8.06%	−4.24%	0.54%
GEM_Innovation	8.89%	1.22%	1.48%	0.16%	38.46%	1.89%	24.48%	36.63%	29.99%	28.24%	72.20%	−0.77%
GEM_Business Services Sector	6.22%	0.07%	−1.36%	0.49%	33.00%	23.77%	23.34%	16.01%	24.22%	6.13%	2.25%	0.00%
GEM_High Status to Successful Entrepreneurs	9.58%	0.12%	0.23%	0.84%	45.61%	3.26%	19.43%	48.55%	29.65%	0.02%	16.10%	−0.46%
GEM_Entrepreneurship as a Good Career Choice	7.28%	44.24%	11.61%	15.77%	48.42%	3.66%	24.86%	11.40%	44.09%	11.07%	−1.02%	−1.57%
GEM_Financing for Entrepreneurs	15.49%	0.13%	2.62%	16.57%	81.74%	1.86%	10.09%	1.22%	86.15%	38.74%	13.29%	0.00%
GEM_Governmental Support and Policies	12.97%	3.16%	65.21%	0.30%	65.49%	8.10%	7.60%	5.63%	64.26%	77.37%	10.07%	51.21%
GEM_Taxes and Bureaucracy	19.59%	0.06%	−0.64%	2.78%	69.14%	10.37%	8.89%	−5.35%	62.06%	91.89%	12.92%	7.39%
GEM_Governmental Programs	19.66%	0.13%	−2.10%	2.87%	74.30%	12.79%	4.71%	−0.47%	72.60%	0.05%	15.29%	0.00%
GEM_Basic School Entrepreneurial Education and Training	21.72%	0.00%	15.41%	0.31%	65.99%	1.46%	5.99%	−4.97%	59.26%	57.37%	1.21%	6.07%
GEM_Post School Entrepreneurial Education and Training	19.47%	0.00%	−0.39%	0.71%	73.53%	−0.63%	4.45%	−4.38%	65.86%	15.15%	2.11%	0.00%
GEM_R&D Transfer	17.71%	0.00%	11.35%	0.42%	77.76%	−0.06%	5.71%	−3.24%	72.38%	5.70%	11.80%	0.00%
GEM_Commercial and Professional Infrastructure	24.71%	0.16%	−1.13%	0.40%	86.46%	13.96%	2.03%	−1.79%	78.52%	1.05%	26.74%	−0.89%
GEM_Internal Market Dynamics	5.42%	0.01%	6.73%	0.47%	40.50%	26.89%	6.29%	0.64%	36.85%	2.14%	2.86%	1.13%
GEM_Internal Market Openness	18.15%	0.00%	−0.39%	9.93%	90.69%	1.34%	5.62%	6.26%	82.43%	9.78%	12.87%	26.95%
GEM_Physical and Services Infrastructure	22.73%	0.00%	−1.61%	0.05%	71.54%	0.19%	3.89%	−3.46%	55.09%	3.77%	13.73%	0.23%
GEM_Cultural and Social Norms	4.91%	0.00%	−4.17%	−0.04%	20.80%	11.93%	7.00%	13.47%	16.33%	5.22%	1.43%	−0.45%
Country	23.50%	7.99%	11.60%	−0.04%	14.17%	1.03%	4.79%	33.29%	10.35%	52.70%	−0.07%	100%
Age	100%	41.09%	100%	20.30%	100%	28.68%	100%	100%	100%	16.60%	80.83%	31.58%
Gender	49.41%	10.63%	29.57%	7.29%	19.98%	23.68%	19.80%	95.45%	30.81%	15.39%	100%	34.78%
DataRobot Metric Score: RMSE
Validation	0.6735	0.6274	0.5929	0.6684	0.6308	0.5653	0.5814	0.5806	0.6341	0.7250	0.8018	0.6032
Cross-Validation	0.6812	0.6466	0.6093	0.6792	0.6245	0.5770	0.5789	0.5775	0.6349	0.7328	0.8035	0.6080
Holdout	0.6982	0.6707	0.6137	0.6657	0.6393	0.5812	0.5788	0.5726	0.6438	0.7155	0.7846	0.5970
DataRobot Metric Score: R2
Validation	0.4286	0.2334	0.1088	0.3629	0.4585	0.3441	0.2154	0.2443	0.3483	0.3082	0.4153	0.4153
Cross-Validation	0.4082	0.2090	0.1077	0.3469	0.4637	0.3398	0.2308	0.2577	0.3246	0.2883	0.4078	0.4158
Holdout	0.4203	0.1903	0.1199	0.3570	0.4595	0.3277	0.2364	0.2760	0.3125	0.3025	0.4385	0.4490

* REP = Representation, DRE = Demand Reconciliation, ToU = Tolerance of Uncertainty. PER = Persuasiveness, IoS = Initiation of Structure, ToF = Tolerance of Freedom, ROA = Role Assumption, CON = Consideration, PEM = Production Emphasis, PAC = Predictive Accuracy, INT = Integration. SOP = Superior Operation.

summarizes the multi-modelling predictions of each LBDQ dimension with the same subset of informative features, as well as the RMSE and R² metric scores for the validation, cross-validation and holdout samples. When we compare both metrics for the validation, cross-validation and holdout samples, results show that each of the modelling algorithms has different scores relating to model accuracy and performance in predicting each preferred leader behavior. This means that using the same subset of datasets corresponds to differing strengths in accuracy and performance for predicting each preferred leader behavior, indicating that the same subset of informative features will have different importance for each target variable. This novelty in using and testing multiple models as an empirical method allows for better prediction of the variable relationships between predictors and leadership preferences. Additionally, the feature effects show which informative features are the most important in each prediction of preferred leadership behavior, where the most important item is given the total of 100% (see Appendix A Figure A2 for feature effects for each LBDQ dimension). Figure 3 visualizes the multi-modeling predictions of each LBDQ dimension on a visual heatmap using a red scale to highlight higher and lower feature associations across the list of features; feature associations range between 0% and 100%, and a feature impact closer to 1 is displayed with a brighter color. We used ChatGPT (version 4o) to create the visual heatmap. The feature’s importance (variables) is interpreted through perturbation [14]: by adding random numbers to the informative features, DataRobot can estimate the impact on prediction accuracy—and thus categorize which of the informative features are most important for predicting the target leader behavior dimension.

3.3. Cultural and Entrepreneurial Predictors of Preferred Leadership

From Table 3 and Figure 3, we find that for the preferred leadership behavior Representation (RMSE validation = 0.6735, cross-validation = 0.6812, holdout = 0.6982; R² validation = 0.4286, cross-validation = 0.4082, holdout = 0.4203), age (β = 1.00), gender (β = 0.49) and ‘GLOBE uncertainty avoidance (as is)’ (β = 0.48) are the most relevant predictors with the strongest visibility on the heatmap, supporting Hypothesis 2.

Demand Reconciliation (RMSE validation = 0.6274, cross-validation = 0.6466, holdout = 0.6707; R² validation = 0.2334, cross-validation = 0.2090, holdout = 0.1903) has ‘GEM perceived opportunities’ (β = 1.00), ‘entrepreneurship as good career choice’ (β = 0.44) and age (β = 0.41) as the most relevant predictors on the visual heatmap, supporting Hypothesis 3.

For Tolerance of Uncertainty (RMSE validation = 0.5929, cross-validation = 0.6093, holdout = 0.6137; R² validation = 0.1088, cross-validation = 0.1077, holdout = 0.1199), age (β = 1.00), ‘Hofstede long-term vs. short-term orientation’ (β = 0.64) and ‘GEM governmental support and policies’ (β = 0.65) are the most relevant predictors, supporting Hypotheses 1 and 3.

Persuasiveness (RMSE validation = 0.6684, cross-validation = 0.6792, holdout = 0.6657; R² validation = 0.3629, cross-validation = 0.3469, holdout = 0.3570) has ‘GLOBE collectivism (institutional, should be)’ (β = 1.00) as the most relevant predictor, supporting Hypothesis 2.

For Initiation of Structure (RMSE validation = 0.6308, cross-validation = 0.6245, holdout = 0.6393; R² validation = 0.4585, cross-validation = 0.4637, holdout = 0.4595), many of the 52 features are relevant predictors, looking at the visual heatmap, with age (β = 1.00), ‘GEM female/male opportunity driven total early stage entrepreneurial activity (TEA)’ (β = 0.91) and ‘internal market openness’ (β = 0.91) appearing prominently in particular, supporting all three hypotheses.

For Tolerance of Freedom (RMSE validation = 0.5653, cross-validation = 0.5770, holdout = 0.5812; R² validation = 0.3441, cross-validation = 0.3398, holdout = 0.3277), ‘GEM female/male opportunity driven TEA’ (β = 1.00) is the most relevant predictor, supporting Hypothesis 3.

Role Assumption (RMSE validation = 0.5814, cross-validation = 0.5789, holdout = 0.5788; R² validation = 0.2154, cross-validation = 0.2308, holdout = 0.2364) has age (β = 1.00), ‘GEM female/male opportunity driven TEA’ (β = 0.52) and ‘GLOBE human orientation (as is)’ (β = 0.52) as the most relevant predictors, supporting Hypotheses 2 and 3.

For Consideration (RMSE validation = 0.5806, cross-validation = 0.5775, holdout = 0.5726; R² validation = 0.2443, cross-validation = 0.2577, holdout = 0.2760), age (β = 1.00), gender (β = 0.95) and ‘GLOBE gender egalitarianism (as is)’ (β = 0.86) are the most relevant predictors, supporting Hypothesis 2.

Production Emphasis (RMSE validation = 0.6341, cross-validation = 0.6349, holdout = 0.6438; R² validation = 0.3483, cross-validation = 0.3246, holdout = 0.3125) has many of the features as relevant predictors, displayed with several levels of visibility in the heatmap, with age (β = 1.00), ‘GEM financing for entrepreneurs’ (β = 0.86) and ‘internal market openness’ (β = 0.82) appearing prominently in particular, supporting Hypothesis 3.

For Predictive Accuracy (RMSE validation = 0.07250, cross-validation = 0.7328, holdout = 0.7155; R² validation = 0.3082, cross-validation = 0.2883, holdout = 0.3025), ‘GLOBE assertiveness (should be)’ (β = 1.00), ‘GEM female/male opportunity driven TEA’ (β = 0.90) and ‘GEM taxes and bureaucracy’ (β = 0.92) are the most relevant predictors, supporting Hypotheses 2 and 3.

Integration (RMSE validation = 0.8018, cross-validation = 0.8035, holdout = 0.7846; R² validation = 0.4153, cross-validation = 0.4078, holdout = 0.4385) has gender (β = 1.00), age (β = 0.81) and ‘Hofstede indulgence vs. restraint’ (β = 0.86) as the most relevant predictors, supporting Hypothesis 1.

Lastly, for Superior Operation (RMSE validation = 0.6032, cross-validation = 0.6080, holdout = 0.5970; R² validation = 0.4153, cross-validation = 0.4158, holdout = 0.4490), country (β = 1.00) and ‘GEM governmental support and policies’ (β = 0.51) appear as the most relevant predictors, supporting Hypothesis 3.

Based on these results, we observe that some features are less relevant as predictors for certain preferred leader behaviors, while we can also observe that a majority of these predictors impact predicted followers’ preferences in leader behaviors. Overall, all of our three hypotheses are supported.

3.4. Feature Performance

Figure 4a–l shows the partial dependence (average partial dependence) of a selection of informative features to each preferred leader behavior to highlight the AML’s feature in evaluating performance for every single feature. The yellow line showcases the informative feature’s marginal effect on the target variable, the preferred leader behavior dimension, and shows if the relationship between the feature and target is linear, ‘monotonic’ or complex [13]. Thus, we can use the partial dependence of the selected informative features to observe how changes in the value of the informative feature (the independent variable) impacts the model’s predictions (assuming all other variables are kept the same). For some of the selected informative features (Figure 4e,g), the charts reveal non-linear relationships between feature and target variables—the independent variables to preferred leader behaviors.

4. Discussion

This article set out to investigate to what degree an automated machine learning model (AML) can estimate cultural and entrepreneurial predictors of preferred leader behavior in international businesses. As future leadership studies have the potential to expand the boundaries of big data analyses using causal and predictive modelling, our study is the first attempt to research and discuss how AML tests such boundaries. Existent studies present deficiencies in predictive models, due to the arbitrary selection of factors and use of limited linear models. This study’s index is developed in a robust three-stage process of data pre-processing, model selection and model interpretation. The use of AML in leadership studies configures an advantage compared to traditional linear and additive models, as it has the capacity to select the best model considering the predictive capacity of many models simultaneously. This study set out to test three hypotheses, namely, to what degree cultural values and entrepreneurial factors predict the preferred leader behaviors. The results indicate that GLOBE indicators increased the likelihood of the preferred leader behavior of “Role Assumption”, “Production Emphasis” and “Initiation of Structure”. Hofstede’s’ Long-Term/Short-Term Orientation is the most critical predictor for “Tolerance of Uncertainty”, “Production Emphasis”, “Predictive Accuracy” and “Initiation of Structure”. The cultural value of restraint impacts how we value a leader with strong skills in “Integration” and “Consideration”. Significant entrepreneurial values indicators have a significant impact on preferences for leaders focused on “Initiation of Structure”, “Production Emphasis” and “Predictive Accuracy”. Findings also support earlier studies that age and gender significantly impact preferences for leader specific behaviors. Consequently, all three hypotheses are supported.

Besides the methodological and theoretical implications of high predictive quality with AML as discussed, our research also offers important practical implications for future leadership studies. First, we offer insights for the potential development and training of global business managers for success in multiple countries, by offering a better understanding of the impact of an entrepreneurial mindset and strong specific cultural values on who we prefer to lead us in global business. For example, followers that prefer cultural practices ‘as it is’ are likely to prefer those leaders that purposefully exercise their leadership role; actual practices in the workplace or a professional setting take the highest priority as these ‘match’ their cultural mindset.

Another finding from the AML results is the importance we place on a leader who shows resilience in handling uncertainty in international business: the results showed that the cultural and entrepreneurial values predicted the preferred leader behaviors of “Initiation of Structure”, “Consideration”, “Role Assumption” and “Production Emphasis”. Followers prefer managers that clearly define roles, consider followers, actively exercise their role, and apply pressure for productive outputs, suggesting followers prefer leaders who show resilience. This finding appears intuitive, given past post-pandemic uncertainties regarding geo-political and economic situations: a follower will prefer a leader that show stronger skills of resilience.

Lastly, followers that have increased entrepreneurial values and practices will strongly prefer leader behavior that establishes clear role definitions and efforts in the interest of productiveness. Such empirical findings support earlier studies that the human factor of leadership is becoming even more critical for international business [37].

5. Conclusions

In this article, we applied an Automated Machine Learning (AML) approach to predict preferred leadership behaviors across cultures, integrating data from the Leader Behavior Description Questionnaire (LBDQXII), Hofstede’s cultural dimensions, the GLOBE project and the Global Entrepreneurship Monitor (GEM). Three hypotheses were tested: (1) cultural values significantly impact followers’ preferred leadership behaviors, (2) global leadership and organizational effectiveness values across cultures influence these preferences and (3) strong entrepreneurial values are correlated with preferred leadership traits. The study confirmed these hypotheses, showing that cultural dimensions and entrepreneurial values significantly predict leadership preferences. Moreover, age and gender were found to influence leadership preferences, further validating previous research.

The key finding of this paper is that cultural dimensions, particularly the “as is” cultural practices from the GLOBE study, have a stronger influence on leadership preferences than normative “should be” values. This reinforces the idea that actual societal practices are more critical to understanding leadership preferences than aspirational values. The study’s contribution to cultural leadership theory is significant, as it introduces AML as a powerful tool for refining predictive models in leadership studies, going beyond traditional linear approaches. By combining cultural and entrepreneurial factors, the paper provides a more nuanced understanding of how leadership behaviors are shaped by culture, offering practical insights for developing global business leaders attuned to cultural and entrepreneurial contexts. The findings of this study have shown that in a world where AI and big data may be used to predict a complex range of business or societal situations, AML can be applied beyond pure technology, statistics, science and medicine to predict cultural and entrepreneurial factors impacting how we may prefer a leader to lead.

All studies have limitations, as is the case with this present study. We have limited datapoints retrieved from a survey of respondents across 18 countries, but also limited databases of cultural and entrepreneurial predictors for AML. Future research could also include a wider range of leadership factors from a larger range of countries, and include more relevant databases for building more AML models. Studies can also explore more on the predictive likelihoods of preferred values between normative and factual practices. Despite the above-mentioned limitations, this study has successfully answered the question ‘can preferred leadership styles be predicted by cultural and entrepreneurial factors in an AML model?’ And, from the results of this study, the answer seems to be a yes!