The current issue and full text archive of this journal is available on Emerald Insight at:

https://www.emerald.com/insight/0143-7720.htm

A human resources analytics and Analytics and

machine-
machine-learning examination of learning
examination
turnover: implications for
theory and practice 1405
Dan Avrahami and Dana Pessach Received 3 December 2020
Revised 29 April 2021
Department of Industrial Engineering and Management, Tel Aviv University, 26 July 2021
Tel Aviv, Israel 28 October 2021
17 November 2021
Gonen Singer 12 January 2022
Accepted 28 January 2022
Faculty of Engineering, Bar-Ilan University, Ramat Gan, Israel, and
Hila Chalutz Ben-Gal
Department of Industrial Engineering and Management,
Afeka Tel-Aviv Academic College of Engineering, Tel Aviv, Israel

Abstract
Purpose – What do antecedents of turnover tell us when examined using human resources (HR) analytics and
machine-learning tools, and what are the respective theoretical and practical implications? Although the
turnover literature is expansive, empirical evidence on turnover antecedents studied using data science tools
remains limited.
Design/methodology/approach – To help reinvigorate research in this field, the authors propose a novel
examination of turnover antecedents—competencies, commitment, trust and cultural values—using big data
tools to develop a granular, case-dependent measure of turnover.
Findings – Using archival data from 700,000 employees of a large organization collected over a period of ten
years, the authors find that turnover is generally associated with varying levels of these antecedents. However,
in more fine-grained analysis, their relation to turnover is contingent upon role, person and cultural
background.
Originality/value – The authors discuss the implications on turnover and strategic HR research and the
potential of Artificial Intelligence and machine-learning methods in the design and implementation of
managerial and HR planning initiatives.
Keywords Turnover, SHRM, HR analytics, Artificial intelligence, Big data, Machine-learning, Data science
Paper type Research paper

Introduction
Turnover has long drawn scholarly and managerial attention. Employee turnover refers to
the number or percentage of employees who leave an organization and are replaced by new
employees or to the transfer of employees across organizational boundaries (Hom et al., 2017).
What do antecedents of turnover tell us when analyzed using HR analytics and machine-
learning tools, and what are the respective theoretical and practical implications?
We believe that progress in understanding people-related processes—such as turnover—
utilizing HR analytics and machine-learning tools has begun to accumulate because of two
core limitations in the extant research. First, given the theorized benefits of antecedents of the
human resources phenomenon (e.g. turnover), it is typically not possible to observe all aspects
of the phenomenon directly, let alone predict it (Allen et al., 2014; Chalutz Ben-Gal, 2019;
International Journal of Manpower
Vol. 43 No. 6, 2022
pp. 1405-1424
This paper was partially supported by the Koret Foundation Grant for Smart Cities and Digital Living © Emerald Publishing Limited
0143-7720
2030. DOI 10.1108/IJM-12-2020-0548
IJM Marler and Boudreau, 2017). Instead, prior studies relying on scholarly efforts to model the
43,6 antecedents of turnover in organizations are likely to be limited because they ignore the
dynamic nature of turnover as a whole, failing to consider the various stakeholders involved
(Tzafrir et al., 2012). Thus, in many cases, these antecedents may correspond poorly to how
they actually come into play in various settings. Therefore, establishing a grounded
framework to understand the complexity of turnover as an organizational phenomenon may
assist in developing a more robust theory.
1406 Second, prior work has typically measured turnover based on surveys, which are
implemented once or at best sporadically. In part for this reason, turnover has generally been
conceptualized as a static construct; for example, in exploring the relationship between
socioeconomic characteristics and turnover (Madariaga et al., 2018). However, turnover that
manifests a dynamic phenomenon is likely to vary from day to day, resulting in complex,
long-term workforce planning challenges (Huselid and Becker, 2005). Without being able to
observe these nuanced and granular fluctuations, it is impossible to know whether—and in
what forms—they might have meaningful managerial implications.
In this article, we seek to address these gaps and help reinvigorate turnover research by
focusing on a specific methodological approach involving HR analytics and machine-learning
tools. We believe that through HR analytics, turnover may be analyzed from a wider, more
suitable perspective, thereby assisting organizations in strategic human capital decisions
(Huselid and Becker, 2005; Malik et al., 2020; Marler and Boudreau, 2017). Guided by these
insights from management scholars, and by recent research calling for more empirical work
in the field (Allen et al., 2014; Chalutz Ben-Gal, 2019; Guenole et al., 2017), we reveal that
turnover is a personalized and granular organizational challenge. By accepting this fact,
organizations can fine-tune the scope and accuracy of managerial attention to turnover and
thus maximize people-centered decision-making processes.
We note, however, that known antecedents of turnover—competencies, commitment, trust
and cultural values—when explored by means of HR analytics are not necessarily aligned with
existing theories. Specifically, we uncover some personalized and disruptive aspects of turnover
that are sometimes counterintuitive. Moreover, we theorize that—all else being equal—
competencies, commitment, trust and cultural values have significant associations with
turnover. However, we also propose that—when utilizing HR analytics tools—turnover may
appear to be a counterintuitive phenomenon, hence the need for further theoretical development.
We test these ideas using a unique and rich longitudinal archival dataset that includes
more than 700,000 employees in the largest public-sector organization over a period of
one decade. The dataset includes rich turnover records, as well as control and independent
variables pertaining to employees. We also use machine-learning techniques to analyze
turnover and its associated features, some serving as control variables in our models and find
strong empirical support for our four hypotheses. We discuss the implications of this novel
approach for research and practice.
Our work yields theoretical and practical contributions. From a theoretical perspective, it
contributes to research on turnover and HR analytics. Our novel approach integrates HR and
data science perspectives on how to analyze turnover. Whereas research has increasingly
recognized the need for a deeper analytical understanding of turnover (Allen et al., 2014), the
core research method in this field—surveys and self-reports—are ill-suited to uncovering
fine-grained organizational variation. This work deepens our understanding of the growing
importance of machine-learning tools. For example, a significant study on prediction based
on machine-learning focused on big data techniques in psychology (Yarkoni and Westfall,
2017). We nuance this account by examining these tools and implementing them in the field of
HR and in turnover research.
From a practical perspective, this work would appear to have important implications for
the design of workforce planning practices in organizations. We illustrate how advanced
tools can be used to extract personalized turnover patterns and shed light on issues far Analytics and
beyond the conventional use of hypothesis testing and regression models. To this end, we machine-
show the benefit of applying data-driven tools to extract actionable recommendations for
strategic decision-making. We propose diagnostic tools which may be useful to globally
learning
distributed workforce and online labor markets (Chalutz Ben-Gal, 2019; Malik et al., 2020). examination

Theory 1407
Turnover
Employee turnover refers to the number or percentage of employees who leave an
organization and are replaced by new employees (Hom et al., 2017). Turnover is a major
concern for management because organizations make enormous investments in their
employees in terms of recruiting, training, developing and retaining them. Turnover has
attracted increased scholarly and managerial attention for several reasons. First, it is usually
associated with potential risk to the organization because it can significantly impact costs.
Such costs stem from recruitment and selection expenses, training costs, and loss of human
capital, which are typically difficult to measure but can be quite substantial (de Oliveira
et al., 2019).
We believe there is a shortage of existing theory to explain, let alone predict, employee
turnover, specifically in light of recent scholarly calls for a broader conceptualization of
analytical mindsets in turnover research and recommendations to incorporate novel elements
of research design, data collection, measurement and analytical strategies (Allen et al., 2014;
Hom et al., 2017). As a result, there arises the need to seek novel theoretical and practical
insights, which is the focus of this study. Thus, we acknowledge that the HR analytics
approach might shed new light on the turnover phenomenon (Chalutz Ben-Gal, 2019).

Turnover and competencies

Earlier studies reveal that employee competencies play an important role in employee
turnover. Competency has been defined as the sum of knowledge, skills and abilities that an
individual possesses (Krausert, 2017).
In organizations, the breadth of the available individual toolkit can have important
turnover implications. For example, it is reported (Chalutz Ben-Gal, 2019) that the future of
work entails novel work arrangements, which are more than ever skill-based. However,
despite the importance attributed to employees’ competencies, studies lack a nuanced view of
how competencies play a specific role in various turnover scenarios. We believe this granular
view may be obtained via an HR analytics approach.
We extend these arguments from the level of the organization as a whole to the person and
role unit of analysis. In the context of HR analytics, rather than traditional analytical tools, we
similarly propose that the relation between competencies and turnover will be contingent
upon person and role because of granularity. Thus, we expect.
H1. Competencies have a positive relation to turnover reduction. However, in the context
of HR analytics, the relation between competencies and turnover will be contingent
upon person and role.

Turnover and commitment

The literature has examined organizational commitment (hereafter OC) and its impact on
employee turnover. Studies have proved that OC tends to extend and prolong the exchange
processes with peers and stakeholders (Chalutz Ben-Gal and Tzafrir, 2011; Meyer and
Parfyonova, 2010). We focus on organizational citizenship behavior (OCB) – i.e. individuals’
voluntary commitment within the company that is not part of contractual tasks (Organ, 1988).
IJM Whereas these prior studies utilized conventional analytical tools, we acknowledge that,
43,6 because of granularity, the effects of voluntary commitment on turnover will be contingent
upon role. In such a context, turnover may be thought of as invoking a granular-based
phenomenon. Thus, we anticipate:
H2. Voluntary commitment (OCB) will have a positive relation to turnover reduction.
However, in the context of HR analytics, the relation between voluntary commitment
1408 and turnover will be contingent upon role.

Turnover and trust

The literature has also examined trust in the workplace in association with turnover theories.
Trust is defined (Mayer et al., 1995) as “the willingness of one party to be vulnerable to the
actions of another party, based on the expectation that the other party will perform particular
actions important to the trustor, irrespective of the first party’s ability to monitor or control
that other party” (p. 712). In this study, we follow the argument (Fineman, 2003) that trust “is
not something that is simply present or absent from a social relationship, but is contextually
or structurally specific” (p. 565). Thus, from this perspective, our study reveals context-
specific patterns related to trust and turnover.
Several scholars have advocated the idea that trust is primarily a characteristic of an
organizational process. For example, Tzafrir and Dolan (2004) contributed to the notion that
trust can be measured at an organizational level (p. 115). It has also been found that trust
boosts the performance of working teams (Bijlsma-Frankema et al., 2008). One recent study
investigated and emphasized the impact of trust on turnover (Purba et al., 2016), confirming
that the trustworthiness of supervisors affects the quality of the relationships between
supervisors and employees and, in turn, affects turnover.
Whereas these prior studies utilized conventional analytical tools, we acknowledge that,
because of granularity, the effects of trust on turnover will be contingent upon role. In such a
context, turnover may be thought of as invoking a granular-based phenomenon. Thus, we
anticipate.
H3. Trust has a negative relation to turnover reduction. However, in the context of HR
analytics, the relation between trust and turnover will be contingent upon role.

Turnover and cultural values

A substantial amount of theoretical and empirical work has paid attention to the meaning of
organizational values and their effects on individuals and on organizations as a whole. For
example, researchers have explored the link between human cultural values and
organizational values (Hofstede, 1984; Schwartz and Bilsky, 1990). In modern
organizations, stakeholders must attain a clear understanding of which cultural values
should be changed and how to successfully implement a process of change (Tzafrir et al.,
2012). It has been shown that employees from heterogeneous cultural backgrounds may
contribute to the organization’s efficiency (Corritore et al., 2019).
We extend these arguments and claim that in the context of HR analytics, turnover can be
thought of as invoking a granular-based phenomenon. In evaluating the extant turnover
literature, we believe there is a shortage of existing theory to explain, let alone predict,
employee turnover. As a result, since current theories are limited in their explanatory power,
the need for new theoretical and practical insights arises, which is the focus of this study.
Moreover, we believe there is a need to complement existing survey-based measures of
turnover and its antecedents, which provide mostly static (or at best episodic) snapshots of
turnover over an organization’s life cycle (Allen et al., 2014) and can be susceptible to various
forms of survey bias with more granular, dynamic measures. Thus, we expect that:
 H4. Cultural values [1] have a positive relation to turnover reduction. However, in the Analytics and
context of HR analytics, the relation between cultural values and turnover will be machine-
contingent upon cultural background.
learning
examination
Methods
Research setting and data
Our research setting is a large public organization which promotes a policy of wide social 1409
diversity. Consequently, the organization recruits a wide range of heterogeneous populations.
Our uniquely rich dataset comprises information on 700,000 employees recruited by the
organization over a period of ten years. Data represent a wide variety of jobs: core jobs (27.2%)
and administrative jobs (72.8%). The average employee turnover rate for the organization was
23.9%, with women averaging 17.8% and men averaging 28.6% in annual turnover. Our
sample comprised 43.5% female and 56.5% male respondents. We collected archival data on
individuals’ personal and demographic characteristics, as well as performance. Some of the key
variables include age, gender, family and marital status, residence, nationality, criminal record,
education, academic performance, job interviews, recruitment test scores, job placement and
medical data. Our expansive dataset enabled the extraction of rules and insights, thereby
extracting large subpopulations and resulting in high prediction significance. Tables 1, 2 and
Figure 1 below summarize some descriptive statistics of the population. Table 1 presents the
categorical values, their frequencies and the corresponding turnover rate. Table 2 maps the
variables used in each of the four hypotheses and Figure 1 illustrates the distribution of key
independent variables.
We preprocessed the data according to standard procedures in clustering analysis and
trained our model using Python 3, a popular software for training models. To assure a
scientifically rigorous study, we undertook wrangling and quality check procedures. Data
wrangling is “the process of acquiring high-quality data and turning ill-structured, messy
data into clean, tidy data ready for analysis” (Braun et al., 2018). We handled missing values

Variable Value Percentage Turnover

All 100.0% 23.9%

Gender Female 43.5% 17.8%
Male 56.5% 28.6%
Core role No 72.8% 22.3%
Yes 27.2% 28.2%
Individuals’ misconduct False 96.9% 23.8%
True 3.1% 35.4%
Volunteer activities False 99.9% 24.4%
True 0.1% 43.3%

Employees’ rated willingness to be placed in a specific designated High 82.1% 21.5%

job Low 17.9% 37.9%

Religion Religion A 95.3% 24.3%

Religion B 4.7% 22.5%
Country of Birth Native 82.4% 24.5%
Immigrant 17.6% 22.6%
Educational Background School type C 13.3% 23.8%
School type D 86.7% 24.2% Table 1.
Leadership Low 72.0% 28.6% Population and
High 28.0% 15.3% turnover
IJM Hypothesis Variable Operationalization
43,6
Competencies (H1) Academic Accumulated school years
Analytical Analytical test score
Demographic background Demographic background score
Physical Physical test score
Leadership Leadership test score
1410 Commitment (H2) Affective commitment Employees’ rated willingness to be placed in a
specific designated job
Organizational citizenship Volunteering activities
behavior (OCB)
Trust (H3) Organizational trust Personal interview score
Individual trust Individuals’ misconduct data
Cultural Language Written language score
Table 2. values (H4) Oral language Oral language score
Independent variables Educational background Education type
and operationalization Culture Country of birth
strategy Religion Religion

in line with acceptable techniques. For example, to ensure data completeness, we applied a
“successful job index” to all employees who were employed for a minimum period. We applied
data-filtering procedures by omitting the records of employees with incomplete data.
In the preprocessing phase, we consolidated 21 data tables for the purpose of masking
sensitive private data and executed feature enrichment processes. We utilized socioeconomic
data extracted from the Central Bureau of Statistics, such as residence socioeconomic score.
Additionally, we consolidated similar positions to obtain homogeneous clusters of positions
and to lump together similar personal characteristics.
Missing values were handled via an imputation process, in which we created a new
separate value for nominal features and zero for most cases of numerical features. Data
records of employees with many missing values were removed.

HR analytics and machine-learning

To examine the research hypotheses, we undertake an inductive exploration approach based
on HR analytics (Garcia-Arroyo and Osca, 2019) and machine-learning tools. We utilize
algorithms in order to find patterns of relationships in unstructured data. We then implement
a classification process: “an iterative procedure by which categories of objects are inductively
discovered and objects are subsequently assigned to them. Machine-learning classifiers, in
essence, let the data “speak for themselves.” They parse the data into categories that the
researcher, presumably, could not have imagined” (Chalutz Ben-Gal, 2019; Goldberg, 2015).

Dependent variable
Leaving an organization or unit is the most critical indicator of turnover. Thus, we extracted
turnover data per sub-population and roles, utilizing the following process. The HR
department examined the documented turnover reason per incumbent. We examined job
changes and classified them as “negative” (e.g. misfit) or “positive” (e.g. promotion, job
enrichment). Following the call to incorporate multiple measurement methods in turnover
research (Allen et al., 2014), we added diverse measurement methods beyond surveys. Our
methodological goal was to conceptualize turnover utilizing the alternative measurement
strategies detailed below (e.g. objective indicators). This enabled us to explore turnover from
a novel perspective. Tables 1 and 2 summarize key turnover characteristics. Table 1 presents
 20.0%
Analytics and
25.0% 17.5%
machine-
20.0%
15.0%
learning
12.5%
examination
Probability

Probability
15.0%
10.0%

10.0% 7.5%

5.0%
5.0% 1411
2.5%

0.0% 0.0%
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Organizational Trust Evaluation Score Analy cal Competencies Score

35.0%

30.0%
80%
%
25.0%
60%
20.0%
Probability

Probability

40% 15.0%

10.0%
20%
5.0%

0% 0.0%
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Oral Language Score Language Score

50%
25.0%

40%
20.0%
Probability
Probability

15.0% 30%

10.0% 20%

5.0% 10% Figure 1.

Independent variables
0.0% 0%
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 distribution
Quality Competencies Score Physical Competencies Score

the categorical values, the frequencies of each of their valid values and that value’s
corresponding turnover rate. Table 2 maps the variables used in each of the hypotheses.

Independent variables
We measure competencies using two operationalization strategies. First, we measure
competencies from an academic perspective: the number of accumulated school years.
Second, an external judge (e.g. HR professional) evaluated competencies (existing data) in
order to obtain a “subjective” perspective. Five dimensions were analyzed as a suitable basis
for measuring competencies: academic (as described above), demographic background,
analytical, physical and leadership. We then used those dimensions as the features of the
model. The intuition behind our measure is that competencies are intended to reflect
employees’ overall ability to fit a job.
We measure commitment utilizing two data sources. First, we used employees’ rated
willingness to be placed in a specific designated job. This operationalization technique
reflects individual preferences driven by positive emotions. Second, we measured
IJM commitment through voluntary activities performed by incumbents, in line with the
43,6 definition of voluntary commitment and OCB (Organ, 1988).
The third independent variable analyzed is trust. In keeping with trust definition, we
operationalize and measure trust at two levels: organizational and individual.
At the organizational level, we measure trust using a subjective organizational evaluation
of candidates following in-depth interviews performed by an HR professional based on a self-
impression rating composed of three items: (1) overall candidate curriculum vitae, (2)
1412 candidate trust to deliver performance indicators, and (3) candidate ethical considerations.
The three items are based on a scale score of 1–10. The HR professional was instructed to
assess the level of trust per candidate based on the scale scores, in addition to an overall
impression score per candidate. Then, at the individual level, we measure trust using
organizational data on individuals’ misconduct, accepting that these data could have a direct
influence on trust.
Lastly, we measured cultural values from the broad perspective of key stakeholders in line
with Tzafrir et al. (2012). Thus, we operationalized cultural values by measuring five
components that compare and contrast individual cultural values and organizational values
associated with culture (Elfenbein and O’Reilly, 2007). The five measured components of
cultural values are language (a two-dimensional measure), educational background, culture
and religion. To develop these measures, we fit a statistical model to the entire dataset in order
to detect the relationship with turnover. Table 2 above sums up the independent variables
used for each of the hypotheses.
Control variables
The control variables were integrated as predictors in our model in order to provide a
prediction tool for turnover in line with acceptable machine-learning techniques. For
example, personal information per employee, interview data, and external data were
integrated into the algorithms. Furthermore, we clustered the data using acceptable
clustering techniques.
Demographics. We measure two types of controls pertaining to demographic data, i.e.
gender and language.
Role. We controlled for functional role type because turnover is likely to vary as a function
of roles and may thus be a role-dependent phenomenon. For the purpose of granular analysis,
two main clusters were constructed: core roles and support (administrative) roles.

Analytical strategy
To test our four hypotheses, we started by executing logistic regressions predicting whether
competencies (H1), commitment (H2), trust (H3) and cultural values (H4) affect turnover. Note
that in this study the terms “effect” and “relation” are used interchangeably. Because we were
interested in the coefficient (in the case of logistic regression, coefficients represent log odds)
of turnover as a function of the independent variables, we ran four logistic regressions with
fixed effects, each one representing a different variable. All models included controls for
demographics and role.
We then applied HR analytics using machine-learning tools. This approach enables us to
detect and model more granular turnover patterns that are otherwise undetectable or hidden
in large datasets (Chalutz Ben-Gal, 2019). For example, big data and prediction algorithms
(e.g. Decision Trees and Bayesian Network), which are well-known probabilistic graphical
models that represent a set of variables and their conditional dependencies (Ben-Gal, 2008),
were applied to uncover patterns associated with employee turnover. Utilizing a data-rich,
algorithm-based prediction methodology enabled us to more deeply probe the nature of the
dependent variable (Allen et al., 2014). Additionally, it enabled the discovery of new insights
otherwise undetectable.
 We generated granular models that addressed various subpopulations. Key results are Analytics and
detailed in the following section. We then calculated the probability of the subpopulation machine-
remaining in a specific role, based on personalized characteristics. Lastly, we formulated
actionable recommendations per role.
learning
examination
Results
Turnover: validating the machine-learning model 1413
There are two common approaches to training and validating machine-learning models:
holdout and cross-validation. We tested our model through a number of scenarios using a
holdout set (data not available in the training phase of the model) in order to avoid “overfitting”
and estimated the accuracy of the model over future unknown data. Applying this approach to
the turnover model provided further evidence that the proposed model performed well at
capturing not only well-known turnover phenomena but also new patterns and insights less
detectable by conventional approaches. Figure 2 illustrates the model validation scheme. The
dataset was split into two datasets—a training set (70% of the data) —and holdout set (30%).
The training set was used to train a machine learning model. We applied the trained machine
learning model on the holdout set, generated predictions and evaluated them compared to the
real turnover values in the holdout set.
The validity checks suggested that our model accurately captured context-relevant
turnover patterns. Training a valid turnover model that performs well at capturing granular
data was key to our subsequent analyses, since we were interested in capturing differences in
turnover patterns and interactions on an individual basis.

Descriptive statistics
As indicated above, Tables 1, 2 and Figure 1 summarize some descriptive statistics of the
population, as well as the distribution of key independent variables. The average turnover
rate in the organization stands at 11%, while the average position dropout rate in the
organization is 24%. Running an Exploratory Data Analysis (EDA) shows that both turnover

Figure 2.
Model validation
strategy
IJM and position dropout rates depend on the employee’s specific role. Figure 3 presents the
43,6 number of roles associated with turnover and position dropout rates. The majority of roles
(100) are associated with low turnover rates (0–10%), while only 14 roles for example are
associated with higher turnover rates (20–30%). Figure 3 illustrates the core advantage of the
proposed approach to measure turnover relative to traditional survey-based measures: even
if collected at a few points in time during the organizational life-cycle, self-reports or surveys
of turnover would simply be unable to capture the fine-grained granular variation that
1414 machine-learning continuous measures, such as ours, can uncover.
Key results
Our first set of analyses focused on the main effect of competencies on turnover. Consistent
with hypothesis 1, we found a positive main effect of competencies on turnover, measured as
the five types of competencies illustrated above. Table 3 presents logistic regressions of
competencies, commitment, trust, cultural values and turnover (Hypotheses 1–4). In Table 3, a
logistic regression model that measures the effects of competencies on turnover (Hypothesis 1)
predicts the turnover rates based on role competencies values. The model shows significant
effects of competencies variables on the turnover rate. However, in the context of HR analytics,
the effects of competencies on turnover rates are contingent on person and role. Thus,
subsequent models indicate that the effects of competencies on turnover is personalized and
role-dependent.
Our findings suggest a non-intuitive interplay between competencies and turnover, as
presented, for illustration purpose, in three different graph types in Figures 4–6. Specifically,
Figure 4 benchmarks the competencies’ relation with turnover for all roles compared to a
specific support role. Figure 5 further analyzes the above comparison with respect to a

Figure 3.
Number of roles per
turnover and position
dropout rates
 Std. P
Analytics and
Regressions Variable Coef. Err. Z value machine-
learning
Regression Model 1: Competencies Academic 0.277 0.007 41.777 0.000
variables –Turnover Analytical 0.006 0.000 13.767 0.000 examination
Demographic background 0.056 0.002 24.748 0.000
Physical 0.006 0.000 34.516 0.000
Leadership 0.957 0.009 105.936 0.000 1415
Intercept 3.597 0.079 45.692 0.000
Regression Model 2: Commitment Affective Commitment 0.7959 0.027 29.956 0.000
variables –Turnover Organizational citizenship 0.6709 0.266 2.523 0.012
behavior (OCB)
Intercept 0.4974 0.023 21.369 0.000
Regression Model 3: Trust Organizational trust 0.0427 0.001 59.346 0.000
variables –Turnover Individual trust 0.1922 0.017 11.273 0.000
Intercept 0.1052 0.018 6.001 0.000 Table 3.
Regression Model 4: Cultural Language 0.2865 0.004 72.093 0.000 Logistic regressions of
values variables – Turnover Oral language 0.1887 0.009 20.217 0.000 competencies,
Educational Background 0.0038 0.009 0.412 0.680 commitment, trust,
Culture 0.1588 0.009 18.091 0.000 cultural values and
Religion 0.1379 0.017 7.916 0.000 turnover (Hypotheses
Intercept 0.2546 0.049 5.187 0.000 1–4)

50
High competencies
45 43
Low competencies
40
35
30
30
Turnover rate [%]

25 21
20
15
15
10
5
0 Figure 4.
All roles S u p p o r t r ol e Competencies’ relation
to turnover for all roles
Note(s): For a specific support role, candidates with high competencies compared to a
demonstrate significantly higher turnover rates (43%) than those with low specific role
competencies (21%)

significant interaction with four demographic background scores. As can be seen in the
Radar chart, not only can the effect of competencies on turnover rate vary, but interaction
with the assigned demographic background scores also varies significantly, ranging from a
turnover rate of 7–25% over all roles, to 20–48% for a specific customer support role.
Figure 6 further highlights the above-mentioned nonlinearity observation: while the
support-role is exposed to a slightly higher turnover rate on average (26 vs. 24%), employees
with higher competencies experience a double turnover rate (21 vs. 43%). Nonetheless,
despite this evidence, Figures 4–6 demonstrate different insights, suggesting that turnover,
in certain cases, is clearly affected by granular feature values related to specific person and
job types. We found that the association between competencies, compared to personal
characteristics, and turnover are role-dependent. Thus, as illustrated for the specific support
IJM All roles - Competnecies' effect on turnover rate [%]
43,6
13
40

30

20
1416 10 High Compentencies
16 0 14 Low Compentencies

15

Support role - Competencies' effect on turnover rate [%]

13
50
40
30
20
10
16 0 14 High Compentencies
Low Compentencies

15
Figure 5.
Competencies’ relation Note(s): The contours represent the turnover rate for both high-
to turnover for all roles competencies level (blue) and low-competencies level (red) over four
compared to a
specific role
demographic background scores (where 13 is the lowest score and 16
is the highest)

role, the aforementioned centrality of competencies with regard to turnover may be reversed.
Figure 6 presents another counterintuitive result pertaining to the association between
competencies and turnover from a decision-tree node perspective. The tree nodes indicate
that on the one hand, for this specific role, the turnover rate is 26%, similar to other roles.
However, when further examining specific competencies—leadership and demographic
background in this example—one can detect that the effect is reversed, resulting in higher
turnover rates associated with higher competencies and vice versa (43 and 21%,
respectively).
Our second set of analyses focused on the main effect of commitment on turnover.
Consistent with hypothesis 2, we found a negative main effect of affective commitment on
turnover, measured by a coefficient value of 0.7959. Note that somewhat surprisingly the
voluntary commitment (i.e. OCB), is positively correlated with the turnover rate. One possible
explanation is that employees involved in volunteering activities associated with less desired
accountabilities are more prone to leave. This is an example of a HR analytics’ granular
 Analytics and
machine-
learning
examination

1417

Figure 6.
Competencies’
nonlinear relation to
turnover for a
specific role

finding that is not addressed in hypothesis 2, and may require further theoretical study. In
Table 3, the logistic regression model 2 predicts turnover based on organizational
commitment. The model shows a significant interaction between commitment and
turnover. However, in the context of HR analytics, the effects of commitment on turnover
are contingent upon role. The subsequent model indicates that the effect of commitment on
turnover is personalized and role-dependent.
Figure 7 indicates a different association between commitment and turnover,
suggesting that employee turnover is, once again, a nonlinear phenomenon that is at
times disruptive and counterintuitive. As illustrated in Figure 7, we found that the
relationship between organizational commitment, measured by motivation level of
employee, and turnover is role-dependent. Thus, in the context of the overall analytics,
the relationship between commitment and turnover is negative: a higher commitment level

80
High Commitment 71
70 67

60 Low Commitment
Turnover Rate [%]

50

40 37

30
20
20

10

0
All Roles Mid Level Management Role
Role
Figure 7.
Note(s): Commitment generates a 17% difference (20% vs. 37%) between job Commitment relation
holders with high commitment and job holders with low commitment levels. to turnover for all roles
compared to a
This trend changes when examining a specific mid-level management role, and specific role
as a result, the difference significantly decreases to 4% only (67% vs. 71%)
IJM reduces the turnover rate. Figure 7 indicates that when examining all roles, commitment
43,6 generates a 17% difference (20 vs. 37%) between jobholders with high commitment and low
commitment levels. This trend however is role dependent. For example, when examining a
specific mid-level management role, the trend changes and results in a much lower decrease
in the turnover rate to 4% (67 vs. 71%). In light of several such observations, it seems that
commitment affects turnover in various roles differently.
Our third set of analyses is focused on the main effect of trust on turnover. Consistent with
1418 hypothesis 3, we found a negative main effect of organizational trust on turnover, measured
by a coefficient value of 0.0427. The logistics regression model in Table 3 predicts turnover
rates based on trust. The model shows a significant effect between trust and turnover.
However, note again that in the context of HR analytics, the effects of trust on turnover are
contingent upon role. The subsequent model indicates that the effect of trust on turnover is
personalized and role-dependent.
Figure 8 indicates a different association between trust, and turnover for specific jobs,
suggesting that employee turnover is sometimes not a straightforward phenomenon. The
obtained observations indicate that the negative association between individual trust and
turnover can fluctuate. Figure 8 shows that when examining all roles, jobholders with high
levels of individual trust experience lower levels of turnover than jobholders with lower levels
of trust (24 and 35%, respectively). However, surprisingly, when examining a specific

40

35
35

30

27

25 24
Turnover Rate [%]

20
20
High individual trust
Low individual trust

15

10

0
Figure 8. All Roles Support Role
Trust relation to
turnover for all roles Role
compared to a
specific role
Note(s): Higher trust leads to lower turnover rates (24% vs. 35%), but for a
specific support role, higher trust results in higher turnover rates (27% vs. 20%)
support role, this effect is reversed. In this case, higher trust results in higher turnover rates Analytics and
and vice versa (27 and 20%, respectively). machine-
Our last set of analyses focused on the main effect of cultural values on turnover.
Consistent with hypothesis 4, we found a significant effect of cultural values on turnover. The
learning
logistics regression model in Table 3, predicts turnover rates based on cultural values. Model examination
4 shows a significant relationship between cultural values and turnover. However, in the
context of HR analytics, the effects of cultural values on turnover are contingent upon
cultural background. The subsequent model indicates that the effect of cultural values on 1419
turnover is personalized and culturally dependent.
Figures 9 and 10 present a diverse relationship between cultural values and turnover,
showing, once again, that turnover is a flexible and granular phenomenon in such a context.
Figure 9 illustrates that when examining the male population, a minor 5% effect of the
jobholder’s cultural background is identified. However, when examining a specific
administrative role, one identifies a surprising substantially larger effect (21%) of cultural
background on the turnover rates of male jobholders.
Figure 9 illustrates how, for male jobholders of a specific religion employed in a specific
administrative role, turnover rates are significantly higher (23 and 44%, respectively).

50%

45% 44%

40%

35%

30% 29%
Turnover Rate [%]

25% 24%
23% Cultural background A
Cultural background B
20%

15%

10%

5%

0%
All male roles Security role
Role
Figure 9.
Note(s): When examining all male roles, a minor 5% effect of the cultural Cultural values’
background of the candidate is identified. Surprisingly, when examining a relation to turnover for
a specific role within
specific administrative role, we identified a substantially larger effect (21%) of the male population
cultural background on turnover rates
IJM
43,6

1420

Figure 10.
Cultural values and
gender effects on
turnover for a
specific role

Figure 10 demonstrates the gender and cultural background relation to turnover rates for a
specific administrative role. Cultural background plays a more significant role in the case of
male jobholders, who demonstrate a 21% gap in turnover rate (23 vs. 44%), compared with
female jobholders, who demonstrate only a 9% gap in turnover rate (28 vs. 37%). These
observations emphasize again the importance of analyzing turnover from a granular
perspective using HR analytics tools.

Discussion
The goal of this paper has been to help reinvigorate turnover research by focusing attention
on a specific methodological approach that involves HR analytics and machine-learning tools.
Given recent calls for more empirical work in the field (Allen et al., 2014; Chalutz Ben-Gal,
2019; Guenole et al., 2017), we analyze turnover using a novel analytical approach, thereby
advancing both theory and practice and moving up to the next level of strategic human
resource management (Huselid and Becker, 2005; Malik et al., 2020; Marler and
Boudreau, 2017).
We answer recent calls to shift the analytical mindset and broaden measurement
strategies (Allen et al., 2014) by analyzing turnover using a granular learning approach,
thereby proposing theoretical and practical implications. Using machine-learning models to
explore turnover, we obtained significant insights into what factors may predict turnover, in
what way, and for which types of populations and roles. We found that many of the HR
analytics outcomes are supportive and fully aligned with previous methodological studies in
the literature. Nonetheless, we also found that, in contrast to previous theoretical
assumptions, turnover is a nonlinear and disruptive phenomenon in certain practical
cases. Using data from over 700,000 employees, we found that, consistent with our
hypotheses, competencies, commitment, trust and cultural values are generally associated
with a positive main effect on turnover. However, in more fine-grained analyses, we found
that by applying machine-learning granular models, one is able to show that the effects of
competencies, commitment, trust and cultural values on turnover are often contingent on role,
gender and cultural background. Thus, turnover depends on context as already indicated
(Johns, 2018), therefore, what we know about turnover may vary depending on a variety of
internal and external variables.
Contributions Analytics and
This work contributes to research on turnover and HR analytics. At the same time, it has machine-
potentially important implications for strategic human resources management initiatives in
organizations. We introduce to turnover research a novel approach that integrates the human
learning
resources and data science perspectives on how people leave organizations and how to examination
analyze (and perhaps predict) the turnover phenomenon. Whereas the literature on turnover
has increasingly recognized the need for a deeper analytical understanding of turnover, the
core research method in this field—self-reports—is ill-suited to uncovering fine-grained 1421
organizational variations. The proposed approach demonstrates the benefits of machine-
learning and HR analytics as complementary means of understanding various granular
facets of turnover, how they vary within various contexts, and what factors give rise to them.
We believe that HR analytics based on machine-learning tools to explore turnover—as
with other HR-related challenges—can open up a number of promising pathways for future
research (Chalutz Ben-Gal et al., 2021). For example, how can organizations predict key
human-related processes—e.g. by effective workforce management—that enhance HR
analytics when it is beneficial to do so? Similarly, in the context of organizations that tend to
become more technological and disperse (Chalutz Ben-Gal, 2019; Chalutz Ben-Gal et al., 2021),
can the introduction of HR analytics tools bring about greater effectiveness and help predict
people-related challenges, using a more relevant and fine-grained approach. Beyond
turnover, we see the potential to develop new machine-learning-based measures of other
established people-related constructs such as predicting behaviors (Yarkoni and
Westfall, 2017).
The Results from this study also provide reasons to question the assumption in many
studies in which hypothesis testing is a central research method (Yarkoni and Westfall, 2017).
Unlike prior work relying primarily on various antecedents of turnover (e.g. competencies,
commitment, trust and cultural values), our approach to possibly predict turnover changes
according to role, person and cultural background.
Lastly, this work is related to a growing body of research connecting data science
techniques—such as machine-learning—with the HR field (Angrave et al., 2016; Edwards,
2019; Minbaeva, 2018). We demonstrate how using these techniques can have consequences
on turnover examination and, ultimately, on possible turnover prediction.
Beyond these contributions to scholarly research, this work would appear to have
potentially important implications for strategic workforce planning practices in
organizations. Our research demonstrates the applicability of how to use big data in
organizations in order to predict and prevent turnover in specific roles (Pessach et al., 2020).
For example, our measure of turnover could be used to predict when individuals are at risk of
leaving and introduce alerts that would enable management to direct turnover in the desired
direction as needed. Such diagnostic tools may be especially useful in the context of a globally
distributed workforce and online labor markets, in which workforce planning often becomes
a challenging task (Chalutz Ben-Gal, 2019).

Limitations
While we believe that this work opens promising avenues for research and practice, we are
also aware of its limitations. Although our data is robust and includes a relatively large
number of participants from diverse backgrounds (in fact, we are not aware of previous study
in this context that was based on such a unique and large dataset), it is important to keep in
mind that our empirical setting involves a single organization. Questions thus remain about
how our results would generalize to other contexts, such as specialized workforce
characteristics and organizations in which the focus is on short-term and freelance
employment.
IJM It is challenging to compare and contrast this work’s results with those of previous
43,6 turnover-related meta-analyses. Methodological restrictions elicit the “apples and oranges”
comparison. Nonetheless, we attempted to overcome this limitation by utilizing a novice and
robust HR analytics machine-learning method that enabled us to reveal otherwise
undetectable aspects of turnover.
Lastly, it is evident that as the use of digital technologies—e.g. HR analytics, machine-
learning and artificial intelligence—become part of strategic HR, they may alter the way in
1422 which managers allocate their attention to organizational challenges (Kadosh and Chalutz
Ben-Gal, 2021). While it is optimistic to assume that the quality of organizational and
managerial processes will rise with the support of big data and digital technology tools, it is
not yet clear which dimensions will be affected and where new problems or biases may
emerge. Scientific evidence suggests that there may be an increase in the abuse of data in
various forms, such as data manipulation, data fairness and ethical considerations, data
secrecy abuse and illegal data handling. Therefore, future studies should focus on studying
how digital technologies may contribute to better decision-making, in line with this work.

Conclusion
In summary, this study demonstrates the value of HR analytics and machine-learning
methods to the study of turnover. Relative to prevailing approaches to studying turnover, HR
analytics and machine-learning tools offer a different way of understanding turnover,
including a granular view of turnover antecedents, and reveal the role of unforeseen
measures in explaining the turnover phenomenon.

Note
1. We use the term “cultural values” based on Hofstede (1984).

References
Allen, D.G., Hancock, J.I., Vardaman, J.M. and Mckee, D.N. (2014), “Analytical mindsets in turnover
research”, Journal of Organizational Behavior, Vol. 35 No. S1, pp. S61-S86, doi: 10.1002/job.1912.
Angrave, D., Charlwood, A., Kirkpatrick, I., Lawrence, M. and Stuart, M. (2016), “HR and analytics:
why HR is set to fail the big data challenge”, Human Resource Management Journal, Vol. 26,
pp. 1-11, doi: 10.1111/1748-8583.12090.
Ben-Gal, I. (2008), Bayesian Networks, Encyclopedia of Statistics in Quality and Reliability, Vol. 1, Wiley
Online Library.
Bijlsma-Frankema, K., de Jong, B. and van de Bunt, G. (2008), “Heed, a missing link between trust,
monitoring and performance in knowledge intensive teams”, International Journal of Human
Resource Management, Vol. 19 No. 1, pp. 19-40, doi: 10.1080/09585190701763800.
Braun, M.T., Kuljanin, G. and DeShon, R.P. (2018), “Special considerations for the acquisition and
wrangling of big data”, Organizational Research Methods, Vol. 21 No. 3, pp. 633-659, doi: 10.
1177/1094428117690235.
Chalutz Ben-Gal, H. (2019), “An ROI-based review of HR analytics: practical implementation tools”,
Personnel Review, Emerald Publishing, Vol. 48 No. 6, pp. 1429-1448, doi: 10.1108/PR-11-2017-0362.
Chalutz Ben-Gal, H., Avrahami, D., Pessach, D. and Singer, G. (2021), “A machine learning
examination of turnover: hidden patterns and new insights”, in Academy of Management
Proceedings, Academy of Management, Briarcliff Manor, NY, Vol. 2021 No. 1, p. 12152.
Chalutz Ben-Gal, H. and Tzafrir, S.S. (2011), “Consultant-client relationship: one of the secrets to
effective organizational change?”, Journal of Organizational Change Management, Vol. 24 No. 5,
pp. 662-679, doi: 10.1108/09534811111158912.
Corritore, M., Goldberg, A. and Srivastava, S.B. (2019), “Duality in diversity: how intrapersonal and Analytics and
interpersonal cultural heterogeneity relate to firm performance”, Administrative Science
Quarterly, Vol. 65 No. 2, pp. 359-394, doi: 10.1177/0001839219844175. machine-
de Oliveira, L.B., Cavazotte, F. and Alan Dunzer, R. (2019), “The interactive effects of organizational
learning
and leadership career management support on job satisfaction and turnover intention”, examination
International Journal of Human Resource Management, Vol. 30 No. 10, pp. 1583-1603, doi: 10.
1080/09585192.2017.1298650.
Edwards, M.R. (2019), “HR metrics and analytics”, in Thite, M. (Ed.), E-HRM: Digital Approaches, 1423
Directions and Applications, Routledge.
Elfenbein, H.A. and O’Reilly, C.A. (2007), “Fitting in: the effects of relational demography and person-
culture fit on group process and performance”, Group and Organization Management, Vol. 32
No. 1, pp. 109-142, doi: 10.1177/1059601106286882.
Fineman, S. (2003), “Emotionalizing organizational learning”, Blackwell Handbook of Organizational
Learning and Knowledge Management, pp. 557-574, available at: http://search.ebscohost.com/
login.aspx?direct5true&db5buh&AN511009422&site5ehost-live.
Garcia-Arroyo, J. and Osca, A. (2019), “Big data contributions to human resource management: a
systematic review”, The International Journal of Human Resource Management, Vol. 32 No. 20,
pp. 4337-4362.
Goldberg, A. (2015), “In defense of forensic social science”, Big Data and Society, Vol. 2, p. 2,
2053951715601145, doi: 10.1177/2053951715601145.
Guenole, N., Ferrar, J. and Feinzig, S. (2017), The Power of People: Learn How Successful Organizations
Use Workforce Analytics to Improve Performance, FT Press, Amazon.Com.
Hofstede, G. (1984), Culture’s Consequences: International Differences in Work-Related Values, SAGE,
Newbury Park, California, Vol. 5.
Hom, P.W., Lee, T.W., Shaw, J.D. and Hausknecht, J.P. (2017), “One hundred years of employee
turnover theory and research”, Journal of Applied Psychology, Vol. 102 No. 3, pp. 530-545, doi: 10.
1037/apl0000103.
Huselid, M.A. and Becker, B.E. (2005), “Improving human resources’ analytical literacy: lessons from
moneyball”, The Future of Human Resource Management, John Wiley & Sons, pp. 278-284.
Johns, G. (2018), “Advances in the treatment of context in organizational research”, Annual Review of
Organizational Psychology and Organizational BehaviorAnnual Reviews, Vol. 5, pp. 21-46.
Kadosh, E. and Chalutz Ben-Gal, H. (2021), “AI acceptance in primary care during COVID-19: a two-
phase study of patients’ perspective”, in Academy of Management Proceedings, Academy of
Management, Briarcliff Manor, NY, Vol. 2021 No. 1, p. 13461.
Krausert, A. (2017), “HR differentiation between professional and managerial employees: broadening
and integrating theoretical perspectives”, Human Resource Management Review, Vol. 27 No. 3,
pp. 442-457, doi: 10.1016/j.hrmr.2016.11.002.
Madariaga, R., Oller, R. and Martori, J.C. (2018), “Discrete choice and survival models in employee
turnover analysis”, Employee Relations, Emerald Publishing, Vol. 40 No. 2, pp. 381-395, doi: 10.
1108/ER-03-2017-0058.
Malik, A., Budhwar, P. and Srikanth, N.R. (2020), “Gig economy, 4IR and artificial intelligence: rethinking
strategic HRM”, Human and Technological Resource Management (HTRM): New Insights into
Revolution 4.0, Emerald Publishing, pp. 75-88, doi: 10.1108/978-1-83867-223-220201005.
Marler, J.H. and Boudreau, J.W. (2017), “An evidence-based review of HR analytics”, International Journal
of Human Resource Management, Vol. 28 No. 1, pp. 3-26, doi: 10.1080/09585192.2016.1244699.
Mayer, R.C., Davis, J.H. and Schoorman, F.D. (1995), “An integrative model of organizational trust”,
Academy of Management Review, Vol. 20 No. 3, pp. 709-734, doi: 10.5465/amr.1995.9508080335.
IJM Meyer, J.P. and Parfyonova, N.M. (2010), “Normative commitment in the workplace: a theoretical
analysis and re-conceptualization”, Human Resource Management Review, Vol. 20 No. 4,
43,6 pp. 283-294, doi: 10.1016/j.hrmr.2009.09.001.
Minbaeva, D.B. (2018), “Building credible human capital analytics for organizational competitive
advantage”, Human Resource Management, Wiley Online Library, Vol. 57 No. 3, pp. 701-713.
Organ, D.W. (1988), Organizational Citizenship Behavior: The Good Soldier Syndrome, Lexington
Books/DC Heath and Com, Lexington, MA.
1424
Pessach, D., Singer, G., Avrahami, D., Ben-Gal, H.C., Shmueli, E. and Ben-Gal, I. (2020), “Employees
recruitment: a prescriptive analytics approach via machine learning and mathematical
programming”, Decision Support Systems, Elsevier, Vol. 134, 113290.
Purba, D.E., Oostrom, J.K., Born, M.P. and Van Der Molen, H.T. (2016), “The relationships between
trust in supervisor, turnover intentions, and voluntary turnover: testing the mediating effect of
on-the-job embeddedness”, Journal of Personnel Psychology, Vol. 15 No. 4, pp. 174-183, doi: 10.
1027/1866-5888/a000165.
Schwartz, S.H. and Bilsky, W. (1990), “Toward a theory of the universal content and structure of
values: extensions and cross-cultural replications”, Journal of Personality and Social Psychology,
Vol. 58 No. 5, pp. 878-891, doi: 10.1037/0022-3514.58.5.878.
Tzafrir, S. and Dolan, S. (2004), “Trust me: a scale for measuring manager-employee trust”,
Management Research: Journal of the Iberoamerican Academy of Management, ISSN:
1536-5433.
Tzafrir, S.S., Chalutz Ben-Gal, H. and Dolan, S.L. (2012), “Exploring the etiology of positive
stakeholder behavior in global downsizing”, Downsizing: Is Less Still More?, pp. 389-417, doi: 10.
1017/CBO9780511791574.019.
Yarkoni, T. and Westfall, J. (2017), “Choosing prediction over explanation in psychology: lessons from
machine learning”, Perspectives on Psychological Science, Sage Publications Sage CA: Los
Angeles, CA, Vol. 12 No. 6, pp. 1100-1122.

Corresponding author
Hila Chalutz Ben-Gal can be contacted at: hilab@afeka.ac.il

For instructions on how to order reprints of this article, please visit our website:
www.emeraldgrouppublishing.com/licensing/reprints.htm
Or contact us for further details: permissions@emeraldinsight.com