Employee turnover presents a significant threat to business sustainability and is a dynamic challenge faced by businesses globally. Despite the economic downturn, turnover remains one of the most relevant and significant issues in human resource management (Robison 2010; Vance, Vaiman & Andersen 2009). The consequences of turnover are likely to be particularly severe in cases where experience is a critical prerequisite for successful work, where replacement is expensive and where turnover is disruptive and demoralising to the remaining workforce (Allen, Bryant & Vardaman 2010).

There are significant costs associated with turnover that include direct recruitment costs incurred whilst searching for a replacement, lost productivity on the part of the departing employee, lost labour between the time the individual quits and a replacement is hired and reduced productivity of the new hire whilst learning the job (Cascio & Boudreau 2008). Organisations often make a significant investment in training and developing their employees and this investment is also lost when employees leave (Mello, 2011). Estimates of the overall cost of turnover range from 90% to 200% of the departing employee’s salary (Cascio & Boudreau 2008).

Other consequences of turnover include a reduction in customer satisfaction, loss of organisational knowledge and ultimately reduced profit (Harter, Schmidt & Keyes 2003). Excessive employee turnover can hurt the overall productivity of a firm and is often a symptom of other difficulties (Iqbal 2010).

In South Africa the demand for skilled employees has continued to increase due to an ageing population and growth in the local economy (Journal of E-Governance 2010). Organisations are forced to compete to attract and retain skilled employees in an environment characterised by a burgeoning skills deficit. The impact of affirmative action and employment equity legislation exacerbates the turnover problem. The retention of skilled employees has thus become a strategic imperative which is critical to business success.

Various explanations have been postulated to explain the growing skills deficit in South Africa, such as rapid economic growth and an increasing need for infrastructure, buildings and electricity. The changing age demographics within the country also play a role. As more baby boomers (born 1946–1964) start retiring, the quantity and quality of available skills becomes a greater problem (Rappaport, Bancroft & Okum 2003).

The Global Competitiveness Report 2011–2012 ranks South Africa 50th in the ranking of competitiveness out of 59 nations, indicating a decline from the previous year. The International Institute for Management Development (IMD) generated the rankings using 331 criteria ranging from gross domestic product growth and unemployment, to the number of Internet users and the price of local cell phone calls. Other criteria included the availability of skilled employees, government regulation, availability of venture capital and various qualitative factors. The IMD report shows that a critical component of competitiveness includes having the appropriate skills quota across a range of occupations and professions as well as the skills to drive leadership in organisations (Journal of E-Governance 2010).

Consequently, in their retention strategies, organisations use both targeted and untargeted approaches in managing employee turnover. Untargeted approaches rely on generic organisational practices to increase organisational commitment and retain employees (Westerman & Yamamura 2007). These approaches include increasing compensation and benefits, promotions, opportunities to learn, special assignments and status-driven incentives. Other untargeted non-monetary retention strategies include increased flexibility and work-at-home options, flexible control over work schedules and additional opportunities to develop skills and knowledge during work time or through employer-funded educational programmes (Westerman & Yamamura 2007).

Targeted approaches rely on identifying employees who are likely to leave and then providing them with either a direct incentive or a customised plan to ensure that they stay. Effective retention strategies hinge on reliable, timely intelligence about prospective leavers (Lyness & Judiesch 2001; Trevor, Gerhart & Boudreau 1997).

Targeted and untargeted retention approaches aimed at managing voluntary turnover form part of turnover risk management and are becoming an increasingly important strategy to ensure organisational stability, sustainability and success.

Variables that generally predict employee turnover have been summarised in three meta-analytic studies (Cotton & Tuttle 1986; Griffeth, Hom & Gaertner 2000; Tett & Meyer 1993). The most recent of these (Griffeth et al. 2000) categorised the predictors as:

Employee demographics.

Demographic characteristics have been advanced in several models as predictors of turnover (Bal et al. 2011; Deumert 2010), but few studies have examined the demographic-turnover correlation in isolation. Other organisational variables, such as job satisfaction, frequently tend to form part of the studies.

Turnover literature has been led by research on how work attitudes (for example commitment and job satisfaction) influence turnover (Mossholder, Setton & Henagan 2005). Whilst this contribution to the literature has been significant, it has been at the expense of studying less traditional turnover variables, such as demographic variables.

Wocke and Heymann (2012) investigated the role of demographic variables as extended factors in the voluntary turnover process in South Africa. Their findings indicate that various demographic variables (for example age, race and gender) influence whether pull or push factors are cited in the turnover process. They also found that level of education has a stronger relationship with employee mobility than race. Based on their results they asserted that employee demographic variables should be considered in the extension of contemporary turnover models.

Vandenberg and Nelson (1999) indicate that identifying predictive precursor variables to turnover is important for the understanding and management thereof. This illustrates the need to develop a predictive model to tackle the problem of employee turnover.

Employee turnover is an important research topic in the fields of organisational behaviour, human resources management and labour economics. Consequently, there is an increasing need to understand the major variables affecting turnover. Such understanding will hopefully lead to the advancement of turnover risk management, which is becoming an important strategy to ensure organisational stability and promote the effective retention of employees.

The relationship between demographic variables and voluntary turnover is a topic of interest for a number of researchers. Although a cursory examination of the literature suggests that consensus exists regarding the nature of this relationship, a more detailed analysis reveals significant disagreement. This lack of consensus leads to different conclusions and limits theoretical development within this research space.

The purpose of this research was to contribute to the practice of turnover risk management by proposing an approach and constructing a model that is able to predict employee turnover solely based on demographic characteristics readily available in a typical human resource information system. It is believed that organisations can use such an approach to model predictors of turnover without having to collect primary data and so efficiently identify employees that are potential flight risks. Once these employees are identified, decisions can be made as to whether and to what extent retention measures need to be put in place to try and retain such employees.

The primary objective of this research was to create a model that could be used to predict employee turnover by using demographic variables readily available to organisations. It was important in this study to make use of data that are already available in organisations, rather than consider variables for which data specifically needs to be collected. This would essentially be a stand-alone model for determining the probability of whether an employee will stay with or leave an organisation. The research also aimed to provoke greater thinking and reflection on the measurement of employee turnover.

The researchers hypothesised that demographic variables could be used to build a statistically significant model to predict and identify employees at risk of turnover.

An exploratory research design was employed. Secondary data weres extracted from an existing human resources database at a general insurance organisation and analysed.

A large insurance organisation, headquartered in Cape Town, South Africa, was selected to serve as a case study. The organisation has approximately 60 branches throughout South Africa and Namibia. The study utilised secondary data collected on employees who were permanently employed with the organisation as of 2008 and who had voluntarily left the organisation between 2008 and 2011. This group represented the sample. Employees who had left the organisation on an involuntary basis were excluded from the study. A total of 2592 employees were included in the sample.

The data were collected from the organisation’s human resources information system. Company records included an observation of the database of the years 2008–2011. The organisation provided 22 data fields for each individual employee record, of which only 14 variables were selected for analysis. The input data contained the demographic characteristics of each employee. The dependent variable was an indicator of whether the employee had voluntarily terminated their employment. Each employee was assigned an identification number and entered into a data file. The SAS program was used to conduct the statistical analysis.

Exploratory data analysis was conducted in order to summarise the relationships of the independent variables and eliminate those that did not relate to voluntary employee turnover.

It would have been impossible to gain consent from all participants in the study; therefore, the custodian of organisation employee data (the Head of Human Resources), was approached and permission requested to gain access to the data.

Permission to conduct this research was granted with the following confidentiality provisions:

The name of the organisation would not be indicated, only the industry.

The Commerce Ethics in Research Committee at UCT further granted ethical permission to conduct this study.

This study considered only limited variables when assessing employee turnover. Critical variables, for example, qualification and salary were excluded. This exclusion was deliberate as the research aimed to isolate the impact of demographic variables to validate their value in predicting turnover. The sample was also limited to one general insurance organisation. It may not be appropriate to generalise the findings across other populations or settings; however, the sample can be viewed as a representative case typical of many other organisations in the same industry.

Table 1 shows the distribution of the sample (n = 2592) by current age, gender, race, language and marital status.

From Table 1, it can be seen that the majority of the employees (n = 871 or 34%) were in the age group 40–50 years, whilst 31% (n = 785) were in the age group 40–50 years. The numbers indicated a mature workforce with the under-30s representing 15% of the sample.

The majority of employees were white people (52%). The ratio of white people to black people (mixed race, Indian and black people combined) employees was 52:48; in an organisation where transformation is one of the strategic drivers, the issue of retention of non-white employees in a highly skilled industry with notable skills scarcity is critical to organisational success.

Afrikaans (52%) and English (36%) were the most common languages used by employees. Most of the sample was married (57%) with a very small number of employees being divorced (9%) or widowed (2%).

Since turnover is a dichotomous variable, a logistic regression model was used in this study (Quinn, Rycraft & Schoech 2002). The forward selection technique began with no variables in the model. Explanatory variables were subsequently added to the model one at a time. At each step, each variable that was not already in the model was tested for inclusion. The likelihood ratio test was used to test if adding a variable was of significant value. The p-values of the test were quoted. Automated model building software packages could not be used because some variables were collinear. Collinear variables, which are not eligible for inclusion in the model, have crosses in place of their statistics. The most significant of these variables were added to the model.

The AIC and SC were used to test how much value a variable added to the model and if its inclusion was warranted. The results of these tests are illustrated in Table 2 and Table 3 respectively.

Step 1 of the model selection process showed that the cost centre band was clearly the most powerful variable. The corresponding small p-value and AIC were much smaller than those calculated for the other variables.

In step 2, the business unit variable could not be fitted to the model because cost-centre had already been included and the two are collinear. This iteration process showed that the number of years of service was the most powerful variable despite the fact that it had the second lowest p-value. The SC calculated for the inclusion of years of service was the smallest and so warranted the inclusion of this variable rather than that of age.

In step 3, the age variable could not be fitted to the model because years of service had already been included and the two are highly correlated. This iteration process showed that the performance score variable was the most powerful variable. The corresponding small p-value and SC were much smaller than those calculated for the other variables.

In step 4, the inclusion of an interaction term was tested (refer to Table 4).

Whilst several variable interactions were tested, the most powerful turned out to be the interaction between the number of dependants and the years of service with a p-value of 0.087 and SC of 1310.

Tests of the statistical significance of each independent variable were conducted. The Wald chi-square test (and its associated p-value) are shown along with the parameter estimate and standardised parameter estimate in Table 5.

To test the prediction, the chi-square test statistic for the predictor performance score was 3.59 with an associated p-value of 0.0581. If the alpha level had been set to 0.05, it would have failed to reject the null hypothesis and concluded that the difference in performance has not been found to be statistically different for employees who leave and those who stay.

The results of the tests can be found in Tables 8, 9 and 10 below.

The pseudo R-squared for the model was low but the other tests (Table 8 and Table 9) seem to indicate that this was a result of its shortcomings in the context of logistic regression.

The literature supports the finding that age influences employee decisions to leave an organisation. For example, mature workers are less likely than younger ones to intend to leave an organisation. Younger employees have higher turnover rates as a result of shifting career paths, flexibility to relocate and fewer family responsibilities and financial obligations (Lewis 1991). Generational differences also encourage younger employees to change jobs and sector of employment frequently after earning their degrees (Peralta & Stark 2006). As Generations X and Y place less value on stability and benefits, there is a greater likelihood of turnover amongst these groups (Lewis & Frank 2002).

In the current logistic regression study, no assumptions were made about the distributions of the explanatory variables. The explanatory variables should not be highly correlated with one another because this could cause problems with estimations. The age variable could not be fitted to the model because years of service was already included and the two are most likely highly correlated. The differences in turnover attributable to age were clear. Employees younger than 35 years had higher turnover rates than their older colleagues, with the propensity to leave actually increasing with age amongst the younger employees. The main argument for this observation was that the available time to amortise the costs related to a change in employment declines with age, thus making a job change more risky (Sousa-Poza & Henneberger 2004).

According to traditional research, years of service has consistently been found to be negatively related to turnover (Griffeth et al. 2000; Steel & Ovalle 1984; Steers 1977). In this study the highest turnover was experienced in the first four years of employment. New employees take time to adjust to their work environment. At the same time, they are still evaluating whether their decision to join the organisation was the right career choice. In addition, in the first few years of employment, new employees would not have financially invested in the organisation. Their pension fund and other benefits would not have accrued to the point that leaving the organisation would disadvantage them financially. Therefore, job mobility becomes easier.

The research confirmed that after every year of service, voluntary turnover decreases. A p-value of 0.0309 meant years of services was included in the model. Whilst age and years of service are positively correlated, they are not identical constructs (Ng & Feldman 2009). Nonetheless, older workers with high years of service may have particularly strong financial reasons to remain with their current employers. Good compensation benefits are offered to employees with long years of service in order to retain them. Ng and Feldman (2009) suggest that these rewards discourage employees from seeking alternative opportunities. Whilst older workers are less likely to leave their organisations in general, those with long years of service are especially less likely to do so because of financial incentives.

Job tenure (referred to as years of service) also seems to have a strong negative influence on turnover (Blau & Kahn 1981; Cotton & Tuttle 1986; Lambert, Hogan & Barton 2001; Lewis 1991; Sorenson 2000). Turnover is greatest at the earliest stages of employment, but it declines rapidly over the first 5 years and then more slowly up to about 15 years of service (Lewis 1991). A principal explanation for this trend is that social interaction in the workplace tends to engender affinity and loyalty toward the organisation and its members, thus reducing the propensity for turnover (Sorenson 2000). As an employee remains with an organisation over time, they will be motivated to respond to changes within it by exiting, voicing concern or remaining loyal. Once an employee has a certain number of years of experience in employment, they may refrain from leaving or retiring early out of consideration for the potential loss in retirement earnings.

Cost centre refers to an organisation unit in which all costs related to a group of people are managed. It could be a specific department but also within a department, a few cost centres may exist. Bands were used to classify cost centres with the lowest and highest turnover on a scale of 0 (lowest) to 7 (highest). The inclusion of cost centre in the model was unexpected but not unexplained. Table 11 shows which business unit contributes most to a particular (turnover) band.

As broker distribution has the largest number of cost centres, it will almost always be the largest contributor to turnover due to the size of the business unit (N = 1066). Table 11 also shows that within business units (e.g. business change), there is very little likelihood of turnover. Table 11 indicates that there is not much turnover in all its cost centres falling into bands 1 and 2. Broker distribution, claims services and specialist business appear to have cost centres that have high turnover. Within these business units, there are specific cost centres that have a higher predictive turnover probability than other cost centres (e.g. band 7).

These cost centres are based in Gauteng. Possible explanations could include all of the traditional turnover factors, including leadership, job satisfaction and commitment, worker engagement, employee benefits and the nature of the work. This study suggests that people working in Gauteng are more susceptible to turnover than people living anywhere else, but it is suggested that this assertion needs to be tested in future research. This could be explained by the fact that Johannesburg offers greater access to alternative career opportunities as large South African and international offices are based there.

The findings suggest that voluntary turnover can be predicted based on performance (p-value = 0.0581). The performance-turnover relationship (i.e. positive, negative or no relationship) is supported by the literature. The literature advises that the performance-turnover relationship should consider performance slopes when predicting turnover (Sturman & Trevor 2001). Alternatively, other researchers suggest that the performance-turnover relationship may be of greater importance in the intermediate stages of employment (Kanfer, Crosby & Brandt 1988). In the current research, turnover was highest within the first four years of employment, which makes this a crucial period in retention.

The findings of this study strongly support the existence of a positive relationship between performance and turnover: the higher the level of performance, the lower the likelihood of turnover. The performance-turnover relationship indicates that the variable is useful in the final model but that its effect is likely to be marginal.

The last component of the model is the relationship between years of service and number of dependants. In order to test for moderating effects, this study adopted a regression procedure. The method is that both categorical (e.g. years of service) and continuous (e.g. organisational years of service) moderator variables can be included. Employees who have same number of years of service and differences in the number of dependants show an equal likelihood of turnover.

In the data exploration phase of the study, there was clear support for the relationship between demographic variables and turnover. However, in the predictive model, there was less support for the demographic variable-turnover relationship.

The study did not consider several variables that have been identified in the management and labour economics literature to be important determinants of turnover. These include wages, wage growth, education status and prior mobility (Ledolter & Power 1984). Of these, wages and wage growth are probably most important for actuarial applications.

A primary limitation of this research was that the model was one-dimensional. It was developed using only data available internally within an organisation. The data did not directly capture influences from the macro and market environments that might have had an impact on employee turnover. Neither did the research focus on any of the traditional turnover models.

Other demographical data and test results may be considered to improve the accuracy of the prediction or generate other potentially useful rules for this model. Further research is suggested to apply this basic model to a greater number of organisations in similar industries. However, in this application, the following should be considered:

Analysis using survival analysis techniques, or duration analysis, could enable more precise prediction.

Alternative data mining techniques such as neural networks could be studied in future research to compare various approaches and these techniques may be integrated for better exploration of complex interrelationships amongst the input personnel variables and target work behaviours (Sexton et al. 2005).