Optimal Employee Turnover Rate: Theory and Evidence*
, Kam-Ki Tang
and Yi-Ping Tseng
Melbourne Institute of Applied Economic and Social Research, University of Melbourne
School of Economics, The University of Queensland & Research School of Pacific and
Asian Studies, The Australian National University
Melbourne Institute Working Paper No. 19/02
ISSN 1328-4991 (Print)
ISSN 1447-5863 (Online)
ISBN 0 7340 1543 7
*This paper is the result of work being undertaken as part of a collaborative research
program entitled: “The Performance of Australian Enterprises: Innovation, Productivity
and Profitability”. The project is generously supported by the Australian Research
Council and the following collaborative partners: the Australian Tax Office, the
Commonwealth Office of Small Business, IBIS Business Information Pty Ltd, the
Productivity Commission, and the Victorian Department of State Development. The
views expressed in this paper, however, represent those of the authors and are not
necessarily shared by the collaborative partners.
Melbourne Institute of Applied Economic and Social Research
The University of Melbourne
Victoria 3010 Australia
Telephone (03) 8344 3701
Fax (03) 8344 5630
WWW Address http://www.melbourneinstitute.com
This paper investigates the quantitative effects of employee turnover on firms’ productivity.
The Australian Business Longitudinal Survey 1995-98, a unique survey providing firm level
data on both production and employee turnover, is used as the data source. Theoretical
studies have advocated that firm specific human capital and job matching to be the two
major, but competing, mechanisms through which turnover affects productivity. Our results
indicate that the effect of job matching dominates when turnover is “low,” while the effect of
firm specific human capital dominates when turnover is “high.” We identify that the optimal
turnover rate − the rate that maximises productivity, controlling for other factors − is about
0.3, well in excess of the sample mean. The finding suggests that further increasing the
flexibility of employment arrangement for small and medium Australian enterprises could
yield substantial productivity gains.
Key words: Employee turnover, productivity, firm specific human capital, job matching,
panel data, unobserved effects, instrumental variable estimation.
It is widely acknowledged in the business community that human resources are a valuable
asset to firms (see, for example, Business Times, 2000, and Business Asia, 1999). Therefore,
it is logical to assume that the flow of this valuable asset − employee turnover − plays a
crucial role in firm performance. In fact, firms (and employees) are burdened with turnover
problems in both good and adverse economic climates.
During economic upturns, employee churning represents one of the greatest difficulties in
business management. For instance, during the “new economy” boom in the U.S., nearly a
quarter of workers were reported to have average tenure of less than a year (Economist
On the other hand, during economic downturns, trimming operating costs through job
retrenchment in order to maintain the firm’s share value is a typical phenomenon.
Nevertheless, downsizing is not a painless option for firms, as they are likely to suffer
adverse consequences, such as low levels of morality and loyalty amongst the remaining
employees. Moreover, firms also bear the risk of not being able to quickly re-establish the
workforce should the economy rebound more swiftly than anticipated.
Given this, employee turnover has been extensively researched across a number of
disciplines, including psychology, sociology, management and economics. Each discipline
has its own focus and, accordingly, employs different research methodologies.
Psychologists and sociologists are generally interested in motivations behind quitting, such as
job satisfaction, organisational commitment and job involvement (see, for example, Carsten
and Spector, 1987, and Muchinsky and Tuttle, 1979). Empirical work in these fields typically
involves case studies using survey data of individual firms or organisations.
In the discipline of management study, high staff turnover has been of great and continuous
concern (see, for example, the symposium in Human Resource Management Review, 9(4)
1999, and Mok and Luk 1995). Similar to the practice in psychology and sociology,
researchers heavily draw on event, or case, studies. Maintaining employee longevity appears
to be especially difficult in the following sectors: information and communication; restaurant
The problem was experienced by high-tech industries as well as the low-tech ones, such as retailing, food
services and call centres.
and catering; warehouse; and retail banking. While reducing employee turnover is a
managerial objective for some firms, the converse is true for others. For example, legal
restrictions and obligations in recruitment and dismissal could prohibit firms from
maintaining a flexible workforce size. This situation is more common in unionised sectors
(Lucifora 1998). The industrial reforms and privatisation in countries such as Australia were
aimed, at least partly, at increasing the flexibility of labour markets.
In contrast, economists focus mainly on the implications of turnover on unemployment. A
strand of matching theories has been developed extensively to explain equilibrium
unemployment, wages and vacancies (Lilien 1982; Lucas and Prescott 1974). National
aggregate time series data are typically employed in this line of research (For recent surveys
on matching theories and their applications see Petrongolo and Pissarides, 2001, and the
symposium in Review of Economic Studies, 61(3) 1994).
Despite turnover being considered crucial to human resource management and production,
there is limited quantitative research on the effect of turnover on firms’ productivity.
omission is probably due to the lack of firm level data on both production and turnover.
Moreover, firm level longitudinal data are typically restricted to individual organisations,
prohibiting researchers from drawing general conclusions.
Utilizing a recently released firm-
level panel data set, based on the Australian Business Longitudinal Survey (BLS), this paper
is therefore able to provide a new dimension to the literature. The BLS data provide an
objective measure of value-added, which is comparable across firms operating in a broad
spectrum of industries. Using data on factor inputs, we can therefore estimate total factor
productivity across industries, and thus investigate the impact of employee turnover on firms’
McLaughlin (1990) examines the relationship between turnover type (quit or layoff) and economy-wide
general productivity growth, but not productivity of individual firms. Shepard et al. (1996) make use of
survey data to estimate the total factor productivity of the pharmaceutical industry; nevertheless, their study is
only concerned with the effect of flexible working hours and not turnover.
For instance, Borland (1997) studies the turnover of a medium-size city-based law firm, Iverson (1999)
examines voluntary turnover of an Australian public hospital, and Glenn, McGarrity and Weller (2001) focus
on major league baseball in the U.S. However, all three studies do not cover the production aspect of the
Using this data set, we establish that employee turnover has a statistically significant and
quantitatively large, but more importantly, non-linear, effect on productivity. We identify the
optimal turnover rate − the rate that maximises productivity, keeping other factors constant −
to be around 0.3. In this paper, the employee turnover rate is defined as the average of total
number of employees newly recruited and departed within a period, divided by the average
number of employees over the period. In other words, the highest productivity appears where
about 30 per cent of total employees changed over the one-year period. The results also
suggest that a substantial productivity gain could be achieved by bringing the employee
turnover rate closer toward the optimal level. As an illustration, if the rate of turnover
increases from the median value of 0.13 to the optimal level, productivity is predicted to
increase by 1.6 per cent. These results could be instrumental in developing human resource
and labour market policies. For instance, if the observed turnover rate is substantially below
the estimated optimal rate, it might reflect the fact that the labour market is too rigid and
deregulation may be warranted.
The rest of the paper is structured as follows. Section 2 reviews two main contending theories
about the linkage between employee turnover and productivity, and formulates the concept of
the optimal turnover rate. In Section 3 the econometric model and the data are briefly
described. Section 4 presents the empirical results and Section 5 concludes.
2. Theories of Turnover and Productivity
There are two main theories on how employee turnover affects firms’ productivity. The first
theory is the firm specific human capital (FSHC) theory, pioneered by Becker (1975). It is
asserted that if firms need to bear the cost of training, their incentives to provide staff training
will be lowered by high quitting rates. The incentive will be even weaker when firm specific
and general training are less separable, as employees have lower opportunity costs of quitting
(Lynch 1993). Therefore, firms’ productivity falls as turnover increases. Even if FSHC is
Recent theoretical studies indicate that employee turnover is important to individual as well as national
welfare. Examples include Chang and Wang (1995) and Cooper (2001); the former emphasises on the channel
of human capital investment, while the latter on that of R&D.
bred through learning-by-doing, its accumulation remains positively related to employees’
tenure. As a result, a higher turnover rate will still lead to lower productivity.
In addition to the direct loss of human capital embodied in the leavers, there are other
negative impacts of turnover on productivity. First of all, a certain amount of output will be
forgone during the vacant and training period. The administrative resources used in
separation, recruitment and training could have been invested in other aspects of the
Moreover, high employee turnover could adversely affect the morale of
the organisation. Using a controlled experiment, Sheehan (1993) records that the leavers alter
the perceptions of the stayers about the organisation and therefore negatively affect its
productivity. As a consequence, warranted (from an employer’s perspective) but involuntary
job separation could trigger unwarranted voluntary employee departure − a snowball effect.
On the other side of the debate, is the job matching theory established by Jovanovic (1979a;
1979b) and Burdett (1978). The key insight of this theory is that firms will search for
employees and job seekers will search for firms until there is a good match for both parties.
However, the conditions for an optimal matching may change over time, leading to
continuous reallocation of labour. For instance, a firm that has upgraded its production
technology will substitute skilled for unskilled labour (for a recent survey on this topic, see
Ahn, 2001). Moreover, established firms also need ‘new blood’ to provide fresh stimulus to
One kind of FSHC is the standardised norm practiced in individual firms. Such organisational norm is a means
to reduce transaction costs within firms as it provides a behavioural guidance to employees (Mintzberg 1979).
If it takes a significant amount of time for new employees to learn the norm (i.e. the socialisation process is
long), high turnover will increase the internal transaction costs and, thus, reduce productivity. A similar
concept is suggested by Nelson and Winter (1982) in that organisational memories are maintained as routines,
which is crucial to production efficiency. However, such memories will be damage d by turnover.
It has been reported that the cost of losing an employee is between half to one and a half times the employee’s
annual salary (Economist 2000).
During the economic downturn in the U.S. in 2001, executives in Charles Schwab and Cisco were reportedly
cutting down their own salaries and setting up charitable funds for laid off staff in order to maintain the
morale of the remaining employees (Economist 2001). Both companies’ efforts were apparently well
received. Fortune (2002) ranked Cisco and Charles Schwab as the 15
best companies to work for,
respectively, during 2001 despite Cisco was reported to lay off 5,500 staff while Charles Schwab 3,800 staff.
the status quo. On the other hand, a worker who has acquired higher qualifications via
education, training, or learning-by-doing may seek a better career opportunity.
Regular employee turnover helps both employers and employees avoiding being locked in
sub-optimal matches permanently. For instance, a study by the U.S. Department of Labor
estimates the cost of a poor hiring decision to be 30 percent of the first year’s potential
earning, and even higher if the mistake is not corrected within six months (Abbasi and
Hollman 2000). Furthermore, Borland (1997) suggests that involuntary turnover can be used
as a mechanism to maintain employees’ incentives. In short, matching theory suggests that
higher turnover adds to productivity.
Although FSHC theory and job matching theory suggest opposite effects of turnover on
productivity, one does not necessarily invalidate the other. In fact, there are empirical
evidence supporting the coexistence of both effects, albeit the effect of FSHC appears to
dominate (see Glenn et al. (2001) and the citations therein). The two theories essentially
answer the question of how to balance the stability and flexibility of the labour force. It is our
contention that, given that FSHC and job matching have opposite effects on productivity
there is the distinct possibility that a certain turnover rate will maximise productivity. A
scenario, in which such an optimal turnover rate exists, is where productivity is a non-linear –
specifically quadratic concave function, of turnover. In the following, we formulate a number
of possible scenarios, and investigate which one is supported by the data.
2.1. Optimal Turnover Rate
Suppose firms’ total factor productivity
, is an additive separable function of employee
turnover rate T, and a set of other variables Z, (such as liquidity; degree of unionisation and
so on) such that:
Another factor that compounds the effect of turnover on productivity is knowledge spillover between firms
(Cooper 2001). Knowledge spill over is more significant if human capital is portable across firms or even
industries. For instance, Megna and Klock (1993) find that increasing research input by one semi-conductor
firm will increase the productivity of rival firms due to labour migration.
If there are feedback effects of productivity on the turnover rate, one should include lagged terms of T in the
equation and/or set up a system of equations. For instance, using U.S. data, Azfar and Danninger (2001) find
that employees participating in profit-sharing schemes are less likely to separate from their jobs, facilitating
. (1) () (); 0fgTT
Suppose is an additive separable function of FSHC and job matching effects: ( )gT
(2) () () ()gT pT qT=+
where ( )
T is a FSHC effect function, and qT a job matching effect function. Both ( ) ( )
and are at least twice differentiable on an interval ( )qT
T of T.
For all T and
∈ T , we can express ( )
T around T using Taylor series expansion:
( ) () '()( ) "()( )/2 ...
pT p p T p T
=+ −+ − +
=+ + +
is the first derivative of (.)
, the second derivative, and so forth.
According to FSHC theory,
is a negative function of T. If terms with orders higher than two
are negligible, for
to be a deceasing function around T , the necessary condition is
, a , and .
Similarly, can be expressed as a Taylor series: ( )qT
( ) () '()( ) "()( )/2 ...
qT q q T q T
=+ −+ − +
=+ + +
According to job matching theory, labour productivity is a positive function of employee
turnover rate. Again, if terms with orders higher than two are negligible, for to be an
increasing function around
T , the necessary condition is b , b , and bb .
Substituting (3) and (4) into (2), and the subsequent result into (1), we can write
00 11 22
( ) ...
( ) ...
the accumulation of FSHC. However, the short time span of our panel data prohibits us from taking this into
account in the empirical analysis.
Conditional on Z, equation (5) is a productivity-turnover (PT) curve. Provided that terms with
orders higher than two are negligible,
there are five scenarios regarding the signs of c
and, thus, the shape of the PT curve (Table 1).
Table 1. Various Scenarios of the Productivity-Turnover Curve
Scenario Shape of PT curve
T ) 0≥
Horizontal FSHC and job matching effects
cancel each other
Inverse U-shaped Matching effects dominate when
T is small, while FSHC effects
dominate when T is large
U-shaped FSHC effects dominate when T is
small, while job matching effects
dominate when T is large
Upward sloping Matching effects dominate Undefined
Downward sloping FSHC effects dominate 0
inverse U-shaped PT curve is the most likely scenario. This is because when
turnover is low, the level of FSHC is relatively high, whereas job-worker matching is less
likely to be optimal. Hence, the marginal benefit of increasing the labour market flexibility
overwhelms the marginal cost of forgoing some FSHC. As a result, productivity rises with
the turnover rate. Due to the law of diminishing marginal returns, the gain in productivity
lessens as turnover increases. Eventually the two effects will net out and further increases in
turnover will then lead to a fall in productivity.
This can be tested empirically and is satisfied for our data.
In the case of an
inverse U-shaped PT curve, the optimal turnover rate is given by .
The rate is not necessarily optimal from the perspective of firms, as competent employees
may leave for a better job opportunity. Neither is it necessarily optimal from the perspective
of employees, as there may be involuntary departure. In essence, turnover represents the fact
that firms are sorting workers and, reciprocally, workers are sorting firms. As a result, the
estimated optimal rate should be interpreted from the production perspective of the economy
as a whole. Moreover, the measurement does not take into account the hidden social costs of
turnover, such as public expenses on re-training and unemployment benefits, and the
searching costs borne by job seekers, and for that matter, hidden social benefits such as
higher social mobility.
3. Data, variables description and empirical models
3.1. Business Longitudinal Survey (BLS)
The BLS is a random sample of business units selected from the Australian Bureau of
Statistics’ business register for inclusion in the first year of the survey. The sample was
stratified by industry and firm size. The sample was selected with the aim of being
representative of all businesses (excluding government agents, public utilities and public
services). The focus is on a balanced panel of small and medium sized businesses. After
excluding businesses with deficient data records, 2,435 businesses are left in our sample
(Further details are presented in the Appendix).
This data source is unique in that it provides firm-level data, including an objective measure
of value-added and structural characteristics. Moreover, individual firms are tracked over a
four-year period from 1994/5 to 1997/8. The panel nature of the data allows us to investigate
the correlation between firm characteristics and productivity, taking into account unobserved
heterogeneity of firms. Although the panel is short, it has the advantages of being nationally
representative and having information on objective measure of value-added and firm
characteristics, which is excellent for our purposes.
3.2. The Econometric Model and Variable Descriptions
In the growth accounting literature, it is typical to use two-stage methods to identify the
determinants of total factor productivity (TFP) growth. At the first stage, TFP growth is
estimated as a residual from regressing value-added against factor inputs:
ln ln ln
it it it it
=+ + + (6)
is value-added of the ith firm in year t; and are capital and labour input,
is the TFP residual. At the second stage, the TFP residual is regressed
against other explanatory variables, as formulated in (5).
The two stages can be combined into one in that value-added is regressed against factor
inputs and other explanatory variables together, which is the approach adopted in this paper.
The econometric model therefore becomes:
ln ln ln
it it it it it it i it
=+ + + + + ++Z φ
where is an unobserved individual effect and an idiosyncratic disturbance term. The
unobserved individual effect accounts for any remaining unobserved entity heterogeneity,
such as management style, competition level of output markets.
The dependent and explanatory variables are briefly described as follows:
V (log value-added): Value-added is defined as sales
purchase + closing stock
opening stock, in financial year t.
K (log capital): Capital is measured as the total book value of non-current assets plus
imputed leasing capital. As reported in Rogers (1999), the importance of leasing capital
relative to owned capital varies significantly with firm size and industry, suggesting that
leasing capital should be included if we are to accurately approximate the total value of
capital employed in the production process. Leasing capital is imputed from data on the
estimated value of rent, leasing and hiring expenses.
L (log labour): Labour input is measured as the number of full-time equivalent
Since employment is a point in time measure, measured at the end of the
survey period (the last pay period in June of each year), we use the average numbers of
full-time equivalent employees in year t and year t-1 for each business as their labour
input in year t.
T (employee turnover rate): Employee turnover rate is measured by the average of new
employees and ceased non-casual employees divided by average non-casual employees at
the end of year t and t-1. The variables are only available from 1995/6 onwards.
shown in Appendix A1, the maximum value of turnover rate is 41. The accuracy of this
figure is questionable because the turnover rate is defined as turnover rate of non-casual
employees only. There is no clear pattern on the characteristics of firms with very high
turnover rate. We suspect that most of those firms with very high turnover rate may have
included the number of newly hired and ceased “casual” employees in the figure of non-
casual counterparts. In that case, considerable measurement errors would be introduced.
Thus, we exclude observations whose labour turnover rates are in the top end of the
distribution. Different cut-off points are experimented with as robustness check.
Leasing capital is imputed using the following formula: leasing capital = leasing expenses/(0.05+r). The
depreciation rate of leasing capital is assumed to be 0.05. Ten-year Treasury bond rate is used as the discount
rate (r). See Rogers (1999) for more detailed discussion.
The BLS only provides data on the number of full-time and part-time employees while the number of work
hours is not available. The full-time equivalent calculation is thus based on estimated average work hours of
part-time and full-time employees for the workforce as a whole, as published by the ABS in its monthly
Labour Force publication (cat. no. 6203.0).
Capital is also a point in time measure. However, capital is far less variable than labour (especially when
measured in terms of its book value), and hence the coefficient of capital is not sensitive to switching between
flow and point-in-time measures.
The questions for the calculation of labour turnover rate are slightly different in 1995/6 questionnaires. We
have tried the estimation that includes only data from the last two waves. However, these results did not affect
our conclusions. Since the panel is already very short, we included three years data in our main estimations.
Z (the control variables):
- employee coverage: there are three variables included in the regression proportion
of employees covered by individual contracts, by registered enterprise agreements,
and by unregistered enterprise agreements. It is expected that the sign on these
variables will be positive, because individual contracts and agreements tend to be
more commonly used with more skilled workers, and also because such agreements
tend to be used in tandem with performance-based pay incentives. Moreover, it is
widely believed that enterprise agreements are positively correlated with productivity
(Tseng and Wooden 2001).
- Union dummies: we include dummy variables that indicate whether a majority of
employees are union members and whether a minority of employees are union
members, respectively. A majority is defined as more than 50 per cent and a minority
being more than zero but less than 50 per cent. The reference category is businesses
without any union members at all.
- Part-time employee to total employee ratio and manager to total employee ratio:
businesses with higher manager to total employee ratio may have higher productivity
because their employees are better monitored. The effect of part-time to total
employee ratio is ambiguous because part-timers may be more efficient due to shorter
work hours, on the other hand, they may be less productive due to lesser accumulation
of human capital.
- A dummy variable that indicates whether a business was “innovative” in the previous
year: Innovation potentially has a long lag effect on productivity. Since the panel is
relatively short, in order to avoid losing observations, we include only a one-year lag.
Moreover, the definition of innovation is very board in the BLS. The coefficient of
innovation dummy is expected to be less significant that it should be.
- Dummy variables that indicate whether a business is an exporter, a family business, or
an incorporated enterprise. Exporters are likely to be positively correlated with
productivity in two aspects. High productivity businesses are more likely to survive in
highly competitive international markets. Secondly, trade may prompt faster
absorption of new foreign technology.
- Borrowing rate (measured at the end of the previous financial year): this variable is
used to measure how highly geared a firm is. The borrowing rate is expected to have a
positive impact on productivity, as the pressure of paying back the debts will motivate
greater efforts in production (Nickell, Wadhwani and Wall 1992).
- Firm age dummies: this variable is to control for any bias associated with the
mismeasurement of capital, as well as to control for industry specific knowledge.
- Industry dummies: industry dummies are included to control for industry specific
factors that may not be captured by the above variables.
4. Empirical Results
4.1. Primary Results
Initial Feasible Generalised Least Squares (FGLS) estimation results show that the t-statistics
of both the capital and labour variables are very large, indicating a strong possibility that
these variables are endogenous. In other words, these variables may be correlated with the
composite disturbance term through the unobserved individual effects. A possible
contributing factor to this is the limited number of control variables available in the data set.
For example, since only one-digit industry classification is available, some industry effects
could be captured in the unobserved individual effect, which is likely to be correlated with
both the capital stock and flow of labour. Moreover, export status may also be endogenous in
that firms in export-oriented industries have higher productivity due to knowledge spillover,
and productive firms are more likely to survive in the highly competitive global market
(Clerides, Lach and Tybout 1998; Bee, Chun and Roberts 2000).
Given this, we consider various estimation techniques that account for such endogeneity.
Firstly, write equation (7) generically as:
A source of measurement bias is the use of the book value of non-current assets. Using the book value will, in
general, lead to the underestimation of the true value of capital due to the treatment of depreciation. As firms
get older, the book value of capital is generally depreciated at a rate greater than the diminution in the true
value of the services provided by the capital stock.
it it i it
=+ ++, (8)
contains both time varying variables, x
, and time invariant ones, f
Hausman and Taylor (1981), it is possible to decompose w
into , where
is a subset of w
it it it
that is independent of the unobserved effect. Generalised Method of
Moments (GMM) estimation can then be based on the orthogonality conditions:
is based upon w
. Using the same partitions as for w
, three versions of this GMM
estimator exist. The Hausman and Taylor (H-T) (1981) estimator uses
Amemiya and MaCurdy (A-M) (1986) uses
. Finally, provided that the
correlation between w
is constant over time, one can use a further orthogonality
condition that deviations from time means are also valid instruments (Breusch, Mizon and
Schmidt 1989) (B-M-S).
In our base set of results, we include capital, labour and exports in w
for all three
estimators. However, different sets of endogenous variables are used as robustness checks
(see Section 4.2). Table 1 presents the results from the three different estimators (H-T, A-M
and B-H-S, respectively) as well as the (inconsistent) random effect FGLS estimator for the
sample with labour turnover rate less than 0.8. That is observations in the top 5 per cent of the
sample distribution are excluded. As mentioned earlier, the labour turnover rate for some
observations are questionably high, we therefore use this trimmed sample (trimmed at 0.8) as
our base sample (Sensitivity test of different cut-off points are also reported in the next
section). The coefficients of variables across the estimators are very similar and the signs are
as expected. We briefly summarise these results before focussing attention on the labour
The coefficient of log capital is very small although this is not surprising due to the use of
non-current assets as a proxy of capital, as explained previously. Because of this, we are
reticent to impose constant returns to scale on the regression. This argument gains support
from the negative coefficients of firm age dummies in that the underestimation of capital is
larger for older firms.
Since both capital and firm age variables are included as control
variables, the mismeasurement of capital should not unduly bias the coefficient of labour
Enterprise bargaining is positively correlated with productivity. This is reflected in the
positive and significant coefficients of the ratio of employees on registered agreements and
individual contract, as well as of highly unionised firms. However, the productivity of firms
with less than 50 per cent of employees unionised is not significantly different from
completely non-unionised firms.
The coefficient of borrowing rate is, as expected, positive, albeit only significant at 10 per
cent level. Manager to total employment ratio has no effect on productivity, whist the effects
of part-time to full-time employee ratio is significantly negative. The coefficients of these
two labour related variables vary considerably across the FGLS and the three consistent
The coefficient of innovation in the previous year is insignificant. This is not surprising as the
effect of innovation generally has long lags. Export firms and incorporated firms have higher
productivity. Family businesses, on average, are 21 per cent less productive than non-family
businesses. Medium and medium large firms have higher productivity than small firms.
If there is no underestimation of capital stock, other things equal, older firms are likely to have higher
productivity due to accumulation of experience.
Table 1 Estimation results from FGLS, H-T, A-M and B-H-S
Turnover Rate squared
Union Dummy (1-49%)
Union Dummy (50%+)
Manager to total Emp. Ratio
Ratio of Employment on
Ratio of Employment on
Ratio of Employment on
Borrowing Rate (t-1)
Ratio of Part time to total
Age (less than 2)
Age (2-5 years)
Age (5-10 years)
Age (10-20 years)
Adjusted R-squared 0.8496 0.8185 0.8190 0.8205
Number of observations 6428 6428 6428 6428
Number of firms 2380 2380 2380 2380
Note: “**” and “*” indicate significance at 5% and 10% level, respectively. Figures in parentheses are standard
4.2. Labour Turnover and Productivity
We now focus on the impact of turnover on productivity.
The coefficients of labour
turnover rate and its square are both significant and are positively and negatively signed,
respectively. This implies an inverse U-shaped productivity-turnover profile. It indicates that,
job matching effects dominate when turnover is low, whereas FSHC effects dominate when
turnover is high. The imputed optimal turnover rates for different estimators lies between
0.32 and 0.33. The differences among the three GMM estimators are negligible. Although the
estimated curvature of the productivity-turnover curve implied by the FGLS results is flatter
than that implied by the three GMM estimators, the estimated optimal turnover rates are very
Two broad sets of robustness checks were carried out. The first set involved changing the
variables that enter w
, the set of potentially endogenous variables. In addition to capital,
labour and exports, we have experimented with turnover (and its square), union density and
innovation. The second set of robustness checks, involves differing cut-off points for the
turnover rate variable using the base specification (that is treating capital, labour and exports
as potentially endogenous variables). In addition to these, we also ran the base specification
on sub-samples split by manufacturing/non-manufacturing and also by firm size. The main
results are reported in Table 2.
The coefficients of labour turnover rate and its squared term remain significant and retain the
same signs when we treat them as endogenous (although their magnitudes are smaller).
Again, there is no significant difference across the three different GMM estimators (as was
the case in all of the robustness checks).
For the second set of robustness checks, we tried a number of different cut-off points of the
turnover rate, from 0.5 to 1.0. Here we only report the results for three different samples: the
full sample, and the sub-samples with cut off points equal to 1.0 and 0.5, respectively.
Recall that the effects of turnover and turnover squared on productivity are essentially the same as those on
value-added as we have controlled for factor inputs.
We have also tried different combination of cut-off points, endogenous variables on full sample and industry
or firm size sub samples. Results can be obtained from the authors on request.
Table 2 Results for robustness checks
Turnover rate < 0.5 0.4721
Turnover rate < 1.0 0.1558
All observations 0.0120
Manufacturing firms 0.2492
Small firms 0.3859
Medium firms 0.1150
Note: all coefficients are significant at 5 % or 10% level (mostly 5%), except those for full sample, and medium
firms. Figures in parentheses are standard errors.
Endogenous variables (A) includes: log capital, log labour, export
Endogenous variables (B) includes: log capital, log labour, export, turnover, turnover squared
Endogenous variables (C) includes: log capital, log labour, export, turnover, turnover squared, union, innovation
Firms with a turnover rate higher than 0.5 are likely to be outliers as our definition of
turnover excluded casual workers.
Altering this cut-off point does change the coefficients
on the turnover variables significantly – becoming smaller as the cut-off point is raised. This
is understandable because if the outliers were due to measurement errors and, thus, not
correlated with productivity, incorporating them would weaken the estimated effect of
employee turnover on productivity. It seems reasonable to assume that the measurement
As a casual benchmark, policy advisers worked for the Australian Government has reportedly to have very
high turnover, mainly due to long hours, high stress and lack of a clear career path (Patrick 2002). Their
turnover rate was found to range from 29 per cent to 47 per cent under the Keating government (1991−1996).
errors are larger at the top end of the distribution, so that the effect of labour turnover rate
weakens as the cut-off point increases.
We also ran separate estimations for manufacturing and non-manufacturing, and small and
medium sized firms, respectively. Due to sample size restrictions, we were unable to further
divide non-manufacturing firms by industry classification. Medium-large firms are excluded
for the same reason. The optimal turnover rate for non-manufacturing firms is 0.38, higher
than the rate of 0.27 for manufacturing firms. Small and medium firms have similar optimal
turnover rates, of around 0.31. However, small firms have higher productivity loss when
moving away from the optimal point. An explanation may be that smaller firms deploy fewer
resources in knowledge management and standardising work processes, so that the departure
of experienced staff will cause a greater interruption on operation.
The effects of turnover rate on productivity for various samples using the B-H-S estimator are
illustrated in Figure 1 (Since the results from the three different estimators are fairly similar,
only one set of results are presented for simplicity of the graph). The diagram is a plot of log
productivity against turnover rate. The turnover-productivity curve can be read as that, in the
base case, increasing labour turnover rate from 0 to the optimal point (0.33), on average,
raises productivity by 4.4 per cent. Excluding the result from the full sample, the optimal
turnover rate ranges from 0.26 to 0.42 for the sub-samples reported in this paper.
The average turnover rate for the base sample is 0.18 and the median 0.13; both are well
below the optimal rate. A possible explanation for the large gap between the estimated
optimal rate and the sample mean and median is that agents (both employers and employees)
are of bounded rationality. Without sufficient information about the possibility of finding a
better substitute (either staff or job), agents will make changes at a rate lower than what
would have been if information were fully revealed. Another plausible explanation is that
there is enormous amount of friction in the dismissal and hiring process, such as legal
restrictions. While the finding cannot pin down exactly what attribute to the gap, it indicates
how much can be gained by bringing the turnover rate toward the optimal level. For instance,
an increase in turnover rate from the sample mean (0.18) to the optimal level (0.33) will
result in a 0.9 per cent rise in value-added, and 1.6 percent if from the sample median (0.13)
to the optimal level. These figures underline the potential gain from further increasing the
flexibility of employment arrangement for small and medium Australian enterprises.
Figure 1 Labour Turnover – Productivity Curve
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
Labour Turnover Rate
Log Value adde
all firms (rturn<0.8) all firms (rturn<0.5)
all firms (rturn<1.0) manufacturing
Although the findings are largely consistent with theory, cautions should be taken in reading
the results. Firstly, the analysis in this paper is based on small and medium firms. The
optimal turnover rate may not be able to extrapolate to large firms. Those very large firms,
particularly conglomerates or multinational enterprises, typically consist of many subunits,
which each one of them can be considered a small firm. Therefore, within-firm mobility may
substitute between-firm mobility.
. Secondly, the potential long-term effects of turnover on
productivity are not able to tested here due to data restriction. For instance, unfavourable
comments on a firm spread by its involuntarily separated employees may damage its
corporate image and, thus, weaken its attraction to quality potential employees. Therefore,
labour turnover may have slightly stronger negative effect in the long run. However, this
reputation effect should not be significant for small and medium firms because of their
In a case study, Lazear (1992) finds that the pattern of within-firm turnover from job to job resembles that of
atomic size in the labour market. The development of a long running panel is necessary for
This paper sets out to quantify the impact of employee turnover on productivity. Between the
two major theoretical arguments about the effect, FSHC theory asserts that high turnover
lowers firms’ incentives to provide staff training programs and, therefore, reduces
productivity. On the other hand, job matching theory postulates that turnover can help
employers and employees to avoid being locked in sub-optimal matches permanently,
subsequently increases productivity. The conflict between retaining workforce stability on the
one hand, and flexibility on the other, gives rise to the quest of an optimal turnover rate.
Using an Australian longitudinal data set, we establish that productivity is a quadratic
function of turnover. The inverse U-shaped productivity-turnover curve is consistent with the
intuition that job matching effects dominate while turnover is low, whereas FSHC effects
dominate while turnover is high. The optimal turnover rate is estimated to be about 0.3. The
result is very robust using various estimation methods, or using small and medium firms
subsamples. However, the non-manufacturing firms subsample has a slightly higher optimal
rate of 0.38, while the manufacturing firms subsample a lower one of 0.27.
The empirical results have significant implications to human resource and labour market
policies. For instance, it is found that an increase in turnover rate from the sample mean of
0.18 to the optimal level of 0.33 will result in a 0.9 per cent rise in value-added; the gain for
small enterprises is even higher. This suggests that further reforms to increase the flexibility
of job markets will yield substantial productivity gains for the Australian economy.
Abbasi, S. M. and K. W. Hollman (2000). “Turnover: The real bottom line.” Public
Ahn, S. (2001). “Firm dynamics and productivity growth: A review of micro evidence from
OECD Economic Department Working Paper # 297.
Amemiya, T. and T. MaCurdy (1986). “Instrumental estimation of an error components
Azfar, O. and S. Danninger (2001). “Profit-sharing, employment stability, and wage growth.”
Industrial and Labor Relations Review 54(3): 619-30.
Becker, G. S. (1975).
Human Capital: A Theoretical and Empirical Analysis, with Special
Reference to Education
. New York, National Bureau of Economic Research.
Bee, Y. A., S. Chun and M. J. Roberts (2000). “Productivity and the decision to export:
Micro evidence from Taiwan and South Korea.”
World Bank Economic Review 14(1):
Borland, J. (1997). “Employee turnover: Evidence from a case study.”
Australian Bulletin of
Breusch, T., G. Mizon and P. Schmidt (1989). “Efficient estimation using panel data.”
Burdett, K. (1978). “A theory of employee job search and quit rates.”
Carsten, J. M. and P. E. Spector (1987). “Unemployment, job satisfaction and employee
turnover: A meta-analytic test of the Muchinshky Model.”
Journal of Applied
Chang, C. and Y. Wang (1995). “A framework for understanding differences in labor
turnover and human capital investment.”
Journal of Economic Behavior and
Clerides, S. K., S. Lach and J. R. Tybout (1998). “Is learning by exporting important? Micro-
dynamic evidence from Colombia, Mexico, and Morocco.”
Quarterly Journal of
Cooper, D. P. (2001). “Innovation and reciprocal externalities: information transmission via
Journal of Economic Behavior and Organization 45: 403-25.
Economist (2000). “Employee turnover: Labours lost.”.
Economist (2001). “Corporate downsizing in America: The jobs challenge.”.
Fortune (2002). “The 100 best companies to work for.”: 21-39.
Glenn, A., J. P. McGarrity and J. Weller (2001). “Firm-specific human capital, job matching,
and turnover: Evidence from major league baseball, 1900-1992.”
Hausman, J. and W. Taylor (1981). “Panel data and unobservable individual effect.”
Econometrica 49: 1377-98.
Iverson, R. D. (1999). “An event history analysis of employee turnover: The case of hospital
employees in Australia.”
Human Resource Management Review 9(4): 397-418.
Jovanovic, B. (1979a). “Job matching and the theory of turnover.”
Journal of Political
Jovanovic, B. (1979b). “Firm-specific capital and turnover.”
Journal of Political Economy
Lazear, E. (1992). “The Job as a Concept” in
Performance Measurement, Evaluation, and
. J. William J. Bruns. Boston, Harvard Business School: 183-215.
Lilien, D. M. (1982). “Sectoral shifts and cycllical unemployment.”
Journal of Political
Lucas, R. E. and E. C. Prescott (1974). “Equilibrium search and unemployment.”
Lucifora, C. (1998). “The impact of unions on labour turnover in Italy: Evidence from
establishment level data.”
International Journal of Industrial Organization 16(3):
Lynch, L. M. (1993). “The economics of youth training in the United States.”
McLaughlin, K. J. (1990). “General productivity growth in a theory of quits and layoffs.”
Journal of Labor Economics 8(1): 75-98.
Megna, P. and M. Klock (1993). “The impact of intangible capital on Tobin's q in the
American Economic Review 83(2): 265-69.
Mintzberg, H. (1979).
The Structuring of Organizations. Englewood Cliffs, NJ, Prentice Hall.
Mok, C. and Y. Luk (1995). “Exit interviews in hotels: Making them a more powerful
International Journal of Hospitality Management 14(2): 187-94.
Muchinsky, P. M. and M. L. Tuttle (1979). “Employee turnover: An empirical and
Journal of Vocational Behaviour 14(43-77).
Nelson, R. and S. Winter (1982).
An Evolutionary Theory of Economic Change. Cambridge,
Mass., Harvard University Press.
Nickell, S., S. Wadhwani and M. Wall (1992). “Productivity growth in U.K. companies,
European Economic Review 36(5): 1055-91.
Patrick, A. (2002). “PM tightens grip on Canberra.”
The Australian Financial Review: 6.
Petrongolo, B. and C. A. Pissarides (2001). “Looking into the black box: A survey of the
Journal of Economic Literature 39(2): 390-431.
Rogers, M. (1999). “The performance of small and medium enterprises: An overview using
the growth and performance survey.”
Melbourne Institute Working Paper (1/99).
Sheehan, E. P. (1993). “The effects of turnover on the productivity of those who stay.”
Journal of Social Psychology 133(5): 699.
Shepard, E. M., T. J. Clifton and D. Kruse (1996). “Flexible working hours and productivity:
Some evidence from the pharmaceutical industry.”
Industrial Relationship 35(1): 123-
Tseng, Y.-P. and M. Wooden (2001). “Enterprise bargaining and productivity: Evidence from
the business longitudinal survey.”
Melbourne Institute Working Paper (8/01).
A.1 Sample Selection
The first wave of BLS was conducted in 1994/5, with a total effective sample size of 8,745
cases. The selection into the 1995/6 sample was not fully random. Businesses that had been
innovative in 1994/95, had exported goods or services in 1994/95, or had increased
employment by at least 10 per cent or sales by 25 per cent between 1993/94 and 1994/95,
were included in the sample. A random selection was then made on all remaining businesses.
These businesses were traced in the surveys of the subsequent two years. In order to maintain
the cross-sectional representativeness of each wave, a sample of about 800 businesses were
drawn from new businesses each year. The sample size in the second, third and fourth waves
are around 5,600. For detailed description of the BLS data set, see Tseng and Wooden
(2001). Due to confidentiality considerations, the complete BLS is not released to the public,
only the Confidentialised Unit Record File (CURF) is available. In the CURF, businesses
exceed 200 employees and another 30 businesses that are regarded as large enterprises using
criteria other than employment are excluded. This leaves around 4,200 businesses in the
We further reduced the number of cases available for analysis by deleting observations that
had been heavily affected by imputation because including them would impose artificial
stability. The cases where complete sections of the questionnaire had been imputed are
discarded. Moreover, we excluded businesses in the finance and insurance industries because
of substantial differences in the measures of value-added and capital.21 In addition,
observations with negative sales and negative liabilities were dropped, as were a small
number of cases where it was reported that there were no employees. In total, this left just
2,435 businesses in our sample. Summary statistics are presented in Table A1.
Since only around 1.6 per cent of businesses in the balanced panel is in the financial sector, it is not feasible to
undertake a separate analysis on this sample.
Table A1. Summary statistics
Mean SD Min Max
Log value-added 7.086 1.458 -0.028 13.554
6.764 1.622 0 13.716
2.811 1.122 -0.979 5.263
0.715 0.451 0 1
0.515 0.500 0 1
Manager to total Emp. Ratio 0.254 0.170 0 0.889
rpart 0.216 0.289 0 1
rturn 0.254 0.670 0 41.000
Ratio of Employment on Individual contract
0.232 0.354 0 1
Ratio of Employment on Unregistered agreement
0.091 0.256 0 1
Ratio of Employment on Registered agreement
0.061 0.207 0 1
Export 0.269 0.444 0 1
Union Dummy (1-49%)
0.206 0.404 0 1
Union Dummy (50%+)
0.083 0.276 0 1
Innovation (t-1) 0.297 0.457 0 1
Borrowing Rate (t-1) 0.752 1.283 0.037 51.064
Firm size dummy: Medium
0.438 0.496 0 1
Firm size dummy: Medium-Large
0.065 0.246 0 1
Age dummy (less than 2)
0.063 0.243 0 1
Age dummy (2-5 years)
0.134 0.340 0 1
Age dummy (5-10 years)
0.251 0.434 0 1
Age dummy (10-20 years)
0.284 0.451 0 1
Age dummy (20 years+) 0.267 0.443 0 1
Mining 0.008 0.089 0 1
Manufacturing 0.428 0.495 0 1
Construction 0.043 0.203 0 1
Wholesale trade 0.182 0.386 0 1
Retail trade 0.106 0.308 0 1
Accommodations, cafes & restaurants 0.036 0.185 0 1
Transport & storage 0.030 0.170 0 1
Finance & insurance 0.012 0.111 0 1
Property & business services 0.118 0.323 0 1
Cultural & recreational services 0.018 0.132 0 1
Personal & other services 0.019 0.137 0 1
No of observations (n×T): 6756