OXFORD BULLETIN OF ECONOMICS AND STATISTICS, 68, 4 (2006) 0305-9049

The Impact of Training on Productivity and

Wages: Evidence from British Panel Data*
Lorraine Dearden, Howard Reedà and
John Van Reenen§

Institute for Fiscal Studies and Institute of Education, London, UK

(e-mail: [email protected])
àInstitute for Public Policy Research, London, UK (e-mail: [email protected])
§Centre for Economic Performance, London School of Economics and CEPR,
London UK (e-mail: [email protected])

Abstract
It is standard in the literature on training to use wages as a sufficient statistic
for productivity. This paper examines the effects of work-related training on
direct measures of productivity. Using a new panel of British industries
1983–96 and a variety of estimation techniques we find that work-related
training is associated with significantly higher productivity. A 1% point
increase in training is associated with an increase in value added per hour of
about 0.6% and an increase in hourly wages of about 0.3%. We also show
evidence using individual-level data sets that is suggestive of training
externalities.

*Financial support from the Economic and Social Research Council and the Leverhulme Trust is
gratefully acknowledged. We would like to thank Fernando Galindo-Ruedo, Amanda Gosling,
Richard Layard, Lisa Lynch, Stephen Machin, Steve McIntosh, Alan Manning, Steve Pischke, two
anonymous referees and participants in many seminars for helpful comments. We are grateful to
the Office for National Statistics (ONS) for supplying the Census of Production data and to the
Organization for Economic Co-operation and Development (OECD) for supplying the Inter-Sectoral
Data Base (ISDB) data. We would like to thank Martin Conyon for supplying the firm-level training
data and Jonathan Haskel for providing us with some of the industry price indices used in this paper.
Material from the Labour Force Survey is Crown Copyright, has been made available by the ONS
through the ESRC Data Archive and has been used by permission. Neither the ONS nor the Data
Archive bears any responsibility for the analysis or interpretation of the data reported here. An earlier
version of this paper (with a longer data description) has appeared as ‘Who gains when workers
train’, Working Paper No. 00/04, Institute for Fiscal Studies.
JEL Classification numbers: C23, D24, J31.
397

Blackwell Publishing Ltd, 2006. Published by Blackwell Publishing Ltd, 9600 Garsington Road, Oxford OX4 2DQ, UK
and 350 Main Street, Malden, MA 02148, USA.
398 Bulletin

I. Introduction
It is a widely held view that Britain needs to increase work-related training to
improve long-term economic performance and address the ‘skills gap’.1
Despite the policy interest and the huge economics literature on human
capital, there are hardly any papers that examine the impact of work-related
training on direct measures of productivity. The primary contribution of our
paper is to provide such evidence for the first time in the UK and for the first
time anywhere over a long period (we have 14 years of data). Analysis of the
impact of training on productivity has focused almost entirely on estimating
the impact of training on wages. Most studies looking at the private return to
work-related training find that training results in workers receiving higher real
wages.2
Although these studies are informative, they only tell half the story as they
ignore the impact on the employer’s productivity. The relationship between
wage increases and productivity gains can vary according to the structure of
the labour and product markets and according to who actually pays the costs
of training. In the simplest neoclassical view of the labour market where the
market is perfectly competitive, wages will be equal to the value of margi-
nal product. Thus the wage can be taken as a direct measure of productivity.
This simple relationship can break down for many reasons. For example, in
Becker’s model of specific human capital, the employer will pay for training,
so there should be no effect of completed training spells on observed wages
even though there may be a large impact on productivity.3
If the labour market is characterized by imperfect competition then the
strict link between wages and productivity is usually broken. Employees can
find themselves being paid less (or more) than their marginal revenue product.
Nevertheless, it is still the case that conditional on a given degree of rent-
sharing or monopsony power; increases in wages have to be paid out of
productivity gains. Therefore, we can assert the general principle that these
1
See Green and Steedman (1997) or National Skills Task Force (1998). In the December 2003 Pre-
Budget Report, the British Chancellor justified the extension of the Employer Training Pilots in order
to help improve the skills gap and UK productivity (http://www.hm-treasury.gov.uk/media//2E3BD/03_
Meeting%20the%20Pro_EF.pdf).
2
See Greenhalgh and Stewart (1987), Booth (1991, 1993) or Blundell, Dearden and Meghir (1996)
for UK evidence. US studies using panel data include Lillard and Tan (1992), Lynch (1992),
Blanchflower and Lynch (1992) and Bartel and Sicherman (1999). Winkelmann (1994) uses German
data and Bartel (1995) looks within a large US manufacturing company.
3
There are many other reasons for a wedge between productivity and wage in a competitive labour
market. First, employees may receive non-pecuniary benefits from training. Secondly, workers may
implicitly pay the costs of a training scheme in the form of lower wages whilst being trained, which
then rise after training is completed – so we might see a greater increase in observed wages than in
productivity. Thirdly, employees’ wages could be lower during training because they are not con-
tributing to firm productivity whilst actually being trained. Fourthly, there may be deferred com-
pensation packages where the employee’s remuneration is ‘backloaded’ towards later post-training
years as a means of ensuring loyalty and/or effort early in the employee’s tenure (e.g. Lazear, 1979).

Blackwell Publishing Ltd 2006

 The impact of training on productivity and wages 399

real wage increases should provide a lower bound on the probable size of
productivity increases. In practice, the productivity gains are likely to be
higher than this. For instance, in a labour market with frictions and some wage
compression (e.g. from a binding minimum wage), there will be productivity
gains even from general training that are not passed on to the employee in
terms of wages but are only reflected in direct measures of productivity.4
Similar results can be found in some bargaining models (e.g. Booth,
Francesconi and Zoega, 1999).
There exist a small number of empirical papers that relate firm productivity
to a measure of training.5 Although a positive correlation is generally found, it
is very difficult to interpret because the training measures are only measured
at a single point of time and could be picking up many unobservable firm-
specific factors correlated with both training and productivity. Black and
Lynch (2001) used an establishment training survey at two points of time.
In the cross-section, they identified some effects of the type of training on
productivity, but they found no significant association when they controlled
for plant-specific effects. Ichniowski, Shaw and Prennushi (1997) investigated
the factors that influence productivity in a panel of US steel finishing mills.
After controlling for fixed effects, they found a role for training only in
combination with a large variety of complementary human resource practices.
Carriou and Jeger (1997), Ballot, Fakhfakh and Taymaz (1998) and Delame
and Kramarz (1997) used French firm-level panel data to look at the effects of
training on value added and found positive and significant effects. Although
these studies are broadly consistent with our own, they do not fully exploit the
potential of their panel data by allowing training to be a choice variable.6
Our contribution in this paper is to advance the literature in at least three
ways. Black and Lynch (2001) emphasize the problems of trying to identify
the effects of training in a short panel (they have only two separate training
observations). Although unobserved heterogeneity can be controlled for
through fixed effects with only two periods, attenuation biases due to
measurement error are exacerbated. To address this, we build a panel that
contains up to 14 consecutive years of training data. Secondly, we explicitly
allow training to be a choice variable by using general method of moments
(GMM) estimators developed to deal with endogenous variables in produc-
tion functions. Thirdly, we combine estimation of the productivity effects of
4
See Acemoglu and Pischke (1999, 2003).
5
Black and Lynch (1996), Bartel (1994), Barrett and O’Connell (2001), De Koning (1994), Boon
and van der Eijken (1997) and Ballot and Taymaz (1998) have objective productivity measures.
Bartel (1995), Holzer (1990), Barron, Black and Zoewenstein (1989) and Krueger and Rouse (1998)
use subjective measures of productivity. Holzer et al. (1993) do find effects of changes in produc-
tivity on changes in one measure of quality – the scrap rate.
6
See Greenhalgh (2002) for a much more extensive review of the French and UK literature in this
area.

Blackwell Publishing Ltd 2006

400 Bulletin

training with estimation of the wage effects of training. Although comparisons

between the production function and the wage equation are becoming more
common for other worker characteristics such as gender and human capital,
this is the first time the strategy has been used for training.7 In principle, this
allows us to examine whether trained workers are paid the value of their
marginal product.
We conduct our main analysis of the effects of training at the industry
level (although we have also performed estimation at the firm and
individual level for comparative purposes). There is simply no alternative
to this strategy if one wishes to use long time series of training and
productivity information. The only publicly available firm-level panel data
in the UK is a sample of about 119 firms in the late 1990s with only very
basic training information (a concise investigation of this is presented in
Dearden, Reed and Van Reenen, 2005, Appendix B). Aggregation has pros
and cons that are discussed in this paper. On the positive side, if there are
important spillovers to training within an industry (e.g. through a faster rate
of innovation) then a firm-level analysis will potentially miss out these
linkages and underestimate the return to human capital.8 On the negative
side, there may be aggregation biases at the sectoral level that could lead
to negative or positive biases on the training coefficient. We follow
Grunfeld and Griliches (1960) in arguing that the pros of aggregation
probably outweigh the cons.
The format of the paper is as follows. Section II describes the simple
economic models of productivity and wages that we will estimate and
section III details the econometric strategy. The data are described in
section IV and the results are presented in section V. Section VI offers some
concluding comments. Dearden et al. (2005) contains more information on
the data and some additional experiments. Our main result is that we find a
statistically and economically significant effect of training on industrial
productivity. A 1% point increase in training is associated with about a 0.6%
increase in productivity and a 0.3% increase in hourly wages. The productivity
effect of training is twice as large as the wage effect, implying that existing
studies have underestimated the benefits of training by focusing on wages.

7
Hellerstein, Neumark and Troske (1999), Hellerstein and Neumark (1999), Hægeland and Klette
(1999) and Jones (2001) examine the differential impact of human capital and gender on wages and
productivity. A recent study that utilizes our methodology and looks at this question using a panel of
French and Swedish firms is Ballot, Fakhfakh and Taymaz (2002). They find that both French and
Swedish firms appropriate a high proportion of the returns to training (82% and 67% respectively).
8
For example, O’Mahony (1998) finds that the coefficient on labour skills in a production function
is more than twice that assumed by traditional growth accounting from relative wages. Other recent
papers that have looked at the impact of human capital on directly measured productivity include
Moretti (2004) on US data and Haskel, Hawkes and Pereira (2003) on UK data.

Blackwell Publishing Ltd 2006

 The impact of training on productivity and wages 401

II. A model of training and productivity

To see our approach, assume that we can characterize a representative plant in
an industry by a Cobb–Douglas production function written in value added
form9
Q ¼ ALa K b ; ð1Þ
where Q is value added, L is effective labour input allowing for quality and
quantity dimensions, K is capital and A is a Hicks neutral efficiency parameter.
We consider that trained workers are more productive than untrained
workers, so that effective labour input can be written as
L ¼ N U þ cN T ; ð2Þ
U T
where N is the number of untrained workers, N is the number of trained
workers and c is a parameter which, if trained workers are more productive
than non-trained workers, will be greater than unity. The total number of
workers, N, is equal to the sum of trained and untrained workers. Substituting
equation (2) into equation (1) gives
Q ¼ A½1 þ ðc
1ÞTRAINa N a K b ; ð3Þ
T
where TRAIN ¼ N /N is the proportion of trained workers in an industry.
Taking natural logarithms, we obtain
ln Q ¼ ln A þ a ln½1 þ ðc
1ÞTRAIN þ a ln N þ b ln K: ð4Þ
This could be estimated by non-linear least squares. If (c
1) TRAIN is
‘small’, we can use the approximation ln(1 + x) ¼ x and rewrite the produc-
tion function as10
ln Q ¼ ln A þ aðc
1ÞTRAIN þ a ln N þ b ln K: ð5Þ
If the industry exhibits constant returns to scale (i.e. a + b ¼ 1) then
equation (5) can be rewritten in terms of labour productivity as



Q K
ln ¼ ln A þ ð1
bÞðc
1ÞTRAIN þ b ln : ð6Þ
N N
If the trained are no more productive than the untrained (c ¼ 1) then the
coefficient on TRAIN will be zero.
This method can be easily extended to a larger number of different types of
heterogeneous workers in the labour quality index. If we index the discrete
9
This should be viewed as a first-order approximation to a more complicated functional form.
It is straightforward to generalize this to more complex functional forms such as translog and
some experiments are included in the empirical results.
10
The results were estimated both by non-linear least squares and by least squares using the
approximation. The results did not significantly differ (see Table 3), so the more convenient log-
linear approximation is used for the baseline results.

Blackwell Publishing Ltd 2006

402 Bulletin

type of labour by k (where until now we have discussed k solely in terms of

the training status of workers) then equation (4) can be written
(
)
X Nk
ln Q ¼ ln A þ a ln 1 þ ðck
1Þ þ a ln N þ b ln K: ð7Þ
k
N
Empirically, we will allow for many other dimensions of labour quality such
as education, occupation, age, tenure and gender.11 There are a large number
of other influences on productivity captured in A, so we allow for differen-
tial hours, worker turnover rates, innovation (as proxied by research and
development expenditures), regional composition and the proportion of small
firms. Labelling these factors as X, imposing constant returns and using the log
approximation, the basic production function becomes

 X


Q Nk K
ln ¼ ð1
bÞ ðck
1Þ þ b ln þ d0 X : ð8Þ
N k
N N
The wage equation that we estimate parallels the productivity equation in (8).
We view the wage equation as more of a descriptive regression than the
structurally derived production function. Under competitive spot markets for
labour, relative wages should equal the relative marginal productivities
of workers of different types. This is because if the relative productivity
of trained workers, c, exceeded the relative wages of trained workers then
employers would only employ trained workers (Hellerstein et al., 1999).
Consider the wage bill, W, for the representative plant in an industry.
Again, take the simplest model where there are only two types of workers:
trained workers paid average wage wT and untrained workers paid average
wage wNT. Relative wages are k ¼ wT/wNT. By definition,
W ¼ wNT ðN
N T Þ þ kwNT N T ¼ wNT ½N þ ðk
1ÞN T : ð9Þ
In logarithms, the average wage (w) is


W
ln w ¼ ln ¼ a þ ln½1 þ ðk
1ÞTRAIN; ð10Þ
N
where a ¼ ln(wNT).
Clearly, estimation of equation (10) can be used to recover the relative
wage mark-up associated with training, k, and then compared with the relative
productivity effect of training, c. Parallel to the productivity equation, we will

11
We follow Hellerstein et al. (1999) by entering these variables in linear proportions. One could
allow a larger number of cells for interactions of the labour quality variables (e.g. the proportion of
educated men – a two-way interaction – or the proportion of educated men who are trained – a three-
way interaction). We experimented with some breakdowns like this on the training variable, but
Labour Force Survey (LFS) cell sizes by industry were generally not large enough.

Blackwell Publishing Ltd 2006

 The impact of training on productivity and wages 403

allow for multiple types of labour quality, capital and other factors to influence
wages. The empirical wage equation to be estimated is therefore:12
X

Nk
ln w ¼ a þ ðkk
1Þ þ bw ln K þ dw0 X : ð11Þ
k
N

III. Econometric modelling strategy

The basic equation we estimate can be written in simplified form as
yit ¼ hxit þ uit ; ð12Þ
where y is Q/N and x is a vector of (suspected endogenous) variables including
training. Subscript i indicates the representative firm in an industry, t is time
and h is the parameter of interest. Assume that the stochastic error term, uit,
takes the form
uit ¼ gi þ st þ xit ;
ð13Þ
xit ¼ qxit
1 þ tit :
The st represent macroeconomic shocks captured by a series of time dum-
mies, gi is an individual effect and tit is a serially uncorrelated mean zero error
term. The other element of the error term, xit, is allowed to have an AR(1)
component (with coefficient q), which could be due to measurement error or
slowly evolving technological change. Substituting equation (13) into equa-
tion (12) gives the dynamic equation
yit ¼ p1 yit
1 þ p2 xit þ p3 xit
1 þ gi þ st þ tit : ð14Þ
The common factor restriction (COMFAC) is p1p2 ¼
p3. Note that st ¼
st
qst
1 and gi ¼ ð1
qÞgi .
In our main results section, we present several econometric estimates of
production functions (random effects, within groups and GMM). The most
rigorous approach follows that recommended by Blundell and Bond (2000),
which uses a ‘system GMM’ approach to estimate equation (14) and then
imposes the COMFAC restrictions by minimum distance. We now turn to
describing the GMM approach in more detail.
How should equation (14) be estimated? If training is strictly exogenous
and there are no dynamics (i.e. q ¼ 0) then the only problem with ordinary
least square (OLS) estimation of equation (12) is the presence of the
12
One could argue that firm variables such as capital intensity and R&D should be excluded from
the wage equation under competitive labour markets. However, these variables are typically quite
informative in wage equations, either because they are picking up some measure of unobserved
labour quality (Hellerstein and Neumark, 1999) or because of departures from perfect competition. In
either case, omitting such variables is likely to cause bias on the training variable and our baseline
specifications will include them.

Blackwell Publishing Ltd 2006

404 Bulletin

individual effects, gi. If these individual effects are uncorrelated with xit then
the random-effects estimator is unbiased and efficient. If the individual
effects are correlated with xit but remain strictly exogenous then although the
random-effects estimator is biased, the within-groups estimator will be
unbiased.
If we allow training to be endogenous (i.e. allowing training decisions to
react to shocks to current productivity), we will require instrumental variables.
In the absence of any obvious natural experiments, we consider moment
conditions that will enable us to construct a GMM estimator for equation (14).
A common method would be to take first differences of equation (14) to
sweep out the fixed effects:
Dyit ¼ p1 Dyit
1 þ p2 Dxit þ p3 Dxit
1 þ Dst þ Dtit : ð15Þ
As tit is serially uncorrelated, the moment condition
Eðxit
2 Dtit Þ ¼ 0 ð16Þ
ensures that instruments dated t
2 and earlier13 are valid and can be used to
construct a GMM estimator for equation (14) in first differences (Arellano and
Bond, 1991). A problem with this estimator is that variables with a high
degree of persistence over time (such as capital) will have very low correlation
between their first difference (Dxit) and the lagged levels being used as
instruments (e.g. xit
2). This problem of weak instruments can lead to
substantial bias in finite samples.
Blundell and Bond (1998) point out that under a restriction on the initial
conditions, another set of moment conditions are available:14
EðDxit
1 ðgi þ tit ÞÞ ¼ 0: ð17Þ
This implies that lags of the first differences of the endogenous variables can
be used to instrument the levels equation (14) directly. The econometric
strategy is then to combine the instruments implied by the moment con-
ditions (16) and (17). We stack the equations in differences and levels, i.e.
equations (14) and (15). We can obtain consistent estimates of the coefficients
and use these to recover the underlying structural parameters in equation (12).
The estimation strategy assumes the absence of serial correlation in the
levels error terms (tit).15 We report serial correlation tests in addition to the
13
Additional instruments dated t
3, t
4, etc. become available as the panel progresses through
time.
14
The restrictions are that the initial change in productivity is uncorrelated with the fixed effect
E(Dyi2gi) ¼ 0 and that initial changes in the endogenous variables are also uncorrelated with the
fixed effect E(Dxi2gi) ¼ 0.
15
If the process is MA(1) instead of MA(0) then the moment conditions in equations (16) and (17)
no longer hold. Nevertheless, E(xit
3Dtit) ¼ 0 and E(Dxit
2(gi + tit)) ¼ 0 remain valid, so earlier-
dated lags could still be used as instruments. This is the situation empirically with the wage
equations.

Blackwell Publishing Ltd 2006

 The impact of training on productivity and wages 405

Sargan–Hansen test of the over-identifying restrictions in all the GMM results

below.16
This GMM ‘system’ estimator has been found to perform well in Monte
Carlo simulations (Blundell and Bond, 1998) and in the context of the esti-
mation of production functions (Blundell and Bond, 2000). The procedure
should also be a way of controlling for transitory measurement error (the fixed
effects control for permanent measurement error). Random measurement error
has been found to be a problem in the returns to human capital literature,
typically generating attenuation bias (see Card, 1999).
In order to assess the importance of biases associated with fixed effects
and endogeneity, we will estimate random effects, within groups and GMM
estimates in section V.
Finally, consider two more issues which are harder to deal with:
aggregation and training stocks vs. training flows. Estimation at the three-
digit industry level has advantages but also disadvantages relative to micro-
level estimation. The production function in equation (1) at the firm level
describes the private impact of training on productivity. However, many
authors, especially in the endogenous growth literature (e.g. Aghion and
Howitt, 1998), have argued that there will be externalities to human capital
acquisition. For example, workers with higher human capital are more likely
to generate new ideas, which may spill over to other firms.17 If spillovers are
industry specific, this implies that there should be additional terms added
to equation (5) representing training in other firms (e.g. the mean number
of trained workers in the industry). In this case, the coefficient on training
in an industry-level production function should exceed that in a firm-level
production function.18 Secondly, grouping by industry may smooth over some
of the measurement error in the micro-data and therefore reduce attenuation
bias.
On the negative side, there may be aggregation biases in industry-level
data. A priori it is not possible to unambiguously sign these biases. We expect
that the fixed effects will control for some of the problem. For example, we are
taking logs of means and not the means of logs in aggregating equation (4),
but so long as the higher-order moments of the distributions are constant over
time in an industry then they will be captured by a fixed effect.19 If the

16
These are based on the first-differenced residuals, so we expect significant first-order serial
correlation but require zero second-order serial correlation for the instruments to be valid. If there
is significant second-order correlation, we need to drop the instruments back a further time period
(this happens to be the case for the wage equation in the results below).
17
Although there are many papers that examine externalities of R&D (e.g. see the survey by
Griliches, 1992) and a few that look at education (Acemoglu and Angrist, 2000; Moretti, 2004), there
are none that focus on training spillovers.
18
For the same argument in the R&D context, see Griliches (1992).
19
If they evolve at the same rate across industries, they will be picked up by the time dummies.

Blackwell Publishing Ltd 2006

406 Bulletin

coefficients are not constant across firms in equation (4), but are actually
random, this will also generate higher-order terms at the industry level. In the
empirical results, we experiment with including higher-order moments and
allowing the coefficients to vary across cross-sectional units.
Turning to the problem of training stocks and flows, note that the model in
equation (1) assumes that we know the stocks of trained workers in an
industry. What we actually have in the data are an estimate of the proportion
of workers in an industry who received training in a given 4-week period (the
training flow). Since individuals are sampled randomly over time in the LFS,
this should be an unbiased estimate of the proportion of people in training in a
given industry in a given year.20 As an alternative to using the flow, we
calculate a stock of training in an analogous way to using investment flows to
calculate a capital stock through the perpetual inventory method (the main
form of depreciation is the turnover rate). This is described in Dearden et al.
(2005), Appendix A.

IV. Data description

The database we construct combines several sources (see Dearden et al.,
2005, Appendix A for full details). The critical individual-level source is the
individual-level UK LFS, which contained about 60,000 households per year.
Most importantly, the LFS has a consistent training measure since 1984 as
well as detailed information on skills, demographics, hours worked, tenure
and wages. We work with this information aggregated by broadly three-digit
industries. The LFS only started asking questions on wages at two points of
time in 1997 (and at one point of time in 1992 when the panel was set up). We
present some individual-level panel wage regressions at the end of section V
for comparison.
The second major data set we use is the Annual Census of Production. This
gives production statistics on capital, wages, labour and output, for industries
in the production sector (manufacturing, mining and utilities). For the services
industries, we drew on the OECD’s ISDB data.
There was a change in Standard Industrial Classification (SIC) classifi-
cation in 1992 which forced us to aggregate some of the industries and
prevented us from using some of the industries after the change. Additionally,
we insisted on having at least 25 individuals in each cell in each year. After
matching the aggregated individual data from LFS, we were left with 94
industry groupings over (a maximum of) 14 years.
20
If there are many multiple training spells in the month, we will underestimate the proportion of
employees who are being trained. If Spring (the LFS quarter we use) is a particularly heavy training
season then we will overestimate the proportion being trained in a year. These biases are likely to be
small and offsetting.

Blackwell Publishing Ltd 2006

 The impact of training on productivity and wages 407

5
Labour productivity

0
0.0 0.1 0.2 0.3 0.4
Proportion training

Figure 1. Labour productivity and training in British industries. Each point is an industry–
year observation. The ordinary least square regression line has a slope of 4.91. Labour pro-
ductivity is ln(value added per employee) aggregated from the Annual census of production
and ISDB; training is the proportion of workers involved in training in the last 4 weeks from
the LFS

The main LFS training question was ‘over the 4 weeks ending Sunday . . .
have you taken part in any education or training connected with your job, or a
job that you might be able to do in the future?’.21 The average proportions of
employees undertaking training grew steadily from about 8% in 1984 to 14%
in 1990 where it stabilized for the next 6 years. Most of this growth was
upgrading within industry rather than between industries.22
Figure 1 gives the scatterplot of labour productivity (log real value added
per worker) against training propensity and Figure 2 repeats the exercise for
log hourly wages. Not surprisingly, training has a strong positive correlation
with both variables, but the association is somewhat weaker for wages than for
productivity.
The outliers in both graphs tend to be in the service sector. Unfortunately,
the published series for real value added and capital stocks are rather unreli-
able in the service sector. For example, in banking and financial services,
measured real value added per person declined every year between 1983
21
Unfortunately, it is not possible to separate out ‘education’ from ‘training’.
22
There is also a question on the length of the training spell, but this was only asked in particular
years and there were too many missing values to use it as a separate regressor. Median spell length
was 2 weeks and the mean was higher.

Blackwell Publishing Ltd 2006

408 Bulletin

2.5

2
Wages

1.5

0.5

0
0.0 0.1 0.2 0.3 0.4
Proportion training

Figure 2. Wages and training in British industries. Each point is an industry–year observation.
The ordinary least square regression line has a slope of 2.95. Wages are ln(hourly wages) from
the ABI and ISDB (wages) and the Labour Force Survey (LFS) (hours); training is the pro-
portion of workers involved in training in the last 4 weeks from the LFS

and 1996. Given the poor quality of the service sector production data,
we reluctantly decided to focus the econometric part of the analysis on
the production side of the economy. This is still a substantial share of the
economy – about 50% of private sector net output in 1986.23 Until robust
measures of service sector productivity are developed, there is simply no
alternative to the empirical strategy of focusing on the production sector.
Care must be taken in interpreting the scatterplots presented in Figures 1
and 2 as they say nothing about the causal impact of training on productivity
or wages. High-training industries are characterized by higher fixed capital
intensity, more professional workers, more educated workers and higher
R&D (see Table A1 in Appendix A of Dearden et al., 2005). We need to
turn to an explicit econometric model to investigate whether there is a causal
effect of training on productivity, and this forms the focus of the rest of the
paper.

23
This led to the loss of only 91 observations and the results are robust to including the service
sector in the unweighted regressions. We generally weight the regressions by the number of LFS
observations in order to reduce sampling variability. In the weighted regressions, including the
service sector does have more substantial effects on the results because of its large employment
shares. Full sets of these results are available on request from the authors.

Blackwell Publishing Ltd 2006

 The impact of training on productivity and wages 409

V. Results
Baseline industry results
In Table 1, we present the basic results for the industry-level regressions
treating all variables as exogenous. The first three columns have productivity
(log real value added per head) as the dependent variable and the last three
columns have wages as the dependent variable.
The first two columns are estimated by random effects; the only
difference is that column (1) does not include the occupational controls. This
omission makes some difference to the ‘no qualifications’ variable, which
has a significantly negative association with productivity in column (1) but
is insignificantly different from zero in column (2) – the occupational
proportions (especially the professional/managerial category) do a better job
at proxying for workforce skill than education.24 The variables generally
take their expected signs, although it is clear that there is some loss of
precision when a full set of fixed effects is added in column (3). Capital per
worker is strongly correlated with productivity, although the coefficient is
lower (0.21–0.25) than capital’s share of value added, which is about 30%.
Worker turnover has a significantly negative association with productivity
and R&D a significantly positive correlation. Younger workers (aged
between 16 and 24) are significantly less productive than the 35- to 44-year-
old group. Most importantly for our purposes, training has a statistically
significant and economically important effect on productivity according to
Table 1. The magnitude of the coefficient falls as we move to the more
rigorous specification which controls for fixed effects, but the change is not
dramatic. The estimates imply that raising the training variable by 1% point
(say, from the 1996 economy-wide mean of about 14–15%) is associated
with an increase in productivity of about 0.7%. We will return to the
plausibility of the magnitude of these effects in the last subsection of
section V.
The last three columns repeat the specifications but instead use
ln(wages) as the dependent variable. The most interesting contrast for
our purpose is the coefficient on training. As with productivity, training
enters the earnings equation with a consistently positive and significant
coefficient across all three columns. The magnitude of the coefficient is
lower in the wage equation than in the productivity equation – about half
the size. At face value, then, estimating the returns to training solely on the

24
This conclusion does not change if we break down the qualifications into four groups. Machin,
Vignoles and Galindo-Ruedo (2003) adopt a much finer classification of education using post-1992
LFS data where there are a larger number of observations. Exploiting the regional and industry
aspects of the aggregated data, they find some role for college proportion, even in fixed-effect
specifications.

Blackwell Publishing Ltd 2006

 TABLE 1
410

Training, productivity and wages

(1) (2) (3) (4) (5) (6)
ln(value added per worker) ln(wages)
Random effects Random effects Within groups Random effects Random effects Within groups
Training 0.788 (0.168) 0.700 (0.169) 0.696 (0.201) 0.425 (0.117) 0.344 (0.119) 0.365 (0.157)
ln(capital/worker) 0.252 (0.020) 0.244 (0.019) 0.212 (0.053) 0.058 (0.012) 0.051 (0.012) 0.069 (0.035)

Blackwell Publishing Ltd 2006

ln(hours/worker) 0.184 (0.181) 0.196 (0.181) 0.275 (0.207) 0.274 (0.123) 0.272 (0.126) 0.310 (0.116)
Lagged R&D intensity 1.628 (0.430) 1.390 (0.428) 1.251 (0.662)
0.212 (0.284)
0.356 (0.281)
0.942 (0.717)
Worker turnover
0.632 (0.206)
0.683 (0.207)
0.430 (0.332) 0.132 (0.143) 0.070 (0.145) 0.163 (0.202)
Occupations: base group is manual workers
Managers 0.487 (0.123) 0.282 (0.131) 0.324 (0.084) 0.195 (0.099)
Clerical 0.366 (0.174)
0.076 (0.190) 0.161 (0.121)
0.126 (0.121)
Personal/security
0.049 (0.355)
0.522 (0.371) 0.504 (0.250) 0.204 (0.223)
Sales people 0.443 (0.276)
0.078 (0.281)
0.037 (0.191)
0.328 (0.190)
Bulletin

No qualifications
0.251 (0.096)
0.036 (0.109) 0.107 (0.096)
0.145 (0.065)
0.033 (0.075) 0.101 (0.069)
Experience: base group age 35–44
Age 16–24
0.579 (0.170)
0.461 (0.172)
0.390 (0.175)
0.315 (0.118)
0.259 (0.121)
0.153 (0.119)
Age 25–34
0.341 (0.155)
0.282 (0.158)
0.314 (0.171)
0.198 (0.109)
0.155 (0.110)
0.196 (0.111)
Age 45–54
0.058 (0.158)
0.042 (0.156)
0.104 (0.160)
0.139 (0.110)
0.148 (0.111)
0.150 (0.101)
Age 55–64 0.178 (0.190) 0.244 (0.192) 0.142 (0.237)
0.263 (0.133)
0.263 (0.136)
0.271 (0.138)
Male 0.037 (0.097) 0.114 (0.099)
0.116 (0.128) 0.293 (0.064) 0.364 (0.065)
0.112 (0.078)
Small firm 0.068 (0.112) 0.016 (0.113) 0.005 (0.127)
0.118 (0.076)
0.126 (0.076)
0.056 (0.074)
Observations 968 968 968 968 968 968
Estimation period 1984–96 1984–96 1984–96 1984–96 1984–96 1984–96
Notes: Standard errors (robust to heteroskedasticity) are given in parentheses under coefficients. In the first three columns the dependent variable is ln(value added
per worker) and in the last three columns the dependent variable is ln(wages). Bold typeface indicates that the variable is significant at the 5% level. All regressions
include a full set of regional dummies (10), time dummies (12) and tenure dummies (6). Observations are weighted by number of individuals in a Labour Force
Survey industry cell. Random effects are estimated by generalized least squares. Within groups are estimated by least squares dummy variables (85 industries).
 The impact of training on productivity and wages 411

basis of wage equations would generate an underestimate of the importance

of work-related training.25
Turning to the other variables in the wage equation, the signs of most
of them are the same as those in the productivity equations, although there
are some differences. As expected, earnings are significantly higher in more
capital-intensive, hours-intensive and highly skilled industries. The R&D
coefficient is surprisingly negative (although insignificantly different from
zero), but this turns out to be because of mis-specified dynamics – including
longer lags of R&D demonstrates there is actually a positive correlation of
technology with wages.26
An important concern with Table 1 is that we do not allow for the
endogeneity of training or other suspected endogenous variables. To deal with
this, we implemented the GMM approach described in section III above.
Table 2 contains a summary of the main results.27 All the same variables are
included in these regressions as in Table 1, but we report only the key
coefficients to preserve space.
In column (1) we present the production function and in column (2) we
present the wage equation. The GMM estimates tell a similar story to the
within-groups estimates. Training has a positive and significant impact on
both productivity and wages, although the training coefficient in the
production function remains almost twice the size of the coefficient in the
wage equation (0.60 vs. 0.35). There are some minor changes to the other
coefficients – the coefficient on capital intensity has risen to 0.33 in the
production function, the R&D coefficient is positively signed in the wage
equation and the coefficient on hours is somewhat larger in magnitude than in
Table 1.
The diagnostics reported at the base of the table are also satisfactory –
there is no sign of second-order serial correlation (in the first-differences
residuals) and the Sargan test of over-identifying restrictions does not
reject. Note that the wage regression uses instruments dated t
3 and
before in the differenced equation (and dated t
2 in the levels equation).
This is because there were some signs of significant second-order serial
correlation using t
2 dated instruments in the wage equation, which
invalidates the Instrumental Variables (IVs) (we dropped one period in

25
A test of the equality of the coefficients on training in the wage and productivity equations
rejected equality at the 0.10 level for training (P-value ¼ 0.068). We would, however, expect the
coefficient on training in the wage equation (k
1) to be lower than in the production function as the
training coefficient in equation (7) is a(c
1). Although a test of the equality between k and c
cannot be rejected at the 0.05 level for the sample as a whole it can be rejected for the ‘low wage’
industries – see footnote 33 below.
26
Consistent with the findings of, inter alia, Bartel and Sicherman (1999).
27
Table B2 in Appendix B of Dearden et al. (2005), has more detailed results, and even more
detailed specifications are available from the authors or in Dearden, Reed and Van Reenen (2000).

Blackwell Publishing Ltd 2006

412 Bulletin

TABLE 2
Production functions and wage equations estimated by general method of moments (GMM)
(1)
ln(real value (2)
added per worker) ln(wages)
Training 0.602 (0.181) 0.351 (0.074)
ln(capital/worker) 0.327 (0.016) 0.106 (0.011)
ln(hours/worker) 0.498 (0.064) 0.489 (0.027)
Lagged R&D intensity 1.905 (0.262) 0.443 (0.182)
Proportion of employees who 0.306 (0.068) 0.160 (0.034)
are professionals or managers
Autocorrelation coefficient (q) 0.741 (0.015) 0.797 (0.013)
LM1 (d.f.)
4.892 (85)
6.053 (85)
[P-value] [0.00] [0.00]
LM2 (d.f.)
0.940 (85)
1.44 (85)
[P-value] [0.347] [0.158]
Sargan (d.f.) 8.819 (121) 11.83 (146)
Instruments (TRAIN)t
2,t
3, ln(Q/N)t
2,t
3, (TRAIN)t
3,..,t
5,
ln(Hrs/N)t
2,t
3, ln(K/N)t
2,t
3 ln(Q/N)t
3,..,t
5,
in differenced equations; ln(Hrs/N)t
3,..,t
5,
D(TRAIN)t
1, D ln(Hrs/N)t
1, ln(K/N)t
3,..,t
5 in
D ln(K/N)t
1 in levels equations differenced equations;
D(TRAIN)t
2,
D ln(Hrs/N)t
2,
D ln(K/N)t
2 in levels
equations
Estimation period 1984–96 1985–96
Observations 898 883
Notes: Estimation by GMM-SYS in Arellano and Bond (1998) DPD-98 package written in
GAUSS; one-step robust estimates reported. All regressions include the current values of all the
variables in columns (3) and (6) of Table 1 (i.e. worker turnover, occupational proportions, quali-
fications, age, tenure, gender, region, firm size and time dummies). Capital intensity, training, hours
and lagged productivity are always treated as endogenous. The other variables are assumed exo-
genous. One-step SEs (robust to arbitrary heteroskedasticity and autocorrelation of unknown form)
are given in parentheses under coefficients (variables significant at 5% level are in bold). LM1 (LM2)
is a Lagrange Multiplier test of first- (second)-order serial correlation distributed N[1, 0] under the
null (see Arellano and Bond, 1991). Sargan is a chi-squared test of the over-identifying restrictions.
Observations are weighted by number of individuals in a Labour Force Survey industry cell. Full
details in Table B2 in Appendix B of Dearden et al. (2005).

order to be able to use the longer lags in estimation). Using the (invalid)
instruments on the longer time period gave a coefficient (SE) on training
of 0.141 (0.067) in the wage equation.28

28
Using the shorter time period with longer-dated instruments in the production function gave a
coefficient (SE) on training of 1.043 (0.325). See Table B2 in Appendix B of Dearden et al. (2005)
for full details.

Blackwell Publishing Ltd 2006

 The impact of training on productivity and wages 413

TABLE 3
Robustness tests of the production function
Row Robustness test Observations Training coefficient (SE)
1 Original training coefficient in 968 0.696 (0.201)
production sector, Table 1 column (3)
2 Using ‘stock’ of trained workers 968 0.775 (0.189)
instead of flows
3 Using the balanced panel only 572 0.508 (0.289)
4 Conditioning on wage in 968 Training:
productivity regression (to 0.659 (0.219)
control for any residual
unobserved worker quality)
Wage coeff.:
0.099 (0.130)
5 Including service sectors 1,059 0.727 (0.206)
6 Include union density (only 547 Training:
available 1989–96) 0.604 (0.266)
Union:

0.177 (0.183)
7 Allow all industries to have 968 Mean of heterogeneous
different training coefficients coefficients: 0.505
8 Allow non-constant returns 968 0.725 (0.201)
9 Estimating a translog production 968 0.703 (0.201)
function
10 Estimation by non-linear least 968 0.518 (0.197)
squares
11 Estimation on 1993–2001 data 1,873 0.436 (0.188)
(region–industry cells)
Notes: These all use the specification in column (3) of Table 1 (unless otherwise specified).
Estimation by within groups; robust SEs in parentheses (except row 10). Bold typeface indicates that
the variable is significant at the 5% level.

Robustness of the results

We conducted a large number of robustness tests on the models in Tables 1
and 2. Table 3 reports some of these. Given the similarity of the within-groups
and GMM results, we performed these tests on the within-groups specifica-
tions of Table 1, column (3). The first row of Table 3 simply reports the
coefficient and SE from that column. Using the stock of trained employees
instead of the flow (calculated allowing for depreciation due to inter-firm
turnover) results in very similar results in row 2.29 Keeping only industries
29
The non-fixed-effects results were significantly different, but the deviations around the fixed effect
in the training stock are dominated by the flow, explaining the similarity of the results. In addition to
using the empirical worker turnover rates, we assumed an exogenous depreciation of training at 40%
per annum (see Appendix A of Dearden et al., 2005). The coefficient was stable to reasonable changes
in these parameters (e.g. increasing the depreciation rate to 50% p.a. increased the coefficient to 0.81,
to 60% to 0.82; decreasing the depreciation rate to 30% reduced the coefficient to 0.70).

Blackwell Publishing Ltd 2006

414 Bulletin

that had 14 continuous years of data (balanced panel) in row 3 means losing
40% of the observations; the coefficient falls, but the change is not significant
(cf. Nickell, 1981). The fourth row includes average wages on the right-hand
side of the production function as a measure of unobserved worker quality;
although wages have a positive coefficient, the training association remains
robust. In row 5 we include all the service sectors, ignoring our concerns over
data quality. The coefficient on training rises to 0.73 and remains significant.
As training is correlated with unionization, we could be picking up
‘collective voice’ effects in the main results. Union membership is only
available in LFS since 1989. Despite the loss in sample in row 6, the training
effect is robust to inclusion of union density (density is insignificantly
negatively associated with productivity). In row 7 we allow the training effect
to be different in each of the 85 industries; the mean of these heterogeneous
coefficients is close to the pooled results. The next two rows allow for more
general functional forms, first relaxing constant returns (row 8) and then
estimating a translog production function (row 9); in both cases, the training
coefficient is essentially unchanged. Row 10 gives the results from a non-
linear least squares estimation of equation (8) again showing no significant
difference.
We also compared our results with a recent paper (Machin et al., 2003)
which has built up similar data to our own covering a more recent period
and exploiting the larger size of the LFS post-1992 to construct industry-by-
region cells. Against these advantages, their data set has a shorter time-
series component (1993–2001) and lacks some of the covariates we use.
Re-estimating identical specifications on their data set gives an estimate of the
training association with productivity of 0.436 with a SE of 0.188 (see row 11
of Table 3). This is lower, but is still significant and remains well within two
SEs of our main results.30 On our data set, we tested whether there was a
tendency for the training coefficient to fall (or rise) over time in the production
function, but we found it to be stable.31
Does the ‘wedge’ between the wage and productivity effect of training
arise from specific human capital or imperfect competition? Under most forms
of imperfect competition, we conjectured that the wedge would be larger in
those industries where workers were earning less than would be implied by
their human capital (i.e. inter-industry wage premiums were low). This could
be because the ‘low-paying’ industries were monopsonistic with large search

30
The specification is identical to column (3) of Table 1 except we drop the occupational pro-
portions and R&D and include employment. On our data, this gives a coefficient (SE) on TRAIN of
0.732 (0.205).
31
For example, interacting TRAIN with a trend in the production function gave a coefficient of
0.003 with a SE of 0.044.

Blackwell Publishing Ltd 2006

 The impact of training on productivity and wages 415

frictions or because workers are more able to capture the quasi-rents from
training in the ‘high-paying’ industries.
In order to identify such industries, we used estimates of inter-industry
wage premiums taken from the US Current Population Survey (CPS).32 We
matched the US industries to the UK industries and split the sample at the
median sectoral wage premium. Allowing an interaction between training and
this industry split revealed that the wedge between the training effect on
productivity and the training effect on wages was solely within the ‘low-wage’
industries. To be precise, including an interaction in the wage equation
between training and low-wage industries gave a coefficient (SE) of
0.664
(0.196) on the interaction and 0.531 (0.113) on the linear training effect.
In the production function, the interaction was 0.332 (0.297) – positive but
insignificant (the linear training term took a coefficient of 0.612 with an SE of
0.172).33 In other words, in the ‘high wage premium’ industries, there was no
significant difference between the impact of training on productivity and the
impact of training on wages. The fact that our results are driven by the wedge
in low-paying sectors is tentative evidence in favour of a monopsony/search
interpretation.
This evidence is open to the critique that firm-specific training may be
systematically more prevalent in the low-wage sectors (although a priori the
usual view is that ‘good jobs’ are more likely to have more specific skills).
There are several questions in LFS that could be interpreted as general
vs. specific training, so we used them to see if the coefficients differed
significantly with training type – they did not. For example, there are
questions related to off-the-job training (more general) and on-the-job training
(more specific). The proportion of off-the-job training produced a coefficient
(SE) of 0.005 (0.018) when added to the wage regression and a coefficient
(SE) of 0.018 (0.029) when added to the production function. We view this
not as any rejection of specific human capital theory per se, but rather as an
indication the type of human capital is intrinsically difficult to measure.
Furthermore, the LFS questions are not asked in all years and have many
missing values.

32
Estimating inter-industry wage premiums from UK wages would have been more problematic as
these could reflect endogenous influences – US wage-setting will be driven by the structural char-
acteristics of the industries in question. These US inter-industry wage premiums were generated from
individual-level wage regressions from the 1986 CPS merged outgoing rotation files. The wage
regressions included years of schooling, a quartic in experience, gender, marital status, gender ·
marriage interactions, race, Standard Metropolitan Statistical Area and regional dummies. The data
were kindly provided by Steve Pischke (see Acemoglu and Pischke, 2003, for details).
33
A test of the equality between the effects of training on wages (k) and on productivity (c) can be
rejected at the 0.05 level for the ‘low wage’ industries (P-value ¼ 0.001), but cannot be rejected for
the ‘high wage’ industries (P-value ¼ 0.752).

Blackwell Publishing Ltd 2006

416 Bulletin

Quantifying the effects of training

Our key qualitative conclusions are: first, there is a significant impact of
training on productivity; and, secondly, the effects of training on produc-
tivity are larger than the effects of training on wages. But, quantitatively, how
economically significant is the magnitude of the training effect?
Interpreting the exact magnitude of the coefficients is difficult, but the
implied effects are large. From Tables 1 and 2, we conservatively take the
coefficient on training in the productivity regressions to be about 0.6 and
the coefficient on training in the wage regressions to be about 0.3. This would
imply that a 10% point increase in the training measure is associated with a
6% increase in productivity and a 3% increase in wages.
Relative to the returns-to-schooling literature, the training impacts appear
high.34 Card (1999) puts the impact of a year of schooling on wages at about
10%, so our baseline impact of 0.3 is about three times as large. Given that the
typical time in training during the 4-week period is under a month (the median
is 2 weeks, the mean is higher), the returns to a month of training appear even
more impressive. For example, an increase in our key variable, TRAIN, of
10% would imply a typical worker only spent 5% extra of his time in training,
if training spells were on average 2 weeks long.
Of course, there may be remaining econometric problems we have not
controlled for generating this difference. But assuming the training effect is
not a statistical artefact, there remain at least two possible explanations for the
training coefficients being larger than conventional estimates of the return to
schooling. First, work-related training may have a higher private return than
schooling as training is more directed at raising productivity in employment.
Training is also likely to have a faster rate of depreciation than schooling,
so it requires a higher year-on-year return in order to give incentives for
investment.35 Secondly, there may be externalities associated with training
that are missed in the conventional schooling literature, which focuses on
private returns whereas we look at returns to the industry as a whole (cf.
Moretti, 2004).
To investigate the externality issue, we estimated some individual-level
wage regressions on the LFS panel. If the private returns to training are higher
than the social returns, we might expect to see a similarly high coefficient in
the individual-level wage regression. We used the individual-level equivalents
of the variables in the industry-level regressions. To construct the proportion
of the year spent in training, we used the LFS panel which follows individuals
34
Compared with existing UK estimates of the training effects on wages (e.g. Booth, 1993;
Blundell et al., 1996), our estimates are actually lower (see Dearden et al., 2000, for a detailed
comparison).
35
See Heckman, Lochner and Todd (2003) for a recent discussion of interpretation of the schooling
coefficient in wage regressions.

Blackwell Publishing Ltd 2006

 The impact of training on productivity and wages 417

for five quarters and asks individuals the training question in each quarter. We
defined a dummy variable (TRAIND4) indicating whether the individual had
been involved in some training in all of the previous four quarters. We also
defined dummies for if the individual had been in training for three quarters
(TRAIND3), two quarters ( TRAIND2), one quarter (TRAIND1) or not at all
(TRAIND0). Using TRAIND0 as the omitted base, the results we obtained
from a typical regression were:36
lnðwageÞ ¼ 0:165ð0:033ÞTRAIND4 þ 0:092ð0:023ÞTRAIND3
þ 0:125ð0:019ÞTRAIND2 þ 0:078ð0:015ÞTRAIND1 þ controls:
Longer lengths of time in training are associated with significantly higher
wages.37 The coefficient on receiving training in all four quarters was 0.165;
this is comparable with the industry-level coefficient of 0.350. Taken literally,
this would suggest that about half of the impact of training on wages at the
industry level is attributable to externalities.
If we include a set of industry dummies (which will include potential
spillovers), the coefficient on TRAIND4 falls from 0.16 to 0.13. If we also
include the initial wage in the first quarter (to control for unobserved
heterogeneity), the coefficient falls even further to 0.079. So these impacts of a
‘year’ of training are rather similar to the conventional impacts of the returns
to a year in school.
Our conclusion from this exercise is that the larger magnitude of the
training effects in this paper primarily reflects our strategy of estimating at a
level above the individual worker. This was forced upon us by the absence of
adequate data on firm productivity and training, but also because of our desire
to incorporate externalities. The results are therefore consistent with a story
that stresses externalities to training.
Even if there remain econometric problems that have caused us to
overestimate the impact of training at the industry level, it is hard to see why
this would not also bias upwards the training coefficient in the production
function and wage equation to a similar extent. Therefore, even if one disputes
the exact quantitative magnitude of the training effect, our key qualitative
conclusion that the productivity impact of training is greater than the wage
impact should still be valid (this is also a feature of the firm-level results in
Appendix B of Dearden et al., 2005).

36
Estimation was by OLS; robust SEs are given in parentheses. Controls include gender, age, areas
(20), employer size, occupational dummies (8), no qualification dummy, and a dummy for part-time
status. Results are for the production sector only. The quarterly LFS panel 1997–98 was used as two
wage observations per individual did not exist in the LFS prior to this. There were 3,998 observa-
tions. Full results are available on request from the authors.
37
The training effects are not monotonic. There is even a perverse fall in the coefficient on being in
training three relative to two quarters, although the coefficients are not significantly different.

Blackwell Publishing Ltd 2006

418 Bulletin

VI. Conclusions
In this paper, we have examined the issue of the impact of private sector
training on productivity. Rather than simply use wages as a measure of
productivity, we have presented (for the first time) estimates of the impact of
training on productivity over a long time period. We have assembled a dataset
that aggregates individual-level data on training and establishment data on
productivity and investment into an industry panel covering 1983–96. We
controlled for unobserved heterogeneity and the potential endogeneity of
training using a variety of methods including GMM system estimation.
Using these new data, we have identified a statistically and economically
significant effect of training on productivity in the UK. An increase of 1%
point in the proportion of employees trained is associated with about a 0.6%
increase in productivity and a 0.3% increase in wages. The impact of training
on productivity is robust to a large number of robustness tests.
We argued that the methodologies in the existing literature may
underestimate the importance of training. The focus on wages as the only
relevant measure of productivity ignores the additional productivity benefits
the firm may capture. The coefficient of training in the production function
was around twice as large as the coefficient in the wage equation. This result
could occur even under standard specific human capital theory. But it could
also arise for a number of other reasons due to imperfect competition in the
labour market (and we have presented some evidence consistent with this
hypothesis). Clearly, further research is needed to distinguish between these
theories.
Finally, a comparison between the industry- and individual-level wage
regressions suggests that our industry-level analysis may capture externalities
from training that are missed out in the micro-level studies. An important
avenue of future research would include probing the returns to training by
combining enterprise data with industry-level data to investigate the external-
ities story in greater detail.

Final Manuscript Received: September 2005

References
Acemoglu, D. and Angrist, J. (2000). ‘How large are the social returns to education? Evidence
from compulsory schooling laws’, NBER Macroeconomics Annual, Vol. 15, pp. 9–59.
Acemoglu, D. and Pischke, S. (1999). ‘The structure of wages and investment in general
training’, Journal of Political Economy, Vol. 107, pp. 539–572.
Acemoglu, D. and Pischke, S. (2003). ‘Minimum wages and on the job training’, Research in
Labor Economics, Vol. 22, pp. 159–202.
Aghion, P. and Howitt, P. (1998). Endogenous Growth Theory, MIT Press, Cambridge, MA.

Blackwell Publishing Ltd 2006

 The impact of training on productivity and wages 419

Arellano, M. and Bond, S. (1991). ‘Some tests of specification for panel data: Monte Carlo
evidence and an application to employment equations’, Review of Economic Studies, Vol. 58,
pp. 277–297.
Arellano, M. and Bond, S. (1998). Dynamic Panel Data Estimation Using DPD for GAUSS,
Mimeo, Institute for Fiscal Studies, London.
Ballot, G. and Taymaz, E. (1998). Firm Sponsored Training and Performance. A Comparison
Between French and Swedish Firms, Mimeo, ERMES, Paris.
Ballot, G., Fakhfakh, F. and Taymaz, E. (1998). ‘Formation continue, recherche et développe-
ment, et performance des entreprises’, Formation Emploi, Vol. 64, pp. 43–55.
Ballot, G., Fakhfakh, F. and Taymaz, E. (2002). Who Benefits from Training and R&D: the
Firm or the Workers?, Economic Research Center Working Papers in Economics 02/01,
Middle East Technical University, Ankara, Turkey.
Barron, J. M., Black, D. A. and Loewenstein, M. A. (1989). ‘Job matching and on-the-job
training’, Journal of Labour Economics, Vol. 7, pp. 1–19.
Bartel, A. P. (1994). ‘Productivity gains from the implementation of employee training pro-
grammes’, Industrial Relations, Vol. 33, pp. 411–425.
Bartel, A. P. (1995). ‘Training, wage growth, and job performance: evidence from a company
database’, Journal of Labour Economics, Vol. 13, pp. 401–425.
Bartel, A. and Sicherman, N. (1999). ‘Technological change and wages: an inter-industry
analysis’, Journal of Political Economy, Vol. 107, pp. 285–325.
Barrett, A. and O’Connell, P. (2001). ‘Does training generally work? The returns to in-company
training’, Industrial and Labor Relations Review, Vol. 54, pp. 647–683.
Black, S. E. and Lynch, L. M. (1996). ‘Human-capital investments and productivity’, American
Economic Review, Vol. 86, pp. 263–267.
Black, S. E. and Lynch, L. M. (2001). ‘How to compete: the impact of workplace practices and
information technology on productivity’, Review of Economics and Statistics, Vol. 83,
pp. 434–445.
Blanchflower, D. and Lynch, L. M. (1992). Training at Work: A Comparison of US and British
Youths, Discussion Paper No. 78, Centre for Economic Performance, London School of
Economics, London.
Blundell, R. and Bond, S. (1998). ‘Initial conditions and moment restrictions in dynamic panel
data models’, Journal of Econometrics, Vol. 87, pp. 115–143.
Blundell, R. and Bond, S. (2000). ‘GMM estimation with persistent panel data: an application
to production functions’, Econometric Reviews, Vol. 19, pp. 321–340.
Blundell, R., Dearden, L. and Meghir, C. (1996). The Determinants of Work-related Training in
Britain, Institute for Fiscal Studies, London.
Boon, M. and van der Eijken, B. (1997). Employee Training and Productivity in Dutch
Manufacturing Firms, Research Paper No. 9716, Department of Statistical Methods,
Statistics Netherlands, Voorburg.
Booth, A. (1991). ‘Job-related formal training: who receives it and what is it worth?’, Oxford
Bulletin of Economics and Statistics, Vol. 53, pp. 281–294.
Booth, A. (1993). ‘Private sector training and graduate earnings’, Review of Economics and
Statistics, Vol. 75, pp. 164–170.
Booth, A., Francesconi, M. and Zoega, G. (1999). ‘Training, rent sharing and unions’, Mimeo,
University of Essex, Essex.
Card, D. (1999). ‘The causal effect of education on earnings’, in Ashenfelter O. and Card D.
(eds), Handbook of Labour Economics, Elsevier, Amsterdam, pp. 1801–1863.
Carriou, Y. and Jeger, F. (1997). ‘La formation continue dans les entreprises et son retour sur
investissement’, Economie et Statistique, Vol. 303, pp. 45–58.

Blackwell Publishing Ltd 2006

420 Bulletin

Dearden, L., Reed, H. and Van Reenen, J. (2000). Who Gains when Workers Train? Training
and Corporate Productivity in a Panel of British Industries, Working Paper No. 00/04,
Institute for Fiscal Studies.
Dearden, L., Reed, H. and Van Reenen, J. (2005). The Impact of Training on Productivity and
Wages: Evidence from British Panel Data, IFS Working Papers No. W05/12, Institute for
Fiscal Studies, August.
Delame, E. and Kramarz, F. (1997). ‘Entreprises et formation continue’, Economie et Prevision,
Vol. 127, pp. 63–82.
Green, F. and Steedman, H. (1997). Into the Twenty First Century: An Assessment of British
Skill Profiles and Prospects, Special Report, Centre for Economic Performance, London
School of Economics, London.
Greenhalgh, C. (2002). ‘Adult vocational training in France and Britain’, Fiscal Studies,
Vol. 23, pp. 223–263.
Greenhalgh, C. and Stewart, M. (1987). ‘The effects and determinants of training’, Oxford
Bulletin of Economics and Statistics, Vol. 49, pp. 171–189.
Griliches, Z. (1992). ‘The search for R&D spillovers’, Scandinavian Journal of Economics,
Vol. 94, pp. S29–S47.
Grunfeld, D. and Griliches, Z. (1960). ‘Is aggregation necessarily bad?’, Review of Economics
and Statistics, Vol. 42, pp. 1–13.
Hægeland, T. and Klette, T. (1999). ‘Do higher wages reflect higher productivity? Education,
gender and experience premiums in a matched plant-worker data set’, in Haltiwanger J., Lane
J., Spletzer J. R., Theeuwes J. and Troske K. (eds), The Creation and Analysis of Employer–
Employee Matched Data, North Holland, Amsterdam.
Haskel, J., Hawkes, D. and Pereira, S. (2003). Skills, Productivity in the UK
Using Matched Establishment, Worker and Workforce Data, Ceriba, Discussion Paper,
May.
Heckman, J., Lochner, L. and Todd, P. (2003). Fifty Years of Mincer Earnings Regressions,
National Bureau of Economic Research, Working Paper No. 9732.
Hellerstein, J. and Neumark, D. (1999). ‘Sex, wages and productivity: an empirical analysis of
Israeli firm level data’, International Economic Review, Vol. 40, pp. 95–123.
Hellerstein, J., Neumark, D. and Troske, K. (1999). ‘Wages, productivity and worker charac-
teristics: evidence from plant level production functions and wage equations’, Journal of
Labor Economics, Vol. 17, pp. 409–446.
Holzer, H. (1990). ‘The determinants of employee productivity and earnings’, Industrial
Relations, Vol. 29, pp. 403–422.
Holzer, H., Block, R., Cheatham, M. and Knott, J. (1993). ‘Are training subsidies for firms
effective?’, Industrial and Labor Relations Review, Vol. 46, pp. 625–636.
Ichniowski, C., Shaw, K. and Prennushi, G. (1997). ‘The effects of human resource manage-
ment practices on productivity’, American Economic Review, Vol. 87, pp. 291–313.
Jones, P. (2001). ‘Are educated workers really more productive?’, Journal of Development
Economics, Vol. 64, pp. 57–79.
de Koning, J. (1994). ‘Evaluating training at the company level’, in McNabb R. and Whitfield K.
(eds), The Market for Training, Avebury, Aldershot, pp. 87–105.
Krueger, A. and Rouse, C. (1998). ‘The effects of workplace education on earnings, turnover
and job performance’, Journal of Labor Economics, Vol. 16, pp. 61–94.
Lazear, E. (1979). ‘Why is there mandatory retirement?’, Journal of Political Economy, Vol. 87,
pp. 1261–1284.
Lillard, L. A. and Tan, H. W. (1992). ‘Private sector training: who gets it and what are its
effects?’, Research in Labour Economics, Vol. 13, pp. 1–62.

Blackwell Publishing Ltd 2006

 The impact of training on productivity and wages 421

Lynch, L. M. (1992). ‘Private sector training and the earnings of young workers’, American
Economic Review, Vol. 82, pp. 299–312.
Machin, S., Vignoles, A. and Galindo-Rueda, F. (2003). Sectoral and Area Analysis of the
Economic Effects of Qualifications and Basic Skills, Research Report No. RR465, Depart-
ment for Education and Skills, London.
Moretti, E. (2004). ‘Workers’ education, spillovers and productivity: evidence from Plant-Level
Production Functions’, American Economic Review, Vol. 94, pp. 656–690.
National Skills Task Force (1998). Towards a National Skills Agenda, Department for Edu-
cation and Employment, Sudbury.
Nickell, S. J. (1981). ‘Biases in dynamic models with fixed effects’, Econometrica, Vol. 49,
pp. 1417–1426.
O’Mahony, M. (1998). ‘Anglo-German productivity differences: the role of broad capital’,
Bulletin of Economic Research, Vol. 50, pp. 19–36.
Winkelmann, R. (1994). Training, Earnings and Mobility in Germany, Discussion Paper No.
982, Centre for Economic Policy Research.

Blackwell Publishing Ltd 2006