Professional Documents
Culture Documents
On-the-Job Training surveys the recent literature from both a theoretical and empirical
perspective. The analysis of how individuals obtain and are paid for their skills is fundamental
to labor economics. The basic idea of human capital theory is that workers and firms invest in
workers’ skills in order to increase their productivity, much as persons invest in financial or
physical assets to earn income. Workers develop many skills through formal education not tied On-the-Job Training
On-the-Job Training focuses on recent literature including empirical research using direct Harley Frazis and Mark Loewenstein
measures of training and theoretical papers inspired by findings from this empirical work. The
authors presents a theoretical model showing that costs and returns to general human capital
may be shared if training increases mobility costs, if there are constraints on lowering wages,
or if there is uncertainty about the value of training at competing employers. This model
analyzes the choice of the amount of training, emphasizing the influence of whether the
employer can commit to training prior to employment. In addition, the model implies that firms
will attempt to match low-turnover workers with training opportunities, which is supported by
the empirical literature.
now
the essence of knowledge
On-the-Job-Training
On-the-Job-Training
Harley Frazis
Mark A. Loewenstein
Boston – Delft
Foundations and Trends
R
in
Microeconomics
All rights reserved. No part of this publication may be reproduced, stored in a retrieval
system, or transmitted in any form or by any means, mechanical, photocopying, recording
or otherwise, without prior written permission of the publishers.
Photocopying. In the USA: This journal is registered at the Copyright Clearance Cen-
ter, Inc., 222 Rosewood Drive, Danvers, MA 01923. Authorization to photocopy items for
internal or personal use, or the internal or personal use of specific clients, is granted by
now Publishers Inc for users registered with the Copyright Clearance Center (CCC). The
‘services’ for users can be found on the internet at: www.copyright.com
For those organizations that have been granted a photocopy license, a separate system
of payment has been arranged. Authorization does not extend to other kinds of copy-
ing, such as that for general distribution, for advertising or promotional purposes, for
creating new collective works, or for resale. In the rest of the world: Permission to pho-
tocopy must be obtained from the copyright owner. Please apply to now Publishers Inc.,
PO Box 1024, Hanover, MA 02339, USA; Tel. +1 781 871 0245; www.nowpublishers.com;
sales@nowpublishers.com
now Publishers Inc. has an exclusive license to publish this material worldwide. Permission
to use this content must be obtained from the copyright license holder. Please apply to now
Publishers, PO Box 179, 2600 AD Delft, The Netherlands, www.nowpublishers.com; e-mail:
sales@nowpublishers.com
Foundations and Trends
R
in
Microeconomics
Volume 2 Issue 5, 2006
Editorial Board
Editor-in-Chief:
W. Kip Viscusi
Vanderbilt University
Editors
Richard Carson, UC San Diego (environmental economics)
Joseph Harrington, Johns Hopkins University (industrial organization)
Tom Kniesner, Syracuse University (labor economics)
Mark V. Pauly, University of Pennsylvania (health economics)
David Wildasin, University of Kentucky (public economics)
Peter Zweifel, University of Zurich (insurance economics)
Editorial Scope
On-the-Job-Training
1
Bureau of Labor Statistics, 2 Massachusetts Ave. NE, Suite 4945,
Washington D.C. 20212, Frazis.Harley@bls.gov
2
Bureau of Labor Statistics, 2 Massachusetts Ave. NE, Suite 4130,
Washington D.C. 20212, Loewenstein.Mark@bls.gov
Abstract
The analysis of how individuals obtain and are paid for their skills is
fundamental to labor economics. The basic idea of human capital theory
is that workers and firms invest in workers’ skills in order to increase
their productivity, much as persons invest in financial or physical assets
to earn income. Workers develop many skills through formal education
not tied to an employer, but an important part of their skills are learned
on the job. This paper is a survey of the recent literature on on-the-job
training, both theoretical and empirical.
Contents
1 Introduction 1
2 Measuring Training 3
ix
6.1 Estimating the Effect of Training on Wages 49
6.2 Estimating the Effect of Training on Productivity 63
6.3 Estimating the Effect of Training on Job Mobility 66
7 Conclusion 71
References 73
1
Introduction
The analysis of how individuals obtain and are paid for their skills is
fundamental to labor economics. The basic idea of human capital theory
is that workers and firms invest in workers’ skills in order to increase
their productivity, much as persons invest in financial or physical assets
to earn income. Workers develop many skills through formal education
not tied to an employer, but an important part of their skills are learned
on the job. This paper is a survey of the recent literature on on-the-job
training, both theoretical and empirical.
While the roots of human capital theory (including the metaphor of
skills as capital) go back at least to Adam Smith (1904) modern human
capital theory was developed in the late 1950s by such economists
as Theodore Schultz (1962), Jacob Mincer (1962), and Gary Becker
(1962). For a period of some two to three decades, the theory of on-
the-job training was dominated by Becker’s (1962) analysis of general
and specific human capital. Empirical work followed the lead of Min-
cer (1962, 1974), who imputed the amount of on-the-job training from
wage-experience profiles.
Because data on the actual amount of on-the-job training were not
available, Mincer’s attempts to measure such training were indirect.
1
2 Introduction
3
4 Measuring Training
1 Fraziset al. (1998) report an anomalously high figure of 70% in SEPT95; this may be due
at least partly to restricting the sample to large establishments and to non-response.
6 Measuring Training
There are two distinct decisions that must be made with respect to
training. First, employers and workers must decide how much train-
ing to undertake. Second, employers and workers must determine how
to share the cost and return to training. Labor turnover considerations
play a fundamental role in shaping these decisions. Unlike physical cap-
ital, workers cannot sell their human capital. When a worker leaves an
employer, the worker’s human capital goes with him. The employer
loses the opportunity to derive any further benefits from it and the
worker loses the opportunity to use it at the employer’s workplace.
Thus, the employer’s and worker’s willingness to invest in training will
depend on the likelihood of a quit or dismissal in the future. Further-
more, as discussed below, decisions about the division of the return and
cost to training between employer and worker will be heavily influenced
by the consequent effects on turnover.
Becker’s (1962) distinction between general and specific human cap-
ital is a key concept in thinking about the relationship between training,
turnover, and the division of returns to training. Specific skills are only
useful at one employer, while general skills make a worker more pro-
ductive at many employers. Training that teaches a worker about an
9
10 The Division of the Cost and Return to Training
1 As examples of turnover bonds, Black and Loewenstein cite the article, “Firms Forcing
Employees to Repay Some Costs if They Quit Too Soon,” in The Wall Street Journal,
Tuesday, 16 July 1985, which indicates that corporations such as Electronic Data Sys-
tems, General Dynamics, McDonald Douglas, and Northrop required employees to repay
relocation costs if they quit within a specified period of time, usually one year, Lockheed
required employees to reimburse educational expenses if they quit within one year, and
American Airlines required pilots to reimburse on a prorated basis their $10,000 training
expense.
2 For a discussion of a fuller set of possible contracts, see Hall and Lazear (1984). Hall and
Lazear focus on demand uncertainty. By way of contrast, Black and Loewenstein (1998)
focus on uncertainty that is entirely match specific. However, the issues are fundamentally
the same. When there is imperfect information, no practical labor market contract will
eliminate all inefficient separations and it becomes necessary to resort to a second best
solution to allocate labor.
12 The Division of the Cost and Return to Training
wA = H + γh. (3.1)
π2 = D + η, (3.2)
D ≡ H + h − w2 (3.3)
13
is the rent the employer extracts when η is zero. The employer dismisses
the worker if π2 < 0. Letting g(·) denote the density function of the
random variable η, the probability of a dismissal or layoff is simply
Z −D
L= g(η)dη. (3.4)
−∞
where f (·) denotes the density function of the random variable ε and
εc = D − (c + (1 − γ)h) (3.6)
is the minimum value of ε such that the worker does not quit. The
expected gain to the worker from his or her match with the employer
is given by
is driven to zero, or
Γ = π + λ(U − U A ), (3.10)
To interpret (3.11), note that the loss imposed on the worker from
a dismissal is the difference between the utility the worker would
have received had the worker stayed with the employer and the utility
the worker receives when moving to another job, or U2 − U2A = (w +
ε) − (H + γh − c) = (H + h − D + ε) − (H + γh − c) = ε − εc . The
amenity value ε is initially unknown, so the worker’s expected loss
R∞
from a dismissal is εc (ε − εc )f (ε)dε. Since the marginal effect of
an increase in the wage on the probability of a dismissal is given
by (∂L/∂D)(∂D/∂w) = g(−D), the right-hand side of (3.11) is the
marginal effect of an increase in the wage on the expected loss to the
15
∂D(η) 1
= < 0, (3.13)
∂η −1 + m0
which implies that ∂w2 /∂η > 0: The second-period wage is increasing in
the productivity shock η. To retain well-matched workers, the employer
offers a share, but generally not all, of the return to a favorable match-
specific productivity shock. As in the fixed-wage model, there are inef-
ficient quits in the no-wage-commitment model. The worker quits when
ε < εc = D − c + (γ − 1)h = m − η − c + (γ − 1)h, but it is only effi-
cient to quit when ε = −η − c + (γ − 1)h. However, there are no inef-
ficient dismissals. Rather than dismiss the worker when η is low, the
employer simply lowers the wage. Following up on this point, Black and
Loewenstein (1997) note that the fixed-wage and no-commitment sce-
narios can be considered as special cases of a more general contract that
specifies a wage floor, but allows the employer to offer a higher wage
should he choose to do so. The wage floor leads to inefficient dismissals,
but limits the inefficient quits that will result from rent extraction by
the employer.
are not binding, ∂w2 /∂h = β, so that the employer and the worker share the return to
training.
3.1. Why Employers May Share the Return to General Training 17
Let D0 denote the optimal value of D, that is, the value that satis-
fies (3.11). Differentiating (3.11), (3.9) and (3.5) yields
where
Z ∞ Z ∞
c 0 c
M ≡ f (ε ) g(η)dη + f (ε ) (D + η)g(η)dη
−D −D
Z ∞ Z ∞ −1
0 c
+ g(−D) f (ε)dε + g (−D) (ε − ε )f (ε)dε
εc εc
Z ∞ Z ∞
c 0 c
× f (ε ) g(η)dη + f (ε ) (D + η)g(η)dη .
−D −D
capital in the production function nor labor market frictions are required for employers to
share the return to training. Rather, he notes that the “presence of specific training cre-
ates quasi-rents that have to be divided ex-post between workers and firms according to
the outside option principle. . . When the surplus is shared, the firm appropriates a share
of the returns on general and specific training ex-post.” Specifically, see the discussion in
20 The Division of the Cost and Return to Training
footnote 3. Note that one does not obtain Balmaceda’s result when one assumes a Nash
bargaining solution.
3.1. Why Employers May Share the Return to General Training 21
and ∂w1 /∂h > −k 0 (h). That is, if it is more costly for higher skilled
workers to change jobs, then trained workers’ higher productivity will
only be partially reflected in the post-training wage and employers will
share the cost of general training.
Acemoglu and Pischke (1999b) point out that costly search gives
prospective employers some monopsony power, preventing workers from
capturing the full value of their marginal product if they move to
another job. Similarly, Stevens (1994) presents a model in which a lim-
ited number of firms bid for a trained worker’s services. In equilibrium,
the worker moves to the employer at which he or she has the highest
value of marginal product, but the worker receives a wage equal to his or
her second highest value of marginal product. Frazis and Loewenstein
(2006) provide another reason why a worker moving to a new employer
will not receive the full value of marginal product: the existence of an
“equity norm” that prevents an employer from paying retained work-
ers less than equally productive experienced workers hired from the
outside. It is easy to imagine that an employer’s senior workers will
be unhappy and put forth less effort, if they receive a lower wage than
other experienced workers who are no more productive, but who simply
began their careers at other firms. Such behavior seems consistent with
both casual observation of the labor market and experimental studies
cited in Akerlof and Yellen (1990). (And equity norm considerations
might help explain why many employers have secrecy rules concerning
workers’ pay.)6
Whether due to monopsony or equity norm considerations, a reduc-
tion in the wage a skilled worker can receive at a new employer reduces
the worker’s optimal share of the cost and return to human capi-
tal investment. Formally, let the wage the worker can receive at an
6 Ransom (1993) finds that controlling for experience, wages at large research universi-
ties decline with tenure, which would seem to be inconsistent with the existence of an
equity norm. As Ransom (1993) and Black and Loewenstein (1991) note, large distances
between universities lead to high mobility costs in the academic labor market. In the con-
text of a multi-period model where employers make take-it-or-leave-it offers, Black and
Loewenstein (1991) show that this can lead to a declining wage profile, as workers who
stay reveal that their moving costs are particularly high. One might hypothesize that
in the absence of equity norm considerations, wages at universities would decline even
more with tenure. Or perhaps equity norm considerations in academe are weaker because
complementarities in production are less pronounced.
22 The Division of the Cost and Return to Training
wA = H + h − DA , (3.10 )
εc = D − DA − (c + (1 − γ)h). (3.60 )
Note too that other than the changed definition of εc , condition (3.11)
is unchanged. This condition implicitly defines D0 as a function of
DA : D0 = χ(DA ). Differentiating D0 with respect to DA , one finds that
χ0 = M : a reduction in the alternative wage causes a partial reduction
in the second-period wage offered by a worker’s initial employer.
In the equity norm model, DA is also a function of D0 : a reduc-
tion in the wage paid to an employer’s senior workers means a fall
in the wage paid to a skilled worker who changes jobs. The equity
norm thus amplifies an initial tendency toward employer sharing of the
return to general human capital acquisition, as the wage compression
and sharing effects reinforce each other. More specifically, competition
for experienced workers will ensure that they do not receive less than
retained workers. If employers hire a mix of inexperienced and experi-
enced workers, labor market equilibrium thus requires that D0 = DA .
Note too that employers will choose to hire a mix of inexperienced
and experienced workers if inexperienced and experienced workers are
strong complements in production, something which seems consistent
with casual observation. For example, it may be efficient to place less
skilled, inexperienced workers in less demanding tasks and let experi-
enced workers concentrate on certain critical tasks for which they are
better suited.
Let D∗ denote the equilibrium value of D. Using (3.60 ) and (3.11),
it is straightforward to show that
(c0 + (1 − γ))
∂D∗ /∂h = . (3.140 )
1−M
Comparing (3.14) and (3.140 ), one sees that the existence of an equity
norm will amplify an initial tendency by employers to share in the costs
and returns to general human capital investment.
3.2. Empirical Evidence on Sharing of General Human Capital 23
27
28 The Choice of Training
wage the employer pays in period 1. The final term in (4.1) captures
this effect.
As discussed above, there are no dismissals when the employer does
not make a wage commitment. Noting that the employer’s second-
period wage offer and hence the worker’s quit probability and reserva-
tion amenity level depend on η, the marginal return to training becomes
Z ∞ Z ∞
θ(h) = (1 − Q(η)g(η)dη + (γ − c0 (h)) (Q(η)
−∞ −∞
c
− f (ε (η))(η + D))g(η)dη, (4.10 )
Note that the employer continues to internalize the worker’s expected
return to training upon separation because the worker is willing to pay
for this in the form of a lower starting wage.
Finally, consider the choice of training when the worker places no
value on the employer’s training commitment in period 1. Formally, this
means that the training the worker expects to receive and the alter-
native wage the worker expects when switching employers are inde-
pendent of the training level that the employer actually chooses. This
corresponds to the noncooperative regime in Acemoglu and Pischke
(1999b). The marginal return to training is now given by
Z ∞
0 c
θ(h) = (1 − L)(1 − Q) − (γ − c (h))f (ε ) (η + D)g(η)dη. (4.100 )
−D
Note that the terms in Eq. (4.100 )also appear in (4.1). However, the
R∞
terms ((1 − L)Q + L)(γ − c0 (h)) and g(−D) εc (ε − εc )f (ε)dε appear
in Eq. (4.1), but not in (4.100 ). Recall that ((1 − L)Q + L)(γ − c0 (h))
is the net increase in the worker’s value of marginal product elsewhere
due to training weighted by the likelihood that the worker will sepa-
rate. When the employer can commit to training, the worker is willing
to accept a lower wage to get this benefit. But an employer who can-
not commit to training will not internalize this benefit to the worker.
Similarly, the employer will not take into account the benefit to the
worker from a lower dismissal probability. It follows immediately that
an inability to write a contract that specifies the amount of training
causes the employer to provide less training.
Note that the more general is the training, the greater is the adverse
impact of the employer’s inability to commit to a specified training
30 The Choice of Training
level. This has led Barron et al. (1999) to argue that employers have an
incentive to replace general training with specific training. If employers
can offer training that is not valuable elsewhere, then they will be more
willing to provide training that market frictions such as the employer’s
inability to commit to training or a liquidity constraint on the part
of young workers (which we will discuss below) prevent workers from
paying for.
We have focused on training investments by the employer, but some
authors have pointed out that workers also make on-the-job human cap-
ital investments. For example, workers may be able to improve their
productivity by investing time and effort into learning about a firm’s
unique production processes or developing relationships with customers
and co-workers. Non-verifiability by third parties will preclude employ-
ers and workers from contracting on the basis of workers’ human capital
investments. The consequent rent extraction by employers will lessen
workers’ incentive to invest in specific human capital. Arrangements
that limit rent extraction thus help encourage workers to make spe-
cific investments in human capital. Kahn and Huberman (1988) argue
that an “up or out” contract with a wage floor is one way of limiting
employer rent extraction and preserving workers’ incentive to invest in
specific human capital. Prendergast (1993) points out that employers
can limit the rent they extract from a match and thereby encourage
workers to invest in specific training if they can credibly commit to
promoting more productive workers to higher paying jobs. He argues
that this commitment is credible if the workers’ human capital invest-
ment makes them sufficiently more productive in the higher paying
job. Alternatively, an employer can run a tournament and commit to
promote a specified number of its most productive workers, although,
as Lazear (1989) notes, tournaments suffer from the disadvantage that
they discourage workers from cooperating with each other.
The same holds true for a wage floor caused by the fact that young
workers are liquidity constrained. By limiting the ability of employers
to reduce the starting wage, a wage floor that binds at the start of the
job forces employers to share the cost and return to general training –
in Bishop’s (1991) words, “general training masquerades as specific
training.” The amount of training is affected because the employer’s
inability to reduce the starting wage means that the employer no longer
internalizes the full value to the worker of a higher alternative wage.
To see this within the context of our model, write the Lagrangean
corresponding to the constrained profit maximization problem when
the employer commits to a future wage as Γ = π + λ(U − U A ) +
µ1 (w1 − wmin ), so that the first-order condition corresponding to choice
of w1 is now given by
∂Γ/∂w1 = −1 + λ + µ1 = 0, (4.2)
Comparing (4.3) and (4.1), one sees that the first-period wage floor
lowers the value the employer places on the worker’s expected second-
period income at other employers, which will cause the marginal return
to training to fall. In addition to this direct effect of the wage floor, there
is an indirect effect stemming from the induced effect on the second-
period wage and turnover. An employer will generally respond to the
first-period wage floor by lowering the second-period wage. (Recall
that the employer is maximizing expected profit subject to providing
the worker with a specified expected utility over two periods. If the
employer pays a higher first-period wage, then other things the same,
the second-period wage must fall if utility is to remain unchanged.) This
will lead to a higher quit probability, which lowers the return to human
32 The Choice of Training
Let
to obtain
Z ∞
0
θ(h) = (1 − γ + c (h)) (1 − Q (η)) g(η)dη
wmin −H−h
Z ∞
0
+ (γ − c (h)) µ2 (η)(D + η)g(η)dη. (4.5)
wmin −H−h
Since µ2 > 0, one sees from the second term in (4.5) that the second-
period wage floor raises the employer’s return to training if one ignores
turnover effects. This is the analog to Acemoglu and Pischke’s wage
compression effect in our model where the employer chooses the second-
period wage to maximize profit. The wage floor forces the employer to
pay a second-period wage that exceeds the profit-maximizing level. An
increase in productivity raises the employer’s desired wage, bringing
it closer to the wage that the employer is constrained to pay. Other
things the same, this raises the employer’s profit. However, as with the
first-period wage floor, there is an ambiguously signed indirect effect
of the wage floor due to turnover, as the wage floor leads to a positive
probability of dismissal but lowers the quit probability.
Similar comments apply to the other contracting assumptions.
For example, when the employer commits to wages and training, the
return to training can be written as θ(h) = (1 − L)(1 − Q) + ((1 − L)
R∞
Q + L)(γ − c0 (h)) + (1 − (γ − c0 (h)))g(−D) εc (ε − εc )f (ε)dε + (γ −
c0 (h))µ2 . A wage floor raises the return to training because training
raises the desired wage closer to the constrained wage. (As above, the
wage floor also indirectly affects training through a higher dismissal
probability and a lower quit probability.)
The empirical findings concerning the effects of the minimum wage
on training are mixed, with different samples and different methods
yielding different results. Neumark and Wascher (2001), using the 1983
and 1991 training supplements to the CPS find a negative effect of the
minimum wage on training. Acemoglu and Pischke (2003), using the
NLSY79, obtain effects that are not statistically significant, although
Neumark and Wascher’s point estimate falls outside the 95% confidence
interval of their preferred specification. Arulampalam et al. (2004a)
find a marginally statistically significant positive effect of the minimum
4.2. Is There Underinvestment in Training? 35
wage, using BHPS data from before and after the imposition of a min-
imum wage in the UK after a period where there was none.
Earlier we listed two reasons why training may have a smaller effect
on workers’ wages at alternative employers than on their productivity at
alternative employers. Asymmetric information may prevent employers
from fully valuing a worker’s previous training and monopsony power
may enable employers to extract rents.
39
40 Matching of High Ability, Low Turnover Workers to High Training Jobs
causing the price of the output in high (low) training jobs to fall (rise).
This in turn means that the value of marginal product, H, will fall in
high training jobs relative to low training jobs, which will cause wages
to fall in high training jobs and rise in low training jobs. In equilibrium,
H will therefore be higher in low training jobs than in high training
jobs. The inverse correlation between H and h across jobs will be just
sufficient to ensure that workers are indifferent between high and low
training jobs. Starting wages will be lower in high training jobs and
period 2 wages will be higher.
Barron et al. (1989) and Barron et al. (1999) find at most a weak
negative relationship between training and the starting wage. One con-
tributing factor behind the inability to find a negative relationship
between the starting wage and training is self-selection. As Barron
et al. (1989) note, there is good reason to believe that the cost of
training is lower for higher ability and more educated workers. Consis-
tent with this hypothesis is the finding by a number of authors that
training is positively correlated with education. The positive correlation
between education and on-the-job training was first noted by Mincer
(1962), but as discussed by Lillard and Tan (1992) and Loewenstein
and Spletzer (1999a), exists in all datasets for both formal and infor-
mal training.
Similarly, worker ability is also likely positively correlated with
training. Cognitive skills (as measured by the Armed Forces Quali-
fying Test) are found to be strongly associated with formal training
incidence by Loewenstein and Spletzer (1997) and Veum (1995). Indi-
rect evidence that education and ability are positively correlated with
training is provided by Neal (1998), who shows that turnover rates are
lower for more able, better educated workers and that this results in
part from a sorting of more able, more educated workers into differ-
ent jobs that presumably require greater investment in human capital.
(We discuss the effect of training on mobility below. Neal suggests that
more able workers are choosing jobs with a higher proportion of specific
to general human capital, but all that is really required for his find-
ings is that the proportion of skills that are specific does not decline
too rapidly as total human capital increases.) Failure to fully control
for the positive correlation between starting human capital, H, and
41
training will obscure the relationship between training and the starting
wage. (In contrast, if one compares workers in the same job, one might
expect a negative correlation between H and h. Presumably, workers
hired for the same job have similar ability. The worker with less pre-
vious training will generally need more training to get up to speed on
the job.)
The cost of training may vary systematically among employers
as well as among workers. For example, Barron et al. (1987, 1989),
Holtman and Idson (1991), Frazis et al. (1995), and Black et al. (1999)
all find that larger firms provide more training than smaller firms, sug-
gesting that their cost of training is lower. This is especially true for
formal training.
Workers in high and low training jobs are likely to differ system-
atically in ways other than just ability. In particular, as Eq. (4.1)
(or (4.100 )) indicates, the return to training is inversely related to a
worker’s quit probability; a given training investment by the employer
will earn a greater return the longer the worker stays with the firm.
Workers with lower quit probabilities will tend to be matched to posi-
tions requiring more training since firms with greater training oppor-
tunities will attempt to hire employees with low propensities to quit
and to have compensation packages and other policies that discourage
turnover.
If there is a sufficiently strong negative correlation between train-
ing and the starting wage, workers may self-select voluntarily into
high and low training jobs since low quit probability workers place
a higher value on the higher wage that an employer will offer in
period 2 than do high quit probability workers. But if the rent shar-
ing term is big enough relative to the k(h) term, then high quit
probability workers will not self select out of the high training jobs.
Instead, as noted by Barron et al. (1993) and Kuhn (1993), employ-
ers will take it upon themselves to screen out high quit probability
workers.
Of course, employers do not directly observe a worker’s quit prob-
ability, but infer it from other observable characteristics. For example,
as noted by Blau and Kahn (2000), male–female differences in the labor
market have been diminishing. However, historically women have had
42 Matching of High Ability, Low Turnover Workers to High Training Jobs
1 Besidesthe variable described in the text, the PSID contains another measure of training
based on the question “Do you feel you are learning things in your job that could lead to
a better job or promotion.” Gronau labels the former variable RQT and the latter OJT.
He interprets RQT as a measure of training acquired on previous jobs, while other PSID-
based research such as Duncan and Hoffman (1979) interprets it as training acquired (or
in the process of being acquired) on the current job. OJT is almost the same for men and
43
women; RQT has a greater effect on wages. In our discussion of Gronau’s findings below,
“training” can be viewed as being synonymous with RQT.
44 Matching of High Ability, Low Turnover Workers to High Training Jobs
where
Many of these sources of bias are relevant for other effects that
economists are interested in estimating, such as the effect of school-
ing on wages. As we shall see, these factors are frequently found to
have a large impact on estimates of the training effect. In contrast,
49
50 Estimating the Effect of Training on Wages, Productivity, and Turnover
where wit is the log real hourly wage of person i at time t, T is the
accumulated stock of hours of employer-provided training, f (·) is the
functional form for training, X is a vector of covariates including the
constant,1 and e is a residual. Details on the data and construction of
the training variable are in FL.
Column 1 of Table 6.1 shows the results of regressing log wages
on the cube root of training without taking into account any of the
2 Asnoted in footnote 1, our specification here includes interactions of other covariates with
a cubic in tenure, so the effect of tenure is adjusted for the covariates.
54 Estimating the Effect of Training on Wages, Productivity, and Turnover
formal training, the relatively low incidence (0.20 in the data used in
the table, for example) and consequent large number of observations
where the reported duration of training is zero complicates the analysis
of measurement error. Reporting no training when in reality there was
training implies a negative measurement error, while reporting training
when there was no training implies positive measurement error – gener-
ating a negative correlation between the true value of training and the
measurement error. This non-classical measurement error implies that
IV is not a consistent estimator (Frazis and Loewenstein 2003a, Kane
et al. 1999, Black et al. 2000). Moreover, even in a fairly simple model
of measurement error where the error in reporting duration conditional
on reporting incidence is uncorrelated with the true value of training,
the direction of the bias of IV is ambiguous (Frazis et al. 1996).
Frazis and Loewenstein (2003a) simplify the problem by estimating
the effect of a binary indicator of formal training. In this case, IV is
unambiguously upwardly biased, but Frazis and Loewenstein develop
a consistent generalized-method-of-moments (GMM) estimator. (Their
estimator is similar to ones developed by Black et al. (2000) and Kane
et al. (1999) where there is a second erroneous measure available.)
Applied to NLSY79 data, this estimator yields a value roughly two
times the OLS (with job fixed-effects) estimate.4 To date there have
been no satisfactory attempts to estimate the returns to training tak-
ing into account measurement error in the mixed discrete–continuous
training variable. One possible approach would be to extend the Frazis
and Loewenstein (2003a) GMM estimator along the lines of Kane
et al. (1999), whose estimator allows multiple discrete levels which could
be used to approximate the continuous training variable.
One final complication in considering measurement error in train-
ing data is the possibility of overestimating the returns to short spells.
To simplify the discussion, assume there is no error in reporting the
4 The dependent variable in Frazis and Loewenstein (2003a) analysis is wage growth.
The instruments are job reallocation rates by 2-digit industry (as a proxy for exoge-
nous turnover) and education controlling for AFQT and 1-digit occupation. Barron
et al. (1997b), using a second measure of training as an instrument, also find that the
effect of log total hours of training on productivity is more than doubled after taking
measurement error into account.
6.1. Estimating the Effect of Training on Wages 57
∂E(βi |T = T0 )
f 0 (T0 ) = E(βi |T = T0 )ϕ0 (T0 ) + ϕ(T0 ), (6.4)
∂T
5 The situation is more complicated in the multiperiod NLSY dataset, where the estimated
return g(T0 ) will partly reflect average returns and partly reflect marginal returns. FL
found that omitting observations with (within-job) accumulated training greater than
zero but less than final observed training did not appreciably change their results.
6.1. Estimating the Effect of Training on Wages 59
|T =T0 )
and which will exceed E(βi |T = T0 )ϕ0 (T0 ), if ∂E(βi∂T > 0: estimation
of ϕ0 is confounded by a composition effect stemming from the fact that
individuals with more training can be expected to have a higher return.
Thus far we have used the term “returns to training” rather loosely,
and have avoided discussing the economic significance of our estimates.
We now consider how and to what extent regressions of wages on train-
ing can be used to estimate the rate of return to the training invest-
ment. It is well known that the rate of return to formal schooling can,
under certain conditions, be determined by the coefficient on years of
schooling in a regression on log wages (Mincer 1970). By comparison,
the estimation of the rate of return to on-the-job training is compli-
cated by a variety of factors: the worker is paid during the period of
training, wages are not adjusted continuously, and the returns may be
split between the worker and the employer.6
To help fix ideas, consider a case where the worker’s wages are
adjusted at the beginning of every year and equal average productivity
during the year. If the worker is trained in the middle of the year, wages
for the year will reflect productivity before, during, and after training.
Productivity after training is higher than pre-training productivity, but
productivity during training is presumably lower or zero. Wages for the
year after training will entirely reflect post-training productivity. FL
show that under these circumstances the return to training (neglecting
direct costs) can be estimated from annually collected data by summing
the coefficients on the stocks of current training, lagged training, and
lead training.
Column 6 of Table 6.1 shows the results for adding the lead and
lagged terms. The sum of the coefficients is 0.0059, which corresponds
to an increase of 0.023 in log wages with 57 h of training. Setting a
work-year equal to 2000 h, we compute the annualized rate of return
as r = (2000)(0.023)/57, approximately 80%. This is a substantially
6 Ifwages are adjusted continuously, we would expect wages to decrease during training
spells that have not been completed. This is not borne out empirically – for example, Lynch
(1992) and Loewenstein and Spletzer (1999b) do not find significant effects of uncompleted
training spells. Given the short length of most training spells, it would require very rapid
adjustment of wages to generate any decline in wages during periods of training.
60 Estimating the Effect of Training on Wages, Productivity, and Turnover
higher rate of return than that found for schooling, which typically is
in the neighborhood of 10% or less.
This rate of return calculation neglects the direct costs of training
in the form of salaries for trainers and other expenses. SEPT95 esti-
mated that, in its sampling frame of firms with 50 or more employees,
wages and salaries of trainers, payments to outside trainers, tuition
reimbursements, and contributions to training funds totaled $300 per
employee in 1994. The survey also estimated that wages and salaries
paid to employees while in formal training totaled $224 over the period
May–October 1995 (Frazis et al., 1998). Pro-rating the wage and salary
cost of employees to a full year, the wages paid to workers receiving
training appear to account for only about 60% of the total costs of
training; other direct costs account for the remaining 40%.7 Apply-
ing this to column 6 gives a rate of return of 48%. This calculation
does not take into account the effect of promotions. However, FL con-
cluded that the most plausible estimate of the returns to formal training
was in the range of 40–50% after making a reasonable adjustment for
promotions.
High estimated returns, especially in combination with the fact that
most employees receive no formal training – for example, only 31% had
received formal training on their current job as of 1994 in our NLSY79
sample – may seem to imply a market failure in training leading to
underinvestment, as claimed by Ahlstrand et al. (2003). We have argued
that our estimates reflect the average return to training of the trained.
This interpretation implies that high estimated returns need not reflect
market failure. Untrained workers may realize much lower returns than
those obtained by workers who actually receive training. Without the
appropriate structural restrictions, it is not possible to estimate the
expected return to training of workers who do not receive training.8
7 Using firm data from Portugal, Almeida and Carneiro (2006) find that foregone production
accounts for less than 25% of training costs.
8 A reviewer argues that the estimated difference between average and marginal returns
is implausibly large if there is no market failure. In response, we note that there may
be extremely high returns to some short spells of training. For example, if an untrained
worker cannot operate a machine essential to production, the productivity of an untrained
worker assigned to the machine is zero and the return to his training is high.
6.1. Estimating the Effect of Training on Wages 61
in both papers are very imprecise, with both zero and values compa-
rable with Frazis and Loewenstein (2005) contained in the confidence
intervals.
data. Barrett and O’Connell (2001) assume that the change in human
capital between two periods is equal to the amount of training, implic-
itly ignoring turnover and depreciation. Almeida and Carneiro (2006),
who have access to turnover data and allow the average level of human
capital in a firm to decline as the result of turnover and depreciation,
explicitly assume that exiting employees have on average the same level
of human capital as employees who stay. The latter paper is the only
one we are aware of that calculates a rate of return to training from
multi-firm or industry level data. After estimating parameters of pro-
duction and training cost functions, the authors find that the average
rate of return is 24% for firms providing training; on the basis of their
observable characteristics, the average rate of return for firms not pro-
viding training is −7%.
Some researchers have used subjective measures of productivity to
estimate the effect of training. The EOPP survey contains the question
for various points in the tenure of the last employee hired, the typical
worker in that employee’s position, and in some cases a second employee
in that position. This allows examination of the effect of training on
productivity growth for an individual employee as well as comparisons
between the training and productivity of different workers in the same
job. A similar question is asked in a 1992 survey sponsored by the Small
Business Administration (SBA).
To what extent these ratings correspond to true productivity is
an obvious issue. Analyses of these data essentially assume that the
observations on the same job are the sum of a component proportional
to true productivity plus a random error. Bishop (1987, 1991) defends
this assumption that the measure is proportional to true productivity
by noting that the coefficient of variation of output observed in EOPP
is similar to the average found in a review of studies that had physical
measures of output (Schmidt and Hunter, 1983).
6.2. Estimating the Effect of Training on Productivity 65
9 Note, however, that if one includes the wage rate in a quit equation, general training
should have a positive effect on the quit probability: holding a worker’s wage constant,
he is more likely to quit the higher the wage he can command elsewhere. Similarly, if one
includes the wage rate in a layoff equation, general training should have a negative effect
on the layoff probability: holding the worker’s wage constant, the worker is less likely to be
laid off if his productivity is higher. In his early paper, Parsons (1972) makes this point.
6.3. Estimating the Effect of Training on Job Mobility 67
very low, then the value of modest amounts of training will be very
high and the employer will provide some training to all of its workers
even though there is a high probability that they will quit. Training
incidence in any period will therefore be higher the greater is turnover,
although presumably the total training that a worker receives will be
lower.
Of course, our model simplifies in assuming that a firm’s workers are
all alike ex ante and are all performing the same job. If one relaxes this
assumption, all workers will not necessarily get trained and training
incidence in any period is not necessarily higher at the high turnover
establishment. But the basic point remains: high turnover establish-
ments will have to provide training to their new workers to replace the
human capital that is lost when experienced workers quit. Consistent
with this argument, Frazis et al. (2000) find that after correcting for
tenure in data from individual employees, establishment level turnover
reduces several measures of formal training incidence and intensity
(however, the coefficient on turnover for informal training is insignifi-
cant and wrong-signed).
Turning to individual level data, Lynch (1991) finds that, for a
sample from the first few years of the NLSY79, on-the-job training
is associated with decreased mobility for women but off-the-job train-
ing with increased mobility. (The same sign pattern holds for men but
is not statistically significant.) As on-the-job training is likely to be
more firm-specific in nature than off-the-job training this finding con-
forms with theory, although it is not clear why off-the-job training
should be associated with increased mobility unless workers are invest-
ing in skills that are more useful at alternative employers. Analyzing
data obtained after the NLSY79 training questions were redesigned,
Loewenstein and Spletzer (1997) find that company training spells and
training in the form of seminars are associated with reduced job mobil-
ity, while the more general “school training” is uncorrelated with
mobility. Levine (1993) finds a negative association between a proxy
for quitting (obtained from a worker’s response about the likelihood of
looking for a new job) and various subjective measures of training in a
matched employee–employer dataset of US and Japanese manufactur-
ing firms. Interestingly, while human capital theory suggests that this
6.3. Estimating the Effect of Training on Job Mobility 69
71
72 Conclusion
73
74 References