Review of Income and Wealth

Series 47, Number 4, December 2001

ON HUMAN CAPITAL AND INDIVIDUAL CAPABILITIES

BY JOOP HARTOG
Uniûersity of Amsterdam

Starting out from a simple conceptual framework running from initial individual abilities to skills
produced in school to the utilization of these skills in the labor market, this paper surveys empirical
studies in labor economics, economics of education and occupational psychology to assess the empiri-
cal strength of the links between these sets of variables. Cognitive and non-cognitive abilities are
relevant for economic success, but make a modest contribution. Occupational psychology is comp-
lementary to economics and supports the notion of interlocking heterogeneity of individuals and jobs.

1. INTRODUCTION
Individuals’ earnings are vitally important, for the individuals themselves
and for assessing distributive justice in a society. Hence, knowing what determines
these earnings is just as important. Moreover, earnings functions are pricing func-
tions of some sort and this makes them also relevant for questions of allocative
efficiency in a market economy. Earnings functions are a worthy research target.
In the empirical economic literature on individual earnings functions, the
standard explanatory variables are education, potential experience, its square,
and gender. For some countries, race and region are commonly added. Individual
abilities are acknowledged to be important, but they are often not observed in
the dataset and only leave their trace in concern over bias in the estimates of the
remaining coefficients. With this small set of variables, the researcher is a long
shot away from the intuitive assessment of economic potential that is often made
in personal contact. If the same researcher is asked to predict future economic
success of his own students, friends or relatives, a whole series of indicators of
skills, abilities and motivations will be used. The field between the economist’s
meagre list of variables and the wide ranging intuitive assessment is taken up by
occupational and educational psychology, as a systematic investigation of individ-
ual heterogeneity in school and labor market performance. In this paper, we will
seek to find out what economists can learn from that research.
In a very simple structure laid out in the next section, the chain from individ-
ual abilities to individual earnings will be discussed: the dimensions of ability in
Section 3, their role within the model of human capital in Section 4, the pro-
duction process within schools in Section 5, modelling the link to earnings in
section 6. Section 7 considers what can be learned from the literature on personnel
selection and section 8 pulls together what has been learned from the march

Note: Work on this paper was begun for a Keynote speech at the European Association of
Labour Economists Meeting, Maastricht in 1993 and was later extended for lectures at Bar-Ilan
University, Tel Aviv and Université Panthéon-Assas, Paris II. Comments on drafts by Wim Groot,
Don Mellenbergh, Hessel Oosterbeek, Coen Teulings, Jules Theeuwes and two anonymous referees
are gratefully acknowledged. A longer version is available as a Scholar Working Paper.

515
through the empirical literature. Section 9 concludes and suggests directions for
future research.

2. A BASIC STRUCTURE
In the simplest possible framework, individuals are characterized by an
ability vector a, which is transferred into a skill vector s through an educational
production matrix E; skills are transferred into outputs q by a productivity matrix
Q, and finally, with output prices p, productivities determine standard earnings
or wages y:
(1) sGEa
(2) qGQs
(3) yGpq
So by substitution:
(4) y GpQEa
In most studies, ability and skills are left undefined, often even without bothering
about any difference. Generally accepted sharp definitions would be useful. Lack-
ing these, this paper will follow the practice of a rather loose understanding.
Ability is an indication of the potential of an individual: mental and physical
characteristics of an individual that determine potential proficiency in human
performance. Skill is the actual proficiency in a specific mental or physical
activity. Productivity is actual output of a commodity or a service. E and Q will
be discussed extensively below.
The simple linear structure of equation (4) ignores all choice and all problems
of information. But it provides an easy link with some of the (early) literature
and it fits in smoothly with the standard neoclassical view of a market economy.
Earnings follow from endowments and prices, with prices determined by scarcity.
The equation can even refer to a simple general equilibrium structure. With fixed
supply of labor by ability, and linear technologies, output supplies are fixed.
Hence, equilibrium prices are determined by the demands for the outputs: con-
sumer demand determines the value of workers’ abilities and hence, the earnings
distribution. The contribution of education to earnings inequality, at given output
prices, production technology and covariance structure of abilities, depends on
the deviation of the education production matrix E from an identity matrix.
Schooling may widen or compress ability differences among pupils. Across
schooling levels it is hard to imagine a reduction, but within schooling levels (or
schools) this is a real policy choice, as will be further discussed below. Of course,
education may also affect output prices, either through supply or through demand
effects.1

1
An obvious problem with ability and skill, and hence E, is the scale of measurement. If a and s
were scaled by the relative market prices, E could in fact be deduced (if it is square). Note that we
use ‘‘schooling’’ and ‘‘education’’ interchangeably, although schooling is a more restricted concept.
The wider meanings of education are not analyzed in this paper.

516
 With linear technologies, a multivariate normal distribution of a leads to
a normal distribution of earnings. However, this clearly clashes with observed
distributions. The clash was once known as Pigou’s Puzzle: if abilities are nor-
mally distributed,2 why is the earnings distribution skewed to the right? Pigou’s
answer was that within occupations, abilities are normally distributed; the overall
skewness arises from aggregation of occupations as non-competing groups. Roy
(1950) argued that output depends on a compound ability that results from multi-
plicative interaction of a number of elementary abilities. Champernowne (1953)
and Rutherford (1955) started from Gibrat’s Law of Proportionate Effect, where
the change of income is a random proportion of income itself. Properties of the
process of change then may generate lognormal or Paretian distributions of
income (see Mincer (1970) and Pen (1971) for discussions; for an application see
Fase (1969)). Hartog (1981) specifies a model where individuals choose to employ
only a proportion of their earnings potential; if effort and earnings potential
are independent lognormal, across individuals, the earnings distribution is also
lognormal.
Some of the simplifications in equation (4) are harmless. If the fixed linear
technologies Q and E were replaced by non-linear functional forms, nothing
essential would change. More substantial changes result because individuals do
not know their own abilities, because of uncertain technologies and because out-
puts are imperfectly observable. The ongoing revolution about the role of infor-
mation in the economy may bring useful models here, but applications to the
distribution of earnings have not yet been developed. Important qualitative differ-
ences also arise from allowing for individual self-selection and comparative
advantage in education and production. These effects will be discussed below.
Schooling is taken to be the only link from ability to productivity. This
ignores all other learning, such as learning from general life experience, learning
in the home environment and learning in the work environment from on-the-job
training and work experience. These are serious omissions, but including them
would simply be too much for a single survey. The first category is routinely
ignored in economic analyses, the second is implicitly covered in the effect of
family background on later achievements and on the third there is a substantial
literature (for reference to work by economists, see e.g. Lynch (1994) and to
work by psychologists, see e.g. Ford (1997)). So far, we have ignored individual
preferences, but they can easily be added (Hartog, 1981). Preferences are some-
what akin to attitudes and motivation, a set of variables also measured by occu-
pational psychology.

3. ABOUT a
Focusing on IQ as the single relevant measure of ability seems unduly re-
strictive. Psychologists have used four main categories of variables to classify
individuals: cognitive abilities, psychomotor abilities, personality variables and
vocational preferences. Vocational preferences (individuals’ preference ranking of

2
For cardinally measured human traits, such as length and weight, the normal distribution is
empirically established. But for ordinal variables, like IQ, normality is imposed by construction.

517
occupations) appear highly stable over age, they are reasonably effective predic-
tors of future occupational classification of persons in broad occupational families
and they have poor predictive value for performance within occupations (Peterson
and Bownas, 1982, p. 83). These results fit the models designed by economists
quite well: tastes are exogenously given, constant and independent from
proficiencies.
As to cognitive ability, Spearman in 1894 introduced the idea of a single
general ability factor g and many specific factors gi . Thorndike later denied the
existence of the g-factor and only recognized the specific factors. Thurstone’s
group factor theory claims that intelligence is made up of six to ten primary or
group factors, such as ‘‘number,’’ ‘‘verbal,’’ ‘‘space,’’ ‘‘word fluency,’’ ‘‘reason-
ing,’’ and ‘‘rote memory’. Later work has suggested that there may be some hier-
archy in the sense that there is a common factor in these primary factors that is
more primitive than these factors.3 Carroll (1993) complicated the picture by
arguing that a single general ability dominates at the top, that in the middle range
ten broad factors are relevant, and at the bottom range some 50 specific factors.
Herrnstein and Murray (1994) made the single dimensional g-factor the key vari-
able in their much debated The Bell Curûe, claiming that this factor is of overrid-
ing importance for social outcomes in the U.S., decisively more important than
family background and mostly determined by heredity. These claims are not
shared by economists (Goldberger and Manski, 1995; Heckman, 1995).4 Heck-
man argues that g is an artifact of linear correlations and that in fact it is always
possible to construct a scalar latent variable that can play the role of a single
factor in factor analysis (Heckman, 1995, p. 1105). He claims that the low contri-
bution of g in explaining test scores and wages implies a lot of room for factors
not measured by psychometric testing. But as measured examples he only specifies
education and experience.
With respect to psychomotor abilities and physical proficiencies, Peterson
and Bownas (1982) conclude from assessing empirical research that for the labor
market allocation problem, some 18 abilities are relevant. These are various types
of physical strength, flexibility, reaction time, dexterity and control.
Personality variables are used to describe an individual’s interpersonal orien-
tation, that is, perceptions and behavior among other individuals. Peterson and
Bownas concluded that a list of 15 variables, such as sociability, impulsiveness
and persistence, appears most useful. In the last decade, consensus has emerged
‘‘that there are five robust factors of personality which can serve as a meaningful
taxonomy for classifying personality attributes’’ (Barrick and Mount, 1991, p. 2).
Their ‘‘Big Five’’ have names that may fluctuate a little; Barrick and Mount call
them Extraversion, Emotional Stability, Agreeableness, Conscientiousness and
Openness to Experience. Conscientiousness is also called Dependability or Con-
formity, or Will to Achieve. Openness to Experience associates with being imagin-
ative, curious, broad-minded. Interestingly, Barrick and Mount mention that

3
In practice, this is established from factor analysis in two stages. In the first stage, primary
factors are found from raw test scores. In the second stage, the general ‘‘g-factor’’ is found from
factor analysis on the primary factor score correlations. See Welland (1976, p. 12).
4
For extensive reviews, see Devlin et al. (1997) and Arrow et al. (2000).

518
these personality dimensions are relatively independent of measures of cognitive
ability.
A good example of applied work on cognitive and psychomotor abilities is
the GATB: the General Aptitude Test Battery developed by the U.S. Employ-
ment Agency. It consists of six tests for cognitive ability (from Thurstone’s pri-
mary factors) and three for psychomotor ability, and it was used to set scoring
norms for occupations. The information is collected in the Dictionary of Occu-
pational Titles. Its usefulness has been shown in many applications (e.g. Bishop,
1989; Hartog, 1980, 1981, see Section 6.2 for results). The empirical relevance of
other ability measures than just standard IQ scores for explaining earnings vari-
ance has been demonstrated by Welland (1976) and by Thurow and Lucas (1972,
cited in Welland, 1976, p. 18); the relevance of ability and personality for job
performance, as demonstrated by occupational psychologists, will be discussed
below.

4. HUMAN CAPITAL THEORY

4.1. Internal Consistency
In the framework of equation (4), human capital theory is a bold shortcut,
focusing only on years of schooling. But there is a price to boldness. If human
capital is homogenous and can be measured in efficiency units, firms are indiffer-
ent on the size distribution of human capital. The assumption is necessary to
apply a single rental rate of human capital. It makes the firm’s demand function
for any particular level of human capital (a particular level of schooling) hori-
zontal, at the relative wage rate reflecting efficiency units. The size distribution of
human capital (the distribution of workers by schooling levels) should then be
determined by supply. But if suppliers maximize net present value, and if these
are equalized in a perfect market, suppliers are indifferent about the investment
volume in human capital. This is inconsistent as, in the words of Sattinger (1980,
p. 20), horizontal supply curves and horizontal demand curves in general do not
meet. The distribution of schooling is thus left unspecified (cf Griliches, 1977,
Section 7). Other considerations should be invoked to explain it, such as ration-
ing, ability constraints and financial constraints for individuals who want to
invest. Human capitalists commonly refer to Becker’s Woitinsky lecture for this
purpose (Becker, 1967). While this model has been revived by Card (1995) for
properly estimating the rate of return to schooling, the failure of human capital
theory as an articulate theory of the joint distribution of schooling and earnings
remains.5

4.2. Vertical Sorting

What is the role of individual ability in the human capital model? Consider
the standard model (e.g. Polachek and Siebert, 1993). An individual, facing the

5
Mincer (1958) studied the shape of the distribution of earnings, but he did so for a given distri-
bution of schooling length. Chiswick and Mincer (1972) apply the human capital model to understand
inequality in observed earnings distributions and the changes over time.

519
choice of schooling length, is assumed to maximize lifetime wealth associated with
schooling length d, Vd , defined by:
d T

(5) Vd G− 冮0
K e −rt dtC 冮We
d
d
−rt
dt

where K is out-of-pocket expenses per schooling period, r is the discount rate, Wd

is earnings obtained after d years of schooling and T is the number of years in
the labor force (i.e. until retirement). Solving the integral, maximizing with respect
to d and assuming T sufficiently large to ignore exp[−r(TAd )], we get:
∂Wd
(6) N≡ Ar(KCWd) G0
∂d
Optimum schooling length d* is determined from equating marginal cost (outlays
and earnings forgone, valued at capital cost) and marginal benefits (increased
earnings). With decreasing marginal benefits and increasing marginal (oppor-
tunity) cost, the optimum is determined at the intersection of the two curves.
Following Becker (1967), the marginal cost and marginal benefit curves (and
hence optimum schooling) may vary with ability and with ‘‘opportunity’’ (such
as family background). Let ability a affect earnings6 :
(7) Wd Gf (d, a)
Then, assuming ∂Wd兾∂aH0, abler individuals have higher opportunity costs
which works towards lower d*. Abler individuals will only choose longer school-
ing if there is some compensation somewhere. From differentiating the optimum
schooling condition we get:

∂d* ∂N
(8) sign 冢 ∂a 冣 Gsign 冢 ∂a 冣
∂N ∂2Wd ∂r ∂K ∂Wd
(9)
∂a
G
∂d ∂a
A(KCWd) Ar
∂a ∂a
C
∂a 冢 冣
With a perfect capital market (∂r兾∂a G0), a separable wage function (∂2 Wd兾
∂d ∂a G0) and direct cost independent of ability (∂K兾∂a G0), abler individuals
invest less in schooling than the less able, because it would be more costly for
them.
The purest model of human capital yields the so called Mincerian earnings
equation7 :
(10) Wd GW0 e rdCK(e rdA1)
If we let ability affect initial earnings, W0 GW0 (a), we have
∂N ∂W0 ∂W0
(11) Gr e rd Ar e rd G0
∂a ∂a ∂a

6
Note that Wd is still wages for education d, not a derivative.
7
It follows by equating Vd as in (5) with V0 : equal present values for d and 0 years of schooling.

520
In this case, ability has no effect on schooling. Everyone chooses the same
amount, because marginal cost and marginal benefit increase in the same pro-
portion if ability increases. An earnings function loglinear in years of schooling
and an ability measure like IQ, as is commonly estimated, also implies that ability
has no effect on the demand for education (unless r is affected).
Vertical sorting, where individuals of higher ability opt for longer schooling,8 will
not necessarily occur. It may depend on the type of ability. It may very well be
that those with higher levels of manual or social ability choose less schooling.
Students with commercial or artistic talent may find advanced schooling a rather
poor investment. Many successful entrepreneurs are school drop-outs (the recent
wave of new ICT entrepreneurs provides striking examples). Unfortunately, there
is no more than anecdotal evidence of such cases. There is, however, substantial
evidence of a positive correlation between length of schooling and general cogni-
tive ability as measured by IQ scores (Taubman, 1975, p. 180; Hartog, 1992, p.
189; Kodde, 1985, p. 199). Given the model as we developed it, this is only com-
patible with a lower discount rate for abler individuals or a positive cross-deriva-
tive in the earnings function.
The effect of ability on the discount rate is probably not very large, although
there may be some indirect effect through the correlation between ability and
family background (poorer family background is associated with lower ability
and higher discount rates). There may be rationing however, with the less able
simply denied access to funds such as grants, scholarships and bank loans.
A positive cross-derivative in the earnings function may arise because with
higher ability a year at school produces more human capital or because ability
has a positive cross-effect on the labor market value of human capital produced
in school. The former is implicit in the commonly adopted Ben-Porath model
(Ben-Porath, 1967, Polachek and Siebert, 1993, p. 25):
(12) It G(θ tHt)a
where It is the output of produced human capital, θ t , is the fraction of the human
capital stock devoted to human capital production, Ht is the stock of human
capital and a is an efficiency parameter reflecting ability. Ability and schooling
interact (abler individuals produce more with a given capital stock, hence get
more out of a year of schooling). To get students out of school for other reasons
than finite working life, we must have aF1. But that implies a declining rate of
growth for human capital with additional years spent in school and is incompat-
ible with the standard Mincer earnings function (requiring constant growth of
human capital for years spent in school).
Abler individuals may indeed need less time to produce a given output of
human capital. But there is very little empirical evidence to test this hypothesis.
Oosterbeek (1992b) finds that in The Netherlands students of lower ability take
more time to obtain a degree in economics than abler students.9 But he also finds
that longer duration (for a given degree) raises earnings (at a rate of return of 8

8
Horizontal sorting, by type of ability, will be discussed later.
9
Siegfried and Fels (1979, p. 955) report that in an introductory economics course, an elasticity
of achievement with respect to ability of 0.89 was found. The elasticity for effort was 0.25.

521
percent!); one may speculate that these individuals have spent their time produc-
ing valuable human capital in other ways, such as work experience or other extra-
curricular activities. First year students of higher ability tend to supply more
effort (though not significantly so) while second year students of higher ability
supply significantly less effort (Oosterbeek 1992b, p. 75; Oosterbeek, 1993).
The evidence of the effect of general cognitive ability on the earnings–school-
ing slope is conflicting, as also noted in surveys by Blaug (1976), Fägerlind (1987)
and Cawley, Heckman, and Vytlacil (2000). Positive interaction is found by Lil-
lard (1977), Oosterbeek (1993), Fägerlind (1987), Hause (1972), Welland (1976),
and Blackburn and Neumark (1993). Independence is reported by Taubman
(1975), Taubman and Wales (1973), Griliches (1976), Cawley et al. (2000), and
Ashenfelter and Rouse (2000). Hanushek (1986) summarizes his readings of the
literature: ‘‘In most studies, however, years of schooling and measures of cogni-
tive ability exhibit independent effects on earnings.’’ The conclusion should be
that the evidence is mixed. Certainly there is no unqualified support for increasing
marginal benefits for individuals of higher intellectual ability (IQ).10
From the perspective of an extended human capital theory, these are rather
unsettling results. Neither increasing marginal benefits nor decreasing marginal
cost for individuals of higher ability has convincingly been established. It may
very well be that the effect of ability materializes through a much cruder mechan-
ism of restricted access to schools based on perceived ability levels, rather than
through unrestricted individual choice of investment. In fact, this calls for better
modeling of investment opportunities facing individuals.

5. ABOUT E
5.1. The Literature
Conceptually, it is easy to think of the school as a production process, but
empirical implementation is complicated. We need measures of output, measures
of input and a perception of the transformation process. Educational psychology
has not yet been able to establish a general theory of learning. There are many
learning theories, all specific to particular situations, but there is no core theory
to flesh out (or replace) the matrix E in equation (1). Outputs are commonly
measured with tests for educational performance. At the primary and to a lesser
extent the secondary level, tests can focus on achievement in a few key areas,
such as reading, writing and arithmetic in primary education. Of course, schools
may produce a lot more than what is measured in these tests. At the more
advanced schooling levels, heterogeneity in curricula, and hence output increases
rapidly, and the production function approach to achievement scores is seldom
used (Hanushek, 1986, p. 1155). The exception, understandably, is a large litera-
ture on teaching college economics using tests for economics apprehension (Sieg-
fried and Fels, 1979). On the input side, distinguishing the role of individual
ability in the school’s contribution to student achievement is hardly feasible, as
ability cannot easily be measured other than through test scores on school related
items. Indeed, the problems of defining exactly what and how to measure has led
10
We note in Section 8 that the ability effect on wages may be sensitive to age (experience).

522
to surveys in which laments on the state of the art take up a large share of the
story (e.g., Hanushek, 1986; Wood, 1987).
Hanushek (1986) draws conclusions from the results of 147 studies of edu-
cational production functions.11 Teachers and schools appear to be dramatically
different in their effectiveness for the educational performance of students: they
have strong fixed effects. But it proves very difficult to find out what it is that
makes the difference. Hanushek concludes that teacher兾student ratios and teacher
education have no effect on student achievement. Teacher experience only has a
significant (positive) effect in one-third of the 109 studies. ‘‘There appears to be
no strong or systematic relationship between school expenditures and student
performance.’’ Family background, on the other hand, is very important in
explaining differences in achievement (for The Netherlands, see Faasse et al.
1987).
Hanushek’s conclusions were received wisdom until Krueger (2000) reana-
lyzed his data and concluded that reducing class size significantly increases test
scores. His main adjustment is a focus on studies rather than on estimates. Hanu-
shek counts every selected estimate as an observation. He has 17 observations
from single-estimate studies, but 123 observations stem from a set of only nine
studies. On the effect of class size, the Tennessee STAR experiment has generated
the methodologically most convincing approach (see Krueger, 1999). Kindergar-
ten children and their teachers were randomly assigned to classes that substan-
tially differed in size, either 13–17 students or 22–25; the class assignments were
maintained throughout the first three years of elementary school. The results
show a persistent 5–7 percentile increase in test scores for children in the small
classes, about a fifth to a quarter of the standard deviation of the average percen-
tile score.
While our conceptual framework separates the school effect on skills from
the skill effect on earnings, there are also shortcuts: estimates of school quality
on earnings. In 1996, Moffitt noted that school inputs apparently had little effect
on test scores, whereas they did significantly affect earnings. Card and Krueger
(1992) estimate the school quality effect at aggregate state level. They relate rates
of return estimated for three birth cohorts by state of birth on school input vari-
ables in the state of birth: pupil兾teacher ratio, teacher salary relative to the state
mean wage, and term length. The former two inputs have statistically and econ-
omically significant effects, the latter has not. In the 1996 REStat symposium
(Moffitt, 1996), these conclusions are challenged on several grounds. Evidence is
presented that aggregation to state level may be responsible for a good deal of
the results, suggesting an effect of state-level omitted variable bias. Heckman,
Layne-Farrar, and Todd (1996) argue that returns to education are not uniform
across the U.S. and that migration follows individuals’ comparative advantage.
They reject models with a separable uniform effect of school quality, favoring
models where region-of-birth interacts with region-of-residence through selective
migration. In particular labor markets for the unskilled are sensitive to regional
shocks. They only find evidence of school quality effects at the college level.

11
They apply to elementary and secondary public education in the U.S.

523
 Differences in length of schooling of course increase differences among indi-
viduals in levels of skill, and hence, in earnings.12 But what about inequality for
a given school level or type—are differences at entry magnified or reduced when
the individuals leave with their diploma? Hartog, Pfann, and Ridder (1989) group
individuals by the realized exit level from the school system (seven levels) and for
each group predict what they would have earned had they chosen any of the other
exit levels. The interesting conclusion is that when simulating identical exit levels
for the different groups, earnings differentials increase with exit level. Had they
all chosen higher education, the earnings differences would be greater than if they
all had chosen basic education only. In other words, at giûen reward structure
on the labor market, earnings differentials increase with schooling levels. The
plausibility of this result is easily illustrated. If randomly selected women compete
in a race over 800 meters, there will be differences at the finish. The more we
train the women, the more the initial differences will be enlarged. If we give most
training to the most gifted runners, this holds even stronger. And this is just what
appears to happen in education.
Brown and Saks (1975), however, argue that there is a choice in the relation
of training intensity to initial ability. They find clear evidence (in Michigan school
districts) that the average experience of instructional staff reduces the dispersion
and weak evidence that level and dispersion of students’ backgrounds13 increase
achievement dispersion. Schools may differ in their policies toward students:
‘‘levellers’’ versus ‘‘elitists.’’ When estimating an educational production function
this may bias the effect of technology with that of unknown preferences. Brown
and Saks (1987) empirically separate technology and teacher tastes. The learning
technology of diminishing marginal returns (smaller advances for pupils who start
out at a higher level) favors a compensation strategy of teachers.
The state of affairs in measuring education has generally been recognized as
unsatisfactory. ‘‘Frankly, I find it hard to conceive of a poorer measure of the
marketable skills a person acquires in school than the number of years he has
been able to endure a classroom-environment. My only justification for such a
crude measure is that I can find nothing better’’ (Welch, 1975, p. 67; cf also
Griliches, 1977, p. 3). Against this background, the effort made by the OECD,
jointly with Statistics Canada, to measure skills directly in the IALS project is
particularly laudable (OECD, 1995). The International Adult Literacy Survey
measures three different skills of information procesing, Prose, Document and
Quantitative Literacy, by the same method for seven countries. The results can
be used, for example, to compare the effects of schooling in different countries.
One of the interesting conclusions is that the mean skill levels do not differ very
much between developed countries, but the dispersion of skills within the popu-
lation differs widely. This is related to the result that the skill level of graduates
of a given schooling level differs markedly between countries. A low education
12
Ram (1990) analyzes the relation between mean and standard deviation of schooling years in
about 100 countries and finds a parabolic pattern: with increasing mean years of schooling in the
labor force, dispersion first increases and then, after a mean of about 7 years, decreases. Dispersion
in schooling years of course translates in earnings dispersion, but Ram has no observations on this
relation.
13
This is measured by average and standard deviation of student’s socio-economic status and by
percent white students in the district.

524
level in the United States comes with much lower levels of skills than the same
schooling level in Europe. This can explain the wider wage dispersion in the U.S.,
as a given difference in schooling simply corresponds with a greater difference in
skills (Leuven, Oosterbeek, and Van Ophem, 1997).

5.2. Horizontal Sorting

We have discussed vertical sorting as emerging from individual choice. But
there is also an important normative question: does it make sense to stratify
schools by ability level, separate schools for the dumb and for the smart? Most
countries have basic education in undifferentiated schools, while at some point
schools start to differentiate, at different ages in different countries. Optimum
conditions will not be developed in this paper, but it is clear that both learning
technology and interaction effects between students play a role. Also, the
reliability of information about individual ability will be relevant.
With more than one type of ability and skill, it makes sense to consider
horizontal sorting: will individuals specialize in developing the ability they are
most gifted with, how much ‘‘ability-type’’ differentiation should there be in the
school system? The answer hinges on the existence of comparative advantage. A
mathematically gifted individual will get a mathematically oriented education,
and a verbally gifted individual an education in languages or humanities, if each
has a comparative advantage in the direction that matches ability endowments.
This is easily demonstrated in a simple model.14
Suppose we have two types of ability, A and B say. Individual α is strongly
endowed with ability A, individual β is strongly endowed with ability B, as in
Figure 1. Wages depend simply and linearly on abilities, and if we draw an iso-
wage function, individuals α and β have the same earnings: their endowments,
prior to schooling, are on the same iso-wage function. There are two schools, or
two curriculums: one exclusively develops ability A, the other one exclusively
develops B. The dashed line indicates how far they would develop ability A in an
A-school or ability B in a B-school. The slope of the lines E indicates the ratio of
ability (or skill) development in the two school types, for individual α and indi-
vidual β . We assume comparative advantage: if you score high on ability A, you
benefit more from the A-education. Now, we get horizontal sorting if for each
individual, developing her ‘‘best ability’’ is most rewarding. It is straightforward
to show that this requires:
∂aβA ∂w兾∂aB ∂aAα
(13) Eβ ≡ F F ≡ Eα
∂aβB ∂w兾∂aA ∂aBα
The condition is feasible because by assumption E β FE α . Horizontal sorting
occurs if the wage slope, the ratio of marginal wage effects of the abilities, lies
between the two educational development ratios. With equal cost for both school
types, this is also the condition that supports the differentiation of the school
system into these two types. In this particular case, the individuals reach the same
wage level after schooling. Hence, their rate of return to schooling is identical.

14
The analysis draws on Hartog (1992, Chapter 4).

525
 Figure 1. Horizontal Sorting

But they take different educations, and if they were to switch to the other school-
ing type, they would both lose. Of course many alternatives and generalizations
are conceivable, with different rates of return to different individuals and with
less than complete specialization by school types, but they will not be developed
here (details are given in Hartog, 1992).
In practice, the extent of differentiation in the school system differs between
countries. If we include curriculum variation among school type variation this
yields a bewildering array of choice. For example, in The Netherlands, individuals
at one time could choose from at least 130 types of higher education (Kodde and
Theunissen, 1984, p. 118). The present analysis may be very relevant for the
choice between academic and vocational training, a topic that draws attention
also for its importance in organizing schooling in developing countries. Neuman
and Ziderman (1991) report that the international literature suggests that
vocational secondary education is not cost-effective relative to academic training.
However, recent U.S. studies would suggest that this changes if one allows for
proper matching between type of curriculum and type of job (e.g., those with
‘‘electricity’’ courses working as electricians). Neuman and Ziderman find similar
results for Israel. In matched occupations, vocational school graduates earn sub-
stantially more than academic school graduates. Lack of data prohibits control
for background and ability, but it would certainly be interesting to attempt such
controls.
Evidence on horizontal sorting seems to be scarce, since the question has
barely been researched. A Dutch study reports that students in secondary school
choose Latin and Greek significantly more often if in elementary school they
scored high at language, and they choose sciences significantly more often if in
elementary school they scored high at mathematics. Similar results hold for type
526
of secondary school chosen (Kodde and Theunissen, 1984). Oosterbeek, Van
Ophem, and Hartog (1993) distinguish secondary and higher education into three
types, α (language and arts), β (health, science, engineering, agriculture) and γ
(humanities, law, economics). The results clearly support vertical and horizontal
sorting.

6. ABOUT y
6.1. The Human Capital Model
In an earnings function, individual earnings can be related to innate abilities,
to augmented abilities or skills, or to schooling. In the simple framework of equa-
tions (1) to (4), these are just different ways of combining the elements. The basic
Mincerian human capital specification only uses years of schooling, and assumes
uniform marginal returns to additional schooling years. This requires very strong
assumptions. The demand side is smoothed away by the assumption of homogen-
eity of human capital, and the coefficient on schooling years (the ‘‘price’’) only
reflects the supply side: the reservation price for postponing earnings. Becker
(1967) allowed for the marginal cost and marginal benefit curves in equation (6)
to vary with individual abilities and family background. Card (1995) revived this
model as a tool to discuss (and assess) econometric issues of endogeneity, omitted
variables and measurement errors (see his survey in Card, 1999). In this approach
both the schooling function and the earnings function are random coefficient
models. Omitted variables, such as abilities, can be hidden in these random coef-
ficients. In Card’s specification, a positive endogeneity-omitted ability bias in the
OLS-estimate of the rate of return is quite plausible.15 The upward bias can be
countered however, by a negative bias from measurement error in the schooling
variable. Card assesses the latter bias at some 10 percent of the estimated mean
returns to schooling, and the former somewhat larger. Hence, on balance, the
OLS estimate of the average return to education would have a slight upward bias.
By contrast, estimates with Instrumental Variables suggest a serious underesti-
mate by OLS. One explanation that has been put forward is that IV estimates
pick out the returns for groups with low schooling, handicapped by high marginal
cost and hence, high marginal returns. For example, the instrument may be
changes in the minimum school leaving age or geographical distance to a college.
Another explanation is publication bias. If estimation results are more likely to
be published if the estimated coefficient is significant, an estimate with a high
standard error only passes the gatekeeping referees and editors if the coefficient
is high enough for the threshold t-value. Correcting for publication bias indeed
reduces the gap between IV and OLS estimates, but does not eliminate it (Ashen-
felter, Harmon, and Oosterbeek, 1999).
15
Sufficient conditions are: a positive covariance between the ability effect on earnings and the
marginal returns from schooling (cov(b, a)H0, in Card’s terms), a negative covariance between the
ability effect on earnings and the marginal cost of schooling (cov(r, a)F0) and a negative covariance
between marginal cost and marginal benefits of schooling (cov(b, r)F0).

527
6.2. The Hedonic Model
The general infeasibility of a single unit price for characteristics is a core
result in the class of hedonic models (Rosen, 1974), where equilibrium matches
between workers and jobs (firms) are tangency points between an iso-profit func-
tion and an iso-utility function.16 The market valuation of a characteristic is the
envelope that connects the common tangency points of the realized matches. The
curvature of the envelope will depend on the distributions of supply and demand.
This essential insight was the core of Tinbergen (1956) and it is elegantly set in
the frame of an Edgeworth Box by Heckman and Scheinkman (1987), who
empirically reject the hypothesis of equal prices. The absence of unit prices in the
labor market was also demonstrated in a series of papers that use job level as a
demand side characteristic. Job level is a variable that measures the complexity
of a job, often expressed as the required ability level of a worker and sometimes
as required education (Hartog, 1985, 1986a, 1986b; Bierens and Hartog, 1988;
Hartog and Bierens, 1988). Hartog (1988) shows how the marginal return to IQ
depends on job level; also, the probability of reaching a higher job level appears
highest for individuals whose earnings gain across job levels is greatest: assign-
ment follows comparative advantage.
It is not easy to estimate a full structural hedonic model (Arguea and Hsiao,
1993). Still, some structural information has been uncovered for The Netherlands,
again using job level as an important demand side variable (Van Ophem and
Hartog, 1993a, 1993b; Van Ophem, Hartog, and Vijverberg, 1993). Separability
of capabilities and job level in the wage function is decisively rejected. The mar-
ginal rate of substitution between job level and wages is increasing in both. The
higher the job level, the more individuals want to be compensated for further
increasing complexity. Similarly, at higher wage levels, individuals are more reluc-
tant to take on more demanding jobs. Hence, ‘‘leisure-on-the-job,’’ obtained by
choosing a less demanding job level, is a normal good, with a positive income
elasticity.
Teulings (1995) manages to estimate a structural model by reducing the
matching problem to a problem of matching two one-dimensional variables, as
in Sattinger (1975), by grouping variables into a skill index and a job complexity
index. He emphasizes that wages can be related to the complexity index or to the
skill index, but the indexes should not appear simultaneously in the wage func-
tion, since in equilibrium they match one-to-one. According to the results for The
Netherlands, wage differentials in the lower tail of the wage distribution are
mainly due to differences in worker skills, while in the upper tail they are mainly
due to assignment of workers to different levels of job complexity (the skill distri-
bution is skewed to the left, the high end of the wage distribution shows the
elongation of the upper part of the skill distribution to fit the upper part of the
job complexity distribution). The model has also been estimated for Portugal
(Teulings and Vieira, 1998).
Hartog (1980, 1981) estimates an earnings function following from Tinbergen
(1956), with earnings related to job requirements, using earnings by job from
16
The hedonic model is set out in standard labor economics texts like Ehrenberg and Smith (1996)
and Filer, Hamermesh, and Rees (1996).

528
the Census (1950, 1960 and 1970) and job requirements for each job from the Dic-
tionary of Occupational Titles. On a standardized scale, intellectual ability had the
highest implicit price, followed by social-commercial ability, with manual ability
having the lowest price. Between 1949 and 1959, relative prices changed, with the
intellectual price lagging behind, the social-commercial ability price increasing a
little, and the manual ability price rising substantially. Comparisons of the earnings
function with a human capital specification proved the latter to be inferior. Also,
evidence was found that the earnings function was indeed non-linear: a number
of interaction levels got significant coefficients. Hartog (1998) and Vieira (1999)
use the Tinbergen model to analyze changes in the wage structure in Portugal
Different, but related, is the earnings function allowing for under- and over-
education (Duncan and Hoffman, 1981). Based on information about the
required education for an individual’s job, the gap between actual and required
years of education is included in the earnings function, separately for positive
values (‘‘over-education’’) and negative values (‘‘under-education’’). The data
favor such an extended earnings function, implying that returns to education
depend on where in the labor market the individual ends up. A proper match
gives the highest return, over-education years are valued less and under-education
generates a penalty. The standard human capital specification, in attained edu-
cation only, is rejected. Hartog (1998, 2000) surveys the international literature
and also points out that there is kinship between the over-education earnings
function and the hedonic model, but that the specification does not directly follow
from it.
The theoretical analysis of the hedonic model has been married to the econo-
metric model of self-selection, where individual choices are highly dependent on
variables that the researcher does not observe. As a model of choice in the labor
market, it goes back to Roy (1951),17 who presented it as a choice between
hunting and fishing.18 Because of individuals’ earnings maximizing choice, we
only observe particular chunks of the joint distribution of potential productivities,
where the selection of those chunks is determined by relative sector output prices,
differences in individual mean sector outputs and by the variances and covari-
ances of outputs. The model has been analyzed extensively by Heckman and
Sedlacek (1985), Heckman and Honoré (1990) and Sattinger (1993). Analytical
results depend critically on variances and covariances of performance in the two
sectors. For empirical work, the key lesson is the selective observation of
individuals in jobs and sectors, and hence, the incomplete view on structural
parameters.
With a sufficient number of observations, one can probably show earnings
functions to have statistically significant different parameters in almost any sector
decomposition (industry, public兾private, union兾non-union, primary兾secondary,

17
Mandelbrot (1962) has developed a similar model to derive a Pareto distribution of earnings
from a multivariate Paretian distribution of abilities. Earnings result from rewards for abilities at
given prices. The ability prices differ by occupation and individuals select the occupation that pays
best.
18
Roy presented his model verbally, but apparently based himself on a fully developed statistical
model. See Sattinger (1993, note 21).

529
level of education, etc).19 The key issue is however, whether the difference is suf-
ficiently large to be economically significant. We know, for example, that wages
differ substantially between industries, in a pattern that is remarkably stable over
time and across countries (Teulings and Hartog, 1998). An interesting question
is then whether the industry effect is a multiplicative constant or varies by ability,
schooling and personality (as Heckman and Scheinkman (1987) found). It seems
also useful to make a distinction between initial entry decisions and (potential)
switches during the career. Any sector choice will tend to imply lock-in effects, in
the sense that potential experience in other sectors is lost. Modeling experience
related switching cost seems a worthwhile exercise.20
So far, the literature has not produced a fully satisfactory earnings model
that is based on a rich theory of heterogeneous individuals facing heterogeneous
jobs and that can be estimated without insurmountable difficulties. The Becker–
Card model, while fully allowing for individual heterogeneity in costs and ben-
efits, is only a partial reduced-form model. It does not specify why individual
costs of schooling differ and it is even less informative about the reasons for
differences in individual returns. With benefits restricted to earnings, the individ-
ual’s earnings simply follow from integrating marginal benefits up to optimal
schooling but applied work usually does not obey this consistency requirement
(see e.g., Blackburn and Neumark, 1995). In fact, the Becker–Card model is never
used for structural identification. The full hedonic and the Tinbergen model are
more structural, by explicitly including labor demand and labor market equilib-
rium, and stressing the strong implications of the impossibility of unbundling.
Individual choice (self selection) in a structurally specified labor market is more
easily envisaged in such a model, even though parameter estimation is not easily
accomplished. It seems then, that econometric specification and economic model-
ling have not yet successfully been matched.

7. ABILITIES AND PERSONNEL SELECTION

Job applicants are often screened by administering paper-and-pencil tests to
predict their productivity on the job. There is a large volume of empirical work
on the predictive value of such tests. For example, Schmidt et al. (1979) refer to
a study on the relation between subtests of the GATB and job performance that
is based on 367 studies published between 1950 and 1966. In fact, personnel psy-
chologists have developed a pragmatic approach to study the relation of output
q with abilities a and skills s in our simple model. As Schmidt et al. (1979) point
out, the approach goes back to work by Brogden published in 1949 (the formulae
are also presented in Cascio, 2000). Starting point is the standard linear regression
model:
(14) Ys Gβ ZsCµyCe
19
In Hartog and Van Ophem (1986) we did not find meaningful results for separate earnings
functions by job level, in a specification that included selectivity bias correction from a multinomial
logit model for job level allocation. But the number of observations within job levels was not very
large.
20
Magnac (1991) distinguishes between testing for comparative advantage and for free entry to a
sector. Both hypotheses are not rejected.

530
where Y is job performance measured in dollar value, Zs is the test performance
in standard score form in the applicant group, µy is the mean job performance in
dollar value of randomly selected employees and e is the error term. If the test
performance is normalized to unit variance, expected job performance for a selec-
ted group s is, using the formula for the regression slope:
(15) r s Grxy σ y Z
Y r xsCµy
where rxy is the correlation between dollar performance and test score where
employees are randomly selected, and σ y is the standard deviation of Y. Hence,
the gain from selection equals the productivity advantage of the selected group
over random selection:
(16) r sAµy Grxy σ y Z
Y r xs
If the test scores are normally distributed, the proportion of applicants selected
is p and F is the ordinate in the N (0, 1) distribution corresponding to p, mean
test score of those selected, Z , is F兾p and we can write21 :
(17) r sAµy Grxy σ y F兾p
Y
Schmidt et al. (1979) discuss the validity of the assumptions (linearity of the
model, normal distribution of dollar performance) at some length, drawing from
extensive empirical work. Correlation coefficients are well documented (e.g. Ghi-
selli, 1966). Schmidt et al.’s innovation is the method they present for estimating
the dispersion of output value σ y . They survey supervisors to present estimates
for the dollar value of performance of three employees: one with average perform-
ance, one performing at the top 15th percentile and one at the bottom 15th per-
centile. Taking differences, they have two estimates of σ y and they can test for
equality as required under the normal distribution assumption.22
Schmidt and Hunter (1983) report estimates of σ y as a percentage of salary
to range from 42 percent to 60 percent and advocate 40 percent as a rule of
thumb. They also present estimates of the standard deviation of output from
published research using physical counts of employee output. It turns out that
the standard deviation of output is about 15 to 22 percent of mean output, and
the estimated values are concentrated in a fairly narrow range.23 The authors
conclude that a standard deviation in output of 20 percent of mean output is a
safe rule of thumb to apply their formula (equation (17)) to assess the value of
selection methods in terms of output. The result can then be compared with the
cost of the selection method, to assess the net effect.24
Bishop (1989) shows output variability to be different by occupation. The
coefficient of variation of output among workers in a job, assessed over a period
21
Note that F兾p is the Heckman correction term in selectivity corrected regression models. It’s
the expected score conditional on scoring above the threshold set for accepting a proportion p of
applicants (cf. e.g. Sattinger, 1993).
22
Their original application was to computer programmers hired for the U.S. government, for
which they received a response from 105 supervisors. The two σ y estimates were not significantly
different. The standard errors for the estimates from the 105 supervisors were 16 and 10 percent.
23
As an interesting aside, they find that piece rate compensation systems yield a smaller disper-
sion: 15 percent of the mean, on average over the studies, compared to 18.5 percent for non-piece
rate compensation.
24
For more detailed discussion and recent developments, see Cascio (2000).

531
of a year or more, is 14 percent for laborers and operatives, 29.8 percent for sales
clerks, 33.8 percent for technicians, and 16.7 percent for routine clerical jobs.
‘‘Standard deviation of output is substantially higher in the more cognitively com-
plex and better paid jobs.’’ (Bishop, 1989, p. 18). Correlations with test scores
also vary across occupations, but not for all predictors of job performance. Infor-
mation on GATB test scores is combined into three measures, Academic Achieve-
ment, Perceptual Speed and Psychomotor Ability. The effect of Academic
Achievement is quite similar in six out of eight occupations25 : a difference of one
standard deviation in achievement has an effect of 25 to 30 percent of the stan-
dardized supervisor performance rating. It is lower for operatives and laborers
(19 percent) and much lower for sales workers, at 12 percent, where it is not even
statistically significant. The variation in the effect of psychomotor skills is modest
on average, but the extremes are far apart: 8 percent for craft workers, 17 percent
for sales clerks. The effect of Perceptual Speed covers a range from 2 percent for
technicians (not significant) to 11 percent for plant operators. Bishop calculates
that if the workers in these eight occupations had been assigned randomly, the
productivity loss would have been 8 percent of mean compensation. Assignment
based on test performance would raise output by 6.9 percent of mean compen-
sation. The magnitude of the benefits from test based job assignment is subject
to debate (Hartigan and Wigdor, 1989; Levin, 1989, 1991, 1993; Mueser and
Maloney, 1991).

8. ABILITY AND PERSONALITY : DOES IT MATTER?

Almost from the beginning of empirical work on the human capital model,
there has been much interest in the potential bias in the schooling coefficient due
to omitted ability variables. Most of this literature suggests that the schooling
coefficient would be reduced by not more than a third if ability variables such as
IQ test scores are included. The central tendency is perhaps in the order of 10 to
15 percent reduction (Welch, 1975, p. 66; Griliches, 1976, p. S78). Ashenfelter
and Rouse (2000) conclude from their data on identical twins that the ability bias
is about 25 percent (but their reported results show wide variation). Taubman,
also using observations on siblings and twins, claims that the bias of ability and
family background combined may be much larger, perhaps up to 70 percent
(Taubman, 1975, p. 297). Two points are noteworthy in this literature. First,
while the emphasis has been on estimating the upward bias in the schooling coef-
ficient when ability is omitted, it has been pointed out that the presence of
measurement errors could very well have the opposite effect (Welch, 1975, p. 67;
Griliches, 1977, p. 12). Second, Hause (1972) shows that while the ability bias
may be negligible in the first years of work experience, it may be substantial after
15 to 20 years of experience. Taubman (1975) also finds that ability effects are
more pronounced for individuals in their late 40s than in their mid-30s. This is
also relevant for the discussion in Section 4.2 on vertical sorting. If ability only
pays off later in the career, the opportunity cost of the more able students is not
higher than for the less able. Ashenfelter et al. (1999) survey studies from the

25
Plant operators, technicians, craft workers, high skill, low skill clerical and service workers.

532
1990s dealing with ability bias, schooling endogeneity and measurement error.
In a meta-analysis, allowing for publication bias (‘‘only significant results are
published’’), they find that estimated returns do not differ by estimation method,
such as OLS or Instrumental Variables. Remarkably, they find that adding ability
controls reduces estimated returns in the U.S. and increases them elsewhere.26
Welland (1976), using scores on 17 tests (relating to verbal, mathematical,
memory and visual abilities) finds that restriction to one or two composites only,
like an IQ measure, is statistically rejected. The effect of ability scores varies
across six occupational groups; ‘‘There does not appear to be one specific skill,
or set of skills, which suffices to characterize cognitive earnings capacity for all
occupations at a given level’’ (Welland, 1976, p. 102). Taubman and Wales (1974)
identify four main factors in the NBER兾Thorndike–Hagen dataset: mathematical
ability, psychomotor coordination, reading兾mechanical principles and spatial–
visual perception. But in their earnings equation, only mathematical ability has a
significant effect.27
Most authors conclude that the contribution of IQ, while significant, is not
very large. Griliches and Mason (1972) report that if schooling and family back-
ground are accounted for, adding an IQ measure does not raise R2 by more than
2 percentage points. Hause (1972), studying four different data sets, also con-
cluded that the contribution of measured ability to explaining differences in earn-
ings is modest. Welland (1976) finds that his 17 ability measures jointly reduce
the standard error of estimate in an earnings equation by 1.7 percent. Within
separate schooling groups, this increases to a maximum of 6.9 percent, within
separate cells for education and occupation it reaches a maximum reduction of
9.9 percent. Cawley et al. (1996) construct a single ability variable (‘‘Spearman’s
g’’) from principal components on a set of test scores and conclude that there is
a modest contribution to explaining wages.
The results for economic relevance tend to be somewhat stronger. Griliches
and Mason (1972), in their sample of men under 35, find that an increase in IQ
score which improves an individual’s position by 1 percentile increases earnings
by about 0.1 percent, given schooling, age and family background. Hause (1972)
reports that for low levels of schooling, ability differentials have negligible effects,
but at high levels, one standard deviation of within-sample-schooling-class meas-
ured ability raises earnings by 10–13 percent for males aged 35–40. Taubman
and Wales (1973) report the earnings differential for scoring on mathematical
ability in the highest or the lowest fifth (again given schooling, age and family
background): 17 percent when individuals are 33 years old on average, 25 percent
when they average 47 years. In Welland’s (1976) results, the change in the pre-
dicted income, for a 1 standard deviation increase in IQ or the Quantitative
Ability composite, within education groups, is 8 percent at most. If an individual
is 1 standard deviation above the mean on all 17 tests, predicted income within
occupations, for given education, may rise by as much as 20 percent (Welland,
1976, pp. 99–100).

26
See the Special issue of Labour Economics, 6(4), 1999 for IV estimates for several countries.
27
Taubman and Wales call the third factor IQ, while Thorndike has indicated his belief that the
first factor would correlate much stronger with IQ (see Taubman, 1975, p. 224).

533
 As mentioned above, personnel psychologists have collected a large amount
of empirical information on the validity of their methods. Based on these studies,
meta-analyses are applied to draw conclusions on true correlations, measurement
error, effects of ‘‘range restriction’’ (observations restricted to a selected group)
etc. Van der Maesen de Sombreff (1992) tabulates results from several meta-
analyses. Mean correlation coefficients with performance are highest for cognitive
abilities, biographical inventories, mini-courses plus test, and structured evalu-
ations of training and experience, all at about 0.50. Lowest predictive values relate
to non-structured evaluations.
Some meta-analyses evaluate the contribution of personality measures to
predict job performance (Barrick and Mount, 1991; Tett, Jackson, and Rotstein,
1991). Barrick and Mount claim progress from focusing on the ‘‘Big Five’’ men-
tioned above. Still the correlation coefficients with performance (corrected for
unreliability in the predictor and the criterion) are not impressive: 0.10 for Extra-
version, 0.07 for Emotional Stability, 0.06 for Agreeableness and even −0.03 for
Openness to Experience; Conscientiousness is an exception, with 0.23. For train-
ing proficiency, the coefficients are never lower, and in some cases substantially
higher.28 They also report correlation coefficients separately for occupational
groups, although, unfortunately they average over predicted variables (job pro-
ficiency, training proficiency and personnel data, such as tenure, turnover and
salary). This differential impact hints at comparative advantage Extraversion cor-
relates much higher for managers and sales personnel than for skilled and semi-
skilled workers (0.18; 0.15 versus 0.01). Agreeableness correlates much better for
police and managers than for professionals and (semi-)skilled (0.10 and 0.10 ver-
sus 0.02 and 0.00). But Conscientiousness correlates about equally for all five
occupational groups.29 This suggests, as Barrick and Mount initially hypothes-
ized, that some personality traits are relevant for all occupations and some are
specially relevant for a subset of occupations. Tett, Jackson, and Rothstein (1991)
report an average correlation coefficient for personality variables that is about
double the value reported by Barrick and Mount (0.24 versus 0.11). An important
reason is their use of absolute values of correlation coefficients, prohibiting the
unwarranted cancelling of positive and negative values. They have limited their
study to job proficiency only. They find correlations for the ‘‘Big Five’’ that are
certainly not negligible: 0.16 for Extraversion, 0.18 for Conscientiousness, 0.23
for Emotional Stability,30 0.27 for Openness, and 0.32 for Agreeableness (Tett et
al., 1991, Table 5, p. 726). Clearly then, personality makes some difference.

9. CONCLUSIONS
This review started from a very simplistic framework to discuss the link
between individual abilities, skills, schooling and earnings, and tried to flesh it
out with evidence from economics and psychology. It turns out that there are
important complementarities that might be exploited for mutual benefit. The
28
0.26; 0.07; 0.10; 0.25; 0.23 respectively. Results taken from Barrick and Mount (1991, Table 3,
p. 15).
29
Barrick and Mount (1991, Table 2, p. 13).
30
Actually, they report −0.23 for Neuroticism.

534
pragmatic approach of occupational psychologists to measure relevant variables
and relations should be stimulating to economists. Focusing on the empirical
relevance of abilities and personality variables the following conclusions can be
drawn:
1. The contribution of ability to earnings differentials is significant; there is
evidence that the monopoly position given by economists to IQ is unwar-
ranted: not all relevant information on ability is covered with just one
summary measure.
2. Although the contribution of abilities to earnings variance is statistically
significant, the increase in explained variance is modest. However, it
increases with the age of individuals, and samples of young workers give
an underestimate of the contribution.
3. The economic significance of abilities, in terms of their effect on predicted
earnings, is certainly not negligible.
4. Personality variables contribute to explaining output variance among
individuals in given jobs, but the contribution is modest.
5. There is clear evidence of interlocking heterogeneity: the relevance of
abilities and personality measures differs by type of occupation. This is in
line with economists’ emphasis on selection processes and comparative
advantage.
The results indicate that there is certainly scope for a richer theory of earnings
differentials between individuals. The theory may start from individual choices
made in schools, in a differentiated school system, allowing for comparative
advantage by type of ability and personality and with stable individual prefer-
ences. It would be useful to add lock-in effects, where sector choice reversal
becomes increasingly costly with growing experience in the sector. More work can
be done on the differentiation of schools, on horizontal sorting and on optimal
differentiation of the schooling system. School systems differ sometimes substan-
tially between countries; an evaluation of the consequences on the basis of inter-
national comparative analyses would be quite informative.31
A important gap is the absence of a good theory of learning. We know very
little about the connection between the school and the labor market and how
school output transforms into labor market input. What is it that is learned in
school that makes more educated workers more productive? Can we get a better
understanding of the process that turns test score achievement into productivity
in jobs? Krueger (1999) indeed applied the link of our basic model to assess the
benefits of reducing class size (estimated at a rate of return of 6 percent), by using
estimated effect of class size on basic math and reading test scores in conjunction
with test score effects on wages. But that is just an example, not an articulate
theory of the link between school and productivity. The labor market may very
well reward other characteristics than test scores, with test scores just correlated
with these characteristics.
Continued attention is needed for the role of imperfect information and the
gradual unfolding of knowledge about individuals’ capabilities as revealed in their
work experience, building on earlier work in the controversy between human

31
Interest in the topic seems to be growing; see Meier (2000).

535
capital and screening models. There are convincing arguments why individuals
reach their labor market destination only after a considerable lapse of time: in
essence, they are continually being tested and it would be remarkable indeed if
a simple battery of psychological tests could accurately predict their ultimate
destination. The older literature assigned a dominant role to stochastic processes.
That seems overdone: success in the labor market is certainly more than just
‘‘luck.’’ But no doubt, the stochastic component is important, as underlined by
the modest explanatory power of the measured variables. The challenge is in
finding the proper mix of stochastic events, imperfect information and rational
choice. Essentially that requires a combination of analyses at different levels. The
work that has been outlined here focuses on the individual, on a deterministic
chain from pupil and worker characteristics to individual earnings, conditioned
by the operation of the labor market. To move ahead, it seems most promising
to consider system characteristics of labor markets. Market imperfections due to
problems of incomplete information, specificity of skills, cost of transition
between jobs are more substantial and more consequential in some market seg-
ments than in others. These features could be analyzed to predict differences in
earnings variance between worker and job categories. That is, it would generate
a theory to predict differences in earnings variances from differences in underlying
variables that could then be combined with the deterministic models of earnings
determination surveyed here. It would make residual variance a target rather than
a measure of ignorance. Indeed, unavoidably incomplete information intrinsically
limits the precision of earnings prediction at individual levels, as this is precisely
the problem that the labor market itself also has to solve. A focus on system
determinants is then inevitable. We do not have to go back to the stochastic
models of the past. It seems much more fruitful to marry the static individual
deterministic models to search models. Postel-Vinay and Robin (2000) estimate a
search model on French data and decompose the variance of log earnings in a
firm effect, a person effect and a search friction effect. The search friction compo-
nent is smallest for the category of ‘‘executives, managers and engineers’’: 22
percent of the variance. For the other six categories, the contribution ranges
between 44 and 52 percent. Thus, system characteristics rather than individual
variables explain almost half the variance of wages! Viewed in that light we should
not be surprised that observable variables explain only a modest proportion of
individual wage variation.

REFERENCES
Arguea, N. and C. Hsiao, ‘‘Econometric Issues of Estimating Hedonic Price Functions,’’ Journal of
Econometrics, 56, 243–67, 1993.
Arrow, K., S. Durlauf, and S. Bowles, eds, Meritocracy and Inequality, Princeton University Press,
Princeton, 2000.
Ashenfelter, O., C. Harmon, and H. Oosterbeek, ‘‘A Review of estimates of the schooling兾earnings
relationship, with tests for publication bias,’’ Labour Economics, 6(4), 453–70, 1999.
Ashenfelter, O. and C. Rouse, ‘‘Schooling, Intelligence and Income in America: Cracks in the Bell
Curve,’’ in Arrow et al., eds, Meritocracy and Inequality, Princeton University Press, Princeton,
2000.
Barrick, M. and M. Mount, ‘‘The Big Five Personality Dimensions and Job Performance: A Meta-
Analysis,’’ Personnel Psychology, 44, 1–26, 1991.

536
Becker, G. S., Human Capital and the Personal Distribution of Income, W. S. Woitinsky lecture no. 1,
University of Michigan, USA, 1967.
Ben-Porath, Y., ‘‘The Production of Human Capital over the Life-cycle,’’ Journal of Political Econ-
omy, 75, 352–65, 1967.
Bierens, H. and J. Hartog, ‘‘Non-linear Regression with Discrete Explanatory Variables, with an
Application to the Earnings Function,’’ Journal of Econometrics, 38, 269–99, 1988.
Bishop, J., The Economics of Employment Testing, Cornell University Working Paper, 88-14, 1989.
Blackburn, M. and D. Neumark, ‘‘Omitted-ability Bias and the Increased Returns to Schooling,’’
Journal of Labor Economics, 11(3), 521–44, 1993.
———, ‘‘Are OLS Estimates of the Return to Schooling Biased Downward? Another look,’’ Reûiew
of Economics and Statistics, LXXVII, 217–30, 1995.
Blaug, M., ‘‘The Empirical Status of Human Capital Theory: A Slightly Jaundiced Survey,’’ Journal
of Economic Literature, XIV(3), 827–55, 1976.
Brown, C., ‘‘Estimating the Determinants of Employee Performance,’’ Journal of Human Resources,
XVII(2), 178–94, 1982.
Brown, B. and D. Saks, ‘‘The Production and Distribution of Cognitive Skills within Schools,’’ Jour-
nal of Political Economy, 83(3), 571–94, 1975.
———, ‘‘The Microeconomics of the Allocation of Teachers’ Time and Student Learning,’’ Economics
of Education Reûiew, 6(4), 319–32, 1987.
Card, D., ‘‘Earnings, Schooling and Ability Revisited,’’ Research in Labor Economics, 14, 23–48,
1995.
———, ‘‘The Causal Effect of Education on Earnings’’ in O. Ashenfelter and D. Card, eds, Handbook
of Labor Economics, Chapter 30, 1801–63, North-Holland, Amsterdam, 1999.
Card, D. and A. Krueger, ‘‘Does School Quality Matter? Returns to Education and the Characteristics
of Public Schools in the United States,’’ Journal of Political Economy, 100(1), 1–40, 1992.
Carroll, J. B., Human Cognitiûe Abilities. A Surûey of Factor-analytic Studies, Cambridge University
Press, Cambridge, 1993.
Cascio, W. F., Costing Human Resources, South-Western College Publishing, Cincinatti, OH, 2000.
Cawley, J., K. Conneely, J. Heckman, and E. Vytlacil, ‘‘Cognitive Ability, Wages and Meritocracy,’’
in Devlin et al., eds, Intelligence, Genes and Success: Scientists Respond to the Bell Curûe, Springer
Verlag, New York, 1996.
Cawley, J., J. Heckman, and E. Vytlacil, ‘‘Understanding the Role of Cognitive Ability in Accounting
for the Recent Rise in the Economic Return to Education,’’ in Arrow et al., eds, Meritocracy
and Inequality, Princeton University Press, Princeton, 2000.
Champernowne, D. G., ‘‘A Model of Income Distribution,’’ Economic Journal, 63, 318–51, 1953.
Chiswick, B. R. and J. Mincer, ‘‘Time Series Changes in Personal Income Inequality in the United
States from 1939, with Projections to 1985,’’ Journal of Political Economy, 80(3), part II, 534–
66, 1972.
Devlin, B., S. Fienberg, D. Resnick, and K. Roeder, eds, Intelligence, Genes and Success: Scientists
Respond to the Bell Curûe, Springer Verlag, New York, 1997.
Duncan, G. J. and S. D. Hoffman, ‘‘The Incidence and Wage Effects of Overeducation,’’ Economics
of Education Reûiew, 1(1), 75–86, 1981.
Ehrenberg, R. G. and R. S. Smith, Modern Labor Economics, Addison-Wesley, Reading, MA, 1996.
Epple, D., ‘‘Hedonic Prices and Implicit Markets: Estimating Demand and Supply Functions for
Differentiated Products,’’ Journal of Political Economy, 98(1), 59–80, 1987.
Faasse, J., B. Bakker, J. Dronkers, and H. Schijf, ‘‘The Impact of Educational Reform: Empirical
Evidence from Two Dutch Generations,’’ Comparatiûe Education, 23, 261–78, 1987.
Fägerlind, I., ‘‘Ability: Effects on Earnings,’’ in G. Psacharopoulos, Economics of Education, Research
and Studies, Pergamon Press, Oxford, 285–6, 1987.
Fase, M. M. G., An Econometric Model of Age–Income Profiles,’’ unpublished Ph.D. thesis, University
of Amsterdam, 1969.
Filer, R. K., D. Hamermesh, and A. Rees, The Economics of Work and Pay, Harper and Row, New
York, 1996.
Ford, J. K., ed., Improûing Training Effectiûes in Work Organizations, Lawrence Erlbaum, Mahwah,
NJ, 1997.
Ghiselli, E. E., The Validity of Occupational Aptitude Tests, Wiley and Sons, New York, 1966.
Goldberger, A. S. and C. Manski, ‘‘Review Article: The Bell Curûe by Herrnstein and Murray,’’
Journal of Economic Literature, XXXIII(2), 762–76, 1995.
Griliches, Z., ‘‘Wages of Very Young Men,’’ Journal of Political Economy, 84(2), S69–86, 1976.
———, ‘‘Estimating the Returns to Schooling: Some Econometric Problems, Econometrica, 45(1), 1–
22, 1977.
Griliches, Z. and W. M. Mason, ‘‘Education, Income and Ability,’’ Journal of Political Economy,
80(3II), S74–103, 1972.

537
Hanushek, E., ‘‘The Economics of Schooling: Production and Efficiency in the Public Schools,’’ Jour-
nal of Economic Literature, XXIV (3), 1141–77, 1986.
Hartigan, J. and A. Wigdor, eds., Fairness in Employment Testing, National Academy Press, Wash-
ington, DC, 1989.
Hartog, J., ‘‘Earnings and Capability Requirements,’’ Reûiew of Economics and Statistics, LXII(2),
230–40, 1980.
———, Personal Income Distribution: A Multicapability Theory, Martinus Nijhoff, Boston, 1981.
———, ‘‘Earnings Functions: Testing for the Demand Side,’’ Economics Letters, 19(3), 281–5, 1985.
———, ‘‘Earnings Functions: Beyond Human Capital,’’ Applied Economics, 18(12), 1291–309, 1986a.
———, ‘‘Allocation and the Earnings Function,’’ Empirical Economics, 11(2), 97–110, 1986b.
———, ‘‘An Ordered Response Model for Allocation and Earnings,’’ Kyklos, 41(1), 112–41, 1988.
———, Capabilities, Allocation and Earnings, Kluwer, Boston, 1992.
———, ‘‘Wandering Along the Hills of ORU Land,’’ Keynote speech for the Conference of the
Applied Econometrics Association, Maastricht, May 15–17, 1997; published in J. A. M. Heijke
and J. Muysken, eds, Education, Training and Employment in the Knowledge Based Economy,
Macmillan, London, 1998.
———, ‘‘Over-education and Earnings: Where Are We, Where Should We Go?,’’ Economics of Edu-
cation Reûiew, Special Issue on Over-education, 19(2), 131–47, 2000.
Hartog, J. and H. Bierens, ‘‘Estimating a Hedonic Earnings Function with a Non-parametric
Method,’’ Empirical Economics, 13, 267–94, 1988.
Hartog, J., G. Pfann, and G. Ridder, ‘‘(Non-)Graduation and the Earnings Function, An Inquiry on
Self-selection,’’ European Economic Reûiew, 33, 1373–95, 1989.
Hartog, J. and J. C. van Ophem, Allocation, Earnings and Efficiency, Research Memorandum 8619,
Universiteit van Amsterdam-FEW, 1986.
Hause, J. C., ‘‘Earnings Profile: Ability and Schooling,’’ Journal of Political Economy, 80(3II), S108–
138, 1972.
Heckman, J., ‘‘Lessons from the Bell Curve,’’ Journal of Political Economy, 103(5), 1091–120, 1995.
Heckman, J. and B. Honoré, ‘‘The Empirical Content of the Roy Model,’’ Econometrica, 58(5), 1121–
49, 1990.
Heckman, J., A. Layne-Farrar, and P. Todd, ‘‘Human Capital Pricing Equations with an Application
to Estimating the Effect of Schooling Quality on Earnings,’’ Reûiew of Economics and Statistics,
78(4), 562–610, 1996.
Heckman, J. and J. Scheinkman, ‘‘The Importance of Bundling in a Gorman–Lancaster model of
Earnings,’’ Reûiew of Economic Studies, LIV, 243–55, 1987.
Heckman, J. and G. Sedlacek, ‘‘Heterogeneity, Aggregation and Market Wage Functions: An Empiri-
cal Model of Self-selection in the Labor Market,’’ Journal of Political Economy, 93(6), 1077–125,
1985.
Herrnstein, R. J. and C. Murray, The Bell Curûe, Free Press, New York, 1994.
Kodde, D., ‘‘Microeconomic Analysis of Demand for Education, Ph.D. thesis, Erasmus University
Rotterdam, 1985.
Kodde, D. and M. Theunissen, De schoolloopbaan ûan de snelle VWO-leerlingen, Economisch Instit-
uut, Katholieke Universiteit Nijmegen, 1984.
Krueger, A., ‘‘Experimental Estimates of Education Production Functions,’’ Quarterly Journal of
Economics, 114(2), 497–532, 1999.
———, Economic Considerations and Class Size, Working paper 447, Princeton University Industrial
Relations Section, presented at the IIPF Meetings, Sevilla, August, 2000.
Leuven, E., H. Oosterbeek, and H. van Ophem, Explaining International Differences in Male Wage
Inequality by Differences in Supply and Demand, FEE Universiteit van Amsterdam, 1997.
Levin, H. M., ‘‘Ability Testing for Job Selection: Are the Economic Claims Justified?,’’ in B. Gifford,
eds, Test Policy and the Policies of Opportunity and Allocation: The Workplace and the Law,
Kluwer Academic Publishers, Boston, 1989.
———, ‘‘Review of: J. A. Hartigan and A. K. Wigdor, eds, Fairness in Employment Testing: Validity
Generalization, Minority Issues and the General Aptitude Test Battery,’’ Journal of Educational
Measurement, 28(4), 358–63, 1991.
———, ‘‘Education and Jobs: A Proactive View,’’ paper for the Education and Work Conference
(Toronto), Stanford University, 1993.
Lillard, L., ‘‘Inequality: Earnings versus Human Wealth,’’ American Economic Reûiew, 67(2), 42–53,
1977.
Lynch, L. M., ed., Training and the Priûate Sector, University of Chicago Press, Chicago, 1994.
van der Maesen de Sombreff, P., ‘‘Het rendement van personeelsselectie,’’ unpublished Ph.D. thesis,
Rijksuniversiteit Groningen, 1992.
Magnac, Th., ‘‘Segmented or Competitive Labor Market?,’’ Econometrica, 59(1), 165–88, 1991.

538
Mandelbrot, B., ‘‘Paretian Distributions and Income Maximization,’’ Quarterly Journal of Economics,
1962.
Meier, V., Choosing Between School Systems, CESifo Working Paper 389, Munich, 2000.
Mincer, J., ‘‘Investments in Human Capital and Personal Income Distribution,’’ Journal of Political
Economy, 66, 281–302, 1958.
———, ‘‘Distribution of Labor Incomes,’’ Journal of Economic Literature, VIII(1), 1–26, 1970.
Moffitt, R., ‘‘Symposium on School Quality and Educational Outcomes: Introduction,’’ Reûiew of
Economics and Statistics, 78(4), 559–561, 1996.
Mueser, P. and T. Maloney, ‘‘Ability, Human Capital and Employer Screening: Reconciling Labor
Market Behavior with Studies of Employers Productivity,’’ Southern Economic Journal, 57(3),
767–89, 1991.
Neuman, S. and A. Ziderman, ‘‘Vocational Schooling, Occupational Matching and Labor Market
Earnings in Israel,’’ Journal of Human Resources, XXVI(2), 256–81, 1991.
OECD兾Statistics Canada, Literacy, Economy and Society, OECD, Paris, 1995.
Oosterbeek, H., ‘‘Study Duration and Earnings,’’ Economics Letters, 40, 223–8, 1992a.
———, ‘‘Essays on Human Capital Theory,’’ Ph.D. dissertation, University of Amsterdam, 1992b.
———, Choosing the Optimum Mix of Duration and Effort in Education, Working Paper, Faculty of
Economics and Econometrics, University of Amsterdam, 1993.
Oosterbeek, H., H. van Ophem, and J. Hartog, Sorting in Higher Education, Working Paper, Faculty
of Economics and Econometrics, University of Amsterdam, 1993.
Pen, J., Income Distribution, Allen Lane, London, 1971.
Peterson, N. and D. Bownas, ‘‘Skills, Task Structure and Performance Acquisition,’’ in M. Dunnette
and E. Fleishman, eds, Human Capability Assessment, Laurence Erlbaum, Hillsdale, NJ, 1982.
Polachek, S. and W. Siebert, The Economics of Earnings, Cambridge University Press, Cambridge,
1993.
Postel-Vinay, F. and J.-M. Robin, ‘‘Wage Dispersion with Heterogeneous Firm Technologies and
Worker Abilities: An Equilibrium Search Model for Matched Employer–Employee Data,’’ Paper
presented at the Conference Assignment and matching models in economics, KNAW兾Tinbergen
Institute, Amsterdam, December 14–16, 2000.
Ram, R., ‘‘Educational Expansion and Schooling Inequality: International Evidence and Some Impli-
cations,’’ Reûiew of Economics and Statistics, 72, 266–74, 1990.
Rosen, S., ‘‘Hedonic Prices and Implicit Markets: Product Differentiation in Pure Competition,’’
Journal of Political Economy, 82(1), 34–55, 1974.
Roy, A., ‘‘The Distribution of Earnings and of Individual Output,’’ Economic Journal, 6a, 489–503,
1950.
———, ‘‘Some Thoughts on the Distribution of Earnings,’’ Oxford Economic Papers, 3, 135–46, 1951.
Rutherford, R. S. G., ‘‘Income Distributions: A New Model,’’ Econometrica, 23, 277–94, 1955.
Sattinger, M., ‘‘Comparative Advantage and the Distributions of Earnings and Abilities,’’ Econo-
metrica, 43(3), 455–68, 1975.
———, Capital and the Distribution of Labor Earnings, North-Holland, Amsterdam, 1980.
———, ‘‘Assignment Models of the Distribution of Earnings,’’ Journal of Economic Literature,
XXXI(2), 851–80, 1993.
Schmidt, F. and J. Hunter, ‘‘Individual Differences in Productivity: An Empirical Test of Estimates
Derived from Studies of Selection Procedure Utility,’’ Journal of Applied Psychology, 68, 407–
14, 1983.
Schmidt, F., J. Hunter, B. McKenzie, and T. Muldrow, ‘‘Impact of Valid Selection Procedures on
Work-force Productivity,’’ Journal of Applied Psychology, 64, 609–26, 1979.
Siegfried, J. and R. Fels, ‘‘Research on Teaching College Economics: A Survey,’’ Journal of Economic
Literature, XVII(3), 923–69, 1979.
Taubman, P. J., Sources of Inequality in Earnings, North-Holland, Amsterdam, 1975.
Taubman, P. and T. Wales, ‘‘Higher Education, Mental Ability and Screening,’ Journal of Political
Economy, LXXXI(1), 28–55, 1973.
———, Higher Education and Earnings, McGraw-Hill, New York, 1974.
Tett, P., D. Jackson, and M. Rothstein, ‘‘Personality Measures as Predictors of Job Performance: A
Meta-analytic Review,’’ Personnel Psychology, 44, 709–42, 1991.
Teulings, C., ‘‘The Wage Distribution in a Model of Matching Between Skills and Jobs,’’ Journal of
Political Economy, 103(2), 280–315, 1995.
Teulings, C. and J. Hartog, Corporatism or Competition?, Cambridge University Press, Cambridge,
1998.
Teulings, C. and J. Vieira, Urban ûersus Rural Returns to Human Capital in Portugal: A Cook-book
Recipe for Applying Assignment Models, Amsterdam兾Rotterdam: Tinbergen Institute Discussion
Paper 98-095兾3, 1998.

539
Thurow, L. C. and R. E. B. Lucas, The American Distribution of Income: A Structural Problem, Study
for the Joint Economic Committee, US Government Printing Office, US Congress, Washington,
1972.
Tinbergen, J., ‘‘On the Theory of Income Distribution,’’ Weltwirtschaftliches Archiû, 77, 155–75, 1956.
Van Ophem, H. and J. Hartog, ‘‘Towards Disentangling Supply and Demand in a Model of Wages
and Job Level,’’ in J. van Ours, G. Pfann, and G. Ridder, eds, Labour Demand and Equilibrium
Wage Formation, North-Holland, Amsterdam, 1993a.
———, Job Complexity and Wages: Confronting Supply and Demand, Working Paper, Faculty of
Economics and Econometrics, University of Amsterdam, 1993b.
Van Ophem, J., J. Hartog, and W. Vijverberg, ‘‘Job Complexity and Wages,’’ International Economic
Reûiew, 34(4), 853–72, 1993.
Vieira, J. C., ‘‘The Evolution of Wage Structures in Portugal, 1982–1992,’’ Ph.D. thesis, University
of Amsterdam, 1999.
Welch, F., ‘‘Linear Synthesis of Skill Distribution,’’ Journal of Human Resources, IV(3), 311–27, 1969.
———, ‘‘Human Capital Theory: Education, Discrimination and Life Cycles,’’ American Economic
Reûiew, Papers and Proceedings, LXV(3), 63–73, 1975.
Welland, J., ‘‘An Investigation of the Role of Abilities in the Individual Earnings Function,’’ Ph.D.
thesis, University of Minnesota, 1976.
Wood, R., ‘‘Achievement Tests,’’ in G. Psacharopoulos, ed., Economics of Education, Research and
Studies, Pergamon Press, Oxford, 59–62, 1987.

540